Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 1 | /* |
Stephen Li | e855c23 | 2018-01-04 14:13:22 +0800 | [diff] [blame] | 2 | * Copyright (c) 2017-2018 ARM Limited. |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in; |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 25 | #include "helpers_cs.h" |
| 26 | |
| 27 | #if defined(DATA_TYPE_FP16) |
| 28 | precision mediump float; |
| 29 | #endif // DATA_TYPE_FP16 |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 30 | |
| 31 | #if defined(DATA_TYPE_FP32) |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 32 | #ifdef GEMM_TRANSPOSE1xW |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 33 | /** This OpenGL ES kernel computes the "vector" 1x4 transposition of input matrix |
| 34 | * |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 35 | * @param[in] src_ptr Pointer to the source matrix. Supported data types: F32 |
| 36 | * @param[in] src_attrs The attributes of the source matrix |
| 37 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr |
| 38 | * @param[in] dst_attrs The attributes of the destination matrix |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 39 | */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 40 | SHADER_PARAMS_DECLARATION |
| 41 | { |
| 42 | ImageAttributes src_attrs; |
| 43 | ImageAttributes dst_attrs; |
| 44 | }; |
| 45 | TENSOR_DECLARATION(1, srcBuffer, float, src_ptr, src_shift, 2, readonly); |
| 46 | TENSOR_DECLARATION(2, dstBuffer, float, dst_ptr, dst_shift, 2, writeonly); |
| 47 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 48 | void main(void) |
| 49 | { |
| 50 | /* Compute address for Matrix B - source */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 51 | ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); |
| 52 | ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(dst_attrs, dst_shift); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 53 | |
| 54 | /* Compute address for Matrix B transposed - destination. X and Y are swapped */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 55 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(dst_iter, gl_GlobalInvocationID.y * uint(16) + gl_GlobalInvocationID.x * dst_attrs.stride_y); |
| 56 | |
| 57 | vec4 b0 = VLOAD4_CURRENT_ITEM(vec4, src_ptr, src_iter); |
| 58 | VSTORE4_CURRENT_ITEM(dst_ptr, dst_iter, b0); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 59 | } |
| 60 | #endif /* GEMM_TRANSPOSE1xW */ |
| 61 | |
| 62 | #ifdef GEMM_INTERLEAVE4x4 |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 63 | /** This OpenGLES kernel reshapes the input matrix interleaving the values |
| 64 | * |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 65 | * @param[in] src_ptr Pointer to the source matrix. Supported data types: F32 |
| 66 | * @param[in] src_attrs The attributes of the source matrix |
| 67 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr |
| 68 | * @param[in] dst_attrs The attributes of the destination matrix |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 69 | */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 70 | SHADER_PARAMS_DECLARATION |
| 71 | { |
| 72 | ImageAttributes src_attrs; |
| 73 | ImageAttributes dst_attrs; |
| 74 | }; |
| 75 | TENSOR_DECLARATION(1, srcBuffer, float, src_ptr, src_shift, 2, readonly); |
| 76 | TENSOR_DECLARATION(2, dstBuffer, float, dst_ptr, dst_shift, 2, writeonly); |
| 77 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 78 | void main(void) |
| 79 | { |
| 80 | /* Compute source and destination addresses */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 81 | ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); |
| 82 | ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR(dst_attrs, dst_shift); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 83 | |
| 84 | int i; |
| 85 | int j; |
| 86 | |
| 87 | for(i = 0; i < 4; ++i) |
| 88 | { |
| 89 | for(j = 0; j < 4; ++j) |
| 90 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 91 | float res = LOAD(src_ptr, IMAGE_OFFSET(src_iter, i, j)); |
| 92 | STORE(dst_ptr, TENSOR_OFFSET_ADVANCE(dst_iter, (i * 4 + j)), res); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 93 | } |
| 94 | } |
| 95 | } |
| 96 | #endif /* GEMM_INTERLEAVE4x4 */ |
| 97 | |
| 98 | #ifdef GEMM_ACCUMULATE_BIASES |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 99 | /** This kernel accumulates each row with the biases vector |
| 100 | * |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 101 | * @param[in, out] accum_ptr Pointer to the accumulate tensor. Supported data type: F32 |
| 102 | * @param[in] accum_attrs The attributes of the accumulate tensor |
| 103 | * @param[in] biases_ptr Pointer to the biases vector. Same as @p accum_ptr |
| 104 | * @param[in] biases_attrs The attributes of the biases tensor |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 105 | */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 106 | SHADER_PARAMS_DECLARATION |
| 107 | { |
| 108 | ImageAttributes accum_attrs; |
| 109 | VectorAttributes biases_attrs; |
| 110 | }; |
| 111 | TENSOR_DECLARATION(1, accumBuffer, float, accum_ptr, accum_shift, 2, restrict); |
| 112 | TENSOR_DECLARATION(2, biasesBuffer, float, biases_ptr, biases_shift, 2, readonly); |
| 113 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 114 | void main(void) |
| 115 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 116 | ImageIterator accum_iter = CONVERT_TO_IMAGE_ITERATOR(accum_attrs, accum_shift); |
| 117 | VectorIterator biases_iter = CONVERT_TO_VECTOR_ITERATOR(biases_attrs, biases_shift); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 118 | |
| 119 | for(int i = 0; i < 16; ++i) |
| 120 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 121 | float accum_value = LOAD(accum_ptr, TENSOR_OFFSET_ADVANCE(accum_iter, i)); |
| 122 | float biases_value = LOAD(biases_ptr, TENSOR_OFFSET_ADVANCE(biases_iter, i)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 123 | accum_value = biases_value + accum_value; |
| 124 | |
| 125 | // Store result in the accummulate buffer |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 126 | STORE(accum_ptr, TENSOR_OFFSET_ADVANCE(accum_iter, i), accum_value); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 127 | } |
| 128 | } |
| 129 | #endif /* GEMM_ACCUMULATE_BIASES */ |
| 130 | |
| 131 | #ifdef GEMM_MM_INTERLEAVED_TRANSPOSED /* unvalidate */ |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 132 | /** This OpenGL ES kernel is optimised for Midgard. It computes the matrix multiplication between matrix A (src0) and matrix B (src1) |
| 133 | * Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_32bit and @ref gemm_transpose1x4 before running the matrix multiplication |
| 134 | * |
| 135 | * @attention The width of matrix B and the alpha's value need to be passed at compile time using WIDTH_MATRIX_B and ALPHA |
| 136 | * |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 137 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 |
| 138 | * @param[in] src0_attrs The attributes of the source matrix |
| 139 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
| 140 | * @param[in] src1_attrs The attributes of the source matrix |
| 141 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
| 142 | * @param[in] dst_attrs The attributes of the destination matrix |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 143 | */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 144 | SHADER_PARAMS_DECLARATION |
| 145 | { |
| 146 | ImageAttributes src0_attrs; |
| 147 | ImageAttributes src1_attrs; |
| 148 | ImageAttributes dst_attrs; |
| 149 | }; |
| 150 | TENSOR_DECLARATION(1, src0Buffer, float, src0_ptr, src0_shift, 2, readonly); |
| 151 | TENSOR_DECLARATION(2, src1Buffer, float, src1_ptr, src1_shift, 2, readonly); |
| 152 | TENSOR_DECLARATION(3, dstBuffer, float, dst_ptr, dst_shift, 2, writeonly); |
| 153 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 154 | void main() |
| 155 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 156 | ImageIterator src0_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(src0_attrs, src0_shift); |
| 157 | ImageIterator src1_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(src1_attrs, src1_shift); |
| 158 | ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR(dst_attrs, dst_shift); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 159 | |
| 160 | /* Compute address for matrix A and B */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 161 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, uint(gl_GlobalInvocationID.y) * (src0_attrs.stride_y)); |
| 162 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, uint(gl_GlobalInvocationID.x) * (src1_attrs.stride_y)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 163 | /* Compute end row address for matrix B */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 164 | int end_row_mtx_b = int(TENSOR_OFFSET_ADVANCE(src1_iter, COLS_B)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 165 | |
| 166 | /* Reset accumulators */ |
| 167 | vec4 c00 = vec4(0.0f); |
| 168 | vec4 c10 = vec4(0.0f); |
| 169 | vec4 c20 = vec4(0.0f); |
| 170 | vec4 c30 = vec4(0.0f); |
| 171 | |
| 172 | // FIXME: loop unrolling really needed for GLES? |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 173 | for(; int(CURRENT_ITEM_OFFSET(src1_iter)) <= (end_row_mtx_b - 8); TENSOR_ITERATOR_ADVANCE(src0_iter, 8), TENSOR_ITERATOR_ADVANCE(src1_iter, 8)) |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 174 | { |
| 175 | /* Load values from matrix A (interleaved) and matrix B (transposed) */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 176 | vec4 a0 = VLOAD4_CURRENT_ITEM(vec4, src0_ptr, src0_iter); |
| 177 | vec4 b0 = VLOAD4_CURRENT_ITEM(vec4, src1_ptr, src1_iter); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 178 | |
| 179 | c00 += vec4(a0.x) * b0; |
| 180 | c10 += vec4(a0.y) * b0; |
| 181 | c20 += vec4(a0.z) * b0; |
| 182 | c30 += vec4(a0.w) * b0; |
| 183 | |
| 184 | /* Load values from matrix A (interleaved) and matrix B (transposed) */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 185 | a0 = VLOAD4(vec4, src0_ptr, TENSOR_OFFSET_ADVANCE(src0_iter, 4)); |
| 186 | b0 = VLOAD4(vec4, src1_ptr, TENSOR_OFFSET_ADVANCE(src1_iter, 4)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 187 | |
| 188 | c00 += vec4(a0.x) * b0; |
| 189 | c10 += vec4(a0.y) * b0; |
| 190 | c20 += vec4(a0.z) * b0; |
| 191 | c30 += vec4(a0.w) * b0; |
| 192 | } |
| 193 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 194 | for(; int(CURRENT_ITEM_OFFSET(src1_iter)) < end_row_mtx_b; TENSOR_ITERATOR_ADVANCE(src0_iter, 4), TENSOR_ITERATOR_ADVANCE(src1_iter, 4)) |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 195 | { |
| 196 | /* Load values from matrix A (interleaved) and matrix B (transposed) */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 197 | vec4 a0 = VLOAD4_CURRENT_ITEM(vec4, src0_ptr, src0_iter); |
| 198 | vec4 b0 = VLOAD4_CURRENT_ITEM(vec4, src1_ptr, src1_iter); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 199 | |
| 200 | c00 += vec4(a0.x) * b0; |
| 201 | c10 += vec4(a0.y) * b0; |
| 202 | c20 += vec4(a0.z) * b0; |
| 203 | c30 += vec4(a0.w) * b0; |
| 204 | } |
| 205 | |
| 206 | /* Multiply by the weight of matrix product */ |
| 207 | c00 = c00 * vec4(ALPHA); |
| 208 | c10 = c10 * vec4(ALPHA); |
| 209 | c20 = c20 * vec4(ALPHA); |
| 210 | c30 = c30 * vec4(ALPHA); |
| 211 | |
| 212 | /* Store 4x4 block */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 213 | VSTORE4(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 0), c00); |
| 214 | VSTORE4(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 1), c10); |
| 215 | VSTORE4(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 2), c20); |
| 216 | VSTORE4(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 3), c30); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 217 | } |
| 218 | #endif /* GEMM_MM_INTERLEAVED_TRANSPOSED */ |
| 219 | |
| 220 | #ifdef GEMM_MM_FLOATING_POINT |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 221 | /** This OpenGL ES kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1) |
| 222 | * Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_32bit and @ref gemm_transpose1x4 before running the matrix multiplication |
| 223 | * |
| 224 | * @attention The width of matrix B and the alpha's value need to be passed at compile time using WIDTH_MATRIX_B and ALPHA |
| 225 | * |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 226 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 |
| 227 | * @param[in] src0_attrs The attributes of the source matrix |
| 228 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
| 229 | * @param[in] src1_attrs The attributes of the source matrix |
| 230 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
| 231 | * @param[in] dst_attrs The attributes of the destination matrix |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 232 | */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 233 | SHADER_PARAMS_DECLARATION |
| 234 | { |
| 235 | ImageAttributes src0_attrs; |
| 236 | ImageAttributes src1_attrs; |
| 237 | ImageAttributes dst_attrs; |
| 238 | }; |
| 239 | TENSOR_DECLARATION(1, src0Buffer, float, src0_ptr, src0_shift, 2, readonly); |
| 240 | TENSOR_DECLARATION(2, src1Buffer, float, src1_ptr, src1_shift, 2, readonly); |
| 241 | TENSOR_DECLARATION(3, dstBuffer, float, dst_ptr, dst_shift, 2, writeonly); |
| 242 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 243 | void main() |
| 244 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 245 | ImageIterator src0_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(src0_attrs, src0_shift); |
| 246 | ImageIterator src1_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(src1_attrs, src1_shift); |
| 247 | ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR(dst_attrs, dst_shift); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 248 | |
| 249 | int idx = int(gl_GlobalInvocationID.x) * int(NUM_ELEMS_PROCESSED_PER_THREAD_X); |
| 250 | /* Compute the address for the vector A and matrix B */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 251 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, uint(gl_GlobalInvocationID.y) * (src0_attrs.stride_y) * uint(NUM_ELEMS_PROCESSED_PER_THREAD_Y)); |
| 252 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, idx * 4); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 253 | |
| 254 | /* Compute end row address for matrix A */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 255 | int end_row_vec_a = int(TENSOR_OFFSET_ADVANCE_IN_BYTES(src0_iter, COLS_A * 4)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 256 | |
| 257 | /* Reset accumulators */ |
| 258 | vec4 acc0 = vec4(0.0f); |
| 259 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 260 | vec4 acc1 = vec4(0.0f); |
| 261 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 262 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 263 | vec4 acc2 = vec4(0.0f); |
| 264 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 265 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 266 | vec4 acc3 = vec4(0.0f); |
| 267 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 268 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 269 | for(; int(CURRENT_ITEM_OFFSET(src0_iter)) <= (end_row_vec_a - 2); TENSOR_ITERATOR_ADVANCE(src0_iter, 2), TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, uint(2) * src1_attrs.stride_y)) |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 270 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 271 | vec2 a0 = VLOAD2_CURRENT_ITEM(vec2, src0_ptr, src0_iter); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 272 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 273 | vec2 a1 = VLOAD2(vec2, src0_ptr, IMAGE_OFFSET(src0_iter, 0, 1)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 274 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 275 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 276 | vec2 a2 = VLOAD2(vec2, src0_ptr, IMAGE_OFFSET(src0_iter, 0, 2)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 277 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 278 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 279 | vec2 a3 = VLOAD2(vec2, src0_ptr, IMAGE_OFFSET(src0_iter, 0, 3)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 280 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 281 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 282 | vec4 b0 = VLOAD4_CURRENT_ITEM(vec4, src1_ptr, src1_iter); |
| 283 | vec4 b1 = VLOAD4(vec4, src1_ptr, IMAGE_OFFSET(src1_iter, 0, 1)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 284 | |
| 285 | acc0 += b0 * vec4(a0.x); |
| 286 | acc0 += b1 * vec4(a0.y); |
| 287 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 288 | acc1 += b0 * vec4(a1.x); |
| 289 | acc1 += b1 * vec4(a1.y); |
| 290 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 291 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 292 | acc2 += b0 * vec4(a2.x); |
| 293 | acc2 += b1 * vec4(a2.y); |
| 294 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 295 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 296 | acc3 += b0 * vec4(a3.x); |
| 297 | acc3 += b1 * vec4(a3.y); |
| 298 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 299 | } |
| 300 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 301 | for(; int(CURRENT_ITEM_OFFSET(src0_iter)) < end_row_vec_a; TENSOR_ITERATOR_ADVANCE(src0_iter, 1), TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, src1_attrs.stride_y)) |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 302 | { |
| 303 | // Load values from matrix A |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 304 | float a0 = LOAD_CURRENT_ITEM(src0_ptr, src0_iter); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 305 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 306 | float a1 = LOAD(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 1)); |
| 307 | //float a1 = 0; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 308 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 309 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 310 | float a2 = LOAD(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 2)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 311 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 312 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 313 | float a3 = LOAD(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 3)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 314 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 315 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 316 | vec4 b0 = VLOAD4_CURRENT_ITEM(vec4, src1_ptr, src1_iter); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 317 | |
| 318 | acc0 += b0 * vec4(a0); |
| 319 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 320 | acc1 += b0 * vec4(a1); |
| 321 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 322 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 323 | acc2 += b0 * vec4(a2); |
| 324 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 325 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 326 | acc3 += b0 * vec4(a3); |
| 327 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 328 | } |
| 329 | |
| 330 | /* Multiply by the weight of vector-matrix product */ |
| 331 | acc0 = acc0 * vec4(ALPHA); |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 332 | VSTORE4_CURRENT_ITEM(dst_ptr, dst_iter, acc0); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 333 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 334 | acc1 = acc1 * vec4(ALPHA); |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 335 | VSTORE4(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 1), acc1); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 336 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 337 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 338 | acc2 = acc2 * vec4(ALPHA); |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 339 | VSTORE4(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 2), acc2); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 340 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 341 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 342 | acc3 = acc3 * vec4(ALPHA); |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 343 | VSTORE4(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 3), acc3); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 344 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 345 | } |
| 346 | #endif /* GEMM_MM_FLOATING_POINT */ |
| 347 | |
| 348 | #ifdef GEMM_MATRIXADDITION |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 349 | /** This OpenGL ES kernel performs the in-place matrix addition between 2 matrices taking into account that the second matrix might be weighted by a scalar value beta: |
| 350 | * |
| 351 | * @attention The beta's value need to be passed at compile time using BETA |
| 352 | * |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 353 | * @param[in] src_ptr Pointer to the source matrix. Supported data types: F32 |
| 354 | * @param[in] src_attrs The attributes of the source matrix |
| 355 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr |
| 356 | * @param[in] dst_attrs The attributes of the destination matrix |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 357 | */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 358 | SHADER_PARAMS_DECLARATION |
| 359 | { |
| 360 | ImageAttributes src_attrs; |
| 361 | ImageAttributes dst_attrs; |
| 362 | }; |
| 363 | TENSOR_DECLARATION(1, srcBuffer, float, src_ptr, src_shift, 2, readonly); |
| 364 | TENSOR_DECLARATION(2, dstBuffer, float, dst_ptr, dst_shift, 2, restrict); |
| 365 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 366 | void main(void) |
| 367 | { |
| 368 | /* Compute source and destination addresses */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 369 | ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); |
| 370 | ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR(dst_attrs, dst_shift); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 371 | |
| 372 | /* Load values from A x B */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 373 | vec4 alpha_ab = VLOAD4_CURRENT_ITEM(vec4, dst_ptr, dst_iter); |
| 374 | vec4 c = VLOAD4_CURRENT_ITEM(vec4, src_ptr, src_iter); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 375 | |
| 376 | /* Computes alpha * axb + beta * c */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 377 | vec4 out1 = alpha_ab + vec4(float(BETA) * c); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 378 | |
| 379 | /* Store final result in axb matrix */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 380 | VSTORE4_CURRENT_ITEM(dst_ptr, dst_iter, out1); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 381 | } |
| 382 | #endif /* GEMM_MATRIXADDITION */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 383 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 384 | #elif defined(DATA_TYPE_FP16) |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 385 | |
Stephen Li | e855c23 | 2018-01-04 14:13:22 +0800 | [diff] [blame] | 386 | #ifdef GEMM_TRANSPOSE1xW |
| 387 | /** This OpenGL ES kernel computes the "vector" 1x8 transposition of input matrix |
| 388 | * |
| 389 | * @param[in] src_ptr Pointer to the source matrix. Supported data types: F16 |
| 390 | * @param[in] src_attrs The attributes of the source matrix |
| 391 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr |
| 392 | * @param[in] dst_attrs The attributes of the destination matrix |
| 393 | */ |
| 394 | SHADER_PARAMS_DECLARATION |
| 395 | { |
| 396 | ImageAttributes src_attrs; |
| 397 | ImageAttributes dst_attrs; |
| 398 | }; |
| 399 | TENSOR_DECLARATION(1, srcBuffer, uvec4, src_ptr, src_shift, 4, readonly); |
| 400 | TENSOR_DECLARATION(2, dstBuffer, uvec4, dst_ptr, dst_shift, 4, writeonly); |
| 401 | |
| 402 | void main(void) |
| 403 | { |
| 404 | /* Compute address for Matrix B - source */ |
| 405 | ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); |
| 406 | ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(dst_attrs, dst_shift); |
| 407 | |
| 408 | /* Compute address for Matrix B transposed - destination. X and Y are swapped */ |
| 409 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(dst_iter, gl_GlobalInvocationID.y * uint(16) + gl_GlobalInvocationID.x * dst_attrs.stride_y); |
| 410 | |
| 411 | STORE_CURRENT_ITEM(dst_ptr, dst_iter, LOAD_CURRENT_ITEM(src_ptr, src_iter)); |
| 412 | } |
| 413 | #endif /* GEMM_TRANSPOSE1xW */ |
| 414 | |
| 415 | #ifdef GEMM_INTERLEAVE4x4 |
| 416 | /** This OpenGLES kernel reshapes the input matrix interleaving the values |
| 417 | * |
| 418 | * @param[in] src_ptr Pointer to the source matrix. Supported data types: F16 |
| 419 | * @param[in] src_attrs The attributes of the source matrix |
| 420 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr |
| 421 | * @param[in] dst_attrs The attributes of the destination matrix |
| 422 | */ |
| 423 | SHADER_PARAMS_DECLARATION |
| 424 | { |
| 425 | ImageAttributes src_attrs; |
| 426 | ImageAttributes dst_attrs; |
| 427 | }; |
| 428 | TENSOR_DECLARATION(1, srcBuffer, uvec4, src_ptr, src_shift, 4, readonly); |
| 429 | TENSOR_DECLARATION(2, dstBuffer, uvec4, dst_ptr, dst_shift, 4, writeonly); |
| 430 | |
| 431 | void main(void) |
| 432 | { |
| 433 | /* Compute source and destination addresses */ |
| 434 | ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); |
| 435 | ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR(dst_attrs, dst_shift); |
| 436 | |
| 437 | vec4 s0[2] = LOAD_UNPACK8_CURRENT_ITEM_HALF(src_ptr, src_iter); |
| 438 | vec4 s1[2] = LOAD_UNPACK8_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, 1)); |
| 439 | vec4 s2[2] = LOAD_UNPACK8_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, 2)); |
| 440 | vec4 s3[2] = LOAD_UNPACK8_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, 3)); |
| 441 | |
| 442 | vec4 s[2]; |
| 443 | s[0] = vec4(s0[0].x, s1[0].x, s2[0].x, s3[0].x); |
| 444 | s[1] = vec4(s0[0].y, s1[0].y, s2[0].y, s3[0].y); |
| 445 | STORE_PACK8_CURRENT_ITEM_HALF(dst_ptr, dst_iter, s); |
| 446 | |
| 447 | s[0] = vec4(s0[0].z, s1[0].z, s2[0].z, s3[0].z); |
| 448 | s[1] = vec4(s0[0].w, s1[0].w, s2[0].w, s3[0].w); |
| 449 | STORE_PACK8_HALF(dst_ptr, TENSOR_OFFSET_ADVANCE(dst_iter, 1u), s); |
| 450 | |
| 451 | s[0] = vec4(s0[1].x, s1[1].x, s2[1].x, s3[1].x); |
| 452 | s[1] = vec4(s0[1].y, s1[1].y, s2[1].y, s3[1].y); |
| 453 | STORE_PACK8_HALF(dst_ptr, TENSOR_OFFSET_ADVANCE(dst_iter, 2u), s); |
| 454 | |
| 455 | s[0] = vec4(s0[1].z, s1[1].z, s2[1].z, s3[1].z); |
| 456 | s[1] = vec4(s0[1].w, s1[1].w, s2[1].w, s3[1].w); |
| 457 | STORE_PACK8_HALF(dst_ptr, TENSOR_OFFSET_ADVANCE(dst_iter, 3u), s); |
| 458 | } |
| 459 | #endif /* GEMM_INTERLEAVE4x4 */ |
| 460 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 461 | #ifdef GEMM_MM_FLOATING_POINT |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 462 | /** This OpenGL ES kernel computes the matrix multiplication between matrix A(src0) and matrix B(src1) |
| 463 | * Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_16bit and @ref gemm_transpose1x4 before running the matrix multiplication |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 464 | * |
| 465 | * @attention The width of matrix B and the alpha's value need to be passed at compile time using WIDTH_MATRIX_B and ALPHA |
| 466 | * |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 467 | * @param[in] src0_ptr Pointer to the source matrix.Supported data types: F16 |
| 468 | * @param[in] src0_attrs The attributes of the source matrix |
| 469 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
| 470 | * @param[in] src1_attrs The attributes of the source matrix |
| 471 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
| 472 | * @param[in] dst_attrs The attributes of the destination matrix |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 473 | */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 474 | SHADER_PARAMS_DECLARATION |
| 475 | { |
| 476 | ImageAttributes src0_attrs; |
| 477 | ImageAttributes src1_attrs; |
| 478 | ImageAttributes dst_attrs; |
| 479 | }; |
| 480 | |
| 481 | #if defined(MM_PROCESS_4X) |
| 482 | TENSOR_DECLARATION(1, src0Buffer, uint, src0_ptr, src0_shift, 2, readonly); |
| 483 | TENSOR_DECLARATION(2, src1Buffer, uvec2, src1_ptr, src1_shift, 3, readonly); |
| 484 | TENSOR_DECLARATION(3, dstBuffer, uvec2, dst_ptr, dst_shift, 3, writeonly); |
| 485 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 486 | void main() |
| 487 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 488 | ImageIterator src0_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(src0_attrs, src0_shift); |
| 489 | ImageIterator src1_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(src1_attrs, src1_shift); |
| 490 | ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR(dst_attrs, dst_shift); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 491 | |
| 492 | int idx = int(gl_GlobalInvocationID.x) * int(NUM_ELEMS_PROCESSED_PER_THREAD_X); |
| 493 | /* Compute the address for the vector A and matrix B */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 494 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, uint(gl_GlobalInvocationID.y) * src0_attrs.stride_y * uint(NUM_ELEMS_PROCESSED_PER_THREAD_Y)); |
| 495 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, uint(idx) * src1_attrs.stride_x); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 496 | |
| 497 | /* Compute end row address for matrix A */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 498 | uint end_row_vec_a = uint(CURRENT_ITEM_OFFSET_IN_BYTES(src0_iter)) + uint(COLS_A << 1); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 499 | |
| 500 | /* Reset accumulators */ |
| 501 | vec4 acc0 = vec4(0.0f); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 502 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 503 | vec4 acc1 = vec4(0.0f); |
| 504 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 505 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 506 | vec4 acc2 = vec4(0.0f); |
| 507 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 508 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 509 | vec4 acc3 = vec4(0.0f); |
| 510 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 511 | |
zhenglin | 2219dea | 2018-01-30 18:15:52 +0800 | [diff] [blame] | 512 | for(; int(CURRENT_ITEM_OFFSET_IN_BYTES(src0_iter)) <= int(end_row_vec_a - uint(4)); |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 513 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, 2 * 2), TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, uint(2) * src1_attrs.stride_y)) |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 514 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 515 | vec2 a0 = LOAD_UNPACK2_CURRENT_ITEM_HALF(src0_ptr, src0_iter); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 516 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 517 | vec2 a1 = LOAD_UNPACK2_HALF(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 1)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 518 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 519 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 520 | vec2 a2 = LOAD_UNPACK2_HALF(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 2)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 521 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 522 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 523 | vec2 a3 = LOAD_UNPACK2_HALF(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 3)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 524 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 525 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 526 | vec4 b0 = LOAD_UNPACK4_CURRENT_ITEM_HALF(src1_ptr, src1_iter); |
| 527 | vec4 b1 = LOAD_UNPACK4_HALF(src1_ptr, IMAGE_OFFSET(src1_iter, 0, 1)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 528 | |
| 529 | acc0 += b0 * vec4(a0.x); |
| 530 | acc0 += b1 * vec4(a0.y); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 531 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 532 | acc1 += b0 * vec4(a1.x); |
| 533 | acc1 += b1 * vec4(a1.y); |
| 534 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 535 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 536 | acc2 += b0 * vec4(a2.x); |
| 537 | acc2 += b1 * vec4(a2.y); |
| 538 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 539 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 540 | acc3 += b0 * vec4(a3.x); |
| 541 | acc3 += b1 * vec4(a3.y); |
| 542 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 543 | } |
| 544 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 545 | for(; int(CURRENT_ITEM_OFFSET_IN_BYTES(src0_iter)) < int(end_row_vec_a); TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, 2 * 2), TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, src1_attrs.stride_y)) |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 546 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 547 | vec2 a0 = LOAD_UNPACK2_CURRENT_ITEM_HALF(src0_ptr, src0_iter); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 548 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 549 | vec2 a1 = LOAD_UNPACK2_HALF(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 1)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 550 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 551 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
zhenglin | 2219dea | 2018-01-30 18:15:52 +0800 | [diff] [blame] | 552 | vec2 a2 = LOAD_UNPACK2_HALF(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 2)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 553 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 554 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 555 | vec2 a3 = LOAD_UNPACK2_HALF(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 3)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 556 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 557 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 558 | vec4 b0 = LOAD_UNPACK4_CURRENT_ITEM_HALF(src1_ptr, src1_iter); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 559 | |
| 560 | acc0 += b0 * (a0.x); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 561 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 562 | acc1 += b0 * (a1.x); |
| 563 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 564 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 565 | acc2 += b0 * (a2.x); |
| 566 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 567 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 568 | acc3 += b0 * (a3.x); |
| 569 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 570 | } |
| 571 | |
| 572 | /* Multiply by the weight of vector-matrix product */ |
| 573 | acc0 = acc0 * vec4(ALPHA); |
| 574 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 575 | STORE_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, acc0); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 576 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 577 | STORE_PACK4_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 1), acc1); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 578 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 579 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 580 | STORE_PACK4_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 2), acc2); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 581 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 582 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 583 | STORE_PACK4_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 3), acc3); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 584 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 585 | } |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 586 | #elif defined(MM_PROCESS_4X_OPTIMIZED) /* PROCESS_4X */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 587 | TENSOR_DECLARATION(1, src0Buffer, uvec4, src0_ptr, src0_shift, 4, readonly); |
| 588 | TENSOR_DECLARATION(2, src1Buffer, uvec2, src1_ptr, src1_shift, 3, readonly); |
| 589 | TENSOR_DECLARATION(3, dstBuffer, uvec2, dst_ptr, dst_shift, 3, writeonly); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 590 | |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 591 | void main() |
| 592 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 593 | ImageIterator src0_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(src0_attrs, src0_shift); |
| 594 | ImageIterator src1_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(src1_attrs, src1_shift); |
| 595 | ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR(dst_attrs, dst_shift); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 596 | |
| 597 | int idx = int(gl_GlobalInvocationID.x) * int(NUM_ELEMS_PROCESSED_PER_THREAD_X); |
| 598 | /* Compute the address for the vector A and matrix B */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 599 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, uint(gl_GlobalInvocationID.y) * src0_attrs.stride_y * uint(NUM_ELEMS_PROCESSED_PER_THREAD_Y)); |
| 600 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, uint(idx) * src1_attrs.stride_x); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 601 | |
| 602 | /* Compute end row address for matrix A */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 603 | uint end_row_vec_a = uint(CURRENT_ITEM_OFFSET_IN_BYTES(src0_iter)) + uint(COLS_A << 1); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 604 | |
| 605 | /* Reset accumulators */ |
| 606 | vec4 acc0 = vec4(0.0f); |
| 607 | |
| 608 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 609 | vec4 acc1 = vec4(0.0f); |
| 610 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 611 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 612 | vec4 acc2 = vec4(0.0f); |
| 613 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 614 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 615 | vec4 acc3 = vec4(0.0f); |
| 616 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 617 | |
zhenglin | 2219dea | 2018-01-30 18:15:52 +0800 | [diff] [blame] | 618 | for(; int(CURRENT_ITEM_OFFSET_IN_BYTES(src0_iter)) <= int(end_row_vec_a - uint(16)); |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 619 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, uint(8) * src0_attrs.stride_x), TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, uint(8) * src1_attrs.stride_y)) |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 620 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 621 | vec4 a0[2] = LOAD_UNPACK8_CURRENT_ITEM_HALF(src0_ptr, src0_iter); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 622 | |
| 623 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 624 | vec4 a1[2] = LOAD_UNPACK8_HALF(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 1)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 625 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 626 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 627 | vec4 a2[2] = LOAD_UNPACK8_HALF(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 2)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 628 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 629 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 630 | vec4 a3[2] = LOAD_UNPACK8_HALF(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 3)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 631 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 632 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 633 | vec4 b; |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 634 | |
| 635 | for(int i = 0; i < 8; i++) |
| 636 | { |
| 637 | int j = i >> 2; |
| 638 | int k = i % 4; |
| 639 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 640 | b = LOAD_UNPACK4_HALF(src1_ptr, IMAGE_OFFSET(src1_iter, 0, i)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 641 | |
| 642 | acc0 += b * vec4(a0[j][k]); |
| 643 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 644 | acc1 += b * vec4(a1[j][k]); |
| 645 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 646 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 647 | acc2 += b * vec4(a2[j][k]); |
| 648 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 649 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 650 | acc3 += b * vec4(a3[j][k]); |
| 651 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 652 | } |
| 653 | } |
| 654 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 655 | for(; int(CURRENT_ITEM_OFFSET_IN_BYTES(src0_iter)) < int(end_row_vec_a); TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, 2 * 8), TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, uint(8) * src1_attrs.stride_y)) |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 656 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 657 | vec4 a0[2] = LOAD_UNPACK8_CURRENT_ITEM_HALF(src0_ptr, src0_iter); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 658 | |
| 659 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 660 | vec4 a1[2] = LOAD_UNPACK8_HALF(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 1)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 661 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 662 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 663 | vec4 a2[2] = LOAD_UNPACK8_HALF(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 2)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 664 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 665 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 666 | vec4 a3[2] = LOAD_UNPACK8_HALF(src0_ptr, IMAGE_OFFSET(src0_iter, 0, 3)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 667 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 668 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 669 | vec4 b; |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 670 | |
| 671 | int leftover = COLS_A % 8; |
| 672 | |
| 673 | for(int i = 0; i < leftover; i++) |
| 674 | { |
| 675 | int j = i >> 2; |
| 676 | int k = i % 4; |
| 677 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 678 | b = LOAD_UNPACK4_HALF(src1_ptr, IMAGE_OFFSET(src1_iter, 0, i)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 679 | |
| 680 | acc0 += b * vec4(a0[j][k]); |
| 681 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 682 | acc1 += b * vec4(a1[j][k]); |
| 683 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 684 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 685 | acc2 += b * vec4(a2[j][k]); |
| 686 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 687 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 688 | acc3 += b * vec4(a3[j][k]); |
| 689 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 690 | } |
| 691 | } |
| 692 | |
| 693 | /* Multiply by the weight of vector-matrix product */ |
| 694 | acc0 = acc0 * vec4(ALPHA); |
| 695 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 696 | STORE_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, acc0); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 697 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 698 | STORE_PACK4_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 1), acc1); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 699 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 |
| 700 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 701 | STORE_PACK4_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 2), acc2); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 702 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 |
| 703 | #if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 704 | STORE_PACK4_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 3), acc3); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 705 | #endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 |
| 706 | } |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 707 | #elif defined(MM_PROCESS_8X) /* PROCESS_8X */ |
| 708 | TENSOR_DECLARATION(1, src0Buffer, uvec4, src0_ptr, src0_shift, 4, readonly); |
| 709 | TENSOR_DECLARATION(2, src1Buffer, uvec4, src1_ptr, src1_shift, 4, readonly); |
| 710 | TENSOR_DECLARATION(3, dstBuffer, uvec4, dst_ptr, dst_shift, 4, writeonly); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 711 | |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 712 | void main() |
| 713 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 714 | ImageIterator src0_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(src0_attrs, src0_shift); |
| 715 | ImageIterator src1_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(src1_attrs, src1_shift); |
| 716 | ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR(dst_attrs, dst_shift); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 717 | |
| 718 | int idx = int(gl_GlobalInvocationID.x) * int(NUM_ELEMS_PROCESSED_PER_THREAD_X); |
| 719 | /* Compute the address for the vector A and matrix B */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 720 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, uint(gl_GlobalInvocationID.y) * src0_attrs.stride_y * uint(NUM_ELEMS_PROCESSED_PER_THREAD_Y)); |
| 721 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, uint(idx) * src1_attrs.stride_x); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 722 | |
| 723 | /* Compute end row address for matrix A */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 724 | uint end_row_vec_a = uint(CURRENT_ITEM_OFFSET_IN_BYTES(src0_iter)) + uint(COLS_A << 1); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 725 | |
| 726 | /* Reset accumulators */ |
| 727 | vec4 acc[2]; |
| 728 | |
| 729 | acc[0] = vec4(0.0f); |
| 730 | acc[1] = vec4(0.0f); |
| 731 | |
zhenglin | 2219dea | 2018-01-30 18:15:52 +0800 | [diff] [blame] | 732 | for(; int(CURRENT_ITEM_OFFSET_IN_BYTES(src0_iter)) <= int(end_row_vec_a - uint(16)); |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 733 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, uint(8) * src0_attrs.stride_x), TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, uint(8) * src1_attrs.stride_y)) |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 734 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 735 | vec4 a[2] = LOAD_UNPACK8_CURRENT_ITEM_HALF(src0_ptr, src0_iter); |
| 736 | vec4 b[2]; |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 737 | |
| 738 | for(int i = 0; i < 8; i++) |
| 739 | { |
| 740 | int j = i >> 2; |
| 741 | int k = i % 4; |
| 742 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 743 | b = LOAD_UNPACK8_HALF(src1_ptr, IMAGE_OFFSET(src1_iter, 0, i)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 744 | |
| 745 | acc[0] += b[0] * vec4(a[j][k]); |
| 746 | acc[1] += b[1] * vec4(a[j][k]); |
| 747 | } |
| 748 | } |
| 749 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 750 | for(; int(CURRENT_ITEM_OFFSET_IN_BYTES(src0_iter)) < int(end_row_vec_a); |
| 751 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, uint(8) * uint(2)), TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, uint(8) * src1_attrs.stride_y)) |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 752 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 753 | vec4 a[2] = LOAD_UNPACK8_CURRENT_ITEM_HALF(src0_ptr, src0_iter); |
| 754 | vec4 b[2]; |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 755 | |
| 756 | int leftover = COLS_A % 8; |
| 757 | |
| 758 | for(int i = 0; i < leftover; i++) |
| 759 | { |
| 760 | int j = i >> 2; |
| 761 | int k = i % 4; |
| 762 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 763 | b = LOAD_UNPACK8_HALF(src1_ptr, IMAGE_OFFSET(src1_iter, 0, i)); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 764 | |
| 765 | acc[0] += b[0] * vec4(a[j][k]); |
| 766 | acc[1] += b[1] * vec4(a[j][k]); |
| 767 | } |
| 768 | } |
| 769 | |
| 770 | /* Multiply by the weight of vector-matrix product */ |
| 771 | acc[0] = acc[0] * vec4(ALPHA); |
| 772 | acc[1] = acc[1] * vec4(ALPHA); |
| 773 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 774 | STORE_PACK8_CURRENT_ITEM_HALF(dst_ptr, dst_iter, acc); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 775 | } |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 776 | #endif /* PROCESS_8X */ |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 777 | #endif /* GEMM_MM_FLOATING_POINT */ |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 778 | |
| 779 | #ifdef GEMM_ACCUMULATE_BIASES |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 780 | #if defined(ACCUM_PROCESS_4X) |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 781 | /** This kernel accumulates each row with the biases vector |
| 782 | * |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 783 | * @param[in, out] accum_ptr Pointer to the accumulate tensor. Supported data type: F16 |
| 784 | * @param[in] accum_attrs The attributes of the accumulate tensor |
| 785 | * @param[in] biases_ptr Pointer to the biases vector. Same as @p accum_ptr |
| 786 | * @param[in] biases_attrs The attributes of the biases tensor |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 787 | */ |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 788 | SHADER_PARAMS_DECLARATION |
| 789 | { |
| 790 | ImageAttributes accum_attrs; |
| 791 | VectorAttributes biases_attrs; |
| 792 | }; |
| 793 | |
| 794 | TENSOR_DECLARATION(1, accumBuffer, uvec2, accum_ptr, accum_shift, 3, restrict); |
| 795 | TENSOR_DECLARATION(2, biasesBuffer, uvec2, biases_ptr, biases_shift, 3, readonly); |
| 796 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 797 | void main(void) |
| 798 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 799 | ImageIterator accum_iter = CONVERT_TO_IMAGE_ITERATOR(accum_attrs, accum_shift); |
| 800 | VectorIterator biases_iter = CONVERT_TO_VECTOR_ITERATOR(biases_attrs, biases_shift); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 801 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 802 | vec4 u[2]; |
| 803 | u[0] = LOAD_UNPACK4_CURRENT_ITEM_HALF(accum_ptr, accum_iter); |
| 804 | u[1] = LOAD_UNPACK4_CURRENT_ITEM_HALF(biases_ptr, biases_iter); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 805 | |
| 806 | vec4 tmp; |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 807 | tmp = u[0] + u[1]; |
| 808 | STORE_PACK4_CURRENT_ITEM_HALF(accum_ptr, accum_iter, tmp); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 809 | } |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 810 | #elif defined(ACCUM_PROCESS_8X) /* ACCUM_PROCESS_8X */ |
| 811 | SHADER_PARAMS_DECLARATION |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 812 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 813 | ImageAttributes accum_attrs; |
| 814 | VectorAttributes biases_attrs; |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 815 | }; |
| 816 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 817 | TENSOR_DECLARATION(1, accumBuffer, uvec4, accum_ptr, accum_shift, 4, restrict); |
| 818 | TENSOR_DECLARATION(2, biasesBuffer, uvec4, biases_ptr, biases_shift, 4, readonly); |
| 819 | |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 820 | void main(void) |
| 821 | { |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 822 | ImageIterator accum_iter = CONVERT_TO_IMAGE_ITERATOR(accum_attrs, accum_shift); |
| 823 | VectorIterator biases_iter = CONVERT_TO_VECTOR_ITERATOR(biases_attrs, biases_shift); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 824 | |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 825 | vec4 u[2] = LOAD_UNPACK8_CURRENT_ITEM_HALF(accum_ptr, accum_iter); |
zhenglin | 2219dea | 2018-01-30 18:15:52 +0800 | [diff] [blame] | 826 | vec4 v[2] = LOAD_UNPACK8_CURRENT_ITEM_HALF(biases_ptr, biases_iter); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 827 | |
| 828 | vec4 r[2]; |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 829 | r[0] = u[0] + v[0]; |
| 830 | r[1] = u[1] + v[1]; |
| 831 | STORE_PACK8_CURRENT_ITEM_HALF(accum_ptr, accum_iter, r); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 832 | } |
zhenglin | 19e9142 | 2018-01-03 12:14:13 +0800 | [diff] [blame] | 833 | #endif /* ACCUM_PROCESS_8X */ |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 834 | #endif /* GEMM_ACCUMULATE_BIASES */ |
Stephen Li | e855c23 | 2018-01-04 14:13:22 +0800 | [diff] [blame] | 835 | |
| 836 | #ifdef GEMM_MM_INTERLEAVED_TRANSPOSED |
| 837 | /** This OpenGL ES kernel is optimised for Midgard. It computes the matrix multiplication between matrix A (src0) and matrix B (src1) |
| 838 | * Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_32bit and @ref gemm_transpose1x4 before running the matrix multiplication |
| 839 | * |
| 840 | * @attention The width of matrix B and the alpha's value need to be passed at compile time using WIDTH_MATRIX_B and ALPHA |
| 841 | * |
| 842 | * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16 |
| 843 | * @param[in] src0_attrs The attributes of the source matrix |
| 844 | * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr |
| 845 | * @param[in] src1_attrs The attributes of the source matrix |
| 846 | * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr |
| 847 | * @param[in] dst_attrs The attributes of the destination matrix |
| 848 | */ |
| 849 | SHADER_PARAMS_DECLARATION |
| 850 | { |
| 851 | ImageAttributes src0_attrs; |
| 852 | ImageAttributes src1_attrs; |
| 853 | ImageAttributes dst_attrs; |
| 854 | }; |
| 855 | TENSOR_DECLARATION(1, src0Buffer, uvec2, src0_ptr, src0_shift, 3, readonly); |
| 856 | TENSOR_DECLARATION(2, src1Buffer, uvec4, src1_ptr, src1_shift, 4, readonly); |
| 857 | TENSOR_DECLARATION(3, dstBuffer, uvec4, dst_ptr, dst_shift, 4, writeonly); |
| 858 | |
| 859 | void main() |
| 860 | { |
| 861 | ImageIterator src0_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(src0_attrs, src0_shift); |
| 862 | ImageIterator src1_iter = CONVERT_TO_IMAGE_ITERATOR_NO_STEP(src1_attrs, src1_shift); |
| 863 | ImageIterator dst_iter = CONVERT_TO_IMAGE_ITERATOR(dst_attrs, dst_shift); |
| 864 | |
| 865 | /* Compute address for matrix A and B */ |
| 866 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, uint(gl_GlobalInvocationID.y) * (src0_attrs.stride_y)); |
| 867 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, uint(gl_GlobalInvocationID.x) * (src1_attrs.stride_y)); |
| 868 | /* Compute end row address for matrix B */ |
| 869 | int end_row_mtx_b = (int(CURRENT_ITEM_OFFSET_IN_BYTES(src1_iter)) >> 1) + int(COLS_B); |
| 870 | |
| 871 | /* Reset accumulators */ |
| 872 | vec4 c00[2]; |
| 873 | vec4 c10[2]; |
| 874 | vec4 c20[2]; |
| 875 | vec4 c30[2]; |
| 876 | c00[0] = vec4(0.0f); |
| 877 | c00[1] = vec4(0.0f); |
| 878 | c10[0] = vec4(0.0f); |
| 879 | c10[1] = vec4(0.0f); |
| 880 | c20[0] = vec4(0.0f); |
| 881 | c20[1] = vec4(0.0f); |
| 882 | c30[0] = vec4(0.0f); |
| 883 | c30[1] = vec4(0.0f); |
| 884 | |
| 885 | // FIXME: loop unrolling really needed for GLES? |
| 886 | for(; (int(CURRENT_ITEM_OFFSET_IN_BYTES(src1_iter)) >> 1) <= (end_row_mtx_b - 16); TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, 16), TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, 32)) |
| 887 | { |
| 888 | /* Load values from matrix A (interleaved) and matrix B (transposed) */ |
| 889 | vec4 a0 = LOAD_UNPACK4_CURRENT_ITEM_HALF(src0_ptr, src0_iter); |
| 890 | vec4 b0[2] = LOAD_UNPACK8_CURRENT_ITEM_HALF(src1_ptr, src1_iter); |
| 891 | |
| 892 | c00[0] += vec4(a0.x) * b0[0]; |
| 893 | c00[1] += vec4(a0.x) * b0[1]; |
| 894 | c10[0] += vec4(a0.y) * b0[0]; |
| 895 | c10[1] += vec4(a0.y) * b0[1]; |
| 896 | c20[0] += vec4(a0.z) * b0[0]; |
| 897 | c20[1] += vec4(a0.z) * b0[1]; |
| 898 | c30[0] += vec4(a0.w) * b0[0]; |
| 899 | c30[1] += vec4(a0.w) * b0[1]; |
| 900 | |
| 901 | /* Load values from matrix A (interleaved) and matrix B (transposed) */ |
| 902 | a0 = LOAD_UNPACK4_HALF(src0_ptr, TENSOR_OFFSET_ADVANCE_IN_BYTES(src0_iter, 8)); |
| 903 | b0 = LOAD_UNPACK8_HALF(src1_ptr, TENSOR_OFFSET_ADVANCE_IN_BYTES(src1_iter, 16)); |
| 904 | |
| 905 | c00[0] += vec4(a0.x) * b0[0]; |
| 906 | c00[1] += vec4(a0.x) * b0[1]; |
| 907 | c10[0] += vec4(a0.y) * b0[0]; |
| 908 | c10[1] += vec4(a0.y) * b0[1]; |
| 909 | c20[0] += vec4(a0.z) * b0[0]; |
| 910 | c20[1] += vec4(a0.z) * b0[1]; |
| 911 | c30[0] += vec4(a0.w) * b0[0]; |
| 912 | c30[1] += vec4(a0.w) * b0[1]; |
| 913 | } |
| 914 | |
| 915 | for(; (int(CURRENT_ITEM_OFFSET_IN_BYTES(src1_iter)) >> 1) < end_row_mtx_b; TENSOR_ITERATOR_ADVANCE_IN_BYTES(src0_iter, 8), TENSOR_ITERATOR_ADVANCE_IN_BYTES(src1_iter, 16)) |
| 916 | { |
| 917 | /* Load values from matrix A (interleaved) and matrix B (transposed) */ |
| 918 | vec4 a0 = LOAD_UNPACK4_CURRENT_ITEM_HALF(src0_ptr, src0_iter); |
| 919 | vec4 b0[2] = LOAD_UNPACK8_CURRENT_ITEM_HALF(src1_ptr, src1_iter); |
| 920 | |
| 921 | c00[0] += vec4(a0.x) * b0[0]; |
| 922 | c00[1] += vec4(a0.x) * b0[1]; |
| 923 | c10[0] += vec4(a0.y) * b0[0]; |
| 924 | c10[1] += vec4(a0.y) * b0[1]; |
| 925 | c20[0] += vec4(a0.z) * b0[0]; |
| 926 | c20[1] += vec4(a0.z) * b0[1]; |
| 927 | c30[0] += vec4(a0.w) * b0[0]; |
| 928 | c30[1] += vec4(a0.w) * b0[1]; |
| 929 | } |
| 930 | |
| 931 | /* Multiply by the weight of matrix product */ |
| 932 | c00[0] = c00[0] * vec4(ALPHA); |
| 933 | c00[1] = c00[1] * vec4(ALPHA); |
| 934 | c10[0] = c10[0] * vec4(ALPHA); |
| 935 | c10[1] = c10[1] * vec4(ALPHA); |
| 936 | c20[0] = c20[0] * vec4(ALPHA); |
| 937 | c20[1] = c20[1] * vec4(ALPHA); |
| 938 | c30[0] = c30[0] * vec4(ALPHA); |
| 939 | c30[1] = c30[1] * vec4(ALPHA); |
| 940 | |
| 941 | /* Store 4x8 block */ |
| 942 | STORE_PACK8_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 0), c00); |
| 943 | STORE_PACK8_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 1), c10); |
| 944 | STORE_PACK8_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 2), c20); |
| 945 | STORE_PACK8_HALF(dst_ptr, IMAGE_OFFSET(dst_iter, 0, 3), c30); |
| 946 | } |
| 947 | #endif /* GEMM_MM_INTERLEAVED_TRANSPOSED */ |
| 948 | #else /* DATA_TYPE_FP16 */ |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 949 | #error Data type not supported |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 950 | #endif /* DATA_TYPE_FP32 */ |