Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2020 Arm Limited. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 24 | #include "gemm_helpers.h" |
Vidhya Sudhan Loganathan | 17b0f8b | 2019-01-08 12:17:03 +0000 | [diff] [blame] | 25 | #include "repeat.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 26 | |
Gian Marco Iodice | 73cdaac | 2020-08-10 21:44:14 +0100 | [diff] [blame^] | 27 | #if defined(M0) && defined(K0) && defined(V0) && defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(PARTIAL_LOAD_M0) && defined(PARTIAL_LOAD_K0) |
Gian Marco Iodice | b87b95e | 2019-01-21 17:14:31 +0000 | [diff] [blame] | 28 | #define INC2 (VEC_DATA_TYPE(uint, 2))(0, 1) |
| 29 | #define INC3 (VEC_DATA_TYPE(uint, 3))(0, 1, 2) |
| 30 | #define INC4 (VEC_DATA_TYPE(uint, 4))(0, 1, 2, 3) |
| 31 | #define INC8 (VEC_DATA_TYPE(uint, 8))(0, 1, 2, 3, 4, 5, 6, 7) |
| 32 | #define INC16 (VEC_DATA_TYPE(uint, 16))(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15) |
| 33 | #define CONCAT_INC(K0) INC##K0 |
| 34 | #define INC(K0) CONCAT_INC(K0) |
| 35 | |
| 36 | #if(SRC_WIDTH % K0) |
| 37 | #define BOUNDARY_CONDITION_X(x, a) \ |
| 38 | ({ \ |
| 39 | a = select(0, a, CONVERT(((x * (VEC_DATA_TYPE(uint, K0))K0 + INC(K0)) < (VEC_DATA_TYPE(uint, K0))SRC_WIDTH), VEC_DATA_TYPE(DATA_TYPE, K0))); \ |
| 40 | }) |
| 41 | #else // (SRC_WIDTH % K0) |
| 42 | #define BOUNDARY_CONDITION_X(x, a) \ |
| 43 | ({}) |
| 44 | #endif // (SRC_WIDTH % K0) |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 45 | |
Gian Marco Iodice | 73cdaac | 2020-08-10 21:44:14 +0100 | [diff] [blame^] | 46 | #define LOAD_TENSOR_BOUNDARY_AWARE_M0XK0(M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \ |
| 47 | ({ \ |
| 48 | if(y * M0 + M0 >= SRC_HEIGHT && PARTIAL_LOAD_M0 != 0) \ |
| 49 | { \ |
| 50 | if(x * K0 + K0 >= SRC_WIDTH && (PARTIAL_LOAD_K0 != 0)) \ |
| 51 | { \ |
| 52 | LOAD_TENSOR_M0XN0(PARTIAL_LOAD_M0, PARTIAL_LOAD_K0, DATA_TYPE, a, input_ptr, src_stride_y, zin); \ |
| 53 | } \ |
| 54 | else \ |
| 55 | { \ |
| 56 | LOAD_TENSOR_M0XN0(PARTIAL_LOAD_M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin); \ |
| 57 | } \ |
| 58 | } \ |
| 59 | else \ |
| 60 | { \ |
| 61 | if(x * K0 + K0 >= SRC_WIDTH && (PARTIAL_LOAD_K0 != 0)) \ |
| 62 | { \ |
| 63 | LOAD_TENSOR_M0XN0(M0, PARTIAL_LOAD_K0, DATA_TYPE, a, input_ptr, src_stride_y, zin); \ |
| 64 | } \ |
| 65 | else \ |
| 66 | { \ |
| 67 | LOAD_TENSOR_M0XN0(M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin); \ |
| 68 | } \ |
| 69 | } \ |
| 70 | }) |
| 71 | |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 72 | /** This OpenCL kernel reshapes the lhs input matrix. The kernel splits the input matrix in blocks of size M0xK0 and stores each one (not transposed) in |
| 73 | * the output matrix unrolling the values. |
| 74 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 75 | * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) |
| 76 | * @note The width of the input tensor must be passed at compile time using -DSRC_WIDTH (e.g. -DSRC_WIDTH=16) |
Gian Marco Iodice | 73cdaac | 2020-08-10 21:44:14 +0100 | [diff] [blame^] | 77 | * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 78 | * @note The block's dimensions (M0 and K0) must be passed at compile time using -DM0 and -DK0 (e.g. -DM0=2, -DK0=2). |
| 79 | * @note The number of M0xK0 vertical blocks to store on the same output row must be passed at compile time using -DV0 (e.g. -DV0=2) |
Gian Marco Iodice | 73cdaac | 2020-08-10 21:44:14 +0100 | [diff] [blame^] | 80 | * @note The size of the partial load block in y must be passed at compile time using -DPARTIAL_LOAD_M0 (e.g. -DPARTIAL_LOAD_M0=1) |
| 81 | * @note The size of the partial load block in x must be passed at compile time using -DPARTIAL_LOAD_K0 (e.g. -DPARTIAL_LOAD_K0=1) |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 82 | * @note Only the following values for M0, K0 and V0 are supported: |
| 83 | * M0: 2,3,4,5,6,7,8 |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 84 | * K0: 2,3,4,8,16 |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 85 | * V0: greater than 0 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 86 | * @note In case the input has to be reinterpreted as a 3D tensor (e.g. input of convolution layer 1x1), the following information must be passed at compile time: |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 87 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
| 88 | * -# HEIGHT_GEMM3D: The height of the input in case it has to be reinterpreted as a 3D tensor. |
| 89 | * -# DEPTH_GEMM3D: The depth of the input in case it has to be reinterpreted as a 3D tensor |
| 90 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 91 | * @note If the M0xK0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time. |
| 92 | * |
Michele Di Giorgio | f6f7876 | 2020-07-06 11:27:21 +0100 | [diff] [blame] | 93 | * @param[in] src_ptr Pointer to the source LHS tensor. Supported data types: All |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 94 | * @param[in] src_stride_x Stride of the source LHS tensor in X dimension (in bytes) |
| 95 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 96 | * @param[in] src_stride_y Stride of the source LHS tensor in Y dimension (in bytes) |
| 97 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 98 | * @param[in] src_stride_z Stride of the source LHS tensor in Z dimension (in bytes) |
| 99 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 100 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source LHS tensor |
| 101 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr |
| 102 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 103 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 104 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 105 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 106 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 107 | * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
| 108 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 109 | * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) |
| 110 | */ |
| 111 | __kernel void gemm_reshape_lhs_matrix_nt(TENSOR3D_DECLARATION(src), |
| 112 | TENSOR3D_DECLARATION(dst) |
| 113 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 114 | , |
| 115 | uint cross_plane_pad |
| 116 | #endif // REINTERPRET_INPUT_AS_3D |
| 117 | ) |
| 118 | { |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 119 | // Block size |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 120 | #define BLOCK_SIZE ((M0) * (K0)) |
| 121 | |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 122 | // Output offset X |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 123 | #if defined(INTERLEAVE) |
| 124 | #define OUTPUT_OFFSET_X (K0) |
| 125 | #else // defined(INTERLEAVE) |
| 126 | #define OUTPUT_OFFSET_X (BLOCK_SIZE) |
| 127 | #endif // defined(INTERLEAVE) |
| 128 | |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 129 | // Output step X |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 130 | #if defined(INTERLEAVE) |
| 131 | #define OUTPUT_STEP_X (K0) * (V0) |
| 132 | #else // Do not interleave |
| 133 | #define OUTPUT_STEP_X (K0) |
| 134 | #endif // defined(INTERLEAVE) |
| 135 | |
| 136 | // Compute source and destination addresses |
| 137 | uint x = get_global_id(0); |
| 138 | uint y = get_global_id(1); |
| 139 | uint z = get_global_id(2); |
| 140 | |
| 141 | // ------------------ Compute input/output addresses --------------------------- |
| 142 | |
| 143 | // Compute the input address |
| 144 | __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)K0 * sizeof(DATA_TYPE) + y * (uint)M0 * src_stride_y; |
| 145 | |
| 146 | // Compute the output address |
| 147 | __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)BLOCK_SIZE * (uint)V0 * sizeof(DATA_TYPE)) + ((y / (uint)V0) * (uint)dst_stride_y) + ((y % V0) * |
| 148 | (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE)); |
| 149 | |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 150 | // Create variables: uint zin0=0, zin1=0, zin2=0...zin(M0-1)=0; |
| 151 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zin, 0); |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 152 | |
| 153 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 154 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 155 | // multiply src_stride_z by DEPTH_GEMM3D |
| 156 | |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 157 | input_ptr += z * (uint)src_stride_z * DEPTH_GEMM3D; |
| 158 | |
| 159 | // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 160 | CALCULATE_Z_OFFSET(M0, uint, zin, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, cross_plane_pad, src_stride_y); |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 161 | |
| 162 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 163 | |
| 164 | input_ptr += z * (uint)src_stride_z; |
| 165 | |
| 166 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 167 | |
| 168 | // Add offset for batched GEMM |
| 169 | output_ptr += z * (uint)dst_stride_z; |
| 170 | |
| 171 | // ---------------------------Load input values -------------------------------- |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 172 | // Load values from the LHS matrix |
Gian Marco Iodice | 73cdaac | 2020-08-10 21:44:14 +0100 | [diff] [blame^] | 173 | REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, K0), a, 0); |
| 174 | |
| 175 | LOAD_TENSOR_BOUNDARY_AWARE_M0XK0(M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin); |
| 176 | |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 177 | // ---------------------------Store output values ------------------------------ |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 178 | REPEAT_VAR_INIT_TO_CONST(16, uint, zout, 0); |
| 179 | STORE_BLOCK(M0, K0, DATA_TYPE, a, output_ptr, OUTPUT_STEP_X * sizeof(DATA_TYPE), zout); |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 180 | |
| 181 | #undef BLOCK_SIZE |
| 182 | #undef OUTPUT_OFFSET_X |
| 183 | #undef OUTPUT_STEP_X |
| 184 | } |
Gian Marco Iodice | 08ddd7b | 2018-12-19 10:01:18 +0000 | [diff] [blame] | 185 | |
| 186 | #if M0 == 2 |
| 187 | #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ |
| 188 | ({ \ |
| 189 | VEC_DATA_TYPE(DATA_TYPE, M0) \ |
| 190 | res = (VEC_DATA_TYPE(DATA_TYPE, M0))(a0.s##i, a1.s##i); \ |
| 191 | VSTORE(M0) \ |
| 192 | (res, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ |
| 193 | }) |
| 194 | #elif M0 == 3 // M0 == 3 |
| 195 | #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ |
| 196 | ({ \ |
| 197 | VEC_DATA_TYPE(DATA_TYPE, M0) \ |
| 198 | res = (VEC_DATA_TYPE(DATA_TYPE, M0))(a0.s##i, a1.s##i, a2.s##i); \ |
| 199 | VSTORE(M0) \ |
| 200 | (res, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ |
| 201 | }) |
| 202 | #elif M0 == 4 // M0 == 4 |
| 203 | #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ |
| 204 | ({ \ |
| 205 | VEC_DATA_TYPE(DATA_TYPE, M0) \ |
| 206 | res = (VEC_DATA_TYPE(DATA_TYPE, M0))(a0.s##i, a1.s##i, a2.s##i, a3.s##i); \ |
| 207 | VSTORE(M0) \ |
| 208 | (res, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ |
| 209 | }) |
| 210 | #elif M0 == 5 // M0 == 5 |
| 211 | #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ |
| 212 | ({ \ |
| 213 | VEC_DATA_TYPE(DATA_TYPE, 4) \ |
| 214 | res0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(a0.s##i, a1.s##i, a2.s##i, a3.s##i); \ |
| 215 | DATA_TYPE res1 = a4.s##i; \ |
| 216 | VSTORE(4) \ |
| 217 | (res0, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ |
| 218 | *((__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE)) + 4) = res1; \ |
| 219 | }) |
| 220 | #elif M0 == 6 // M0 == 6 |
| 221 | #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ |
| 222 | ({ \ |
| 223 | VEC_DATA_TYPE(DATA_TYPE, 4) \ |
| 224 | res0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(a0.s##i, a1.s##i, a2.s##i, a3.s##i); \ |
| 225 | VEC_DATA_TYPE(DATA_TYPE, 2) \ |
| 226 | res1 = (VEC_DATA_TYPE(DATA_TYPE, 2))(a4.s##i, a5.s##i); \ |
| 227 | VSTORE(4) \ |
| 228 | (res0, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ |
| 229 | VSTORE(2) \ |
| 230 | (res1, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE)) + 4); \ |
| 231 | }) |
| 232 | #elif M0 == 7 // M0 == 7 |
| 233 | #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ |
| 234 | ({ \ |
| 235 | VEC_DATA_TYPE(DATA_TYPE, 4) \ |
| 236 | res0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(a0.s##i, a1.s##i, a2.s##i, a3.s##i); \ |
| 237 | VEC_DATA_TYPE(DATA_TYPE, 3) \ |
| 238 | res1 = (VEC_DATA_TYPE(DATA_TYPE, 3))(a4.s##i, a5.s##i, a6.s##i); \ |
| 239 | VSTORE(4) \ |
| 240 | (res0, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ |
| 241 | VSTORE(3) \ |
| 242 | (res1, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE)) + 4); \ |
| 243 | }) |
| 244 | #elif M0 == 8 // M0 == 8 |
| 245 | #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ |
| 246 | ({ \ |
| 247 | VEC_DATA_TYPE(DATA_TYPE, M0) \ |
| 248 | res = (VEC_DATA_TYPE(DATA_TYPE, M0))(a0.s##i, a1.s##i, a2.s##i, a3.s##i, a4.s##i, a5.s##i, a6.s##i, a7.s##i); \ |
| 249 | VSTORE(M0) \ |
| 250 | (res, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ |
| 251 | }) |
| 252 | #else // M0 not supported |
| 253 | #error "M0 value not supported" |
| 254 | #endif // N0 conditions |
| 255 | |
| 256 | /** This OpenCL kernel reshapes the lhs input matrix. The kernel splits the input matrix in blocks of size M0xK0 and stores each one (transposed) in |
| 257 | * the output matrix unrolling the values. |
| 258 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 259 | * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) |
| 260 | * @note The width of the input tensor must be passed at compile time using -DSRC_WIDTH (e.g. -DSRC_WIDTH=16) |
Gian Marco Iodice | 73cdaac | 2020-08-10 21:44:14 +0100 | [diff] [blame^] | 261 | * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 262 | * @note The block's dimensions (M0 and K0) must be passed at compile time using -DM0 and -DK0 (e.g. -DM0=2, -DK0=2). |
| 263 | * @note The number of M0xK0 vertical blocks to store on the same output row must be passed at compile time using -DV0 (e.g. -DV0=2) |
Gian Marco Iodice | 73cdaac | 2020-08-10 21:44:14 +0100 | [diff] [blame^] | 264 | * @note The size of the partial load block in y must be passed at compile time using -DPARTIAL_LOAD_M0 (e.g. -DPARTIAL_LOAD_M0=1) |
| 265 | * @note The size of the partial load block in x must be passed at compile time using -DPARTIAL_LOAD_K0 (e.g. -DPARTIAL_LOAD_K0=1) |
Gian Marco Iodice | 08ddd7b | 2018-12-19 10:01:18 +0000 | [diff] [blame] | 266 | * @note Only the following values for M0, K0 and V0 are supported: |
| 267 | * M0: 2,3,4,5,6,7,8 |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 268 | * K0: 2,3,4,8,16 |
Gian Marco Iodice | 08ddd7b | 2018-12-19 10:01:18 +0000 | [diff] [blame] | 269 | * V0: greater than 0 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 270 | * @note In case the input has to be reinterpreted as a 3D tensor (e.g. input of convolution layer 1x1), the following information must be passed at compile time: |
Gian Marco Iodice | 08ddd7b | 2018-12-19 10:01:18 +0000 | [diff] [blame] | 271 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
| 272 | * -# HEIGHT_GEMM3D: The height of the input in case it has to be reinterpreted as a 3D tensor. |
| 273 | * -# DEPTH_GEMM3D: The depth of the input in case it has to be reinterpreted as a 3D tensor |
| 274 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 275 | * @note If the M0xK0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time. |
| 276 | * |
Michele Di Giorgio | f6f7876 | 2020-07-06 11:27:21 +0100 | [diff] [blame] | 277 | * @param[in] src_ptr Pointer to the source LHS tensor. Supported data types: All |
Gian Marco Iodice | 08ddd7b | 2018-12-19 10:01:18 +0000 | [diff] [blame] | 278 | * @param[in] src_stride_x Stride of the source LHS tensor in X dimension (in bytes) |
| 279 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 280 | * @param[in] src_stride_y Stride of the source LHS tensor in Y dimension (in bytes) |
| 281 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 282 | * @param[in] src_stride_z Stride of the source LHS tensor in Z dimension (in bytes) |
| 283 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 284 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source LHS tensor |
| 285 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr |
| 286 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 287 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 288 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 289 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 290 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 291 | * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
| 292 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 293 | * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) |
| 294 | */ |
| 295 | __kernel void gemm_reshape_lhs_matrix_t(TENSOR3D_DECLARATION(src), |
| 296 | TENSOR3D_DECLARATION(dst) |
| 297 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 298 | , |
| 299 | uint cross_plane_pad |
| 300 | #endif // REINTERPRET_INPUT_AS_3D |
| 301 | ) |
| 302 | { |
| 303 | // Block size |
| 304 | #define BLOCK_SIZE ((M0) * (K0)) |
| 305 | |
| 306 | // Output offset X |
| 307 | #if defined(INTERLEAVE) |
| 308 | #define OUTPUT_OFFSET_X (M0) |
| 309 | #else // defined(INTERLEAVE) |
| 310 | #define OUTPUT_OFFSET_X (BLOCK_SIZE) |
| 311 | #endif // defined(INTERLEAVE) |
| 312 | |
| 313 | // Output step X |
| 314 | #if defined(INTERLEAVE) |
| 315 | #define OUTPUT_STEP_X (M0) * (V0) |
| 316 | #else // Do not interleave |
| 317 | #define OUTPUT_STEP_X (M0) |
| 318 | #endif // defined(INTERLEAVE) |
| 319 | |
| 320 | // Compute source and destination addresses |
| 321 | uint x = get_global_id(0); |
| 322 | uint y = get_global_id(1); |
| 323 | uint z = get_global_id(2); |
| 324 | |
| 325 | // ------------------ Compute input/output addresses --------------------------- |
| 326 | |
| 327 | // Compute the input address |
| 328 | __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)K0 * sizeof(DATA_TYPE) + y * (uint)M0 * src_stride_y; |
| 329 | |
| 330 | // Compute the output address |
| 331 | __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)BLOCK_SIZE * (uint)V0 * sizeof(DATA_TYPE)) + ((y / (uint)V0) * (uint)dst_stride_y) + ((y % V0) * |
| 332 | (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE)); |
| 333 | |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 334 | // Create variables: uint zin0=0, zin1=0, zin2=0...zin(M0-1)=0; |
| 335 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zin, 0); |
Gian Marco Iodice | 08ddd7b | 2018-12-19 10:01:18 +0000 | [diff] [blame] | 336 | |
| 337 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 338 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 339 | // multiply src_stride_z by DEPTH_GEMM3D |
| 340 | |
Gian Marco Iodice | 08ddd7b | 2018-12-19 10:01:18 +0000 | [diff] [blame] | 341 | input_ptr += z * (uint)src_stride_z * DEPTH_GEMM3D; |
| 342 | |
| 343 | // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 344 | CALCULATE_Z_OFFSET(M0, uint, zin, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, cross_plane_pad, src_stride_y); |
Gian Marco Iodice | 08ddd7b | 2018-12-19 10:01:18 +0000 | [diff] [blame] | 345 | |
| 346 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 347 | |
| 348 | input_ptr += z * (uint)src_stride_z; |
| 349 | |
| 350 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 351 | |
| 352 | // Add offset for batched GEMM |
| 353 | output_ptr += z * (uint)dst_stride_z; |
| 354 | |
| 355 | // ---------------------------Load input values -------------------------------- |
Gian Marco Iodice | 73cdaac | 2020-08-10 21:44:14 +0100 | [diff] [blame^] | 356 | REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, K0), a, 0); |
Gian Marco Iodice | 08ddd7b | 2018-12-19 10:01:18 +0000 | [diff] [blame] | 357 | |
Gian Marco Iodice | 73cdaac | 2020-08-10 21:44:14 +0100 | [diff] [blame^] | 358 | LOAD_TENSOR_BOUNDARY_AWARE_M0XK0(M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin); |
| 359 | |
Gian Marco Iodice | 08ddd7b | 2018-12-19 10:01:18 +0000 | [diff] [blame] | 360 | // ---------------------------Transpose and store block ----------------------- |
| 361 | |
| 362 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 0); |
| 363 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 1); |
| 364 | #if K0 > 2 |
| 365 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 2); |
Gian Marco Iodice | 08ddd7b | 2018-12-19 10:01:18 +0000 | [diff] [blame] | 366 | #endif // K0 > 2 |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 367 | #if K0 > 3 |
| 368 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 3); |
| 369 | #endif // K0 > 3 |
Gian Marco Iodice | 08ddd7b | 2018-12-19 10:01:18 +0000 | [diff] [blame] | 370 | #if K0 > 4 |
| 371 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 4); |
| 372 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 5); |
| 373 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 6); |
| 374 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 7); |
| 375 | #endif // K0 > 4 |
| 376 | #if K0 > 8 |
| 377 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 8); |
| 378 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 9); |
| 379 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, A); |
| 380 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, B); |
| 381 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, C); |
| 382 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, D); |
| 383 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, E); |
| 384 | TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, F); |
| 385 | #endif // K0 > 8 |
| 386 | |
| 387 | #undef BLOCK_SIZE |
| 388 | #undef OUTPUT_OFFSET_X |
| 389 | #undef OUTPUT_STEP_X |
| 390 | } |
Gian Marco Iodice | 73cdaac | 2020-08-10 21:44:14 +0100 | [diff] [blame^] | 391 | #endif // defined(M0) && defined(K0) && defined(V0) && defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(PARTIAL_LOAD_M0) && defined(PARTIAL_LOAD_K0) |
Gian Marco Iodice | 5ba5e09 | 2018-12-06 17:13:09 +0000 | [diff] [blame] | 392 | |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 393 | #if defined(K0) && defined(N0) && defined(H0) && defined(DATA_TYPE) && defined(SRC_HEIGHT) |
| 394 | /** This OpenCL kernel reshapes the rhs input matrix. The kernel splits the input matrix in blocks of size K0xN0 and stores each one (not transposed) in |
| 395 | * the output matrix unrolling the values. |
| 396 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 397 | * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) |
| 398 | * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16) |
| 399 | * @note The block's dimensions (K0 and N0) must be passed at compile time using -DK0 and -DN0 (e.g. -DK0=2, -DN0=2). |
| 400 | * @note The number of K0xN0 vertical blocks to store on the same output row must be passed at compile time using -DH0 (e.g. -DH0=2) |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 401 | * @note If the K0xN0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time. |
| 402 | * @note Only the following values for K0, N0 and H0 are supported: |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 403 | * N0: 2,3,4,8,16 |
| 404 | * K0: 1,2,3,4,8,16 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 405 | * H0: greater than 0 |
| 406 | * |
Michele Di Giorgio | f6f7876 | 2020-07-06 11:27:21 +0100 | [diff] [blame] | 407 | * @param[in] src_ptr Pointer to the source RHS tensor. Supported data types: All |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 408 | * @param[in] src_stride_x Stride of the source RHS tensor in X dimension (in bytes) |
| 409 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 410 | * @param[in] src_stride_y Stride of the source RHS tensor in Y dimension (in bytes) |
| 411 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 412 | * @param[in] src_stride_z Stride of the source RHS tensor in Z dimension (in bytes) |
| 413 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 414 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source RHS tensor |
| 415 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr |
| 416 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 417 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 418 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 419 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 420 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 421 | * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
| 422 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 423 | */ |
| 424 | __kernel void gemm_reshape_rhs_matrix_nt(TENSOR3D_DECLARATION(src), |
| 425 | TENSOR3D_DECLARATION(dst)) |
| 426 | { |
| 427 | // Block size |
| 428 | #define BLOCK_SIZE ((K0) * (N0)) |
| 429 | |
| 430 | // Output offset X |
| 431 | #if defined(INTERLEAVE) |
| 432 | #define OUTPUT_OFFSET_X (N0) |
| 433 | #else // defined(INTERLEAVE) |
| 434 | #define OUTPUT_OFFSET_X (BLOCK_SIZE) |
| 435 | #endif // defined(INTERLEAVE) |
| 436 | |
| 437 | // Output step X |
| 438 | #if defined(INTERLEAVE) |
| 439 | #define OUTPUT_STEP_X (N0) * (H0) |
| 440 | #else // Do not interleave |
| 441 | #define OUTPUT_STEP_X (N0) |
| 442 | #endif // defined(INTERLEAVE) |
| 443 | |
| 444 | // Compute source and destination addresses |
| 445 | uint x = get_global_id(0); |
| 446 | uint y = get_global_id(1); |
| 447 | uint z = get_global_id(2); |
| 448 | |
| 449 | // ------------------ Compute input/output addresses --------------------------- |
| 450 | |
| 451 | // Compute the input address |
| 452 | __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)N0 * sizeof(DATA_TYPE) + y * (uint)K0 * src_stride_y + z * (uint)src_stride_z; |
| 453 | |
| 454 | // Compute the output address |
| 455 | __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (y * (uint)BLOCK_SIZE * (uint)H0 * sizeof(DATA_TYPE)) + ((x % (uint)H0) * (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE)) + (( |
| 456 | x / (uint)H0) |
| 457 | * (uint)dst_stride_y) |
| 458 | + z * (uint)dst_stride_z; |
| 459 | |
| 460 | // ---------------------------Load input values -------------------------------- |
| 461 | |
Vidhya Sudhan Loganathan | 17b0f8b | 2019-01-08 12:17:03 +0000 | [diff] [blame] | 462 | REPEAT_VAR_INIT_TO_CONST(K0, VEC_DATA_TYPE(DATA_TYPE, N0), a, 0); ////uint a0=0, a1=0, a2=0...a(M0-1)=0; |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 463 | |
| 464 | // Load values from the RHS matrix |
| 465 | a0 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 0 * src_stride_y)); |
| 466 | #if K0 > 1 |
| 467 | if(y * (uint)K0 + 1 < SRC_HEIGHT) |
| 468 | { |
| 469 | a1 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 1 * src_stride_y)); |
| 470 | } |
| 471 | #endif // K0 > 1 |
| 472 | #if K0 > 2 |
| 473 | if(y * (uint)K0 + 2 < SRC_HEIGHT) |
| 474 | { |
| 475 | a2 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 2 * src_stride_y)); |
| 476 | } |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 477 | #endif // K0 > 2 |
| 478 | #if K0 > 3 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 479 | if(y * (uint)K0 + 3 < SRC_HEIGHT) |
| 480 | { |
| 481 | a3 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 3 * src_stride_y)); |
| 482 | } |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 483 | #endif // K0 > 3 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 484 | #if K0 > 4 |
| 485 | if(y * (uint)K0 + 4 < SRC_HEIGHT) |
| 486 | { |
| 487 | a4 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 4 * src_stride_y)); |
| 488 | } |
| 489 | if(y * (uint)K0 + 5 < SRC_HEIGHT) |
| 490 | { |
| 491 | a5 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 5 * src_stride_y)); |
| 492 | } |
| 493 | if(y * (uint)K0 + 6 < SRC_HEIGHT) |
| 494 | { |
| 495 | a6 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 6 * src_stride_y)); |
| 496 | } |
| 497 | if(y * (uint)K0 + 7 < SRC_HEIGHT) |
| 498 | { |
| 499 | a7 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 7 * src_stride_y)); |
| 500 | } |
| 501 | #endif // K0 > 4 |
| 502 | #if K0 > 8 |
Gian Marco Iodice | 08ddd7b | 2018-12-19 10:01:18 +0000 | [diff] [blame] | 503 | if(y * (uint)K0 + 8 < SRC_HEIGHT) |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 504 | { |
| 505 | a8 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 8 * src_stride_y)); |
| 506 | } |
| 507 | if(y * (uint)K0 + 9 < SRC_HEIGHT) |
| 508 | { |
| 509 | a9 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 9 * src_stride_y)); |
| 510 | } |
| 511 | if(y * (uint)K0 + 10 < SRC_HEIGHT) |
| 512 | { |
| 513 | aA = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 10 * src_stride_y)); |
| 514 | } |
| 515 | if(y * (uint)K0 + 11 < SRC_HEIGHT) |
| 516 | { |
| 517 | aB = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 11 * src_stride_y)); |
| 518 | } |
| 519 | if(y * (uint)K0 + 12 < SRC_HEIGHT) |
| 520 | { |
| 521 | aC = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 12 * src_stride_y)); |
| 522 | } |
| 523 | if(y * (uint)K0 + 13 < SRC_HEIGHT) |
| 524 | { |
| 525 | aD = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 13 * src_stride_y)); |
| 526 | } |
| 527 | if(y * (uint)K0 + 14 < SRC_HEIGHT) |
| 528 | { |
| 529 | aE = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 14 * src_stride_y)); |
| 530 | } |
| 531 | if(y * (uint)K0 + 15 < SRC_HEIGHT) |
| 532 | { |
| 533 | aF = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 15 * src_stride_y)); |
| 534 | } |
| 535 | #endif // K0 > 8 |
| 536 | |
| 537 | // ---------------------------Store output values ------------------------------ |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 538 | REPEAT_VAR_INIT_TO_CONST(16, uint, zout, 0); |
| 539 | STORE_BLOCK(K0, N0, DATA_TYPE, a, output_ptr, OUTPUT_STEP_X * sizeof(DATA_TYPE), zout); |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 540 | |
| 541 | #undef BLOCK_SIZE |
| 542 | #undef OUTPUT_OFFSET_X |
| 543 | #undef OUTPUT_STEP_X |
| 544 | } |
| 545 | |
| 546 | #if defined(TRANSPOSE) |
| 547 | /** This OpenCL kernel reshapes the rhs input matrix. The kernel splits the input matrix in blocks of size K0xN0 and stores each one (transposed) in |
| 548 | * the output matrix unrolling the values. |
| 549 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 550 | * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) |
| 551 | * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16) |
| 552 | * @note The block's dimensions (K0 and N0) must be passed at compile time using -DK0 and -DN0 (e.g. -DK0=2, -DN0=2). |
| 553 | * @note The number of K0xN0 vertical blocks to store on the same output row must be passed at compile time using -DH0 (e.g. -DH0=2) |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 554 | * @note If the K0xN0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time. |
| 555 | * @note The option -DTRANSPOSE must passed at compile time. |
| 556 | * @note Only the following values for K0, N0 and H0 are supported: |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 557 | * N0: 2,3,4,8,16 |
| 558 | * K0: 2,3,4,8,16 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 559 | * H0: greater than 0 |
| 560 | * |
Michele Di Giorgio | f6f7876 | 2020-07-06 11:27:21 +0100 | [diff] [blame] | 561 | * @param[in] src_ptr Pointer to the source RHS tensor. Supported data types: All |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 562 | * @param[in] src_stride_x Stride of the source RHS tensor in X dimension (in bytes) |
| 563 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 564 | * @param[in] src_stride_y Stride of the source RHS tensor in Y dimension (in bytes) |
| 565 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 566 | * @param[in] src_stride_z Stride of the source RHS tensor in Z dimension (in bytes) |
| 567 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 568 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source RHS tensor |
| 569 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr |
| 570 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 571 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 572 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 573 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 574 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 575 | * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
| 576 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 577 | */ |
| 578 | __kernel void gemm_reshape_rhs_matrix_t(TENSOR3D_DECLARATION(src), |
| 579 | TENSOR3D_DECLARATION(dst)) |
| 580 | { |
| 581 | // Block size |
| 582 | #define BLOCK_SIZE ((K0) * (N0)) |
| 583 | |
| 584 | // Output offset X |
| 585 | #if defined(INTERLEAVE) |
| 586 | #define OUTPUT_OFFSET_X (K0) |
| 587 | #else // defined(INTERLEAVE) |
| 588 | #define OUTPUT_OFFSET_X (BLOCK_SIZE) |
| 589 | #endif // defined(INTERLEAVE) |
| 590 | |
| 591 | // Output step X |
| 592 | #if defined(INTERLEAVE) |
| 593 | #define OUTPUT_STEP_X (K0) * (H0) |
| 594 | #else // Do not interleave |
| 595 | #define OUTPUT_STEP_X (K0) |
| 596 | #endif // defined(INTERLEAVE) |
| 597 | |
| 598 | // Compute source and destination addresses |
| 599 | uint x = get_global_id(0); |
| 600 | uint y = get_global_id(1); |
| 601 | uint z = get_global_id(2); |
| 602 | |
| 603 | // ------------------ Compute input/output addresses --------------------------- |
| 604 | |
| 605 | // Compute the input address |
| 606 | __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)N0 * sizeof(DATA_TYPE) + y * (uint)K0 * src_stride_y + z * (uint)src_stride_z; |
| 607 | |
| 608 | // Compute the output address |
| 609 | __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (y * (uint)BLOCK_SIZE * (uint)H0 * sizeof(DATA_TYPE)) + ((x % H0) * (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE)) + ((x / |
| 610 | (uint)H0) * (uint)dst_stride_y) + z * (uint)dst_stride_z; |
| 611 | |
| 612 | // ---------------------------Load input values -------------------------------- |
Vidhya Sudhan Loganathan | 17b0f8b | 2019-01-08 12:17:03 +0000 | [diff] [blame] | 613 | REPEAT_VAR_INIT_TO_CONST(K0, VEC_DATA_TYPE(DATA_TYPE, N0), a, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) a0=0, a1=0, ... a(K0-1)=0; |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 614 | |
| 615 | // Load values from the RHS matrix |
| 616 | a0 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 0 * src_stride_y)); |
| 617 | if(y * (uint)K0 + 1 < SRC_HEIGHT) |
| 618 | { |
| 619 | a1 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 1 * src_stride_y)); |
| 620 | } |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 621 | #if K0 > 2 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 622 | if(y * (uint)K0 + 2 < SRC_HEIGHT) |
| 623 | { |
| 624 | a2 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 2 * src_stride_y)); |
| 625 | } |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 626 | #endif // K0 > 2 |
| 627 | #if K0 > 3 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 628 | if(y * (uint)K0 + 3 < SRC_HEIGHT) |
| 629 | { |
| 630 | a3 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 3 * src_stride_y)); |
| 631 | } |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 632 | #endif // K0 > 3 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 633 | #if K0 > 4 |
| 634 | if(y * (uint)K0 + 4 < SRC_HEIGHT) |
| 635 | { |
| 636 | a4 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 4 * src_stride_y)); |
| 637 | } |
| 638 | if(y * (uint)K0 + 5 < SRC_HEIGHT) |
| 639 | { |
| 640 | a5 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 5 * src_stride_y)); |
| 641 | } |
| 642 | if(y * (uint)K0 + 6 < SRC_HEIGHT) |
| 643 | { |
| 644 | a6 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 6 * src_stride_y)); |
| 645 | } |
| 646 | if(y * (uint)K0 + 7 < SRC_HEIGHT) |
| 647 | { |
| 648 | a7 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 7 * src_stride_y)); |
| 649 | } |
| 650 | #endif // K0 > 4 |
| 651 | #if K0 > 8 |
Gian Marco Iodice | 8912434 | 2018-12-19 14:17:22 +0000 | [diff] [blame] | 652 | if(y * (uint)K0 + 8 < SRC_HEIGHT) |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 653 | { |
| 654 | a8 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 8 * src_stride_y)); |
| 655 | } |
| 656 | if(y * (uint)K0 + 9 < SRC_HEIGHT) |
| 657 | { |
| 658 | a9 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 9 * src_stride_y)); |
| 659 | } |
| 660 | if(y * (uint)K0 + 10 < SRC_HEIGHT) |
| 661 | { |
| 662 | aA = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 10 * src_stride_y)); |
| 663 | } |
| 664 | if(y * (uint)K0 + 11 < SRC_HEIGHT) |
| 665 | { |
| 666 | aB = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 11 * src_stride_y)); |
| 667 | } |
| 668 | if(y * (uint)K0 + 12 < SRC_HEIGHT) |
| 669 | { |
| 670 | aC = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 12 * src_stride_y)); |
| 671 | } |
| 672 | if(y * (uint)K0 + 13 < SRC_HEIGHT) |
| 673 | { |
| 674 | aD = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 13 * src_stride_y)); |
| 675 | } |
| 676 | if(y * (uint)K0 + 14 < SRC_HEIGHT) |
| 677 | { |
| 678 | aE = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 14 * src_stride_y)); |
| 679 | } |
| 680 | if(y * (uint)K0 + 15 < SRC_HEIGHT) |
| 681 | { |
| 682 | aF = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 15 * src_stride_y)); |
| 683 | } |
| 684 | #endif // K0 > 8 |
| 685 | |
| 686 | // ---------------------------Transpose the block ------------------------------ |
Vidhya Sudhan Loganathan | 17b0f8b | 2019-01-08 12:17:03 +0000 | [diff] [blame] | 687 | REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), res, 0); //VEC_DATA_TYPE(DATA_TYPE, K0) res0=0, res1=0, res2=0,... res(N0-1)=0; |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 688 | |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 689 | #if K0 == 2 |
| 690 | // This part computes the following transpositions: |
| 691 | // 2x2 -> 2x2 |
| 692 | // 2x4 -> 4x2 |
| 693 | // 2x8 -> 8x2 |
| 694 | // 2x16 -> 16x2 |
| 695 | res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0); |
| 696 | res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1); |
| 697 | #if N0 > 2 |
| 698 | res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2); |
| 699 | #endif // N0 > 2 |
| 700 | #if N0 > 3 |
| 701 | res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3); |
| 702 | #endif // N0 > 3 |
| 703 | #if N0 > 4 |
| 704 | res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4); |
| 705 | res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5); |
| 706 | res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6); |
| 707 | res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7); |
| 708 | #endif // N0 > 4 |
| 709 | #if N0 > 8 |
| 710 | res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8); |
| 711 | res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9); |
| 712 | resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA); |
| 713 | resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB); |
| 714 | resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC); |
| 715 | resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD); |
| 716 | resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE); |
| 717 | resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF); |
| 718 | #endif // N0 > 8 |
| 719 | |
| 720 | #elif K0 == 3 // K0 == 2 |
| 721 | // This part computes the following transpositions: |
| 722 | // 3x2 -> 2x3 |
| 723 | // 3x4 -> 4x3 |
| 724 | // 3x8 -> 8x3 |
| 725 | // 3x16 -> 16x3 |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 726 | res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0); |
| 727 | res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1); |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 728 | #if N0 > 2 |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 729 | res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2); |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 730 | #endif // N0 > 2 |
| 731 | #if N0 > 3 |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 732 | res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3); |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 733 | #endif // N0 > 3 |
| 734 | #if N0 > 4 |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 735 | res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4); |
| 736 | res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5); |
| 737 | res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6); |
| 738 | res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7); |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 739 | #endif // N0 > 4 |
| 740 | #if N0 > 8 |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 741 | res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8); |
| 742 | res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9); |
| 743 | resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA); |
| 744 | resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB); |
| 745 | resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC); |
| 746 | resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD); |
| 747 | resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE); |
| 748 | resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF); |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 749 | #endif // N0 > 8 |
| 750 | |
| 751 | #elif K0 == 4 // K0 == 4 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 752 | // This part computes the following transpositions: |
| 753 | // 4x2 -> 2x4 |
| 754 | // 4x4 -> 4x4 |
| 755 | // 4x8 -> 8x4 |
| 756 | // 4x16 -> 16x4 |
| 757 | res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0, a3.s0); |
| 758 | res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1, a3.s1); |
| 759 | #if N0 > 2 |
| 760 | res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2, a3.s2); |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 761 | #endif // N0 > 2 |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 762 | #if N0 > 3 |
| 763 | res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3, a3.s3); |
| 764 | #endif // N0 > 3 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 765 | #if N0 > 4 |
| 766 | res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4, a3.s4); |
| 767 | res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5, a3.s5); |
| 768 | res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6, a3.s6); |
| 769 | res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7, a3.s7); |
| 770 | #endif // N0 > 4 |
| 771 | #if N0 > 8 |
| 772 | res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8, a3.s8); |
| 773 | res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9, a3.s9); |
| 774 | resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA, a3.sA); |
| 775 | resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB, a3.sB); |
| 776 | resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC, a3.sC); |
| 777 | resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD, a3.sD); |
| 778 | resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE, a3.sE); |
| 779 | resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF, a3.sF); |
| 780 | #endif // N0 > 8 |
| 781 | |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 782 | #elif K0 == 8 // K0 == 8 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 783 | // This part computes the following transpositions: |
| 784 | // 8x2 -> 2x8 |
| 785 | // 8x4 -> 4x8 |
| 786 | // 8x8 -> 8x8 |
| 787 | // 8x16 -> 16x8 |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 788 | res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0, a3.s0, a4.s0, a5.s0, a6.s0, a7.s0); |
| 789 | res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1, a3.s1, a4.s1, a5.s1, a6.s1, a7.s1); |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 790 | #if N0 > 2 |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 791 | res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2, a3.s2, a4.s2, a5.s2, a6.s2, a7.s2); |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 792 | #endif // N0 > 2 |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 793 | #if N0 > 3 |
| 794 | res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3, a3.s3, a4.s3, a5.s3, a6.s3, a7.s3); |
| 795 | #endif // N0 > 3 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 796 | #if N0 > 4 |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 797 | res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4, a3.s4, a4.s4, a5.s4, a6.s4, a7.s4); |
| 798 | res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5, a3.s5, a4.s5, a5.s5, a6.s5, a7.s5); |
| 799 | res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6, a3.s6, a4.s6, a5.s6, a6.s6, a7.s6); |
| 800 | res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7, a3.s7, a4.s7, a5.s7, a6.s7, a7.s7); |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 801 | #endif // N0 > 4 |
| 802 | #if N0 > 8 |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 803 | res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8, a3.s8, a4.s8, a5.s8, a6.s8, a7.s8); |
| 804 | res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9, a3.s9, a4.s9, a5.s9, a6.s9, a7.s9); |
| 805 | resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA, a3.sA, a4.sA, a5.sA, a6.sA, a7.sA); |
| 806 | resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB, a3.sB, a4.sB, a5.sB, a6.sB, a7.sB); |
| 807 | resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC, a3.sC, a4.sC, a5.sC, a6.sC, a7.sC); |
| 808 | resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD, a3.sD, a4.sD, a5.sD, a6.sD, a7.sD); |
| 809 | resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE, a3.sE, a4.sE, a5.sE, a6.sE, a7.sE); |
| 810 | resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF, a3.sF, a4.sF, a5.sF, a6.sF, a7.sF); |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 811 | #endif // N0 > 8 |
| 812 | |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 813 | #elif K0 == 16 // K0 == 16 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 814 | |
| 815 | // This part computes the following transpositions: |
| 816 | // 16x2 -> 2x16 |
| 817 | // 16x4 -> 4x16 |
| 818 | // 16x8 -> 8x16 |
| 819 | // 16x16 -> 16x16 |
| 820 | res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0, a3.s0, a4.s0, a5.s0, a6.s0, a7.s0, |
| 821 | a8.s0, a9.s0, aA.s0, aB.s0, aC.s0, aD.s0, aE.s0, aF.s0); |
| 822 | res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1, a3.s1, a4.s1, a5.s1, a6.s1, a7.s1, |
| 823 | a8.s1, a9.s1, aA.s1, aB.s1, aC.s1, aD.s1, aE.s1, aF.s1); |
| 824 | #if N0 > 2 |
| 825 | res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2, a3.s2, a4.s2, a5.s2, a6.s2, a7.s2, |
| 826 | a8.s2, a9.s2, aA.s2, aB.s2, aC.s2, aD.s2, aE.s2, aF.s2); |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 827 | #endif // N0 > 2 |
| 828 | #if N0 > 3 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 829 | res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3, a3.s3, a4.s3, a5.s3, a6.s3, a7.s3, |
| 830 | a8.s3, a9.s3, aA.s3, aB.s3, aC.s3, aD.s3, aE.s3, aF.s3); |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 831 | #endif // N0 > 3 |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 832 | #if N0 > 4 |
| 833 | res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4, a3.s4, a4.s4, a5.s4, a6.s4, a7.s4, |
| 834 | a8.s4, a9.s4, aA.s4, aB.s4, aC.s4, aD.s4, aE.s4, aF.s4); |
| 835 | res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5, a3.s5, a4.s5, a5.s5, a6.s5, a7.s5, |
| 836 | a8.s5, a9.s5, aA.s5, aB.s5, aC.s5, aD.s5, aE.s5, aF.s5); |
| 837 | res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6, a3.s6, a4.s6, a5.s6, a6.s6, a7.s6, |
| 838 | a8.s6, a9.s6, aA.s6, aB.s6, aC.s6, aD.s6, aE.s6, aF.s6); |
| 839 | res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7, a3.s7, a4.s7, a5.s7, a6.s7, a7.s7, |
| 840 | a8.s7, a9.s7, aA.s7, aB.s7, aC.s7, aD.s7, aE.s7, aF.s7); |
| 841 | #endif // N0 > 4 |
| 842 | #if N0 > 8 |
| 843 | res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8, a3.s8, a4.s8, a5.s8, a6.s8, a7.s8, |
| 844 | a8.s8, a9.s8, aA.s8, aB.s8, aC.s8, aD.s8, aE.s8, aF.s8); |
| 845 | res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9, a3.s9, a4.s9, a5.s9, a6.s9, a7.s9, |
| 846 | a8.s9, a9.s9, aA.s9, aB.s9, aC.s9, aD.s9, aE.s9, aF.s9); |
| 847 | resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA, a3.sA, a4.sA, a5.sA, a6.sA, a7.sA, |
| 848 | a8.sA, a9.sA, aA.sA, aB.sA, aC.sA, aD.sA, aE.sA, aF.sA); |
| 849 | resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB, a3.sB, a4.sB, a5.sB, a6.sB, a7.sB, |
| 850 | a8.sB, a9.sB, aA.sB, aB.sB, aC.sB, aD.sB, aE.sB, aF.sB); |
| 851 | resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC, a3.sC, a4.sC, a5.sC, a6.sC, a7.sC, |
| 852 | a8.sC, a9.sC, aA.sC, aB.sC, aC.sC, aD.sC, aE.sC, aF.sC); |
| 853 | resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD, a3.sD, a4.sD, a5.sD, a6.sD, a7.sD, |
| 854 | a8.sD, a9.sD, aA.sD, aB.sD, aC.sD, aD.sD, aE.sD, aF.sD); |
| 855 | resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE, a3.sE, a4.sE, a5.sE, a6.sE, a7.sE, |
| 856 | a8.sE, a9.sE, aA.sE, aB.sE, aC.sE, aD.sE, aE.sE, aF.sE); |
| 857 | resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF, a3.sF, a4.sF, a5.sF, a6.sF, a7.sF, |
| 858 | a8.sF, a9.sF, aA.sF, aB.sF, aC.sF, aD.sF, aE.sF, aF.sF); |
| 859 | #endif // N0 > 8 |
| 860 | |
| 861 | #else // N0 == 16 |
| 862 | #error "Not supported N0 value" |
| 863 | #endif // N0 > 2 |
| 864 | |
| 865 | // ---------------------------Store the output values ------------------------------ |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 866 | REPEAT_VAR_INIT_TO_CONST(16, uint, zout, 0); |
| 867 | STORE_BLOCK(N0, K0, DATA_TYPE, res, output_ptr, OUTPUT_STEP_X * sizeof(DATA_TYPE), zout); |
Gian Marco Iodice | 3b0a265 | 2018-12-07 11:18:09 +0000 | [diff] [blame] | 868 | |
| 869 | #undef BLOCK_SIZE |
| 870 | #undef OUTPUT_OFFSET_X |
| 871 | #undef OUTPUT_STEP_X |
| 872 | } |
| 873 | #endif // defined(TRANSPOSE) |
| 874 | #endif // defined(K0) && defined(N0) && defined(H0) && defined(DATA_TYPE) && defined(SRC_HEIGHT) |
| 875 | |
Gian Marco Iodice | b0c5037 | 2019-03-15 10:13:05 +0000 | [diff] [blame] | 876 | #if defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(M) && defined(N) && defined(K) |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 877 | |
| 878 | #define CONCAT(a, b) a##b |
| 879 | |
| 880 | #define ARM_DOT1(a, b, c) \ |
| 881 | ({ \ |
| 882 | c = fma(a, b, c); \ |
| 883 | }) |
| 884 | #define ARM_DOT2(a, b, c) \ |
| 885 | ({ \ |
| 886 | c = fma(a.s0, b.s0, c); \ |
| 887 | c = fma(a.s1, b.s1, c); \ |
| 888 | }) |
| 889 | #define ARM_DOT3(a, b, c) \ |
| 890 | ({ \ |
| 891 | ARM_DOT2(a, b, c); \ |
| 892 | c = fma((a.s2), (b.s2), c); \ |
| 893 | }) |
| 894 | #define ARM_DOT4(a, b, c) \ |
| 895 | ({ \ |
| 896 | ARM_DOT3(a, b, c); \ |
| 897 | c = fma((a.s3), (b.s3), c); \ |
| 898 | }) |
| 899 | #define ARM_DOT8(a, b, c) \ |
| 900 | ({ \ |
| 901 | ARM_DOT4((a.lo), (b.lo), c); \ |
| 902 | ARM_DOT4((a.hi), (b.hi), c); \ |
| 903 | }) |
| 904 | #define ARM_DOT16(a, b, c) \ |
| 905 | ({ \ |
| 906 | ARM_DOT8((a.lo), (b.lo), c); \ |
| 907 | ARM_DOT8((a.hi), (b.hi), c); \ |
| 908 | }) |
| 909 | |
| 910 | #if N0 == 2 |
| 911 | #define ARM_DOT_K0XN0(k0, a, b, c) \ |
| 912 | ({ \ |
| 913 | CONCAT(ARM_DOT, k0) \ |
| 914 | ((a), (b##0), (c.s0)); \ |
| 915 | CONCAT(ARM_DOT, k0) \ |
| 916 | ((a), (b##1), (c.s1)); \ |
| 917 | }) |
| 918 | #elif N0 == 3 // N0 == 3 |
| 919 | #define ARM_DOT_K0XN0(k0, a, b, c) \ |
| 920 | ({ \ |
| 921 | CONCAT(ARM_DOT, k0) \ |
| 922 | ((a), (b##0), (c.s0)); \ |
| 923 | CONCAT(ARM_DOT, k0) \ |
| 924 | ((a), (b##1), (c.s1)); \ |
| 925 | CONCAT(ARM_DOT, k0) \ |
| 926 | ((a), (b##2), (c.s2)); \ |
| 927 | }) |
| 928 | #elif N0 == 4 // N0 == 4 |
| 929 | #define ARM_DOT_K0XN0(k0, a, b, c) \ |
| 930 | ({ \ |
| 931 | CONCAT(ARM_DOT, k0) \ |
| 932 | ((a), (b##0), (c.s0)); \ |
| 933 | CONCAT(ARM_DOT, k0) \ |
| 934 | ((a), (b##1), (c.s1)); \ |
| 935 | CONCAT(ARM_DOT, k0) \ |
| 936 | ((a), (b##2), (c.s2)); \ |
| 937 | CONCAT(ARM_DOT, k0) \ |
| 938 | ((a), (b##3), (c.s3)); \ |
| 939 | }) |
| 940 | #elif N0 == 8 // N0 == 8 |
| 941 | #define ARM_DOT_K0XN0(k0, a, b, c) \ |
| 942 | ({ \ |
| 943 | CONCAT(ARM_DOT, k0) \ |
| 944 | ((a), (b##0), (c.s0)); \ |
| 945 | CONCAT(ARM_DOT, k0) \ |
| 946 | ((a), (b##1), (c.s1)); \ |
| 947 | CONCAT(ARM_DOT, k0) \ |
| 948 | ((a), (b##2), (c.s2)); \ |
| 949 | CONCAT(ARM_DOT, k0) \ |
| 950 | ((a), (b##3), (c.s3)); \ |
| 951 | CONCAT(ARM_DOT, k0) \ |
| 952 | ((a), (b##4), (c.s4)); \ |
| 953 | CONCAT(ARM_DOT, k0) \ |
| 954 | ((a), (b##5), (c.s5)); \ |
| 955 | CONCAT(ARM_DOT, k0) \ |
| 956 | ((a), (b##6), (c.s6)); \ |
| 957 | CONCAT(ARM_DOT, k0) \ |
| 958 | ((a), (b##7), (c.s7)); \ |
| 959 | }) |
| 960 | #elif N0 == 16 // N0 == 16 |
| 961 | #define ARM_DOT_K0XN0(k0, a, b, c) \ |
| 962 | ({ \ |
| 963 | CONCAT(ARM_DOT, k0) \ |
| 964 | ((a), (b##0), (c.s0)); \ |
| 965 | CONCAT(ARM_DOT, k0) \ |
| 966 | ((a), (b##1), (c.s1)); \ |
| 967 | CONCAT(ARM_DOT, k0) \ |
| 968 | ((a), (b##2), (c.s2)); \ |
| 969 | CONCAT(ARM_DOT, k0) \ |
| 970 | ((a), (b##3), (c.s3)); \ |
| 971 | CONCAT(ARM_DOT, k0) \ |
| 972 | ((a), (b##4), (c.s4)); \ |
| 973 | CONCAT(ARM_DOT, k0) \ |
| 974 | ((a), (b##5), (c.s5)); \ |
| 975 | CONCAT(ARM_DOT, k0) \ |
| 976 | ((a), (b##6), (c.s6)); \ |
| 977 | CONCAT(ARM_DOT, k0) \ |
| 978 | ((a), (b##7), (c.s7)); \ |
| 979 | CONCAT(ARM_DOT, k0) \ |
| 980 | ((a), (b##8), (c.s8)); \ |
| 981 | CONCAT(ARM_DOT, k0) \ |
| 982 | ((a), (b##9), (c.s9)); \ |
| 983 | CONCAT(ARM_DOT, k0) \ |
| 984 | ((a), (b##A), (c.sA)); \ |
| 985 | CONCAT(ARM_DOT, k0) \ |
| 986 | ((a), (b##B), (c.sB)); \ |
| 987 | CONCAT(ARM_DOT, k0) \ |
| 988 | ((a), (b##C), (c.sC)); \ |
| 989 | CONCAT(ARM_DOT, k0) \ |
| 990 | ((a), (b##D), (c.sD)); \ |
| 991 | CONCAT(ARM_DOT, k0) \ |
| 992 | ((a), (b##E), (c.sE)); \ |
| 993 | CONCAT(ARM_DOT, k0) \ |
| 994 | ((a), (b##F), (c.sF)); \ |
| 995 | }) |
| 996 | #else // N0 not supported |
| 997 | #error "N0 value not supported" |
| 998 | #endif // N0 conditions |
| 999 | |
| 1000 | /** This OpenCL kernel computes the matrix multiplication between 2 matrices. |
| 1001 | * The LHS matrix is NOT reshaped |
| 1002 | * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed |
| 1003 | * |
Gian Marco Iodice | b0c5037 | 2019-03-15 10:13:05 +0000 | [diff] [blame] | 1004 | * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 1005 | * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90) |
| 1006 | * @note The number of columns of LHS matrix must be passed at compile time using -DK (e.g. -DK=64) |
| 1007 | * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4). |
| 1008 | * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2) |
| 1009 | * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2) |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1010 | * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time. |
SiCong Li | 3a50166 | 2020-06-26 10:02:06 +0100 | [diff] [blame] | 1011 | * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) |
| 1012 | * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1013 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 1014 | * - M0 = 1, 2, 3, 4, 5, 6, 7, 8 |
| 1015 | * - N0 = 2, 3, 4, 8, 16 |
| 1016 | * - K0 = 2, 3, 4, 8, 16 |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1017 | * - H0 >= 1 |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1018 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 1019 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1020 | * The activation function is performed after the bias addition |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1021 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 1022 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
| 1023 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 1024 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 1025 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 1026 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix |
| 1027 | * |
Sheri Zhang | 1a37810 | 2020-04-30 12:59:39 +0100 | [diff] [blame] | 1028 | * @param[in] lhs_ptr Pointer to the LHS matrix. Supported data type: F16/F32 |
| 1029 | * @param[in] lhs_stride_x Stride of the LHS matrix in X dimension (in bytes) |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1030 | * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
Sheri Zhang | 1a37810 | 2020-04-30 12:59:39 +0100 | [diff] [blame] | 1031 | * @param[in] lhs_stride_y Stride of the LHS matrix in Y dimension (in bytes) |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1032 | * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
Sheri Zhang | 1a37810 | 2020-04-30 12:59:39 +0100 | [diff] [blame] | 1033 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS matrix |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1034 | * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr |
| 1035 | * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes) |
| 1036 | * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1037 | * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes) |
| 1038 | * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1039 | * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1040 | * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 1041 | * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 1042 | * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes) |
| 1043 | * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 1044 | * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1045 | * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1046 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr |
| 1047 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 1048 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1049 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 1050 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1051 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Sheri Zhang | 1a37810 | 2020-04-30 12:59:39 +0100 | [diff] [blame] | 1052 | * @param[in] lhs_stride_z Stride of the LHS matrix in Z dimension (in bytes) |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1053 | * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1054 | * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1055 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 1056 | * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) |
| 1057 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1058 | */ |
| 1059 | __kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs), |
| 1060 | IMAGE_DECLARATION(rhs), |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1061 | #if defined(BETA) |
| 1062 | IMAGE_DECLARATION(bias), |
| 1063 | #endif // defined(BETA) |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1064 | IMAGE_DECLARATION(dst), |
| 1065 | uint lhs_stride_z, |
| 1066 | uint rhs_stride_z, |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1067 | #if defined(BETA) |
| 1068 | uint bias_stride_z, |
| 1069 | #endif //defined(BETA) |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1070 | uint dst_stride_z |
| 1071 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 1072 | , |
| 1073 | uint lhs_cross_plane_pad |
| 1074 | #endif // REINTERPRET_INPUT_AS_3D |
| 1075 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 1076 | , |
| 1077 | uint dst_cross_plane_pad |
| 1078 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 1079 | ) |
| 1080 | { |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1081 | // Block size |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1082 | #define RHS_BLOCK_SIZE ((K0) * (N0)) |
| 1083 | |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1084 | // RHS offset and step X |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1085 | #if defined(RHS_INTERLEAVE) |
| 1086 | #define RHS_OFFSET_X (K0) |
| 1087 | #define RHS_STEP_X ((K0) * (H0)) |
| 1088 | #define RHS_STEP_LOOP (1) |
| 1089 | #else // defined(RHS_INTERLEAVE) |
| 1090 | #define RHS_OFFSET_X (RHS_BLOCK_SIZE) |
| 1091 | #define RHS_STEP_X (K0) |
| 1092 | #define RHS_STEP_LOOP (H0) |
| 1093 | #endif // defined(RHS_INTERLEAVE) |
| 1094 | |
| 1095 | uint x = get_global_id(0); |
| 1096 | uint y = get_global_id(1); |
| 1097 | uint z = get_global_id(2); |
| 1098 | |
Gian Marco Iodice | b0c5037 | 2019-03-15 10:13:05 +0000 | [diff] [blame] | 1099 | #if defined(DUMMY_WORK_ITEMS) |
| 1100 | if((x * N0 >= N) || (y * M0 >= M)) |
| 1101 | { |
| 1102 | return; |
| 1103 | } |
| 1104 | #endif // defined(DUMMY_WORK_ITEMS) |
| 1105 | |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1106 | // Compute LHS matrix address |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 1107 | uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y; |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1108 | |
Sheri Zhang | 1a37810 | 2020-04-30 12:59:39 +0100 | [diff] [blame] | 1109 | // Compute RHS reshaped matrix address |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1110 | uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y; |
| 1111 | |
| 1112 | #if defined(MATRIX_B_DEPTH) |
| 1113 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 1114 | rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z; |
| 1115 | #else // defined(MATRIX_B_DEPTH) |
| 1116 | rhs_offset += z * rhs_stride_z; |
| 1117 | #endif // defined(MATRIX_B_DEPTH) |
| 1118 | |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 1119 | REPEAT_VAR_INIT_TO_CONST(8, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0; |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1120 | REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1121 | |
| 1122 | #if defined(REINTERPRET_INPUT_AS_3D) |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 1123 | // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
| 1124 | CALCULATE_Z_OFFSET(M0, uint, zlhs, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1125 | |
| 1126 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 1127 | // multiply lhs_stride_z by DEPTH_GEMM3D |
| 1128 | lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D; |
| 1129 | |
| 1130 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 1131 | |
| 1132 | // Add offset for batched GEMM |
| 1133 | lhs_offset += z * lhs_stride_z; |
| 1134 | |
| 1135 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 1136 | |
| 1137 | // Initialize the accumulators |
| 1138 | REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(M0-1)=0; |
| 1139 | |
| 1140 | int i = 0; |
| 1141 | for(; i <= (K - K0); i += K0) |
| 1142 | { |
| 1143 | // Supported cases (M0, K0): |
| 1144 | // 1,2 - 1,3 - 1,4 - 1,8 - 1,16 |
| 1145 | // 2,2 - 2,3 - 2,4 - 2,8 - 2,16 |
| 1146 | // 3,2 - 3,3 - 3,4 - 3,8 - 3,16 |
| 1147 | // 4,2 - 4,3 - 4,4 - 4,8 - 4,16 |
| 1148 | // 5,2 - 5,3 - 5,4 - 5,8 - 5,16 |
| 1149 | // 6,2 - 6,3 - 6,4 - 6,8 - 6,16 |
| 1150 | // 7,2 - 7,3 - 7,4 - 7,8 - 7,16 |
| 1151 | // 8,2 - 8,3 - 8,4 - 8,8 - 8,16 |
| 1152 | // Load values from LHS matrix |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 1153 | LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1154 | |
Sheri Zhang | 1a37810 | 2020-04-30 12:59:39 +0100 | [diff] [blame] | 1155 | // Load values from RHS reshaped matrix |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1156 | LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X * sizeof(DATA_TYPE), zero); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1157 | |
| 1158 | // Accumulate |
| 1159 | ARM_DOT_K0XN0(K0, a0, b, c0); |
| 1160 | #if M0 > 1 |
| 1161 | ARM_DOT_K0XN0(K0, a1, b, c1); |
| 1162 | #endif // M0 > 1 |
| 1163 | #if M0 > 2 |
| 1164 | ARM_DOT_K0XN0(K0, a2, b, c2); |
| 1165 | #endif // M0 > 2 |
| 1166 | #if M0 > 3 |
| 1167 | ARM_DOT_K0XN0(K0, a3, b, c3); |
| 1168 | #endif // M0 > 3 |
| 1169 | #if M0 > 4 |
| 1170 | ARM_DOT_K0XN0(K0, a4, b, c4); |
| 1171 | #endif // M0 > 4 |
| 1172 | #if M0 > 5 |
| 1173 | ARM_DOT_K0XN0(K0, a5, b, c5); |
| 1174 | #endif // M0 > 5 |
| 1175 | #if M0 > 6 |
| 1176 | ARM_DOT_K0XN0(K0, a6, b, c6); |
| 1177 | #endif // M0 > 6 |
| 1178 | #if M0 > 7 |
| 1179 | ARM_DOT_K0XN0(K0, a7, b, c7); |
| 1180 | #endif // M0 > 7 |
| 1181 | |
| 1182 | lhs_offset += K0 * sizeof(DATA_TYPE); |
| 1183 | rhs_offset += (N0 * RHS_STEP_X * RHS_STEP_LOOP) * sizeof(DATA_TYPE); |
| 1184 | } |
| 1185 | |
| 1186 | // Left-over accumulations |
| 1187 | for(; i < K; ++i) |
| 1188 | { |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1189 | // Load values from LHS matrix |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 1190 | LOAD_BLOCK(M0, 1, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1191 | |
Sheri Zhang | 1a37810 | 2020-04-30 12:59:39 +0100 | [diff] [blame] | 1192 | // Load values from RHS reshaped matrix |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1193 | LOAD_BLOCK(N0, 1, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X * sizeof(DATA_TYPE), zero); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1194 | |
| 1195 | // Accumulate |
| 1196 | ARM_DOT_K0XN0(1, a0, b, c0); |
| 1197 | #if M0 > 1 |
| 1198 | ARM_DOT_K0XN0(1, a1, b, c1); |
| 1199 | #endif // M0 > 1 |
| 1200 | #if M0 > 2 |
| 1201 | ARM_DOT_K0XN0(1, a2, b, c2); |
| 1202 | #endif // M0 > 2 |
| 1203 | #if M0 > 3 |
| 1204 | ARM_DOT_K0XN0(1, a3, b, c3); |
| 1205 | #endif // M0 > 3 |
| 1206 | #if M0 > 4 |
| 1207 | ARM_DOT_K0XN0(1, a4, b, c4); |
| 1208 | #endif // M0 > 4 |
| 1209 | #if M0 > 5 |
| 1210 | ARM_DOT_K0XN0(1, a5, b, c5); |
| 1211 | #endif // M0 > 5 |
| 1212 | #if M0 > 6 |
| 1213 | ARM_DOT_K0XN0(1, a6, b, c6); |
| 1214 | #endif // M0 > 6 |
| 1215 | #if M0 > 7 |
| 1216 | ARM_DOT_K0XN0(1, a7, b, c7); |
| 1217 | #endif // M0 > 7 |
| 1218 | |
| 1219 | lhs_offset += sizeof(DATA_TYPE); |
| 1220 | rhs_offset += sizeof(DATA_TYPE); |
| 1221 | } |
| 1222 | |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 1223 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1224 | |
| 1225 | REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; |
| 1226 | |
| 1227 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1228 | |
| 1229 | // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 1230 | CALCULATE_Z_OFFSET(M0, uint, zout, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1231 | |
| 1232 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 1233 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 1234 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
| 1235 | |
| 1236 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 1237 | |
| 1238 | // Add offset for batched GEMM |
| 1239 | dst_addr += z * dst_stride_z; |
| 1240 | |
| 1241 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 1242 | |
| 1243 | // Multiply by the weight of matrix-matrix product and store the result |
| 1244 | #if defined(ALPHA) |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 1245 | SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1246 | #endif // defined(ALPHA) |
| 1247 | |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1248 | // Add beta*bias |
| 1249 | #if defined(BETA) |
| 1250 | #if defined(BROADCAST_BIAS) |
| 1251 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); |
| 1252 | |
| 1253 | LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 1254 | |
| 1255 | #ifndef UNIT_BETA |
| 1256 | SCALE_BLOCK(1, DATA_TYPE, bias, BETA); |
| 1257 | #endif // UNIT_BIAS |
| 1258 | |
| 1259 | // c = c + bias[broadcasted] |
| 1260 | ADD_BLOCK_BROADCAST(M0, c, bias0); |
| 1261 | |
| 1262 | #else // defined(BROADCAST_BIAS) |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 1263 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z; |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1264 | |
| 1265 | LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 1266 | |
| 1267 | #ifndef UNIT_BETA |
| 1268 | SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); |
| 1269 | #endif // UNIT_BIAS |
| 1270 | |
| 1271 | // c = c + bias |
| 1272 | ADD_BLOCK(M0, c, bias); |
| 1273 | |
| 1274 | #endif // defined(BROADCAST_BIAS) |
| 1275 | #endif // defined(BETA) |
| 1276 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1277 | #if defined(ACTIVATION_TYPE) |
| 1278 | ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, c, A_VAL, B_VAL); |
| 1279 | #endif // defined(ACTIVATION_TYPE) |
| 1280 | |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1281 | // Store output block |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 1282 | STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, y, x); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1283 | |
| 1284 | #undef RHS_BLOCK_SIZE |
| 1285 | #undef RHS_OFFSET_X |
| 1286 | #undef RHS_STEP_X |
| 1287 | } |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1288 | |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1289 | #if defined(OPENCL_IMAGE_SUPPORT) |
| 1290 | /** This OpenCL kernel computes the matrix multiplication between 2 matrices. The RHS matrix is stored in OpenCL image |
| 1291 | * The LHS matrix is NOT reshaped |
| 1292 | * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed |
| 1293 | * |
| 1294 | * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel |
| 1295 | * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. |
| 1296 | * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90) |
| 1297 | * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32) |
| 1298 | * Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT |
| 1299 | * could be different from the value returned by get_image_height(rhs_img). |
| 1300 | * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4). |
| 1301 | * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2) |
| 1302 | * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2) |
| 1303 | * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time. |
SiCong Li | 3a50166 | 2020-06-26 10:02:06 +0100 | [diff] [blame] | 1304 | * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) |
| 1305 | * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1306 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 1307 | * - M0 = 1, 2, 3, 4, 5, 6, 7, 8 |
| 1308 | * - N0 = 4, 8, 16 |
| 1309 | * - K0 = 4, 8, 16 |
| 1310 | * - H0 >= 1 |
| 1311 | * |
| 1312 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 1313 | * The activation function is performed after the bias addition |
| 1314 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 1315 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
| 1316 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 1317 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 1318 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 1319 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix |
| 1320 | * |
| 1321 | * @param[in] lhs_ptr Pointer to the LHS matrix. Supported data type: F32 |
| 1322 | * @param[in] lhs_stride_x Stride of the LHS matrix in X dimension (in bytes) |
| 1323 | * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1324 | * @param[in] lhs_stride_y Stride of the LHS matrix in Y dimension (in bytes) |
| 1325 | * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1326 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS matrix |
| 1327 | * @param[in] rhs_img The RHS reshaped matrix as OpenCL image object. Supported data type: same as @p lhs_ptr |
| 1328 | * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 1329 | * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 1330 | * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes) |
| 1331 | * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 1332 | * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1333 | * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
| 1334 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr |
| 1335 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 1336 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1337 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 1338 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1339 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 1340 | * @param[in] lhs_stride_z Stride of the LHS matrix in Z dimension (in bytes) |
| 1341 | * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) |
| 1342 | * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
| 1343 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 1344 | * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) |
| 1345 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
| 1346 | */ |
| 1347 | __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs), |
| 1348 | __read_only image2d_t rhs_img, |
| 1349 | #if defined(BETA) |
| 1350 | IMAGE_DECLARATION(bias), |
| 1351 | #endif // defined(BETA) |
| 1352 | IMAGE_DECLARATION(dst), |
| 1353 | uint lhs_stride_z, |
| 1354 | uint rhs_stride_z, |
| 1355 | #if defined(BETA) |
| 1356 | uint bias_stride_z, |
| 1357 | #endif //defined(BETA) |
| 1358 | uint dst_stride_z |
| 1359 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 1360 | , |
| 1361 | uint lhs_cross_plane_pad |
| 1362 | #endif // REINTERPRET_INPUT_AS_3D |
| 1363 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 1364 | , |
| 1365 | uint dst_cross_plane_pad |
| 1366 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 1367 | ) |
| 1368 | { |
| 1369 | // Pixel unit |
| 1370 | #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(K0) |
| 1371 | |
| 1372 | #define LEFTOVER_K (K % K0) |
| 1373 | |
| 1374 | // Block size |
| 1375 | #define RHS_BLOCK_SIZE (PIXEL_UNIT * (N0)) |
| 1376 | |
| 1377 | // RHS offset and step X |
| 1378 | #if defined(RHS_INTERLEAVE) |
| 1379 | #define RHS_OFFSET_X (PIXEL_UNIT) |
| 1380 | #define RHS_STEP_X (PIXEL_UNIT * (H0)) |
| 1381 | #define RHS_STEP_LOOP (1) |
| 1382 | #else // defined(RHS_INTERLEAVE) |
| 1383 | #define RHS_OFFSET_X (RHS_BLOCK_SIZE) |
| 1384 | #define RHS_STEP_X PIXEL_UNIT |
| 1385 | #define RHS_STEP_LOOP (H0) |
| 1386 | #endif // defined(RHS_INTERLEAVE) |
| 1387 | |
| 1388 | uint x = get_global_id(0); |
| 1389 | uint y = get_global_id(1); |
| 1390 | uint z = get_global_id(2); |
| 1391 | |
| 1392 | #if defined(DUMMY_WORK_ITEMS) |
| 1393 | if((x * N0 >= N) || (y * M0 >= M)) |
| 1394 | { |
| 1395 | return; |
| 1396 | } |
| 1397 | #endif // defined(DUMMY_WORK_ITEMS) |
| 1398 | |
| 1399 | // Compute LHS matrix address |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 1400 | uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y; |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1401 | |
| 1402 | #if defined(MATRIX_B_DEPTH) |
| 1403 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 1404 | const uint z_rhs = (get_global_id(2) % MATRIX_B_DEPTH); |
| 1405 | #else // defined(MATRIX_B_DEPTH) |
| 1406 | const uint z_rhs = get_global_id(2); |
| 1407 | #endif // defined(MATRIX_B_DEPTH) |
| 1408 | |
| 1409 | // Compute RHS matrix coordinates |
| 1410 | uint x_rhs = (get_global_id(0) % H0) * (uint)RHS_OFFSET_X; |
| 1411 | const uint y_rhs = (get_global_id(0) / (uint)H0) + z_rhs * RHS_HEIGHT; |
| 1412 | |
| 1413 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0); |
| 1414 | REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); |
| 1415 | |
| 1416 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 1417 | // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
| 1418 | CALCULATE_Z_OFFSET(M0, uint, zlhs, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); |
| 1419 | |
| 1420 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 1421 | // multiply lhs_stride_z by DEPTH_GEMM3D |
| 1422 | lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D; |
| 1423 | |
| 1424 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 1425 | |
| 1426 | // Add offset for batched GEMM |
| 1427 | lhs_offset += z * lhs_stride_z; |
| 1428 | |
| 1429 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 1430 | |
| 1431 | // Initialize the accumulators |
| 1432 | REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); |
| 1433 | |
| 1434 | int i = 0; |
| 1435 | for(; i <= (K - K0); i += K0) |
| 1436 | { |
| 1437 | // Load values from LHS matrix |
| 1438 | LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs); |
| 1439 | |
| 1440 | // Load values from RHS matrix stored in a cl_image |
| 1441 | REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0); |
| 1442 | LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0); |
| 1443 | |
| 1444 | // Accumulate |
| 1445 | ARM_DOT_K0XN0(K0, a0, b, c0); |
| 1446 | #if M0 > 1 |
| 1447 | ARM_DOT_K0XN0(K0, a1, b, c1); |
| 1448 | #endif // M0 > 1 |
| 1449 | #if M0 > 2 |
| 1450 | ARM_DOT_K0XN0(K0, a2, b, c2); |
| 1451 | #endif // M0 > 2 |
| 1452 | #if M0 > 3 |
| 1453 | ARM_DOT_K0XN0(K0, a3, b, c3); |
| 1454 | #endif // M0 > 3 |
| 1455 | #if M0 > 4 |
| 1456 | ARM_DOT_K0XN0(K0, a4, b, c4); |
| 1457 | #endif // M0 > 4 |
| 1458 | #if M0 > 5 |
| 1459 | ARM_DOT_K0XN0(K0, a5, b, c5); |
| 1460 | #endif // M0 > 5 |
| 1461 | #if M0 > 6 |
| 1462 | ARM_DOT_K0XN0(K0, a6, b, c6); |
| 1463 | #endif // M0 > 6 |
| 1464 | #if M0 > 7 |
| 1465 | ARM_DOT_K0XN0(K0, a7, b, c7); |
| 1466 | #endif // M0 > 7 |
| 1467 | |
| 1468 | lhs_offset += K0 * sizeof(DATA_TYPE); |
| 1469 | x_rhs += N0 * RHS_STEP_X * RHS_STEP_LOOP; |
| 1470 | } |
| 1471 | |
| 1472 | #if LEFTOVER_K != 0 |
| 1473 | // Note: We cannot read out-of-bound elements from the RHS matrix because |
| 1474 | // the RHS width is always multiple of K0. This is not be true for the LHS matrix |
| 1475 | |
| 1476 | union UNION_VEC_TYPE |
| 1477 | { |
| 1478 | DATA_TYPE s[K0]; |
| 1479 | VEC_DATA_TYPE(DATA_TYPE, K0) |
| 1480 | v; |
| 1481 | }; |
| 1482 | |
| 1483 | union UNION_VEC_TYPE a0 = {.v = 0 }; |
| 1484 | #if M0 > 1 |
| 1485 | union UNION_VEC_TYPE a1 = {.v = 0 }; |
| 1486 | #endif // M0 > 1 |
| 1487 | #if M0 > 2 |
| 1488 | union UNION_VEC_TYPE a2 = {.v = 0 }; |
| 1489 | #endif // M0 > 2 |
| 1490 | #if M0 > 3 |
| 1491 | union UNION_VEC_TYPE a3 = {.v = 0 }; |
| 1492 | #endif // M0 > 3 |
| 1493 | #if M0 > 4 |
| 1494 | union UNION_VEC_TYPE a4 = {.v = 0 }; |
| 1495 | #endif // M0 > 4 |
| 1496 | #if M0 > 5 |
| 1497 | union UNION_VEC_TYPE a5 = {.v = 0 }; |
| 1498 | #endif // M0 > 5 |
| 1499 | #if M0 > 6 |
| 1500 | union UNION_VEC_TYPE a6 = {.v = 0 }; |
| 1501 | #endif // M0 > 6 |
| 1502 | #if M0 > 7 |
| 1503 | union UNION_VEC_TYPE a7 = {.v = 0 }; |
| 1504 | #endif // M0 > 7 |
| 1505 | |
| 1506 | REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0); |
| 1507 | |
| 1508 | // Load from RHS matrix |
| 1509 | LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0); |
| 1510 | |
| 1511 | // Load from LHS matrix |
| 1512 | for(int k = 0; k < LEFTOVER_K; ++k) |
| 1513 | { |
| 1514 | a0.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0); |
| 1515 | #if M0 > 1 |
| 1516 | a1.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1); |
| 1517 | #endif // M0 > 1 |
| 1518 | #if M0 > 2 |
| 1519 | a2.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2); |
| 1520 | #endif // M0 > 2 |
| 1521 | #if M0 > 3 |
| 1522 | a3.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3); |
| 1523 | #endif // M0 > 3 |
| 1524 | #if M0 > 4 |
| 1525 | a4.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4); |
| 1526 | #endif // M0 > 4 |
| 1527 | #if M0 > 5 |
| 1528 | a5.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5); |
| 1529 | #endif // M0 > 5 |
| 1530 | #if M0 > 6 |
| 1531 | a6.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6); |
| 1532 | #endif // M0 > 6 |
| 1533 | #if M0 > 7 |
| 1534 | a7.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7); |
| 1535 | #endif // M0 > 7 |
| 1536 | |
| 1537 | lhs_offset += sizeof(DATA_TYPE); |
| 1538 | } |
| 1539 | |
| 1540 | // Accumulate |
| 1541 | ARM_DOT_K0XN0(K0, a0.v, b, c0); |
| 1542 | #if M0 > 1 |
| 1543 | ARM_DOT_K0XN0(K0, a1.v, b, c1); |
| 1544 | #endif // M0 > 1 |
| 1545 | #if M0 > 2 |
| 1546 | ARM_DOT_K0XN0(K0, a2.v, b, c2); |
| 1547 | #endif // M0 > 2 |
| 1548 | #if M0 > 3 |
| 1549 | ARM_DOT_K0XN0(K0, a3.v, b, c3); |
| 1550 | #endif // M0 > 3 |
| 1551 | #if M0 > 4 |
| 1552 | ARM_DOT_K0XN0(K0, a4.v, b, c4); |
| 1553 | #endif // M0 > 4 |
| 1554 | #if M0 > 5 |
| 1555 | ARM_DOT_K0XN0(K0, a5.v, b, c5); |
| 1556 | #endif // M0 > 5 |
| 1557 | #if M0 > 6 |
| 1558 | ARM_DOT_K0XN0(K0, a6.v, b, c6); |
| 1559 | #endif // M0 > 6 |
| 1560 | #if M0 > 7 |
| 1561 | ARM_DOT_K0XN0(K0, a7.v, b, c7); |
| 1562 | #endif // M0 > 7 |
| 1563 | |
| 1564 | #endif // LEFTOVER_K != 0 |
| 1565 | |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 1566 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y); |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1567 | |
| 1568 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; |
| 1569 | |
| 1570 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 1571 | |
| 1572 | // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
| 1573 | CALCULATE_Z_OFFSET(M0, uint, zout, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); |
| 1574 | |
| 1575 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 1576 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 1577 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
| 1578 | |
| 1579 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 1580 | |
| 1581 | // Add offset for batched GEMM |
| 1582 | dst_addr += z * dst_stride_z; |
| 1583 | |
| 1584 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 1585 | |
| 1586 | // Multiply by the weight of matrix-matrix product and store the result |
| 1587 | #if defined(ALPHA) |
| 1588 | SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); |
| 1589 | #endif // defined(ALPHA) |
| 1590 | |
| 1591 | // Add beta*bias |
| 1592 | #if defined(BETA) |
| 1593 | #if defined(BROADCAST_BIAS) |
| 1594 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); |
| 1595 | |
| 1596 | LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 1597 | |
| 1598 | #ifndef UNIT_BETA |
| 1599 | SCALE_BLOCK(1, DATA_TYPE, bias, BETA); |
| 1600 | #endif // UNIT_BIAS |
| 1601 | |
| 1602 | // c = c + bias[broadcasted] |
| 1603 | ADD_BLOCK_BROADCAST(M0, c, bias0); |
| 1604 | |
| 1605 | #else // defined(BROADCAST_BIAS) |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 1606 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z; |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1607 | |
| 1608 | LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 1609 | |
| 1610 | #ifndef UNIT_BETA |
| 1611 | SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); |
| 1612 | #endif // UNIT_BIAS |
| 1613 | |
| 1614 | // c = c + bias |
| 1615 | ADD_BLOCK(M0, c, bias); |
| 1616 | |
| 1617 | #endif // defined(BROADCAST_BIAS) |
| 1618 | #endif // defined(BETA) |
| 1619 | |
| 1620 | #if defined(ACTIVATION_TYPE) |
| 1621 | ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, c, A_VAL, B_VAL); |
| 1622 | #endif // defined(ACTIVATION_TYPE) |
| 1623 | |
| 1624 | // Store output block |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 1625 | STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, y, x); |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1626 | |
| 1627 | #undef RHS_BLOCK_SIZE |
| 1628 | #undef RHS_OFFSET_X |
| 1629 | #undef RHS_STEP_X |
| 1630 | #undef LEFTOVER_K |
| 1631 | #undef PIXEL_UNIT |
| 1632 | } |
| 1633 | #endif // defined(OPENCL_IMAGE_SUPPORT) |
| 1634 | |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1635 | #define VFMA(a, b, c) \ |
| 1636 | ({ \ |
| 1637 | c = fma(a, b, c); \ |
| 1638 | }) |
| 1639 | |
| 1640 | #if M0 == 1 |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1641 | #define VFMA_M0xN0(i, a, b, c) \ |
| 1642 | ({ \ |
| 1643 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1644 | }) |
| 1645 | #elif M0 == 2 // M0 == 2 |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1646 | #define VFMA_M0xN0(i, a, b, c) \ |
| 1647 | ({ \ |
| 1648 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 1649 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1650 | }) |
| 1651 | #elif M0 == 3 // M0 == 3 |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1652 | #define VFMA_M0xN0(i, a, b, c) \ |
| 1653 | ({ \ |
| 1654 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 1655 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
| 1656 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1657 | }) |
| 1658 | #elif M0 == 4 // M0 == 4 |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1659 | #define VFMA_M0xN0(i, a, b, c) \ |
| 1660 | ({ \ |
| 1661 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 1662 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
| 1663 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ |
| 1664 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1665 | }) |
| 1666 | #elif M0 == 5 // M0 == 5 |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1667 | #define VFMA_M0xN0(i, a, b, c) \ |
| 1668 | ({ \ |
| 1669 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 1670 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
| 1671 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ |
| 1672 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ |
| 1673 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1674 | }) |
| 1675 | #elif M0 == 6 // M0 == 6 |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1676 | #define VFMA_M0xN0(i, a, b, c) \ |
| 1677 | ({ \ |
| 1678 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 1679 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
| 1680 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ |
| 1681 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ |
| 1682 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ |
| 1683 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1684 | }) |
| 1685 | #elif M0 == 7 // M0 == 7 |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1686 | #define VFMA_M0xN0(i, a, b, c) \ |
| 1687 | ({ \ |
| 1688 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 1689 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
| 1690 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ |
| 1691 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ |
| 1692 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ |
| 1693 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ |
| 1694 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \ |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1695 | }) |
| 1696 | #elif M0 == 8 // M0 == 8 |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1697 | #define VFMA_M0xN0(i, a, b, c) \ |
| 1698 | ({ \ |
| 1699 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 1700 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
| 1701 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ |
| 1702 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ |
| 1703 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ |
| 1704 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ |
| 1705 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \ |
| 1706 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##7).s##i), b, (c##7)); \ |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1707 | }) |
| 1708 | #else // M0 not supported |
| 1709 | #error "M0 not supported" |
| 1710 | #endif // M0 not supported |
| 1711 | |
| 1712 | /** This OpenCL kernel computes the matrix multiplication between 2 matrices. |
| 1713 | * The LHS matrix is NOT reshaped |
| 1714 | * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is NOT transposed |
| 1715 | * |
Gian Marco Iodice | b0c5037 | 2019-03-15 10:13:05 +0000 | [diff] [blame] | 1716 | * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 1717 | * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90). |
| 1718 | * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4). |
| 1719 | * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2) |
| 1720 | * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2) |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1721 | * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time. |
SiCong Li | 3a50166 | 2020-06-26 10:02:06 +0100 | [diff] [blame] | 1722 | * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) |
| 1723 | * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1724 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 1725 | * - M0 = 1, 2, 3, 4, 5, 6, 7, 8 |
| 1726 | * - N0 = 2, 3, 4, 8, 16 |
| 1727 | * - K0 = 2, 3, 4, 8, 16 |
Gian Marco Iodice | b0c5037 | 2019-03-15 10:13:05 +0000 | [diff] [blame] | 1728 | * - H0 >= 1 |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1729 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 1730 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1731 | * The activation function is performed after the bias addition |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1732 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 1733 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
| 1734 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 1735 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 1736 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 1737 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix |
| 1738 | * |
Sheri Zhang | 1a37810 | 2020-04-30 12:59:39 +0100 | [diff] [blame] | 1739 | * @param[in] lhs_ptr Pointer to the LHS matrix. Supported data type: F16/F32 |
| 1740 | * @param[in] lhs_stride_x Stride of the LHS matrix in X dimension (in bytes) |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1741 | * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
Sheri Zhang | 1a37810 | 2020-04-30 12:59:39 +0100 | [diff] [blame] | 1742 | * @param[in] lhs_stride_y Stride of the LHS matrix in Y dimension (in bytes) |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1743 | * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
Sheri Zhang | 1a37810 | 2020-04-30 12:59:39 +0100 | [diff] [blame] | 1744 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS matrix |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1745 | * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr |
| 1746 | * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes) |
| 1747 | * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1748 | * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes) |
| 1749 | * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1750 | * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1751 | * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 1752 | * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1753 | * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes) |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1754 | * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1755 | * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1756 | * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
| 1757 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr |
| 1758 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 1759 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1760 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 1761 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1762 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Sheri Zhang | 1a37810 | 2020-04-30 12:59:39 +0100 | [diff] [blame] | 1763 | * @param[in] lhs_stride_z Stride of the LHS matrix in Z dimension (in bytes) |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1764 | * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1765 | * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1766 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 1767 | * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) |
| 1768 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1769 | */ |
| 1770 | __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs), |
| 1771 | IMAGE_DECLARATION(rhs), |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1772 | #if defined(BETA) |
| 1773 | IMAGE_DECLARATION(bias), |
| 1774 | #endif // defined(BETA) |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1775 | IMAGE_DECLARATION(dst), |
| 1776 | uint lhs_stride_z, |
| 1777 | uint rhs_stride_z, |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1778 | #if defined(BETA) |
| 1779 | uint bias_stride_z, |
| 1780 | #endif //defined(BETA) |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1781 | uint dst_stride_z |
| 1782 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 1783 | , |
| 1784 | uint lhs_cross_plane_pad |
| 1785 | #endif // REINTERPRET_INPUT_AS_3D |
| 1786 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 1787 | , |
| 1788 | uint dst_cross_plane_pad |
| 1789 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 1790 | ) |
| 1791 | { |
| 1792 | // Block size |
| 1793 | #define RHS_BLOCK_SIZE ((K0) * (N0)) |
| 1794 | |
| 1795 | // RHS offset and step X |
| 1796 | #if defined(RHS_INTERLEAVE) |
| 1797 | #define RHS_OFFSET_X (N0) |
| 1798 | #define RHS_STEP_X ((N0) * (H0)) |
| 1799 | #define RHS_STEP_LOOP (1) |
| 1800 | #else // defined(RHS_INTERLEAVE) |
| 1801 | #define RHS_OFFSET_X (RHS_BLOCK_SIZE) |
| 1802 | #define RHS_STEP_X (N0) |
| 1803 | #define RHS_STEP_LOOP (H0) |
| 1804 | #endif // defined(RHS_INTERLEAVE) |
| 1805 | |
| 1806 | uint x = get_global_id(0); |
| 1807 | uint y = get_global_id(1); |
| 1808 | uint z = get_global_id(2); |
| 1809 | |
Gian Marco Iodice | b0c5037 | 2019-03-15 10:13:05 +0000 | [diff] [blame] | 1810 | #if defined(DUMMY_WORK_ITEMS) |
| 1811 | if((x * N0 >= N) || (y * M0 >= M)) |
| 1812 | { |
| 1813 | return; |
| 1814 | } |
| 1815 | #endif // defined(DUMMY_WORK_ITEMS) |
| 1816 | |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1817 | // Compute LHS matrix address |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 1818 | uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y; |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1819 | |
Sheri Zhang | 1a37810 | 2020-04-30 12:59:39 +0100 | [diff] [blame] | 1820 | // Compute RHS reshaped matrix address |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1821 | uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y; |
| 1822 | |
| 1823 | #if defined(MATRIX_B_DEPTH) |
| 1824 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 1825 | rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z; |
| 1826 | #else // defined(MATRIX_B_DEPTH) |
| 1827 | rhs_offset += z * rhs_stride_z; |
| 1828 | #endif // defined(MATRIX_B_DEPTH) |
| 1829 | |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1830 | REPEAT_VAR_INIT_TO_CONST(8, uint, zin, 0); //uint zin0=0,zin1=0,zin2=0,... zin7=0; |
| 1831 | REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); //uint zero0=0,zero1=0,zero2=0,... zero7=0; |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1832 | |
| 1833 | #if defined(REINTERPRET_INPUT_AS_3D) |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1834 | |
| 1835 | // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 1836 | CALCULATE_Z_OFFSET(M0, uint, zin, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1837 | |
| 1838 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 1839 | // multiply lhs_stride_z by DEPTH_GEMM3D |
| 1840 | lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D; |
| 1841 | |
| 1842 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 1843 | |
| 1844 | // Add offset for batched GEMM |
| 1845 | lhs_offset += z * lhs_stride_z; |
| 1846 | |
| 1847 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 1848 | |
| 1849 | // Initialize the accumulators |
| 1850 | REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(N0-1)=0; |
| 1851 | |
| 1852 | int i = 0; |
| 1853 | for(; i <= (K - K0); i += K0) |
| 1854 | { |
| 1855 | // Supported cases (M0, K0): |
| 1856 | // 1,2 - 1,3 - 1,4 - 1,8 - 1,16 |
| 1857 | // 2,2 - 2,3 - 2,4 - 2,8 - 2,16 |
| 1858 | // 3,2 - 3,3 - 3,4 - 3,8 - 3,16 |
| 1859 | // 4,2 - 4,3 - 4,4 - 4,8 - 4,16 |
| 1860 | // 5,2 - 5,3 - 5,4 - 5,8 - 5,16 |
| 1861 | // 6,2 - 6,3 - 6,4 - 6,8 - 6,16 |
| 1862 | // 7,2 - 7,3 - 7,4 - 7,8 - 7,16 |
| 1863 | // 8,2 - 8,3 - 8,4 - 8,8 - 8,16 |
| 1864 | // Load values from LHS matrix |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 1865 | LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zin); |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1866 | |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1867 | VEC_DATA_TYPE(DATA_TYPE, N0) |
| 1868 | b0; |
| 1869 | |
| 1870 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1871 | VFMA_M0xN0(0, a, b0, c); |
| 1872 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 1 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1873 | VFMA_M0xN0(1, a, b0, c); |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1874 | #if K0 > 2 |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1875 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 2 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1876 | VFMA_M0xN0(2, a, b0, c); |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1877 | #endif // K0 > 2 |
| 1878 | #if K0 > 3 |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1879 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 3 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1880 | VFMA_M0xN0(3, a, b0, c); |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1881 | #endif // K0 > 3 |
| 1882 | #if K0 > 4 |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1883 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 4 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1884 | VFMA_M0xN0(4, a, b0, c); |
| 1885 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 5 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1886 | VFMA_M0xN0(5, a, b0, c); |
| 1887 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 6 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1888 | VFMA_M0xN0(6, a, b0, c); |
| 1889 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 7 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1890 | VFMA_M0xN0(7, a, b0, c); |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1891 | #endif // K0 > 4 |
| 1892 | #if K0 > 8 |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1893 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 8 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1894 | VFMA_M0xN0(8, a, b0, c); |
| 1895 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 9 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1896 | VFMA_M0xN0(9, a, b0, c); |
| 1897 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 10 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1898 | VFMA_M0xN0(A, a, b0, c); |
| 1899 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 11 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1900 | VFMA_M0xN0(B, a, b0, c); |
| 1901 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 12 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1902 | VFMA_M0xN0(C, a, b0, c); |
| 1903 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 13 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1904 | VFMA_M0xN0(D, a, b0, c); |
| 1905 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 14 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1906 | VFMA_M0xN0(E, a, b0, c); |
| 1907 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 15 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1908 | VFMA_M0xN0(F, a, b0, c); |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1909 | #endif // K0 > 8 |
| 1910 | |
| 1911 | lhs_offset += K0 * sizeof(DATA_TYPE); |
| 1912 | rhs_offset += K0 * RHS_STEP_X * RHS_STEP_LOOP * sizeof(DATA_TYPE); |
| 1913 | } |
| 1914 | |
| 1915 | // Left-over accumulations |
| 1916 | for(; i < K; ++i) |
| 1917 | { |
| 1918 | // Load values from LHS matrix |
| 1919 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 1920 | a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zin0)); |
| 1921 | #if M0 > 1 |
| 1922 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 1923 | a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1)); |
| 1924 | #endif // M0 > 1 |
| 1925 | #if M0 > 2 |
| 1926 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 1927 | a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2)); |
| 1928 | #endif // M0 > 2 |
| 1929 | #if M0 > 3 |
| 1930 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 1931 | a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3)); |
| 1932 | #endif // M0 > 3 |
| 1933 | #if M0 > 4 |
| 1934 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 1935 | a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4)); |
| 1936 | #endif // M0 > 4 |
| 1937 | #if M0 > 5 |
| 1938 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 1939 | a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5)); |
| 1940 | #endif // M0 > 5 |
| 1941 | #if M0 > 6 |
| 1942 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 1943 | a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6)); |
| 1944 | #endif // M0 > 6 |
| 1945 | #if M0 > 7 |
| 1946 | VEC_DATA_TYPE(DATA_TYPE, 2) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1947 | a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7)); |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1948 | #endif // M0 > 7 |
| 1949 | |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1950 | VEC_DATA_TYPE(DATA_TYPE, N0) |
| 1951 | b0; |
| 1952 | |
| 1953 | b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE))); |
| 1954 | VFMA_M0xN0(0, a, b0, c); |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1955 | |
| 1956 | lhs_offset += sizeof(DATA_TYPE); |
| 1957 | rhs_offset += RHS_STEP_X * sizeof(DATA_TYPE); |
| 1958 | } |
| 1959 | |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 1960 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y); |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1961 | |
| 1962 | REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; |
| 1963 | |
| 1964 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1965 | // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 1966 | CALCULATE_Z_OFFSET(M0, uint, zout, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1967 | |
| 1968 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 1969 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 1970 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
| 1971 | |
| 1972 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 1973 | |
| 1974 | // Add offset for batched GEMM |
| 1975 | dst_addr += z * dst_stride_z; |
| 1976 | |
| 1977 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 1978 | |
| 1979 | // Multiply by the weight of matrix-matrix product and store the result |
| 1980 | #if defined(ALPHA) |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 1981 | SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 1982 | #endif // defined(ALPHA) |
| 1983 | |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1984 | // Add beta*bias |
| 1985 | #if defined(BETA) |
| 1986 | #if defined(BROADCAST_BIAS) |
| 1987 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); |
| 1988 | |
| 1989 | LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 1990 | |
| 1991 | #ifndef UNIT_BETA |
| 1992 | SCALE_BLOCK(1, DATA_TYPE, bias, BETA); |
| 1993 | #endif // UNIT_BIAS |
| 1994 | |
| 1995 | // c = c + bias[broadcasted] |
| 1996 | ADD_BLOCK_BROADCAST(M0, c, bias0); |
| 1997 | |
| 1998 | #else // defined(BROADCAST_BIAS) |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 1999 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z; |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 2000 | |
| 2001 | LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 2002 | |
| 2003 | #ifndef UNIT_BETA |
| 2004 | SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); |
| 2005 | #endif // UNIT_BIAS |
| 2006 | |
| 2007 | // c = c + bias |
| 2008 | ADD_BLOCK(M0, c, bias); |
| 2009 | |
| 2010 | #endif // defined(BROADCAST_BIAS) |
| 2011 | #endif // defined(BETA) |
| 2012 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 2013 | #if defined(ACTIVATION_TYPE) |
| 2014 | ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, c, A_VAL, B_VAL); |
| 2015 | #endif // defined(ACTIVATION_TYPE) |
| 2016 | |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 2017 | // Store output block |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 2018 | STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, y, x); |
Gian Marco Iodice | ba5e096 | 2019-03-11 12:17:44 +0000 | [diff] [blame] | 2019 | |
| 2020 | #undef RHS_BLOCK_SIZE |
| 2021 | #undef RHS_OFFSET_X |
| 2022 | #undef RHS_STEP_X |
| 2023 | } |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 2024 | |
| 2025 | #if defined(OPENCL_IMAGE_SUPPORT) |
| 2026 | /** This OpenCL kernel computes the matrix multiplication between 2 matrices. |
| 2027 | * The LHS matrix is NOT reshaped |
| 2028 | * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is NOT transposed |
| 2029 | * |
| 2030 | * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel |
| 2031 | * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. |
| 2032 | * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90). |
| 2033 | * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32) |
| 2034 | * Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT |
| 2035 | * could be different from the value returned by get_image_height(rhs_img). |
| 2036 | * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4). |
| 2037 | * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2) |
| 2038 | * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2) |
| 2039 | * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time. |
SiCong Li | 3a50166 | 2020-06-26 10:02:06 +0100 | [diff] [blame] | 2040 | * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) |
| 2041 | * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 2042 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 2043 | * - M0 = 1, 2, 3, 4, 5, 6, 7, 8 |
| 2044 | * - N0 = 4, 8, 16 |
| 2045 | * - K0 = 4, 8, 16 |
| 2046 | * - H0 >= 1 |
| 2047 | * |
| 2048 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 2049 | * The activation function is performed after the bias addition |
| 2050 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 2051 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
| 2052 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 2053 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 2054 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 2055 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix |
| 2056 | * |
| 2057 | * @param[in] lhs_ptr Pointer to the LHS matrix. Supported data type: F32 |
| 2058 | * @param[in] lhs_stride_x Stride of the LHS matrix in X dimension (in bytes) |
| 2059 | * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 2060 | * @param[in] lhs_stride_y Stride of the LHS matrix in Y dimension (in bytes) |
| 2061 | * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2062 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS matrix |
| 2063 | * @param[in] rhs_img The RHS reshaped matrix as OpenCL image object. Supported data type: same as @p lhs_ptr |
| 2064 | * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 2065 | * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 2066 | * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes) |
| 2067 | * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 2068 | * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2069 | * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
| 2070 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr |
| 2071 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 2072 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 2073 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 2074 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2075 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 2076 | * @param[in] lhs_stride_z Stride of the LHS matrix in Z dimension (in bytes) |
| 2077 | * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) |
| 2078 | * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
| 2079 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 2080 | * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) |
| 2081 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
| 2082 | */ |
| 2083 | __kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs), |
| 2084 | __read_only image2d_t rhs_img, |
| 2085 | #if defined(BETA) |
| 2086 | IMAGE_DECLARATION(bias), |
| 2087 | #endif // defined(BETA) |
| 2088 | IMAGE_DECLARATION(dst), |
| 2089 | uint lhs_stride_z, |
| 2090 | uint rhs_stride_z, |
| 2091 | #if defined(BETA) |
| 2092 | uint bias_stride_z, |
| 2093 | #endif //defined(BETA) |
| 2094 | uint dst_stride_z |
| 2095 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 2096 | , |
| 2097 | uint lhs_cross_plane_pad |
| 2098 | #endif // REINTERPRET_INPUT_AS_3D |
| 2099 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 2100 | , |
| 2101 | uint dst_cross_plane_pad |
| 2102 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 2103 | ) |
| 2104 | { |
| 2105 | // Pixel unit |
| 2106 | #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(N0) |
| 2107 | |
| 2108 | // Block size |
| 2109 | #define RHS_BLOCK_SIZE ((K0) * (PIXEL_UNIT)) |
| 2110 | |
| 2111 | // RHS offset and step X |
| 2112 | #if defined(RHS_INTERLEAVE) |
| 2113 | #define RHS_OFFSET_X (PIXEL_UNIT) |
| 2114 | #define RHS_STEP_X ((PIXEL_UNIT) * (H0)) |
| 2115 | #else // defined(RHS_INTERLEAVE) |
| 2116 | #define RHS_OFFSET_X (RHS_BLOCK_SIZE) |
| 2117 | #define RHS_STEP_X (PIXEL_UNIT) |
| 2118 | #endif // defined(RHS_INTERLEAVE) |
| 2119 | |
| 2120 | uint x = get_global_id(0); |
| 2121 | uint y = get_global_id(1); |
| 2122 | uint z = get_global_id(2); |
| 2123 | |
| 2124 | #if defined(DUMMY_WORK_ITEMS) |
| 2125 | if((x * N0 >= N) || (y * M0 >= M)) |
| 2126 | { |
| 2127 | return; |
| 2128 | } |
| 2129 | #endif // defined(DUMMY_WORK_ITEMS) |
| 2130 | |
| 2131 | // Compute LHS matrix address |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 2132 | uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y; |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 2133 | |
| 2134 | #if defined(MATRIX_B_DEPTH) |
| 2135 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 2136 | const uint z_rhs = (z % MATRIX_B_DEPTH); |
| 2137 | #else // defined(MATRIX_B_DEPTH) |
| 2138 | const uint z_rhs = z; |
| 2139 | #endif // defined(MATRIX_B_DEPTH) |
| 2140 | |
| 2141 | // Compute RHS matrix coordinates |
| 2142 | uint x_rhs = (x % H0) * (uint)RHS_OFFSET_X; |
| 2143 | const uint y_rhs = (x / (uint)H0) + z_rhs * RHS_HEIGHT; |
| 2144 | |
| 2145 | REPEAT_VAR_INIT_TO_CONST(8, uint, zin, 0); |
| 2146 | REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); |
| 2147 | |
| 2148 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 2149 | |
| 2150 | // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
| 2151 | CALCULATE_Z_OFFSET(M0, uint, zin, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); |
| 2152 | |
| 2153 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 2154 | // multiply lhs_stride_z by DEPTH_GEMM3D |
| 2155 | lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D; |
| 2156 | |
| 2157 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 2158 | |
| 2159 | // Add offset for batched GEMM |
| 2160 | lhs_offset += z * lhs_stride_z; |
| 2161 | |
| 2162 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 2163 | |
| 2164 | // Initialize the accumulators |
| 2165 | REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); |
| 2166 | |
| 2167 | int i = 0; |
| 2168 | for(; i <= (K - K0); i += K0) |
| 2169 | { |
| 2170 | // Load values from LHS matrix |
| 2171 | LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zin); |
| 2172 | |
| 2173 | VEC_DATA_TYPE(DATA_TYPE, N0) |
| 2174 | b0; |
| 2175 | |
| 2176 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs)); |
| 2177 | VFMA_M0xN0(0, a, b0, c); |
| 2178 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 1 * RHS_STEP_X), (y_rhs)); |
| 2179 | VFMA_M0xN0(1, a, b0, c); |
| 2180 | #if K0 > 2 |
| 2181 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 2 * RHS_STEP_X), (y_rhs)); |
| 2182 | VFMA_M0xN0(2, a, b0, c); |
| 2183 | #endif // K0 > 2 |
| 2184 | #if K0 > 3 |
| 2185 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 3 * RHS_STEP_X), (y_rhs)); |
| 2186 | VFMA_M0xN0(3, a, b0, c); |
| 2187 | #endif // K0 > 3 |
| 2188 | #if K0 > 4 |
| 2189 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 4 * RHS_STEP_X), (y_rhs)); |
| 2190 | VFMA_M0xN0(4, a, b0, c); |
| 2191 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 5 * RHS_STEP_X), (y_rhs)); |
| 2192 | VFMA_M0xN0(5, a, b0, c); |
| 2193 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 6 * RHS_STEP_X), (y_rhs)); |
| 2194 | VFMA_M0xN0(6, a, b0, c); |
| 2195 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 7 * RHS_STEP_X), (y_rhs)); |
| 2196 | VFMA_M0xN0(7, a, b0, c); |
| 2197 | #endif // K0 > 4 |
| 2198 | #if K0 > 8 |
| 2199 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 8 * RHS_STEP_X), (y_rhs)); |
| 2200 | VFMA_M0xN0(8, a, b0, c); |
| 2201 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 9 * RHS_STEP_X), (y_rhs)); |
| 2202 | VFMA_M0xN0(9, a, b0, c); |
| 2203 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 10 * RHS_STEP_X), (y_rhs)); |
| 2204 | VFMA_M0xN0(A, a, b0, c); |
| 2205 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 11 * RHS_STEP_X), (y_rhs)); |
| 2206 | VFMA_M0xN0(B, a, b0, c); |
| 2207 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 12 * RHS_STEP_X), (y_rhs)); |
| 2208 | VFMA_M0xN0(C, a, b0, c); |
| 2209 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 13 * RHS_STEP_X), (y_rhs)); |
| 2210 | VFMA_M0xN0(D, a, b0, c); |
| 2211 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 14 * RHS_STEP_X), (y_rhs)); |
| 2212 | VFMA_M0xN0(E, a, b0, c); |
| 2213 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 15 * RHS_STEP_X), (y_rhs)); |
| 2214 | VFMA_M0xN0(F, a, b0, c); |
| 2215 | #endif // K0 > 8 |
| 2216 | |
| 2217 | lhs_offset += K0 * sizeof(DATA_TYPE); |
| 2218 | x_rhs += K0 * RHS_STEP_X * RHS_STEP_LOOP; |
| 2219 | } |
| 2220 | |
| 2221 | // Left-over accumulations |
| 2222 | for(; i < K; ++i) |
| 2223 | { |
| 2224 | // Load values from LHS matrix |
| 2225 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 2226 | a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zin0)); |
| 2227 | #if M0 > 1 |
| 2228 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 2229 | a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1)); |
| 2230 | #endif // M0 > 1 |
| 2231 | #if M0 > 2 |
| 2232 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 2233 | a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2)); |
| 2234 | #endif // M0 > 2 |
| 2235 | #if M0 > 3 |
| 2236 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 2237 | a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3)); |
| 2238 | #endif // M0 > 3 |
| 2239 | #if M0 > 4 |
| 2240 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 2241 | a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4)); |
| 2242 | #endif // M0 > 4 |
| 2243 | #if M0 > 5 |
| 2244 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 2245 | a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5)); |
| 2246 | #endif // M0 > 5 |
| 2247 | #if M0 > 6 |
| 2248 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 2249 | a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6)); |
| 2250 | #endif // M0 > 6 |
| 2251 | #if M0 > 7 |
| 2252 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 2253 | a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7)); |
| 2254 | #endif // M0 > 7 |
| 2255 | |
| 2256 | VEC_DATA_TYPE(DATA_TYPE, N0) |
| 2257 | b0; |
| 2258 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs)); |
| 2259 | |
| 2260 | VFMA_M0xN0(0, a, b0, c); |
| 2261 | |
| 2262 | lhs_offset += sizeof(DATA_TYPE); |
| 2263 | x_rhs += RHS_STEP_X; |
| 2264 | } |
| 2265 | |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 2266 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y); |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 2267 | |
| 2268 | REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; |
| 2269 | |
| 2270 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 2271 | // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
| 2272 | CALCULATE_Z_OFFSET(M0, uint, zout, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); |
| 2273 | |
| 2274 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 2275 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 2276 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
| 2277 | |
| 2278 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 2279 | |
| 2280 | // Add offset for batched GEMM |
| 2281 | dst_addr += z * dst_stride_z; |
| 2282 | |
| 2283 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 2284 | |
| 2285 | // Multiply by the weight of matrix-matrix product and store the result |
| 2286 | #if defined(ALPHA) |
| 2287 | SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); |
| 2288 | #endif // defined(ALPHA) |
| 2289 | |
| 2290 | // Add beta*bias |
| 2291 | #if defined(BETA) |
| 2292 | #if defined(BROADCAST_BIAS) |
| 2293 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); |
| 2294 | |
| 2295 | LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 2296 | |
| 2297 | #ifndef UNIT_BETA |
| 2298 | SCALE_BLOCK(1, DATA_TYPE, bias, BETA); |
| 2299 | #endif // UNIT_BIAS |
| 2300 | |
| 2301 | // c = c + bias[broadcasted] |
| 2302 | ADD_BLOCK_BROADCAST(M0, c, bias0); |
| 2303 | |
| 2304 | #else // defined(BROADCAST_BIAS) |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 2305 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z; |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 2306 | |
| 2307 | LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 2308 | |
| 2309 | #ifndef UNIT_BETA |
| 2310 | SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); |
| 2311 | #endif // UNIT_BIAS |
| 2312 | |
| 2313 | // c = c + bias |
| 2314 | ADD_BLOCK(M0, c, bias); |
| 2315 | |
| 2316 | #endif // defined(BROADCAST_BIAS) |
| 2317 | #endif // defined(BETA) |
| 2318 | |
| 2319 | #if defined(ACTIVATION_TYPE) |
| 2320 | ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, c, A_VAL, B_VAL); |
| 2321 | #endif // defined(ACTIVATION_TYPE) |
| 2322 | |
| 2323 | // Store output block |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 2324 | STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, y, x); |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 2325 | |
| 2326 | #undef RHS_BLOCK_SIZE |
| 2327 | #undef RHS_OFFSET_X |
| 2328 | #undef RHS_STEP_X |
| 2329 | } |
| 2330 | #endif // defined(OPENCL_IMAGE_SUPPORT) |
Gian Marco Iodice | b0c5037 | 2019-03-15 10:13:05 +0000 | [diff] [blame] | 2331 | #endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(M) && defined(N) && defined(K) |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 2332 | |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 2333 | #if defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE) && defined(DATA_TYPE_ACCUMULATOR) && defined(M) && defined(N) |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2334 | |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 2335 | #if defined(MIXED_PRECISION) |
| 2336 | #if K0 == 2 |
| 2337 | #define ARM_DOT_K0(a, b, c) \ |
| 2338 | ({ \ |
| 2339 | c += a.s0 * b.s0; \ |
| 2340 | c += a.s1 * b.s1; \ |
| 2341 | }) |
| 2342 | #elif K0 == 3 // K0 == 3 |
| 2343 | #define ARM_DOT_K0(a, b, c) \ |
| 2344 | ({ \ |
| 2345 | c += a.s0 * b.s0; \ |
| 2346 | c += a.s1 * b.s1; \ |
| 2347 | c += a.s2 * b.s2; \ |
| 2348 | }) |
| 2349 | #elif K0 == 4 // K0 == 4 |
| 2350 | #define ARM_DOT_K0(a, b, c) \ |
| 2351 | ({ \ |
| 2352 | c += a.s0 * b.s0; \ |
| 2353 | c += a.s1 * b.s1; \ |
| 2354 | c += a.s2 * b.s2; \ |
| 2355 | c += a.s3 * b.s3; \ |
| 2356 | }) |
| 2357 | #elif K0 == 8 // K0 == 8 |
| 2358 | #define ARM_DOT_K0(a, b, c) \ |
| 2359 | ({ \ |
| 2360 | c += a.s0 * b.s0; \ |
| 2361 | c += a.s1 * b.s1; \ |
| 2362 | c += a.s2 * b.s2; \ |
| 2363 | c += a.s3 * b.s3; \ |
| 2364 | c += a.s4 * b.s4; \ |
| 2365 | c += a.s5 * b.s5; \ |
| 2366 | c += a.s6 * b.s6; \ |
| 2367 | c += a.s7 * b.s7; \ |
| 2368 | }) |
| 2369 | #elif K0 == 16 // K0 == 16 |
| 2370 | #define ARM_DOT_K0(a, b, c) \ |
| 2371 | ({ \ |
| 2372 | c += a.s0 * b.s0; \ |
| 2373 | c += a.s1 * b.s1; \ |
| 2374 | c += a.s2 * b.s2; \ |
| 2375 | c += a.s3 * b.s3; \ |
| 2376 | c += a.s4 * b.s4; \ |
| 2377 | c += a.s5 * b.s5; \ |
| 2378 | c += a.s6 * b.s6; \ |
| 2379 | c += a.s7 * b.s7; \ |
| 2380 | c += a.s8 * b.s8; \ |
| 2381 | c += a.s9 * b.s9; \ |
| 2382 | c += a.sA * b.sA; \ |
| 2383 | c += a.sB * b.sB; \ |
| 2384 | c += a.sC * b.sC; \ |
| 2385 | c += a.sD * b.sD; \ |
| 2386 | c += a.sE * b.sE; \ |
| 2387 | c += a.sF * b.sF; \ |
| 2388 | }) |
| 2389 | #else // K0 not supported |
| 2390 | #error "K0 value not supported" |
| 2391 | #endif // K0 conditions |
| 2392 | #else // defined(MIXED_PRECISION) |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 2393 | #if K0 == 2 |
| 2394 | #define ARM_DOT_K0(a, b, c) \ |
| 2395 | ({ \ |
| 2396 | c = fma(a.s0, b.s0, c); \ |
| 2397 | c = fma(a.s1, b.s1, c); \ |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2398 | }) |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 2399 | #elif K0 == 3 // K0 == 3 |
| 2400 | #define ARM_DOT_K0(a, b, c) \ |
| 2401 | ({ \ |
| 2402 | c = fma(a.s0, b.s0, c); \ |
| 2403 | c = fma(a.s1, b.s1, c); \ |
| 2404 | c = fma(a.s2, b.s2, c); \ |
| 2405 | }) |
| 2406 | #elif K0 == 4 // K0 == 4 |
| 2407 | #define ARM_DOT_K0(a, b, c) \ |
| 2408 | ({ \ |
| 2409 | c = fma(a.s0, b.s0, c); \ |
| 2410 | c = fma(a.s1, b.s1, c); \ |
| 2411 | c = fma(a.s2, b.s2, c); \ |
| 2412 | c = fma(a.s3, b.s3, c); \ |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2413 | }) |
| 2414 | #elif K0 == 8 // K0 == 8 |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 2415 | #define ARM_DOT_K0(a, b, c) \ |
| 2416 | ({ \ |
| 2417 | c = fma(a.s0, b.s0, c); \ |
| 2418 | c = fma(a.s1, b.s1, c); \ |
| 2419 | c = fma(a.s2, b.s2, c); \ |
| 2420 | c = fma(a.s3, b.s3, c); \ |
| 2421 | c = fma(a.s4, b.s4, c); \ |
| 2422 | c = fma(a.s5, b.s5, c); \ |
| 2423 | c = fma(a.s6, b.s6, c); \ |
| 2424 | c = fma(a.s7, b.s7, c); \ |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2425 | }) |
| 2426 | #elif K0 == 16 // K0 == 16 |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 2427 | #define ARM_DOT_K0(a, b, c) \ |
| 2428 | ({ \ |
| 2429 | c = fma(a.s0, b.s0, c); \ |
| 2430 | c = fma(a.s1, b.s1, c); \ |
| 2431 | c = fma(a.s2, b.s2, c); \ |
| 2432 | c = fma(a.s3, b.s3, c); \ |
| 2433 | c = fma(a.s4, b.s4, c); \ |
| 2434 | c = fma(a.s5, b.s5, c); \ |
| 2435 | c = fma(a.s6, b.s6, c); \ |
| 2436 | c = fma(a.s7, b.s7, c); \ |
| 2437 | c = fma(a.s8, b.s8, c); \ |
| 2438 | c = fma(a.s9, b.s9, c); \ |
| 2439 | c = fma(a.sA, b.sA, c); \ |
| 2440 | c = fma(a.sB, b.sB, c); \ |
| 2441 | c = fma(a.sC, b.sC, c); \ |
| 2442 | c = fma(a.sD, b.sD, c); \ |
| 2443 | c = fma(a.sE, b.sE, c); \ |
| 2444 | c = fma(a.sF, b.sF, c); \ |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2445 | }) |
| 2446 | #else // K0 not supported |
| 2447 | #error "K0 value not supported" |
| 2448 | #endif // K0 conditions |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 2449 | #endif // defined(MIXED_PRECISION) |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2450 | |
| 2451 | #if N0 == 2 |
| 2452 | #define ARM_DOT_K0XN0(a, b, c) \ |
| 2453 | ({ \ |
| 2454 | ARM_DOT_K0((a), (b##0), (c.s0)); \ |
| 2455 | ARM_DOT_K0((a), (b##1), (c.s1)); \ |
| 2456 | }) |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 2457 | #elif N0 == 3 // N0 == 3 |
| 2458 | #define ARM_DOT_K0XN0(a, b, c) \ |
| 2459 | ({ \ |
| 2460 | ARM_DOT_K0((a), (b##0), (c.s0)); \ |
| 2461 | ARM_DOT_K0((a), (b##1), (c.s1)); \ |
| 2462 | ARM_DOT_K0((a), (b##2), (c.s2)); \ |
| 2463 | }) |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2464 | #elif N0 == 4 // N0 == 4 |
| 2465 | #define ARM_DOT_K0XN0(a, b, c) \ |
| 2466 | ({ \ |
| 2467 | ARM_DOT_K0((a), (b##0), (c.s0)); \ |
| 2468 | ARM_DOT_K0((a), (b##1), (c.s1)); \ |
| 2469 | ARM_DOT_K0((a), (b##2), (c.s2)); \ |
| 2470 | ARM_DOT_K0((a), (b##3), (c.s3)); \ |
| 2471 | }) |
| 2472 | #elif N0 == 8 // N0 == 8 |
| 2473 | #define ARM_DOT_K0XN0(a, b, c) \ |
| 2474 | ({ \ |
| 2475 | ARM_DOT_K0((a), (b##0), (c.s0)); \ |
| 2476 | ARM_DOT_K0((a), (b##1), (c.s1)); \ |
| 2477 | ARM_DOT_K0((a), (b##2), (c.s2)); \ |
| 2478 | ARM_DOT_K0((a), (b##3), (c.s3)); \ |
| 2479 | ARM_DOT_K0((a), (b##4), (c.s4)); \ |
| 2480 | ARM_DOT_K0((a), (b##5), (c.s5)); \ |
| 2481 | ARM_DOT_K0((a), (b##6), (c.s6)); \ |
| 2482 | ARM_DOT_K0((a), (b##7), (c.s7)); \ |
| 2483 | }) |
| 2484 | #elif N0 == 16 // N0 == 16 |
| 2485 | #define ARM_DOT_K0XN0(a, b, c) \ |
| 2486 | ({ \ |
| 2487 | ARM_DOT_K0((a), (b##0), (c.s0)); \ |
| 2488 | ARM_DOT_K0((a), (b##1), (c.s1)); \ |
| 2489 | ARM_DOT_K0((a), (b##2), (c.s2)); \ |
| 2490 | ARM_DOT_K0((a), (b##3), (c.s3)); \ |
| 2491 | ARM_DOT_K0((a), (b##4), (c.s4)); \ |
| 2492 | ARM_DOT_K0((a), (b##5), (c.s5)); \ |
| 2493 | ARM_DOT_K0((a), (b##6), (c.s6)); \ |
| 2494 | ARM_DOT_K0((a), (b##7), (c.s7)); \ |
| 2495 | ARM_DOT_K0((a), (b##8), (c.s8)); \ |
| 2496 | ARM_DOT_K0((a), (b##9), (c.s9)); \ |
| 2497 | ARM_DOT_K0((a), (b##A), (c.sA)); \ |
| 2498 | ARM_DOT_K0((a), (b##B), (c.sB)); \ |
| 2499 | ARM_DOT_K0((a), (b##C), (c.sC)); \ |
| 2500 | ARM_DOT_K0((a), (b##D), (c.sD)); \ |
| 2501 | ARM_DOT_K0((a), (b##E), (c.sE)); \ |
| 2502 | ARM_DOT_K0((a), (b##F), (c.sF)); \ |
| 2503 | }) |
| 2504 | #else // N0 not supported |
| 2505 | #error "N0 value not supported" |
| 2506 | #endif // N0 conditions |
| 2507 | |
| 2508 | /** This OpenCL kernel computes the matrix multiplication between 2 matrices. |
| 2509 | * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be NOT transposed |
| 2510 | * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be transposed |
| 2511 | * |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 2512 | * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) |
| 2513 | * @note The data type used for the accumulators must be passed at compile time using -DDATA_TYPE_ACCUMULATOR (e.g. -DDATA_TYPE_ACCUMULATOR=float) |
| 2514 | * @note The F16 computation also supports mixed precision through the option -DMIXED_PRECISION passed at compile time. If enabled, DATA_TYPE_ACCUMULATOR should be set to float |
Gian Marco Iodice | b0c5037 | 2019-03-15 10:13:05 +0000 | [diff] [blame] | 2515 | * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 2516 | * @note The GEMM's dimensions M, N and K must be passed at compile time using -DM, -DN and -DK (e.g. -DM=52, -DN=90 and -DK=24). |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 2517 | * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (e.g. -DM0=4, -DN0=8, -DK0=4). |
| 2518 | * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (e.g. -DV0=2) |
| 2519 | * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2) |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2520 | * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time. |
| 2521 | * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time. |
| 2522 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 2523 | * - M0 = 2, 3, 4, 5, 6, 7, 8 |
Gian Marco Iodice | bacfec5 | 2019-01-11 11:30:55 +0000 | [diff] [blame] | 2524 | * - N0 = 2, 3, 4, 8, 16 |
| 2525 | * - K0 = 2, 3, 4, 8, 16 |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 2526 | * - V0 >= 1 |
| 2527 | * - H0 >= 1 |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2528 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 2529 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 2530 | * The activation function is performed after the bias addition |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 2531 | * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2532 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 2533 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 2534 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 2535 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped |
| 2536 | * |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 2537 | * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: F16/F32 |
| 2538 | * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes) |
| 2539 | * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 2540 | * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes) |
| 2541 | * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2542 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix |
| 2543 | * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr |
| 2544 | * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes) |
| 2545 | * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 2546 | * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes) |
| 2547 | * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2548 | * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix |
| 2549 | * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 2550 | * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 2551 | * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes) |
| 2552 | * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 2553 | * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2554 | * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
| 2555 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr |
| 2556 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 2557 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 2558 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 2559 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2560 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Gian Marco Iodice | e5563d9 | 2020-06-25 17:18:36 +0100 | [diff] [blame] | 2561 | * @param[in] k Number of columns in LHS matrix and rows in RHS matrix not reshaped. |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 2562 | * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) |
| 2563 | * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) |
| 2564 | * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
| 2565 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 2566 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2567 | */ |
| 2568 | __kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs), |
| 2569 | IMAGE_DECLARATION(rhs), |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 2570 | #if defined(BETA) |
| 2571 | IMAGE_DECLARATION(bias), |
| 2572 | #endif // defined(BETA) |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2573 | IMAGE_DECLARATION(dst), |
Gian Marco Iodice | e5563d9 | 2020-06-25 17:18:36 +0100 | [diff] [blame] | 2574 | uint k, |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2575 | uint lhs_stride_z, |
| 2576 | uint rhs_stride_z, |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 2577 | #if defined(BETA) |
| 2578 | uint bias_stride_z, |
| 2579 | #endif //defined(BETA) |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2580 | uint dst_stride_z |
| 2581 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 2582 | , |
| 2583 | uint dst_cross_plane_pad |
| 2584 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 2585 | ) |
| 2586 | { |
| 2587 | // Block size |
| 2588 | #define LHS_BLOCK_SIZE ((K0) * (M0)) |
| 2589 | |
| 2590 | #if defined(LHS_INTERLEAVE) |
| 2591 | #define LHS_OFFSET_X (K0) |
| 2592 | #define LHS_STEP_X ((K0) * (V0)) |
| 2593 | #define LHS_STEP_LOOP (1) |
| 2594 | #else // defined(INTERLEAVE) |
| 2595 | #define LHS_OFFSET_X (LHS_BLOCK_SIZE) |
| 2596 | #define LHS_STEP_X (K0) |
| 2597 | #define LHS_STEP_LOOP (V0) |
| 2598 | #endif // defined(INTERLEAVE) |
| 2599 | |
| 2600 | // Block size |
| 2601 | #define RHS_BLOCK_SIZE ((K0) * (N0)) |
| 2602 | |
| 2603 | // RHS offset and step X |
| 2604 | #if defined(RHS_INTERLEAVE) |
| 2605 | #define RHS_OFFSET_X (K0) |
| 2606 | #define RHS_STEP_X ((K0) * (H0)) |
| 2607 | #define RHS_STEP_LOOP (1) |
| 2608 | #else // defined(RHS_INTERLEAVE) |
| 2609 | #define RHS_OFFSET_X (RHS_BLOCK_SIZE) |
| 2610 | #define RHS_STEP_X (K0) |
| 2611 | #define RHS_STEP_LOOP (H0) |
| 2612 | #endif // defined(RHS_INTERLEAVE) |
| 2613 | |
Gian Marco Iodice | b0c5037 | 2019-03-15 10:13:05 +0000 | [diff] [blame] | 2614 | #if defined(DUMMY_WORK_ITEMS) |
| 2615 | if((get_global_id(0) * N0 >= N) || (get_global_id(1) * M0 >= M)) |
| 2616 | { |
| 2617 | return; |
| 2618 | } |
| 2619 | #endif // defined(DUMMY_WORK_ITEMS) |
| 2620 | |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2621 | // Compute LHS matrix address |
| 2622 | __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (get_global_id(1) % V0) * (uint)LHS_OFFSET_X * sizeof(DATA_TYPE) + (get_global_id(1) / V0) * (uint)lhs_stride_y + |
| 2623 | (get_global_id(2) * lhs_stride_z); |
| 2624 | |
| 2625 | // Compute RHS matrix address |
| 2626 | __global uchar *rhs_addr = rhs_ptr + rhs_offset_first_element_in_bytes + (get_global_id(0) % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (get_global_id(0) / (uint)H0) * rhs_stride_y; |
| 2627 | |
| 2628 | #if defined(MATRIX_B_DEPTH) |
| 2629 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 2630 | rhs_addr += (get_global_id(2) % MATRIX_B_DEPTH) * rhs_stride_z; |
| 2631 | #else // defined(MATRIX_B_DEPTH) |
| 2632 | rhs_addr += get_global_id(2) * rhs_stride_z; |
| 2633 | #endif // defined(MATRIX_B_DEPTH) |
| 2634 | |
| 2635 | // Initialize the accumulators |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 2636 | REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0), c, 0); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2637 | |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 2638 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0; |
| 2639 | REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 2640 | |
Gian Marco Iodice | e5563d9 | 2020-06-25 17:18:36 +0100 | [diff] [blame] | 2641 | for(int i = 0; i < k; i += K0) |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2642 | { |
| 2643 | // Supported cases (M0, K0): |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 2644 | // 1,2 - 1,3 - 1,4 - 1,8 - 1,16 |
| 2645 | // 2,2 - 2,3 - 2,4 - 2,8 - 2,16 |
| 2646 | // 3,2 - 3,3 - 3,4 - 3,8 - 3,16 |
| 2647 | // 4,2 - 4,3 - 4,4 - 4,8 - 4,16 |
| 2648 | // 5,2 - 5,3 - 5,4 - 5,8 - 5,16 |
| 2649 | // 6,2 - 6,3 - 6,4 - 6,8 - 6,16 |
| 2650 | // 7,2 - 7,3 - 7,4 - 7,8 - 7,16 |
| 2651 | // 8,2 - 8,3 - 8,4 - 8,8 - 8,16 |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2652 | // Load values from LHS matrix |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 2653 | LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_addr, 0, LHS_STEP_X * sizeof(DATA_TYPE), zlhs); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2654 | |
| 2655 | // Load values from RHS matrix |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 2656 | LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_addr, 0, RHS_STEP_X * sizeof(DATA_TYPE), zero); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2657 | |
| 2658 | // Accumulate |
| 2659 | ARM_DOT_K0XN0(a0, b, c0); |
| 2660 | #if M0 > 1 |
| 2661 | ARM_DOT_K0XN0(a1, b, c1); |
| 2662 | #endif // M0 > 1 |
| 2663 | #if M0 > 2 |
| 2664 | ARM_DOT_K0XN0(a2, b, c2); |
| 2665 | #endif // M0 > 2 |
| 2666 | #if M0 > 3 |
| 2667 | ARM_DOT_K0XN0(a3, b, c3); |
| 2668 | #endif // M0 > 3 |
| 2669 | #if M0 > 4 |
| 2670 | ARM_DOT_K0XN0(a4, b, c4); |
| 2671 | #endif // M0 > 4 |
| 2672 | #if M0 > 5 |
| 2673 | ARM_DOT_K0XN0(a5, b, c5); |
| 2674 | #endif // M0 > 5 |
| 2675 | #if M0 > 6 |
| 2676 | ARM_DOT_K0XN0(a6, b, c6); |
| 2677 | #endif // M0 > 6 |
| 2678 | #if M0 > 7 |
| 2679 | ARM_DOT_K0XN0(a7, b, c7); |
| 2680 | #endif // M0 > 7 |
| 2681 | |
| 2682 | lhs_addr += (M0 * LHS_STEP_X * LHS_STEP_LOOP) * sizeof(DATA_TYPE); |
| 2683 | rhs_addr += (N0 * RHS_STEP_X * RHS_STEP_LOOP) * sizeof(DATA_TYPE); |
| 2684 | } |
| 2685 | |
| 2686 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * dst_stride_y); |
| 2687 | |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 2688 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2689 | |
| 2690 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2691 | |
| 2692 | // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 2693 | CALCULATE_Z_OFFSET(M0, uint, zout, get_global_id(1), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2694 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 2695 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 2696 | dst_addr += get_global_id(2) * dst_stride_z * DEPTH_GEMM3D; |
| 2697 | |
| 2698 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 2699 | |
| 2700 | // Add offset for batched GEMM |
| 2701 | dst_addr += get_global_id(2) * dst_stride_z; |
| 2702 | |
| 2703 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 2704 | |
| 2705 | // Multiply by the weight of matrix-matrix product and store the result |
| 2706 | #if defined(ALPHA) |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 2707 | SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2708 | #endif // defined(ALPHA) |
| 2709 | |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 2710 | // Add beta*bias |
| 2711 | #if defined(BETA) |
| 2712 | #if defined(BROADCAST_BIAS) |
| 2713 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); |
| 2714 | |
| 2715 | LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 2716 | |
| 2717 | #ifndef UNIT_BETA |
| 2718 | SCALE_BLOCK(1, DATA_TYPE, bias, BETA); |
| 2719 | #endif // UNIT_BIAS |
| 2720 | |
| 2721 | // c = c + bias[broadcasted] |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 2722 | #if defined(MIXED_PRECISION) |
| 2723 | CONVERT_BLOCK(1, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp); |
| 2724 | ADD_BLOCK_BROADCAST(M0, c, bias_hp0); |
| 2725 | #else // defined(MIXED_PRECISION) |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 2726 | ADD_BLOCK_BROADCAST(M0, c, bias0); |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 2727 | #endif // defined(MIXED_PRECISION) |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 2728 | |
| 2729 | #else // defined(BROADCAST_BIAS) |
| 2730 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id( |
| 2731 | 2) * bias_stride_z; |
| 2732 | |
| 2733 | LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 2734 | |
| 2735 | #ifndef UNIT_BETA |
| 2736 | SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); |
| 2737 | #endif // UNIT_BIAS |
| 2738 | |
| 2739 | // c = c + bias |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 2740 | #if defined(MIXED_PRECISION) |
| 2741 | CONVERT_BLOCK(M0, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp); |
| 2742 | ADD_BLOCK(M0, c, bias_hp); |
| 2743 | #else // defined(MIXED_PRECISION) |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 2744 | ADD_BLOCK(M0, c, bias); |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 2745 | #endif // defined(MIXED_PRECISION) |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 2746 | |
| 2747 | #endif // defined(BROADCAST_BIAS) |
| 2748 | #endif // defined(BETA) |
| 2749 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 2750 | #if defined(ACTIVATION_TYPE) |
Georgios Pinitas | a07ce15 | 2019-10-11 17:38:50 +0100 | [diff] [blame] | 2751 | #if defined(MIXED_PRECISION) |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 2752 | ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE_ACCUMULATOR, c, A_VAL, B_VAL); |
Georgios Pinitas | a07ce15 | 2019-10-11 17:38:50 +0100 | [diff] [blame] | 2753 | #else // defined(MIXED_PRECISION) |
| 2754 | ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, c, A_VAL, B_VAL); |
| 2755 | #endif // defined(MIXED_PRECISION) |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 2756 | #endif // defined(ACTIVATION_TYPE) |
| 2757 | |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2758 | // Store output block |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 2759 | #if defined(MIXED_PRECISION) |
| 2760 | CONVERT_STORE_BLOCK(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout); |
| 2761 | #else // defined(MIXED_PRECISION) |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 2762 | STORE_BLOCK(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout); |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 2763 | #endif // defined(MIXED_PRECISION) |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 2764 | |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2765 | #undef LHS_BLOCK_SIZE |
| 2766 | #undef LHS_OFFSET_X |
| 2767 | #undef LHS_STEP_X |
| 2768 | #undef RHS_BLOCK_SIZE |
| 2769 | #undef RHS_OFFSET_X |
| 2770 | #undef RHS_STEP_X |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 2771 | #undef LHS_STEP_LOOP |
| 2772 | #undef RHS_STEP_LOOP |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 2773 | } |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2774 | |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 2775 | #if defined(OPENCL_IMAGE_SUPPORT) |
| 2776 | /** This OpenCL kernel computes the matrix multiplication between 2 matrices. The RHS matrix is stored in OpenCL image object. |
| 2777 | * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be NOT transposed |
| 2778 | * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be transposed |
| 2779 | * |
| 2780 | * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel |
| 2781 | * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) |
| 2782 | * @note The data type used for the accumulators must be passed at compile time using -DDATA_TYPE_ACCUMULATOR (e.g. -DDATA_TYPE_ACCUMULATOR=float) |
| 2783 | * @note The F16 computation also supports mixed precision through the option -DMIXED_PRECISION passed at compile time. If enabled, DATA_TYPE_ACCUMULATOR should be set to float |
| 2784 | * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. |
| 2785 | * @note The GEMM's dimensions M, N and K must be passed at compile time using -DM, -DN and -DK (e.g. -DM=52, -DN=90 and -DK=24). |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 2786 | * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32) |
| 2787 | * Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT |
| 2788 | * could be different from the value returned by get_image_height(rhs_img). |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 2789 | * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (e.g. -DM0=4, -DN0=8, -DK0=4). |
| 2790 | * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (e.g. -DV0=2) |
| 2791 | * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2) |
| 2792 | * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time. |
| 2793 | * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time. |
| 2794 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 2795 | * - M0 = 2, 3, 4, 5, 6, 7, 8 |
| 2796 | * - N0 = 4, 8, 16 |
| 2797 | * - K0 = 4, 8, 16 |
| 2798 | * - V0 >= 1 |
| 2799 | * - H0 >= 1 |
| 2800 | * |
| 2801 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 2802 | * The activation function is performed after the bias addition |
| 2803 | * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: |
| 2804 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 2805 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 2806 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 2807 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped |
| 2808 | * |
| 2809 | * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: F32 |
| 2810 | * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes) |
| 2811 | * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 2812 | * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes) |
| 2813 | * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2814 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix |
| 2815 | * @param[in] rhs_img The RHS reshaped matrix as OpenCL image object. Supported data type: same as @p lhs_ptr |
| 2816 | * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 2817 | * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 2818 | * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes) |
| 2819 | * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 2820 | * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2821 | * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
| 2822 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr |
| 2823 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 2824 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 2825 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 2826 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2827 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Gian Marco Iodice | e5563d9 | 2020-06-25 17:18:36 +0100 | [diff] [blame] | 2828 | * @param[in] k Number of columns in LHS matrix and rows in RHS matrix not reshaped. |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 2829 | * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) |
| 2830 | * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) |
| 2831 | * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
| 2832 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 2833 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
| 2834 | */ |
| 2835 | __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture(IMAGE_DECLARATION(lhs), |
| 2836 | __read_only image2d_t rhs_img, |
| 2837 | #if defined(BETA) |
| 2838 | IMAGE_DECLARATION(bias), |
| 2839 | #endif // defined(BETA) |
| 2840 | IMAGE_DECLARATION(dst), |
Gian Marco Iodice | e5563d9 | 2020-06-25 17:18:36 +0100 | [diff] [blame] | 2841 | uint k, |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 2842 | uint lhs_stride_z, |
| 2843 | uint rhs_stride_z, |
| 2844 | #if defined(BETA) |
| 2845 | uint bias_stride_z, |
| 2846 | #endif //defined(BETA) |
| 2847 | uint dst_stride_z |
| 2848 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 2849 | , |
| 2850 | uint dst_cross_plane_pad |
| 2851 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 2852 | ) |
| 2853 | { |
| 2854 | // Pixel unit |
| 2855 | #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(K0) |
| 2856 | |
| 2857 | // Block size |
| 2858 | #define LHS_BLOCK_SIZE ((K0) * (M0)) |
| 2859 | |
| 2860 | #if defined(LHS_INTERLEAVE) |
| 2861 | #define LHS_OFFSET_X (K0) |
| 2862 | #define LHS_STEP_X ((K0) * (V0)) |
| 2863 | #define LHS_STEP_LOOP (1) |
| 2864 | #else // defined(INTERLEAVE) |
| 2865 | #define LHS_OFFSET_X (LHS_BLOCK_SIZE) |
| 2866 | #define LHS_STEP_X (K0) |
| 2867 | #define LHS_STEP_LOOP (V0) |
| 2868 | #endif // defined(INTERLEAVE) |
| 2869 | |
| 2870 | // Block size |
| 2871 | #define RHS_BLOCK_SIZE (PIXEL_UNIT * (N0)) |
| 2872 | |
| 2873 | // RHS offset and step X |
| 2874 | #if defined(RHS_INTERLEAVE) |
| 2875 | #define RHS_OFFSET_X (PIXEL_UNIT) |
| 2876 | #define RHS_STEP_X (PIXEL_UNIT * (H0)) |
| 2877 | #define RHS_STEP_LOOP (1) |
| 2878 | #else // defined(RHS_INTERLEAVE) |
| 2879 | #define RHS_OFFSET_X (RHS_BLOCK_SIZE) |
| 2880 | #define RHS_STEP_X PIXEL_UNIT |
| 2881 | #define RHS_STEP_LOOP (H0) |
| 2882 | #endif // defined(RHS_INTERLEAVE) |
| 2883 | |
| 2884 | #if defined(DUMMY_WORK_ITEMS) |
| 2885 | if((get_global_id(0) * N0 >= N) || (get_global_id(1) * M0 >= M)) |
| 2886 | { |
| 2887 | return; |
| 2888 | } |
| 2889 | #endif // defined(DUMMY_WORK_ITEMS) |
| 2890 | |
| 2891 | // Compute LHS matrix address |
| 2892 | __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (get_global_id(1) % V0) * (uint)LHS_OFFSET_X * sizeof(DATA_TYPE) + (get_global_id(1) / V0) * (uint)lhs_stride_y + |
| 2893 | (get_global_id(2) * lhs_stride_z); |
| 2894 | |
| 2895 | #if defined(MATRIX_B_DEPTH) |
| 2896 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 2897 | const uint z_rhs = (get_global_id(2) % MATRIX_B_DEPTH); |
| 2898 | #else // defined(MATRIX_B_DEPTH) |
| 2899 | const uint z_rhs = get_global_id(2); |
| 2900 | #endif // defined(MATRIX_B_DEPTH) |
| 2901 | |
| 2902 | // Compute RHS matrix coordinates |
| 2903 | uint x_rhs = (get_global_id(0) % H0) * (uint)RHS_OFFSET_X; |
| 2904 | const uint y_rhs = (get_global_id(0) / (uint)H0) + z_rhs * RHS_HEIGHT; |
| 2905 | |
| 2906 | // Initialize the accumulators |
| 2907 | REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0), c, 0); |
| 2908 | |
| 2909 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0; |
| 2910 | REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); |
| 2911 | |
| 2912 | for(int i = 0; i < K; i += K0) |
| 2913 | { |
| 2914 | // Load values from LHS matrix |
| 2915 | LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_addr, 0, LHS_STEP_X * sizeof(DATA_TYPE), zlhs); |
| 2916 | |
| 2917 | // Load values from RHS matrix stored in a cl_image |
| 2918 | REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0); |
| 2919 | LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0); |
| 2920 | |
| 2921 | // Accumulate |
| 2922 | ARM_DOT_K0XN0(a0, b, c0); |
| 2923 | #if M0 > 1 |
| 2924 | ARM_DOT_K0XN0(a1, b, c1); |
| 2925 | #endif // M0 > 1 |
| 2926 | #if M0 > 2 |
| 2927 | ARM_DOT_K0XN0(a2, b, c2); |
| 2928 | #endif // M0 > 2 |
| 2929 | #if M0 > 3 |
| 2930 | ARM_DOT_K0XN0(a3, b, c3); |
| 2931 | #endif // M0 > 3 |
| 2932 | #if M0 > 4 |
| 2933 | ARM_DOT_K0XN0(a4, b, c4); |
| 2934 | #endif // M0 > 4 |
| 2935 | #if M0 > 5 |
| 2936 | ARM_DOT_K0XN0(a5, b, c5); |
| 2937 | #endif // M0 > 5 |
| 2938 | #if M0 > 6 |
| 2939 | ARM_DOT_K0XN0(a6, b, c6); |
| 2940 | #endif // M0 > 6 |
| 2941 | #if M0 > 7 |
| 2942 | ARM_DOT_K0XN0(a7, b, c7); |
| 2943 | #endif // M0 > 7 |
| 2944 | |
| 2945 | lhs_addr += (M0 * LHS_STEP_X * LHS_STEP_LOOP) * sizeof(DATA_TYPE); |
| 2946 | |
| 2947 | x_rhs += N0 * RHS_STEP_X * RHS_STEP_LOOP; |
| 2948 | } |
| 2949 | |
| 2950 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * dst_stride_y); |
| 2951 | |
| 2952 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); |
| 2953 | |
| 2954 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 2955 | |
| 2956 | // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
| 2957 | CALCULATE_Z_OFFSET(M0, uint, zout, get_global_id(1), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); |
| 2958 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 2959 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 2960 | dst_addr += get_global_id(2) * dst_stride_z * DEPTH_GEMM3D; |
| 2961 | |
| 2962 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 2963 | |
| 2964 | // Add offset for batched GEMM |
| 2965 | dst_addr += get_global_id(2) * dst_stride_z; |
| 2966 | |
| 2967 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 2968 | |
| 2969 | // Multiply by the weight of matrix-matrix product and store the result |
| 2970 | #if defined(ALPHA) |
| 2971 | SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); |
| 2972 | #endif // defined(ALPHA) |
| 2973 | |
| 2974 | // Add beta*bias |
| 2975 | #if defined(BETA) |
| 2976 | #if defined(BROADCAST_BIAS) |
| 2977 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); |
| 2978 | |
| 2979 | LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 2980 | |
| 2981 | #ifndef UNIT_BETA |
| 2982 | SCALE_BLOCK(1, DATA_TYPE, bias, BETA); |
| 2983 | #endif // UNIT_BIAS |
| 2984 | |
| 2985 | // c = c + bias[broadcasted] |
| 2986 | #if defined(MIXED_PRECISION) |
| 2987 | CONVERT_BLOCK(1, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp); |
| 2988 | ADD_BLOCK_BROADCAST(M0, c, bias_hp0); |
| 2989 | #else // defined(MIXED_PRECISION) |
| 2990 | ADD_BLOCK_BROADCAST(M0, c, bias0); |
| 2991 | #endif // defined(MIXED_PRECISION) |
| 2992 | |
| 2993 | #else // defined(BROADCAST_BIAS) |
| 2994 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id( |
| 2995 | 2) * bias_stride_z; |
| 2996 | |
| 2997 | LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 2998 | |
| 2999 | #ifndef UNIT_BETA |
| 3000 | SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); |
| 3001 | #endif // UNIT_BIAS |
| 3002 | |
| 3003 | // c = c + bias |
| 3004 | #if defined(MIXED_PRECISION) |
| 3005 | CONVERT_BLOCK(M0, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp); |
| 3006 | ADD_BLOCK(M0, c, bias_hp); |
| 3007 | #else // defined(MIXED_PRECISION) |
| 3008 | ADD_BLOCK(M0, c, bias); |
| 3009 | #endif // defined(MIXED_PRECISION) |
| 3010 | |
| 3011 | #endif // defined(BROADCAST_BIAS) |
| 3012 | #endif // defined(BETA) |
| 3013 | |
| 3014 | #if defined(ACTIVATION_TYPE) |
| 3015 | #if defined(MIXED_PRECISION) |
| 3016 | ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE_ACCUMULATOR, c, A_VAL, B_VAL); |
| 3017 | #else // defined(MIXED_PRECISION) |
| 3018 | ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, c, A_VAL, B_VAL); |
| 3019 | #endif // defined(MIXED_PRECISION) |
| 3020 | #endif // defined(ACTIVATION_TYPE) |
| 3021 | |
| 3022 | // Store output block |
| 3023 | #if defined(MIXED_PRECISION) |
| 3024 | CONVERT_STORE_BLOCK(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout); |
| 3025 | #else // defined(MIXED_PRECISION) |
| 3026 | STORE_BLOCK(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout); |
| 3027 | #endif // defined(MIXED_PRECISION) |
| 3028 | |
| 3029 | #undef LHS_BLOCK_SIZE |
| 3030 | #undef LHS_OFFSET_X |
| 3031 | #undef LHS_STEP_X |
| 3032 | #undef RHS_BLOCK_SIZE |
| 3033 | #undef RHS_OFFSET_X |
| 3034 | #undef RHS_STEP_X |
| 3035 | #undef PIXEL_UNIT |
| 3036 | #undef LHS_STEP_LOOP |
| 3037 | #undef RHS_STEP_LOOP |
| 3038 | } |
| 3039 | #endif // defined(OPENCL_IMAGE_SUPPORT) |
| 3040 | |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3041 | #if defined(LHS_TRANSPOSE) |
| 3042 | |
| 3043 | #define VTYPE(TYPE, SIZE) VEC_DATA_TYPE(TYPE, SIZE) |
| 3044 | |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3045 | #if defined(MIXED_PRECISION) |
| 3046 | |
| 3047 | #if(GPU_ARCH == GPU_ARCH_MIDGARD) |
| 3048 | #define ARM_VFMA(N0, a, b, c) c += (CONVERT(a, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0))) * (CONVERT(b, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0))); |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3049 | #else // GPU_ARCH == GPU_ARCH_MIDGARD |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3050 | #define ARM_VFMA(N0, a, b, c) c = fma((CONVERT(a, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0))), (CONVERT(b, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0))), (c)); |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3051 | #endif // GPU_ARCH == GPU_ARCH_MIDGARD |
| 3052 | |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3053 | #else // defined(MIXED_PRECISION |
| 3054 | |
| 3055 | #if(GPU_ARCH == GPU_ARCH_MIDGARD) |
| 3056 | #define ARM_VFMA(N0, a, b, c) c += (a) * (b); |
| 3057 | #else // GPU_ARCH == GPU_ARCH_MIDGARD |
| 3058 | #define ARM_VFMA(N0, a, b, c) c = fma((a), (b), (c)); |
| 3059 | #endif // GPU_ARCH == GPU_ARCH_MIDGARD |
| 3060 | |
| 3061 | #endif // defined(MIXED_PRECISION) |
| 3062 | |
| 3063 | #define ARM_VVM_T_NT_1xN0x1(N0, TYPE, a, b, C) \ |
| 3064 | ({ \ |
| 3065 | ARM_VFMA(N0, (VTYPE(TYPE, N0))(a), b, (C##0)); \ |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3066 | }) |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3067 | #define ARM_VVM_T_NT_2xN0x1(N0, TYPE, a, b, C) \ |
| 3068 | ({ \ |
| 3069 | ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s0), b, (C##0)); \ |
| 3070 | ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s1), b, (C##1)); \ |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3071 | }) |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3072 | #define ARM_VVM_T_NT_3xN0x1(N0, TYPE, a, b, C) \ |
| 3073 | ({ \ |
| 3074 | ARM_VVM_T_NT_2xN0x1(N0, TYPE, a, b, C); \ |
| 3075 | ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s2), b, (C##2)); \ |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3076 | }) |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3077 | #define ARM_VVM_T_NT_4xN0x1(N0, TYPE, a, b, C) \ |
| 3078 | ({ \ |
| 3079 | ARM_VVM_T_NT_3xN0x1(N0, TYPE, a, b, C); \ |
| 3080 | ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s3), b, (C##3)); \ |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3081 | }) |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3082 | #define ARM_VVM_T_NT_8xN0x1(N0, TYPE, a, b, C) \ |
| 3083 | ({ \ |
| 3084 | ARM_VVM_T_NT_4xN0x1(N0, TYPE, a, b, C); \ |
| 3085 | ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s4), b, (C##4)); \ |
| 3086 | ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s5), b, (C##5)); \ |
| 3087 | ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s6), b, (C##6)); \ |
| 3088 | ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s7), b, (C##7)); \ |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3089 | }) |
| 3090 | |
| 3091 | // Factory macro for the column-vector (transposed) by row-vector (not transposed) multiplication. K0 = 1 |
| 3092 | // a is the column-vector (transposed) |
| 3093 | // b is the row-vector (not transposed) |
| 3094 | // C is the output matrix |
| 3095 | // Lower case is a vector (a, b) |
| 3096 | // Upper case is a matrix (C) |
| 3097 | #define ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, a, b, C) ARM_VVM_T_NT_##M0##xN0x1(N0, TYPE, a, b, C) |
| 3098 | |
| 3099 | #define ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, A, B, C) \ |
| 3100 | ({ \ |
| 3101 | ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##0), (B##0), C); \ |
| 3102 | }) |
| 3103 | #define ARM_MM_T_NT_M0xN0x2(M0, N0, TYPE, A, B, C) \ |
| 3104 | ({ \ |
| 3105 | ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, A, B, C); \ |
| 3106 | ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##1), (B##1), C); \ |
| 3107 | }) |
| 3108 | #define ARM_MM_T_NT_M0xN0x3(M0, N0, TYPE, A, B, C) \ |
| 3109 | ({ \ |
| 3110 | ARM_MM_T_NT_M0xN0x2(M0, N0, TYPE, A, B, C); \ |
| 3111 | ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##2), (B##2), C); \ |
| 3112 | }) |
| 3113 | #define ARM_MM_T_NT_M0xN0x4(M0, N0, TYPE, A, B, C) \ |
| 3114 | ({ \ |
| 3115 | ARM_MM_T_NT_M0xN0x3(M0, N0, TYPE, A, B, C); \ |
| 3116 | ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##3), (B##3), C); \ |
| 3117 | }) |
| 3118 | #define ARM_MM_T_NT_M0xN0x8(M0, N0, TYPE, A, B, C) \ |
| 3119 | ({ \ |
| 3120 | ARM_MM_T_NT_M0xN0x4(M0, N0, TYPE, A, B, C); \ |
| 3121 | ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##4), (B##4), C); \ |
| 3122 | ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##5), (B##5), C); \ |
| 3123 | ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##6), (B##6), C); \ |
| 3124 | ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##7), (B##7), C); \ |
| 3125 | }) |
| 3126 | #define ARM_MM_T_NT_M0xN0x16(M0, N0, TYPE, A, B, C) \ |
| 3127 | ({ \ |
| 3128 | ARM_MM_T_NT_M0xN0x8(M0, N0, TYPE, A, B, C); \ |
| 3129 | ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##8), (B##8), C); \ |
| 3130 | ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##9), (B##9), C); \ |
| 3131 | ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##A), (B##A), C); \ |
| 3132 | ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##B), (B##B), C); \ |
| 3133 | ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##C), (B##C), C); \ |
| 3134 | ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##D), (B##D), C); \ |
| 3135 | ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##E), (B##E), C); \ |
| 3136 | ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##F), (B##F), C); \ |
| 3137 | }) |
| 3138 | |
| 3139 | // Factory macro for the matrix (transposed) by matrix (not transposed) multiplication. |
| 3140 | // The dimensions for this matrix multiplications are defined through M0, N0 and K0 |
| 3141 | // The dimensions supported are: |
| 3142 | // M0: 1, 2, 3, 4, 8 |
| 3143 | // N0: 1, 2, 3, 4, 8, 16 |
| 3144 | // K0: 1, 2, 3, 4, 8, 16 |
| 3145 | // This macro calls the vector-by-matrix macro K0 times |
| 3146 | // A, B and C are matrices |
Gian Marco Iodice | 05639f6 | 2019-09-24 12:05:06 +0100 | [diff] [blame] | 3147 | #define ARM_MM_T_NT(M0, N0, K0, TYPE, A, B, C) \ |
| 3148 | CONCAT(ARM_MM_T_NT_M0xN0x, K0) \ |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3149 | (M0, N0, TYPE, A, B, C) |
| 3150 | |
| 3151 | /** This OpenCL kernel computes the matrix multiplication between 2 matrices. |
| 3152 | * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be transposed |
| 3153 | * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be NOT transposed |
| 3154 | * |
| 3155 | * @note LHS_TRANSPOSE should be passed at compile time in order to compile this OpenCL kernel (e.g. -DLHS_TRANSPOSE). |
| 3156 | * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 3157 | * @note The GEMM's dimensions M, N and K must be passed at compile time using -DM, -DN and -DK (e.g. -DM=52, -DN=90 and -DK=24). |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3158 | * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (e.g. -DM0=4, -DN0=8, -DK0=4). |
| 3159 | * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (e.g. -DV0=2) |
| 3160 | * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2) |
| 3161 | * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time. |
| 3162 | * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time. |
| 3163 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 3164 | * - M0 = 2, 3, 4, 8 |
| 3165 | * - N0 = 2, 3, 4, 8, 16 |
| 3166 | * - K0 = 2, 3, 4, 8, 16 |
| 3167 | * - V0 >= 1 |
| 3168 | * - H0 >= 1 |
| 3169 | * |
| 3170 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 3171 | * The activation function is performed after the bias addition |
| 3172 | * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: |
| 3173 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 3174 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 3175 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 3176 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped |
| 3177 | * |
| 3178 | * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: F16/F32 |
| 3179 | * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes) |
| 3180 | * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 3181 | * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes) |
| 3182 | * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3183 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix |
| 3184 | * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr |
| 3185 | * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes) |
| 3186 | * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 3187 | * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes) |
| 3188 | * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3189 | * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix |
| 3190 | * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 3191 | * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 3192 | * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes) |
| 3193 | * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 3194 | * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3195 | * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
| 3196 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr |
| 3197 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 3198 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 3199 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 3200 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3201 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Gian Marco Iodice | e5563d9 | 2020-06-25 17:18:36 +0100 | [diff] [blame] | 3202 | * @param[in] k Number of columns in LHS matrix and rows in RHS matrix not reshaped. |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3203 | * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) |
| 3204 | * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) |
| 3205 | * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
| 3206 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 3207 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
| 3208 | */ |
| 3209 | __kernel void gemm_mm_reshaped_lhs_t_rhs_nt(IMAGE_DECLARATION(lhs), |
| 3210 | IMAGE_DECLARATION(rhs), |
| 3211 | #if defined(BETA) |
| 3212 | IMAGE_DECLARATION(bias), |
| 3213 | #endif // defined(BETA) |
| 3214 | IMAGE_DECLARATION(dst), |
Gian Marco Iodice | e5563d9 | 2020-06-25 17:18:36 +0100 | [diff] [blame] | 3215 | uint k, |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3216 | uint lhs_stride_z, |
| 3217 | uint rhs_stride_z, |
| 3218 | #if defined(BETA) |
| 3219 | uint bias_stride_z, |
| 3220 | #endif //defined(BETA) |
| 3221 | uint dst_stride_z |
| 3222 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 3223 | , |
| 3224 | uint dst_cross_plane_pad |
| 3225 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 3226 | ) |
| 3227 | { |
| 3228 | // Block size |
| 3229 | #define LHS_BLOCK_SIZE ((K0) * (M0)) |
| 3230 | |
| 3231 | #if defined(LHS_INTERLEAVE) |
| 3232 | #define LHS_OFFSET_X (M0) |
| 3233 | #define LHS_STEP_X ((M0) * (V0)) |
| 3234 | #define LHS_STEP_LOOP (1) |
| 3235 | #else // defined(INTERLEAVE) |
| 3236 | #define LHS_OFFSET_X (LHS_BLOCK_SIZE) |
| 3237 | #define LHS_STEP_X (M0) |
| 3238 | #define LHS_STEP_LOOP (V0) |
| 3239 | #endif // defined(INTERLEAVE) |
| 3240 | |
| 3241 | // Block size |
| 3242 | #define RHS_BLOCK_SIZE ((K0) * (N0)) |
| 3243 | |
| 3244 | // RHS offset and step X |
| 3245 | #if defined(RHS_INTERLEAVE) |
| 3246 | #define RHS_OFFSET_X (N0) |
| 3247 | #define RHS_STEP_X ((N0) * (H0)) |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3248 | #else // defined(RHS_INTERLEAVE) |
| 3249 | #define RHS_OFFSET_X (RHS_BLOCK_SIZE) |
| 3250 | #define RHS_STEP_X (N0) |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3251 | #endif // defined(RHS_INTERLEAVE) |
| 3252 | |
| 3253 | const uint x = get_global_id(0); |
| 3254 | const uint y = get_global_id(1); |
| 3255 | const uint z = get_global_id(2); |
| 3256 | |
| 3257 | #if defined(DUMMY_WORK_ITEMS) |
| 3258 | if((x * N0 >= N) || (y * M0 >= M)) |
| 3259 | { |
| 3260 | return; |
| 3261 | } |
| 3262 | #endif // defined(DUMMY_WORK_ITEMS) |
| 3263 | |
| 3264 | // Compute LHS matrix address |
| 3265 | __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (y % V0) * (uint)LHS_OFFSET_X * sizeof(DATA_TYPE) + (y / V0) * (uint)lhs_stride_y + (z * lhs_stride_z); |
| 3266 | |
| 3267 | // Compute RHS matrix address |
| 3268 | __global uchar *rhs_addr = rhs_ptr + rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y; |
| 3269 | |
| 3270 | #if defined(MATRIX_B_DEPTH) |
| 3271 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 3272 | rhs_addr += (z % MATRIX_B_DEPTH) * rhs_stride_z; |
| 3273 | #else // defined(MATRIX_B_DEPTH) |
| 3274 | rhs_addr += z * rhs_stride_z; |
| 3275 | #endif // defined(MATRIX_B_DEPTH) |
| 3276 | |
| 3277 | // Initialize the accumulators |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3278 | REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0), c, 0); |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3279 | |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3280 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0); |
| 3281 | |
Gian Marco Iodice | 05639f6 | 2019-09-24 12:05:06 +0100 | [diff] [blame] | 3282 | __global DATA_TYPE *lhs = (__global DATA_TYPE *)(lhs_addr); |
| 3283 | __global DATA_TYPE *rhs = (__global DATA_TYPE *)(rhs_addr); |
| 3284 | |
Gian Marco Iodice | e5563d9 | 2020-06-25 17:18:36 +0100 | [diff] [blame] | 3285 | for(int i = 0; i < k; i += K0) |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3286 | { |
Gian Marco Iodice | 05639f6 | 2019-09-24 12:05:06 +0100 | [diff] [blame] | 3287 | VEC_DATA_TYPE(DATA_TYPE, M0) |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 3288 | a0; |
Gian Marco Iodice | 05639f6 | 2019-09-24 12:05:06 +0100 | [diff] [blame] | 3289 | VEC_DATA_TYPE(DATA_TYPE, N0) |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 3290 | b0; |
| 3291 | |
| 3292 | a0 = VLOAD(M0)(0, lhs); |
Gian Marco Iodice | 05639f6 | 2019-09-24 12:05:06 +0100 | [diff] [blame] | 3293 | b0 = VLOAD(N0)(0, rhs); |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3294 | |
Gian Marco Iodice | 05639f6 | 2019-09-24 12:05:06 +0100 | [diff] [blame] | 3295 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3296 | |
Gian Marco Iodice | 05639f6 | 2019-09-24 12:05:06 +0100 | [diff] [blame] | 3297 | lhs += LHS_STEP_X; |
| 3298 | rhs += RHS_STEP_X; |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3299 | |
Gian Marco Iodice | 05639f6 | 2019-09-24 12:05:06 +0100 | [diff] [blame] | 3300 | #if K0 > 1 |
| 3301 | a0 = VLOAD(M0)(0, lhs); |
| 3302 | b0 = VLOAD(N0)(0, rhs); |
| 3303 | |
| 3304 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3305 | |
| 3306 | lhs += LHS_STEP_X; |
| 3307 | rhs += RHS_STEP_X; |
| 3308 | #endif // K0 > 1 |
| 3309 | |
| 3310 | #if K0 > 2 |
| 3311 | a0 = VLOAD(M0)(0, lhs); |
| 3312 | b0 = VLOAD(N0)(0, rhs); |
| 3313 | |
| 3314 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3315 | |
| 3316 | lhs += LHS_STEP_X; |
| 3317 | rhs += RHS_STEP_X; |
| 3318 | #endif // K0 > 2 |
| 3319 | |
| 3320 | #if K0 > 3 |
| 3321 | a0 = VLOAD(M0)(0, lhs); |
| 3322 | b0 = VLOAD(N0)(0, rhs); |
| 3323 | |
| 3324 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3325 | |
| 3326 | lhs += LHS_STEP_X; |
| 3327 | rhs += RHS_STEP_X; |
| 3328 | #endif // K0 > 3 |
| 3329 | |
| 3330 | #if K0 > 4 |
| 3331 | a0 = VLOAD(M0)(0, lhs); |
| 3332 | b0 = VLOAD(N0)(0, rhs); |
| 3333 | |
| 3334 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3335 | |
| 3336 | lhs += LHS_STEP_X; |
| 3337 | rhs += RHS_STEP_X; |
| 3338 | |
| 3339 | a0 = VLOAD(M0)(0, lhs); |
| 3340 | b0 = VLOAD(N0)(0, rhs); |
| 3341 | |
| 3342 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3343 | |
| 3344 | lhs += LHS_STEP_X; |
| 3345 | rhs += RHS_STEP_X; |
| 3346 | |
| 3347 | a0 = VLOAD(M0)(0, lhs); |
| 3348 | b0 = VLOAD(N0)(0, rhs); |
| 3349 | |
| 3350 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3351 | |
| 3352 | lhs += LHS_STEP_X; |
| 3353 | rhs += RHS_STEP_X; |
| 3354 | |
| 3355 | a0 = VLOAD(M0)(0, lhs); |
| 3356 | b0 = VLOAD(N0)(0, rhs); |
| 3357 | |
| 3358 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3359 | |
| 3360 | lhs += LHS_STEP_X; |
| 3361 | rhs += RHS_STEP_X; |
| 3362 | #endif // K0 > 4 |
| 3363 | |
| 3364 | #if K0 > 8 |
| 3365 | a0 = VLOAD(M0)(0, lhs); |
| 3366 | b0 = VLOAD(N0)(0, rhs); |
| 3367 | |
| 3368 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3369 | |
| 3370 | lhs += LHS_STEP_X; |
| 3371 | rhs += RHS_STEP_X; |
| 3372 | |
| 3373 | a0 = VLOAD(M0)(0, lhs); |
| 3374 | b0 = VLOAD(N0)(0, rhs); |
| 3375 | |
| 3376 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3377 | |
| 3378 | lhs += LHS_STEP_X; |
| 3379 | rhs += RHS_STEP_X; |
| 3380 | |
| 3381 | a0 = VLOAD(M0)(0, lhs); |
| 3382 | b0 = VLOAD(N0)(0, rhs); |
| 3383 | |
| 3384 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3385 | |
| 3386 | lhs += LHS_STEP_X; |
| 3387 | rhs += RHS_STEP_X; |
| 3388 | |
| 3389 | a0 = VLOAD(M0)(0, lhs); |
| 3390 | b0 = VLOAD(N0)(0, rhs); |
| 3391 | |
| 3392 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3393 | |
| 3394 | lhs += LHS_STEP_X; |
| 3395 | rhs += RHS_STEP_X; |
| 3396 | |
| 3397 | a0 = VLOAD(M0)(0, lhs); |
| 3398 | b0 = VLOAD(N0)(0, rhs); |
| 3399 | |
| 3400 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3401 | |
| 3402 | lhs += LHS_STEP_X; |
| 3403 | rhs += RHS_STEP_X; |
| 3404 | |
| 3405 | a0 = VLOAD(M0)(0, lhs); |
| 3406 | b0 = VLOAD(N0)(0, rhs); |
| 3407 | |
| 3408 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3409 | |
| 3410 | lhs += LHS_STEP_X; |
| 3411 | rhs += RHS_STEP_X; |
| 3412 | |
| 3413 | a0 = VLOAD(M0)(0, lhs); |
| 3414 | b0 = VLOAD(N0)(0, rhs); |
| 3415 | |
| 3416 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3417 | |
| 3418 | lhs += LHS_STEP_X; |
| 3419 | rhs += RHS_STEP_X; |
| 3420 | |
| 3421 | a0 = VLOAD(M0)(0, lhs); |
| 3422 | b0 = VLOAD(N0)(0, rhs); |
| 3423 | |
| 3424 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3425 | |
| 3426 | lhs += LHS_STEP_X; |
| 3427 | rhs += RHS_STEP_X; |
| 3428 | #endif // K0 > 8 |
| 3429 | |
| 3430 | #ifndef LHS_INTERLEAVE |
| 3431 | lhs += (M0 * K0 * (V0 - 1)); |
| 3432 | #endif // LHS_INTERLEAVE |
| 3433 | |
| 3434 | #ifndef RHS_INTERLEAVE |
| 3435 | rhs += (N0 * K0 * (H0 - 1)); |
| 3436 | #endif // RHS_INTERLEAVE |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3437 | } |
| 3438 | |
| 3439 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * (uint)M0 * dst_stride_y); |
| 3440 | |
| 3441 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); |
| 3442 | |
| 3443 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 3444 | |
| 3445 | // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
| 3446 | CALCULATE_Z_OFFSET(M0, uint, zout, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); |
| 3447 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 3448 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 3449 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
| 3450 | |
| 3451 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 3452 | |
| 3453 | // Add offset for batched GEMM |
| 3454 | dst_addr += z * dst_stride_z; |
| 3455 | |
| 3456 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 3457 | |
| 3458 | // Multiply by the weight of matrix-matrix product and store the result |
| 3459 | #if defined(ALPHA) |
| 3460 | SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); |
| 3461 | #endif // defined(ALPHA) |
| 3462 | |
| 3463 | // Add beta*bias |
| 3464 | #if defined(BETA) |
| 3465 | #if defined(BROADCAST_BIAS) |
| 3466 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)); |
| 3467 | |
| 3468 | LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 3469 | |
| 3470 | #ifndef UNIT_BETA |
| 3471 | SCALE_BLOCK(1, DATA_TYPE, bias, BETA); |
| 3472 | #endif // UNIT_BIAS |
| 3473 | |
| 3474 | // c = c + bias[broadcasted] |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3475 | #if defined(MIXED_PRECISION) |
| 3476 | CONVERT_BLOCK(1, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp); |
| 3477 | ADD_BLOCK_BROADCAST(M0, c, bias_hp0); |
| 3478 | #else // defined(MIXED_PRECISION) |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3479 | ADD_BLOCK_BROADCAST(M0, c, bias0); |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3480 | #endif // defined(MIXED_PRECISION) |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3481 | |
| 3482 | #else // defined(BROADCAST_BIAS) |
| 3483 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * (uint)M0 * bias_stride_y) + z * bias_stride_z; |
| 3484 | |
| 3485 | LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 3486 | |
| 3487 | #ifndef UNIT_BETA |
| 3488 | SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); |
| 3489 | #endif // UNIT_BIAS |
| 3490 | |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3491 | #if defined(MIXED_PRECISION) |
| 3492 | CONVERT_BLOCK(M0, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp); |
| 3493 | ADD_BLOCK(M0, c, bias_hp); |
| 3494 | #else // defined(MIXED_PRECISION) |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3495 | ADD_BLOCK(M0, c, bias); |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3496 | #endif // defined(MIXED_PRECISION) |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3497 | |
| 3498 | #endif // defined(BROADCAST_BIAS) |
| 3499 | #endif // defined(BETA) |
| 3500 | |
| 3501 | #if defined(ACTIVATION_TYPE) |
Georgios Pinitas | a07ce15 | 2019-10-11 17:38:50 +0100 | [diff] [blame] | 3502 | #if defined(MIXED_PRECISION) |
| 3503 | ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE_ACCUMULATOR, c, A_VAL, B_VAL); |
| 3504 | #else // defined(MIXED_PRECISION) |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3505 | ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, c, A_VAL, B_VAL); |
Georgios Pinitas | a07ce15 | 2019-10-11 17:38:50 +0100 | [diff] [blame] | 3506 | #endif // defined(MIXED_PRECISION) |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3507 | #endif // defined(ACTIVATION_TYPE) |
| 3508 | |
| 3509 | // Store output block |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3510 | #if defined(MIXED_PRECISION) |
| 3511 | CONVERT_STORE_BLOCK(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout); |
| 3512 | #else // defined(MIXED_PRECISION) |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3513 | STORE_BLOCK(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout); |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 3514 | #endif // defined(MIXED_PRECISION) |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3515 | |
| 3516 | #undef LHS_BLOCK_SIZE |
| 3517 | #undef LHS_OFFSET_X |
| 3518 | #undef LHS_STEP_X |
| 3519 | #undef RHS_BLOCK_SIZE |
| 3520 | #undef RHS_OFFSET_X |
| 3521 | #undef RHS_STEP_X |
| 3522 | } |
| 3523 | |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 3524 | #if defined(OPENCL_IMAGE_SUPPORT) |
| 3525 | /** This OpenCL kernel computes the matrix multiplication between 2 matrices. The RHS matrix is stored in OpenCL image object. |
| 3526 | * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be transposed |
| 3527 | * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be NOT transposed |
| 3528 | * |
| 3529 | * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel |
| 3530 | * @note LHS_TRANSPOSE should be passed at compile time in order to compile this OpenCL kernel (e.g. -DLHS_TRANSPOSE). |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 3531 | * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. |
| 3532 | * @note The GEMM's dimensions M, N and K must be passed at compile time using -DM, -DN and -DK (e.g. -DM=52, -DN=90 and -DK=24). |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 3533 | * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32) |
| 3534 | * Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT |
| 3535 | * could be different from the value returned by get_image_height(rhs_img). |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 3536 | * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (e.g. -DM0=4, -DN0=8, -DK0=4). |
| 3537 | * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (e.g. -DV0=2) |
| 3538 | * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2) |
| 3539 | * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time. |
| 3540 | * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time. |
| 3541 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 3542 | * - M0 = 2, 3, 4, 8 |
| 3543 | * - N0 = 4, 8, 16 |
| 3544 | * - K0 = 4, 8, 16 |
| 3545 | * - V0 >= 1 |
| 3546 | * - H0 >= 1 |
| 3547 | * |
| 3548 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 3549 | * The activation function is performed after the bias addition |
| 3550 | * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: |
| 3551 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 3552 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 3553 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 3554 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped |
| 3555 | * |
| 3556 | * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: F32 |
| 3557 | * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes) |
| 3558 | * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 3559 | * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes) |
| 3560 | * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3561 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix |
| 3562 | * @param[in] rhs_img The RHS reshaped matrix as cl_image 2d. Supported data type: same as @p lhs_ptr |
| 3563 | * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 3564 | * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 3565 | * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes) |
| 3566 | * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 3567 | * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3568 | * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
| 3569 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr |
| 3570 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 3571 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 3572 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 3573 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3574 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Gian Marco Iodice | e5563d9 | 2020-06-25 17:18:36 +0100 | [diff] [blame] | 3575 | * @param[in] k Number of columns in LHS matrix and rows in RHS matrix not reshaped. |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 3576 | * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) |
| 3577 | * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) |
| 3578 | * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
| 3579 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 3580 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
| 3581 | */ |
| 3582 | __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture(IMAGE_DECLARATION(lhs), |
| 3583 | __read_only image2d_t rhs_img, |
| 3584 | #if defined(BETA) |
| 3585 | IMAGE_DECLARATION(bias), |
| 3586 | #endif // defined(BETA) |
| 3587 | IMAGE_DECLARATION(dst), |
Gian Marco Iodice | e5563d9 | 2020-06-25 17:18:36 +0100 | [diff] [blame] | 3588 | uint k, |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 3589 | uint lhs_stride_z, |
| 3590 | uint rhs_stride_z, |
| 3591 | #if defined(BETA) |
| 3592 | uint bias_stride_z, |
| 3593 | #endif //defined(BETA) |
| 3594 | uint dst_stride_z |
| 3595 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 3596 | , |
| 3597 | uint dst_cross_plane_pad |
| 3598 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 3599 | ) |
| 3600 | { |
| 3601 | // Pixel unit |
| 3602 | #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(N0) |
| 3603 | |
| 3604 | // Block size |
| 3605 | #define LHS_BLOCK_SIZE ((K0) * (M0)) |
| 3606 | |
| 3607 | #if defined(LHS_INTERLEAVE) |
| 3608 | #define LHS_OFFSET_X (M0) |
| 3609 | #define LHS_STEP_X ((M0) * (V0)) |
| 3610 | #define LHS_STEP_LOOP (1) |
| 3611 | #else // defined(INTERLEAVE) |
| 3612 | #define LHS_OFFSET_X (LHS_BLOCK_SIZE) |
| 3613 | #define LHS_STEP_X (M0) |
| 3614 | #define LHS_STEP_LOOP (V0) |
| 3615 | #endif // defined(INTERLEAVE) |
| 3616 | |
| 3617 | // Block size |
| 3618 | #define RHS_BLOCK_SIZE ((K0) * (PIXEL_UNIT)) |
| 3619 | |
| 3620 | // RHS offset and step X |
| 3621 | #if defined(RHS_INTERLEAVE) |
| 3622 | #define RHS_OFFSET_X (PIXEL_UNIT) |
| 3623 | #define RHS_STEP_X ((PIXEL_UNIT) * (H0)) |
| 3624 | #else // defined(RHS_INTERLEAVE) |
| 3625 | #define RHS_OFFSET_X (RHS_BLOCK_SIZE) |
| 3626 | #define RHS_STEP_X (PIXEL_UNIT) |
| 3627 | #endif // defined(RHS_INTERLEAVE) |
| 3628 | |
| 3629 | const uint x = get_global_id(0); |
| 3630 | const uint y = get_global_id(1); |
| 3631 | const uint z = get_global_id(2); |
| 3632 | |
| 3633 | #if defined(DUMMY_WORK_ITEMS) |
| 3634 | if((x * N0 >= N) || (y * M0 >= M)) |
| 3635 | { |
| 3636 | return; |
| 3637 | } |
| 3638 | #endif // defined(DUMMY_WORK_ITEMS) |
| 3639 | |
| 3640 | // Compute LHS matrix address |
| 3641 | __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (y % V0) * (uint)LHS_OFFSET_X * sizeof(DATA_TYPE) + (y / V0) * (uint)lhs_stride_y + (z * lhs_stride_z); |
| 3642 | |
| 3643 | #if defined(MATRIX_B_DEPTH) |
| 3644 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 3645 | const uint z_rhs = (z % MATRIX_B_DEPTH); |
| 3646 | #else // defined(MATRIX_B_DEPTH) |
| 3647 | const uint z_rhs = z; |
| 3648 | #endif // defined(MATRIX_B_DEPTH) |
| 3649 | |
| 3650 | // Compute RHS matrix coordinates |
| 3651 | uint x_rhs = (x % H0) * (uint)RHS_OFFSET_X; |
| 3652 | const uint y_rhs = (x / (uint)H0) + z_rhs * RHS_HEIGHT; |
| 3653 | |
| 3654 | // Initialize the accumulators |
| 3655 | REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0), c, 0); |
| 3656 | |
| 3657 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0); |
| 3658 | |
| 3659 | __global DATA_TYPE *lhs = (__global DATA_TYPE *)(lhs_addr); |
| 3660 | |
| 3661 | for(int i = 0; i < K; i += K0) |
| 3662 | { |
| 3663 | VEC_DATA_TYPE(DATA_TYPE, M0) |
| 3664 | a0; |
| 3665 | VEC_DATA_TYPE(DATA_TYPE, N0) |
| 3666 | b0; |
| 3667 | |
| 3668 | a0 = VLOAD(M0)(0, lhs); |
| 3669 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs)); |
| 3670 | |
| 3671 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3672 | |
| 3673 | lhs += LHS_STEP_X; |
| 3674 | |
| 3675 | #if K0 > 1 |
| 3676 | a0 = VLOAD(M0)(0, lhs); |
| 3677 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 1 * RHS_STEP_X), (y_rhs)); |
| 3678 | |
| 3679 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3680 | |
| 3681 | lhs += LHS_STEP_X; |
| 3682 | #endif // K0 > 1 |
| 3683 | |
| 3684 | #if K0 > 2 |
| 3685 | a0 = VLOAD(M0)(0, lhs); |
| 3686 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 2 * RHS_STEP_X), (y_rhs)); |
| 3687 | |
| 3688 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3689 | |
| 3690 | lhs += LHS_STEP_X; |
| 3691 | #endif // K0 > 2 |
| 3692 | |
| 3693 | #if K0 > 3 |
| 3694 | a0 = VLOAD(M0)(0, lhs); |
| 3695 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 3 * RHS_STEP_X), (y_rhs)); |
| 3696 | |
| 3697 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3698 | |
| 3699 | lhs += LHS_STEP_X; |
| 3700 | #endif // K0 > 3 |
| 3701 | |
| 3702 | #if K0 > 4 |
| 3703 | a0 = VLOAD(M0)(0, lhs); |
| 3704 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 4 * RHS_STEP_X), (y_rhs)); |
| 3705 | |
| 3706 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3707 | |
| 3708 | lhs += LHS_STEP_X; |
| 3709 | |
| 3710 | a0 = VLOAD(M0)(0, lhs); |
| 3711 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 5 * RHS_STEP_X), (y_rhs)); |
| 3712 | |
| 3713 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3714 | |
| 3715 | lhs += LHS_STEP_X; |
| 3716 | |
| 3717 | a0 = VLOAD(M0)(0, lhs); |
| 3718 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 6 * RHS_STEP_X), (y_rhs)); |
| 3719 | |
| 3720 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3721 | |
| 3722 | lhs += LHS_STEP_X; |
| 3723 | |
| 3724 | a0 = VLOAD(M0)(0, lhs); |
| 3725 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 7 * RHS_STEP_X), (y_rhs)); |
| 3726 | |
| 3727 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3728 | |
| 3729 | lhs += LHS_STEP_X; |
| 3730 | #endif // K0 > 4 |
| 3731 | |
| 3732 | #if K0 > 8 |
| 3733 | a0 = VLOAD(M0)(0, lhs); |
| 3734 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 8 * RHS_STEP_X), (y_rhs)); |
| 3735 | |
| 3736 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3737 | |
| 3738 | lhs += LHS_STEP_X; |
| 3739 | |
| 3740 | a0 = VLOAD(M0)(0, lhs); |
| 3741 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 9 * RHS_STEP_X), (y_rhs)); |
| 3742 | |
| 3743 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3744 | |
| 3745 | lhs += LHS_STEP_X; |
| 3746 | |
| 3747 | a0 = VLOAD(M0)(0, lhs); |
| 3748 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 10 * RHS_STEP_X), (y_rhs)); |
| 3749 | |
| 3750 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3751 | |
| 3752 | lhs += LHS_STEP_X; |
| 3753 | |
| 3754 | a0 = VLOAD(M0)(0, lhs); |
| 3755 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 11 * RHS_STEP_X), (y_rhs)); |
| 3756 | |
| 3757 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3758 | |
| 3759 | lhs += LHS_STEP_X; |
| 3760 | |
| 3761 | a0 = VLOAD(M0)(0, lhs); |
| 3762 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 12 * RHS_STEP_X), (y_rhs)); |
| 3763 | |
| 3764 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3765 | |
| 3766 | lhs += LHS_STEP_X; |
| 3767 | |
| 3768 | a0 = VLOAD(M0)(0, lhs); |
| 3769 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 13 * RHS_STEP_X), (y_rhs)); |
| 3770 | |
| 3771 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3772 | |
| 3773 | lhs += LHS_STEP_X; |
| 3774 | |
| 3775 | a0 = VLOAD(M0)(0, lhs); |
| 3776 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 14 * RHS_STEP_X), (y_rhs)); |
| 3777 | |
| 3778 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3779 | |
| 3780 | lhs += LHS_STEP_X; |
| 3781 | |
| 3782 | a0 = VLOAD(M0)(0, lhs); |
| 3783 | b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 15 * RHS_STEP_X), (y_rhs)); |
| 3784 | |
| 3785 | ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c); |
| 3786 | |
| 3787 | lhs += LHS_STEP_X; |
| 3788 | #endif // K0 > 8 |
| 3789 | |
| 3790 | #ifndef LHS_INTERLEAVE |
| 3791 | lhs += (M0 * K0 * (V0 - 1)); |
| 3792 | #endif // LHS_INTERLEAVE |
| 3793 | |
| 3794 | x_rhs += K0 * RHS_STEP_X; |
| 3795 | #ifndef RHS_INTERLEAVE |
| 3796 | x_rhs += (PIXEL_UNIT * K0 * (H0 - 1)); |
| 3797 | #endif // RHS_INTERLEAVE |
| 3798 | } |
| 3799 | |
| 3800 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * (uint)M0 * dst_stride_y); |
| 3801 | |
| 3802 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); |
| 3803 | |
| 3804 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 3805 | |
| 3806 | // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
| 3807 | CALCULATE_Z_OFFSET(M0, uint, zout, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); |
| 3808 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 3809 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 3810 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
| 3811 | |
| 3812 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 3813 | |
| 3814 | // Add offset for batched GEMM |
| 3815 | dst_addr += z * dst_stride_z; |
| 3816 | |
| 3817 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 3818 | |
| 3819 | // Multiply by the weight of matrix-matrix product and store the result |
| 3820 | #if defined(ALPHA) |
| 3821 | SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); |
| 3822 | #endif // defined(ALPHA) |
| 3823 | |
| 3824 | // Add beta*bias |
| 3825 | #if defined(BETA) |
| 3826 | #if defined(BROADCAST_BIAS) |
| 3827 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)); |
| 3828 | |
| 3829 | LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 3830 | |
| 3831 | #ifndef UNIT_BETA |
| 3832 | SCALE_BLOCK(1, DATA_TYPE, bias, BETA); |
| 3833 | #endif // UNIT_BIAS |
| 3834 | |
| 3835 | // c = c + bias[broadcasted] |
| 3836 | #if defined(MIXED_PRECISION) |
| 3837 | CONVERT_BLOCK(1, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp); |
| 3838 | ADD_BLOCK_BROADCAST(M0, c, bias_hp0); |
| 3839 | #else // defined(MIXED_PRECISION) |
| 3840 | ADD_BLOCK_BROADCAST(M0, c, bias0); |
| 3841 | #endif // defined(MIXED_PRECISION) |
| 3842 | |
| 3843 | #else // defined(BROADCAST_BIAS) |
| 3844 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * (uint)M0 * bias_stride_y) + z * bias_stride_z; |
| 3845 | |
| 3846 | LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 3847 | |
| 3848 | #ifndef UNIT_BETA |
| 3849 | SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); |
| 3850 | #endif // UNIT_BIAS |
| 3851 | |
| 3852 | #if defined(MIXED_PRECISION) |
| 3853 | CONVERT_BLOCK(M0, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp); |
| 3854 | ADD_BLOCK(M0, c, bias_hp); |
| 3855 | #else // defined(MIXED_PRECISION) |
| 3856 | ADD_BLOCK(M0, c, bias); |
| 3857 | #endif // defined(MIXED_PRECISION) |
| 3858 | |
| 3859 | #endif // defined(BROADCAST_BIAS) |
| 3860 | #endif // defined(BETA) |
| 3861 | |
| 3862 | #if defined(ACTIVATION_TYPE) |
| 3863 | #if defined(MIXED_PRECISION) |
| 3864 | ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE_ACCUMULATOR, c, A_VAL, B_VAL); |
| 3865 | #else // defined(MIXED_PRECISION) |
| 3866 | ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, c, A_VAL, B_VAL); |
| 3867 | #endif // defined(MIXED_PRECISION) |
| 3868 | #endif // defined(ACTIVATION_TYPE) |
| 3869 | |
| 3870 | // Store output block |
| 3871 | #if defined(MIXED_PRECISION) |
| 3872 | CONVERT_STORE_BLOCK(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout); |
| 3873 | #else // defined(MIXED_PRECISION) |
| 3874 | STORE_BLOCK(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout); |
| 3875 | #endif // defined(MIXED_PRECISION) |
| 3876 | |
| 3877 | #undef LHS_BLOCK_SIZE |
| 3878 | #undef LHS_OFFSET_X |
| 3879 | #undef LHS_STEP_X |
| 3880 | #undef RHS_BLOCK_SIZE |
| 3881 | #undef RHS_OFFSET_X |
| 3882 | #undef RHS_STEP_X |
| 3883 | #undef PIXEL_UNIT |
| 3884 | #undef LHS_STEP_LOOP |
| 3885 | #undef RHS_STEP_LOOP |
| 3886 | } |
| 3887 | #endif // defined(OPENCL_IMAGE_SUPPORT) |
| 3888 | |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 3889 | #endif // defined(LHS_TRANSPOSE) |
| 3890 | |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 3891 | #endif // defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(K) && defined(DATA_TYPE) |
| 3892 | |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 3893 | #if defined(M0) && defined(N0) && defined(K0) && defined(K) && defined(DATA_TYPE) |
| 3894 | |
| 3895 | #define VFMA(a, b, c) \ |
| 3896 | ({ \ |
| 3897 | c = fma(a, b, c); \ |
| 3898 | }) |
| 3899 | |
| 3900 | #if M0 == 1 |
| 3901 | #define RHS_VFMA_M0xN0(i, a, b, c) \ |
| 3902 | ({ \ |
| 3903 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 3904 | }) |
| 3905 | #elif M0 == 2 // M0 == 2 |
| 3906 | #define RHS_VFMA_M0xN0(i, a, b, c) \ |
| 3907 | ({ \ |
| 3908 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 3909 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
| 3910 | }) |
| 3911 | #elif M0 == 3 // M0 == 3 |
| 3912 | #define RHS_VFMA_M0xN0(i, a, b, c) \ |
| 3913 | ({ \ |
| 3914 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 3915 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
| 3916 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ |
| 3917 | }) |
| 3918 | #elif M0 == 4 // M0 == 4 |
| 3919 | #define RHS_VFMA_M0xN0(i, a, b, c) \ |
| 3920 | ({ \ |
| 3921 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 3922 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
| 3923 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ |
| 3924 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ |
| 3925 | }) |
| 3926 | #elif M0 == 5 // M0 == 5 |
| 3927 | #define RHS_VFMA_M0xN0(i, a, b, c) \ |
| 3928 | ({ \ |
| 3929 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 3930 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
| 3931 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ |
| 3932 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ |
| 3933 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ |
| 3934 | }) |
| 3935 | #elif M0 == 6 // M0 == 6 |
| 3936 | #define RHS_VFMA_M0xN0(i, a, b, c) \ |
| 3937 | ({ \ |
| 3938 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 3939 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
| 3940 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ |
| 3941 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ |
| 3942 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ |
| 3943 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ |
| 3944 | }) |
| 3945 | #elif M0 == 7 // M0 == 7 |
| 3946 | #define RHS_VFMA_M0xN0(i, a, b, c) \ |
| 3947 | ({ \ |
| 3948 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 3949 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
| 3950 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ |
| 3951 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ |
| 3952 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ |
| 3953 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ |
| 3954 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \ |
| 3955 | }) |
| 3956 | #elif M0 == 8 // M0 == 8 |
| 3957 | #define RHS_VFMA_M0xN0(i, a, b, c) \ |
| 3958 | ({ \ |
| 3959 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ |
| 3960 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ |
| 3961 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ |
| 3962 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ |
| 3963 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ |
| 3964 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ |
| 3965 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \ |
| 3966 | VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##7).s##i), b, (c##7)); \ |
| 3967 | }) |
| 3968 | #else // M0 not supported |
| 3969 | #error "M0 not supported" |
| 3970 | #endif // M0 not supported |
| 3971 | |
| 3972 | /** This OpenCL kernel computes the matrix multiplication between 2 matrices. |
| 3973 | * The LHS matrix is NOT reshaped |
| 3974 | * The RHS matrix is NOT reshaped |
| 3975 | * |
| 3976 | * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 3977 | * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90) |
| 3978 | * @note The number of columns of LHS matrix must be passed at compile time using -DK (e.g. -DK=64) |
| 3979 | * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2) |
| 3980 | * @note The number of K0 partial accumulations must be passed at compile time using -DK0 (e.g., -DK0=2) |
| 3981 | * @note The number of N0 columns to process must be passed at compile time using -DN0 (e.g. -DN0=2) |
SiCong Li | 3a50166 | 2020-06-26 10:02:06 +0100 | [diff] [blame] | 3982 | * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) |
| 3983 | * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 3984 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 3985 | * - M0 = 1, 2, 3, 4, 5, 6, 7, 8 |
| 3986 | * - N0 = 2, 3, 4, 8, 16 |
| 3987 | * - K0 = 2, 3, 4, 8, 16 |
| 3988 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 3989 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 3990 | * The activation function is performed after the bias addition |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 3991 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 3992 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
| 3993 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 3994 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 3995 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 3996 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix |
| 3997 | * |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 3998 | * @param[in] lhs_ptr Pointer to the LHS matrix. Supported data type: F16/F32 |
| 3999 | * @param[in] lhs_stride_x Stride of the LHS matrix in X dimension (in bytes) |
| 4000 | * @param[in] lhs_step_x lhs_stride_x * number of elements along X processed per workitem(in bytes) |
| 4001 | * @param[in] lhs_stride_y Stride of the LHS matrix in Y dimension (in bytes) |
| 4002 | * @param[in] lhs_step_y lhs_stride_y * number of elements along Y processed per workitem(in bytes) |
| 4003 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS matrix |
| 4004 | * @param[in] rhs_ptr Pointer to the RHS matrix. Supported data type: same as @p lhs_ptr |
| 4005 | * @param[in] rhs_stride_x Stride of the RHS matrix in X dimension (in bytes) |
| 4006 | * @param[in] rhs_step_x rhs_stride_x * number of elements along X processed per workitem(in bytes) |
| 4007 | * @param[in] rhs_stride_y Stride of the RHS matrix in Y dimension (in bytes) |
| 4008 | * @param[in] rhs_step_y rhs_stride_y * number of elements along Y processed per workitem(in bytes) |
| 4009 | * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS matrix |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 4010 | * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 4011 | * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 4012 | * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes) |
| 4013 | * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 4014 | * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes) |
| 4015 | * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
| 4016 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr |
| 4017 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 4018 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 4019 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 4020 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 4021 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 4022 | * @param[in] lhs_stride_z Stride of the LHS matrix in Z dimension (in bytes) |
| 4023 | * @param[in] rhs_stride_z Stride of the RHS matrix in Z dimension (in bytes) |
| 4024 | * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
| 4025 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 4026 | * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) |
| 4027 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 4028 | */ |
| 4029 | __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs), |
| 4030 | IMAGE_DECLARATION(rhs), |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 4031 | #if defined(BETA) |
| 4032 | IMAGE_DECLARATION(bias), |
| 4033 | #endif // defined(BETA) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 4034 | IMAGE_DECLARATION(dst), |
| 4035 | uint lhs_stride_z, |
| 4036 | uint rhs_stride_z, |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 4037 | #if defined(BETA) |
| 4038 | uint bias_stride_z, |
| 4039 | #endif //defined(BETA) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 4040 | uint dst_stride_z |
| 4041 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 4042 | , |
| 4043 | uint lhs_cross_plane_pad |
| 4044 | #endif // REINTERPRET_INPUT_AS_3D |
| 4045 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 4046 | , |
| 4047 | uint dst_cross_plane_pad |
| 4048 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 4049 | ) |
| 4050 | { |
| 4051 | // Block size |
| 4052 | #define RHS_BLOCK_SIZE ((K0) * (N0)) |
| 4053 | |
| 4054 | // RHS offset and step X |
| 4055 | #define RHS_OFFSET_X (RHS_BLOCK_SIZE) |
| 4056 | |
| 4057 | uint x = get_global_id(0); |
| 4058 | uint y = get_global_id(1); |
| 4059 | uint z = get_global_id(2); |
| 4060 | |
| 4061 | #if defined(DUMMY_WORK_ITEMS) |
| 4062 | if((x * N0 >= N) || (y * M0 >= M)) |
| 4063 | { |
| 4064 | return; |
| 4065 | } |
| 4066 | #endif // defined(DUMMY_WORK_ITEMS) |
| 4067 | |
| 4068 | // Compute LHS matrix address |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 4069 | uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 4070 | |
| 4071 | // Compute RHS matrix address |
| 4072 | uint rhs_offset = rhs_offset_first_element_in_bytes + x * N0 * sizeof(DATA_TYPE); |
| 4073 | |
| 4074 | #if defined(MATRIX_B_DEPTH) |
| 4075 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 4076 | rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z; |
| 4077 | #else // defined(MATRIX_B_DEPTH) |
| 4078 | rhs_offset += z * rhs_stride_z; |
| 4079 | #endif // defined(MATRIX_B_DEPTH) |
| 4080 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 4081 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0); |
| 4082 | REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 4083 | |
| 4084 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 4085 | // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
| 4086 | CALCULATE_Z_OFFSET(M0, uint, zlhs, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); |
| 4087 | |
| 4088 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 4089 | // multiply lhs_stride_z by DEPTH_GEMM3D |
| 4090 | lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D; |
| 4091 | |
| 4092 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 4093 | |
| 4094 | // Add offset for batched GEMM |
| 4095 | lhs_offset += z * lhs_stride_z; |
| 4096 | |
| 4097 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 4098 | |
| 4099 | // Initialize the accumulators |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 4100 | REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(M0-1)=0; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 4101 | |
| 4102 | int i = 0; |
| 4103 | for(; i <= (K - K0); i += K0) |
| 4104 | { |
| 4105 | // Supported cases (M0, K0): |
| 4106 | // 1,2 - 1,3 - 1,4 - 1,8 - 1,16 |
| 4107 | // 2,2 - 2,3 - 2,4 - 2,8 - 2,16 |
| 4108 | // 3,2 - 3,3 - 3,4 - 3,8 - 3,16 |
| 4109 | // 4,2 - 4,3 - 4,4 - 4,8 - 4,16 |
| 4110 | // 5,2 - 5,3 - 5,4 - 5,8 - 5,16 |
| 4111 | // 6,2 - 6,3 - 6,4 - 6,8 - 6,16 |
| 4112 | // 7,2 - 7,3 - 7,4 - 7,8 - 7,16 |
| 4113 | // 8,2 - 8,3 - 8,4 - 8,8 - 8,16 |
| 4114 | // Load values from LHS matrix |
| 4115 | LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs); |
| 4116 | |
| 4117 | // Load values from RHS matrix |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 4118 | LOAD_BLOCK(K0, N0, DATA_TYPE, b, rhs_ptr, rhs_offset, rhs_stride_y, zero); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 4119 | |
| 4120 | RHS_VFMA_M0xN0(0, a, b0, c); |
| 4121 | RHS_VFMA_M0xN0(1, a, b1, c); |
| 4122 | #if K0 > 2 |
| 4123 | RHS_VFMA_M0xN0(2, a, b2, c); |
| 4124 | #endif // K0 > 2 |
| 4125 | #if K0 > 3 |
| 4126 | RHS_VFMA_M0xN0(3, a, b3, c); |
| 4127 | #endif // K0 > 3 |
| 4128 | #if K0 > 4 |
| 4129 | RHS_VFMA_M0xN0(4, a, b4, c); |
| 4130 | RHS_VFMA_M0xN0(5, a, b5, c); |
| 4131 | RHS_VFMA_M0xN0(6, a, b6, c); |
| 4132 | RHS_VFMA_M0xN0(7, a, b7, c); |
| 4133 | #endif // K0 > 4 |
| 4134 | #if K0 > 8 |
| 4135 | RHS_VFMA_M0xN0(8, a, b8, c); |
| 4136 | RHS_VFMA_M0xN0(9, a, b9, c); |
Gian Marco Iodice | 7b9d7ca | 2019-09-19 16:37:39 +0100 | [diff] [blame] | 4137 | RHS_VFMA_M0xN0(A, a, bA, c); |
| 4138 | RHS_VFMA_M0xN0(B, a, bB, c); |
| 4139 | RHS_VFMA_M0xN0(C, a, bC, c); |
| 4140 | RHS_VFMA_M0xN0(D, a, bD, c); |
| 4141 | RHS_VFMA_M0xN0(E, a, bE, c); |
| 4142 | RHS_VFMA_M0xN0(F, a, bF, c); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 4143 | #endif // K0 > 8 |
| 4144 | |
| 4145 | lhs_offset += K0 * sizeof(DATA_TYPE); |
| 4146 | rhs_offset += K0 * rhs_stride_y; |
| 4147 | } |
| 4148 | |
| 4149 | // Left-over accumulations |
| 4150 | for(; i < K; ++i) |
| 4151 | { |
| 4152 | // Load values from LHS matrix |
| 4153 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 4154 | a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0)); |
| 4155 | #if M0 > 1 |
| 4156 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 4157 | a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1)); |
| 4158 | #endif // M0 > 1 |
| 4159 | #if M0 > 2 |
| 4160 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 4161 | a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2)); |
| 4162 | #endif // M0 > 2 |
| 4163 | #if M0 > 3 |
| 4164 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 4165 | a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3)); |
| 4166 | #endif // M0 > 3 |
| 4167 | #if M0 > 4 |
| 4168 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 4169 | a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4)); |
| 4170 | #endif // M0 > 4 |
| 4171 | #if M0 > 5 |
| 4172 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 4173 | a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5)); |
| 4174 | #endif // M0 > 5 |
| 4175 | #if M0 > 6 |
| 4176 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 4177 | a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6)); |
| 4178 | #endif // M0 > 6 |
| 4179 | #if M0 > 7 |
| 4180 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 4181 | a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7)); |
| 4182 | #endif // M0 > 7 |
| 4183 | |
| 4184 | VEC_DATA_TYPE(DATA_TYPE, N0) |
| 4185 | b = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * rhs_stride_y)); |
| 4186 | RHS_VFMA_M0xN0(0, a, b, c); |
| 4187 | |
| 4188 | lhs_offset += sizeof(DATA_TYPE); |
| 4189 | rhs_offset += rhs_stride_y; |
| 4190 | } |
| 4191 | |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 4192 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 4193 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 4194 | REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 4195 | |
| 4196 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 4197 | // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
| 4198 | CALCULATE_Z_OFFSET(M0, uint, zout, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); |
| 4199 | |
| 4200 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 4201 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 4202 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
| 4203 | |
| 4204 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 4205 | |
| 4206 | // Add offset for batched GEMM |
| 4207 | dst_addr += z * dst_stride_z; |
| 4208 | |
| 4209 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 4210 | |
| 4211 | // Multiply by the weight of matrix-matrix product and store the result |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 4212 | #if defined(ALPHA) |
| 4213 | SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); |
| 4214 | #endif // defined(ALPHA) |
| 4215 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 4216 | // Add beta*bias |
| 4217 | #if defined(BETA) |
| 4218 | #if defined(BROADCAST_BIAS) |
| 4219 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); |
| 4220 | |
| 4221 | LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 4222 | |
| 4223 | #ifndef UNIT_BETA |
| 4224 | SCALE_BLOCK(1, DATA_TYPE, bias, BETA); |
| 4225 | #endif // UNIT_BIAS |
| 4226 | |
| 4227 | // c = c + bias[broadcasted] |
| 4228 | ADD_BLOCK_BROADCAST(M0, c, bias0); |
| 4229 | |
| 4230 | #else // defined(BROADCAST_BIAS) |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 4231 | __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z; |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 4232 | |
| 4233 | LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero); |
| 4234 | |
| 4235 | #ifndef UNIT_BETA |
| 4236 | SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); |
| 4237 | #endif // UNIT_BIAS |
| 4238 | |
| 4239 | // c = c + bias |
| 4240 | ADD_BLOCK(M0, c, bias); |
| 4241 | |
| 4242 | #endif // defined(BROADCAST_BIAS) |
| 4243 | #endif // defined(BETA) |
| 4244 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 4245 | #if defined(ACTIVATION_TYPE) |
| 4246 | ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, c, A_VAL, B_VAL); |
| 4247 | #endif // defined(ACTIVATION_TYPE) |
| 4248 | |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 4249 | // Store output block |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 4250 | STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, N, y, x); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 4251 | |
| 4252 | #undef RHS_BLOCK_SIZE |
| 4253 | #undef RHS_OFFSET_X |
| 4254 | #undef RHS_STEP_X |
| 4255 | } |
| 4256 | #endif // defined(M0) && defined(N0) && defined(K0) && defined(K) && defined(DATA_TYPE) |
| 4257 | |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4258 | #if defined(COLS_B) && defined(MULT_TRANSPOSE1XW_WIDTH) && defined(MULT_INTERLEAVE4X4_HEIGHT) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4259 | /** This OpenCL kernel is optimised for Midgard. It computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1) |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 4260 | * |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 4261 | * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4262 | * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (e.g. -DMULT_TRANSPOSE1XW_WIDTH=2) |
| 4263 | * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (e.g. -DMULT_INTERLEAVE4X4_HEIGHT=2) |
| 4264 | * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) |
| 4265 | * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4266 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4267 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 4268 | * The activation function is performed after the bias addition |
| 4269 | * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4270 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 4271 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 4272 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 4273 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 4274 | * |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4275 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 |
| 4276 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 4277 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 4278 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 4279 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 4280 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 4281 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4282 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 4283 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 4284 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 4285 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 4286 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4287 | * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 4288 | * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 4289 | * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) |
| 4290 | * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 4291 | * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) |
| 4292 | * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 4293 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4294 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4295 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4296 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4297 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4298 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4299 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 4300 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4301 | * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4302 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 4303 | * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4304 | */ |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 4305 | __kernel void gemm_mm_interleaved_transposed_f32(IMAGE_DECLARATION(src0), |
| 4306 | IMAGE_DECLARATION(src1), |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4307 | #if defined(BETA) |
| 4308 | IMAGE_DECLARATION(src2), |
| 4309 | #endif // defined(BETA) |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 4310 | IMAGE_DECLARATION(dst), |
| 4311 | uint src0_stride_z, |
| 4312 | uint src1_stride_z, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4313 | #if defined(BETA) |
| 4314 | uint src2_stride_z, |
| 4315 | #endif //defined(BETA) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4316 | uint dst_stride_z |
| 4317 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 4318 | , |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 4319 | uint cross_plane_pad |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4320 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 4321 | ) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4322 | { |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4323 | int x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH; |
| 4324 | int y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT; |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 4325 | int z = get_global_id(2); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4326 | |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4327 | // Offset |
| 4328 | const int offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4; |
| 4329 | const int offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 4; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4330 | |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4331 | // src_addr_a = address of matrix A |
| 4332 | // src_addr_b = address of matrix B |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 4333 | int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; |
| 4334 | int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; |
| 4335 | |
| 4336 | #if defined(MATRIX_B_DEPTH) |
| 4337 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 4338 | src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; |
| 4339 | #else // defined(MATRIX_B_DEPTH) |
| 4340 | src1_addr_in_bytes += z * src1_stride_z; |
| 4341 | #endif // defined(MATRIX_B_DEPTH) |
| 4342 | |
| 4343 | __global float *src_addr_a = (__global float *)(src0_ptr + src0_addr_in_bytes); |
| 4344 | __global float *src_addr_b = (__global float *)(src1_ptr + src1_addr_in_bytes); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4345 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 4346 | // Compute end row address for matrix B |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4347 | __global float *src_end_addr_b = src_addr_b + COLS_B; |
| 4348 | |
| 4349 | src_addr_a += offset_row_a; |
| 4350 | src_addr_b += offset_row_b; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4351 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 4352 | // Reset accumulators |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4353 | float4 c0 = 0.0f; |
| 4354 | float4 c1 = 0.0f; |
| 4355 | float4 c2 = 0.0f; |
| 4356 | float4 c3 = 0.0f; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4357 | |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4358 | for(; src_addr_b <= (src_end_addr_b - (int)(8 * MULT_TRANSPOSE1XW_WIDTH)); src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4359 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 4360 | // Load values from matrix A (interleaved) and matrix B (transposed) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4361 | float4 a0 = vload4(0, src_addr_a); |
| 4362 | float4 b0 = vload4(0, src_addr_b); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4363 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4364 | c0 += (float4)a0.s0 * b0; |
| 4365 | c1 += (float4)a0.s1 * b0; |
| 4366 | c2 += (float4)a0.s2 * b0; |
| 4367 | c3 += (float4)a0.s3 * b0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4368 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 4369 | // Load values from matrix A (interleaved) and matrix B (transposed) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4370 | a0 = vload4(0, src_addr_a + 4 * MULT_INTERLEAVE4X4_HEIGHT); |
| 4371 | b0 = vload4(0, src_addr_b + 4 * MULT_TRANSPOSE1XW_WIDTH); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4372 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4373 | c0 += (float4)a0.s0 * b0; |
| 4374 | c1 += (float4)a0.s1 * b0; |
| 4375 | c2 += (float4)a0.s2 * b0; |
| 4376 | c3 += (float4)a0.s3 * b0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4377 | } |
| 4378 | |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4379 | for(; src_addr_b < src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4380 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 4381 | // Load values from matrix A (interleaved) and matrix B (transposed) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4382 | float4 a0 = vload4(0, src_addr_a); |
| 4383 | float4 b0 = vload4(0, src_addr_b); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4384 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4385 | c0 += (float4)a0.s0 * b0; |
| 4386 | c1 += (float4)a0.s1 * b0; |
| 4387 | c2 += (float4)a0.s2 * b0; |
| 4388 | c3 += (float4)a0.s3 * b0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4389 | } |
| 4390 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 4391 | // Compute destination address |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4392 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 4393 | |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 4394 | // Compute dst address |
| 4395 | __global uchar *dst_addr = offset(&dst, 0, 0); |
| 4396 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4397 | uint4 zout = 0; |
| 4398 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4399 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 4400 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 4401 | // in order to take into account the presence of possible cross plane paddings |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4402 | // |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 4403 | // | | |
| 4404 | // | plane0 | |
| 4405 | // | | |
| 4406 | // |__________________| |
| 4407 | // |******************| |
| 4408 | // | cross_plane_pad | |
| 4409 | // |******************| |
| 4410 | // | | |
| 4411 | // | plane1 | |
| 4412 | // | | |
| 4413 | // |__________________| |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4414 | |
| 4415 | // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4416 | zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D; |
| 4417 | zout = min(DEPTH_GEMM3D - 1, zout); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4418 | |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 4419 | // Add offset due to the cross plane paddings |
| 4420 | zout *= (cross_plane_pad * dst_stride_y); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4421 | |
| 4422 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 4423 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 4424 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4425 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 4426 | // Add offset for batched GEMM |
| 4427 | dst_addr += z * dst_stride_z; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4428 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 4429 | |
| 4430 | // Multiply by the weight of matrix-matrix product and store the result |
| 4431 | #if defined(ALPHA) |
| 4432 | SCALE_BLOCK(4, float, c, ALPHA); |
| 4433 | #endif // defined(ALPHA) |
| 4434 | |
| 4435 | // Add beta*bias |
| 4436 | #if defined(BETA) |
| 4437 | REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0); |
| 4438 | |
| 4439 | #if defined(BROADCAST_BIAS) |
| 4440 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)); |
| 4441 | |
| 4442 | LOAD_BLOCK(1, 4, float, bias, src2_addr, 0, src2_stride_y, zero); |
| 4443 | |
| 4444 | #ifndef UNIT_BETA |
| 4445 | SCALE_BLOCK(1, float, bias, BETA); |
| 4446 | #endif // UNIT_BIAS |
| 4447 | |
| 4448 | // c = c + bias[broadcasted] |
| 4449 | ADD_BLOCK_BROADCAST(4, c, bias0); |
| 4450 | |
| 4451 | #else // defined(BROADCAST_BIAS) |
| 4452 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id( |
| 4453 | 2) * src2_stride_z; |
| 4454 | |
| 4455 | LOAD_BLOCK(4, 4, float, bias, src2_addr, 0, src2_stride_y, zero); |
| 4456 | |
| 4457 | #ifndef UNIT_BETA |
| 4458 | SCALE_BLOCK(4, float, bias, BETA); |
| 4459 | #endif // UNIT_BIAS |
| 4460 | |
| 4461 | // c = c + bias |
| 4462 | ADD_BLOCK(4, c, bias); |
| 4463 | |
| 4464 | #endif // defined(BROADCAST_BIAS) |
| 4465 | #endif // defined(BETA) |
| 4466 | |
| 4467 | #if defined(ACTIVATION_TYPE) |
| 4468 | ACTIVATION_BLOCK(4, ACTIVATION_TYPE, float, c, A_VAL, B_VAL); |
| 4469 | #endif // defined(ACTIVATION_TYPE) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 4470 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 4471 | // Store 4x4 block |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4472 | vstore4(c0, 0, (__global float *)(dst_addr + 0 * dst_stride_y + zout.s0)); |
| 4473 | vstore4(c1, 0, (__global float *)(dst_addr + 1 * dst_stride_y + zout.s1)); |
| 4474 | vstore4(c2, 0, (__global float *)(dst_addr + 2 * dst_stride_y + zout.s2)); |
| 4475 | vstore4(c3, 0, (__global float *)(dst_addr + 3 * dst_stride_y + zout.s3)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4476 | } |
| 4477 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4478 | /** This OpenCL kernel is optimized for Bifrost and tt computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4479 | * |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 4480 | * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4481 | * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (e.g. -DMULT_TRANSPOSE1XW_WIDTH=2) |
| 4482 | * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (e.g. -DMULT_INTERLEAVE4X4_HEIGHT=2) |
| 4483 | * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (e.g. -DMULT_INTERLEAVE4X4_HEIGHT=2) |
| 4484 | * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) |
| 4485 | * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4486 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4487 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 4488 | * The activation function is performed after the bias addition |
| 4489 | * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4490 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 4491 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 4492 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 4493 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 4494 | * |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4495 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 |
| 4496 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 4497 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 4498 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 4499 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 4500 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 4501 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4502 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 4503 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 4504 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 4505 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 4506 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4507 | * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 4508 | * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 4509 | * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) |
| 4510 | * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 4511 | * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) |
| 4512 | * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 4513 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4514 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4515 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4516 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4517 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4518 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4519 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 4520 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4521 | * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4522 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 4523 | * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4524 | */ |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 4525 | __kernel void gemm_mm_interleaved_transposed_f32_bifrost(IMAGE_DECLARATION(src0), |
| 4526 | IMAGE_DECLARATION(src1), |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4527 | #if defined(BETA) |
| 4528 | IMAGE_DECLARATION(src2), |
| 4529 | #endif // defined(BETA) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 4530 | IMAGE_DECLARATION(dst), |
| 4531 | uint src0_stride_z, |
| 4532 | uint src1_stride_z, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4533 | #if defined(BETA) |
| 4534 | uint src2_stride_z, |
| 4535 | #endif //defined(BETA) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4536 | uint dst_stride_z |
| 4537 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 4538 | , |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 4539 | uint cross_plane_pad |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4540 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 4541 | ) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4542 | { |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4543 | int x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH; |
| 4544 | int y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT; |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 4545 | int z = get_global_id(2); |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4546 | |
| 4547 | // Offset |
| 4548 | const int offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4; |
| 4549 | const int offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 4; |
| 4550 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4551 | // src_addr_a = address of matrix A |
| 4552 | // src_addr_b = address of matrix B |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 4553 | int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; |
| 4554 | int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; |
| 4555 | |
| 4556 | #if defined(MATRIX_B_DEPTH) |
| 4557 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 4558 | src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; |
| 4559 | #else // defined(MATRIX_B_DEPTH) |
| 4560 | src1_addr_in_bytes += z * src1_stride_z; |
| 4561 | #endif // defined(MATRIX_B_DEPTH) |
| 4562 | |
| 4563 | __global float *src_addr_a = (__global float *)(src0_ptr + src0_addr_in_bytes); |
| 4564 | __global float *src_addr_b = (__global float *)(src1_ptr + src1_addr_in_bytes); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4565 | |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4566 | src_addr_a += offset_row_a; |
| 4567 | src_addr_b += offset_row_b; |
| 4568 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4569 | // Reset accumulators |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4570 | float4 c0 = 0.0f; |
| 4571 | float4 c1 = 0.0f; |
| 4572 | float4 c2 = 0.0f; |
| 4573 | float4 c3 = 0.0f; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4574 | |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 4575 | #define COLS_MTX_B (COLS_B / (4 * MULT_TRANSPOSE1XW_WIDTH)) |
| 4576 | |
| 4577 | int i = 0; |
| 4578 | for(; i <= (int)(COLS_MTX_B - 4); i += 4) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4579 | { |
| 4580 | // Load values from matrix A (interleaved) and matrix B (transposed) |
| 4581 | float4 a0 = vload4(0, src_addr_a); |
| 4582 | float4 b0 = vload4(0, src_addr_b); |
| 4583 | |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 4584 | src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT; |
| 4585 | src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4586 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4587 | c0.s0 = fma(a0.s0, b0.s0, c0.s0); |
| 4588 | c0.s1 = fma(a0.s0, b0.s1, c0.s1); |
| 4589 | c0.s2 = fma(a0.s0, b0.s2, c0.s2); |
| 4590 | c0.s3 = fma(a0.s0, b0.s3, c0.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4591 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4592 | c1.s0 = fma(a0.s1, b0.s0, c1.s0); |
| 4593 | c1.s1 = fma(a0.s1, b0.s1, c1.s1); |
| 4594 | c1.s2 = fma(a0.s1, b0.s2, c1.s2); |
| 4595 | c1.s3 = fma(a0.s1, b0.s3, c1.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4596 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4597 | c2.s0 = fma(a0.s2, b0.s0, c2.s0); |
| 4598 | c2.s1 = fma(a0.s2, b0.s1, c2.s1); |
| 4599 | c2.s2 = fma(a0.s2, b0.s2, c2.s2); |
| 4600 | c2.s3 = fma(a0.s2, b0.s3, c2.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4601 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4602 | c3.s0 = fma(a0.s3, b0.s0, c3.s0); |
| 4603 | c3.s1 = fma(a0.s3, b0.s1, c3.s1); |
| 4604 | c3.s2 = fma(a0.s3, b0.s2, c3.s2); |
| 4605 | c3.s3 = fma(a0.s3, b0.s3, c3.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4606 | |
| 4607 | // Load values from matrix A (interleaved) and matrix B (transposed) |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 4608 | a0 = vload4(0, src_addr_a); |
| 4609 | b0 = vload4(0, src_addr_b); |
| 4610 | |
| 4611 | src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT; |
| 4612 | src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4613 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4614 | c0.s0 = fma(a0.s0, b0.s0, c0.s0); |
| 4615 | c0.s1 = fma(a0.s0, b0.s1, c0.s1); |
| 4616 | c0.s2 = fma(a0.s0, b0.s2, c0.s2); |
| 4617 | c0.s3 = fma(a0.s0, b0.s3, c0.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4618 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4619 | c1.s0 = fma(a0.s1, b0.s0, c1.s0); |
| 4620 | c1.s1 = fma(a0.s1, b0.s1, c1.s1); |
| 4621 | c1.s2 = fma(a0.s1, b0.s2, c1.s2); |
| 4622 | c1.s3 = fma(a0.s1, b0.s3, c1.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4623 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4624 | c2.s0 = fma(a0.s2, b0.s0, c2.s0); |
| 4625 | c2.s1 = fma(a0.s2, b0.s1, c2.s1); |
| 4626 | c2.s2 = fma(a0.s2, b0.s2, c2.s2); |
| 4627 | c2.s3 = fma(a0.s2, b0.s3, c2.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4628 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4629 | c3.s0 = fma(a0.s3, b0.s0, c3.s0); |
| 4630 | c3.s1 = fma(a0.s3, b0.s1, c3.s1); |
| 4631 | c3.s2 = fma(a0.s3, b0.s2, c3.s2); |
| 4632 | c3.s3 = fma(a0.s3, b0.s3, c3.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4633 | |
| 4634 | // Load values from matrix A (interleaved) and matrix B (transposed) |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 4635 | a0 = vload4(0, src_addr_a); |
| 4636 | b0 = vload4(0, src_addr_b); |
| 4637 | |
| 4638 | src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT; |
| 4639 | src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH; |
| 4640 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4641 | c0.s0 = fma(a0.s0, b0.s0, c0.s0); |
| 4642 | c0.s1 = fma(a0.s0, b0.s1, c0.s1); |
| 4643 | c0.s2 = fma(a0.s0, b0.s2, c0.s2); |
| 4644 | c0.s3 = fma(a0.s0, b0.s3, c0.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 4645 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4646 | c1.s0 = fma(a0.s1, b0.s0, c1.s0); |
| 4647 | c1.s1 = fma(a0.s1, b0.s1, c1.s1); |
| 4648 | c1.s2 = fma(a0.s1, b0.s2, c1.s2); |
| 4649 | c1.s3 = fma(a0.s1, b0.s3, c1.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 4650 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4651 | c2.s0 = fma(a0.s2, b0.s0, c2.s0); |
| 4652 | c2.s1 = fma(a0.s2, b0.s1, c2.s1); |
| 4653 | c2.s2 = fma(a0.s2, b0.s2, c2.s2); |
| 4654 | c2.s3 = fma(a0.s2, b0.s3, c2.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 4655 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4656 | c3.s0 = fma(a0.s3, b0.s0, c3.s0); |
| 4657 | c3.s1 = fma(a0.s3, b0.s1, c3.s1); |
| 4658 | c3.s2 = fma(a0.s3, b0.s2, c3.s2); |
| 4659 | c3.s3 = fma(a0.s3, b0.s3, c3.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 4660 | |
| 4661 | // Load values from matrix A (interleaved) and matrix B (transposed) |
| 4662 | a0 = vload4(0, src_addr_a); |
| 4663 | b0 = vload4(0, src_addr_b); |
| 4664 | |
| 4665 | src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT; |
| 4666 | src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4667 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4668 | c0.s0 = fma(a0.s0, b0.s0, c0.s0); |
| 4669 | c0.s1 = fma(a0.s0, b0.s1, c0.s1); |
| 4670 | c0.s2 = fma(a0.s0, b0.s2, c0.s2); |
| 4671 | c0.s3 = fma(a0.s0, b0.s3, c0.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4672 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4673 | c1.s0 = fma(a0.s1, b0.s0, c1.s0); |
| 4674 | c1.s1 = fma(a0.s1, b0.s1, c1.s1); |
| 4675 | c1.s2 = fma(a0.s1, b0.s2, c1.s2); |
| 4676 | c1.s3 = fma(a0.s1, b0.s3, c1.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4677 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4678 | c2.s0 = fma(a0.s2, b0.s0, c2.s0); |
| 4679 | c2.s1 = fma(a0.s2, b0.s1, c2.s1); |
| 4680 | c2.s2 = fma(a0.s2, b0.s2, c2.s2); |
| 4681 | c2.s3 = fma(a0.s2, b0.s3, c2.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4682 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4683 | c3.s0 = fma(a0.s3, b0.s0, c3.s0); |
| 4684 | c3.s1 = fma(a0.s3, b0.s1, c3.s1); |
| 4685 | c3.s2 = fma(a0.s3, b0.s2, c3.s2); |
| 4686 | c3.s3 = fma(a0.s3, b0.s3, c3.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4687 | } |
| 4688 | |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 4689 | for(; i < (int)(COLS_MTX_B); ++i) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4690 | { |
| 4691 | // Load values from matrix A (interleaved) and matrix B (transposed) |
| 4692 | float4 a0 = vload4(0, src_addr_a); |
| 4693 | float4 b0 = vload4(0, src_addr_b); |
| 4694 | |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 4695 | src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT; |
| 4696 | src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH; |
| 4697 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4698 | c0.s0 = fma(a0.s0, b0.s0, c0.s0); |
| 4699 | c0.s1 = fma(a0.s0, b0.s1, c0.s1); |
| 4700 | c0.s2 = fma(a0.s0, b0.s2, c0.s2); |
| 4701 | c0.s3 = fma(a0.s0, b0.s3, c0.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4702 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4703 | c1.s0 = fma(a0.s1, b0.s0, c1.s0); |
| 4704 | c1.s1 = fma(a0.s1, b0.s1, c1.s1); |
| 4705 | c1.s2 = fma(a0.s1, b0.s2, c1.s2); |
| 4706 | c1.s3 = fma(a0.s1, b0.s3, c1.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4707 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4708 | c2.s0 = fma(a0.s2, b0.s0, c2.s0); |
| 4709 | c2.s1 = fma(a0.s2, b0.s1, c2.s1); |
| 4710 | c2.s2 = fma(a0.s2, b0.s2, c2.s2); |
| 4711 | c2.s3 = fma(a0.s2, b0.s3, c2.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4712 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4713 | c3.s0 = fma(a0.s3, b0.s0, c3.s0); |
| 4714 | c3.s1 = fma(a0.s3, b0.s1, c3.s1); |
| 4715 | c3.s2 = fma(a0.s3, b0.s2, c3.s2); |
| 4716 | c3.s3 = fma(a0.s3, b0.s3, c3.s3); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4717 | } |
| 4718 | |
| 4719 | // Compute destination address |
| 4720 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 4721 | |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 4722 | // Compute dst address |
| 4723 | __global uchar *dst_addr = offset(&dst, 0, 0); |
| 4724 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4725 | uint4 zout = 0; |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 4726 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4727 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 4728 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 4729 | // in order to take into account the presence of possible cross plane paddings |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4730 | // |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 4731 | // | | |
| 4732 | // | plane0 | |
| 4733 | // | | |
| 4734 | // |__________________| |
| 4735 | // |******************| |
| 4736 | // | cross_plane_pad | |
| 4737 | // |******************| |
| 4738 | // | | |
| 4739 | // | plane1 | |
| 4740 | // | | |
| 4741 | // |__________________| |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4742 | |
| 4743 | // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4744 | zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D; |
| 4745 | zout = min(DEPTH_GEMM3D - 1, zout); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4746 | |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 4747 | // Add offset due to the cross plane paddings |
| 4748 | zout *= (cross_plane_pad * dst_stride_y); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4749 | |
| 4750 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 4751 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 4752 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4753 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 4754 | // Add offset for batched GEMM |
| 4755 | dst_addr += z * dst_stride_z; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4756 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 4757 | |
| 4758 | // Multiply by the weight of matrix-matrix product and store the result |
| 4759 | #if defined(ALPHA) |
| 4760 | SCALE_BLOCK(4, float, c, ALPHA); |
| 4761 | #endif // defined(ALPHA) |
| 4762 | |
| 4763 | // Add beta*bias |
| 4764 | #if defined(BETA) |
| 4765 | REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0); |
| 4766 | |
| 4767 | #if defined(BROADCAST_BIAS) |
| 4768 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)); |
| 4769 | |
| 4770 | LOAD_BLOCK(1, 4, float, bias, src2_addr, 0, src2_stride_y, zero); |
| 4771 | |
| 4772 | #ifndef UNIT_BETA |
| 4773 | SCALE_BLOCK(1, float, bias, BETA); |
| 4774 | #endif // UNIT_BIAS |
| 4775 | |
| 4776 | // c = c + bias[broadcasted] |
| 4777 | ADD_BLOCK_BROADCAST(4, c, bias0); |
| 4778 | |
| 4779 | #else // defined(BROADCAST_BIAS) |
| 4780 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id( |
| 4781 | 2) * src2_stride_z; |
| 4782 | |
| 4783 | LOAD_BLOCK(4, 4, float, bias, src2_addr, 0, src2_stride_y, zero); |
| 4784 | |
| 4785 | #ifndef UNIT_BETA |
| 4786 | SCALE_BLOCK(4, float, bias, BETA); |
| 4787 | #endif // UNIT_BIAS |
| 4788 | |
| 4789 | // c = c + bias |
| 4790 | ADD_BLOCK(4, c, bias); |
| 4791 | |
| 4792 | #endif // defined(BROADCAST_BIAS) |
| 4793 | #endif // defined(BETA) |
| 4794 | |
| 4795 | #if defined(ACTIVATION_TYPE) |
| 4796 | ACTIVATION_BLOCK(4, ACTIVATION_TYPE, float, c, A_VAL, B_VAL); |
| 4797 | #endif // defined(ACTIVATION_TYPE) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 4798 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4799 | // Store 4x4 block |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4800 | vstore4(c0, 0, (__global float *)(dst_addr + 0 * dst_stride_y + zout.s0)); |
| 4801 | vstore4(c1, 0, (__global float *)(dst_addr + 1 * dst_stride_y + zout.s1)); |
| 4802 | vstore4(c2, 0, (__global float *)(dst_addr + 2 * dst_stride_y + zout.s2)); |
| 4803 | vstore4(c3, 0, (__global float *)(dst_addr + 3 * dst_stride_y + zout.s3)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4804 | } |
| 4805 | |
Georgios Pinitas | 8422558 | 2018-05-14 12:00:05 +0100 | [diff] [blame] | 4806 | // Undefine local defines |
| 4807 | #undef COLS_MTX_B |
| 4808 | |
Matthew Bentham | 6f31f8c | 2017-10-27 11:50:06 +0100 | [diff] [blame] | 4809 | #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4810 | /** This OpenCL kernel computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1) |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 4811 | * |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 4812 | * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4813 | * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (e.g. -DMULT_TRANSPOSE1XW_WIDTH=2) |
| 4814 | * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (e.g. -DMULT_INTERLEAVE4X4_HEIGHT=2) |
| 4815 | * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) |
| 4816 | * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4817 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4818 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 4819 | * The activation function is performed after the bias addition |
| 4820 | * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4821 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 4822 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 4823 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 4824 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 4825 | * |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4826 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16 |
| 4827 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 4828 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 4829 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 4830 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 4831 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 4832 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4833 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 4834 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 4835 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 4836 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 4837 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4838 | * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 4839 | * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 4840 | * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) |
| 4841 | * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 4842 | * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) |
| 4843 | * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 4844 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4845 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4846 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4847 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4848 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4849 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4850 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 4851 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4852 | * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4853 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 4854 | * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4855 | */ |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 4856 | __kernel void gemm_mm_interleaved_transposed_f16(IMAGE_DECLARATION(src0), |
| 4857 | IMAGE_DECLARATION(src1), |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4858 | #if defined(BETA) |
| 4859 | IMAGE_DECLARATION(src2), |
| 4860 | #endif // defined(BETA) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 4861 | IMAGE_DECLARATION(dst), |
| 4862 | uint src0_stride_z, |
| 4863 | uint src1_stride_z, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4864 | #if defined(BETA) |
| 4865 | uint src2_stride_z, |
| 4866 | #endif //defined(BETA) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4867 | uint dst_stride_z |
| 4868 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 4869 | , |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 4870 | uint cross_plane_pad |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4871 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 4872 | ) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4873 | { |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4874 | int x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH; |
| 4875 | int y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT; |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 4876 | int z = get_global_id(2); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4877 | |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4878 | // Offset |
| 4879 | const int offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4; |
| 4880 | const int offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 8; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4881 | |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4882 | // src_addr_a = address of matrix A |
| 4883 | // src_addr_b = address of matrix B |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 4884 | int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; |
| 4885 | int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; |
| 4886 | |
| 4887 | #if defined(MATRIX_B_DEPTH) |
| 4888 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 4889 | src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; |
| 4890 | #else // defined(MATRIX_B_DEPTH) |
| 4891 | src1_addr_in_bytes += z * src1_stride_z; |
| 4892 | #endif // defined(MATRIX_B_DEPTH) |
| 4893 | |
| 4894 | __global half *src_addr_a = (__global half *)(src0_ptr + src0_addr_in_bytes); |
| 4895 | __global half *src_addr_b = (__global half *)(src1_ptr + src1_addr_in_bytes); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4896 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 4897 | // Compute end row address for matrix B |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4898 | __global half *src_end_addr_b = src_addr_b + COLS_B; |
| 4899 | |
| 4900 | src_addr_a += offset_row_a; |
| 4901 | src_addr_b += offset_row_b; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4902 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 4903 | // Reset accumulators |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4904 | half8 c0 = 0.0f; |
| 4905 | half8 c1 = 0.0f; |
| 4906 | half8 c2 = 0.0f; |
| 4907 | half8 c3 = 0.0f; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4908 | |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4909 | for(; src_addr_b <= (src_end_addr_b - (int)(16 * MULT_TRANSPOSE1XW_WIDTH)); src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 16 * MULT_TRANSPOSE1XW_WIDTH) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4910 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 4911 | // Load values from matrix A (interleaved) and matrix B (transposed) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4912 | half4 a0 = vload4(0, src_addr_a); |
| 4913 | half8 b0 = vload8(0, src_addr_b); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4914 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4915 | c0 += (half8)a0.s0 * b0; |
| 4916 | c1 += (half8)a0.s1 * b0; |
| 4917 | c2 += (half8)a0.s2 * b0; |
| 4918 | c3 += (half8)a0.s3 * b0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4919 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 4920 | // Load values from matrix A (interleaved) and matrix B (transposed) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4921 | a0 = vload4(0, src_addr_a + 4 * MULT_INTERLEAVE4X4_HEIGHT); |
| 4922 | b0 = vload8(0, src_addr_b + 8 * MULT_TRANSPOSE1XW_WIDTH); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4923 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4924 | c0 += (half8)a0.s0 * b0; |
| 4925 | c1 += (half8)a0.s1 * b0; |
| 4926 | c2 += (half8)a0.s2 * b0; |
| 4927 | c3 += (half8)a0.s3 * b0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4928 | } |
| 4929 | |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4930 | for(; src_addr_b < src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4931 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 4932 | // Load values from matrix A (interleaved) and matrix B (transposed) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 4933 | half4 a0 = vload4(0, src_addr_a); |
| 4934 | half8 b0 = vload8(0, src_addr_b); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4935 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4936 | c0 += (half8)a0.s0 * b0; |
| 4937 | c1 += (half8)a0.s1 * b0; |
| 4938 | c2 += (half8)a0.s2 * b0; |
| 4939 | c3 += (half8)a0.s3 * b0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4940 | } |
| 4941 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 4942 | // Compute destination address |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 4943 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 4944 | |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 4945 | // Compute dst address |
| 4946 | __global uchar *dst_addr = offset(&dst, 0, 0); |
| 4947 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4948 | uint4 zout = 0; |
| 4949 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4950 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 4951 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 4952 | // in order to take into account the presence of possible cross plane paddings |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4953 | // |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 4954 | // | | |
| 4955 | // | plane0 | |
| 4956 | // | | |
| 4957 | // |__________________| |
| 4958 | // |******************| |
| 4959 | // | cross_plane_pad | |
| 4960 | // |******************| |
| 4961 | // | | |
| 4962 | // | plane1 | |
| 4963 | // | | |
| 4964 | // |__________________| |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4965 | |
| 4966 | // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4967 | zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D; |
| 4968 | zout = min(DEPTH_GEMM3D - 1, zout); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4969 | |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 4970 | // Add offset due to the cross plane paddings |
| 4971 | zout *= (cross_plane_pad * dst_stride_y); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4972 | |
| 4973 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 4974 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 4975 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 4976 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 4977 | // Add offset for batched GEMM |
| 4978 | dst_addr += z * dst_stride_z; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 4979 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 4980 | |
| 4981 | // Multiply by the weight of matrix-matrix product and store the result |
| 4982 | #if defined(ALPHA) |
| 4983 | SCALE_BLOCK(4, half, c, ALPHA); |
| 4984 | #endif // defined(ALPHA) |
| 4985 | |
| 4986 | // Add beta*bias |
| 4987 | #if defined(BETA) |
| 4988 | REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0); |
| 4989 | |
| 4990 | #if defined(BROADCAST_BIAS) |
| 4991 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)); |
| 4992 | |
| 4993 | LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero); |
| 4994 | |
| 4995 | #ifndef UNIT_BETA |
| 4996 | SCALE_BLOCK(1, half, bias, BETA); |
| 4997 | #endif // UNIT_BIAS |
| 4998 | |
| 4999 | // c = c + bias[broadcasted] |
| 5000 | ADD_BLOCK_BROADCAST(4, c, bias0); |
| 5001 | |
| 5002 | #else // defined(BROADCAST_BIAS) |
| 5003 | |
| 5004 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id( |
| 5005 | 2) * src2_stride_z; |
| 5006 | |
| 5007 | LOAD_BLOCK(4, 8, half, bias, src2_addr, 0, src2_stride_y, zero); |
| 5008 | |
| 5009 | #ifndef UNIT_BETA |
| 5010 | SCALE_BLOCK(4, half, bias, BETA); |
| 5011 | #endif // UNIT_BIAS |
| 5012 | |
| 5013 | // c = c + bias |
| 5014 | ADD_BLOCK(4, c, bias); |
| 5015 | |
| 5016 | #endif // defined(BROADCAST_BIAS) |
| 5017 | #endif // defined(BETA) |
| 5018 | |
| 5019 | #if defined(ACTIVATION_TYPE) |
| 5020 | ACTIVATION_BLOCK(4, ACTIVATION_TYPE, half, c, A_VAL, B_VAL); |
| 5021 | #endif // defined(ACTIVATION_TYPE) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 5022 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 5023 | // Store 4x8 block |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5024 | vstore8(c0, 0, (__global half *)(dst_addr + 0 * dst_stride_y + zout.s0)); |
| 5025 | vstore8(c1, 0, (__global half *)(dst_addr + 1 * dst_stride_y + zout.s1)); |
| 5026 | vstore8(c2, 0, (__global half *)(dst_addr + 2 * dst_stride_y + zout.s2)); |
| 5027 | vstore8(c3, 0, (__global half *)(dst_addr + 3 * dst_stride_y + zout.s3)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5028 | } |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5029 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5030 | /** This OpenCL kernel computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1) while accumulating the result in a 32 floating point variable. |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 5031 | * |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5032 | * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5033 | * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (e.g. -DMULT_TRANSPOSE1XW_WIDTH=2) |
| 5034 | * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (e.g. -DMULT_INTERLEAVE4X4_HEIGHT=2) |
| 5035 | * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) |
| 5036 | * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5037 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5038 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 5039 | * The activation function is performed after the bias addition |
| 5040 | * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5041 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 5042 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 5043 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 5044 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 5045 | * |
| 5046 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16 |
| 5047 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 5048 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 5049 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 5050 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5051 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 5052 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
| 5053 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 5054 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 5055 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 5056 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5057 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5058 | * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 5059 | * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 5060 | * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) |
| 5061 | * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 5062 | * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5063 | * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5064 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
| 5065 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 5066 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 5067 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 5068 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5069 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 5070 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 5071 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5072 | * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5073 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 5074 | * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
| 5075 | */ |
| 5076 | __kernel void gemm_mm_interleaved_transposed_f16_acc32(IMAGE_DECLARATION(src0), |
| 5077 | IMAGE_DECLARATION(src1), |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5078 | #if defined(BETA) |
| 5079 | IMAGE_DECLARATION(src2), |
| 5080 | #endif // defined(BETA) |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5081 | IMAGE_DECLARATION(dst), |
| 5082 | uint src0_stride_z, |
| 5083 | uint src1_stride_z, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5084 | #if defined(BETA) |
| 5085 | uint src2_stride_z, |
| 5086 | #endif //defined(BETA) |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5087 | uint dst_stride_z |
| 5088 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 5089 | , |
| 5090 | uint cross_plane_pad |
| 5091 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 5092 | ) |
| 5093 | { |
| 5094 | int x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH; |
| 5095 | int y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT; |
| 5096 | int z = get_global_id(2); |
| 5097 | |
| 5098 | // Offset |
| 5099 | const int offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4; |
| 5100 | const int offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 8; |
| 5101 | |
| 5102 | // src_addr_a = address of matrix A |
| 5103 | // src_addr_b = address of matrix B |
| 5104 | int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; |
| 5105 | int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; |
| 5106 | |
| 5107 | #if defined(MATRIX_B_DEPTH) |
| 5108 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 5109 | src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; |
| 5110 | #else // defined(MATRIX_B_DEPTH) |
| 5111 | src1_addr_in_bytes += z * src1_stride_z; |
| 5112 | #endif // defined(MATRIX_B_DEPTH) |
| 5113 | |
| 5114 | __global half *src_addr_a = (__global half *)(src0_ptr + src0_addr_in_bytes); |
| 5115 | __global half *src_addr_b = (__global half *)(src1_ptr + src1_addr_in_bytes); |
| 5116 | |
| 5117 | // Compute end row address for matrix B |
| 5118 | __global half *src_end_addr_b = src_addr_b + COLS_B; |
| 5119 | |
| 5120 | src_addr_a += offset_row_a; |
| 5121 | src_addr_b += offset_row_b; |
| 5122 | |
| 5123 | // Reset accumulators |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5124 | float8 c0 = 0.0f; |
| 5125 | float8 c1 = 0.0f; |
| 5126 | float8 c2 = 0.0f; |
| 5127 | float8 c3 = 0.0f; |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5128 | |
| 5129 | for(; src_addr_b <= (src_end_addr_b - (int)(16 * MULT_TRANSPOSE1XW_WIDTH)); src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 16 * MULT_TRANSPOSE1XW_WIDTH) |
| 5130 | { |
| 5131 | // Load values from matrix A (interleaved) and matrix B (transposed) |
| 5132 | float4 a0 = convert_float4(vload4(0, src_addr_a)); |
| 5133 | float8 b0 = convert_float8(vload8(0, src_addr_b)); |
| 5134 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5135 | c0 += (float8)a0.s0 * b0; |
| 5136 | c1 += (float8)a0.s1 * b0; |
| 5137 | c2 += (float8)a0.s2 * b0; |
| 5138 | c3 += (float8)a0.s3 * b0; |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5139 | |
| 5140 | // Load values from matrix A (interleaved) and matrix B (transposed) |
| 5141 | a0 = convert_float4(vload4(0, src_addr_a + 4 * MULT_INTERLEAVE4X4_HEIGHT)); |
| 5142 | b0 = convert_float8(vload8(0, src_addr_b + 8 * MULT_TRANSPOSE1XW_WIDTH)); |
| 5143 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5144 | c0 += (float8)a0.s0 * b0; |
| 5145 | c1 += (float8)a0.s1 * b0; |
| 5146 | c2 += (float8)a0.s2 * b0; |
| 5147 | c3 += (float8)a0.s3 * b0; |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5148 | } |
| 5149 | |
| 5150 | for(; src_addr_b < src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH) |
| 5151 | { |
| 5152 | // Load values from matrix A (interleaved) and matrix B (transposed) |
| 5153 | float4 a0 = convert_float4(vload4(0, src_addr_a)); |
| 5154 | float8 b0 = convert_float8(vload8(0, src_addr_b)); |
| 5155 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5156 | c0 += (float8)a0.s0 * b0; |
| 5157 | c1 += (float8)a0.s1 * b0; |
| 5158 | c2 += (float8)a0.s2 * b0; |
| 5159 | c3 += (float8)a0.s3 * b0; |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5160 | } |
| 5161 | |
| 5162 | // Compute destination address |
| 5163 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 5164 | |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5165 | // Compute dst address |
| 5166 | __global uchar *dst_addr = offset(&dst, 0, 0); |
| 5167 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5168 | uint4 zout = 0; |
| 5169 | |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5170 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 5171 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
| 5172 | // in order to take into account the presence of possible cross plane paddings |
| 5173 | // |
| 5174 | // | | |
| 5175 | // | plane0 | |
| 5176 | // | | |
| 5177 | // |__________________| |
| 5178 | // |******************| |
| 5179 | // | cross_plane_pad | |
| 5180 | // |******************| |
| 5181 | // | | |
| 5182 | // | plane1 | |
| 5183 | // | | |
| 5184 | // |__________________| |
| 5185 | |
| 5186 | // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5187 | zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D; |
| 5188 | zout = min(DEPTH_GEMM3D - 1, zout); |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5189 | |
| 5190 | // Add offset due to the cross plane paddings |
| 5191 | zout *= (cross_plane_pad * dst_stride_y); |
| 5192 | |
| 5193 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 5194 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 5195 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5196 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 5197 | // Add offset for batched GEMM |
| 5198 | dst_addr += z * dst_stride_z; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5199 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 5200 | |
| 5201 | // Multiply by the weight of matrix-matrix product and store the result |
| 5202 | #if defined(ALPHA) |
| 5203 | SCALE_BLOCK(4, float, c, ALPHA); |
| 5204 | #endif // defined(ALPHA) |
| 5205 | |
| 5206 | #if defined(BETA) |
| 5207 | REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0); |
| 5208 | |
| 5209 | #if defined(BROADCAST_BIAS) |
| 5210 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)); |
| 5211 | |
| 5212 | LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero); |
| 5213 | |
| 5214 | float8 bias_f0 = convert_float8(bias0); |
| 5215 | |
| 5216 | #ifndef UNIT_BETA |
| 5217 | SCALE_BLOCK(1, float, bias_f, BETA); |
| 5218 | #endif // UNIT_BIAS |
| 5219 | |
| 5220 | // c = c + bias[broadcasted] |
| 5221 | ADD_BLOCK_BROADCAST(4, c, bias_f0); |
| 5222 | |
| 5223 | #else // defined(BROADCAST_BIAS) |
| 5224 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id( |
| 5225 | 2) * src2_stride_z; |
| 5226 | |
| 5227 | LOAD_BLOCK(4, 8, half, bias, src2_addr, 0, src2_stride_y, zero); |
| 5228 | |
| 5229 | float8 bias_f0 = convert_float8(bias0); |
| 5230 | float8 bias_f1 = convert_float8(bias1); |
| 5231 | float8 bias_f2 = convert_float8(bias2); |
| 5232 | float8 bias_f3 = convert_float8(bias3); |
| 5233 | |
| 5234 | #ifndef UNIT_BETA |
| 5235 | SCALE_BLOCK(4, float, bias_f, BETA); |
| 5236 | #endif // UNIT_BIAS |
| 5237 | |
| 5238 | // c = c + bias |
| 5239 | ADD_BLOCK(4, c, bias_f); |
| 5240 | |
| 5241 | #endif // defined(BROADCAST_BIAS) |
| 5242 | #endif // defined(BETA) |
| 5243 | |
| 5244 | half8 c_h0 = convert_half8(c0); |
| 5245 | half8 c_h1 = convert_half8(c1); |
| 5246 | half8 c_h2 = convert_half8(c2); |
| 5247 | half8 c_h3 = convert_half8(c3); |
| 5248 | |
| 5249 | #if defined(ACTIVATION_TYPE) |
| 5250 | ACTIVATION_BLOCK(4, ACTIVATION_TYPE, half, c_h, A_VAL, B_VAL); |
| 5251 | #endif // defined(ACTIVATION_TYPE) |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5252 | |
| 5253 | // Store 4x8 block |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5254 | vstore8(c_h0, 0, (__global half *)(dst_addr + 0 * dst_stride_y + zout.s0)); |
| 5255 | vstore8(c_h1, 0, (__global half *)(dst_addr + 1 * dst_stride_y + zout.s1)); |
| 5256 | vstore8(c_h2, 0, (__global half *)(dst_addr + 2 * dst_stride_y + zout.s2)); |
| 5257 | vstore8(c_h3, 0, (__global half *)(dst_addr + 3 * dst_stride_y + zout.s3)); |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 5258 | } |
| 5259 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5260 | /** This OpenCL kernel optimized for Bifrost architectures computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1) |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 5261 | * |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5262 | * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5263 | * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (e.g. -DMULT_TRANSPOSE1XW_WIDTH=2) |
| 5264 | * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (e.g. -DMULT_INTERLEAVE4X4_HEIGHT=2) |
| 5265 | * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) |
| 5266 | * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5267 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5268 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 5269 | * The activation function is performed after the bias addition |
| 5270 | * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5271 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 5272 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 5273 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 5274 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 5275 | * |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5276 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16 |
| 5277 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 5278 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 5279 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 5280 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5281 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 5282 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
| 5283 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 5284 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 5285 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 5286 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5287 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5288 | * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 5289 | * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 5290 | * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) |
| 5291 | * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 5292 | * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5293 | * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5294 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
| 5295 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 5296 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 5297 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 5298 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5299 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5300 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 5301 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 5302 | * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5303 | * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5304 | */ |
| 5305 | __kernel void gemm_mm_interleaved_transposed_f16_bifrost(IMAGE_DECLARATION(src0), |
| 5306 | IMAGE_DECLARATION(src1), |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5307 | #if defined(BETA) |
| 5308 | IMAGE_DECLARATION(src2), |
| 5309 | #endif // defined(BETA) |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5310 | IMAGE_DECLARATION(dst), |
| 5311 | uint src0_stride_z, |
| 5312 | uint src1_stride_z, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5313 | #if defined(BETA) |
| 5314 | uint src2_stride_z, |
| 5315 | #endif //defined(BETA) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5316 | uint dst_stride_z |
| 5317 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 5318 | , |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 5319 | uint cross_plane_pad |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5320 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 5321 | ) |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5322 | { |
| 5323 | int x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH; |
| 5324 | int y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT; |
| 5325 | int z = get_global_id(2); |
| 5326 | |
| 5327 | // Offset |
| 5328 | const int offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4; |
| 5329 | const int offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 8; |
| 5330 | |
| 5331 | // src_addr_a = address of matrix A |
| 5332 | // src_addr_b = address of matrix B |
| 5333 | int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; |
| 5334 | int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; |
| 5335 | |
| 5336 | #if defined(MATRIX_B_DEPTH) |
| 5337 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 5338 | src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; |
| 5339 | #else // defined(MATRIX_B_DEPTH) |
| 5340 | src1_addr_in_bytes += z * src1_stride_z; |
| 5341 | #endif // defined(MATRIX_B_DEPTH) |
| 5342 | |
| 5343 | __global half *src_addr_a = (__global half *)(src0_ptr + src0_addr_in_bytes); |
| 5344 | __global half *src_addr_b = (__global half *)(src1_ptr + src1_addr_in_bytes); |
| 5345 | |
| 5346 | // Compute end row address for matrix B |
| 5347 | __global half *src_end_addr_b = src_addr_b + COLS_B; |
| 5348 | |
| 5349 | src_addr_a += offset_row_a; |
| 5350 | src_addr_b += offset_row_b; |
| 5351 | |
| 5352 | // Reset accumulators |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5353 | half8 c0 = 0.0f; |
| 5354 | half8 c1 = 0.0f; |
| 5355 | half8 c2 = 0.0f; |
| 5356 | half8 c3 = 0.0f; |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5357 | |
| 5358 | #define COLS_MTX_B (COLS_B / (8 * MULT_TRANSPOSE1XW_WIDTH)) |
| 5359 | |
| 5360 | int i = 0; |
| 5361 | for(; i <= (int)(COLS_MTX_B - 4); i += 4) |
| 5362 | { |
| 5363 | #if MULT_INTERLEAVE4X4_HEIGHT == 1 |
| 5364 | // Load values from matrix A (interleaved) and matrix B (transposed) |
| 5365 | half8 a0 = vload8(0, src_addr_a); |
| 5366 | half8 b0 = vload8(0, src_addr_b); |
| 5367 | |
| 5368 | src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT; |
| 5369 | src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH; |
| 5370 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5371 | c0 = fma((half8)a0.s0, b0, c0); |
| 5372 | c1 = fma((half8)a0.s1, b0, c1); |
| 5373 | c2 = fma((half8)a0.s2, b0, c2); |
| 5374 | c3 = fma((half8)a0.s3, b0, c3); |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5375 | |
| 5376 | // Load values from matrix B (transposed) |
| 5377 | b0 = vload8(0, src_addr_b); |
| 5378 | |
| 5379 | src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH; |
| 5380 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5381 | c0 = fma((half8)a0.s4, b0, c0); |
| 5382 | c1 = fma((half8)a0.s5, b0, c1); |
| 5383 | c2 = fma((half8)a0.s6, b0, c2); |
| 5384 | c3 = fma((half8)a0.s7, b0, c3); |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5385 | |
| 5386 | // Load values from matrix A (interleaved) and matrix B (transposed) |
| 5387 | a0 = vload8(0, src_addr_a); |
| 5388 | b0 = vload8(0, src_addr_b); |
| 5389 | |
| 5390 | src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT; |
| 5391 | src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH; |
| 5392 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5393 | c0 = fma((half8)a0.s0, b0, c0); |
| 5394 | c1 = fma((half8)a0.s1, b0, c1); |
| 5395 | c2 = fma((half8)a0.s2, b0, c2); |
| 5396 | c3 = fma((half8)a0.s3, b0, c3); |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5397 | |
| 5398 | // Load values from matrix B (transposed) |
| 5399 | b0 = vload8(0, src_addr_b); |
| 5400 | |
| 5401 | src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH; |
| 5402 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5403 | c0 = fma((half8)a0.s4, b0, c0); |
| 5404 | c1 = fma((half8)a0.s5, b0, c1); |
| 5405 | c2 = fma((half8)a0.s6, b0, c2); |
| 5406 | c3 = fma((half8)a0.s7, b0, c3); |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5407 | #else // MULT_INTERLEAVE4X4_HEIGHT == 1 |
| 5408 | // Load values from matrix A (interleaved) and matrix B (transposed) |
| 5409 | half4 a0 = vload4(0, src_addr_a); |
| 5410 | half8 b0 = vload8(0, src_addr_b); |
| 5411 | |
| 5412 | src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT; |
| 5413 | src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH; |
| 5414 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5415 | c0 = fma((half8)a0.s0, b0, c0); |
| 5416 | c1 = fma((half8)a0.s1, b0, c1); |
| 5417 | c2 = fma((half8)a0.s2, b0, c2); |
| 5418 | c3 = fma((half8)a0.s3, b0, c3); |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5419 | |
| 5420 | // Load values from matrix A (interleaved) and matrix B (transposed) |
| 5421 | a0 = vload4(0, src_addr_a); |
| 5422 | b0 = vload8(0, src_addr_b); |
| 5423 | |
| 5424 | src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT; |
| 5425 | src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH; |
| 5426 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5427 | c0 = fma((half8)a0.s0, b0, c0); |
| 5428 | c1 = fma((half8)a0.s1, b0, c1); |
| 5429 | c2 = fma((half8)a0.s2, b0, c2); |
| 5430 | c3 = fma((half8)a0.s3, b0, c3); |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5431 | |
| 5432 | // Load values from matrix A (interleaved) and matrix B (transposed) |
| 5433 | a0 = vload4(0, src_addr_a); |
| 5434 | b0 = vload8(0, src_addr_b); |
| 5435 | |
| 5436 | src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT; |
| 5437 | src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH; |
| 5438 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5439 | c0 = fma((half8)a0.s0, b0, c0); |
| 5440 | c1 = fma((half8)a0.s1, b0, c1); |
| 5441 | c2 = fma((half8)a0.s2, b0, c2); |
| 5442 | c3 = fma((half8)a0.s3, b0, c3); |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5443 | |
| 5444 | // Load values from matrix A (interleaved) and matrix B (transposed) |
| 5445 | a0 = vload4(0, src_addr_a); |
| 5446 | b0 = vload8(0, src_addr_b); |
| 5447 | |
| 5448 | src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT; |
| 5449 | src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH; |
| 5450 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5451 | c0 = fma((half8)a0.s0, b0, c0); |
| 5452 | c1 = fma((half8)a0.s1, b0, c1); |
| 5453 | c2 = fma((half8)a0.s2, b0, c2); |
| 5454 | c3 = fma((half8)a0.s3, b0, c3); |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5455 | #endif // MULT_INTERLEAVE4X4_HEIGHT == 1 |
| 5456 | } |
| 5457 | |
| 5458 | for(; i < (int)(COLS_MTX_B); ++i) |
| 5459 | { |
| 5460 | // Load values from matrix A (interleaved) and matrix B (transposed) |
| 5461 | half4 a0 = vload4(0, src_addr_a); |
| 5462 | half8 b0 = vload8(0, src_addr_b); |
| 5463 | |
| 5464 | src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT; |
| 5465 | src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH; |
| 5466 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5467 | c0 = fma((half8)a0.s0, b0, c0); |
| 5468 | c1 = fma((half8)a0.s1, b0, c1); |
| 5469 | c2 = fma((half8)a0.s2, b0, c2); |
| 5470 | c3 = fma((half8)a0.s3, b0, c3); |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5471 | } |
| 5472 | |
| 5473 | // Compute destination address |
| 5474 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 5475 | |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5476 | // Compute dst address |
| 5477 | __global uchar *dst_addr = offset(&dst, 0, 0); |
| 5478 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5479 | uint4 zout = 0; |
| 5480 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5481 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 5482 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 5483 | // in order to take into account the presence of possible cross plane paddings |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5484 | // |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 5485 | // | | |
| 5486 | // | plane0 | |
| 5487 | // | | |
| 5488 | // |__________________| |
| 5489 | // |******************| |
| 5490 | // | cross_plane_pad | |
| 5491 | // |******************| |
| 5492 | // | | |
| 5493 | // | plane1 | |
| 5494 | // | | |
| 5495 | // |__________________| |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5496 | |
| 5497 | // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5498 | zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D; |
| 5499 | zout = min(DEPTH_GEMM3D - 1, zout); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5500 | |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 5501 | // Add offset due to the cross plane paddings |
| 5502 | zout *= (cross_plane_pad * dst_stride_y); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5503 | |
| 5504 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 5505 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 5506 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5507 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 5508 | // Add offset for batched GEMM |
| 5509 | dst_addr += z * dst_stride_z; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5510 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 5511 | |
| 5512 | // Multiply by the weight of matrix-matrix product and store the result |
| 5513 | #if defined(ALPHA) |
| 5514 | SCALE_BLOCK(4, half, c, ALPHA); |
| 5515 | #endif // defined(ALPHA) |
| 5516 | |
| 5517 | // Add beta*bias |
| 5518 | #if defined(BETA) |
| 5519 | REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0); |
| 5520 | |
| 5521 | #if defined(BROADCAST_BIAS) |
| 5522 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)); |
| 5523 | |
| 5524 | LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero); |
| 5525 | |
| 5526 | #ifndef UNIT_BETA |
| 5527 | SCALE_BLOCK(1, half, bias, BETA); |
| 5528 | #endif // UNIT_BIAS |
| 5529 | |
| 5530 | // c = c + bias[broadcasted] |
| 5531 | ADD_BLOCK_BROADCAST(4, c, bias0); |
| 5532 | |
| 5533 | #else // defined(BROADCAST_BIAS) |
| 5534 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id( |
| 5535 | 2) * src2_stride_z; |
| 5536 | |
| 5537 | LOAD_BLOCK(4, 8, half, bias, src2_addr, 0, src2_stride_y, zero); |
| 5538 | |
| 5539 | #ifndef UNIT_BETA |
| 5540 | SCALE_BLOCK(4, half, bias, BETA); |
| 5541 | #endif // UNIT_BIAS |
| 5542 | |
| 5543 | // c = c + bias |
| 5544 | ADD_BLOCK(4, c, bias); |
| 5545 | |
| 5546 | #endif // defined(BROADCAST_BIAS) |
| 5547 | #endif // defined(BETA) |
| 5548 | |
| 5549 | #if defined(ACTIVATION_TYPE) |
| 5550 | ACTIVATION_BLOCK(4, ACTIVATION_TYPE, half, c, A_VAL, B_VAL); |
| 5551 | #endif // defined(ACTIVATION_TYPE) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5552 | |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5553 | // Store 4x8 block |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5554 | vstore8(c0, 0, (__global half *)(dst_addr + 0 * dst_stride_y + zout.s0)); |
| 5555 | vstore8(c1, 0, (__global half *)(dst_addr + 1 * dst_stride_y + zout.s1)); |
| 5556 | vstore8(c2, 0, (__global half *)(dst_addr + 2 * dst_stride_y + zout.s2)); |
| 5557 | vstore8(c3, 0, (__global half *)(dst_addr + 3 * dst_stride_y + zout.s3)); |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 5558 | } |
Georgios Pinitas | 8422558 | 2018-05-14 12:00:05 +0100 | [diff] [blame] | 5559 | |
| 5560 | // Undefine local defines |
| 5561 | #undef COLS_MTX_B |
| 5562 | |
Matthew Bentham | 6f31f8c | 2017-10-27 11:50:06 +0100 | [diff] [blame] | 5563 | #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5564 | |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 5565 | #endif // defined(COLS_B) && defined(MULT_TRANSPOSE1XW_WIDTH) && defined(MULT_INTERLEAVE4X4_HEIGHT) |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 5566 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5567 | #if defined(COLS_A) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && (NUM_ELEMS_PROCESSED_PER_THREAD_Y) |
| 5568 | #if defined(DATA_TYPE) |
| 5569 | #define VECTOR_TYPE VEC_DATA_TYPE(DATA_TYPE, NUM_ELEMS_PROCESSED_PER_THREAD_X) |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 5570 | /** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not been reshaped. |
| 5571 | * |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5572 | * @note This OpenCL kernel works with floating point data types (F16/F32) |
| 5573 | * @note The floating point data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) |
| 5574 | * @note The number of elements processed along the x and y directions must be passed at compile time using -DNUM_ELEMS_PROCESSED_PER_THREAD_X and -DNUM_ELEMS_PROCESSED_PER_THREAD_Y |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 5575 | * @note The number of matrix A columns and the optional alpha's value need to be passed at compile time using -DCOLS_A and -DALPHA |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5576 | * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) |
| 5577 | * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5578 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5579 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 5580 | * The activation function is performed after the bias addition |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5581 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 5582 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5583 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 5584 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 5585 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 5586 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 5587 | * |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5588 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16/F32 |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5589 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 5590 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 5591 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 5592 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5593 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 5594 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5595 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 5596 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 5597 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 5598 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5599 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5600 | * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 5601 | * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 5602 | * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) |
| 5603 | * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 5604 | * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5605 | * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 5606 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5607 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 5608 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 5609 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 5610 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5611 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5612 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 5613 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5614 | * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5615 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5616 | * @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D) |
| 5617 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements for the output tensor (only if defined REINTERPRET_OUTPUT_AS_3D) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5618 | */ |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5619 | __kernel void gemm_mm_floating_point(IMAGE_DECLARATION(src0), |
| 5620 | IMAGE_DECLARATION(src1), |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5621 | #if defined(BETA) |
| 5622 | IMAGE_DECLARATION(src2), |
| 5623 | #endif // defined(BETA) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 5624 | IMAGE_DECLARATION(dst), |
| 5625 | uint src0_stride_z, |
| 5626 | uint src1_stride_z, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5627 | #if defined(BETA) |
| 5628 | uint src2_stride_z, |
| 5629 | #endif //defined(BETA) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5630 | uint dst_stride_z |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5631 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 5632 | , |
| 5633 | uint src_cross_plane_pad |
| 5634 | #endif // REINTERPRET_INPUT_AS_3D |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5635 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 5636 | , |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5637 | uint dst_cross_plane_pad |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5638 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 5639 | ) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5640 | { |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5641 | int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5642 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5643 | // Compute starting address for matrix A and Matrix B |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5644 | int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5645 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5646 | // Update address for the matrix A |
| 5647 | src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5648 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5649 | // Update address for the matrix B |
| 5650 | src_addr.s1 += idx * sizeof(DATA_TYPE); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5651 | |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5652 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 5653 | // Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension |
| 5654 | // in order to take into account the presence of possible cross plane paddings |
| 5655 | // |
| 5656 | // | | |
| 5657 | // | plane0 | |
| 5658 | // | | |
| 5659 | // |__________________| |
| 5660 | // |******************| |
| 5661 | // | cross_plane_pad | |
| 5662 | // |******************| |
| 5663 | // | | |
| 5664 | // | plane1 | |
| 5665 | // | | |
| 5666 | // |__________________| |
| 5667 | |
| 5668 | // The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
| 5669 | uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
| 5670 | zin = min(DEPTH_GEMM3D - 1, zin); |
| 5671 | |
| 5672 | // Add offset due to the cross plane paddings |
| 5673 | zin *= (src_cross_plane_pad * src0_stride_y); |
| 5674 | |
| 5675 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 5676 | // multiply src0_stride_z by DEPTH_GEMM3D |
| 5677 | src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D; |
| 5678 | |
| 5679 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 5680 | |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 5681 | // Add offset for batched GEMM |
| 5682 | src_addr.s0 += get_global_id(2) * src0_stride_z; |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 5683 | |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5684 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 5685 | |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 5686 | #if defined(MATRIX_B_DEPTH) |
| 5687 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 5688 | src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; |
| 5689 | #else // defined(MATRIX_B_DEPTH) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 5690 | src_addr.s1 += get_global_id(2) * src1_stride_z; |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 5691 | #endif // defined(MATRIX_B_DEPTH) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 5692 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5693 | int end_row_vec_a = src_addr.s0 + (COLS_A * sizeof(DATA_TYPE)); |
| 5694 | |
| 5695 | VECTOR_TYPE acc0 = 0.0f; |
| 5696 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 5697 | VECTOR_TYPE acc1 = 0.0f; |
| 5698 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 5699 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 5700 | VECTOR_TYPE acc2 = 0.0f; |
| 5701 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 5702 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 5703 | VECTOR_TYPE acc3 = 0.0f; |
| 5704 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 5705 | |
Georgios Pinitas | 96880cf | 2017-10-20 18:52:20 +0100 | [diff] [blame] | 5706 | for(; src_addr.s0 <= (end_row_vec_a - 2 * (int)sizeof(DATA_TYPE)); src_addr += (int2)(2 * sizeof(DATA_TYPE), 2 * src1_stride_y)) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5707 | { |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5708 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 5709 | // Load values from matrix A |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 5710 | LOAD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, 2, DATA_TYPE, a, src0_ptr, src_addr.s0, src0_stride_y, zin.s); |
| 5711 | #else // defined(REINTERPRET_INPUT_AS_3D) |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5712 | // Load values from matrix A |
| 5713 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 5714 | a0 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); |
| 5715 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 5716 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 5717 | a1 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); |
| 5718 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 5719 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 5720 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 5721 | a2 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); |
| 5722 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 5723 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 5724 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 5725 | a3 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); |
| 5726 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5727 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 5728 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5729 | // Load values from matrix B |
| 5730 | VECTOR_TYPE b0 = VLOAD(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1)); |
| 5731 | VECTOR_TYPE b1 = VLOAD(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1 + src1_stride_y)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5732 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5733 | // Accumulate |
| 5734 | acc0 += b0 * (VECTOR_TYPE)a0.s0; |
| 5735 | acc0 += b1 * (VECTOR_TYPE)a0.s1; |
| 5736 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 5737 | acc1 += b0 * (VECTOR_TYPE)a1.s0; |
| 5738 | acc1 += b1 * (VECTOR_TYPE)a1.s1; |
| 5739 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 5740 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 5741 | acc2 += b0 * (VECTOR_TYPE)a2.s0; |
| 5742 | acc2 += b1 * (VECTOR_TYPE)a2.s1; |
| 5743 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 5744 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 5745 | acc3 += b0 * (VECTOR_TYPE)a3.s0; |
| 5746 | acc3 += b1 * (VECTOR_TYPE)a3.s1; |
| 5747 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5748 | } |
| 5749 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5750 | for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(DATA_TYPE), src1_stride_y)) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5751 | { |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5752 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 5753 | // Load values from matrix A |
| 5754 | DATA_TYPE a0 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0)); |
| 5755 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 5756 | DATA_TYPE a1 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1)); |
| 5757 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 5758 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 5759 | DATA_TYPE a2 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2)); |
| 5760 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 5761 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 5762 | DATA_TYPE a3 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3)); |
| 5763 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 5764 | #else // defined(REINTERPRET_INPUT_AS_3D) |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5765 | // Load values from matrix A |
| 5766 | DATA_TYPE a0 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); |
| 5767 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 5768 | DATA_TYPE a1 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); |
| 5769 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 5770 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 5771 | DATA_TYPE a2 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); |
| 5772 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 5773 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 5774 | DATA_TYPE a3 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); |
| 5775 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5776 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 5777 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5778 | // Load values from matrix B |
| 5779 | VECTOR_TYPE b0 = VLOAD(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5780 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5781 | // Accumulate |
| 5782 | acc0 += b0 * (VECTOR_TYPE)a0; |
| 5783 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 5784 | acc1 += b0 * (VECTOR_TYPE)a1; |
| 5785 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 5786 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 5787 | acc2 += b0 * (VECTOR_TYPE)a2; |
| 5788 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 5789 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 5790 | acc3 += b0 * (VECTOR_TYPE)a3; |
| 5791 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5792 | } |
| 5793 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5794 | int z = get_global_id(2); |
| 5795 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5796 | // Compute destination address |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5797 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 5798 | |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 5799 | // Compute dst address |
| 5800 | __global uchar *dst_addr = offset(&dst, 0, 0); |
| 5801 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5802 | uint4 zout = 0; |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5803 | |
| 5804 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5805 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5806 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 5807 | // in order to take into account the presence of possible cross plane paddings |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5808 | // |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 5809 | // | | |
| 5810 | // | plane0 | |
| 5811 | // | | |
| 5812 | // |__________________| |
| 5813 | // |******************| |
| 5814 | // | cross_plane_pad | |
| 5815 | // |******************| |
| 5816 | // | | |
| 5817 | // | plane1 | |
| 5818 | // | | |
| 5819 | // |__________________| |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5820 | |
| 5821 | // The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5822 | zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
| 5823 | zout = min(DEPTH_GEMM3D - 1, zout); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5824 | |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 5825 | // Add offset due to the cross plane paddings |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5826 | zout *= (dst_cross_plane_pad * dst_stride_y); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5827 | |
| 5828 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 5829 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 5830 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5831 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 5832 | // Add offset for batched GEMM |
| 5833 | dst_addr += z * dst_stride_z; |
| 5834 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 5835 | |
| 5836 | // Multiply by the weight of matrix-matrix product and store the result |
| 5837 | #if defined(ALPHA) |
| 5838 | SCALE_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, DATA_TYPE, acc, ALPHA); |
| 5839 | #endif // defined(ALPHA) |
| 5840 | |
| 5841 | // Add beta*bias |
| 5842 | #if defined(BETA) |
| 5843 | REPEAT_VAR_INIT_TO_CONST(NUM_ELEMS_PROCESSED_PER_THREAD_Y, uint, zero, 0); |
| 5844 | |
| 5845 | #if defined(BROADCAST_BIAS) |
| 5846 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)NUM_ELEMS_PROCESSED_PER_THREAD_X * sizeof(DATA_TYPE)); |
| 5847 | |
| 5848 | LOAD_BLOCK(1, NUM_ELEMS_PROCESSED_PER_THREAD_X, DATA_TYPE, bias, src2_addr, 0, src2_stride_y, zero); |
| 5849 | |
| 5850 | #ifndef UNIT_BETA |
| 5851 | SCALE_BLOCK(1, DATA_TYPE, bias, BETA); |
| 5852 | #endif // UNIT_BIAS |
| 5853 | |
| 5854 | // c = c + bias[broadcasted] |
| 5855 | ADD_BLOCK_BROADCAST(NUM_ELEMS_PROCESSED_PER_THREAD_Y, acc, bias0); |
| 5856 | |
| 5857 | #else // defined(BROADCAST_BIAS) |
| 5858 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)NUM_ELEMS_PROCESSED_PER_THREAD_X * sizeof(DATA_TYPE)) + (get_global_id(1) * |
| 5859 | (uint)NUM_ELEMS_PROCESSED_PER_THREAD_Y * src2_stride_y) + get_global_id(2) * src2_stride_z; |
| 5860 | |
| 5861 | LOAD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, NUM_ELEMS_PROCESSED_PER_THREAD_X, DATA_TYPE, bias, src2_addr, 0, src2_stride_y, zero); |
| 5862 | |
| 5863 | #ifndef UNIT_BETA |
| 5864 | SCALE_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, DATA_TYPE, bias, BETA); |
| 5865 | #endif // UNIT_BIAS |
| 5866 | |
| 5867 | // c = c + bias |
| 5868 | ADD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, acc, bias); |
| 5869 | |
| 5870 | #endif // defined(BROADCAST_BIAS) |
| 5871 | #endif // defined(BETA) |
| 5872 | |
| 5873 | #if defined(ACTIVATION_TYPE) |
| 5874 | ACTIVATION_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, ACTIVATION_TYPE, DATA_TYPE, acc, A_VAL, B_VAL); |
| 5875 | #endif // defined(ACTIVATION_TYPE) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5876 | |
| 5877 | // Store output block |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 5878 | STORE_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, NUM_ELEMS_PROCESSED_PER_THREAD_X, DATA_TYPE, acc, dst_addr, dst_stride_y, zout.s); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 5879 | } |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 5880 | #endif // defined(DATA_TYPE) |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 5881 | |
Michele Di Giorgio | f6f08da | 2018-04-26 10:24:30 +0100 | [diff] [blame] | 5882 | /** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not been reshaped |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 5883 | * |
| 5884 | * @note This OpenCL kernel works with the 32-bit floating point data type (float) and uses the fma units. |
| 5885 | * @note The number of elements processed along the x and y directions must be passed at compile time using -DNUM_ELEMS_PROCESSED_PER_THREAD_X and -DNUM_ELEMS_PROCESSED_PER_THREAD_Y. |
| 5886 | * This kernel optimally uses -DNUM_ELEMS_PROCESSED_PER_THREAD_X=4. |
| 5887 | * @note The number of matrix A columns must be passed at compile time using -DCOLS_A. |
| 5888 | * @note The optional value of scalar alpha is passed at compile time using -DALPHA=alpha |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5889 | * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) |
| 5890 | * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 5891 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5892 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 5893 | * The activation function is performed after the bias addition |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5894 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 5895 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5896 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 5897 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 5898 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 5899 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 5900 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5901 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 5902 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 5903 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 5904 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 5905 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5906 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 5907 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
| 5908 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 5909 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 5910 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 5911 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5912 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5913 | * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 5914 | * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 5915 | * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) |
| 5916 | * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 5917 | * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5918 | * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 5919 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
| 5920 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 5921 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 5922 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 5923 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 5924 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5925 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 5926 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5927 | * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5928 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5929 | * @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D) |
| 5930 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 5931 | */ |
| 5932 | __kernel void gemm_mm_floating_point_f32_bifrost(IMAGE_DECLARATION(src0), |
| 5933 | IMAGE_DECLARATION(src1), |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5934 | #if defined(BETA) |
| 5935 | IMAGE_DECLARATION(src2), |
| 5936 | #endif // defined(BETA) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 5937 | IMAGE_DECLARATION(dst), |
| 5938 | uint src0_stride_z, |
| 5939 | uint src1_stride_z, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 5940 | #if defined(BETA) |
| 5941 | uint src2_stride_z, |
| 5942 | #endif //defined(BETA) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5943 | uint dst_stride_z |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5944 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 5945 | , |
| 5946 | uint src_cross_plane_pad |
| 5947 | #endif // REINTERPRET_INPUT_AS_3D |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5948 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 5949 | , |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5950 | uint dst_cross_plane_pad |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 5951 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 5952 | ) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 5953 | { |
| 5954 | int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; |
| 5955 | |
| 5956 | // Compute starting address for matrix A and matrix B |
| 5957 | int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); |
| 5958 | |
| 5959 | // Update address for matrix A |
| 5960 | src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y; |
| 5961 | |
| 5962 | // Update address for matrix B |
| 5963 | src_addr.s1 += idx * sizeof(float); |
| 5964 | |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5965 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 5966 | // Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension |
| 5967 | // in order to take into account the presence of possible cross plane paddings |
| 5968 | // |
| 5969 | // | | |
| 5970 | // | plane0 | |
| 5971 | // | | |
| 5972 | // |__________________| |
| 5973 | // |******************| |
| 5974 | // | cross_plane_pad | |
| 5975 | // |******************| |
| 5976 | // | | |
| 5977 | // | plane1 | |
| 5978 | // | | |
| 5979 | // |__________________| |
| 5980 | |
| 5981 | // The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
| 5982 | uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
| 5983 | zin = min(DEPTH_GEMM3D - 1, zin); |
| 5984 | |
| 5985 | // Add offset due to the cross plane paddings |
| 5986 | zin *= (src_cross_plane_pad * src0_stride_y); |
| 5987 | |
| 5988 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 5989 | // multiply src0_stride_z by DEPTH_GEMM3D |
| 5990 | src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D; |
| 5991 | |
| 5992 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 5993 | |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 5994 | // Add offset for batched GEMM |
| 5995 | src_addr.s0 += get_global_id(2) * src0_stride_z; |
| 5996 | |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 5997 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 5998 | |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 5999 | #if defined(MATRIX_B_DEPTH) |
| 6000 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 6001 | src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; |
| 6002 | #else // defined(MATRIX_B_DEPTH) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 6003 | src_addr.s1 += get_global_id(2) * src1_stride_z; |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 6004 | #endif // defined(MATRIX_B_DEPTH) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 6005 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6006 | // Initialize accumulators |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6007 | float4 acc0 = 0.0f; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6008 | |
| 6009 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6010 | float4 acc1 = 0.0f; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6011 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6012 | |
| 6013 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6014 | float4 acc2 = 0.0f; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6015 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6016 | |
| 6017 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6018 | float4 acc3 = 0.0f; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6019 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6020 | |
| 6021 | // A and B src indices get incremented at the same time. |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6022 | int i = 0; |
| 6023 | for(; i <= ((int)COLS_A - 4); i += 4) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6024 | { |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6025 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 6026 | // Load values from matrix A and matrix B |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 6027 | LOAD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, 4, float, a, src0_ptr, src_addr.s0, src0_stride_y, zin.s); |
| 6028 | #else // defined(REINTERPRET_INPUT_AS_3D) |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6029 | // Load values from matrix A and matrix B |
| 6030 | float4 a0 = vload4(0, (__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6031 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6032 | float4 a1 = vload4(0, (__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6033 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6034 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6035 | float4 a2 = vload4(0, (__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6036 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6037 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6038 | float4 a3 = vload4(0, (__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6039 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6040 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 6041 | |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6042 | float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 6043 | src_addr.s1 += src1_stride_y; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6044 | |
| 6045 | // Multiply and accumulate |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6046 | acc0.s0 = fma(a0.s0, b0.s0, acc0.s0); |
| 6047 | acc0.s1 = fma(a0.s0, b0.s1, acc0.s1); |
| 6048 | acc0.s2 = fma(a0.s0, b0.s2, acc0.s2); |
| 6049 | acc0.s3 = fma(a0.s0, b0.s3, acc0.s3); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6050 | |
| 6051 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6052 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6053 | acc1.s0 = fma(a1.s0, b0.s0, acc1.s0); |
| 6054 | acc1.s1 = fma(a1.s0, b0.s1, acc1.s1); |
| 6055 | acc1.s2 = fma(a1.s0, b0.s2, acc1.s2); |
| 6056 | acc1.s3 = fma(a1.s0, b0.s3, acc1.s3); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6057 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6058 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6059 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6060 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6061 | acc2.s0 = fma(a2.s0, b0.s0, acc2.s0); |
| 6062 | acc2.s1 = fma(a2.s0, b0.s1, acc2.s1); |
| 6063 | acc2.s2 = fma(a2.s0, b0.s2, acc2.s2); |
| 6064 | acc2.s3 = fma(a2.s0, b0.s3, acc2.s3); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6065 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6066 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6067 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6068 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6069 | acc3.s0 = fma(a3.s0, b0.s0, acc3.s0); |
| 6070 | acc3.s1 = fma(a3.s0, b0.s1, acc3.s1); |
| 6071 | acc3.s2 = fma(a3.s0, b0.s2, acc3.s2); |
| 6072 | acc3.s3 = fma(a3.s0, b0.s3, acc3.s3); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6073 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6074 | |
| 6075 | // Load values from matrix A and matrix B |
| 6076 | b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 6077 | src_addr.s1 += src1_stride_y; |
| 6078 | |
| 6079 | // Multiply and accumulate |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6080 | acc0.s0 = fma(a0.s1, b0.s0, acc0.s0); |
| 6081 | acc0.s1 = fma(a0.s1, b0.s1, acc0.s1); |
| 6082 | acc0.s2 = fma(a0.s1, b0.s2, acc0.s2); |
| 6083 | acc0.s3 = fma(a0.s1, b0.s3, acc0.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6084 | |
| 6085 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6086 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6087 | acc1.s0 = fma(a1.s1, b0.s0, acc1.s0); |
| 6088 | acc1.s1 = fma(a1.s1, b0.s1, acc1.s1); |
| 6089 | acc1.s2 = fma(a1.s1, b0.s2, acc1.s2); |
| 6090 | acc1.s3 = fma(a1.s1, b0.s3, acc1.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6091 | |
| 6092 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6093 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6094 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6095 | acc2.s0 = fma(a2.s1, b0.s0, acc2.s0); |
| 6096 | acc2.s1 = fma(a2.s1, b0.s1, acc2.s1); |
| 6097 | acc2.s2 = fma(a2.s1, b0.s2, acc2.s2); |
| 6098 | acc2.s3 = fma(a2.s1, b0.s3, acc2.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6099 | |
| 6100 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6101 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6102 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6103 | acc3.s0 = fma(a3.s1, b0.s0, acc3.s0); |
| 6104 | acc3.s1 = fma(a3.s1, b0.s1, acc3.s1); |
| 6105 | acc3.s2 = fma(a3.s1, b0.s2, acc3.s2); |
| 6106 | acc3.s3 = fma(a3.s1, b0.s3, acc3.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6107 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6108 | |
| 6109 | // Load values from matrix A and matrix B |
| 6110 | b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 6111 | src_addr.s1 += src1_stride_y; |
| 6112 | |
| 6113 | // Multiply and accumulate |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6114 | acc0.s0 = fma(a0.s2, b0.s0, acc0.s0); |
| 6115 | acc0.s1 = fma(a0.s2, b0.s1, acc0.s1); |
| 6116 | acc0.s2 = fma(a0.s2, b0.s2, acc0.s2); |
| 6117 | acc0.s3 = fma(a0.s2, b0.s3, acc0.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6118 | |
| 6119 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6120 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6121 | acc1.s0 = fma(a1.s2, b0.s0, acc1.s0); |
| 6122 | acc1.s1 = fma(a1.s2, b0.s1, acc1.s1); |
| 6123 | acc1.s2 = fma(a1.s2, b0.s2, acc1.s2); |
| 6124 | acc1.s3 = fma(a1.s2, b0.s3, acc1.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6125 | |
| 6126 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6127 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6128 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6129 | acc2.s0 = fma(a2.s2, b0.s0, acc2.s0); |
| 6130 | acc2.s1 = fma(a2.s2, b0.s1, acc2.s1); |
| 6131 | acc2.s2 = fma(a2.s2, b0.s2, acc2.s2); |
| 6132 | acc2.s3 = fma(a2.s2, b0.s3, acc2.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6133 | |
| 6134 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6135 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6136 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6137 | acc3.s0 = fma(a3.s2, b0.s0, acc3.s0); |
| 6138 | acc3.s1 = fma(a3.s2, b0.s1, acc3.s1); |
| 6139 | acc3.s2 = fma(a3.s2, b0.s2, acc3.s2); |
| 6140 | acc3.s3 = fma(a3.s2, b0.s3, acc3.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6141 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6142 | |
| 6143 | // Load values from matrix A and matrix B |
| 6144 | b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 6145 | src_addr.s1 += src1_stride_y; |
| 6146 | |
| 6147 | // Multiply and accumulate |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6148 | acc0.s0 = fma(a0.s3, b0.s0, acc0.s0); |
| 6149 | acc0.s1 = fma(a0.s3, b0.s1, acc0.s1); |
| 6150 | acc0.s2 = fma(a0.s3, b0.s2, acc0.s2); |
| 6151 | acc0.s3 = fma(a0.s3, b0.s3, acc0.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6152 | |
| 6153 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6154 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6155 | acc1.s0 = fma(a1.s3, b0.s0, acc1.s0); |
| 6156 | acc1.s1 = fma(a1.s3, b0.s1, acc1.s1); |
| 6157 | acc1.s2 = fma(a1.s3, b0.s2, acc1.s2); |
| 6158 | acc1.s3 = fma(a1.s3, b0.s3, acc1.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6159 | |
| 6160 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6161 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6162 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6163 | acc2.s0 = fma(a2.s3, b0.s0, acc2.s0); |
| 6164 | acc2.s1 = fma(a2.s3, b0.s1, acc2.s1); |
| 6165 | acc2.s2 = fma(a2.s3, b0.s2, acc2.s2); |
| 6166 | acc2.s3 = fma(a2.s3, b0.s3, acc2.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6167 | |
| 6168 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6169 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6170 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6171 | acc3.s0 = fma(a3.s3, b0.s0, acc3.s0); |
| 6172 | acc3.s1 = fma(a3.s3, b0.s1, acc3.s1); |
| 6173 | acc3.s2 = fma(a3.s3, b0.s2, acc3.s2); |
| 6174 | acc3.s3 = fma(a3.s3, b0.s3, acc3.s3); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6175 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6176 | |
| 6177 | src_addr.s0 += 4 * sizeof(float); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6178 | } |
| 6179 | |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6180 | for(; i < (int)COLS_A; ++i) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6181 | { |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6182 | #if defined(REINTERPRET_INPUT_AS_3D) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6183 | // Load values from matrix A |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6184 | float a0 = *((__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0)); |
| 6185 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6186 | float a1 = *((__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1)); |
| 6187 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6188 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6189 | float a2 = *((__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2)); |
| 6190 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6191 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6192 | float a3 = *((__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3)); |
| 6193 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6194 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 6195 | // Load values from matrix A |
| 6196 | float a0 = *((__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6197 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6198 | float a1 = *((__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); |
| 6199 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6200 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6201 | float a2 = *((__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); |
| 6202 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6203 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6204 | float a3 = *((__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); |
| 6205 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6206 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 6207 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6208 | // Load values from matrix B |
| 6209 | float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6210 | src_addr.s1 += src1_stride_y; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6211 | |
| 6212 | // Multiply and accumulate |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6213 | acc0.s0 = fma(a0, b0.s0, acc0.s0); |
| 6214 | acc0.s1 = fma(a0, b0.s1, acc0.s1); |
| 6215 | acc0.s2 = fma(a0, b0.s2, acc0.s2); |
| 6216 | acc0.s3 = fma(a0, b0.s3, acc0.s3); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6217 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6218 | acc1.s0 = fma(a1, b0.s0, acc1.s0); |
| 6219 | acc1.s1 = fma(a1, b0.s1, acc1.s1); |
| 6220 | acc1.s2 = fma(a1, b0.s2, acc1.s2); |
| 6221 | acc1.s3 = fma(a1, b0.s3, acc1.s3); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6222 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6223 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6224 | acc2.s0 = fma(a2, b0.s0, acc2.s0); |
| 6225 | acc2.s1 = fma(a2, b0.s1, acc2.s1); |
| 6226 | acc2.s2 = fma(a2, b0.s2, acc2.s2); |
| 6227 | acc2.s3 = fma(a2, b0.s3, acc2.s3); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6228 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6229 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6230 | acc3.s0 = fma(a3, b0.s0, acc3.s0); |
| 6231 | acc3.s1 = fma(a3, b0.s1, acc3.s1); |
| 6232 | acc3.s2 = fma(a3, b0.s2, acc3.s2); |
| 6233 | acc3.s3 = fma(a3, b0.s3, acc3.s3); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6234 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6235 | |
| 6236 | src_addr.s0 += sizeof(float); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6237 | } |
| 6238 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6239 | int z = get_global_id(2); |
| 6240 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6241 | // Compute destination address |
| 6242 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 6243 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6244 | // Compute dst address |
| 6245 | __global uchar *dst_addr = offset(&dst, 0, 0); |
| 6246 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6247 | uint4 zout = 0; |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 6248 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6249 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 6250 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 6251 | // in order to take into account the presence of possible cross plane paddings |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6252 | // |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 6253 | // | | |
| 6254 | // | plane0 | |
| 6255 | // | | |
| 6256 | // |__________________| |
| 6257 | // |******************| |
| 6258 | // | cross_plane_pad | |
| 6259 | // |******************| |
| 6260 | // | | |
| 6261 | // | plane1 | |
| 6262 | // | | |
| 6263 | // |__________________| |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6264 | |
| 6265 | // The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6266 | zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
| 6267 | zout = min(DEPTH_GEMM3D - 1, zout); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6268 | |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 6269 | // Add offset due to the cross plane paddings |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6270 | zout *= (dst_cross_plane_pad * dst_stride_y); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6271 | |
| 6272 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 6273 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 6274 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6275 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6276 | // Add offset for batched GEMM |
| 6277 | dst_addr += z * dst_stride_z; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6278 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 6279 | |
| 6280 | // Multiply by the weight of matrix-matrix product and store the result |
| 6281 | #if defined(ALPHA) |
| 6282 | SCALE_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, float, acc, ALPHA); |
| 6283 | #endif // defined(ALPHA) |
| 6284 | |
| 6285 | // Add beta*bias |
| 6286 | #if defined(BETA) |
| 6287 | REPEAT_VAR_INIT_TO_CONST(NUM_ELEMS_PROCESSED_PER_THREAD_Y, uint, zero, 0); |
| 6288 | |
| 6289 | #if defined(BROADCAST_BIAS) |
| 6290 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)); |
| 6291 | |
| 6292 | LOAD_BLOCK(1, 4, float, bias, src2_addr, 0, src2_stride_y, zero); |
| 6293 | |
| 6294 | #ifndef UNIT_BETA |
| 6295 | SCALE_BLOCK(1, float, bias, BETA); |
| 6296 | #endif // UNIT_BIAS |
| 6297 | |
| 6298 | // acc = acc + bias[broadcasted] |
| 6299 | ADD_BLOCK_BROADCAST(NUM_ELEMS_PROCESSED_PER_THREAD_Y, acc, bias0); |
| 6300 | |
| 6301 | #else // defined(BROADCAST_BIAS) |
| 6302 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)) + (get_global_id(1) * |
| 6303 | (uint)NUM_ELEMS_PROCESSED_PER_THREAD_Y * src2_stride_y) + get_global_id(2) * src2_stride_z; |
| 6304 | |
| 6305 | LOAD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, 4, float, bias, src2_addr, 0, src2_stride_y, zero); |
| 6306 | |
| 6307 | #ifndef UNIT_BETA |
| 6308 | SCALE_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, float, bias, BETA); |
| 6309 | #endif // UNIT_BIAS |
| 6310 | |
| 6311 | // acc = acc + bias |
| 6312 | ADD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, acc, bias); |
| 6313 | |
| 6314 | #endif // defined(BROADCAST_BIAS) |
| 6315 | #endif // defined(BETA) |
| 6316 | |
| 6317 | #if defined(ACTIVATION_TYPE) |
| 6318 | ACTIVATION_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, ACTIVATION_TYPE, float, acc, A_VAL, B_VAL); |
| 6319 | #endif // defined(ACTIVATION_TYPE) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6320 | |
| 6321 | // Store the output block |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6322 | vstore4(acc0, 0, (__global float *)(dst_addr + 0 * dst_stride_y + zout.s0)); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6323 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6324 | vstore4(acc1, 0, (__global float *)(dst_addr + 1 * dst_stride_y + zout.s1)); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6325 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6326 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6327 | vstore4(acc2, 0, (__global float *)(dst_addr + 2 * dst_stride_y + zout.s2)); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6328 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6329 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6330 | vstore4(acc3, 0, (__global float *)(dst_addr + 3 * dst_stride_y + zout.s3)); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6331 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6332 | } |
| 6333 | |
| 6334 | /** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not been reshaped |
| 6335 | * |
| 6336 | * @note This OpenCL kernel works with the 32-bit floating point data type (float) and uses the fma units. |
| 6337 | * This OpenCL kernel is optimized for Bifrost when the number of matrix B columns is less or equal to 1000. |
| 6338 | * @note The number of elements processed along the x and y directions must be passed at compile time using -DNUM_ELEMS_PROCESSED_PER_THREAD_X and -DNUM_ELEMS_PROCESSED_PER_THREAD_Y. |
| 6339 | * This kernel optimally uses -DNUM_ELEMS_PROCESSED_PER_THREAD_X=2. |
| 6340 | * @note The number of matrix A columns must be passed at compile time using -DCOLS_A. |
| 6341 | * @note The optional value of scalar alpha is passed at compile time using -DALPHA=alpha if alpha!=1.0f. |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6342 | * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) |
| 6343 | * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6344 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6345 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 6346 | * The activation function is performed after the bias addition |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6347 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 6348 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6349 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 6350 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 6351 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 6352 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 6353 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6354 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6355 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 6356 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 6357 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 6358 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 6359 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 6360 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
| 6361 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 6362 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 6363 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 6364 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 6365 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6366 | * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 6367 | * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 6368 | * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) |
| 6369 | * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 6370 | * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) |
| 6371 | * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6372 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
| 6373 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 6374 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 6375 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 6376 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 6377 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6378 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 6379 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6380 | * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6381 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6382 | * @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D) |
| 6383 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6384 | */ |
| 6385 | __kernel void gemm_mm_floating_point_f32_bifrost_1000(IMAGE_DECLARATION(src0), |
| 6386 | IMAGE_DECLARATION(src1), |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6387 | #if defined(BETA) |
| 6388 | IMAGE_DECLARATION(src2), |
| 6389 | #endif // defined(BETA) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 6390 | IMAGE_DECLARATION(dst), |
| 6391 | uint src0_stride_z, |
| 6392 | uint src1_stride_z, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6393 | #if defined(BETA) |
| 6394 | uint src2_stride_z, |
| 6395 | #endif //defined(BETA) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6396 | uint dst_stride_z |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6397 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 6398 | , |
| 6399 | uint src_cross_plane_pad |
| 6400 | #endif // REINTERPRET_INPUT_AS_3D |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6401 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 6402 | , |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6403 | uint dst_cross_plane_pad |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6404 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 6405 | ) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6406 | { |
| 6407 | // Requires 2 NUM_ELEMS_PROCESSED_PER_THREAD_X, C vect2, A vect4, B (2 vload2) // to fix for NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6408 | int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; |
| 6409 | |
| 6410 | // Compute starting address for matrix A and Matrix B |
| 6411 | int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); |
| 6412 | |
| 6413 | // Update address for the matrix A |
| 6414 | src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y; |
| 6415 | |
| 6416 | // Update address for the matrix B |
| 6417 | src_addr.s1 += idx * sizeof(float); |
| 6418 | |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6419 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 6420 | // Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension |
| 6421 | // in order to take into account the presence of possible cross plane paddings |
| 6422 | // |
| 6423 | // | | |
| 6424 | // | plane0 | |
| 6425 | // | | |
| 6426 | // |__________________| |
| 6427 | // |******************| |
| 6428 | // | cross_plane_pad | |
| 6429 | // |******************| |
| 6430 | // | | |
| 6431 | // | plane1 | |
| 6432 | // | | |
| 6433 | // |__________________| |
| 6434 | |
| 6435 | // The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
| 6436 | uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
| 6437 | zin = min(DEPTH_GEMM3D - 1, zin); |
| 6438 | |
| 6439 | // Add offset due to the cross plane paddings |
| 6440 | zin *= (src_cross_plane_pad * src0_stride_y); |
| 6441 | |
| 6442 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 6443 | // multiply src0_stride_z by DEPTH_GEMM3D |
| 6444 | src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D; |
| 6445 | |
| 6446 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 6447 | |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 6448 | // Add offset for batched GEMM |
| 6449 | src_addr.s0 += get_global_id(2) * src0_stride_z; |
| 6450 | |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6451 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 6452 | |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 6453 | #if defined(MATRIX_B_DEPTH) |
| 6454 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 6455 | src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; |
| 6456 | #else // defined(MATRIX_B_DEPTH) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 6457 | src_addr.s1 += get_global_id(2) * src1_stride_z; |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 6458 | #endif // defined(MATRIX_B_DEPTH) |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 6459 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6460 | // Initialize accumulators |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6461 | float2 acc0 = 0.0f; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6462 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6463 | float2 acc1 = 0.0f; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6464 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6465 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6466 | float2 acc2 = 0.0f; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6467 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6468 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6469 | float2 acc3 = 0.0f; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6470 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6471 | |
| 6472 | // A and B src indices get incremented at the same time. |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6473 | int i = 0; |
| 6474 | for(; i <= ((int)COLS_A - 8); i += 8) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6475 | { |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6476 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 6477 | // Load values from matrix A |
| 6478 | float8 a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + zin.s0)); |
| 6479 | #else // defined(REINTERPRET_INPUT_AS_3D) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6480 | // Load values from matrix A |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6481 | float8 a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0)); |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6482 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6483 | |
| 6484 | // Load values from matrix B |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6485 | float2 b0 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 6486 | src_addr.s1 += src1_stride_y; |
| 6487 | float2 b1 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 6488 | src_addr.s1 += src1_stride_y; |
| 6489 | float2 b2 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 6490 | src_addr.s1 += src1_stride_y; |
| 6491 | float2 b3 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 6492 | src_addr.s1 += src1_stride_y; |
| 6493 | float2 b4 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 6494 | src_addr.s1 += src1_stride_y; |
| 6495 | float2 b5 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 6496 | src_addr.s1 += src1_stride_y; |
| 6497 | float2 b6 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 6498 | src_addr.s1 += src1_stride_y; |
| 6499 | float2 b7 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 6500 | src_addr.s1 += src1_stride_y; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6501 | |
| 6502 | // Multiply and accumulate |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6503 | acc0.s0 = fma(a0.s0, b0.s0, acc0.s0); |
| 6504 | acc0.s0 = fma(a0.s1, b1.s0, acc0.s0); |
| 6505 | acc0.s0 = fma(a0.s2, b2.s0, acc0.s0); |
| 6506 | acc0.s0 = fma(a0.s3, b3.s0, acc0.s0); |
| 6507 | acc0.s0 = fma(a0.s4, b4.s0, acc0.s0); |
| 6508 | acc0.s0 = fma(a0.s5, b5.s0, acc0.s0); |
| 6509 | acc0.s0 = fma(a0.s6, b6.s0, acc0.s0); |
| 6510 | acc0.s0 = fma(a0.s7, b7.s0, acc0.s0); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6511 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6512 | acc0.s1 = fma(a0.s0, b0.s1, acc0.s1); |
| 6513 | acc0.s1 = fma(a0.s1, b1.s1, acc0.s1); |
| 6514 | acc0.s1 = fma(a0.s2, b2.s1, acc0.s1); |
| 6515 | acc0.s1 = fma(a0.s3, b3.s1, acc0.s1); |
| 6516 | acc0.s1 = fma(a0.s4, b4.s1, acc0.s1); |
| 6517 | acc0.s1 = fma(a0.s5, b5.s1, acc0.s1); |
| 6518 | acc0.s1 = fma(a0.s6, b6.s1, acc0.s1); |
| 6519 | acc0.s1 = fma(a0.s7, b7.s1, acc0.s1); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6520 | |
| 6521 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6522 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 6523 | a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1)); |
| 6524 | #else // defined(REINTERPRET_INPUT_AS_3D) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6525 | a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6526 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6527 | acc1.s0 = fma(a0.s0, b0.s0, acc1.s0); |
| 6528 | acc1.s0 = fma(a0.s1, b1.s0, acc1.s0); |
| 6529 | acc1.s0 = fma(a0.s2, b2.s0, acc1.s0); |
| 6530 | acc1.s0 = fma(a0.s3, b3.s0, acc1.s0); |
| 6531 | acc1.s0 = fma(a0.s4, b4.s0, acc1.s0); |
| 6532 | acc1.s0 = fma(a0.s5, b5.s0, acc1.s0); |
| 6533 | acc1.s0 = fma(a0.s6, b6.s0, acc1.s0); |
| 6534 | acc1.s0 = fma(a0.s7, b7.s0, acc1.s0); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6535 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6536 | acc1.s1 = fma(a0.s0, b0.s1, acc1.s1); |
| 6537 | acc1.s1 = fma(a0.s1, b1.s1, acc1.s1); |
| 6538 | acc1.s1 = fma(a0.s2, b2.s1, acc1.s1); |
| 6539 | acc1.s1 = fma(a0.s3, b3.s1, acc1.s1); |
| 6540 | acc1.s1 = fma(a0.s4, b4.s1, acc1.s1); |
| 6541 | acc1.s1 = fma(a0.s5, b5.s1, acc1.s1); |
| 6542 | acc1.s1 = fma(a0.s6, b6.s1, acc1.s1); |
| 6543 | acc1.s1 = fma(a0.s7, b7.s1, acc1.s1); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6544 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6545 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6546 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 6547 | a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2)); |
| 6548 | #else // defined(REINTERPRET_INPUT_AS_3D) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6549 | a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6550 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6551 | acc2.s0 = fma(a0.s0, b0.s0, acc2.s0); |
| 6552 | acc2.s0 = fma(a0.s1, b1.s0, acc2.s0); |
| 6553 | acc2.s0 = fma(a0.s2, b2.s0, acc2.s0); |
| 6554 | acc2.s0 = fma(a0.s3, b3.s0, acc2.s0); |
| 6555 | acc2.s0 = fma(a0.s4, b4.s0, acc2.s0); |
| 6556 | acc2.s0 = fma(a0.s5, b5.s0, acc2.s0); |
| 6557 | acc2.s0 = fma(a0.s6, b6.s0, acc2.s0); |
| 6558 | acc2.s0 = fma(a0.s7, b7.s0, acc2.s0); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6559 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6560 | acc2.s1 = fma(a0.s0, b0.s1, acc2.s1); |
| 6561 | acc2.s1 = fma(a0.s1, b1.s1, acc2.s1); |
| 6562 | acc2.s1 = fma(a0.s2, b2.s1, acc2.s1); |
| 6563 | acc2.s1 = fma(a0.s3, b3.s1, acc2.s1); |
| 6564 | acc2.s1 = fma(a0.s4, b4.s1, acc2.s1); |
| 6565 | acc2.s1 = fma(a0.s5, b5.s1, acc2.s1); |
| 6566 | acc2.s1 = fma(a0.s6, b6.s1, acc2.s1); |
| 6567 | acc2.s1 = fma(a0.s7, b7.s1, acc2.s1); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6568 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6569 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6570 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 6571 | a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3)); |
| 6572 | #else // defined(REINTERPRET_INPUT_AS_3D) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6573 | a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6574 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6575 | acc3.s0 = fma(a0.s0, b0.s0, acc3.s0); |
| 6576 | acc3.s0 = fma(a0.s1, b1.s0, acc3.s0); |
| 6577 | acc3.s0 = fma(a0.s2, b2.s0, acc3.s0); |
| 6578 | acc3.s0 = fma(a0.s3, b3.s0, acc3.s0); |
| 6579 | acc3.s0 = fma(a0.s4, b4.s0, acc3.s0); |
| 6580 | acc3.s0 = fma(a0.s5, b5.s0, acc3.s0); |
| 6581 | acc3.s0 = fma(a0.s6, b6.s0, acc3.s0); |
| 6582 | acc3.s0 = fma(a0.s7, b7.s0, acc3.s0); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6583 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6584 | acc3.s1 = fma(a0.s0, b0.s1, acc3.s1); |
| 6585 | acc3.s1 = fma(a0.s1, b1.s1, acc3.s1); |
| 6586 | acc3.s1 = fma(a0.s2, b2.s1, acc3.s1); |
| 6587 | acc3.s1 = fma(a0.s3, b3.s1, acc3.s1); |
| 6588 | acc3.s1 = fma(a0.s4, b4.s1, acc3.s1); |
| 6589 | acc3.s1 = fma(a0.s5, b5.s1, acc3.s1); |
| 6590 | acc3.s1 = fma(a0.s6, b6.s1, acc3.s1); |
| 6591 | acc3.s1 = fma(a0.s7, b7.s1, acc3.s1); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6592 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6593 | |
| 6594 | src_addr.s0 += sizeof(float) * 8; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6595 | } |
| 6596 | // float size increment |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6597 | for(; i < (int)COLS_A; ++i) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6598 | { |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6599 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 6600 | // Load values from matrix A |
| 6601 | float a0 = *((__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0)); |
| 6602 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6603 | float a1 = *((__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1)); |
| 6604 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6605 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6606 | float a2 = *((__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2)); |
| 6607 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6608 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6609 | float a3 = *((__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3)); |
| 6610 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6611 | #else // defined(REINTERPRET_INPUT_AS_3D) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6612 | // Load values from matrix A |
| 6613 | float a0 = *((__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); |
| 6614 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6615 | float a1 = *((__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); |
| 6616 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6617 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6618 | float a2 = *((__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); |
| 6619 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6620 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6621 | float a3 = *((__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); |
| 6622 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6623 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 6624 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6625 | // Load values from matrix B |
| 6626 | float2 b0 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6627 | src_addr.s1 += src1_stride_y; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6628 | |
| 6629 | // Multiply and accumulate |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6630 | acc0.s0 = fma(a0, b0.s0, acc0.s0); |
| 6631 | acc0.s1 = fma(a0, b0.s1, acc0.s1); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6632 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6633 | acc1.s0 = fma(a1, b0.s0, acc1.s0); |
| 6634 | acc1.s1 = fma(a1, b0.s1, acc1.s1); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6635 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6636 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6637 | acc2.s0 = fma(a2, b0.s0, acc2.s0); |
| 6638 | acc2.s1 = fma(a2, b0.s1, acc2.s1); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6639 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6640 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6641 | acc3.s0 = fma(a3, b0.s0, acc3.s0); |
| 6642 | acc3.s1 = fma(a3, b0.s1, acc3.s1); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6643 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | c9c62c2 | 2018-04-06 10:00:10 +0100 | [diff] [blame] | 6644 | |
| 6645 | src_addr.s0 += sizeof(float); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6646 | } |
| 6647 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6648 | int z = get_global_id(2); |
| 6649 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6650 | // Compute destination address |
| 6651 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 6652 | |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 6653 | // Compute dst address |
| 6654 | __global uchar *dst_addr = offset(&dst, 0, 0); |
| 6655 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6656 | uint4 zout = 0; |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 6657 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6658 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6659 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6660 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 6661 | // in order to take into account the presence of possible cross plane paddings |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6662 | // |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 6663 | // | | |
| 6664 | // | plane0 | |
| 6665 | // | | |
| 6666 | // |__________________| |
| 6667 | // |******************| |
| 6668 | // | cross_plane_pad | |
| 6669 | // |******************| |
| 6670 | // | | |
| 6671 | // | plane1 | |
| 6672 | // | | |
| 6673 | // |__________________| |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 6674 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6675 | // The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6676 | zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
| 6677 | zout = min(DEPTH_GEMM3D - 1, zout); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6678 | |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 6679 | // Add offset due to the cross plane paddings |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 6680 | zout *= (dst_cross_plane_pad * dst_stride_y); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6681 | |
| 6682 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 6683 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 6684 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6685 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6686 | // Add offset for batched GEMM |
| 6687 | dst_addr += z * dst_stride_z; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6688 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 6689 | |
| 6690 | // Multiply by the weight of matrix-matrix product and store the result |
| 6691 | #if defined(ALPHA) |
| 6692 | SCALE_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, float, acc, ALPHA); |
| 6693 | #endif // defined(ALPHA) |
| 6694 | |
| 6695 | // Add beta*bias |
| 6696 | #if defined(BETA) |
| 6697 | REPEAT_VAR_INIT_TO_CONST(NUM_ELEMS_PROCESSED_PER_THREAD_Y, uint, zero, 0); |
| 6698 | |
| 6699 | #if defined(BROADCAST_BIAS) |
| 6700 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)2 * sizeof(float)); |
| 6701 | |
| 6702 | LOAD_BLOCK(1, 2, float, bias, src2_addr, 0, src2_stride_y, zero); |
| 6703 | |
| 6704 | #ifndef UNIT_BETA |
| 6705 | SCALE_BLOCK(1, float, bias, BETA); |
| 6706 | #endif // UNIT_BIAS |
| 6707 | |
| 6708 | // acc = acc + bias[broadcasted] |
| 6709 | ADD_BLOCK_BROADCAST(NUM_ELEMS_PROCESSED_PER_THREAD_Y, acc, bias0); |
| 6710 | |
| 6711 | #else // defined(BROADCAST_BIAS) |
| 6712 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)2 * sizeof(float)) + (get_global_id(1) * |
| 6713 | (uint)NUM_ELEMS_PROCESSED_PER_THREAD_Y * src2_stride_y) + get_global_id(2) * src2_stride_z; |
| 6714 | |
| 6715 | LOAD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, 2, float, bias, src2_addr, 0, src2_stride_y, zero); |
| 6716 | |
| 6717 | #ifndef UNIT_BETA |
| 6718 | SCALE_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, float, bias, BETA); |
| 6719 | #endif // UNIT_BIAS |
| 6720 | |
| 6721 | // acc = acc + bias |
| 6722 | ADD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, acc, bias); |
| 6723 | |
| 6724 | #endif // defined(BROADCAST_BIAS) |
| 6725 | #endif // defined(BETA) |
| 6726 | |
| 6727 | #if defined(ACTIVATION_TYPE) |
| 6728 | ACTIVATION_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, ACTIVATION_TYPE, float, acc, A_VAL, B_VAL); |
| 6729 | #endif // defined(ACTIVATION_TYPE) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6730 | |
| 6731 | // Store the output block |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6732 | vstore2(acc0, 0, (__global float *)(dst_addr + 0 * dst_stride_y + zout.s0)); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6733 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6734 | vstore2(acc1, 0, (__global float *)(dst_addr + 1 * dst_stride_y + zout.s1)); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6735 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6736 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6737 | vstore2(acc2, 0, (__global float *)(dst_addr + 2 * dst_stride_y + zout.s2)); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6738 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6739 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6740 | vstore2(acc3, 0, (__global float *)(dst_addr + 3 * dst_stride_y + zout.s3)); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 6741 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 6742 | } |
| 6743 | |
Vidhya Sudhan Loganathan | bdff491 | 2018-05-22 15:03:09 +0100 | [diff] [blame] | 6744 | #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 6745 | /** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped |
| 6746 | * |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 6747 | * @note This OpenCL kernel works with the 16-bit floating point data type (half) and accumulating the result in a 32 floating point variable. |
| 6748 | * @note The number of elements processed along the x and y directions must be passed at compile time using -DNUM_ELEMS_PROCESSED_PER_THREAD_X and -DNUM_ELEMS_PROCESSED_PER_THREAD_Y. |
| 6749 | * This kernel optimally uses -DNUM_ELEMS_PROCESSED_PER_THREAD_X=4. |
| 6750 | * @note The number of matrix A columns must be passed at compile time using -DCOLS_A. |
| 6751 | * @note The optional value of scalar alpha is passed at compile time using -DALPHA=alpha |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6752 | * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) |
| 6753 | * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 6754 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6755 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 6756 | * The activation function is performed after the bias addition |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 6757 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 6758 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
| 6759 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 6760 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 6761 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 6762 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 6763 | * |
| 6764 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16 |
| 6765 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 6766 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 6767 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 6768 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 6769 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 6770 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
| 6771 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 6772 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 6773 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 6774 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 6775 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6776 | * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 6777 | * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 6778 | * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) |
| 6779 | * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 6780 | * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) |
| 6781 | * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 6782 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
| 6783 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 6784 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 6785 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 6786 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 6787 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 6788 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 6789 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6790 | * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 6791 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 6792 | * @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D) |
| 6793 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
| 6794 | */ |
| 6795 | __kernel void gemm_mm_floating_point_f16_bifrost_acc32(IMAGE_DECLARATION(src0), |
| 6796 | IMAGE_DECLARATION(src1), |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6797 | #if defined(BETA) |
| 6798 | IMAGE_DECLARATION(src2), |
| 6799 | #endif // defined(BETA) |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 6800 | IMAGE_DECLARATION(dst), |
| 6801 | uint src0_stride_z, |
| 6802 | uint src1_stride_z, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 6803 | #if defined(BETA) |
| 6804 | uint src2_stride_z, |
| 6805 | #endif //defined(BETA) |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 6806 | uint dst_stride_z |
| 6807 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 6808 | , |
| 6809 | uint src_cross_plane_pad |
| 6810 | #endif // REINTERPRET_INPUT_AS_3D |
| 6811 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 6812 | , |
| 6813 | uint dst_cross_plane_pad |
| 6814 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 6815 | ) |
| 6816 | { |
| 6817 | int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; |
| 6818 | |
| 6819 | // Compute starting address for matrix A and Matrix B |
| 6820 | int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); |
| 6821 | |
| 6822 | // Update address for the matrix A |
| 6823 | src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y; |
| 6824 | |
| 6825 | // Update address for the matrix B |
| 6826 | src_addr.s1 += idx * sizeof(half); |
| 6827 | |
| 6828 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 6829 | // Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension |
| 6830 | // in order to take into account the presence of possible cross plane paddings |
| 6831 | // |
| 6832 | // | | |
| 6833 | // | plane0 | |
| 6834 | // | | |
| 6835 | // |__________________| |
| 6836 | // |******************| |
| 6837 | // | cross_plane_pad | |
| 6838 | // |******************| |
| 6839 | // | | |
| 6840 | // | plane1 | |
| 6841 | // | | |
| 6842 | // |__________________| |
| 6843 | |
| 6844 | // The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
| 6845 | uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
| 6846 | zin = min(DEPTH_GEMM3D - 1, zin); |
| 6847 | |
| 6848 | // Add offset due to the cross plane paddings |
| 6849 | zin *= (src_cross_plane_pad * src0_stride_y); |
| 6850 | |
| 6851 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 6852 | // multiply src0_stride_z by DEPTH_GEMM3D |
| 6853 | src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D; |
| 6854 | |
| 6855 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 6856 | |
| 6857 | // Add offset for batched GEMM |
| 6858 | src_addr.s0 += get_global_id(2) * src0_stride_z; |
| 6859 | |
| 6860 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 6861 | |
| 6862 | #if defined(MATRIX_B_DEPTH) |
| 6863 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 6864 | src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; |
| 6865 | #else // defined(MATRIX_B_DEPTH) |
| 6866 | src_addr.s1 += get_global_id(2) * src1_stride_z; |
| 6867 | #endif // defined(MATRIX_B_DEPTH) |
| 6868 | |
| 6869 | float8 acc0 = 0.0h; |
| 6870 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6871 | float8 acc1 = 0.0h; |
| 6872 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6873 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6874 | float8 acc2 = 0.0h; |
| 6875 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6876 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6877 | float8 acc3 = 0.0h; |
| 6878 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6879 | |
| 6880 | int i = 0; |
| 6881 | for(; i <= ((int)COLS_A - 4); i += 4) |
| 6882 | { |
| 6883 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 6884 | // Load values from matrix A |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 6885 | LOAD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, 4, half, a, src0_ptr, src_addr.s0, src0_stride_y, zin.s); |
| 6886 | #else // defined(REINTERPRET_INPUT_AS_3D) |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 6887 | // Load values from matrix A |
| 6888 | half4 a0 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); |
| 6889 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6890 | half4 a1 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); |
| 6891 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6892 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6893 | half4 a2 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); |
| 6894 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6895 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6896 | half4 a3 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); |
| 6897 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6898 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 6899 | |
| 6900 | // Load values from matrix B |
| 6901 | float8 b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1))); |
| 6902 | src_addr.s1 += src1_stride_y; |
| 6903 | |
| 6904 | // Accumulate |
| 6905 | acc0 = fma(b0, (float8)a0.s0, acc0); |
| 6906 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6907 | acc1 = fma(b0, (float8)a1.s0, acc1); |
| 6908 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6909 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6910 | acc2 = fma(b0, (float8)a2.s0, acc2); |
| 6911 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6912 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6913 | acc3 = fma(b0, (float8)a3.s0, acc3); |
| 6914 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6915 | |
| 6916 | b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1))); |
| 6917 | src_addr.s1 += src1_stride_y; |
| 6918 | acc0 = fma(b0, (float8)a0.s1, acc0); |
| 6919 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6920 | acc1 = fma(b0, (float8)a1.s1, acc1); |
| 6921 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6922 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6923 | acc2 = fma(b0, (float8)a2.s1, acc2); |
| 6924 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6925 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6926 | acc3 = fma(b0, (float8)a3.s1, acc3); |
| 6927 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6928 | |
| 6929 | b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1))); |
| 6930 | src_addr.s1 += src1_stride_y; |
| 6931 | acc0 = fma(b0, (float8)a0.s2, acc0); |
| 6932 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6933 | acc1 = fma(b0, (float8)a1.s2, acc1); |
| 6934 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6935 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6936 | acc2 = fma(b0, (float8)a2.s2, acc2); |
| 6937 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6938 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6939 | acc3 = fma(b0, (float8)a3.s2, acc3); |
| 6940 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6941 | |
| 6942 | b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1))); |
| 6943 | src_addr.s1 += src1_stride_y; |
| 6944 | acc0 = fma(b0, (float8)a0.s3, acc0); |
| 6945 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6946 | acc1 = fma(b0, (float8)a1.s3, acc1); |
| 6947 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6948 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6949 | acc2 = fma(b0, (float8)a2.s3, acc2); |
| 6950 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6951 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6952 | acc3 = fma(b0, (float8)a3.s3, acc3); |
| 6953 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6954 | |
| 6955 | src_addr.s0 += 4 * sizeof(half); |
| 6956 | } |
| 6957 | |
| 6958 | for(; i < (int)COLS_A; ++i) |
| 6959 | { |
| 6960 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 6961 | // Load values from matrix A |
| 6962 | half a0 = *((__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0)); |
| 6963 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6964 | half a1 = *((__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1)); |
| 6965 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6966 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6967 | half a2 = *((__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2)); |
| 6968 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6969 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6970 | half a3 = *((__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3)); |
| 6971 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6972 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 6973 | // Load values from matrix A |
| 6974 | half a0 = *((__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); |
| 6975 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6976 | half a1 = *((__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); |
| 6977 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6978 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6979 | half a2 = *((__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); |
| 6980 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6981 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6982 | half a3 = *((__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); |
| 6983 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 6984 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 6985 | |
| 6986 | // Load values from matrix B |
| 6987 | float8 b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1))); |
| 6988 | |
| 6989 | src_addr += (int2)(sizeof(half), src1_stride_y); |
| 6990 | |
| 6991 | // Accumulate |
| 6992 | acc0 = fma(b0, (float8)a0, acc0); // b0 * (half8)a0; |
| 6993 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6994 | acc1 = fma(b0, (float8)a1, acc1); // b0 * (half8)a1; |
| 6995 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 6996 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6997 | acc2 = fma(b0, (float8)a2, acc2); // b0 * (half8)a2; |
| 6998 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 6999 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7000 | acc3 = fma(b0, (float8)a3, acc3); // b0 * (half8)a3; |
| 7001 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7002 | } |
| 7003 | |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 7004 | int z = get_global_id(2); |
| 7005 | |
| 7006 | // Compute destination address |
| 7007 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 7008 | |
| 7009 | // Compute dst address |
| 7010 | __global uchar *dst_addr = offset(&dst, 0, 0); |
| 7011 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7012 | uint4 zout = 0; |
| 7013 | |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 7014 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7015 | |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 7016 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
| 7017 | // in order to take into account the presence of possible cross plane paddings |
| 7018 | // |
| 7019 | // | | |
| 7020 | // | plane0 | |
| 7021 | // | | |
| 7022 | // |__________________| |
| 7023 | // |******************| |
| 7024 | // | cross_plane_pad | |
| 7025 | // |******************| |
| 7026 | // | | |
| 7027 | // | plane1 | |
| 7028 | // | | |
| 7029 | // |__________________| |
| 7030 | |
| 7031 | // The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7032 | zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
| 7033 | zout = min(DEPTH_GEMM3D - 1, zout); |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 7034 | |
| 7035 | // Add offset due to the cross plane paddings |
| 7036 | zout *= (dst_cross_plane_pad * dst_stride_y); |
| 7037 | |
| 7038 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 7039 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 7040 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7041 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 7042 | // Add offset for batched GEMM |
| 7043 | dst_addr += z * dst_stride_z; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7044 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 7045 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7046 | // Multiply by the weight of matrix-matrix product and store the result |
| 7047 | #if defined(ALPHA) |
| 7048 | SCALE_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, float, acc, ALPHA); |
| 7049 | #endif // defined(ALPHA) |
| 7050 | |
| 7051 | #if defined(BETA) |
| 7052 | REPEAT_VAR_INIT_TO_CONST(NUM_ELEMS_PROCESSED_PER_THREAD_Y, uint, zero, 0); |
| 7053 | |
| 7054 | #if defined(BROADCAST_BIAS) |
| 7055 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)); |
| 7056 | |
| 7057 | LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero); |
| 7058 | |
| 7059 | float8 bias_f0 = convert_float8(bias0); |
| 7060 | |
| 7061 | #ifndef UNIT_BETA |
| 7062 | SCALE_BLOCK(1, float, bias_f, BETA); |
| 7063 | #endif // UNIT_BIAS |
| 7064 | |
| 7065 | // acc = acc + bias[broadcasted] |
| 7066 | ADD_BLOCK_BROADCAST(NUM_ELEMS_PROCESSED_PER_THREAD_Y, acc, bias_f0); |
| 7067 | |
| 7068 | #else // defined(BROADCAST_BIAS) |
| 7069 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (get_global_id(1) * |
| 7070 | (uint)NUM_ELEMS_PROCESSED_PER_THREAD_Y * src2_stride_y) + get_global_id(2) * src2_stride_z; |
| 7071 | |
| 7072 | LOAD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, 8, half, bias, src2_addr, 0, src2_stride_y, zero); |
| 7073 | |
| 7074 | float8 bias_f0 = convert_float8(bias0); |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 7075 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7076 | float8 bias_f1 = convert_float8(bias1); |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 7077 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7078 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7079 | float8 bias_f2 = convert_float8(bias2); |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 7080 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7081 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7082 | float8 bias_f3 = convert_float8(bias3); |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 7083 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7084 | |
| 7085 | #ifndef UNIT_BETA |
| 7086 | SCALE_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, float, bias_f, BETA); |
| 7087 | #endif // UNIT_BIAS |
| 7088 | |
| 7089 | // acc = acc + bias |
| 7090 | ADD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, acc, bias_f); |
| 7091 | |
| 7092 | #endif // defined(BROADCAST_BIAS) |
| 7093 | #endif // defined(BETA) |
| 7094 | |
| 7095 | half8 acc_h0 = convert_half8(acc0); |
| 7096 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7097 | half8 acc_h1 = convert_half8(acc1); |
| 7098 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7099 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7100 | half8 acc_h2 = convert_half8(acc2); |
| 7101 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7102 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7103 | half8 acc_h3 = convert_half8(acc3); |
| 7104 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7105 | |
| 7106 | #if defined(ACTIVATION_TYPE) |
| 7107 | ACTIVATION_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, ACTIVATION_TYPE, half, acc_h, A_VAL, B_VAL); |
| 7108 | #endif // defined(ACTIVATION_TYPE) |
| 7109 | |
| 7110 | // Store the output block |
| 7111 | STORE_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, 8, half, acc_h, dst_addr, dst_stride_y, zout.s); |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 7112 | } |
| 7113 | |
| 7114 | /** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped |
| 7115 | * |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7116 | * @note This OpenCL kernel works with the 16-bit floating point data type (half) and uses the fma units. |
| 7117 | * @note The number of elements processed along the x and y directions must be passed at compile time using -DNUM_ELEMS_PROCESSED_PER_THREAD_X and -DNUM_ELEMS_PROCESSED_PER_THREAD_Y. |
| 7118 | * This kernel optimally uses -DNUM_ELEMS_PROCESSED_PER_THREAD_X=4. |
| 7119 | * @note The number of matrix A columns must be passed at compile time using -DCOLS_A. |
| 7120 | * @note The optional value of scalar alpha is passed at compile time using -DALPHA=alpha |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7121 | * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) |
| 7122 | * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7123 | * |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7124 | * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. |
| 7125 | * The activation function is performed after the bias addition |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 7126 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 7127 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7128 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 7129 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 7130 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 7131 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 7132 | * |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7133 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16 |
| 7134 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 7135 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 7136 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 7137 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 7138 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 7139 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
| 7140 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 7141 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 7142 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 7143 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 7144 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7145 | * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr |
| 7146 | * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) |
| 7147 | * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) |
| 7148 | * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) |
| 7149 | * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) |
| 7150 | * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7151 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
| 7152 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 7153 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 7154 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 7155 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 7156 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7157 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 7158 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7159 | * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7160 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 7161 | * @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D) |
| 7162 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7163 | */ |
| 7164 | __kernel void gemm_mm_floating_point_f16_bifrost(IMAGE_DECLARATION(src0), |
| 7165 | IMAGE_DECLARATION(src1), |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7166 | #if defined(BETA) |
| 7167 | IMAGE_DECLARATION(src2), |
| 7168 | #endif // defined(BETA) |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7169 | IMAGE_DECLARATION(dst), |
| 7170 | uint src0_stride_z, |
| 7171 | uint src1_stride_z, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7172 | #if defined(BETA) |
| 7173 | uint src2_stride_z, |
| 7174 | #endif //defined(BETA) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7175 | uint dst_stride_z |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 7176 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 7177 | , |
| 7178 | uint src_cross_plane_pad |
| 7179 | #endif // REINTERPRET_INPUT_AS_3D |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7180 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 7181 | , |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 7182 | uint dst_cross_plane_pad |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7183 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 7184 | ) |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7185 | { |
| 7186 | int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; |
| 7187 | |
| 7188 | // Compute starting address for matrix A and Matrix B |
| 7189 | int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); |
| 7190 | |
| 7191 | // Update address for the matrix A |
| 7192 | src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y; |
| 7193 | |
| 7194 | // Update address for the matrix B |
| 7195 | src_addr.s1 += idx * sizeof(half); |
| 7196 | |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 7197 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 7198 | // Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension |
| 7199 | // in order to take into account the presence of possible cross plane paddings |
| 7200 | // |
| 7201 | // | | |
| 7202 | // | plane0 | |
| 7203 | // | | |
| 7204 | // |__________________| |
| 7205 | // |******************| |
| 7206 | // | cross_plane_pad | |
| 7207 | // |******************| |
| 7208 | // | | |
| 7209 | // | plane1 | |
| 7210 | // | | |
| 7211 | // |__________________| |
| 7212 | |
| 7213 | // The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
| 7214 | uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
| 7215 | zin = min(DEPTH_GEMM3D - 1, zin); |
| 7216 | |
| 7217 | // Add offset due to the cross plane paddings |
| 7218 | zin *= (src_cross_plane_pad * src0_stride_y); |
| 7219 | |
| 7220 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 7221 | // multiply src0_stride_z by DEPTH_GEMM3D |
| 7222 | src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D; |
| 7223 | |
| 7224 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 7225 | |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7226 | // Add offset for batched GEMM |
| 7227 | src_addr.s0 += get_global_id(2) * src0_stride_z; |
| 7228 | |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 7229 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 7230 | |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7231 | #if defined(MATRIX_B_DEPTH) |
| 7232 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 7233 | src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; |
| 7234 | #else // defined(MATRIX_B_DEPTH) |
| 7235 | src_addr.s1 += get_global_id(2) * src1_stride_z; |
| 7236 | #endif // defined(MATRIX_B_DEPTH) |
| 7237 | |
| 7238 | half8 acc0 = 0.0h; |
| 7239 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7240 | half8 acc1 = 0.0h; |
| 7241 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7242 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7243 | half8 acc2 = 0.0h; |
| 7244 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7245 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7246 | half8 acc3 = 0.0h; |
| 7247 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7248 | |
| 7249 | int i = 0; |
| 7250 | for(; i <= ((int)COLS_A - 4); i += 4) |
| 7251 | { |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 7252 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 7253 | // Load values from matrix A |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 7254 | LOAD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, 4, half, a, src0_ptr, src_addr.s0, src0_stride_y, zin.s); |
| 7255 | #else // defined(REINTERPRET_INPUT_AS_3D) |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7256 | // Load values from matrix A |
| 7257 | half4 a0 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); |
| 7258 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7259 | half4 a1 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); |
| 7260 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7261 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7262 | half4 a2 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); |
| 7263 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7264 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7265 | half4 a3 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); |
| 7266 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 7267 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 7268 | |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7269 | // Load values from matrix B |
| 7270 | half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1)); |
| 7271 | src_addr.s1 += src1_stride_y; |
| 7272 | |
| 7273 | // Accumulate |
| 7274 | acc0 = fma(b0, (half8)a0.s0, acc0); |
| 7275 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7276 | acc1 = fma(b0, (half8)a1.s0, acc1); |
| 7277 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7278 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7279 | acc2 = fma(b0, (half8)a2.s0, acc2); |
| 7280 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7281 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7282 | acc3 = fma(b0, (half8)a3.s0, acc3); |
| 7283 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7284 | |
| 7285 | b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1)); |
| 7286 | src_addr.s1 += src1_stride_y; |
| 7287 | acc0 = fma(b0, (half8)a0.s1, acc0); |
| 7288 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7289 | acc1 = fma(b0, (half8)a1.s1, acc1); |
| 7290 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7291 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7292 | acc2 = fma(b0, (half8)a2.s1, acc2); |
| 7293 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7294 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7295 | acc3 = fma(b0, (half8)a3.s1, acc3); |
| 7296 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7297 | |
| 7298 | b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1)); |
| 7299 | src_addr.s1 += src1_stride_y; |
| 7300 | acc0 = fma(b0, (half8)a0.s2, acc0); |
| 7301 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7302 | acc1 = fma(b0, (half8)a1.s2, acc1); |
| 7303 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7304 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7305 | acc2 = fma(b0, (half8)a2.s2, acc2); |
| 7306 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7307 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7308 | acc3 = fma(b0, (half8)a3.s2, acc3); |
| 7309 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7310 | |
| 7311 | b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1)); |
| 7312 | src_addr.s1 += src1_stride_y; |
| 7313 | acc0 = fma(b0, (half8)a0.s3, acc0); |
| 7314 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7315 | acc1 = fma(b0, (half8)a1.s3, acc1); |
| 7316 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7317 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7318 | acc2 = fma(b0, (half8)a2.s3, acc2); |
| 7319 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7320 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7321 | acc3 = fma(b0, (half8)a3.s3, acc3); |
| 7322 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7323 | |
| 7324 | src_addr.s0 += 4 * sizeof(half); |
| 7325 | } |
| 7326 | |
| 7327 | for(; i < (int)COLS_A; ++i) |
| 7328 | { |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 7329 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 7330 | // Load values from matrix A |
| 7331 | half a0 = *((__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0)); |
| 7332 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7333 | half a1 = *((__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1)); |
| 7334 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7335 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7336 | half a2 = *((__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2)); |
| 7337 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7338 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7339 | half a3 = *((__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3)); |
| 7340 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7341 | #else // defined(REINTERPRET_INPUT_AS_3D) |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7342 | // Load values from matrix A |
| 7343 | half a0 = *((__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); |
| 7344 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7345 | half a1 = *((__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); |
| 7346 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7347 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7348 | half a2 = *((__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); |
| 7349 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7350 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7351 | half a3 = *((__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); |
| 7352 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 7353 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 7354 | |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7355 | // Load values from matrix B |
| 7356 | half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1)); |
| 7357 | |
| 7358 | src_addr += (int2)(sizeof(half), src1_stride_y); |
| 7359 | |
| 7360 | // Accumulate |
| 7361 | acc0 = fma(b0, (half8)a0, acc0); // b0 * (half8)a0; |
| 7362 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7363 | acc1 = fma(b0, (half8)a1, acc1); // b0 * (half8)a1; |
| 7364 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 7365 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7366 | acc2 = fma(b0, (half8)a2, acc2); // b0 * (half8)a2; |
| 7367 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 7368 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7369 | acc3 = fma(b0, (half8)a3, acc3); // b0 * (half8)a3; |
| 7370 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 7371 | } |
| 7372 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7373 | int z = get_global_id(2); |
| 7374 | |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7375 | // Compute destination address |
| 7376 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 7377 | |
| 7378 | // Compute dst address |
| 7379 | __global uchar *dst_addr = offset(&dst, 0, 0); |
| 7380 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7381 | uint4 zout = 0; |
| 7382 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7383 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7384 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7385 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 7386 | // in order to take into account the presence of possible cross plane paddings |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7387 | // |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 7388 | // | | |
| 7389 | // | plane0 | |
| 7390 | // | | |
| 7391 | // |__________________| |
| 7392 | // |******************| |
| 7393 | // | cross_plane_pad | |
| 7394 | // |******************| |
| 7395 | // | | |
| 7396 | // | plane1 | |
| 7397 | // | | |
| 7398 | // |__________________| |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7399 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7400 | // The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7401 | zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
| 7402 | zout = min(DEPTH_GEMM3D - 1, zout); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7403 | |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 7404 | // Add offset due to the cross plane paddings |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 7405 | zout *= (dst_cross_plane_pad * dst_stride_y); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7406 | |
| 7407 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 7408 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 7409 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 7410 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 7411 | // Add offset for batched GEMM |
| 7412 | dst_addr += z * dst_stride_z; |
| 7413 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 7414 | |
| 7415 | // Multiply by the weight of matrix-matrix product and store the result |
| 7416 | #if defined(ALPHA) |
| 7417 | SCALE_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, half, acc, ALPHA); |
| 7418 | #endif // defined(ALPHA) |
| 7419 | |
| 7420 | // Add beta*bias |
| 7421 | #if defined(BETA) |
| 7422 | REPEAT_VAR_INIT_TO_CONST(NUM_ELEMS_PROCESSED_PER_THREAD_Y, uint, zero, 0); |
| 7423 | |
| 7424 | #if defined(BROADCAST_BIAS) |
| 7425 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)); |
| 7426 | |
| 7427 | LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero); |
| 7428 | |
| 7429 | #ifndef UNIT_BETA |
| 7430 | SCALE_BLOCK(1, half, bias, BETA); |
| 7431 | #endif // UNIT_BIAS |
| 7432 | |
| 7433 | // acc = acc + bias[broadcasted] |
| 7434 | ADD_BLOCK_BROADCAST(NUM_ELEMS_PROCESSED_PER_THREAD_Y, acc, bias0); |
| 7435 | |
| 7436 | #else // defined(BROADCAST_BIAS) |
| 7437 | __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (get_global_id(1) * |
| 7438 | (uint)NUM_ELEMS_PROCESSED_PER_THREAD_Y * src2_stride_y) + get_global_id(2) * src2_stride_z; |
| 7439 | |
| 7440 | LOAD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, 8, half, bias, src2_addr, 0, src2_stride_y, zero); |
| 7441 | |
| 7442 | #ifndef UNIT_BETA |
| 7443 | SCALE_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, half, bias, BETA); |
| 7444 | #endif // UNIT_BIAS |
| 7445 | |
| 7446 | // acc = acc + bias |
| 7447 | ADD_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, acc, bias); |
| 7448 | |
| 7449 | #endif // defined(BROADCAST_BIAS) |
| 7450 | #endif // defined(BETA) |
| 7451 | |
| 7452 | #if defined(ACTIVATION_TYPE) |
| 7453 | ACTIVATION_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, ACTIVATION_TYPE, half, acc, A_VAL, B_VAL); |
| 7454 | #endif // defined(ACTIVATION_TYPE) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7455 | |
| 7456 | // Store the output block |
Usama Arif | 0681e3b | 2019-04-25 14:28:07 +0100 | [diff] [blame] | 7457 | STORE_BLOCK(NUM_ELEMS_PROCESSED_PER_THREAD_Y, 8, half, acc, dst_addr, dst_stride_y, zout.s); |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7458 | } |
Vidhya Sudhan Loganathan | bdff491 | 2018-05-22 15:03:09 +0100 | [diff] [blame] | 7459 | #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 7460 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 7461 | #endif // defined(COLS_A) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && (NUM_ELEMS_PROCESSED_PER_THREAD_Y) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7462 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7463 | #if defined(BETA) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7464 | /** This OpenCL kernel performs the in-place matrix addition between 2 matrices taking into account that the second matrix might be weighted by a scalar value beta: |
| 7465 | * |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 7466 | * @note The beta's value need to be passed at compile time using -DBETA |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7467 | * |
| 7468 | * @param[in] src_ptr Pointer to the source matrix. Supported data types: F32 |
| 7469 | * @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes) |
| 7470 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 7471 | * @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 7472 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7473 | * @param[in] src_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 7474 | * @param[in] src_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7475 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 7476 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7477 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 7478 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 7479 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 7480 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7481 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 7482 | * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7483 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 7484 | */ |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7485 | __kernel void gemm_ma_f32(TENSOR3D_DECLARATION(src), |
| 7486 | TENSOR3D_DECLARATION(dst)) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7487 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7488 | // Compute source and destination addresses |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7489 | Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); |
| 7490 | Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7491 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7492 | // Load values from A x B |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7493 | float4 alpha_ab = vload4(0, (__global float *)dst.ptr); |
| 7494 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7495 | // Load values from Matrix C |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7496 | float4 c = vload4(0, (__global float *)src.ptr); |
| 7497 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7498 | // Computes alpha * axb + beta * c |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7499 | float4 out = alpha_ab + (float4)BETA * c; |
| 7500 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7501 | // Store final result in axb matrix |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7502 | vstore4(out, 0, (__global float *)dst.ptr); |
| 7503 | } |
| 7504 | |
Vidhya Sudhan Loganathan | 76c8564 | 2018-05-25 13:53:02 +0100 | [diff] [blame] | 7505 | #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7506 | /** This OpenCL kernel performs the in-place matrix addition between 2 matrices taking into account that the second matrix might be weighted by a scalar value beta: |
| 7507 | * |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 7508 | * @note The beta's value need to be passed at compile time using -DBETA |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 7509 | * |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7510 | * @param[in] src_ptr Pointer to the source matrix. Supported data types: F16 |
| 7511 | * @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes) |
| 7512 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 7513 | * @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 7514 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7515 | * @param[in] src_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 7516 | * @param[in] src_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7517 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 7518 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7519 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 7520 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 7521 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 7522 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7523 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 7524 | * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7525 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 7526 | */ |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7527 | __kernel void gemm_ma_f16(TENSOR3D_DECLARATION(src), |
| 7528 | TENSOR3D_DECLARATION(dst)) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7529 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7530 | // Compute source and destination addresses |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 7531 | Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); |
| 7532 | Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7533 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7534 | // Load values from A x B |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7535 | half8 alpha_ab = vload8(0, (__global half *)dst.ptr); |
| 7536 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7537 | // Load values from Matrix C |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7538 | half8 c = vload8(0, (__global half *)src.ptr); |
| 7539 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7540 | // Computes alpha * axb + beta * c |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7541 | half8 out = alpha_ab + (half8)BETA * c; |
| 7542 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7543 | // Store final result in axb matrix |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7544 | vstore8(out, 0, (__global half *)dst.ptr); |
| 7545 | } |
Vidhya Sudhan Loganathan | 76c8564 | 2018-05-25 13:53:02 +0100 | [diff] [blame] | 7546 | #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7547 | #endif // defined(BETA) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7548 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7549 | #if defined(WIDTH_VECTOR_A) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7550 | /** This OpenCL kernel computes the vector by matrix multiplication between each row of A (src0) and matrix B (src1) used for locally connected layer |
| 7551 | * |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 7552 | * @note The width of A need to be passed at compile time using -DWIDTH_VECTOR_A |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7553 | * |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 7554 | * @note The input A and matrix B must not be reshaped |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7555 | * |
| 7556 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 |
| 7557 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 7558 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 7559 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 7560 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 7561 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 7562 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7563 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 7564 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 7565 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 7566 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 7567 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 7568 | * @param[in] src1_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 7569 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 7570 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7571 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 7572 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 7573 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 7574 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 7575 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 7576 | */ |
| 7577 | __kernel void gemm_lc_vm_f32(IMAGE_DECLARATION(src0), |
| 7578 | TENSOR3D_DECLARATION(src1), |
| 7579 | IMAGE_DECLARATION(dst)) |
| 7580 | { |
| 7581 | int idx = get_global_id(0) * 4; |
| 7582 | int idy = get_global_id(1); |
| 7583 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7584 | // Compute the address for the vector A and matrix B |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7585 | int2 src_addr = ((int2)(src0_offset_first_element_in_bytes + src0_stride_y * idy, src1_offset_first_element_in_bytes + src1_stride_z * idy)); |
| 7586 | src_addr.s1 += idx * sizeof(float); |
| 7587 | |
| 7588 | int end_row_vec_a = src_addr.s0 + (WIDTH_VECTOR_A * sizeof(float)); |
| 7589 | |
| 7590 | float4 acc = 0.0f; |
| 7591 | |
Georgios Pinitas | 96880cf | 2017-10-20 18:52:20 +0100 | [diff] [blame] | 7592 | for(; src_addr.s0 <= (end_row_vec_a - 2 * (int)sizeof(float)); src_addr += (int2)(2 * sizeof(float), 2 * src1_stride_y)) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7593 | { |
| 7594 | float2 a0 = vload2(0, (__global float *)(src0_ptr + src_addr.s0)); |
| 7595 | float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 7596 | float4 b1 = vload4(0, (__global float *)(src1_ptr + src_addr.s1 + src1_stride_y)); |
| 7597 | |
| 7598 | acc += b0 * (float4)a0.s0; |
| 7599 | acc += b1 * (float4)a0.s1; |
| 7600 | } |
| 7601 | |
| 7602 | for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(float), src1_stride_y)) |
| 7603 | { |
| 7604 | float a0 = *((__global float *)(src0_ptr + src_addr.s0)); |
| 7605 | float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); |
| 7606 | |
| 7607 | acc += b0 * (float4)a0; |
| 7608 | } |
| 7609 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7610 | // Compute destination address |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 7611 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 7612 | |
| 7613 | vstore4(acc, 0, (__global float *)(offset(&dst, 0, 0))); |
| 7614 | } |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 7615 | #endif // defined(WIDTH_VECTOR_A) |