Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 1 | /* |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2019 ARM Limited. |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [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 | */ |
| 24 | #include "helpers.h" |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 25 | #include "helpers_asymm.h" |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 26 | #include "repeat.h" |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 27 | |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 28 | #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
| 29 | #if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 30 | #define ARM_DOT(x, y, val) val = arm_dot_acc((x), (y), (val)); |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 31 | #else // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 32 | #define ARM_DOT(x, y, val) val += arm_dot((x), (y)); |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 33 | #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) |
| 34 | #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
Giorgio Arena | c50da38 | 2018-07-26 15:50:09 +0100 | [diff] [blame] | 35 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 36 | #if defined(COLS_B) && defined(MULT_INTERLEAVE4X4_HEIGHT) && defined(TRANSPOSE1XW_WIDTH_STEP) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 37 | /** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1) |
Gian Marco Iodice | 5fc07aa | 2019-05-15 17:08:02 +0100 | [diff] [blame] | 38 | * Matrix A and matrix B must be reshaped respectively with @ref CLGEMMReshapeLHSMatrixKernel and @ref CLGEMMReshapeRHSMatrixKernel before running the matrix multiplication |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 39 | * |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 40 | * @note The number of matrix B columns needs to be passed at compile time using -DCOLS_B: e.g. -DCOLS_B=1024 |
| 41 | * @note The transposition width step (mult_transpose1xW_width * 4) must be passed at compile time using -DTRANSPOSE1XW_WIDTH_STEP (i.e. -DTRANSPOSE1XW_WIDTH_STEP=2) |
| 42 | * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 43 | * |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 44 | * @note In case the output has to be reinterpreted as a 3D tensor (i.e. output of convolution layer), the following information must be passed at compile time: |
| 45 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 46 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 47 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 48 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 49 | * |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 50 | * @param[in] src0_ptr Pointer to the source matrix. Supported data type: QASYMM8 |
| 51 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 52 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 53 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 54 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 55 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 56 | * @param[in] src1_ptr Pointer to the source matrix. Supported data type: same as @p src0_ptr |
| 57 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 58 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 59 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 60 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 61 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 62 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: S32 |
| 63 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 64 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 65 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 66 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 67 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 68 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 69 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 70 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 71 | * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 72 | */ |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 73 | __kernel void gemmlowp_mm_interleaved_transposed_midgard(IMAGE_DECLARATION(src0), |
| 74 | IMAGE_DECLARATION(src1), |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 75 | IMAGE_DECLARATION(dst), |
| 76 | uint src0_stride_z, |
| 77 | uint src1_stride_z, |
| 78 | uint dst_stride_z |
| 79 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 80 | , |
| 81 | uint cross_plane_pad |
| 82 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 83 | ) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 84 | { |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 85 | const int x = get_global_id(0) / TRANSPOSE1XW_WIDTH_STEP; |
| 86 | const int y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT; |
| 87 | const int z = get_global_id(2); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 88 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 89 | // Offset |
| 90 | const int offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4; |
| 91 | const int offset_row_b = (get_global_id(0) % TRANSPOSE1XW_WIDTH_STEP) * 4; |
| 92 | |
| 93 | // src_addr_a = address of matrix A |
| 94 | // src_addr_b = address of matrix B |
Isabella Gottardi | b92805b | 2018-09-28 18:24:27 +0100 | [diff] [blame] | 95 | __global uchar *src_addr_a = (__global uchar *)(src0_ptr + z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes); |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 96 | __global uchar *src_addr_b = (__global uchar *)(src1_ptr + x * src1_stride_y + src1_offset_first_element_in_bytes); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 97 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 98 | #if defined(MATRIX_B_DEPTH) |
| 99 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 100 | src_addr_b += (z % MATRIX_B_DEPTH) * src1_stride_z; |
| 101 | #else // defined(MATRIX_B_DEPTH) |
| 102 | src_addr_b += z * src1_stride_z; |
| 103 | #endif // defined(MATRIX_B_DEPTH) |
| 104 | |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 105 | // Compute end row address for matrix B |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 106 | __global uchar *src_end_addr_b = src_addr_b + COLS_B; |
| 107 | |
| 108 | src_addr_a += offset_row_a; |
| 109 | src_addr_b += offset_row_b; |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 110 | |
| 111 | // Reset accumulators |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 112 | int4 c00 = 0; |
| 113 | int4 c10 = 0; |
| 114 | int4 c20 = 0; |
| 115 | int4 c30 = 0; |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 116 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 117 | for(; src_addr_b <= (src_end_addr_b - (int)(8 * TRANSPOSE1XW_WIDTH_STEP)); src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 8 * TRANSPOSE1XW_WIDTH_STEP) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 118 | { |
| 119 | // Load values from matrix A (interleaved) and matrix B (transposed) |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 120 | int4 a0 = convert_int4(vload4(0, src_addr_a)); |
| 121 | int4 b0 = convert_int4(vload4(0, src_addr_b)); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 122 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 123 | c00 += (int4)a0.s0 * b0; |
| 124 | c10 += (int4)a0.s1 * b0; |
| 125 | c20 += (int4)a0.s2 * b0; |
| 126 | c30 += (int4)a0.s3 * b0; |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 127 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 128 | a0 = convert_int4(vload4(0, src_addr_a + 4 * MULT_INTERLEAVE4X4_HEIGHT)); |
| 129 | b0 = convert_int4(vload4(0, src_addr_b + 4 * TRANSPOSE1XW_WIDTH_STEP)); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 130 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 131 | c00 += (int4)a0.s0 * b0; |
| 132 | c10 += (int4)a0.s1 * b0; |
| 133 | c20 += (int4)a0.s2 * b0; |
| 134 | c30 += (int4)a0.s3 * b0; |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 135 | } |
| 136 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 137 | for(; src_addr_b < src_end_addr_b; src_addr_a += (4 * MULT_INTERLEAVE4X4_HEIGHT), src_addr_b += (4 * TRANSPOSE1XW_WIDTH_STEP)) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 138 | { |
| 139 | // Load values from matrix A (interleaved) and matrix B (transposed) |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 140 | int4 a0 = convert_int4(vload4(0, src_addr_a)); |
| 141 | int4 b0 = convert_int4(vload4(0, src_addr_b)); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 142 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 143 | c00 += (int4)a0.s0 * b0; |
| 144 | c10 += (int4)a0.s1 * b0; |
| 145 | c20 += (int4)a0.s2 * b0; |
| 146 | c30 += (int4)a0.s3 * b0; |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 147 | } |
| 148 | |
| 149 | // Compute destination address |
| 150 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 151 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 152 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 153 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
| 154 | // in order to take into account the presence of possible cross plane paddings |
| 155 | // |
| 156 | // | | |
| 157 | // | plane0 | |
| 158 | // | | |
| 159 | // |__________________| |
| 160 | // |******************| |
| 161 | // | cross_plane_pad | |
| 162 | // |******************| |
| 163 | // | | |
| 164 | // | plane1 | |
| 165 | // | | |
| 166 | // |__________________| |
| 167 | |
| 168 | // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D |
| 169 | uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D; |
| 170 | zout = min(DEPTH_GEMM3D - 1, zout); |
| 171 | |
| 172 | // Add offset due to the cross plane paddings |
| 173 | zout *= (cross_plane_pad * dst_stride_y); |
| 174 | |
| 175 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 176 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 177 | dst.ptr += z * dst_stride_z * DEPTH_GEMM3D; |
| 178 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 179 | // Store 4x4 block |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 180 | vstore4(c00, 0, (__global int *)(dst.ptr + 0 * dst_stride_y + zout.s0)); |
| 181 | vstore4(c10, 0, (__global int *)(dst.ptr + 1 * dst_stride_y + zout.s1)); |
| 182 | vstore4(c20, 0, (__global int *)(dst.ptr + 2 * dst_stride_y + zout.s2)); |
| 183 | vstore4(c30, 0, (__global int *)(dst.ptr + 3 * dst_stride_y + zout.s3)); |
| 184 | |
| 185 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 186 | // Add offset for batched GEMM |
| 187 | dst.ptr += z * dst_stride_z; |
| 188 | |
| 189 | // Store 4x4 block |
| 190 | vstore4(c00, 0, (__global int *)(dst.ptr + 0 * dst_stride_y)); |
| 191 | vstore4(c10, 0, (__global int *)(dst.ptr + 1 * dst_stride_y)); |
| 192 | vstore4(c20, 0, (__global int *)(dst.ptr + 2 * dst_stride_y)); |
| 193 | vstore4(c30, 0, (__global int *)(dst.ptr + 3 * dst_stride_y)); |
| 194 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 195 | } |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 196 | #endif // defined(COLS_B) && defined(MULT_INTERLEAVE4X4_HEIGHT) && defined(TRANSPOSE1XW_WIDTH_STEP) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 197 | |
| 198 | #if defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_Y) && defined(COLS_A) |
| 199 | #define VECTOR_UCHAR VEC_DATA_TYPE(uchar, NUM_ELEMS_PROCESSED_PER_THREAD_X) |
| 200 | #define VECTOR_UINT VEC_DATA_TYPE(uint, NUM_ELEMS_PROCESSED_PER_THREAD_X) |
| 201 | #define VECTOR_INT VEC_DATA_TYPE(int, NUM_ELEMS_PROCESSED_PER_THREAD_X) |
| 202 | /** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped |
| 203 | * |
| 204 | * @attention The number of matrix A columns needs to be passed at compile time using -DCOLS_A |
| 205 | * |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 206 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 207 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
| 208 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 209 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 210 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 211 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 212 | * |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 213 | * @param[in] src0_ptr Pointer to the source matrix. Supported data type: QASYMM8 |
| 214 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 215 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 216 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 217 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 218 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 219 | * @param[in] src1_ptr Pointer to the source matrix. Supported data type: same as @p src0_ptr |
| 220 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 221 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 222 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 223 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 224 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 225 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: S32 |
| 226 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 227 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 228 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 229 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 230 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 231 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 232 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 233 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 234 | * @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D) |
| 235 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements for the output tensor (only if defined REINTERPRET_OUTPUT_AS_3D) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 236 | */ |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 237 | __kernel void gemmlowp_mm_midgard(IMAGE_DECLARATION(src0), |
| 238 | IMAGE_DECLARATION(src1), |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 239 | IMAGE_DECLARATION(dst), |
| 240 | uint src0_stride_z, |
| 241 | uint src1_stride_z, |
| 242 | uint dst_stride_z |
| 243 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 244 | , |
| 245 | uint src_cross_plane_pad |
| 246 | #endif // REINTERPRET_INPUT_AS_3D |
| 247 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 248 | , |
| 249 | uint dst_cross_plane_pad |
| 250 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 251 | ) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 252 | { |
| 253 | int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; |
| 254 | |
| 255 | // Compute starting address for matrix A and Matrix B |
| 256 | int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); |
| 257 | |
| 258 | // Update address for the matrix A |
| 259 | src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y; |
| 260 | |
| 261 | // Update address for the matrix B |
| 262 | src_addr.s1 += idx; |
| 263 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 264 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 265 | // Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension |
| 266 | // in order to take into account the presence of possible cross plane paddings |
| 267 | // |
| 268 | // | | |
| 269 | // | plane0 | |
| 270 | // | | |
| 271 | // |__________________| |
| 272 | // |******************| |
| 273 | // | cross_plane_pad | |
| 274 | // |******************| |
| 275 | // | | |
| 276 | // | plane1 | |
| 277 | // | | |
| 278 | // |__________________| |
| 279 | |
| 280 | // The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
| 281 | uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
| 282 | zin = min(DEPTH_GEMM3D - 1, zin); |
| 283 | |
| 284 | // Add offset due to the cross plane paddings |
| 285 | zin *= (src_cross_plane_pad * src0_stride_y); |
| 286 | |
| 287 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 288 | // multiply src0_stride_z by DEPTH_GEMM3D |
| 289 | src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D; |
| 290 | |
| 291 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 292 | |
| 293 | // Add offset for batched GEMM |
| 294 | src_addr.s0 += get_global_id(2) * src0_stride_z; |
| 295 | |
| 296 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 297 | |
| 298 | #if defined(MATRIX_B_DEPTH) |
| 299 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 300 | src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; |
| 301 | #else // defined(MATRIX_B_DEPTH) |
| 302 | src_addr.s1 += get_global_id(2) * src1_stride_z; |
| 303 | #endif // defined(MATRIX_B_DEPTH) |
| 304 | |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 305 | int end_row_vec_a = src_addr.s0 + COLS_A; |
| 306 | |
| 307 | VECTOR_UINT acc0 = 0; |
| 308 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 309 | VECTOR_UINT acc1 = 0; |
| 310 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 311 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 312 | VECTOR_UINT acc2 = 0; |
| 313 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 314 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 315 | VECTOR_UINT acc3 = 0; |
| 316 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 317 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 318 | VECTOR_UINT acc4 = 0; |
| 319 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 320 | |
| 321 | for(; src_addr.s0 <= (end_row_vec_a - 2); src_addr += (int2)(2, 2 * src1_stride_y)) |
| 322 | { |
| 323 | // Load values from matrix A |
| 324 | uchar2 a0 = vload2(0, src0_ptr + src_addr.s0 + 0 * src0_stride_y); |
| 325 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 326 | uchar2 a1 = vload2(0, src0_ptr + src_addr.s0 + 1 * src0_stride_y); |
| 327 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 328 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 329 | uchar2 a2 = vload2(0, src0_ptr + src_addr.s0 + 2 * src0_stride_y); |
| 330 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 331 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 332 | uchar2 a3 = vload2(0, src0_ptr + src_addr.s0 + 3 * src0_stride_y); |
| 333 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 334 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 335 | uchar2 a4 = vload2(0, src0_ptr + src_addr.s0 + 4 * src0_stride_y); |
| 336 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 337 | // Load values from matrix B |
| 338 | VECTOR_UCHAR b0 = VLOAD(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, src1_ptr + src_addr.s1); |
| 339 | VECTOR_UCHAR b1 = VLOAD(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, src1_ptr + src_addr.s1 + src1_stride_y); |
| 340 | |
| 341 | // Accumulate |
| 342 | acc0 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a0.s0; |
| 343 | acc0 += CONVERT(b1, VECTOR_UINT) * (VECTOR_UINT)a0.s1; |
| 344 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 345 | acc1 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a1.s0; |
| 346 | acc1 += CONVERT(b1, VECTOR_UINT) * (VECTOR_UINT)a1.s1; |
| 347 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 348 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 349 | acc2 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a2.s0; |
| 350 | acc2 += CONVERT(b1, VECTOR_UINT) * (VECTOR_UINT)a2.s1; |
| 351 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 352 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 353 | acc3 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a3.s0; |
| 354 | acc3 += CONVERT(b1, VECTOR_UINT) * (VECTOR_UINT)a3.s1; |
| 355 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 356 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 357 | acc4 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a4.s0; |
| 358 | acc4 += CONVERT(b1, VECTOR_UINT) * (VECTOR_UINT)a4.s1; |
| 359 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 360 | } |
| 361 | |
| 362 | for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(1, src1_stride_y)) |
| 363 | { |
| 364 | // Load values from matrix A |
| 365 | uchar a0 = *(src0_ptr + src_addr.s0 + 0 * src0_stride_y); |
| 366 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 367 | uchar a1 = *(src0_ptr + src_addr.s0 + 1 * src0_stride_y); |
| 368 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 369 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 370 | uchar a2 = *(src0_ptr + src_addr.s0 + 2 * src0_stride_y); |
| 371 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 372 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 373 | uchar a3 = *(src0_ptr + src_addr.s0 + 3 * src0_stride_y); |
| 374 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 375 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 376 | uchar a4 = *(src0_ptr + src_addr.s0 + 4 * src0_stride_y); |
| 377 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 378 | // Load values from matrix B |
| 379 | VECTOR_UCHAR b0 = VLOAD(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, src1_ptr + src_addr.s1); |
| 380 | |
| 381 | // Accumulate |
| 382 | acc0 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a0; |
| 383 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 384 | acc1 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a1; |
| 385 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 386 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 387 | acc2 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a2; |
| 388 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 389 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 390 | acc3 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a3; |
| 391 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 392 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 393 | acc4 += CONVERT(b0, VECTOR_UINT) * (VECTOR_UINT)a4; |
| 394 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 395 | } |
| 396 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 397 | const int z = get_global_id(2); |
| 398 | |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 399 | // Compute destination address |
| 400 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 401 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 402 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 403 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
| 404 | // in order to take into account the presence of possible cross plane paddings |
| 405 | // |
| 406 | // | | |
| 407 | // | plane0 | |
| 408 | // | | |
| 409 | // |__________________| |
| 410 | // |******************| |
| 411 | // | cross_plane_pad | |
| 412 | // |******************| |
| 413 | // | | |
| 414 | // | plane1 | |
| 415 | // | | |
| 416 | // |__________________| |
| 417 | |
| 418 | // The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
| 419 | uint8 zout = ((uint8)(0, 1, 2, 3, 4, 5, 6, 7) + (uint8)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint8)HEIGHT_GEMM3D; |
| 420 | zout = min(DEPTH_GEMM3D - 1, zout); |
| 421 | |
| 422 | // Add offset due to the cross plane paddings |
| 423 | zout *= (dst_cross_plane_pad * dst_stride_y); |
| 424 | |
| 425 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 426 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 427 | dst.ptr += z * dst_stride_z * DEPTH_GEMM3D; |
| 428 | |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 429 | // Store the result |
| 430 | VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 431 | (CONVERT(acc0, VECTOR_INT), 0, (__global int *)(dst.ptr + 0 * dst_stride_y + zout.s0)); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 432 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 433 | VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 434 | (CONVERT(acc1, VECTOR_INT), 0, (__global int *)(dst.ptr + 1 * dst_stride_y + zout.s1)); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 435 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 436 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 437 | VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 438 | (CONVERT(acc2, VECTOR_INT), 0, (__global int *)(dst.ptr + 2 * dst_stride_y + zout.s2)); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 439 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 440 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 441 | VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 442 | (CONVERT(acc3, VECTOR_INT), 0, (__global int *)(dst.ptr + 3 * dst_stride_y + zout.s3)); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 443 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 444 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 445 | VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 446 | (CONVERT(acc4, VECTOR_INT), 0, (__global int *)(dst.ptr + 4 * dst_stride_y + zout.s4)); |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 447 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 448 | |
| 449 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 450 | // Add offset for batched GEMM |
| 451 | dst.ptr += z * dst_stride_z; |
| 452 | |
| 453 | // Store the result |
| 454 | VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) |
| 455 | (CONVERT(acc0, VECTOR_INT), 0, (__global int *)(dst.ptr + 0 * dst_stride_y)); |
| 456 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 457 | VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) |
| 458 | (CONVERT(acc1, VECTOR_INT), 0, (__global int *)(dst.ptr + 1 * dst_stride_y)); |
| 459 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 460 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 461 | VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) |
| 462 | (CONVERT(acc2, VECTOR_INT), 0, (__global int *)(dst.ptr + 2 * dst_stride_y)); |
| 463 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 464 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 465 | VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) |
| 466 | (CONVERT(acc3, VECTOR_INT), 0, (__global int *)(dst.ptr + 3 * dst_stride_y)); |
| 467 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 468 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 469 | VSTORE(NUM_ELEMS_PROCESSED_PER_THREAD_X) |
| 470 | (CONVERT(acc4, VECTOR_INT), 0, (__global int *)(dst.ptr + 4 * dst_stride_y)); |
| 471 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 472 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 473 | } |
| 474 | |
| 475 | /** OpenCL kernel optimized for Bifrost architectures that computes the matrix multiplication between matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped |
| 476 | * |
| 477 | * @attention The number of matrix A columns needs to be passed at compile time using -DCOLS_A |
| 478 | * |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 479 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 480 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
| 481 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 482 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 483 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 484 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 485 | * |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 486 | * @param[in] src0_ptr Pointer to the source matrix. Supported data type: QASYMM8 |
| 487 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 488 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 489 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 490 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 491 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 492 | * @param[in] src1_ptr Pointer to the source matrix. Supported data type: same as @p src0_ptr |
| 493 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 494 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 495 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 496 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 497 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 498 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: S32 |
| 499 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 500 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 501 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 502 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 503 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 504 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 505 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 506 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 507 | * @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D) |
| 508 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements for the output tensor (only if defined REINTERPRET_OUTPUT_AS_3D) |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 509 | */ |
| 510 | __kernel void gemmlowp_mm_bifrost(IMAGE_DECLARATION(src0), |
| 511 | IMAGE_DECLARATION(src1), |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 512 | IMAGE_DECLARATION(dst), |
| 513 | uint src0_stride_z, |
| 514 | uint src1_stride_z, |
| 515 | uint dst_stride_z |
| 516 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 517 | , |
| 518 | uint src_cross_plane_pad |
| 519 | #endif // REINTERPRET_INPUT_AS_3D |
| 520 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 521 | , |
| 522 | uint dst_cross_plane_pad |
| 523 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 524 | ) |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 525 | { |
| 526 | int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; |
| 527 | |
| 528 | // Compute starting address for matrix A and Matrix B |
| 529 | int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); |
| 530 | |
| 531 | // Update address for the matrix A |
| 532 | src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y; |
| 533 | |
| 534 | // Update address for the matrix B |
| 535 | src_addr.s1 += idx; |
| 536 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 537 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 538 | // Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension |
| 539 | // in order to take into account the presence of possible cross plane paddings |
| 540 | // |
| 541 | // | | |
| 542 | // | plane0 | |
| 543 | // | | |
| 544 | // |__________________| |
| 545 | // |******************| |
| 546 | // | cross_plane_pad | |
| 547 | // |******************| |
| 548 | // | | |
| 549 | // | plane1 | |
| 550 | // | | |
| 551 | // |__________________| |
| 552 | |
| 553 | // The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
| 554 | uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
| 555 | zin = min(DEPTH_GEMM3D - 1, zin); |
| 556 | |
| 557 | // Add offset due to the cross plane paddings |
| 558 | zin *= (src_cross_plane_pad * src0_stride_y); |
| 559 | |
| 560 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 561 | // multiply src0_stride_z by DEPTH_GEMM3D |
| 562 | src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D; |
| 563 | |
| 564 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 565 | |
| 566 | // Add offset for batched GEMM |
| 567 | src_addr.s0 += get_global_id(2) * src0_stride_z; |
| 568 | |
| 569 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 570 | |
| 571 | #if defined(MATRIX_B_DEPTH) |
| 572 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 573 | src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; |
| 574 | #else // defined(MATRIX_B_DEPTH) |
| 575 | src_addr.s1 += get_global_id(2) * src1_stride_z; |
| 576 | #endif // defined(MATRIX_B_DEPTH) |
| 577 | |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 578 | int end_row_vec_a = src_addr.s0 + COLS_A; |
| 579 | |
| 580 | uint acc00 = 0; |
| 581 | uint acc01 = 0; |
| 582 | uint acc02 = 0; |
| 583 | uint acc03 = 0; |
| 584 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 585 | uint acc10 = 0; |
| 586 | uint acc11 = 0; |
| 587 | uint acc12 = 0; |
| 588 | uint acc13 = 0; |
| 589 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 590 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 591 | uint acc20 = 0; |
| 592 | uint acc21 = 0; |
| 593 | uint acc22 = 0; |
| 594 | uint acc23 = 0; |
| 595 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 596 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 597 | uint acc30 = 0; |
| 598 | uint acc31 = 0; |
| 599 | uint acc32 = 0; |
| 600 | uint acc33 = 0; |
| 601 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 602 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 603 | uint acc40 = 0; |
| 604 | uint acc41 = 0; |
| 605 | uint acc42 = 0; |
| 606 | uint acc43 = 0; |
| 607 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 608 | |
| 609 | for(; src_addr.s0 <= (end_row_vec_a - 4); src_addr += (int2)(4, 4 * src1_stride_y)) |
| 610 | { |
| 611 | // Load values from matrix A |
| 612 | uchar4 a0 = vload4(0, src0_ptr + src_addr.s0 + 0 * src0_stride_y); |
| 613 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 614 | uchar4 a1 = vload4(0, src0_ptr + src_addr.s0 + 1 * src0_stride_y); |
| 615 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 616 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 617 | uchar4 a2 = vload4(0, src0_ptr + src_addr.s0 + 2 * src0_stride_y); |
| 618 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 619 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 620 | uchar4 a3 = vload4(0, src0_ptr + src_addr.s0 + 3 * src0_stride_y); |
| 621 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 622 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 623 | uchar4 a4 = vload4(0, src0_ptr + src_addr.s0 + 4 * src0_stride_y); |
| 624 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 625 | // Load values from matrix B |
| 626 | uchar4 b0 = vload4(0, src1_ptr + src_addr.s1 + 0 * src1_stride_y); |
| 627 | uchar4 b1 = vload4(0, src1_ptr + src_addr.s1 + 1 * src1_stride_y); |
| 628 | uchar4 b2 = vload4(0, src1_ptr + src_addr.s1 + 2 * src1_stride_y); |
| 629 | uchar4 b3 = vload4(0, src1_ptr + src_addr.s1 + 3 * src1_stride_y); |
| 630 | |
| 631 | { |
| 632 | // Accumulate |
| 633 | ushort tmp0 = (ushort)b0.s0 * (ushort)a0.s0; |
| 634 | ushort tmp1 = (ushort)b0.s1 * (ushort)a0.s0; |
| 635 | ushort tmp2 = (ushort)b0.s2 * (ushort)a0.s0; |
| 636 | ushort tmp3 = (ushort)b0.s3 * (ushort)a0.s0; |
| 637 | |
| 638 | ushort tmp4 = (ushort)b1.s0 * (ushort)a0.s1; |
| 639 | ushort tmp5 = (ushort)b1.s1 * (ushort)a0.s1; |
| 640 | ushort tmp6 = (ushort)b1.s2 * (ushort)a0.s1; |
| 641 | ushort tmp7 = (ushort)b1.s3 * (ushort)a0.s1; |
| 642 | |
| 643 | ushort tmp8 = (ushort)b2.s0 * (ushort)a0.s2; |
| 644 | ushort tmp9 = (ushort)b2.s1 * (ushort)a0.s2; |
| 645 | ushort tmpA = (ushort)b2.s2 * (ushort)a0.s2; |
| 646 | ushort tmpB = (ushort)b2.s3 * (ushort)a0.s2; |
| 647 | |
| 648 | ushort tmpC = (ushort)b3.s0 * (ushort)a0.s3; |
| 649 | ushort tmpD = (ushort)b3.s1 * (ushort)a0.s3; |
| 650 | ushort tmpE = (ushort)b3.s2 * (ushort)a0.s3; |
| 651 | ushort tmpF = (ushort)b3.s3 * (ushort)a0.s3; |
| 652 | |
| 653 | acc00 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC); |
| 654 | acc01 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD); |
| 655 | acc02 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE); |
| 656 | acc03 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF); |
| 657 | } |
| 658 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 659 | { |
| 660 | // Accumulate |
| 661 | ushort tmp0 = (ushort)b0.s0 * (ushort)a1.s0; |
| 662 | ushort tmp1 = (ushort)b0.s1 * (ushort)a1.s0; |
| 663 | ushort tmp2 = (ushort)b0.s2 * (ushort)a1.s0; |
| 664 | ushort tmp3 = (ushort)b0.s3 * (ushort)a1.s0; |
| 665 | |
| 666 | ushort tmp4 = (ushort)b1.s0 * (ushort)a1.s1; |
| 667 | ushort tmp5 = (ushort)b1.s1 * (ushort)a1.s1; |
| 668 | ushort tmp6 = (ushort)b1.s2 * (ushort)a1.s1; |
| 669 | ushort tmp7 = (ushort)b1.s3 * (ushort)a1.s1; |
| 670 | |
| 671 | ushort tmp8 = (ushort)b2.s0 * (ushort)a1.s2; |
| 672 | ushort tmp9 = (ushort)b2.s1 * (ushort)a1.s2; |
| 673 | ushort tmpA = (ushort)b2.s2 * (ushort)a1.s2; |
| 674 | ushort tmpB = (ushort)b2.s3 * (ushort)a1.s2; |
| 675 | |
| 676 | ushort tmpC = (ushort)b3.s0 * (ushort)a1.s3; |
| 677 | ushort tmpD = (ushort)b3.s1 * (ushort)a1.s3; |
| 678 | ushort tmpE = (ushort)b3.s2 * (ushort)a1.s3; |
| 679 | ushort tmpF = (ushort)b3.s3 * (ushort)a1.s3; |
| 680 | |
| 681 | acc10 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC); |
| 682 | acc11 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD); |
| 683 | acc12 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE); |
| 684 | acc13 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF); |
| 685 | } |
| 686 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 687 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 688 | { |
| 689 | // Accumulate |
| 690 | ushort tmp0 = (ushort)b0.s0 * (ushort)a2.s0; |
| 691 | ushort tmp1 = (ushort)b0.s1 * (ushort)a2.s0; |
| 692 | ushort tmp2 = (ushort)b0.s2 * (ushort)a2.s0; |
| 693 | ushort tmp3 = (ushort)b0.s3 * (ushort)a2.s0; |
| 694 | |
| 695 | ushort tmp4 = (ushort)b1.s0 * (ushort)a2.s1; |
| 696 | ushort tmp5 = (ushort)b1.s1 * (ushort)a2.s1; |
| 697 | ushort tmp6 = (ushort)b1.s2 * (ushort)a2.s1; |
| 698 | ushort tmp7 = (ushort)b1.s3 * (ushort)a2.s1; |
| 699 | |
| 700 | ushort tmp8 = (ushort)b2.s0 * (ushort)a2.s2; |
| 701 | ushort tmp9 = (ushort)b2.s1 * (ushort)a2.s2; |
| 702 | ushort tmpA = (ushort)b2.s2 * (ushort)a2.s2; |
| 703 | ushort tmpB = (ushort)b2.s3 * (ushort)a2.s2; |
| 704 | |
| 705 | ushort tmpC = (ushort)b3.s0 * (ushort)a2.s3; |
| 706 | ushort tmpD = (ushort)b3.s1 * (ushort)a2.s3; |
| 707 | ushort tmpE = (ushort)b3.s2 * (ushort)a2.s3; |
| 708 | ushort tmpF = (ushort)b3.s3 * (ushort)a2.s3; |
| 709 | |
| 710 | acc20 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC); |
| 711 | acc21 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD); |
| 712 | acc22 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE); |
| 713 | acc23 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF); |
| 714 | } |
| 715 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 716 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 717 | { |
| 718 | // Accumulate |
| 719 | ushort tmp0 = (ushort)b0.s0 * (ushort)a3.s0; |
| 720 | ushort tmp1 = (ushort)b0.s1 * (ushort)a3.s0; |
| 721 | ushort tmp2 = (ushort)b0.s2 * (ushort)a3.s0; |
| 722 | ushort tmp3 = (ushort)b0.s3 * (ushort)a3.s0; |
| 723 | |
| 724 | ushort tmp4 = (ushort)b1.s0 * (ushort)a3.s1; |
| 725 | ushort tmp5 = (ushort)b1.s1 * (ushort)a3.s1; |
| 726 | ushort tmp6 = (ushort)b1.s2 * (ushort)a3.s1; |
| 727 | ushort tmp7 = (ushort)b1.s3 * (ushort)a3.s1; |
| 728 | |
| 729 | ushort tmp8 = (ushort)b2.s0 * (ushort)a3.s2; |
| 730 | ushort tmp9 = (ushort)b2.s1 * (ushort)a3.s2; |
| 731 | ushort tmpA = (ushort)b2.s2 * (ushort)a3.s2; |
| 732 | ushort tmpB = (ushort)b2.s3 * (ushort)a3.s2; |
| 733 | |
| 734 | ushort tmpC = (ushort)b3.s0 * (ushort)a3.s3; |
| 735 | ushort tmpD = (ushort)b3.s1 * (ushort)a3.s3; |
| 736 | ushort tmpE = (ushort)b3.s2 * (ushort)a3.s3; |
| 737 | ushort tmpF = (ushort)b3.s3 * (ushort)a3.s3; |
| 738 | |
| 739 | acc30 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC); |
| 740 | acc31 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD); |
| 741 | acc32 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE); |
| 742 | acc33 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF); |
| 743 | } |
| 744 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 745 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 746 | { |
| 747 | // Accumulate |
| 748 | ushort tmp0 = (ushort)b0.s0 * (ushort)a4.s0; |
| 749 | ushort tmp1 = (ushort)b0.s1 * (ushort)a4.s0; |
| 750 | ushort tmp2 = (ushort)b0.s2 * (ushort)a4.s0; |
| 751 | ushort tmp3 = (ushort)b0.s3 * (ushort)a4.s0; |
| 752 | |
| 753 | ushort tmp4 = (ushort)b1.s0 * (ushort)a4.s1; |
| 754 | ushort tmp5 = (ushort)b1.s1 * (ushort)a4.s1; |
| 755 | ushort tmp6 = (ushort)b1.s2 * (ushort)a4.s1; |
| 756 | ushort tmp7 = (ushort)b1.s3 * (ushort)a4.s1; |
| 757 | |
| 758 | ushort tmp8 = (ushort)b2.s0 * (ushort)a4.s2; |
| 759 | ushort tmp9 = (ushort)b2.s1 * (ushort)a4.s2; |
| 760 | ushort tmpA = (ushort)b2.s2 * (ushort)a4.s2; |
| 761 | ushort tmpB = (ushort)b2.s3 * (ushort)a4.s2; |
| 762 | |
| 763 | ushort tmpC = (ushort)b3.s0 * (ushort)a4.s3; |
| 764 | ushort tmpD = (ushort)b3.s1 * (ushort)a4.s3; |
| 765 | ushort tmpE = (ushort)b3.s2 * (ushort)a4.s3; |
| 766 | ushort tmpF = (ushort)b3.s3 * (ushort)a4.s3; |
| 767 | |
| 768 | acc40 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC); |
| 769 | acc41 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD); |
| 770 | acc42 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE); |
| 771 | acc43 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF); |
| 772 | } |
| 773 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 774 | } |
| 775 | |
| 776 | for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(1, src1_stride_y)) |
| 777 | { |
| 778 | // Load values from matrix A |
| 779 | uchar a0 = *(src0_ptr + src_addr.s0 + 0 * src0_stride_y); |
| 780 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 781 | uchar a1 = *(src0_ptr + src_addr.s0 + 1 * src0_stride_y); |
| 782 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 783 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 784 | uchar a2 = *(src0_ptr + src_addr.s0 + 2 * src0_stride_y); |
| 785 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 786 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 787 | uchar a3 = *(src0_ptr + src_addr.s0 + 3 * src0_stride_y); |
| 788 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 789 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 790 | uchar a4 = *(src0_ptr + src_addr.s0 + 4 * src0_stride_y); |
| 791 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 792 | // Load values from matrix B |
| 793 | uchar4 b0 = vload4(0, src1_ptr + src_addr.s1); |
| 794 | |
| 795 | // Accumulate |
| 796 | { |
| 797 | // Accumulate |
| 798 | ushort tmp0 = (ushort)b0.s0 * (ushort)a0; |
| 799 | ushort tmp1 = (ushort)b0.s1 * (ushort)a0; |
| 800 | ushort tmp2 = (ushort)b0.s2 * (ushort)a0; |
| 801 | ushort tmp3 = (ushort)b0.s3 * (ushort)a0; |
| 802 | |
| 803 | acc00 += ((uint)tmp0); |
| 804 | acc01 += ((uint)tmp1); |
| 805 | acc02 += ((uint)tmp2); |
| 806 | acc03 += ((uint)tmp3); |
| 807 | } |
| 808 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 809 | { |
| 810 | // Accumulate |
| 811 | ushort tmp0 = (ushort)b0.s0 * (ushort)a1; |
| 812 | ushort tmp1 = (ushort)b0.s1 * (ushort)a1; |
| 813 | ushort tmp2 = (ushort)b0.s2 * (ushort)a1; |
| 814 | ushort tmp3 = (ushort)b0.s3 * (ushort)a1; |
| 815 | |
| 816 | acc10 += ((uint)tmp0); |
| 817 | acc11 += ((uint)tmp1); |
| 818 | acc12 += ((uint)tmp2); |
| 819 | acc13 += ((uint)tmp3); |
| 820 | } |
| 821 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 822 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 823 | { |
| 824 | // Accumulate |
| 825 | ushort tmp0 = (ushort)b0.s0 * (ushort)a2; |
| 826 | ushort tmp1 = (ushort)b0.s1 * (ushort)a2; |
| 827 | ushort tmp2 = (ushort)b0.s2 * (ushort)a2; |
| 828 | ushort tmp3 = (ushort)b0.s3 * (ushort)a2; |
| 829 | |
| 830 | acc20 += ((uint)tmp0); |
| 831 | acc21 += ((uint)tmp1); |
| 832 | acc22 += ((uint)tmp2); |
| 833 | acc23 += ((uint)tmp3); |
| 834 | } |
| 835 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 836 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 837 | { |
| 838 | // Accumulate |
| 839 | ushort tmp0 = (ushort)b0.s0 * (ushort)a3; |
| 840 | ushort tmp1 = (ushort)b0.s1 * (ushort)a3; |
| 841 | ushort tmp2 = (ushort)b0.s2 * (ushort)a3; |
| 842 | ushort tmp3 = (ushort)b0.s3 * (ushort)a3; |
| 843 | |
| 844 | acc30 += ((uint)tmp0); |
| 845 | acc31 += ((uint)tmp1); |
| 846 | acc32 += ((uint)tmp2); |
| 847 | acc33 += ((uint)tmp3); |
| 848 | } |
| 849 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 850 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 851 | { |
| 852 | // Accumulate |
| 853 | ushort tmp0 = (ushort)b0.s0 * (ushort)a4; |
| 854 | ushort tmp1 = (ushort)b0.s1 * (ushort)a4; |
| 855 | ushort tmp2 = (ushort)b0.s2 * (ushort)a4; |
| 856 | ushort tmp3 = (ushort)b0.s3 * (ushort)a4; |
| 857 | |
| 858 | acc40 += ((uint)tmp0); |
| 859 | acc41 += ((uint)tmp1); |
| 860 | acc42 += ((uint)tmp2); |
| 861 | acc43 += ((uint)tmp3); |
| 862 | } |
| 863 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 864 | } |
| 865 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 866 | const int z = get_global_id(2); |
| 867 | |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 868 | // Compute destination address |
| 869 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 870 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 871 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 872 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
| 873 | // in order to take into account the presence of possible cross plane paddings |
| 874 | // |
| 875 | // | | |
| 876 | // | plane0 | |
| 877 | // | | |
| 878 | // |__________________| |
| 879 | // |******************| |
| 880 | // | cross_plane_pad | |
| 881 | // |******************| |
| 882 | // | | |
| 883 | // | plane1 | |
| 884 | // | | |
| 885 | // |__________________| |
| 886 | |
| 887 | // The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
| 888 | uint8 zout = ((uint8)(0, 1, 2, 3, 4, 5, 6, 7) + (uint8)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint8)HEIGHT_GEMM3D; |
| 889 | zout = min(DEPTH_GEMM3D - 1, zout); |
| 890 | |
| 891 | // Add offset due to the cross plane paddings |
| 892 | zout *= (dst_cross_plane_pad * dst_stride_y); |
| 893 | |
| 894 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 895 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 896 | dst.ptr += z * dst_stride_z * DEPTH_GEMM3D; |
| 897 | |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 898 | // Store the result |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 899 | vstore4((int4)(acc00, acc01, acc02, acc03), 0, (__global int *)(dst.ptr + 0 * dst_stride_y + zout.s0)); |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 900 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 901 | vstore4((int4)(acc10, acc11, acc12, acc13), 0, (__global int *)(dst.ptr + 1 * dst_stride_y + zout.s1)); |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 902 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 903 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 904 | vstore4((int4)(acc20, acc21, acc22, acc23), 0, (__global int *)(dst.ptr + 2 * dst_stride_y + zout.s2)); |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 905 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 906 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 907 | vstore4((int4)(acc30, acc31, acc32, acc33), 0, (__global int *)(dst.ptr + 3 * dst_stride_y + zout.s3)); |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 908 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 909 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 910 | vstore4((int4)(acc40, acc41, acc42, acc43), 0, (__global int *)(dst.ptr + 4 * dst_stride_y + zout.s4)); |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 911 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 912 | |
| 913 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 914 | // Add offset for batched GEMM |
| 915 | dst.ptr += z * dst_stride_z; |
| 916 | |
| 917 | // Store the result |
| 918 | vstore4((int4)(acc00, acc01, acc02, acc03), 0, (__global int *)(dst.ptr + 0 * dst_stride_y)); |
| 919 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 920 | vstore4((int4)(acc10, acc11, acc12, acc13), 0, (__global int *)(dst.ptr + 1 * dst_stride_y)); |
| 921 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 922 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 923 | vstore4((int4)(acc20, acc21, acc22, acc23), 0, (__global int *)(dst.ptr + 2 * dst_stride_y)); |
| 924 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 925 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 926 | vstore4((int4)(acc30, acc31, acc32, acc33), 0, (__global int *)(dst.ptr + 3 * dst_stride_y)); |
| 927 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 928 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 929 | vstore4((int4)(acc40, acc41, acc42, acc43), 0, (__global int *)(dst.ptr + 4 * dst_stride_y)); |
| 930 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4 |
| 931 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 932 | } |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 933 | |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 934 | #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 935 | /** OpenCL kernel optimized to use dot product that computes the matrix multiplication between matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped |
| 936 | * |
| 937 | * @attention The number of matrix A columns needs to be passed at compile time using -DCOLS_A |
| 938 | * |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 939 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 940 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
| 941 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 942 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 943 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 944 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped |
| 945 | * |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 946 | * @param[in] src0_ptr Pointer to the source matrix. Supported data type: QASYMM8 |
| 947 | * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) |
| 948 | * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 949 | * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 950 | * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 951 | * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 952 | * @param[in] src1_ptr Pointer to the source matrix. Supported data type: same as @p src0_ptr |
| 953 | * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) |
| 954 | * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 955 | * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 956 | * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 957 | * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 958 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: S32 |
| 959 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 960 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 961 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 962 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 963 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 964 | * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 965 | * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) |
| 966 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 967 | * @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D) |
| 968 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements for the output tensor (only if defined REINTERPRET_OUTPUT_AS_3D) |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 969 | */ |
| 970 | __kernel void gemmlowp_mm_bifrost_dot8(IMAGE_DECLARATION(src0), |
| 971 | IMAGE_DECLARATION(src1), |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 972 | IMAGE_DECLARATION(dst), |
| 973 | uint src0_stride_z, |
| 974 | uint src1_stride_z, |
| 975 | uint dst_stride_z |
| 976 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 977 | , |
| 978 | uint src_cross_plane_pad |
| 979 | #endif // REINTERPRET_INPUT_AS_3D |
| 980 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 981 | , |
| 982 | uint dst_cross_plane_pad |
| 983 | #endif // REINTERPRET_OUTPUT_AS_3D) |
| 984 | ) |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 985 | { |
| 986 | int idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X; |
| 987 | |
| 988 | // Compute starting address for matrix A and Matrix B |
| 989 | int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); |
| 990 | |
| 991 | // Update address for the matrix A |
| 992 | src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y; |
| 993 | |
| 994 | // Update address for the matrix B |
| 995 | src_addr.s1 += idx; |
| 996 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 997 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 998 | // Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension |
| 999 | // in order to take into account the presence of possible cross plane paddings |
| 1000 | // |
| 1001 | // | | |
| 1002 | // | plane0 | |
| 1003 | // | | |
| 1004 | // |__________________| |
| 1005 | // |******************| |
| 1006 | // | cross_plane_pad | |
| 1007 | // |******************| |
| 1008 | // | | |
| 1009 | // | plane1 | |
| 1010 | // | | |
| 1011 | // |__________________| |
| 1012 | |
| 1013 | // The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
| 1014 | uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
| 1015 | zin = min(DEPTH_GEMM3D - 1, zin); |
| 1016 | |
| 1017 | // Add offset due to the cross plane paddings |
| 1018 | zin *= (src_cross_plane_pad * src0_stride_y); |
| 1019 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1020 | zin += ((uint4)(0, 1, 2, 3)) * src0_stride_y; |
| 1021 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 1022 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 1023 | // multiply src0_stride_z by DEPTH_GEMM3D |
| 1024 | src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D; |
| 1025 | |
| 1026 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 1027 | |
| 1028 | // Add offset for batched GEMM |
| 1029 | src_addr.s0 += get_global_id(2) * src0_stride_z; |
| 1030 | |
| 1031 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 1032 | |
| 1033 | #if defined(MATRIX_B_DEPTH) |
| 1034 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 1035 | src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; |
| 1036 | #else // defined(MATRIX_B_DEPTH) |
| 1037 | src_addr.s1 += get_global_id(2) * src1_stride_z; |
| 1038 | #endif // defined(MATRIX_B_DEPTH) |
| 1039 | |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1040 | uint acc00 = 0; |
| 1041 | uint acc01 = 0; |
| 1042 | uint acc02 = 0; |
| 1043 | uint acc03 = 0; |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1044 | uint acc04 = 0; |
| 1045 | uint acc05 = 0; |
| 1046 | uint acc06 = 0; |
| 1047 | uint acc07 = 0; |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1048 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1049 | uint acc10 = 0; |
| 1050 | uint acc11 = 0; |
| 1051 | uint acc12 = 0; |
| 1052 | uint acc13 = 0; |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1053 | uint acc14 = 0; |
| 1054 | uint acc15 = 0; |
| 1055 | uint acc16 = 0; |
| 1056 | uint acc17 = 0; |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1057 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1058 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1059 | uint acc20 = 0; |
| 1060 | uint acc21 = 0; |
| 1061 | uint acc22 = 0; |
| 1062 | uint acc23 = 0; |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1063 | uint acc24 = 0; |
| 1064 | uint acc25 = 0; |
| 1065 | uint acc26 = 0; |
| 1066 | uint acc27 = 0; |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1067 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1068 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 1069 | uint acc30 = 0; |
| 1070 | uint acc31 = 0; |
| 1071 | uint acc32 = 0; |
| 1072 | uint acc33 = 0; |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1073 | uint acc34 = 0; |
| 1074 | uint acc35 = 0; |
| 1075 | uint acc36 = 0; |
| 1076 | uint acc37 = 0; |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1077 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1078 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1079 | // A and B src indices get incremented at the same time. |
| 1080 | int i = 0; |
| 1081 | for(; i <= ((int)COLS_A - 8); i += 8) |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1082 | { |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1083 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 1084 | // Load values from matrix A and matrix B |
| 1085 | uchar8 a0 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + zin.s0)); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1086 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1087 | uchar8 a1 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + zin.s1)); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1088 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1089 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1090 | uchar8 a2 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + zin.s2)); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1091 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1092 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1093 | uchar8 a3 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + zin.s3)); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1094 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1095 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 1096 | // Load values from matrix A and matrix B |
| 1097 | uchar8 a0 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); |
| 1098 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1099 | uchar8 a1 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); |
| 1100 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1101 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1102 | uchar8 a2 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); |
| 1103 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1104 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 1105 | uchar8 a3 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); |
| 1106 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 1107 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1108 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1109 | uchar8 b0 = vload8(0, src1_ptr + src_addr.s1 + 0 * src1_stride_y); |
| 1110 | uchar8 b1 = vload8(0, src1_ptr + src_addr.s1 + 1 * src1_stride_y); |
| 1111 | uchar8 b2 = vload8(0, src1_ptr + src_addr.s1 + 2 * src1_stride_y); |
| 1112 | uchar8 b3 = vload8(0, src1_ptr + src_addr.s1 + 3 * src1_stride_y); |
| 1113 | src_addr.s1 += 4 * src1_stride_y; |
| 1114 | |
| 1115 | ARM_DOT(a0.s0123, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc00); |
| 1116 | ARM_DOT(a0.s0123, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc01); |
| 1117 | ARM_DOT(a0.s0123, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc02); |
| 1118 | ARM_DOT(a0.s0123, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc03); |
| 1119 | ARM_DOT(a0.s0123, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc04); |
| 1120 | ARM_DOT(a0.s0123, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc05); |
| 1121 | ARM_DOT(a0.s0123, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc06); |
| 1122 | ARM_DOT(a0.s0123, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc07); |
| 1123 | |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1124 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1125 | ARM_DOT(a1.s0123, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc10); |
| 1126 | ARM_DOT(a1.s0123, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc11); |
| 1127 | ARM_DOT(a1.s0123, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc12); |
| 1128 | ARM_DOT(a1.s0123, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc13); |
| 1129 | ARM_DOT(a1.s0123, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc14); |
| 1130 | ARM_DOT(a1.s0123, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc15); |
| 1131 | ARM_DOT(a1.s0123, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc16); |
| 1132 | ARM_DOT(a1.s0123, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc17); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1133 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1134 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1135 | ARM_DOT(a2.s0123, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc20); |
| 1136 | ARM_DOT(a2.s0123, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc21); |
| 1137 | ARM_DOT(a2.s0123, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc22); |
| 1138 | ARM_DOT(a2.s0123, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc23); |
| 1139 | ARM_DOT(a2.s0123, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc24); |
| 1140 | ARM_DOT(a2.s0123, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc25); |
| 1141 | ARM_DOT(a2.s0123, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc26); |
| 1142 | ARM_DOT(a2.s0123, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc27); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1143 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1144 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1145 | ARM_DOT(a3.s0123, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc30); |
| 1146 | ARM_DOT(a3.s0123, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc31); |
| 1147 | ARM_DOT(a3.s0123, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc32); |
| 1148 | ARM_DOT(a3.s0123, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc33); |
| 1149 | ARM_DOT(a3.s0123, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc34); |
| 1150 | ARM_DOT(a3.s0123, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc35); |
| 1151 | ARM_DOT(a3.s0123, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc36); |
| 1152 | ARM_DOT(a3.s0123, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc37); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1153 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1154 | |
| 1155 | b0 = vload8(0, src1_ptr + src_addr.s1 + 0 * src1_stride_y); |
| 1156 | b1 = vload8(0, src1_ptr + src_addr.s1 + 1 * src1_stride_y); |
| 1157 | b2 = vload8(0, src1_ptr + src_addr.s1 + 2 * src1_stride_y); |
| 1158 | b3 = vload8(0, src1_ptr + src_addr.s1 + 3 * src1_stride_y); |
| 1159 | src_addr.s1 += 4 * src1_stride_y; |
| 1160 | |
| 1161 | ARM_DOT(a0.s4567, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc00); |
| 1162 | ARM_DOT(a0.s4567, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc01); |
| 1163 | ARM_DOT(a0.s4567, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc02); |
| 1164 | ARM_DOT(a0.s4567, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc03); |
| 1165 | ARM_DOT(a0.s4567, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc04); |
| 1166 | ARM_DOT(a0.s4567, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc05); |
| 1167 | ARM_DOT(a0.s4567, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc06); |
| 1168 | ARM_DOT(a0.s4567, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc07); |
| 1169 | |
| 1170 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1171 | ARM_DOT(a1.s4567, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc10); |
| 1172 | ARM_DOT(a1.s4567, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc11); |
| 1173 | ARM_DOT(a1.s4567, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc12); |
| 1174 | ARM_DOT(a1.s4567, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc13); |
| 1175 | ARM_DOT(a1.s4567, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc14); |
| 1176 | ARM_DOT(a1.s4567, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc15); |
| 1177 | ARM_DOT(a1.s4567, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc16); |
| 1178 | ARM_DOT(a1.s4567, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc17); |
| 1179 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1180 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1181 | ARM_DOT(a2.s4567, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc20); |
| 1182 | ARM_DOT(a2.s4567, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc21); |
| 1183 | ARM_DOT(a2.s4567, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc22); |
| 1184 | ARM_DOT(a2.s4567, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc23); |
| 1185 | ARM_DOT(a2.s4567, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc24); |
| 1186 | ARM_DOT(a2.s4567, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc25); |
| 1187 | ARM_DOT(a2.s4567, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc26); |
| 1188 | ARM_DOT(a2.s4567, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc27); |
| 1189 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1190 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 1191 | ARM_DOT(a3.s4567, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc30); |
| 1192 | ARM_DOT(a3.s4567, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc31); |
| 1193 | ARM_DOT(a3.s4567, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc32); |
| 1194 | ARM_DOT(a3.s4567, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc33); |
| 1195 | ARM_DOT(a3.s4567, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc34); |
| 1196 | ARM_DOT(a3.s4567, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc35); |
| 1197 | ARM_DOT(a3.s4567, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc36); |
| 1198 | ARM_DOT(a3.s4567, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc37); |
| 1199 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 1200 | |
| 1201 | src_addr.s0 += 8; |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1202 | } |
| 1203 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1204 | for(; i < (int)COLS_A; ++i) |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1205 | { |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1206 | #if defined(REINTERPRET_INPUT_AS_3D) |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1207 | // Load values from matrix A |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1208 | uchar a0 = *((__global uchar *)(src0_ptr + src_addr.s0 + zin.s0)); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1209 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1210 | uchar a1 = *((__global uchar *)(src0_ptr + src_addr.s0 + zin.s1)); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1211 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1212 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1213 | uchar a2 = *((__global uchar *)(src0_ptr + src_addr.s0 + zin.s2)); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1214 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1215 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1216 | uchar a3 = *((__global uchar *)(src0_ptr + src_addr.s0 + zin.s3)); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1217 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1218 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 1219 | // Load values from matrix A |
| 1220 | uchar a0 = *((__global uchar *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); |
| 1221 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1222 | uchar a1 = *((__global uchar *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); |
| 1223 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1224 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1225 | uchar a2 = *((__global uchar *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); |
| 1226 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1227 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 1228 | uchar a3 = *((__global uchar *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); |
| 1229 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 1230 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 1231 | |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1232 | // Load values from matrix B |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1233 | uchar8 b0 = vload8(0, src1_ptr + src_addr.s1); |
| 1234 | src_addr.s1 += src1_stride_y; |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1235 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1236 | acc00 += (uint)a0 * b0.s0; |
| 1237 | acc01 += (uint)a0 * b0.s1; |
| 1238 | acc02 += (uint)a0 * b0.s2; |
| 1239 | acc03 += (uint)a0 * b0.s3; |
| 1240 | acc04 += (uint)a0 * b0.s4; |
| 1241 | acc05 += (uint)a0 * b0.s5; |
| 1242 | acc06 += (uint)a0 * b0.s6; |
| 1243 | acc07 += (uint)a0 * b0.s7; |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1244 | |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1245 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1246 | acc10 += (uint)a1 * b0.s0; |
| 1247 | acc11 += (uint)a1 * b0.s1; |
| 1248 | acc12 += (uint)a1 * b0.s2; |
| 1249 | acc13 += (uint)a1 * b0.s3; |
| 1250 | acc14 += (uint)a1 * b0.s4; |
| 1251 | acc15 += (uint)a1 * b0.s5; |
| 1252 | acc16 += (uint)a1 * b0.s6; |
| 1253 | acc17 += (uint)a1 * b0.s7; |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1254 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1255 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1256 | acc20 += (uint)a2 * b0.s0; |
| 1257 | acc21 += (uint)a2 * b0.s1; |
| 1258 | acc22 += (uint)a2 * b0.s2; |
| 1259 | acc23 += (uint)a2 * b0.s3; |
| 1260 | acc24 += (uint)a2 * b0.s4; |
| 1261 | acc25 += (uint)a2 * b0.s5; |
| 1262 | acc26 += (uint)a2 * b0.s6; |
| 1263 | acc27 += (uint)a2 * b0.s7; |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1264 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1265 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1266 | acc30 += (uint)a3 * b0.s0; |
| 1267 | acc31 += (uint)a3 * b0.s1; |
| 1268 | acc32 += (uint)a3 * b0.s2; |
| 1269 | acc33 += (uint)a3 * b0.s3; |
| 1270 | acc34 += (uint)a3 * b0.s4; |
| 1271 | acc35 += (uint)a3 * b0.s5; |
| 1272 | acc36 += (uint)a3 * b0.s6; |
| 1273 | acc37 += (uint)a3 * b0.s7; |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1274 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1275 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1276 | src_addr.s0 += 1; |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1277 | } |
| 1278 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1279 | int z = get_global_id(2); |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 1280 | |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1281 | // Compute destination address |
| 1282 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 1283 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1284 | // Compute dst address |
| 1285 | __global uchar *dst_addr = dst.ptr; |
| 1286 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 1287 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 1288 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
| 1289 | // in order to take into account the presence of possible cross plane paddings |
| 1290 | // |
| 1291 | // | | |
| 1292 | // | plane0 | |
| 1293 | // | | |
| 1294 | // |__________________| |
| 1295 | // |******************| |
| 1296 | // | cross_plane_pad | |
| 1297 | // |******************| |
| 1298 | // | | |
| 1299 | // | plane1 | |
| 1300 | // | | |
| 1301 | // |__________________| |
| 1302 | |
| 1303 | // The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1304 | uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D; |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 1305 | zout = min(DEPTH_GEMM3D - 1, zout); |
| 1306 | |
| 1307 | // Add offset due to the cross plane paddings |
| 1308 | zout *= (dst_cross_plane_pad * dst_stride_y); |
| 1309 | |
| 1310 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 1311 | // multiply dst_stride_z by DEPTH_GEMM3D |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1312 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 1313 | |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1314 | // Store the result |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1315 | vstore4((int4)(acc00, acc01, acc02, acc03), 0, (__global int *)(dst_addr + 0 * dst_stride_y + zout.s0)); |
| 1316 | vstore4((int4)(acc04, acc05, acc06, acc07), 1, (__global int *)(dst_addr + 0 * dst_stride_y + zout.s0)); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1317 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1318 | vstore4((int4)(acc10, acc11, acc12, acc13), 0, (__global int *)(dst_addr + 1 * dst_stride_y + zout.s1)); |
| 1319 | vstore4((int4)(acc14, acc15, acc16, acc17), 1, (__global int *)(dst_addr + 1 * dst_stride_y + zout.s1)); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1320 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1321 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1322 | vstore4((int4)(acc20, acc21, acc22, acc23), 0, (__global int *)(dst_addr + 2 * dst_stride_y + zout.s2)); |
| 1323 | vstore4((int4)(acc24, acc25, acc26, acc27), 1, (__global int *)(dst_addr + 2 * dst_stride_y + zout.s2)); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1324 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1325 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1326 | vstore4((int4)(acc30, acc31, acc32, acc33), 0, (__global int *)(dst_addr + 3 * dst_stride_y + zout.s3)); |
| 1327 | vstore4((int4)(acc34, acc35, acc36, acc37), 0, (__global int *)(dst_addr + 3 * dst_stride_y + zout.s3)); |
Giorgio Arena | 6200fa4 | 2018-07-06 17:06:36 +0100 | [diff] [blame] | 1328 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 1329 | |
| 1330 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 1331 | // Add offset for batched GEMM |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1332 | dst_addr += z * dst_stride_z; |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 1333 | |
| 1334 | // Store the result |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1335 | vstore4((int4)(acc00, acc01, acc02, acc03), 0, (__global int *)(dst_addr + 0 * dst_stride_y)); |
| 1336 | vstore4((int4)(acc04, acc05, acc06, acc07), 1, (__global int *)(dst_addr + 0 * dst_stride_y)); |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 1337 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1338 | vstore4((int4)(acc10, acc11, acc12, acc13), 0, (__global int *)(dst_addr + 1 * dst_stride_y)); |
| 1339 | vstore4((int4)(acc14, acc15, acc16, acc17), 1, (__global int *)(dst_addr + 1 * dst_stride_y)); |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 1340 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 1341 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1342 | vstore4((int4)(acc20, acc21, acc22, acc23), 0, (__global int *)(dst_addr + 2 * dst_stride_y)); |
| 1343 | vstore4((int4)(acc24, acc25, acc26, acc27), 1, (__global int *)(dst_addr + 2 * dst_stride_y)); |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 1344 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 1345 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1346 | vstore4((int4)(acc30, acc31, acc32, acc33), 0, (__global int *)(dst_addr + 3 * dst_stride_y)); |
| 1347 | vstore4((int4)(acc34, acc35, acc36, acc37), 0, (__global int *)(dst_addr + 3 * dst_stride_y)); |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 1348 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 1349 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 1350 | } |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 1351 | #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 1352 | #endif // defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_Y) && defined(COLS_A) |
| 1353 | |
Gian Marco Iodice | b0c5037 | 2019-03-15 10:13:05 +0000 | [diff] [blame] | 1354 | #if defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(M) && defined(N) |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1355 | |
| 1356 | #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
| 1357 | |
| 1358 | #if K0 == 2 |
| 1359 | #define ARM_DOT_K0(a, b, c) \ |
| 1360 | ({ \ |
| 1361 | ARM_DOT((uchar4)(a, (uchar2)0), (uchar4)(b, (uchar2)0), c); \ |
| 1362 | }) |
| 1363 | #elif K0 == 3 // K0 == 3 |
| 1364 | #define ARM_DOT_K0(a, b, c) \ |
| 1365 | ({ \ |
| 1366 | ARM_DOT((uchar4)(a, (uchar)0), (uchar4)(b, (uchar)0), c); \ |
| 1367 | }) |
| 1368 | #elif K0 == 4 // K0 == 4 |
| 1369 | #define ARM_DOT_K0(a, b, c) \ |
| 1370 | ({ \ |
| 1371 | ARM_DOT(a, b, c); \ |
| 1372 | }) |
| 1373 | #elif K0 == 8 // K0 == 8 |
| 1374 | #define ARM_DOT_K0(a, b, c) \ |
| 1375 | ({ \ |
| 1376 | ARM_DOT(a.s0123, b.s0123, c); \ |
| 1377 | ARM_DOT(a.s4567, b.s4567, c); \ |
| 1378 | }) |
| 1379 | #elif K0 == 16 // K0 == 16 |
| 1380 | #define ARM_DOT_K0(a, b, c) \ |
| 1381 | ({ \ |
| 1382 | ARM_DOT(a.s0123, b.s0123, c); \ |
| 1383 | ARM_DOT(a.s4567, b.s4567, c); \ |
| 1384 | ARM_DOT(a.s89AB, b.s89AB, c); \ |
| 1385 | ARM_DOT(a.sCDEF, b.sCDEF, c); \ |
| 1386 | }) |
| 1387 | #else // K0 not supported |
| 1388 | #error "K0 value not supported" |
| 1389 | #endif // K0 |
| 1390 | |
| 1391 | #else // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
| 1392 | |
| 1393 | #if K0 == 2 |
| 1394 | #define ARM_DOT_K0(a, b, c) \ |
| 1395 | ({ \ |
| 1396 | c += (uint)a.s0 * b.s0; \ |
| 1397 | c += (uint)a.s1 * b.s1; \ |
| 1398 | }) |
| 1399 | #elif K0 == 3 // K0 == 3 |
| 1400 | #define ARM_DOT_K0(a, b, c) \ |
| 1401 | ({ \ |
| 1402 | c += (uint)a.s0 * b.s0; \ |
| 1403 | c += (uint)a.s1 * b.s1; \ |
| 1404 | c += (uint)a.s2 * b.s2; \ |
| 1405 | }) |
| 1406 | #elif K0 == 4 // K0 == 4 |
| 1407 | #define ARM_DOT_K0(a, b, c) \ |
| 1408 | ({ \ |
| 1409 | c += (uint)a.s0 * b.s0; \ |
| 1410 | c += (uint)a.s1 * b.s1; \ |
| 1411 | c += (uint)a.s2 * b.s2; \ |
| 1412 | c += (uint)a.s3 * b.s3; \ |
| 1413 | }) |
| 1414 | #elif K0 == 8 // K0 == 8 |
| 1415 | #define ARM_DOT_K0(a, b, c) \ |
| 1416 | ({ \ |
| 1417 | c += (uint)a.s0 * b.s0; \ |
| 1418 | c += (uint)a.s1 * b.s1; \ |
| 1419 | c += (uint)a.s2 * b.s2; \ |
| 1420 | c += (uint)a.s3 * b.s3; \ |
| 1421 | c += (uint)a.s4 * b.s4; \ |
| 1422 | c += (uint)a.s5 * b.s5; \ |
| 1423 | c += (uint)a.s6 * b.s6; \ |
| 1424 | c += (uint)a.s7 * b.s7; \ |
| 1425 | }) |
| 1426 | #elif K0 == 16 // K0 == 16 |
| 1427 | #define ARM_DOT_K0(a, b, c) \ |
| 1428 | ({ \ |
| 1429 | c += (uint)a.s0 * b.s0; \ |
| 1430 | c += (uint)a.s1 * b.s1; \ |
| 1431 | c += (uint)a.s2 * b.s2; \ |
| 1432 | c += (uint)a.s3 * b.s3; \ |
| 1433 | c += (uint)a.s4 * b.s4; \ |
| 1434 | c += (uint)a.s5 * b.s5; \ |
| 1435 | c += (uint)a.s6 * b.s6; \ |
| 1436 | c += (uint)a.s7 * b.s7; \ |
| 1437 | c += (uint)a.s8 * b.s8; \ |
| 1438 | c += (uint)a.s9 * b.s9; \ |
| 1439 | c += (uint)a.sA * b.sA; \ |
| 1440 | c += (uint)a.sB * b.sB; \ |
| 1441 | c += (uint)a.sC * b.sC; \ |
| 1442 | c += (uint)a.sD * b.sD; \ |
| 1443 | c += (uint)a.sE * b.sE; \ |
| 1444 | c += (uint)a.sF * b.sF; \ |
| 1445 | }) |
| 1446 | #else // K0 not supported |
| 1447 | #error "K0 value not supported" |
| 1448 | #endif // K0 |
| 1449 | |
| 1450 | #endif //defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
| 1451 | |
| 1452 | #if N0 == 2 |
| 1453 | #define ARM_DOT_K0XN0(a, b, c) \ |
| 1454 | ({ \ |
| 1455 | ARM_DOT_K0((a), (b##0), (c.s0)); \ |
| 1456 | ARM_DOT_K0((a), (b##1), (c.s1)); \ |
| 1457 | }) |
| 1458 | #elif N0 == 3 // N0 == 3 |
| 1459 | #define ARM_DOT_K0XN0(a, b, c) \ |
| 1460 | ({ \ |
| 1461 | ARM_DOT_K0((a), (b##0), (c.s0)); \ |
| 1462 | ARM_DOT_K0((a), (b##1), (c.s1)); \ |
| 1463 | ARM_DOT_K0((a), (b##2), (c.s2)); \ |
| 1464 | }) |
| 1465 | #elif N0 == 4 // N0 == 4 |
| 1466 | #define ARM_DOT_K0XN0(a, b, c) \ |
| 1467 | ({ \ |
| 1468 | ARM_DOT_K0((a), (b##0), (c.s0)); \ |
| 1469 | ARM_DOT_K0((a), (b##1), (c.s1)); \ |
| 1470 | ARM_DOT_K0((a), (b##2), (c.s2)); \ |
| 1471 | ARM_DOT_K0((a), (b##3), (c.s3)); \ |
| 1472 | }) |
| 1473 | #elif N0 == 8 // N0 == 8 |
| 1474 | #define ARM_DOT_K0XN0(a, b, c) \ |
| 1475 | ({ \ |
| 1476 | ARM_DOT_K0((a), (b##0), (c.s0)); \ |
| 1477 | ARM_DOT_K0((a), (b##1), (c.s1)); \ |
| 1478 | ARM_DOT_K0((a), (b##2), (c.s2)); \ |
| 1479 | ARM_DOT_K0((a), (b##3), (c.s3)); \ |
| 1480 | ARM_DOT_K0((a), (b##4), (c.s4)); \ |
| 1481 | ARM_DOT_K0((a), (b##5), (c.s5)); \ |
| 1482 | ARM_DOT_K0((a), (b##6), (c.s6)); \ |
| 1483 | ARM_DOT_K0((a), (b##7), (c.s7)); \ |
| 1484 | }) |
| 1485 | #elif N0 == 16 // N0 == 16 |
| 1486 | #define ARM_DOT_K0XN0(a, b, c) \ |
| 1487 | ({ \ |
| 1488 | ARM_DOT_K0((a), (b##0), (c.s0)); \ |
| 1489 | ARM_DOT_K0((a), (b##1), (c.s1)); \ |
| 1490 | ARM_DOT_K0((a), (b##2), (c.s2)); \ |
| 1491 | ARM_DOT_K0((a), (b##3), (c.s3)); \ |
| 1492 | ARM_DOT_K0((a), (b##4), (c.s4)); \ |
| 1493 | ARM_DOT_K0((a), (b##5), (c.s5)); \ |
| 1494 | ARM_DOT_K0((a), (b##6), (c.s6)); \ |
| 1495 | ARM_DOT_K0((a), (b##7), (c.s7)); \ |
| 1496 | ARM_DOT_K0((a), (b##8), (c.s8)); \ |
| 1497 | ARM_DOT_K0((a), (b##9), (c.s9)); \ |
| 1498 | ARM_DOT_K0((a), (b##A), (c.sA)); \ |
| 1499 | ARM_DOT_K0((a), (b##B), (c.sB)); \ |
| 1500 | ARM_DOT_K0((a), (b##C), (c.sC)); \ |
| 1501 | ARM_DOT_K0((a), (b##D), (c.sD)); \ |
| 1502 | ARM_DOT_K0((a), (b##E), (c.sE)); \ |
| 1503 | ARM_DOT_K0((a), (b##F), (c.sF)); \ |
| 1504 | }) |
| 1505 | #else // N0 not supported |
| 1506 | #error "N0 value not supported" |
| 1507 | #endif // N0 conditions |
| 1508 | |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1509 | /** This OpenCL kernel computes the matrix multiplication between 2 matrices with QASYMM data type . |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1510 | * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be NOT transposed |
| 1511 | * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be transposed |
| 1512 | * |
Gian Marco Iodice | b0c5037 | 2019-03-15 10:13:05 +0000 | [diff] [blame] | 1513 | * @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. |
| 1514 | * @note The GEMM's dimensions M and N must be passed at compile time using -DM and -DN (i.e. -DM=52 and -DN=90). |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1515 | * @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 (i.e. -DM0=4, -DN0=8, -DK0=4). |
| 1516 | * @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 (i.e. -DV0=2) |
| 1517 | * @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 (i.e. -DH0=2) |
| 1518 | * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time. |
| 1519 | * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time. |
| 1520 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 1521 | * - M0 = 2, 3, 4, 5, 6, 7, 8 |
| 1522 | * - N0 = 2, 3, 4, 8, 16 |
| 1523 | * - K0 = 2, 3, 4, 8, 16 |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1524 | * - V0 >= 1 |
| 1525 | * - H0 >= 1 |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1526 | * |
| 1527 | * @note In case the output has to be reinterpreted as a 3D tensor (i.e. output of convolution layer), the following information must be passed at compile time: |
| 1528 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 1529 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 1530 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 1531 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped |
| 1532 | * |
| 1533 | * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: QASYMM8 |
| 1534 | * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes) |
| 1535 | * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1536 | * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes) |
| 1537 | * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1538 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix |
| 1539 | * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr |
| 1540 | * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes) |
| 1541 | * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1542 | * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes) |
| 1543 | * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1544 | * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix |
| 1545 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr |
| 1546 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 1547 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1548 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 1549 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1550 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 1551 | * @param[in] k Number of columns in LHS matrix and rows in RHS matrix not reshaped. |
| 1552 | * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) |
| 1553 | * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) |
| 1554 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 1555 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
| 1556 | */ |
| 1557 | __kernel void gemmlowp_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs), |
| 1558 | IMAGE_DECLARATION(rhs), |
| 1559 | IMAGE_DECLARATION(dst), |
| 1560 | uint k, |
| 1561 | uint lhs_stride_z, |
| 1562 | uint rhs_stride_z, |
| 1563 | uint dst_stride_z |
| 1564 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 1565 | , |
| 1566 | uint dst_cross_plane_pad |
| 1567 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 1568 | ) |
| 1569 | { |
| 1570 | // Block size |
| 1571 | #define LHS_BLOCK_SIZE ((K0) * (M0)) |
| 1572 | |
| 1573 | #if defined(LHS_INTERLEAVE) |
| 1574 | #define LHS_OFFSET_X (K0) |
| 1575 | #define LHS_STEP_X ((K0) * (V0)) |
| 1576 | #define LHS_STEP_LOOP (1) |
| 1577 | #else // defined(INTERLEAVE) |
| 1578 | #define LHS_OFFSET_X (LHS_BLOCK_SIZE) |
| 1579 | #define LHS_STEP_X (K0) |
| 1580 | #define LHS_STEP_LOOP (V0) |
| 1581 | #endif // defined(INTERLEAVE) |
| 1582 | |
| 1583 | // Block size |
| 1584 | #define RHS_BLOCK_SIZE ((K0) * (N0)) |
| 1585 | |
| 1586 | // RHS offset and step X |
| 1587 | #if defined(RHS_INTERLEAVE) |
| 1588 | #define RHS_OFFSET_X (K0) |
| 1589 | #define RHS_STEP_X ((K0) * (H0)) |
| 1590 | #define RHS_STEP_LOOP (1) |
| 1591 | #else // defined(RHS_INTERLEAVE) |
| 1592 | #define RHS_OFFSET_X (RHS_BLOCK_SIZE) |
| 1593 | #define RHS_STEP_X (K0) |
| 1594 | #define RHS_STEP_LOOP (H0) |
| 1595 | #endif // defined(RHS_INTERLEAVE) |
| 1596 | |
Gian Marco Iodice | b0c5037 | 2019-03-15 10:13:05 +0000 | [diff] [blame] | 1597 | #if defined(DUMMY_WORK_ITEMS) |
| 1598 | if((get_global_id(0) * N0 >= N) || (get_global_id(1) * M0 >= M)) |
| 1599 | { |
| 1600 | return; |
| 1601 | } |
| 1602 | #endif // defined(DUMMY_WORK_ITEMS) |
| 1603 | |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1604 | // Compute LHS matrix address |
| 1605 | __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (get_global_id(1) % V0) * (uint)LHS_OFFSET_X + (get_global_id(1) / V0) * (uint)lhs_stride_y + (get_global_id( |
| 1606 | 2) |
| 1607 | * lhs_stride_z); |
| 1608 | |
| 1609 | // Compute RHS matrix address |
| 1610 | __global uchar *rhs_addr = rhs_ptr + rhs_offset_first_element_in_bytes + (get_global_id(0) % H0) * (uint)RHS_OFFSET_X + (get_global_id(0) / (uint)H0) * rhs_stride_y; |
| 1611 | |
| 1612 | #if defined(MATRIX_B_DEPTH) |
| 1613 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 1614 | rhs_addr += (get_global_id(2) % MATRIX_B_DEPTH) * rhs_stride_z; |
| 1615 | #else // defined(MATRIX_B_DEPTH) |
| 1616 | rhs_addr += get_global_id(2) * rhs_stride_z; |
| 1617 | #endif // defined(MATRIX_B_DEPTH) |
| 1618 | |
| 1619 | // Initialize the accumulators |
| 1620 | REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(uint, N0), c, 0); //VEC_DATA_TYPE(uint, N0) c0=0,c1=0,c2=0,... c(M0-1)=0; |
| 1621 | |
| 1622 | for(int i = 0; i < k; i += K0) |
| 1623 | { |
| 1624 | // Supported cases (M0, K0): |
| 1625 | // 2,4 - 2,8 - 2,16 |
| 1626 | // 3,4 - 3,8 - 3,16 |
| 1627 | // 4,4 - 4,8 - 4,16 |
| 1628 | // 5,4 - 5,8 - 5,16 |
| 1629 | // 6,4 - 6,8 - 6,16 |
| 1630 | // Load values from LHS matrix |
| 1631 | VEC_DATA_TYPE(uchar, K0) |
| 1632 | a0 = VLOAD(K0)(0, lhs_addr + 0 * LHS_STEP_X); |
| 1633 | #if M0 > 1 |
| 1634 | VEC_DATA_TYPE(uchar, K0) |
| 1635 | a1 = VLOAD(K0)(0, lhs_addr + 1 * LHS_STEP_X); |
| 1636 | #endif // M0 > 1 |
| 1637 | #if M0 > 2 |
| 1638 | VEC_DATA_TYPE(uchar, K0) |
| 1639 | a2 = VLOAD(K0)(0, lhs_addr + 2 * LHS_STEP_X); |
| 1640 | #endif // M0 > 2 |
| 1641 | #if M0 > 3 |
| 1642 | VEC_DATA_TYPE(uchar, K0) |
| 1643 | a3 = VLOAD(K0)(0, lhs_addr + 3 * LHS_STEP_X); |
| 1644 | #endif // M0 > 3 |
| 1645 | #if M0 > 4 |
| 1646 | VEC_DATA_TYPE(uchar, K0) |
| 1647 | a4 = VLOAD(K0)(0, lhs_addr + 4 * LHS_STEP_X); |
| 1648 | #endif // M0 > 4 |
| 1649 | #if M0 > 5 |
| 1650 | VEC_DATA_TYPE(uchar, K0) |
| 1651 | a5 = VLOAD(K0)(0, lhs_addr + 5 * LHS_STEP_X); |
| 1652 | #endif // M0 > 5 |
| 1653 | #if M0 > 6 |
| 1654 | VEC_DATA_TYPE(uchar, K0) |
| 1655 | a6 = VLOAD(K0)(0, lhs_addr + 6 * LHS_STEP_X); |
| 1656 | #endif // M0 > 6 |
| 1657 | #if M0 > 7 |
| 1658 | VEC_DATA_TYPE(uchar, K0) |
| 1659 | a7 = VLOAD(K0)(0, lhs_addr + 7 * LHS_STEP_X); |
| 1660 | #endif // M0 > 7 |
| 1661 | |
| 1662 | // Load values from RHS matrix |
| 1663 | VEC_DATA_TYPE(uchar, K0) |
| 1664 | b0 = VLOAD(K0)(0, rhs_addr + 0 * RHS_STEP_X); |
| 1665 | VEC_DATA_TYPE(uchar, K0) |
| 1666 | b1 = VLOAD(K0)(0, rhs_addr + 1 * RHS_STEP_X); |
| 1667 | #if N0 > 2 |
| 1668 | VEC_DATA_TYPE(uchar, K0) |
| 1669 | b2 = VLOAD(K0)(0, rhs_addr + 2 * RHS_STEP_X); |
| 1670 | #endif // N0 > 2 |
| 1671 | #if N0 > 3 |
| 1672 | VEC_DATA_TYPE(uchar, K0) |
| 1673 | b3 = VLOAD(K0)(0, rhs_addr + 3 * RHS_STEP_X); |
| 1674 | #endif // N0 > 3 |
| 1675 | #if N0 > 4 |
| 1676 | VEC_DATA_TYPE(uchar, K0) |
| 1677 | b4 = VLOAD(K0)(0, rhs_addr + 4 * RHS_STEP_X); |
| 1678 | VEC_DATA_TYPE(uchar, K0) |
| 1679 | b5 = VLOAD(K0)(0, rhs_addr + 5 * RHS_STEP_X); |
| 1680 | VEC_DATA_TYPE(uchar, K0) |
| 1681 | b6 = VLOAD(K0)(0, rhs_addr + 6 * RHS_STEP_X); |
| 1682 | VEC_DATA_TYPE(uchar, K0) |
| 1683 | b7 = VLOAD(K0)(0, rhs_addr + 7 * RHS_STEP_X); |
| 1684 | #endif // N0 > 4 |
| 1685 | #if N0 > 8 |
| 1686 | VEC_DATA_TYPE(uchar, K0) |
| 1687 | b8 = VLOAD(K0)(0, rhs_addr + 8 * RHS_STEP_X); |
| 1688 | VEC_DATA_TYPE(uchar, K0) |
| 1689 | b9 = VLOAD(K0)(0, rhs_addr + 9 * RHS_STEP_X); |
| 1690 | VEC_DATA_TYPE(uchar, K0) |
| 1691 | bA = VLOAD(K0)(0, rhs_addr + 10 * RHS_STEP_X); |
| 1692 | VEC_DATA_TYPE(uchar, K0) |
| 1693 | bB = VLOAD(K0)(0, rhs_addr + 11 * RHS_STEP_X); |
| 1694 | VEC_DATA_TYPE(uchar, K0) |
| 1695 | bC = VLOAD(K0)(0, rhs_addr + 12 * RHS_STEP_X); |
| 1696 | VEC_DATA_TYPE(uchar, K0) |
| 1697 | bD = VLOAD(K0)(0, rhs_addr + 13 * RHS_STEP_X); |
| 1698 | VEC_DATA_TYPE(uchar, K0) |
| 1699 | bE = VLOAD(K0)(0, rhs_addr + 14 * RHS_STEP_X); |
| 1700 | VEC_DATA_TYPE(uchar, K0) |
| 1701 | bF = VLOAD(K0)(0, rhs_addr + 15 * RHS_STEP_X); |
| 1702 | #endif // N0 > 8 |
| 1703 | |
| 1704 | // Accumulate |
| 1705 | ARM_DOT_K0XN0(a0, b, c0); |
| 1706 | #if M0 > 1 |
| 1707 | ARM_DOT_K0XN0(a1, b, c1); |
| 1708 | #endif // M0 > 1 |
| 1709 | #if M0 > 2 |
| 1710 | ARM_DOT_K0XN0(a2, b, c2); |
| 1711 | #endif // M0 > 2 |
| 1712 | #if M0 > 3 |
| 1713 | ARM_DOT_K0XN0(a3, b, c3); |
| 1714 | #endif // M0 > 3 |
| 1715 | #if M0 > 4 |
| 1716 | ARM_DOT_K0XN0(a4, b, c4); |
| 1717 | #endif // M0 > 4 |
| 1718 | #if M0 > 5 |
| 1719 | ARM_DOT_K0XN0(a5, b, c5); |
| 1720 | #endif // M0 > 5 |
| 1721 | #if M0 > 6 |
| 1722 | ARM_DOT_K0XN0(a6, b, c6); |
| 1723 | #endif // M0 > 6 |
| 1724 | #if M0 > 7 |
| 1725 | ARM_DOT_K0XN0(a7, b, c7); |
| 1726 | #endif // M0 > 7 |
| 1727 | |
| 1728 | lhs_addr += (M0 * LHS_STEP_X * LHS_STEP_LOOP); |
| 1729 | rhs_addr += (N0 * RHS_STEP_X * RHS_STEP_LOOP); |
| 1730 | } |
| 1731 | |
| 1732 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(int)) + (get_global_id(1) * (uint)M0 * dst_stride_y); |
| 1733 | |
| 1734 | REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; |
| 1735 | |
| 1736 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 1737 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
| 1738 | // in order to take into account the presence of possible cross plane paddings |
| 1739 | // |
| 1740 | // | | |
| 1741 | // | plane0 | |
| 1742 | // | | |
| 1743 | // |__________________| |
| 1744 | // |******************| |
| 1745 | // | cross_plane_pad | |
| 1746 | // |******************| |
| 1747 | // | | |
| 1748 | // | plane1 | |
| 1749 | // | | |
| 1750 | // |__________________| |
| 1751 | |
| 1752 | // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
| 1753 | zout0 = (0 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 1754 | zout0 = min((uint)(DEPTH_GEMM3D - 1), zout0); |
| 1755 | zout0 *= (dst_cross_plane_pad * dst_stride_y); |
| 1756 | #if M0 > 1 |
| 1757 | zout1 = (1 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 1758 | zout1 = min((uint)(DEPTH_GEMM3D - 1), zout1); |
| 1759 | zout1 *= (dst_cross_plane_pad * dst_stride_y); |
| 1760 | #endif // M0 > 1 |
| 1761 | #if M0 > 2 |
| 1762 | zout2 = (2 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 1763 | zout2 = min((uint)(DEPTH_GEMM3D - 1), zout2); |
| 1764 | zout2 *= (dst_cross_plane_pad * dst_stride_y); |
| 1765 | #endif // M0 > 2 |
| 1766 | #if M0 > 3 |
| 1767 | zout3 = (3 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 1768 | zout3 = min((uint)(DEPTH_GEMM3D - 1), zout3); |
| 1769 | zout3 *= (dst_cross_plane_pad * dst_stride_y); |
| 1770 | #endif // M0 > 3 |
| 1771 | #if M0 > 4 |
| 1772 | zout4 = (4 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 1773 | zout4 = min((uint)(DEPTH_GEMM3D - 1), zout4); |
| 1774 | zout4 *= (dst_cross_plane_pad * dst_stride_y); |
| 1775 | #endif // M0 > 4 |
| 1776 | #if M0 > 5 |
| 1777 | zout5 = (5 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 1778 | zout5 = min((uint)(DEPTH_GEMM3D - 1), zout5); |
| 1779 | zout5 *= (dst_cross_plane_pad * dst_stride_y); |
| 1780 | #endif // M0 > 5 |
| 1781 | #if M0 > 6 |
| 1782 | zout6 = (6 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 1783 | zout6 = min((uint)(DEPTH_GEMM3D - 1), zout6); |
| 1784 | zout6 *= (dst_cross_plane_pad * dst_stride_y); |
| 1785 | #endif // M0 > 6 |
| 1786 | #if M0 > 7 |
| 1787 | zout7 = (7 + (uint)(get_global_id(1) * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 1788 | zout7 = min((uint)(DEPTH_GEMM3D - 1), zout7); |
| 1789 | zout7 *= (dst_cross_plane_pad * dst_stride_y); |
| 1790 | #endif // M0 > 7 |
| 1791 | |
| 1792 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 1793 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 1794 | dst_addr += get_global_id(2) * dst_stride_z * DEPTH_GEMM3D; |
| 1795 | |
| 1796 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 1797 | |
| 1798 | // Add offset for batched GEMM |
| 1799 | dst_addr += get_global_id(2) * dst_stride_z; |
| 1800 | |
| 1801 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 1802 | |
| 1803 | // Store output block |
| 1804 | VSTORE(N0) |
| 1805 | (CONVERT_SAT(c0, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 0 * dst_stride_y + zout0)); |
| 1806 | #if M0 > 1 |
| 1807 | VSTORE(N0) |
| 1808 | (CONVERT_SAT(c1, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 1 * dst_stride_y + zout1)); |
| 1809 | #endif // M0 > 1 |
| 1810 | #if M0 > 2 |
| 1811 | VSTORE(N0) |
| 1812 | (CONVERT_SAT(c2, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 2 * dst_stride_y + zout2)); |
| 1813 | #endif // M0 > 2 |
| 1814 | #if M0 > 3 |
| 1815 | VSTORE(N0) |
| 1816 | (CONVERT_SAT(c3, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 3 * dst_stride_y + zout3)); |
| 1817 | #endif // M0 > 3 |
| 1818 | #if M0 > 4 |
| 1819 | VSTORE(N0) |
| 1820 | (CONVERT_SAT(c4, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 4 * dst_stride_y + zout4)); |
| 1821 | #endif // M0 > 4 |
| 1822 | #if M0 > 5 |
| 1823 | VSTORE(N0) |
| 1824 | (CONVERT_SAT(c5, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 5 * dst_stride_y + zout5)); |
| 1825 | #endif // M0 > 5 |
| 1826 | #if M0 > 6 |
| 1827 | VSTORE(N0) |
| 1828 | (CONVERT_SAT(c6, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 6 * dst_stride_y + zout6)); |
| 1829 | #endif // M0 > 6 |
| 1830 | #if M0 > 7 |
| 1831 | VSTORE(N0) |
| 1832 | (CONVERT_SAT(c7, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 7 * dst_stride_y + zout7)); |
| 1833 | #endif // M0 > 7 |
| 1834 | |
| 1835 | #undef LHS_BLOCK_SIZE |
| 1836 | #undef LHS_OFFSET_X |
| 1837 | #undef LHS_STEP_X |
| 1838 | #undef RHS_BLOCK_SIZE |
| 1839 | #undef RHS_OFFSET_X |
| 1840 | #undef RHS_STEP_X |
| 1841 | } |
| 1842 | |
| 1843 | #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1844 | /** This OpenCL kernel computes the matrix multiplication between 2 matrices with QASYMM8 data type using the dot8 instruction. |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1845 | * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be NOT transposed |
| 1846 | * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be transposed |
| 1847 | * |
| 1848 | * @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 (i.e. -DM0=4, -DN0=8, -DK0=4). |
| 1849 | * @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 (i.e. -DV0=2) |
| 1850 | * @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 (i.e. -DH0=2) |
| 1851 | * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time. |
| 1852 | * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time. |
| 1853 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 1854 | * - M0 = 2, 3, 4, 5, 6, 7, 8 |
| 1855 | * - N0 = 2, 3, 4, 8, 16 |
| 1856 | * - K0 = 2, 3, 4, 8, 16 |
| 1857 | * |
| 1858 | * @note In case the output has to be reinterpreted as a 3D tensor (i.e. output of convolution layer), the following information must be passed at compile time: |
| 1859 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 1860 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 1861 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 1862 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped |
| 1863 | * |
| 1864 | * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: QASYMM8 |
| 1865 | * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes) |
| 1866 | * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1867 | * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes) |
| 1868 | * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1869 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix |
| 1870 | * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr |
| 1871 | * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes) |
| 1872 | * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1873 | * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes) |
| 1874 | * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1875 | * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix |
| 1876 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr |
| 1877 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 1878 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1879 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 1880 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1881 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 1882 | * @param[in] k Number of columns in LHS matrix and rows in RHS matrix not reshaped. |
| 1883 | * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) |
| 1884 | * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) |
| 1885 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 1886 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
| 1887 | */ |
| 1888 | __kernel void gemmlowp_mm_reshaped_lhs_nt_rhs_t_dot8(IMAGE_DECLARATION(lhs), |
| 1889 | IMAGE_DECLARATION(rhs), |
| 1890 | IMAGE_DECLARATION(dst), |
| 1891 | uint k, |
| 1892 | uint lhs_stride_z, |
| 1893 | uint rhs_stride_z, |
| 1894 | uint dst_stride_z |
| 1895 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 1896 | , |
| 1897 | uint dst_cross_plane_pad |
| 1898 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 1899 | ) |
| 1900 | { |
| 1901 | // Note: ARM_DOT_K0XN0 is generated with the dot8 instruction |
| 1902 | gemmlowp_mm_reshaped_lhs_nt_rhs_t(lhs_ptr, |
| 1903 | lhs_stride_x, |
| 1904 | lhs_step_x, |
| 1905 | lhs_stride_y, |
| 1906 | lhs_step_y, |
| 1907 | lhs_offset_first_element_in_bytes, |
| 1908 | rhs_ptr, |
| 1909 | rhs_stride_x, |
| 1910 | rhs_step_x, |
| 1911 | rhs_stride_y, |
| 1912 | rhs_step_y, |
| 1913 | rhs_offset_first_element_in_bytes, |
| 1914 | dst_ptr, |
| 1915 | dst_stride_x, |
| 1916 | dst_step_x, |
| 1917 | dst_stride_y, |
| 1918 | dst_step_y, |
| 1919 | dst_offset_first_element_in_bytes, |
| 1920 | k, |
| 1921 | lhs_stride_z, |
| 1922 | rhs_stride_z, |
| 1923 | dst_stride_z |
| 1924 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 1925 | , |
| 1926 | dst_cross_plane_pad |
| 1927 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 1928 | ); |
| 1929 | } |
| 1930 | #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
| 1931 | #endif // defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(K) |
| 1932 | |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1933 | #if defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(K) |
| 1934 | |
| 1935 | #define CONCAT(a, b) a##b |
| 1936 | |
| 1937 | #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
| 1938 | |
| 1939 | #define ARM_DOT1(a, b, c) \ |
| 1940 | ({ \ |
| 1941 | ARM_DOT((uchar4)(a, (uchar3)0), (uchar4)(b, (uchar3)0), c); \ |
| 1942 | }) |
| 1943 | #define ARM_DOT2(a, b, c) \ |
| 1944 | ({ \ |
| 1945 | ARM_DOT((uchar4)(a, (uchar2)0), (uchar4)(b, (uchar2)0), c); \ |
| 1946 | }) |
| 1947 | #define ARM_DOT3(a, b, c) \ |
| 1948 | ({ \ |
| 1949 | ARM_DOT((uchar4)(a, (uchar)0), (uchar4)(b, (uchar)0), c); \ |
| 1950 | }) |
| 1951 | #define ARM_DOT4(a, b, c) \ |
| 1952 | ({ \ |
| 1953 | ARM_DOT(a, b, c); \ |
| 1954 | }) |
| 1955 | #define ARM_DOT8(a, b, c) \ |
| 1956 | ({ \ |
| 1957 | ARM_DOT4((a.lo), (b.lo), c); \ |
| 1958 | ARM_DOT4((a.hi), (b.hi), c); \ |
| 1959 | }) |
| 1960 | #define ARM_DOT16(a, b, c) \ |
| 1961 | ({ \ |
| 1962 | ARM_DOT8((a.lo), (b.lo), c); \ |
| 1963 | ARM_DOT8((a.hi), (b.hi), c); \ |
| 1964 | }) |
| 1965 | |
| 1966 | #else // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
| 1967 | |
| 1968 | #define ARM_DOT1(a, b, c) \ |
| 1969 | ({ \ |
| 1970 | c += (uint)a.s0 * b.s0; \ |
| 1971 | }) |
| 1972 | #define ARM_DOT2(a, b, c) \ |
| 1973 | ({ \ |
| 1974 | ARM_DOT1(a, b, c); \ |
| 1975 | c += (uint)a.s1 * b.s1; \ |
| 1976 | }) |
| 1977 | #define ARM_DOT3(a, b, c) \ |
| 1978 | ({ \ |
| 1979 | ARM_DOT2(a, b, c); \ |
| 1980 | c += (uint)a.s2 * b.s2; \ |
| 1981 | }) |
| 1982 | #define ARM_DOT4(a, b, c) \ |
| 1983 | ({ \ |
| 1984 | ARM_DOT3(a, b, c); \ |
| 1985 | c += (uint)a.s3 * b.s3; \ |
| 1986 | }) |
| 1987 | #define ARM_DOT8(a, b, c) \ |
| 1988 | ({ \ |
| 1989 | ARM_DOT4((a.lo), (b.lo), c); \ |
| 1990 | ARM_DOT4((a.hi), (b.hi), c); \ |
| 1991 | }) |
| 1992 | #define ARM_DOT16(a, b, c) \ |
| 1993 | ({ \ |
| 1994 | ARM_DOT8((a.lo), (b.lo), c); \ |
| 1995 | ARM_DOT8((a.hi), (b.hi), c); \ |
| 1996 | }) |
| 1997 | #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
| 1998 | |
| 1999 | #if N0 == 2 |
| 2000 | #define ARM_DOT_K0XN0(k0, a, b, c) \ |
| 2001 | ({ \ |
| 2002 | CONCAT(ARM_DOT, k0) \ |
| 2003 | ((a), (b##0), (c.s0)); \ |
| 2004 | CONCAT(ARM_DOT, k0) \ |
| 2005 | ((a), (b##1), (c.s1)); \ |
| 2006 | }) |
| 2007 | #elif N0 == 3 // N0 == 3 |
| 2008 | #define ARM_DOT_K0XN0(k0, a, b, c) \ |
| 2009 | ({ \ |
| 2010 | CONCAT(ARM_DOT, k0) \ |
| 2011 | ((a), (b##0), (c.s0)); \ |
| 2012 | CONCAT(ARM_DOT, k0) \ |
| 2013 | ((a), (b##1), (c.s1)); \ |
| 2014 | CONCAT(ARM_DOT, k0) \ |
| 2015 | ((a), (b##2), (c.s2)); \ |
| 2016 | }) |
| 2017 | #elif N0 == 4 // N0 == 4 |
| 2018 | #define ARM_DOT_K0XN0(k0, a, b, c) \ |
| 2019 | ({ \ |
| 2020 | CONCAT(ARM_DOT, k0) \ |
| 2021 | ((a), (b##0), (c.s0)); \ |
| 2022 | CONCAT(ARM_DOT, k0) \ |
| 2023 | ((a), (b##1), (c.s1)); \ |
| 2024 | CONCAT(ARM_DOT, k0) \ |
| 2025 | ((a), (b##2), (c.s2)); \ |
| 2026 | CONCAT(ARM_DOT, k0) \ |
| 2027 | ((a), (b##3), (c.s3)); \ |
| 2028 | }) |
| 2029 | #elif N0 == 8 // N0 == 8 |
| 2030 | #define ARM_DOT_K0XN0(k0, a, b, c) \ |
| 2031 | ({ \ |
| 2032 | CONCAT(ARM_DOT, k0) \ |
| 2033 | ((a), (b##0), (c.s0)); \ |
| 2034 | CONCAT(ARM_DOT, k0) \ |
| 2035 | ((a), (b##1), (c.s1)); \ |
| 2036 | CONCAT(ARM_DOT, k0) \ |
| 2037 | ((a), (b##2), (c.s2)); \ |
| 2038 | CONCAT(ARM_DOT, k0) \ |
| 2039 | ((a), (b##3), (c.s3)); \ |
| 2040 | CONCAT(ARM_DOT, k0) \ |
| 2041 | ((a), (b##4), (c.s4)); \ |
| 2042 | CONCAT(ARM_DOT, k0) \ |
| 2043 | ((a), (b##5), (c.s5)); \ |
| 2044 | CONCAT(ARM_DOT, k0) \ |
| 2045 | ((a), (b##6), (c.s6)); \ |
| 2046 | CONCAT(ARM_DOT, k0) \ |
| 2047 | ((a), (b##7), (c.s7)); \ |
| 2048 | }) |
| 2049 | #elif N0 == 16 // N0 == 16 |
| 2050 | #define ARM_DOT_K0XN0(k0, a, b, c) \ |
| 2051 | ({ \ |
| 2052 | CONCAT(ARM_DOT, k0) \ |
| 2053 | ((a), (b##0), (c.s0)); \ |
| 2054 | CONCAT(ARM_DOT, k0) \ |
| 2055 | ((a), (b##1), (c.s1)); \ |
| 2056 | CONCAT(ARM_DOT, k0) \ |
| 2057 | ((a), (b##2), (c.s2)); \ |
| 2058 | CONCAT(ARM_DOT, k0) \ |
| 2059 | ((a), (b##3), (c.s3)); \ |
| 2060 | CONCAT(ARM_DOT, k0) \ |
| 2061 | ((a), (b##4), (c.s4)); \ |
| 2062 | CONCAT(ARM_DOT, k0) \ |
| 2063 | ((a), (b##5), (c.s5)); \ |
| 2064 | CONCAT(ARM_DOT, k0) \ |
| 2065 | ((a), (b##6), (c.s6)); \ |
| 2066 | CONCAT(ARM_DOT, k0) \ |
| 2067 | ((a), (b##7), (c.s7)); \ |
| 2068 | CONCAT(ARM_DOT, k0) \ |
| 2069 | ((a), (b##8), (c.s8)); \ |
| 2070 | CONCAT(ARM_DOT, k0) \ |
| 2071 | ((a), (b##9), (c.s9)); \ |
| 2072 | CONCAT(ARM_DOT, k0) \ |
| 2073 | ((a), (b##A), (c.sA)); \ |
| 2074 | CONCAT(ARM_DOT, k0) \ |
| 2075 | ((a), (b##B), (c.sB)); \ |
| 2076 | CONCAT(ARM_DOT, k0) \ |
| 2077 | ((a), (b##C), (c.sC)); \ |
| 2078 | CONCAT(ARM_DOT, k0) \ |
| 2079 | ((a), (b##D), (c.sD)); \ |
| 2080 | CONCAT(ARM_DOT, k0) \ |
| 2081 | ((a), (b##E), (c.sE)); \ |
| 2082 | CONCAT(ARM_DOT, k0) \ |
| 2083 | ((a), (b##F), (c.sF)); \ |
| 2084 | }) |
| 2085 | #else // N0 not supported |
| 2086 | #error "N0 value not supported" |
| 2087 | #endif // N0 conditions |
| 2088 | |
| 2089 | /** This OpenCL kernel computes the matrix multiplication between 2 matrices. |
| 2090 | * The LHS matrix is NOT reshaped |
| 2091 | * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed |
| 2092 | * |
| 2093 | * @note The number of columns of LHS matrix must be passed at compile time using -DK (i.e. -DK=64) |
| 2094 | * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (i.e. -DN0=8, -DK0=4). |
| 2095 | * @note The number of M0 rows to process must be passed at compile time using -DM0 (i.e. -DM0=2) |
| 2096 | * @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 (i.e. -DH0=2) |
| 2097 | * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time. |
| 2098 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 2099 | * - M0 = 1, 2, 3, 4, 5, 6, 7, 8 |
| 2100 | * - N0 = 2, 3, 4, 8, 16 |
| 2101 | * - K0 = 2, 3, 4, 8, 16 |
| 2102 | * - H0 >= 1 |
| 2103 | * |
| 2104 | * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: |
| 2105 | * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D |
| 2106 | * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D |
| 2107 | * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. |
| 2108 | * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor |
| 2109 | * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix |
| 2110 | * |
| 2111 | * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: F16/F32 |
| 2112 | * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes) |
| 2113 | * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 2114 | * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes) |
| 2115 | * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2116 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix |
| 2117 | * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr |
| 2118 | * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes) |
| 2119 | * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 2120 | * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes) |
| 2121 | * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2122 | * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix |
| 2123 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr |
| 2124 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 2125 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 2126 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 2127 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2128 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 2129 | * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) |
| 2130 | * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) |
| 2131 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 2132 | * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) |
| 2133 | * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) |
| 2134 | */ |
| 2135 | __kernel void gemmlowp_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs), |
| 2136 | IMAGE_DECLARATION(rhs), |
| 2137 | IMAGE_DECLARATION(dst), |
| 2138 | uint lhs_stride_z, |
| 2139 | uint rhs_stride_z, |
| 2140 | uint dst_stride_z |
| 2141 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 2142 | , |
| 2143 | uint lhs_cross_plane_pad |
| 2144 | #endif // REINTERPRET_INPUT_AS_3D |
| 2145 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 2146 | , |
| 2147 | uint dst_cross_plane_pad |
| 2148 | #endif // REINTERPRET_OUTPUT_AS_3D |
| 2149 | ) |
| 2150 | { |
| 2151 | // Block size |
| 2152 | #define RHS_BLOCK_SIZE ((K0) * (N0)) |
| 2153 | |
| 2154 | // RHS offset and step X |
| 2155 | #if defined(RHS_INTERLEAVE) |
| 2156 | #define RHS_OFFSET_X (K0) |
| 2157 | #define RHS_STEP_X ((K0) * (H0)) |
| 2158 | #define RHS_STEP_LOOP (1) |
| 2159 | #else // defined(RHS_INTERLEAVE) |
| 2160 | #define RHS_OFFSET_X (RHS_BLOCK_SIZE) |
| 2161 | #define RHS_STEP_X (K0) |
| 2162 | #define RHS_STEP_LOOP (H0) |
| 2163 | #endif // defined(RHS_INTERLEAVE) |
| 2164 | |
| 2165 | uint x = get_global_id(0); |
| 2166 | uint y = get_global_id(1); |
| 2167 | uint z = get_global_id(2); |
| 2168 | |
Gian Marco Iodice | 86cfffe | 2019-04-02 11:02:20 +0100 | [diff] [blame] | 2169 | #if defined(DUMMY_WORK_ITEMS) |
| 2170 | if((x * N0 >= N) || (y * M0 >= M)) |
| 2171 | { |
| 2172 | return; |
| 2173 | } |
| 2174 | #endif // defined(DUMMY_WORK_ITEMS) |
| 2175 | |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 2176 | // Compute LHS matrix address |
| 2177 | uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y; |
| 2178 | |
| 2179 | // Compute RHS matrix address |
| 2180 | uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X + (x / (uint)H0) * rhs_stride_y; |
| 2181 | |
| 2182 | #if defined(MATRIX_B_DEPTH) |
| 2183 | // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 |
| 2184 | rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z; |
| 2185 | #else // defined(MATRIX_B_DEPTH) |
| 2186 | rhs_offset += z * rhs_stride_z; |
| 2187 | #endif // defined(MATRIX_B_DEPTH) |
| 2188 | |
| 2189 | REPEAT_VAR_INIT_TO_CONST(8, uint, zin, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; |
| 2190 | |
| 2191 | #if defined(REINTERPRET_INPUT_AS_3D) |
| 2192 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
| 2193 | // in order to take into account the presence of possible cross plane paddings |
| 2194 | // |
| 2195 | // | | |
| 2196 | // | plane0 | |
| 2197 | // | | |
| 2198 | // |__________________| |
| 2199 | // |******************| |
| 2200 | // | cross_plane_pad | |
| 2201 | // |******************| |
| 2202 | // | | |
| 2203 | // | plane1 | |
| 2204 | // | | |
| 2205 | // |__________________| |
| 2206 | |
| 2207 | // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
| 2208 | zin0 = (0 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2209 | zin0 = min((uint)(DEPTH_GEMM3D - 1), zin0); |
| 2210 | zin0 *= (lhs_cross_plane_pad * lhs_stride_y); |
| 2211 | #if M0 > 1 |
| 2212 | zin1 = (1 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2213 | zin1 = min((uint)(DEPTH_GEMM3D - 1), zin1); |
| 2214 | zin1 *= (lhs_cross_plane_pad * lhs_stride_y); |
| 2215 | #endif // M0 > 1 |
| 2216 | #if M0 > 2 |
| 2217 | zin2 = (2 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2218 | zin2 = min((uint)(DEPTH_GEMM3D - 1), zin2); |
| 2219 | zin2 *= (lhs_cross_plane_pad * lhs_stride_y); |
| 2220 | #endif // M0 > 2 |
| 2221 | #if M0 > 3 |
| 2222 | zin3 = (3 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2223 | zin3 = min((uint)(DEPTH_GEMM3D - 1), zin3); |
| 2224 | zin3 *= (lhs_cross_plane_pad * lhs_stride_y); |
| 2225 | #endif // M0 > 3 |
| 2226 | #if M0 > 4 |
| 2227 | zin4 = (4 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2228 | zin4 = min((uint)(DEPTH_GEMM3D - 1), zin4); |
| 2229 | zin4 *= (lhs_cross_plane_pad * lhs_stride_y); |
| 2230 | #endif // M0 > 4 |
| 2231 | #if M0 > 5 |
| 2232 | zin5 = (5 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2233 | zin5 = min((uint)(DEPTH_GEMM3D - 1), zin5); |
| 2234 | zin5 *= (lhs_cross_plane_pad * lhs_stride_y); |
| 2235 | #endif // M0 > 5 |
| 2236 | #if M0 > 6 |
| 2237 | zin6 = (6 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2238 | zin6 = min((uint)(DEPTH_GEMM3D - 1), zin6); |
| 2239 | zin6 *= (lhs_cross_plane_pad * lhs_stride_y); |
| 2240 | #endif // M0 > 6 |
| 2241 | #if M0 > 7 |
| 2242 | zin7 = (7 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2243 | zin7 = min((uint)(DEPTH_GEMM3D - 1), zout7); |
| 2244 | zin7 *= (lhs_cross_plane_pad * lhs_stride_y); |
| 2245 | #endif // M0 > 7 |
| 2246 | |
| 2247 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 2248 | // multiply lhs_stride_z by DEPTH_GEMM3D |
| 2249 | lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D; |
| 2250 | |
| 2251 | #else // defined(REINTERPRET_INPUT_AS_3D) |
| 2252 | |
| 2253 | // Add offset for batched GEMM |
| 2254 | lhs_offset += z * lhs_stride_z; |
| 2255 | |
| 2256 | #endif // defined(REINTERPRET_INPUT_AS_3D) |
| 2257 | |
| 2258 | // Initialize the accumulators |
| 2259 | REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(uint, N0), c, 0); //VEC_DATA_TYPE(uint, N0) c0=0,c1=0,c2=0,... c(N0-1)=0; |
| 2260 | |
| 2261 | for(int i = 0; i < K; i += K0) |
| 2262 | { |
| 2263 | // Supported cases (M0, K0): |
| 2264 | // 1,2 - 1,3 - 1,4 - 1,8 - 1,16 |
| 2265 | // 2,2 - 2,3 - 2,4 - 2,8 - 2,16 |
| 2266 | // 3,2 - 3,3 - 3,4 - 3,8 - 3,16 |
| 2267 | // 4,2 - 4,3 - 4,4 - 4,8 - 4,16 |
| 2268 | // 5,2 - 5,3 - 5,4 - 5,8 - 5,16 |
| 2269 | // 6,2 - 6,3 - 6,4 - 6,8 - 6,16 |
| 2270 | // 7,2 - 7,3 - 7,4 - 7,8 - 7,16 |
| 2271 | // 8,2 - 8,3 - 8,4 - 8,8 - 8,16 |
| 2272 | // Load values from LHS matrix |
| 2273 | VEC_DATA_TYPE(uchar, K0) |
| 2274 | a0 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 0 * lhs_stride_y + zin0); |
| 2275 | #if M0 > 1 |
| 2276 | VEC_DATA_TYPE(uchar, K0) |
| 2277 | a1 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1); |
| 2278 | #endif // M0 > 1 |
| 2279 | #if M0 > 2 |
| 2280 | VEC_DATA_TYPE(uchar, K0) |
| 2281 | a2 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2); |
| 2282 | #endif // M0 > 2 |
| 2283 | #if M0 > 3 |
| 2284 | VEC_DATA_TYPE(uchar, K0) |
| 2285 | a3 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3); |
| 2286 | #endif // M0 > 3 |
| 2287 | #if M0 > 4 |
| 2288 | VEC_DATA_TYPE(uchar, K0) |
| 2289 | a4 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4); |
| 2290 | #endif // M0 > 4 |
| 2291 | #if M0 > 5 |
| 2292 | VEC_DATA_TYPE(uchar, K0) |
| 2293 | a5 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5); |
| 2294 | #endif // M0 > 5 |
| 2295 | #if M0 > 6 |
| 2296 | VEC_DATA_TYPE(uchar, K0) |
| 2297 | a6 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6); |
| 2298 | #endif // M0 > 6 |
| 2299 | #if M0 > 7 |
| 2300 | VEC_DATA_TYPE(uchar, K0) |
| 2301 | a7 = VLOAD(K0)(0, lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7); |
| 2302 | #endif // M0 > 7 |
| 2303 | |
| 2304 | // Load values from RHS matrix |
| 2305 | VEC_DATA_TYPE(uchar, K0) |
| 2306 | b0 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 0 * RHS_STEP_X); |
| 2307 | VEC_DATA_TYPE(uchar, K0) |
| 2308 | b1 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 1 * RHS_STEP_X); |
| 2309 | #if N0 > 2 |
| 2310 | VEC_DATA_TYPE(uchar, K0) |
| 2311 | b2 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 2 * RHS_STEP_X); |
| 2312 | #endif // N0 > 2 |
| 2313 | #if N0 > 3 |
| 2314 | VEC_DATA_TYPE(uchar, K0) |
| 2315 | b3 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 3 * RHS_STEP_X); |
| 2316 | #endif // N0 > 3 |
| 2317 | #if N0 > 4 |
| 2318 | VEC_DATA_TYPE(uchar, K0) |
| 2319 | b4 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 4 * RHS_STEP_X); |
| 2320 | VEC_DATA_TYPE(uchar, K0) |
| 2321 | b5 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 5 * RHS_STEP_X); |
| 2322 | VEC_DATA_TYPE(uchar, K0) |
| 2323 | b6 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 6 * RHS_STEP_X); |
| 2324 | VEC_DATA_TYPE(uchar, K0) |
| 2325 | b7 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 7 * RHS_STEP_X); |
| 2326 | #endif // N0 > 4 |
| 2327 | #if N0 > 8 |
| 2328 | VEC_DATA_TYPE(uchar, K0) |
| 2329 | b8 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 8 * RHS_STEP_X); |
| 2330 | VEC_DATA_TYPE(uchar, K0) |
| 2331 | b9 = VLOAD(K0)(0, rhs_ptr + rhs_offset + 9 * RHS_STEP_X); |
| 2332 | VEC_DATA_TYPE(uchar, K0) |
| 2333 | bA = VLOAD(K0)(0, rhs_ptr + rhs_offset + 10 * RHS_STEP_X); |
| 2334 | VEC_DATA_TYPE(uchar, K0) |
| 2335 | bB = VLOAD(K0)(0, rhs_ptr + rhs_offset + 11 * RHS_STEP_X); |
| 2336 | VEC_DATA_TYPE(uchar, K0) |
| 2337 | bC = VLOAD(K0)(0, rhs_ptr + rhs_offset + 12 * RHS_STEP_X); |
| 2338 | VEC_DATA_TYPE(uchar, K0) |
| 2339 | bD = VLOAD(K0)(0, rhs_ptr + rhs_offset + 13 * RHS_STEP_X); |
| 2340 | VEC_DATA_TYPE(uchar, K0) |
| 2341 | bE = VLOAD(K0)(0, rhs_ptr + rhs_offset + 14 * RHS_STEP_X); |
| 2342 | VEC_DATA_TYPE(uchar, K0) |
| 2343 | bF = VLOAD(K0)(0, rhs_ptr + rhs_offset + 15 * RHS_STEP_X); |
| 2344 | #endif // N0 > 8 |
| 2345 | |
| 2346 | // Accumulate |
| 2347 | ARM_DOT_K0XN0(K0, a0, b, c0); |
| 2348 | #if M0 > 1 |
| 2349 | ARM_DOT_K0XN0(K0, a1, b, c1); |
| 2350 | #endif // M0 > 1 |
| 2351 | #if M0 > 2 |
| 2352 | ARM_DOT_K0XN0(K0, a2, b, c2); |
| 2353 | #endif // M0 > 2 |
| 2354 | #if M0 > 3 |
| 2355 | ARM_DOT_K0XN0(K0, a3, b, c3); |
| 2356 | #endif // M0 > 3 |
| 2357 | #if M0 > 4 |
| 2358 | ARM_DOT_K0XN0(K0, a4, b, c4); |
| 2359 | #endif // M0 > 4 |
| 2360 | #if M0 > 5 |
| 2361 | ARM_DOT_K0XN0(K0, a5, b, c5); |
| 2362 | #endif // M0 > 5 |
| 2363 | #if M0 > 6 |
| 2364 | ARM_DOT_K0XN0(K0, a6, b, c6); |
| 2365 | #endif // M0 > 6 |
| 2366 | #if M0 > 7 |
| 2367 | ARM_DOT_K0XN0(K0, a7, b, c7); |
| 2368 | #endif // M0 > 7 |
| 2369 | |
| 2370 | lhs_offset += K0; |
| 2371 | rhs_offset += N0 * RHS_STEP_X * RHS_STEP_LOOP; |
| 2372 | } |
| 2373 | |
| 2374 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0) * sizeof(int) + (y * (uint)M0 * dst_stride_y); |
| 2375 | |
| 2376 | REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; |
| 2377 | |
| 2378 | #if defined(REINTERPRET_OUTPUT_AS_3D) |
| 2379 | // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension |
| 2380 | // in order to take into account the presence of possible cross plane paddings |
| 2381 | // |
| 2382 | // | | |
| 2383 | // | plane0 | |
| 2384 | // | | |
| 2385 | // |__________________| |
| 2386 | // |******************| |
| 2387 | // | cross_plane_pad | |
| 2388 | // |******************| |
| 2389 | // | | |
| 2390 | // | plane1 | |
| 2391 | // | | |
| 2392 | // |__________________| |
| 2393 | |
| 2394 | // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D |
| 2395 | zout0 = (0 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2396 | zout0 = min((uint)(DEPTH_GEMM3D - 1), zout0); |
| 2397 | zout0 *= (dst_cross_plane_pad * dst_stride_y); |
| 2398 | #if M0 > 1 |
| 2399 | zout1 = (1 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2400 | zout1 = min((uint)(DEPTH_GEMM3D - 1), zout1); |
| 2401 | zout1 *= (dst_cross_plane_pad * dst_stride_y); |
| 2402 | #endif // M0 > 1 |
| 2403 | #if M0 > 2 |
| 2404 | zout2 = (2 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2405 | zout2 = min((uint)(DEPTH_GEMM3D - 1), zout2); |
| 2406 | zout2 *= (dst_cross_plane_pad * dst_stride_y); |
| 2407 | #endif // M0 > 2 |
| 2408 | #if M0 > 3 |
| 2409 | zout3 = (3 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2410 | zout3 = min((uint)(DEPTH_GEMM3D - 1), zout3); |
| 2411 | zout3 *= (dst_cross_plane_pad * dst_stride_y); |
| 2412 | #endif // M0 > 3 |
| 2413 | #if M0 > 4 |
| 2414 | zout4 = (4 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2415 | zout4 = min((uint)(DEPTH_GEMM3D - 1), zout4); |
| 2416 | zout4 *= (dst_cross_plane_pad * dst_stride_y); |
| 2417 | #endif // M0 > 4 |
| 2418 | #if M0 > 5 |
| 2419 | zout5 = (5 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2420 | zout5 = min((uint)(DEPTH_GEMM3D - 1), zout5); |
| 2421 | zout5 *= (dst_cross_plane_pad * dst_stride_y); |
| 2422 | #endif // M0 > 5 |
| 2423 | #if M0 > 6 |
| 2424 | zout6 = (6 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2425 | zout6 = min((uint)(DEPTH_GEMM3D - 1), zout6); |
| 2426 | zout6 *= (dst_cross_plane_pad * dst_stride_y); |
| 2427 | #endif // M0 > 6 |
| 2428 | #if M0 > 7 |
| 2429 | zout7 = (7 + (uint)(y * (uint)M0)) / (uint)HEIGHT_GEMM3D; |
| 2430 | zout7 = min((uint)(DEPTH_GEMM3D - 1), zout7); |
| 2431 | zout7 *= (dst_cross_plane_pad * dst_stride_y); |
| 2432 | #endif // M0 > 7 |
| 2433 | |
| 2434 | // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we |
| 2435 | // multiply dst_stride_z by DEPTH_GEMM3D |
| 2436 | dst_addr += z * dst_stride_z * DEPTH_GEMM3D; |
| 2437 | |
| 2438 | #else // defined(REINTERPRET_OUTPUT_AS_3D) |
| 2439 | |
| 2440 | // Add offset for batched GEMM |
| 2441 | dst_addr += z * dst_stride_z; |
| 2442 | |
| 2443 | #endif // defined(REINTERPRET_OUTPUT_AS_3D) |
| 2444 | |
| 2445 | // Store output block |
| 2446 | VSTORE(N0) |
| 2447 | (CONVERT_SAT(c0, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 0 * dst_stride_y + zout0)); |
| 2448 | #if M0 > 1 |
| 2449 | VSTORE(N0) |
| 2450 | (CONVERT_SAT(c1, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 1 * dst_stride_y + zout1)); |
| 2451 | #endif // M0 > 1 |
| 2452 | #if M0 > 2 |
| 2453 | VSTORE(N0) |
| 2454 | (CONVERT_SAT(c2, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 2 * dst_stride_y + zout2)); |
| 2455 | #endif // M0 > 2 |
| 2456 | #if M0 > 3 |
| 2457 | VSTORE(N0) |
| 2458 | (CONVERT_SAT(c3, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 3 * dst_stride_y + zout3)); |
| 2459 | #endif // M0 > 3 |
| 2460 | #if M0 > 4 |
| 2461 | VSTORE(N0) |
| 2462 | (CONVERT_SAT(c4, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 4 * dst_stride_y + zout4)); |
| 2463 | #endif // M0 > 4 |
| 2464 | #if M0 > 5 |
| 2465 | VSTORE(N0) |
| 2466 | (CONVERT_SAT(c5, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 5 * dst_stride_y + zout5)); |
| 2467 | #endif // M0 > 5 |
| 2468 | #if M0 > 6 |
| 2469 | VSTORE(N0) |
| 2470 | (CONVERT_SAT(c6, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 6 * dst_stride_y + zout6)); |
| 2471 | #endif // M0 > 6 |
| 2472 | #if M0 > 7 |
| 2473 | VSTORE(N0) |
| 2474 | (CONVERT_SAT(c7, VEC_DATA_TYPE(int, N0)), 0, (__global int *)(dst_addr + 7 * dst_stride_y + zout7)); |
| 2475 | #endif // M0 > 7 |
| 2476 | |
| 2477 | #undef RHS_BLOCK_SIZE |
| 2478 | #undef RHS_OFFSET_X |
| 2479 | #undef RHS_STEP_X |
| 2480 | } |
| 2481 | #endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(K) |
| 2482 | |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2483 | #if defined(COLS_A) |
| 2484 | /** OpenCL kernel used to compute the row-vectors of sums of all the entries in each row of Matrix A. |
| 2485 | * |
| 2486 | * @note This stage is needed to handle the offset of matrix product |
| 2487 | * https://github.com/google/gemmlowp/blob/master/doc/low-precision.md |
| 2488 | * |
| 2489 | * @attention The number of matrix A columns needs to be passed at compile time using -DCOLS_A |
| 2490 | * |
| 2491 | * @param[in] src_ptr Pointer to the source tensor. Supported data type: QASYMM8 |
| 2492 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 2493 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 2494 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 2495 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2496 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 2497 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 2498 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 2499 | * @param[out] dst_ptr Pointer to the destination tensor Supported data type: S32 |
| 2500 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 2501 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 2502 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 2503 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2504 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 2505 | */ |
| 2506 | __kernel void gemmlowp_matrix_a_reduction(TENSOR3D_DECLARATION(src), |
| 2507 | IMAGE_DECLARATION(dst)) |
| 2508 | { |
| 2509 | // Compute source and destination addresses |
| 2510 | Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); |
| 2511 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 2512 | |
| 2513 | uint4 sum_row_u32 = (uint4)0; |
| 2514 | uint sum_row = 0; |
| 2515 | |
| 2516 | __global const uchar *matrix_a = (__global const uchar *)(src.ptr + get_global_id(0) * src_stride_y + get_global_id(1) * src_stride_z); |
| 2517 | |
| 2518 | int i = 0; |
| 2519 | |
| 2520 | // This for loop performs 16 accumulations |
| 2521 | for(; i <= ((int)COLS_A - 16); i += 16) |
| 2522 | { |
| 2523 | const uchar16 a0_u8 = vload16(0, matrix_a + i); |
| 2524 | |
| 2525 | sum_row_u32 += convert_uint4(a0_u8.s0123) + convert_uint4(a0_u8.s4567) + convert_uint4(a0_u8.s89AB) + convert_uint4(a0_u8.sCDEF); |
| 2526 | } |
| 2527 | |
| 2528 | // This for loop performs the leftover accumulations |
| 2529 | for(; i < COLS_A; ++i) |
| 2530 | { |
| 2531 | sum_row += matrix_a[i]; |
| 2532 | } |
| 2533 | |
| 2534 | sum_row += sum_row_u32.s0 + sum_row_u32.s1 + sum_row_u32.s2 + sum_row_u32.s3; |
| 2535 | |
| 2536 | *((__global int *)dst.ptr) = (int)sum_row; |
| 2537 | } |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 2538 | |
| 2539 | #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
| 2540 | /** OpenCL kernel used to compute the row-vectors of sums of all the entries in each row of Matrix A using the arm dot product instruction |
| 2541 | * |
| 2542 | * @note This stage is needed to handle the offset of matrix product |
| 2543 | * https://github.com/google/gemmlowp/blob/master/doc/low-precision.md |
| 2544 | * |
| 2545 | * @attention The number of matrix A columns needs to be passed at compile time using -DCOLS_A |
| 2546 | * |
| 2547 | * @param[in] src_ptr Pointer to the source tensor. Supported data type: QASYMM8 |
| 2548 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 2549 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 2550 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 2551 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2552 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 2553 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 2554 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 2555 | * @param[out] dst_ptr Pointer to the destination tensor Supported data type: S32 |
| 2556 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 2557 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 2558 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 2559 | * @param[in] dst_step_y dst_gx_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 tensor |
| 2561 | */ |
| 2562 | __kernel void gemmlowp_matrix_a_reduction_dot8(TENSOR3D_DECLARATION(src), |
| 2563 | IMAGE_DECLARATION(dst)) |
| 2564 | { |
| 2565 | // Compute source and destination addresses |
| 2566 | Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); |
| 2567 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 2568 | |
| 2569 | uint sum_row = 0; |
| 2570 | |
| 2571 | __global const uchar *matrix_a = (__global const uchar *)(src.ptr + get_global_id(0) * src_stride_y + get_global_id(1) * src_stride_z); |
| 2572 | |
| 2573 | int i = 0; |
| 2574 | |
| 2575 | // This for loop performs 16 accumulations |
| 2576 | for(; i <= ((int)COLS_A - 32); i += 32) |
| 2577 | { |
| 2578 | uchar16 a0_u8 = vload16(0, matrix_a + i); |
| 2579 | |
| 2580 | sum_row += arm_dot(a0_u8.s0123, (uchar4)(1)); |
| 2581 | sum_row += arm_dot(a0_u8.s4567, (uchar4)(1)); |
| 2582 | sum_row += arm_dot(a0_u8.s89AB, (uchar4)(1)); |
| 2583 | sum_row += arm_dot(a0_u8.sCDEF, (uchar4)(1)); |
| 2584 | |
| 2585 | a0_u8 = vload16(1, matrix_a + i); |
| 2586 | |
| 2587 | sum_row += arm_dot(a0_u8.s0123, (uchar4)(1)); |
| 2588 | sum_row += arm_dot(a0_u8.s4567, (uchar4)(1)); |
| 2589 | sum_row += arm_dot(a0_u8.s89AB, (uchar4)(1)); |
| 2590 | sum_row += arm_dot(a0_u8.sCDEF, (uchar4)(1)); |
| 2591 | } |
| 2592 | |
| 2593 | // This for loop performs the leftover accumulations |
| 2594 | for(; i < COLS_A; ++i) |
| 2595 | { |
| 2596 | sum_row += matrix_a[i]; |
| 2597 | } |
| 2598 | |
| 2599 | *((__global int *)dst.ptr) = (int)sum_row; |
| 2600 | } |
| 2601 | #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2602 | #endif // defined(COLS_A) |
| 2603 | |
| 2604 | #if defined(COLS_B) && defined(ROWS_B) |
| 2605 | /** OpenCL kernel used to compute the row-vectors of sums of all the entries in each column of Matrix B. |
| 2606 | * |
| 2607 | * @note This stage is needed to handle the offset of matrix product |
| 2608 | * https://github.com/google/gemmlowp/blob/master/doc/low-precision.md |
| 2609 | * |
| 2610 | * @attention The number of matrix B columns and rows needs to be passed at compile time using -DCOLS_B and -DROWS_B |
| 2611 | * |
| 2612 | * @param[in] src_ptr Pointer to the source tensor. Supported data type: QASYMM8 |
| 2613 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 2614 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 2615 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 2616 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2617 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 2618 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 2619 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 2620 | * @param[out] dst_ptr Pointer to the destination tensor Supported data type: S32 |
| 2621 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 2622 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 2623 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 2624 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2625 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 2626 | */ |
| 2627 | __kernel void gemmlowp_matrix_b_reduction(TENSOR3D_DECLARATION(src), |
| 2628 | IMAGE_DECLARATION(dst)) |
| 2629 | { |
| 2630 | // Compute source and destination addresses |
| 2631 | Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); |
| 2632 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 2633 | |
| 2634 | uint16 sum_col_u32 = (uint16)0; |
| 2635 | |
| 2636 | __global const uchar *matrix_b = (__global const uchar *)(src.ptr + get_global_id(1) * src_stride_z); |
| 2637 | |
| 2638 | int i = 0; |
| 2639 | // This for loop performs 4 accumulations |
| 2640 | for(; i <= ((int)ROWS_B - 4); i += 4) |
| 2641 | { |
| 2642 | const uchar16 b0_u8 = vload16(0, matrix_b + 0 * src_stride_y); |
| 2643 | const uchar16 b1_u8 = vload16(0, matrix_b + 1 * src_stride_y); |
| 2644 | const uchar16 b2_u8 = vload16(0, matrix_b + 2 * src_stride_y); |
| 2645 | const uchar16 b3_u8 = vload16(0, matrix_b + 3 * src_stride_y); |
| 2646 | |
| 2647 | sum_col_u32 += convert_uint16(b0_u8) + convert_uint16(b1_u8) + convert_uint16(b2_u8) + convert_uint16(b3_u8); |
| 2648 | |
| 2649 | matrix_b += 4 * src_stride_y; |
| 2650 | } |
| 2651 | |
| 2652 | // This for loop perfoms the leftover accumulations |
| 2653 | for(; i < (int)ROWS_B; ++i) |
| 2654 | { |
| 2655 | const uchar16 b0_u8 = vload16(0, matrix_b); |
| 2656 | |
| 2657 | sum_col_u32 += convert_uint16(b0_u8); |
| 2658 | |
| 2659 | matrix_b += src_stride_y; |
| 2660 | } |
| 2661 | |
| 2662 | vstore16(convert_int16(sum_col_u32), 0, (__global int *)dst.ptr); |
| 2663 | } |
| 2664 | #endif // defined(COLS_B) && defined(ROWS_B) |
| 2665 | |
| 2666 | #if defined(K_OFFSET) |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 2667 | |
| 2668 | /* Helper function used to calculate the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel. |
| 2669 | * |
| 2670 | * This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel), |
| 2671 | * and calculates the offset contribution of matrix A and matrix B. |
| 2672 | * |
| 2673 | * @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200) |
| 2674 | * @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1) |
| 2675 | * @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6) |
| 2676 | * @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches |
| 2677 | * |
| 2678 | * @param[in] x get_global_id(0) * 4 |
| 2679 | * @param[in] y get_global_id(1) |
| 2680 | * @param[in] z get_global_id(2) |
| 2681 | * @param[in] sum_col_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr |
| 2682 | * @param[in] sum_col_stride_x (Optional) Stride of the source tensor in X dimension (in bytes) |
| 2683 | * @param[in] sum_col_step_x (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes) |
| 2684 | * @param[in] sum_col_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes) |
| 2685 | * @param[in] sum_col_step_y (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2686 | * @param[in] sum_col_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor |
| 2687 | * @param[in] sum_row_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr |
| 2688 | * @param[in] sum_row_stride_x (Optional) Stride of the source tensor in X dimension (in bytes) |
| 2689 | * @param[in] sum_row_step_x (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes) |
| 2690 | * @param[in] sum_row_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes) |
| 2691 | * @param[in] sum_row_step_y (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2692 | * @param[in] sum_row_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor |
| 2693 | * @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr |
| 2694 | * @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes) |
| 2695 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 2696 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor |
| 2697 | */ |
| 2698 | inline int4 offset_contribution( |
| 2699 | int x, |
| 2700 | int y, |
| 2701 | int z |
| 2702 | #if defined(A_OFFSET) |
| 2703 | , |
| 2704 | IMAGE_DECLARATION(sum_col) |
| 2705 | #endif // defined(A_OFFSET) |
| 2706 | #if defined(B_OFFSET) |
| 2707 | , |
| 2708 | IMAGE_DECLARATION(sum_row) |
| 2709 | #endif // defined(B_OFFSET) |
| 2710 | #if defined(ADD_BIAS) |
| 2711 | , |
| 2712 | VECTOR_DECLARATION(biases) |
| 2713 | #endif // defined(ADD_BIAS) |
| 2714 | ) |
| 2715 | { |
| 2716 | int4 a_offset_s32 = (int4)0; |
| 2717 | int4 b_offset_s32 = (int4)0; |
| 2718 | |
| 2719 | int batch_id = z; |
| 2720 | #if defined(DEPTH_INPUT3D) |
| 2721 | batch_id /= (int)DEPTH_INPUT3D; |
| 2722 | #endif // defined(DEPTH_INPUT3D) |
| 2723 | |
| 2724 | #if defined(A_OFFSET) |
| 2725 | // Compute the offset contribution due to A_OFFSET |
| 2726 | __global uchar *sum_col_addr = sum_col_ptr + sum_col_offset_first_element_in_bytes + x * sizeof(int); |
| 2727 | |
| 2728 | // Compute the offset contribution due to A_OFFSET |
| 2729 | #if defined(SUM_COL_HAS_BATCHES) |
| 2730 | a_offset_s32 = vload4(0, (__global int *)(sum_col_addr + batch_id * sum_col_stride_y)); |
| 2731 | #else // defined(SUM_COL_HAS_BATCHES) |
| 2732 | a_offset_s32 = vload4(0, (__global int *)sum_col_addr); |
| 2733 | #endif // defined(SUM_COL_HAS_BATCHES) |
| 2734 | |
| 2735 | a_offset_s32 *= (int4)A_OFFSET; |
| 2736 | #endif // defined(A_OFFSET) |
| 2737 | |
| 2738 | #if defined(B_OFFSET) |
| 2739 | // Compute the offset contribution due to A_OFFSET |
| 2740 | __global uchar *sum_row_addr = sum_row_ptr + sum_row_offset_first_element_in_bytes + y * sizeof(int); |
| 2741 | |
| 2742 | // Compute the offset contribution due to B_OFFSET |
| 2743 | #if defined(HEIGHT_INPUT3D) && defined(DEPTH_INPUT3D) |
| 2744 | b_offset_s32 = (int4) * (((__global int *)(sum_row_addr + batch_id * sum_row_stride_y)) + (z % (int)DEPTH_INPUT3D) * (int)HEIGHT_INPUT3D); |
| 2745 | #else // defined(HEIGHT_INPUT3D) && defined(DEPTH_INPUT3D) |
| 2746 | b_offset_s32 = (int4) * (((__global int *)(sum_row_addr + batch_id * sum_row_stride_y))); |
| 2747 | #endif // defined(HEIGHT_INPUT3D) && defined(DEPTH_INPUT3D) |
| 2748 | b_offset_s32 *= (int4)B_OFFSET; |
| 2749 | #endif // defined(B_OFFSET) |
| 2750 | |
| 2751 | #if defined(ADD_BIAS) |
| 2752 | // Add bias |
| 2753 | __global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(int); |
| 2754 | |
| 2755 | int4 biases_values = vload4(0, (__global int *)bias_addr); |
| 2756 | b_offset_s32 += (int4)biases_values; |
| 2757 | #endif // defined(ADD_BIAS) |
| 2758 | |
| 2759 | return (int4)K_OFFSET + a_offset_s32 + b_offset_s32; |
| 2760 | } |
| 2761 | |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2762 | /* OpenCL kernel used to add the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel. The computation is performed in-place |
| 2763 | * |
| 2764 | * This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel), |
| 2765 | * and adds to it the offset contribution of matrix A and matrix B in-place. |
| 2766 | * |
| 2767 | * @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200) |
| 2768 | * @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1) |
| 2769 | * @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6) |
Chunosov | 5124be5 | 2017-11-22 20:42:13 +0700 | [diff] [blame] | 2770 | * @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2771 | * |
| 2772 | * The final result is: |
| 2773 | * |
| 2774 | * mm_result[i][k] = mm_result[i][k] + |
| 2775 | * (sum_col[k] * A_OFFSET) + |
| 2776 | * (sum_row[i] * B_OFFSET) + |
| 2777 | * (K_OFFSET) |
| 2778 | * |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 2779 | * @param[in] mm_result_ptr Pointer to the source tensor. Supported data type: S32 |
| 2780 | * @param[in] mm_result_stride_x Stride of the source tensor in X dimension (in bytes) |
| 2781 | * @param[in] mm_result_step_x mm_result_stride_x * number of elements along X processed per workitem(in bytes) |
| 2782 | * @param[in] mm_result_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 2783 | * @param[in] mm_result_step_y mm_result_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2784 | * @param[in] mm_result_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 2785 | * @param[in] mm_result_step_z mm_result_stride_z * number of elements along Z processed per workitem(in bytes) |
| 2786 | * @param[in] mm_result_offset_first_element_in_bytes The offset of the first element in the source tensor |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 2787 | * @param[in] sum_col_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr |
| 2788 | * @param[in] sum_col_stride_x (Optional) Stride of the source tensor in X dimension (in bytes) |
| 2789 | * @param[in] sum_col_step_x (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes) |
| 2790 | * @param[in] sum_col_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes) |
| 2791 | * @param[in] sum_col_step_y (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2792 | * @param[in] sum_col_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor |
| 2793 | * @param[in] sum_row_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr |
| 2794 | * @param[in] sum_row_stride_x (Optional) Stride of the source tensor in X dimension (in bytes) |
| 2795 | * @param[in] sum_row_step_x (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes) |
| 2796 | * @param[in] sum_row_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes) |
| 2797 | * @param[in] sum_row_step_y (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2798 | * @param[in] sum_row_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor |
| 2799 | * @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr |
| 2800 | * @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes) |
| 2801 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 2802 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2803 | */ |
| 2804 | __kernel void gemmlowp_offset_contribution(TENSOR3D_DECLARATION(mm_result) |
| 2805 | #if defined(A_OFFSET) |
| 2806 | , |
| 2807 | IMAGE_DECLARATION(sum_col) |
| 2808 | #endif // defined(A_OFFSET) |
| 2809 | #if defined(B_OFFSET) |
| 2810 | , |
| 2811 | IMAGE_DECLARATION(sum_row) |
| 2812 | #endif // defined(B_OFFSET) |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 2813 | #if defined(ADD_BIAS) |
| 2814 | , |
| 2815 | VECTOR_DECLARATION(biases) |
| 2816 | #endif // defined(ADD_BIAS)) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2817 | ) |
| 2818 | { |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 2819 | const int x = get_global_id(0) * 4; |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 2820 | const int y = get_global_id(1); |
| 2821 | const int z = get_global_id(2); |
| 2822 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 2823 | // Compute offset contribution |
| 2824 | int4 offset_term_s32 = offset_contribution( |
| 2825 | x, y, z |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2826 | #if defined(A_OFFSET) |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 2827 | , |
| 2828 | sum_col_ptr, |
| 2829 | sum_col_stride_x, |
| 2830 | sum_col_step_x, |
| 2831 | sum_col_stride_y, |
| 2832 | sum_col_step_y, |
| 2833 | sum_col_offset_first_element_in_bytes |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2834 | #endif // defined(A_OFFSET) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2835 | #if defined(B_OFFSET) |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 2836 | , |
| 2837 | sum_row_ptr, |
| 2838 | sum_row_stride_x, |
| 2839 | sum_row_step_x, |
| 2840 | sum_row_stride_y, |
| 2841 | sum_row_step_y, |
| 2842 | sum_row_offset_first_element_in_bytes |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2843 | #endif // defined(B_OFFSET) |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 2844 | #if defined(ADD_BIAS) |
| 2845 | , |
| 2846 | biases_ptr, |
| 2847 | biases_stride_x, |
| 2848 | biases_step_x, |
| 2849 | biases_offset_first_element_in_bytes |
| 2850 | #endif // defined(ADD_BIAS) |
| 2851 | ); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2852 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 2853 | __global uchar *mm_result_addr = mm_result_ptr + mm_result_offset_first_element_in_bytes + x * sizeof(int) + y * mm_result_stride_y + z * mm_result_stride_z; |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2854 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 2855 | int4 in_s32 = vload4(0, (__global int *)mm_result_addr); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2856 | |
| 2857 | // Add the offset terms to GEMM's result |
| 2858 | in_s32 += offset_term_s32; |
| 2859 | |
| 2860 | // Store the result with the offset contribution |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 2861 | vstore4(in_s32, 0, (__global int *)mm_result_addr); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 2862 | } |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 2863 | |
| 2864 | #if defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT) |
| 2865 | /* OpenCL kernel used to add the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel and it quantizes down to uint8. |
| 2866 | * |
| 2867 | * This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel), adds to it the offset contribution of matrix A and matrix B and quantizes to uint8 through the output stage. |
| 2868 | * |
| 2869 | * |
| 2870 | * @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200) |
| 2871 | * @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1) |
| 2872 | * @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6) |
| 2873 | * @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches |
| 2874 | * |
| 2875 | * The result before the output stage is: |
| 2876 | * |
| 2877 | * mm_result[i][k] = mm_result[i][k] + |
| 2878 | * (sum_col[k] * A_OFFSET) + |
| 2879 | * (sum_row[i] * B_OFFSET) + |
| 2880 | * (K_OFFSET) |
| 2881 | * |
| 2882 | * This result is quantized down to uint8 using the output stage. The output stage computes the following operations: |
| 2883 | * |
| 2884 | * -# Add offset terms to final result |
| 2885 | * -# Multiply each entry of result by result_mult_int |
| 2886 | * -# Add bias to final result (if -DADD_BIAS is passed at compile time) |
| 2887 | * -# Shift the int32 accumulator by result_shift |
| 2888 | * -# Clamp the value between the specified min and max bounds (if -DMIN_BOUND and/or -DMAX_BOUND are passed at compile time) |
| 2889 | * -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8. |
| 2890 | * |
| 2891 | * @attention The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_OFFSET, -RESULT_MULT_INT and -DRESULT_SHIFT |
| 2892 | * |
| 2893 | * @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time |
| 2894 | * @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND. |
| 2895 | * These values can be used to implement "rectified linear unit" activation functions |
| 2896 | * |
| 2897 | * @param[in] mm_result_ptr Pointer to the source tensor. Supported data type: S32 |
| 2898 | * @param[in] mm_result_stride_x Stride of the source tensor in X dimension (in bytes) |
| 2899 | * @param[in] mm_result_step_x mm_result_stride_x * number of elements along X processed per workitem(in bytes) |
| 2900 | * @param[in] mm_result_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 2901 | * @param[in] mm_result_step_y mm_result_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2902 | * @param[in] mm_result_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 2903 | * @param[in] mm_result_step_z mm_result_stride_z * number of elements along Z processed per workitem(in bytes) |
| 2904 | * @param[in] mm_result_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 2905 | * @param[in] sum_col_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr |
| 2906 | * @param[in] sum_col_stride_x (Optional) Stride of the source tensor in X dimension (in bytes) |
| 2907 | * @param[in] sum_col_step_x (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes) |
| 2908 | * @param[in] sum_col_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes) |
| 2909 | * @param[in] sum_col_step_y (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2910 | * @param[in] sum_col_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor |
| 2911 | * @param[in] sum_row_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr |
| 2912 | * @param[in] sum_row_stride_x (Optional) Stride of the source tensor in X dimension (in bytes) |
| 2913 | * @param[in] sum_row_step_x (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes) |
| 2914 | * @param[in] sum_row_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes) |
| 2915 | * @param[in] sum_row_step_y (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2916 | * @param[in] sum_row_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor |
| 2917 | * @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr |
| 2918 | * @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes) |
| 2919 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 2920 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor |
| 2921 | * @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8 |
| 2922 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 2923 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 2924 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 2925 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 2926 | * @param[in] dst_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 2927 | * @param[in] dst_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 2928 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 2929 | */ |
| 2930 | __kernel void gemmlowp_offset_contribution_quantize_down(TENSOR3D_DECLARATION(mm_result) |
| 2931 | #if defined(A_OFFSET) |
| 2932 | , |
| 2933 | IMAGE_DECLARATION(sum_col) |
| 2934 | #endif // defined(A_OFFSET) |
| 2935 | #if defined(B_OFFSET) |
| 2936 | , |
| 2937 | IMAGE_DECLARATION(sum_row) |
| 2938 | #endif // defined(B_OFFSET) |
| 2939 | , |
| 2940 | #if defined(ADD_BIAS) |
| 2941 | VECTOR_DECLARATION(biases), |
| 2942 | #endif // defined(ADD_BIAS) |
| 2943 | TENSOR3D_DECLARATION(dst)) |
| 2944 | { |
| 2945 | const int x = get_global_id(0) * 4; |
| 2946 | const int y = get_global_id(1); |
| 2947 | const int z = get_global_id(2); |
| 2948 | |
| 2949 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z; |
| 2950 | |
| 2951 | // Compute offset contribution |
| 2952 | int4 offset_term_s32 = offset_contribution( |
| 2953 | x, y, z |
| 2954 | #if defined(A_OFFSET) |
| 2955 | , |
| 2956 | sum_col_ptr, |
| 2957 | sum_col_stride_x, |
| 2958 | sum_col_step_x, |
| 2959 | sum_col_stride_y, |
| 2960 | sum_col_step_y, |
| 2961 | sum_col_offset_first_element_in_bytes |
| 2962 | #endif // defined(A_OFFSET) |
| 2963 | #if defined(B_OFFSET) |
| 2964 | , |
| 2965 | sum_row_ptr, |
| 2966 | sum_row_stride_x, |
| 2967 | sum_row_step_x, |
| 2968 | sum_row_stride_y, |
| 2969 | sum_row_step_y, |
| 2970 | sum_row_offset_first_element_in_bytes |
| 2971 | #endif // defined(B_OFFSET) |
| 2972 | #if defined(ADD_BIAS) |
| 2973 | , |
| 2974 | biases_ptr, |
| 2975 | biases_stride_x, |
| 2976 | biases_step_x, |
| 2977 | biases_offset_first_element_in_bytes |
| 2978 | #endif // defined(ADD_BIAS) |
| 2979 | ); |
| 2980 | |
| 2981 | __global uchar *mm_result_addr = mm_result_ptr + mm_result_offset_first_element_in_bytes + x * sizeof(int) + y * mm_result_stride_y + z * mm_result_stride_z; |
| 2982 | |
| 2983 | int4 in_s32 = vload4(0, (__global int *)mm_result_addr); |
| 2984 | |
| 2985 | // Add the offset terms to GEMM's result |
| 2986 | in_s32 += offset_term_s32; |
| 2987 | |
| 2988 | // -------------- OUTPUT STAGE |
| 2989 | |
| 2990 | // Add the offset terms to GEMM's result |
| 2991 | in_s32 += (int4)RESULT_OFFSET; |
| 2992 | |
| 2993 | // Multiply by result_mult_int and shift |
| 2994 | in_s32 *= RESULT_MULTIPLIER; |
| 2995 | |
| 2996 | in_s32 >>= RESULT_SHIFT; |
| 2997 | |
| 2998 | uchar4 res = convert_uchar4_sat(in_s32); |
| 2999 | |
| 3000 | #if defined(MIN_BOUND) |
| 3001 | res = max(res, (uchar4)MIN_BOUND); |
| 3002 | #endif // defined(MIN_BOUND) |
| 3003 | #if defined(MAX_BOUND) |
| 3004 | res = min(res, (uchar4)MAX_BOUND); |
| 3005 | #endif // defined(MAX_BOUND) |
| 3006 | |
| 3007 | // Store the result |
| 3008 | vstore4(res, 0, dst_addr); |
| 3009 | } |
| 3010 | |
| 3011 | /* OpenCL kernel used to add the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel and it quantizes down to uint8. |
| 3012 | * |
| 3013 | * This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel), adds to it the offset contribution of matrix A and matrix B and quantizes to uint8 through the output stage. |
| 3014 | * |
| 3015 | * |
| 3016 | * @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200) |
| 3017 | * @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1) |
| 3018 | * @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6) |
| 3019 | * @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches |
| 3020 | * |
| 3021 | * The result before the output stage is: |
| 3022 | * |
| 3023 | * mm_result[i][k] = mm_result[i][k] + |
| 3024 | * (sum_col[k] * A_OFFSET) + |
| 3025 | * (sum_row[i] * B_OFFSET) + |
| 3026 | * (K_OFFSET) |
| 3027 | * |
| 3028 | * This result is quantized down to uint8 using the output stage. The output stage computes the following operations: |
| 3029 | * |
| 3030 | * -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier |
| 3031 | * -# Add bias to final result if bias tensor is not a nullptr |
| 3032 | * -# Round to nearest division by a power-of-two using result_shift |
| 3033 | * -# Add offset to each result |
| 3034 | * -# Clamp the value between the specified min and max bounds |
| 3035 | * -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8. |
| 3036 | * |
| 3037 | * @attention The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_OFFSET, -RESULT_MULT_INT and -DRESULT_SHIFT |
| 3038 | * |
| 3039 | * @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time |
| 3040 | * @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND. |
| 3041 | * These values can be used to implement "rectified linear unit" activation functions |
| 3042 | * |
| 3043 | * @param[in] mm_result_ptr Pointer to the source tensor. Supported data type: S32 |
| 3044 | * @param[in] mm_result_stride_x Stride of the source tensor in X dimension (in bytes) |
| 3045 | * @param[in] mm_result_step_x mm_result_stride_x * number of elements along X processed per workitem(in bytes) |
| 3046 | * @param[in] mm_result_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 3047 | * @param[in] mm_result_step_y mm_result_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3048 | * @param[in] mm_result_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 3049 | * @param[in] mm_result_step_z mm_result_stride_z * number of elements along Z processed per workitem(in bytes) |
| 3050 | * @param[in] mm_result_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 3051 | * @param[in] sum_col_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr |
| 3052 | * @param[in] sum_col_stride_x (Optional) Stride of the source tensor in X dimension (in bytes) |
| 3053 | * @param[in] sum_col_step_x (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes) |
| 3054 | * @param[in] sum_col_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes) |
| 3055 | * @param[in] sum_col_step_y (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3056 | * @param[in] sum_col_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor |
| 3057 | * @param[in] sum_row_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr |
| 3058 | * @param[in] sum_row_stride_x (Optional) Stride of the source tensor in X dimension (in bytes) |
| 3059 | * @param[in] sum_row_step_x (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes) |
| 3060 | * @param[in] sum_row_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes) |
| 3061 | * @param[in] sum_row_step_y (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3062 | * @param[in] sum_row_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor |
| 3063 | * @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr |
| 3064 | * @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes) |
| 3065 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 3066 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor |
| 3067 | * @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8 |
| 3068 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 3069 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 3070 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 3071 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3072 | * @param[in] dst_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 3073 | * @param[in] dst_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 3074 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 3075 | */ |
| 3076 | __kernel void gemmlowp_offset_contribution_quantize_down_fixedpoint(TENSOR3D_DECLARATION(mm_result) |
| 3077 | #if defined(A_OFFSET) |
| 3078 | , |
| 3079 | IMAGE_DECLARATION(sum_col) |
| 3080 | #endif // defined(A_OFFSET) |
| 3081 | #if defined(B_OFFSET) |
| 3082 | , |
| 3083 | IMAGE_DECLARATION(sum_row) |
| 3084 | #endif // defined(B_OFFSET) |
| 3085 | , |
| 3086 | #if defined(ADD_BIAS) |
| 3087 | VECTOR_DECLARATION(biases), |
| 3088 | #endif // defined(ADD_BIAS) |
| 3089 | TENSOR3D_DECLARATION(dst)) |
| 3090 | { |
| 3091 | const int x = get_global_id(0) * 4; |
| 3092 | const int y = get_global_id(1); |
| 3093 | const int z = get_global_id(2); |
| 3094 | |
| 3095 | // Compute offset contribution |
| 3096 | int4 offset_term_s32 = offset_contribution( |
| 3097 | x, y, z |
| 3098 | #if defined(A_OFFSET) |
| 3099 | , |
| 3100 | sum_col_ptr, |
| 3101 | sum_col_stride_x, |
| 3102 | sum_col_step_x, |
| 3103 | sum_col_stride_y, |
| 3104 | sum_col_step_y, |
| 3105 | sum_col_offset_first_element_in_bytes |
| 3106 | #endif // defined(A_OFFSET) |
| 3107 | #if defined(B_OFFSET) |
| 3108 | , |
| 3109 | sum_row_ptr, |
| 3110 | sum_row_stride_x, |
| 3111 | sum_row_step_x, |
| 3112 | sum_row_stride_y, |
| 3113 | sum_row_step_y, |
| 3114 | sum_row_offset_first_element_in_bytes |
| 3115 | #endif // defined(B_OFFSET) |
| 3116 | #if defined(ADD_BIAS) |
| 3117 | , |
| 3118 | biases_ptr, |
| 3119 | biases_stride_x, |
| 3120 | biases_step_x, |
| 3121 | biases_offset_first_element_in_bytes |
| 3122 | #endif // defined(ADD_BIAS) |
| 3123 | ); |
| 3124 | |
| 3125 | __global uchar *mm_result_addr = mm_result_ptr + mm_result_offset_first_element_in_bytes + x * sizeof(int) + y * mm_result_stride_y + z * mm_result_stride_z; |
| 3126 | |
| 3127 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z; |
| 3128 | |
| 3129 | int4 in_s32 = vload4(0, (__global int *)mm_result_addr); |
| 3130 | |
| 3131 | // Add the offset terms to GEMM's result |
| 3132 | in_s32 += offset_term_s32; |
| 3133 | |
| 3134 | // -------------- OUTPUT STAGE |
| 3135 | |
| 3136 | // Multiply by result_mult_int and shift |
| 3137 | in_s32 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(in_s32, RESULT_MULTIPLIER, RESULT_SHIFT, 4); |
| 3138 | |
| 3139 | // Add the offset terms to GEMM's result |
| 3140 | in_s32 += (int4)RESULT_OFFSET; |
| 3141 | |
| 3142 | uchar4 res = convert_uchar4_sat(in_s32); |
| 3143 | |
| 3144 | #if defined(MIN_BOUND) |
| 3145 | res = max(res, (uchar4)MIN_BOUND); |
| 3146 | #endif // defined(MIN_BOUND) |
| 3147 | #if defined(MAX_BOUND) |
| 3148 | res = min(res, (uchar4)MAX_BOUND); |
| 3149 | #endif // defined(MAX_BOUND) |
| 3150 | |
| 3151 | // Store the result |
| 3152 | vstore4(res, 0, dst_addr); |
| 3153 | } |
| 3154 | #endif // defined(K_OFFSET) && defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 3155 | #endif // defined(K_OFFSET) |
| 3156 | |
| 3157 | #if defined(RESULT_OFFSET) && defined(RESULT_MULT_INT) && defined(RESULT_SHIFT) |
| 3158 | /** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8 |
| 3159 | * |
| 3160 | * This kernel takes a final int32 accumulator value and processes it to obtain the final QASYMM8 value. |
| 3161 | * The following computations will be performed by the kernel: |
| 3162 | * |
| 3163 | * -# Add offset terms to final result |
| 3164 | * -# Multiply each entry of result by result_mult_int |
| 3165 | * -# Add bias to final result (if -DADD_BIAS is passed at compile time) |
| 3166 | * -# Shift the int32 accumulator by result_shift |
| 3167 | * -# Clamp the value between the specified min and max bounds (if -DMIN_BOUND and/or -DMAX_BOUND are passed at compile time) |
| 3168 | * -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8. |
| 3169 | * |
| 3170 | * @attention The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_OFFSET, -RESULT_MULT_INT and -DRESULT_SHIFT |
| 3171 | * |
| 3172 | * @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time |
| 3173 | * @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND. |
| 3174 | * These values can be used to implement "rectified linear unit" activation functions |
| 3175 | * |
| 3176 | * @param[in] src_ptr Pointer to the source tensor. Supported data type: S32 |
| 3177 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 3178 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 3179 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 3180 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3181 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 3182 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 3183 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3184 | * @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr |
| 3185 | * @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes) |
| 3186 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 3187 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 3188 | * @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8 |
| 3189 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 3190 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 3191 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 3192 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3193 | * @param[in] dst_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 3194 | * @param[in] dst_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 3195 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 3196 | */ |
| 3197 | __kernel void gemmlowp_output_stage_quantize_down(TENSOR3D_DECLARATION(src), |
| 3198 | #if defined(ADD_BIAS) |
| 3199 | VECTOR_DECLARATION(biases), |
| 3200 | #endif // defined(ADD_BIAS) |
| 3201 | TENSOR3D_DECLARATION(dst)) |
| 3202 | { |
| 3203 | // Compute source and destination addresses |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3204 | int x = get_global_id(0) * 4; |
| 3205 | int y = get_global_id(1); |
| 3206 | int z = get_global_id(2); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 3207 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3208 | __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(int) + y * src_stride_y + z * src_stride_z; |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 3209 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3210 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z; |
| 3211 | |
| 3212 | int4 input_values = vload4(0, (__global int *)src_addr); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3213 | |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 3214 | #if defined(ADD_BIAS) |
| 3215 | // Add bias |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3216 | __global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(int); |
| 3217 | |
| 3218 | int4 biases_values = vload4(0, (__global int *)bias_addr); |
| 3219 | input_values += (int4)biases_values; |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 3220 | #endif // defined(ADD_BIAS) |
| 3221 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3222 | // Add the offset terms to GEMM's result |
| 3223 | input_values += (int4)RESULT_OFFSET; |
| 3224 | |
Georgios Pinitas | 45bcc3a | 2017-11-29 11:06:49 +0000 | [diff] [blame] | 3225 | // Multiply by result_mult_int and shift |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3226 | input_values *= RESULT_MULT_INT; |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 3227 | |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3228 | input_values >>= RESULT_SHIFT; |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 3229 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3230 | uchar4 res = convert_uchar4_sat(input_values); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 3231 | |
| 3232 | #if defined(MIN_BOUND) |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3233 | res = max(res, (uchar4)MIN_BOUND); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 3234 | #endif // defined(MIN_BOUND) |
| 3235 | #if defined(MAX_BOUND) |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3236 | res = min(res, (uchar4)MAX_BOUND); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 3237 | #endif // defined(MAX_BOUND) |
| 3238 | |
| 3239 | // Store the result |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3240 | vstore4(res, 0, dst_addr); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 3241 | } |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3242 | #endif // defined(RESULT_OFFSET) && defined(RESULT_MULT_INT) && defined(RESULT_SHIFT) |
| 3243 | |
| 3244 | #if defined(RESULT_OFFSET_AFTER_SHIFT) && defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT) |
| 3245 | /** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8 |
| 3246 | * |
| 3247 | * This kernel takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 value. |
| 3248 | * The following computations will be performed by the kernel: |
| 3249 | * |
| 3250 | * -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier |
| 3251 | * -# Add bias to final result if bias tensor is not a nullptr |
| 3252 | * -# Round to nearest division by a power-of-two using result_shift |
| 3253 | * -# Add offset to each result |
| 3254 | * -# Clamp the value between the specified min and max bounds |
| 3255 | * -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8. |
| 3256 | * |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3257 | * @attention The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_OFFSET_AFTER_SHIFT, -DRESULT_FIXEDPOINT_MULTIPLIER and -DRESULT_SHIFT |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3258 | * |
| 3259 | * @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time |
| 3260 | * @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND. |
| 3261 | * These values can be used to implement "rectified linear unit" activation functions |
| 3262 | * |
| 3263 | * @param[in] src_ptr Pointer to the source tensor. Supported data type: S32 |
| 3264 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 3265 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 3266 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 3267 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3268 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 3269 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 3270 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3271 | * @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr |
| 3272 | * @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes) |
| 3273 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 3274 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3275 | * @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8 |
| 3276 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 3277 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 3278 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 3279 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3280 | * @param[in] dst_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 3281 | * @param[in] dst_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 3282 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 3283 | */ |
| 3284 | __kernel void gemmlowp_output_stage_quantize_down_fixedpoint(TENSOR3D_DECLARATION(src), |
| 3285 | #if defined(ADD_BIAS) |
| 3286 | VECTOR_DECLARATION(biases), |
| 3287 | #endif // defined(ADD_BIAS) |
| 3288 | TENSOR3D_DECLARATION(dst)) |
| 3289 | { |
| 3290 | // Compute source and destination addresses |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3291 | int x = get_global_id(0) * 4; |
| 3292 | int y = get_global_id(1); |
| 3293 | int z = get_global_id(2); |
Georgios Pinitas | 932491f | 2018-09-21 16:33:15 +0100 | [diff] [blame] | 3294 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3295 | __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(int) + y * src_stride_y + z * src_stride_z; |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3296 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3297 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z; |
| 3298 | |
| 3299 | int4 input_values = vload4(0, (__global int *)src_addr); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3300 | |
| 3301 | #if defined(ADD_BIAS) |
| 3302 | // Add bias |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3303 | __global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(int); |
| 3304 | |
| 3305 | int4 biases_values = vload4(0, (__global int *)bias_addr); |
| 3306 | input_values += (int4)biases_values; |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3307 | #endif // defined(ADD_BIAS) |
| 3308 | |
| 3309 | // Multiply by result_mult_int and shift |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3310 | input_values = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(input_values, RESULT_FIXEDPOINT_MULTIPLIER, RESULT_SHIFT, 4); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3311 | |
| 3312 | // Add the offset terms to GEMM's result |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3313 | input_values += (int4)RESULT_OFFSET_AFTER_SHIFT; |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3314 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3315 | uchar4 res = convert_uchar4_sat(input_values); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3316 | |
| 3317 | #if defined(MIN_BOUND) |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3318 | res = max(res, (uchar4)MIN_BOUND); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3319 | #endif // defined(MIN_BOUND) |
| 3320 | #if defined(MAX_BOUND) |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3321 | res = min(res, (uchar4)MAX_BOUND); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3322 | #endif // defined(MAX_BOUND) |
| 3323 | |
| 3324 | // Store the result |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 3325 | vstore4(res, 0, dst_addr); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 3326 | } |
Chunosov | 5124be5 | 2017-11-22 20:42:13 +0700 | [diff] [blame] | 3327 | #endif // defined(RESULT_OFFSET_AFTER_SHIFT) && defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT) |
Georgios Pinitas | 51e53a3 | 2018-10-22 13:49:08 +0100 | [diff] [blame] | 3328 | |
| 3329 | #if defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET) |
| 3330 | /** This OpenCL kernel is used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8 |
| 3331 | * |
| 3332 | * This kernel takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 value. |
| 3333 | * The following computations will be performed by the kernel: |
| 3334 | * |
| 3335 | * -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier |
| 3336 | * -# Add bias to final result if bias tensor is not a nullptr |
| 3337 | * -# Requantize |
| 3338 | * -# Add offset to each result |
| 3339 | * -# Clamp the value between the specified min and max bounds |
| 3340 | * -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8. |
| 3341 | * |
| 3342 | * @attention The offset and scalar scale factor must be passed at compile time using -DRESULT_OFFSET, -DREAL_MULTIPLIER |
| 3343 | * |
| 3344 | * @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time |
| 3345 | * @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND. |
| 3346 | * These values can be used to implement "rectified linear unit" activation functions |
| 3347 | * |
| 3348 | * @param[in] src_ptr Pointer to the source tensor. Supported data type: S32 |
| 3349 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 3350 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 3351 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 3352 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3353 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 3354 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 3355 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 3356 | * @param[in] biases_ptr Pointer to the biases tensor. Supported data type: same as @p src_ptr |
| 3357 | * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) |
| 3358 | * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 3359 | * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor |
| 3360 | * @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8 |
| 3361 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 3362 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 3363 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 3364 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 3365 | * @param[in] dst_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 3366 | * @param[in] dst_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 3367 | * @param[in] dst_stride_w Stride of the source tensor in W dimension (in bytes) |
| 3368 | * @param[in] dst_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 3369 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 3370 | */ |
| 3371 | __kernel void gemmlowp_output_stage_quantize_down_float(TENSOR3D_DECLARATION(src), |
| 3372 | #if defined(ADD_BIAS) |
| 3373 | VECTOR_DECLARATION(biases), |
| 3374 | #endif // defined(ADD_BIAS) |
| 3375 | #if defined(DST_HEIGHT) |
| 3376 | TENSOR4D_DECLARATION(dst)) |
| 3377 | #else // defined(DST_HEIGHT) |
| 3378 | TENSOR3D_DECLARATION(dst)) |
| 3379 | #endif // defined(DST_HEIGHT) |
| 3380 | { |
| 3381 | // Compute source and destination addresses |
Gian Marco Iodice | 0c54a62 | 2018-10-30 12:20:03 +0000 | [diff] [blame] | 3382 | int x = get_global_id(0) * 4; |
| 3383 | int y = get_global_id(1); |
| 3384 | int z = get_global_id(2); |
Georgios Pinitas | 51e53a3 | 2018-10-22 13:49:08 +0100 | [diff] [blame] | 3385 | |
Gian Marco Iodice | 0c54a62 | 2018-10-30 12:20:03 +0000 | [diff] [blame] | 3386 | __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(int) + y * src_stride_y + z * src_stride_z; |
Georgios Pinitas | 51e53a3 | 2018-10-22 13:49:08 +0100 | [diff] [blame] | 3387 | |
Gian Marco Iodice | 0c54a62 | 2018-10-30 12:20:03 +0000 | [diff] [blame] | 3388 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z; |
| 3389 | |
| 3390 | int4 input_values = vload4(0, (__global int *)src_addr); |
Georgios Pinitas | 51e53a3 | 2018-10-22 13:49:08 +0100 | [diff] [blame] | 3391 | |
| 3392 | #if defined(ADD_BIAS) |
| 3393 | // Add bias |
Gian Marco Iodice | 0c54a62 | 2018-10-30 12:20:03 +0000 | [diff] [blame] | 3394 | __global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(int); |
| 3395 | |
| 3396 | int4 biases_values = vload4(0, (__global int *)bias_addr); |
| 3397 | input_values += (int4)biases_values; |
Georgios Pinitas | 51e53a3 | 2018-10-22 13:49:08 +0100 | [diff] [blame] | 3398 | #endif // defined(ADD_BIAS) |
| 3399 | |
| 3400 | // Convert to float |
Gian Marco Iodice | 0c54a62 | 2018-10-30 12:20:03 +0000 | [diff] [blame] | 3401 | float16 input_values_f = convert_float4(input_values); |
Georgios Pinitas | 51e53a3 | 2018-10-22 13:49:08 +0100 | [diff] [blame] | 3402 | input_values_f = round(input_values_f * (float)REAL_MULTIPLIER + (float)OUTPUT_OFFSET); |
| 3403 | |
Gian Marco Iodice | 0c54a62 | 2018-10-30 12:20:03 +0000 | [diff] [blame] | 3404 | uchar4 res = convert_uchar4_sat(input_values_f); |
Georgios Pinitas | 51e53a3 | 2018-10-22 13:49:08 +0100 | [diff] [blame] | 3405 | |
| 3406 | #if defined(MIN_BOUND) |
Gian Marco Iodice | 0c54a62 | 2018-10-30 12:20:03 +0000 | [diff] [blame] | 3407 | res = max(res, (uchar4)MIN_BOUND); |
Georgios Pinitas | 51e53a3 | 2018-10-22 13:49:08 +0100 | [diff] [blame] | 3408 | #endif // defined(MIN_BOUND) |
| 3409 | #if defined(MAX_BOUND) |
Gian Marco Iodice | 0c54a62 | 2018-10-30 12:20:03 +0000 | [diff] [blame] | 3410 | res = min(res, (uchar4)MAX_BOUND); |
Georgios Pinitas | 51e53a3 | 2018-10-22 13:49:08 +0100 | [diff] [blame] | 3411 | #endif // defined(MAX_BOUND) |
| 3412 | |
| 3413 | // Store the result |
Gian Marco Iodice | 0c54a62 | 2018-10-30 12:20:03 +0000 | [diff] [blame] | 3414 | vstore4(res, 0, dst_addr); |
Georgios Pinitas | 51e53a3 | 2018-10-22 13:49:08 +0100 | [diff] [blame] | 3415 | } |
Gian Marco Iodice | db18a6f | 2019-05-30 09:53:10 +0100 | [diff] [blame] | 3416 | #endif // defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET) |