Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 1 | /* |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [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 | |
| 25 | layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in; |
| 26 | |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 27 | #include "helpers_cs.h" |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 28 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 29 | #ifdef FUSED_ACTIVATION |
| 30 | #include "activation_layer_helpers_cs.h" |
| 31 | #endif /* FUSED_ACTIVATION */ |
| 32 | |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 33 | #if defined(DATA_TYPE_FP16) |
| 34 | precision mediump float; |
| 35 | #endif // DATA_TYPE_FP16 |
| 36 | |
| 37 | /** This kernel performs a direct convolution to convolve the low three dimensions. |
| 38 | * |
| 39 | * @note The data type must be passed at compile time using "#define DATA_TYPE_NAME". e.g. "#define DATA_TYPE_FP32" |
Xinghang Zhou | 4af62a0 | 2017-11-02 16:37:24 +0800 | [diff] [blame] | 40 | * @note This kernel has multiple optimized direct convolution options for FP16. |
| 41 | * The direct convolution option must be passed at compile time using "#define PROCESS_nX_nY_nZ" e.g. "#define PROCESS_8X_1Y_1Z" |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 42 | * @note The convolution stride x must be passed at compile time using "#define STRIDE_X n" e.g. "#define STRIDE_X 1" |
| 43 | * @note In case biases will be added to the convolution "#define HAS_BIAS" has to be passed to append the final matrix with 1 in each row. |
| 44 | * |
| 45 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 |
| 46 | * @param[in] src_attrs The attributes of the source tensor |
| 47 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 48 | * @param[in] dst_attrs The attributes of the destination tensor |
| 49 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr |
| 50 | * @param[in] weights_attrs The attributes of the weights tensor |
| 51 | * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr |
| 52 | * @param[in] biases_attrs The attributes of the weights tensor |
| 53 | * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension |
| 54 | * @param[in] weights_depth The third dimensions of the weights tensors |
| 55 | */ |
| 56 | SHADER_PARAMS_DECLARATION |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 57 | { |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 58 | Tensor3DAttributes src_attrs; |
| 59 | Tensor3DAttributes dst_attrs; |
| 60 | Tensor3DAttributes weights_attrs; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 61 | #ifdef BIAS |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 62 | VectorAttributes biases_attrs; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 63 | #endif /* BIAS */ |
| 64 | uint weights_stride_w; |
| 65 | uint weights_depth; |
| 66 | }; |
| 67 | |
| 68 | #if defined(DATA_TYPE_FP32) |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 69 | TENSOR_DECLARATION(1, srcBuffer, float, src_ptr, src_shift, 2, readonly); |
| 70 | TENSOR_DECLARATION(2, dstBuffer, float, dst_ptr, dst_shift, 2, writeonly); |
| 71 | TENSOR_DECLARATION(3, weightsBuffer, float, weights_ptr, weights_shift, 2, readonly); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 72 | #ifdef BIAS |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 73 | TENSOR_DECLARATION(4, biasesBuffer, float, biases_ptr, biases_shift, 2, readonly); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 74 | #endif /* BIAS */ |
| 75 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 76 | void main() |
| 77 | { |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 78 | ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); |
| 79 | Tensor3DIterator weights_iter = CONVERT_TO_TENSOR3D_ITERATOR_NO_STEP(weights_attrs, weights_shift); |
| 80 | Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 81 | |
| 82 | #ifdef BIAS |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 83 | VectorIterator biases_iter = CONVERT_TO_VECTOR_ITERATOR_NO_STEP(biases_attrs, biases_shift); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 84 | #endif /* BIAS */ |
| 85 | |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 86 | float pixels = 0.f; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 87 | uint z_index = gl_GlobalInvocationID.z; |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 88 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_index * weights_stride_w); |
| 89 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 90 | float temp; |
| 91 | float temp_weight; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 92 | for(int d = 0; d < int(weights_depth); ++d) |
| 93 | { |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 94 | temp = LOAD_CURRENT_ITEM(src_ptr, src_iter); |
| 95 | temp_weight = LOAD_CURRENT_ITEM(weights_ptr, weights_iter); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 96 | pixels += temp * temp_weight; |
| 97 | |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 98 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 99 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 100 | } |
| 101 | |
| 102 | #ifdef BIAS |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 103 | pixels += LOAD(biases_ptr, VECTOR_OFFSET(biases_iter, z_index)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 104 | #endif /* BIAS */ |
| 105 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 106 | #ifdef FUSED_ACTIVATION |
| 107 | pixels = ACT_OP(pixels); |
| 108 | #endif /* FUSED_ACTIVATION */ |
| 109 | |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 110 | STORE_CURRENT_ITEM(dst_ptr, dst_iter, pixels); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 111 | } |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 112 | |
Xinghang Zhou | 4af62a0 | 2017-11-02 16:37:24 +0800 | [diff] [blame] | 113 | #elif defined(DATA_TYPE_FP16) |
| 114 | #if defined(PROCESS_4X_1Y_1Z) |
| 115 | TENSOR_DECLARATION(1, srcBuffer, uvec2, src_ptr, src_shift, 3, readonly); |
| 116 | TENSOR_DECLARATION(2, dstBuffer, uvec2, dst_ptr, dst_shift, 3, writeonly); |
| 117 | TENSOR_DECLARATION(3, weightsBuffer, uint, weights_ptr, weights_shift, 2, readonly); |
| 118 | #ifdef BIAS |
| 119 | TENSOR_DECLARATION(4, biasesBuffer, uint, biases_ptr, biases_shift, 2, readonly); |
| 120 | #endif /* BIAS */ |
| 121 | |
| 122 | #if STRIDE_X == 2 |
| 123 | #define CONVOLVE(s, w) convolve_stride2(s, w) |
| 124 | #elif STRIDE_X == 1 /* STRIDE_X == 1 */ |
| 125 | #define CONVOLVE(s, w) convolve_stride1(s, w) |
| 126 | #else /* STRIDE_X not equals 1 or 2 */ |
| 127 | #error STRIDE_X larger than 2 is not supported |
| 128 | #endif /* STRIDE_X == 2 */ |
| 129 | |
| 130 | vec4 convolve_stride1(ImageIterator src_iter, float w) |
| 131 | { |
| 132 | vec4 s; |
| 133 | s = LOAD_UNPACK4_CURRENT_ITEM_HALF(src_ptr, src_iter); |
| 134 | |
| 135 | s *= w; |
| 136 | |
| 137 | return s; |
| 138 | } |
| 139 | |
| 140 | vec4 convolve_stride2(ImageIterator src_iter, float w) |
| 141 | { |
| 142 | vec4 s[2]; |
| 143 | vec4 r; |
| 144 | |
| 145 | s[0] = LOAD_UNPACK4_CURRENT_ITEM_HALF(src_ptr, src_iter); |
| 146 | s[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 4, 0)); |
| 147 | r = vec4(s[0].xz, s[1].xz); |
| 148 | |
| 149 | r *= w; |
| 150 | |
| 151 | return r; |
| 152 | } |
| 153 | |
| 154 | void main() |
| 155 | { |
| 156 | ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); |
| 157 | Tensor3DIterator weights_iter = CONVERT_TO_TENSOR3D_ITERATOR_NO_STEP(weights_attrs, weights_shift); |
| 158 | Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); |
| 159 | |
| 160 | #ifdef BIAS |
| 161 | VectorIterator biases_iter = CONVERT_TO_VECTOR_ITERATOR_NO_STEP(biases_attrs, biases_shift); |
| 162 | #endif /* BIAS */ |
| 163 | |
| 164 | vec4 pixels = vec4(0.f); |
| 165 | |
| 166 | uint z_index = gl_GlobalInvocationID.z; |
| 167 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_index * weights_stride_w); |
| 168 | |
| 169 | #ifdef WEIGHTS_OPTIMIZATION |
| 170 | float w1, w2; |
| 171 | int nums = (int(weights_depth)) / 2; |
| 172 | for(int d = 0; d < nums; ++d) |
| 173 | { |
| 174 | vec2 vec2_w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter); |
| 175 | |
| 176 | w1 = vec2_w.x; |
| 177 | vec4 r1 = CONVOLVE(src_iter, w1); |
| 178 | pixels += r1; |
| 179 | |
| 180 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 181 | |
| 182 | w2 = vec2_w.y; |
| 183 | vec4 r2 = CONVOLVE(src_iter, w2); |
| 184 | pixels += r2; |
| 185 | |
| 186 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 187 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z * uint(2)); |
| 188 | } |
| 189 | #else /* WEIGHTS_OPTIMIZATION */ |
| 190 | float w; |
| 191 | for(int d = 0; d < int(weights_depth); ++d) |
| 192 | { |
| 193 | w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter).x; |
| 194 | |
| 195 | vec4 r = CONVOLVE(src_iter, w); |
| 196 | pixels += r; |
| 197 | |
| 198 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 199 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z); |
| 200 | } |
| 201 | #endif /* WEIGHTS_OPTIMIZATION */ |
| 202 | |
| 203 | #ifdef BIAS |
| 204 | vec2 vec2_b; |
| 205 | float b; |
| 206 | |
| 207 | vec2_b = LOAD_UNPACK2_HALF(biases_ptr, VECTOR_OFFSET(biases_iter, z_index)); |
| 208 | |
| 209 | if(z_index % uint(2) == uint(0)) |
| 210 | { |
| 211 | b = vec2_b.x; |
| 212 | } |
| 213 | else |
| 214 | { |
| 215 | b = vec2_b.y; |
| 216 | } |
| 217 | |
| 218 | pixels += b; |
| 219 | #endif /* BIAS */ |
| 220 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 221 | #ifdef FUSED_ACTIVATION |
| 222 | pixels = ACT_OP(pixels); |
| 223 | #endif /* FUSED_ACTIVATION */ |
| 224 | |
Xinghang Zhou | 4af62a0 | 2017-11-02 16:37:24 +0800 | [diff] [blame] | 225 | STORE_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, pixels); |
| 226 | } |
| 227 | #elif defined(PROCESS_4X_2Y_1Z) |
| 228 | TENSOR_DECLARATION(1, srcBuffer, uvec2, src_ptr, src_shift, 3, readonly); |
| 229 | TENSOR_DECLARATION(2, dstBuffer, uvec2, dst_ptr, dst_shift, 3, writeonly); |
| 230 | TENSOR_DECLARATION(3, weightsBuffer, uint, weights_ptr, weights_shift, 2, readonly); |
| 231 | #ifdef BIAS |
| 232 | TENSOR_DECLARATION(4, biasesBuffer, uint, biases_ptr, biases_shift, 2, readonly); |
| 233 | #endif /* BIAS */ |
| 234 | |
| 235 | #if STRIDE_X == 2 |
| 236 | #define CONVOLVE(s, w) convolve_stride2(s, w) |
| 237 | #elif STRIDE_X == 1 /* STRIDE_X == 1 */ |
| 238 | #define CONVOLVE(s, w) convolve_stride1(s, w) |
| 239 | #else /* STRIDE_X not equals 1 or 2 */ |
| 240 | #error STRIDE_X larger than 2 is not supported |
| 241 | #endif /* STRIDE_X == 2 */ |
| 242 | |
| 243 | vec4[2] convolve_stride1(ImageIterator src_iter, float w) |
| 244 | { |
| 245 | vec4 s[2]; |
| 246 | s[0] = LOAD_UNPACK4_CURRENT_ITEM_HALF(src_ptr, src_iter); |
| 247 | s[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, int(STRIDE_Y))); |
| 248 | |
| 249 | s[0] *= w; |
| 250 | s[1] *= w; |
| 251 | |
| 252 | return s; |
| 253 | } |
| 254 | |
| 255 | vec4[2] convolve_stride2(ImageIterator src_iter, float w) |
| 256 | { |
| 257 | vec4 s1[2]; |
| 258 | vec4 s2[2]; |
| 259 | vec4 r[2]; |
| 260 | |
| 261 | s1[0] = LOAD_UNPACK4_CURRENT_ITEM_HALF(src_ptr, src_iter); |
| 262 | s1[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 4, 0)); |
| 263 | r[0] = vec4(s1[0].xz, s1[1].xz); |
| 264 | |
| 265 | s2[0] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, int(STRIDE_Y))); |
| 266 | s2[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 4, int(STRIDE_Y))); |
| 267 | r[1] = vec4(s2[0].xz, s2[1].xz); |
| 268 | |
| 269 | r[0] *= w; |
| 270 | r[1] *= w; |
| 271 | |
| 272 | return r; |
| 273 | } |
| 274 | |
| 275 | void main() |
| 276 | { |
| 277 | ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); |
| 278 | Tensor3DIterator weights_iter = CONVERT_TO_TENSOR3D_ITERATOR_NO_STEP(weights_attrs, weights_shift); |
| 279 | Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); |
| 280 | |
| 281 | #ifdef BIAS |
| 282 | VectorIterator biases_iter = CONVERT_TO_VECTOR_ITERATOR_NO_STEP(biases_attrs, biases_shift); |
| 283 | #endif /* BIAS */ |
| 284 | |
| 285 | vec4 pixels[2]; |
| 286 | pixels[0] = vec4(0.f); |
| 287 | pixels[1] = vec4(0.f); |
| 288 | |
| 289 | uint z_index = gl_GlobalInvocationID.z; |
| 290 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_index * weights_stride_w); |
| 291 | |
| 292 | #ifdef WEIGHTS_OPTIMIZATION |
| 293 | float w1, w2; |
| 294 | int nums = (int(weights_depth)) / 2; |
| 295 | for(int d = 0; d < nums; ++d) |
| 296 | { |
| 297 | vec2 vec2_w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter); |
| 298 | |
| 299 | w1 = vec2_w.x; |
| 300 | vec4 r1[2] = CONVOLVE(src_iter, w1); |
| 301 | pixels[0] += r1[0]; |
| 302 | pixels[1] += r1[1]; |
| 303 | |
| 304 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 305 | |
| 306 | w2 = vec2_w.y; |
| 307 | vec4 r2[2] = CONVOLVE(src_iter, w2); |
| 308 | pixels[0] += r2[0]; |
| 309 | pixels[1] += r2[1]; |
| 310 | |
| 311 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 312 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z * uint(2)); |
| 313 | } |
| 314 | #else /* WEIGHTS_OPTIMIZATION */ |
| 315 | float w; |
| 316 | for(int d = 0; d < int(weights_depth); ++d) |
| 317 | { |
| 318 | w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter).x; |
| 319 | |
| 320 | vec4 r[2] = CONVOLVE(src_iter, w); |
| 321 | pixels[0] += r[0]; |
| 322 | pixels[1] += r[1]; |
| 323 | |
| 324 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 325 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z); |
| 326 | } |
| 327 | #endif /* WEIGHTS_OPTIMIZATION */ |
| 328 | |
| 329 | #ifdef BIAS |
| 330 | vec2 vec2_b; |
| 331 | float b; |
| 332 | |
| 333 | vec2_b = LOAD_UNPACK2_HALF(biases_ptr, VECTOR_OFFSET(biases_iter, z_index)); |
| 334 | |
| 335 | if(z_index % uint(2) == uint(0)) |
| 336 | { |
| 337 | b = vec2_b.x; |
| 338 | } |
| 339 | else |
| 340 | { |
| 341 | b = vec2_b.y; |
| 342 | } |
| 343 | |
| 344 | pixels[0] += b; |
| 345 | pixels[1] += b; |
| 346 | #endif /* BIAS */ |
| 347 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 348 | #ifdef FUSED_ACTIVATION |
| 349 | pixels[0] = ACT_OP(pixels[0]); |
| 350 | pixels[1] = ACT_OP(pixels[1]); |
| 351 | #endif /* FUSED_ACTIVATION */ |
| 352 | |
Xinghang Zhou | 4af62a0 | 2017-11-02 16:37:24 +0800 | [diff] [blame] | 353 | STORE_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, pixels[0]); |
| 354 | STORE_PACK4_HALF(dst_ptr, TENSOR3D_OFFSET(dst_iter, 0, 1, 0), pixels[1]); |
| 355 | } |
| 356 | #elif defined(PROCESS_4X_3Y_1Z) |
| 357 | TENSOR_DECLARATION(1, srcBuffer, uvec2, src_ptr, src_shift, 3, readonly); |
| 358 | TENSOR_DECLARATION(2, dstBuffer, uvec2, dst_ptr, dst_shift, 3, writeonly); |
| 359 | TENSOR_DECLARATION(3, weightsBuffer, uint, weights_ptr, weights_shift, 2, readonly); |
| 360 | #ifdef BIAS |
| 361 | TENSOR_DECLARATION(4, biasesBuffer, uint, biases_ptr, biases_shift, 2, readonly); |
| 362 | #endif /* BIAS */ |
| 363 | |
| 364 | #if STRIDE_X == 2 |
| 365 | #define CONVOLVE(s, w) convolve_stride2(s, w) |
| 366 | #elif STRIDE_X == 1 /* STRIDE_X == 1 */ |
| 367 | #define CONVOLVE(s, w) convolve_stride1(s, w) |
| 368 | #else /* STRIDE_X not equals 1 or 2 */ |
| 369 | #error STRIDE_X larger than 2 is not supported |
| 370 | #endif /* STRIDE_X == 2 */ |
| 371 | |
| 372 | vec4[3] convolve_stride1(ImageIterator src_iter, float w) |
| 373 | { |
| 374 | vec4 s[3]; |
| 375 | s[0] = LOAD_UNPACK4_CURRENT_ITEM_HALF(src_ptr, src_iter); |
| 376 | s[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, int(STRIDE_Y))); |
| 377 | s[2] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, (2 * int(STRIDE_Y)))); |
| 378 | |
| 379 | s[0] *= w; |
| 380 | s[1] *= w; |
| 381 | s[2] *= w; |
| 382 | |
| 383 | return s; |
| 384 | } |
| 385 | |
| 386 | vec4[3] convolve_stride2(ImageIterator src_iter, float w) |
| 387 | { |
| 388 | vec4 s1[2]; |
| 389 | vec4 s2[2]; |
| 390 | vec4 s3[2]; |
| 391 | vec4 r[3]; |
| 392 | |
| 393 | s1[0] = LOAD_UNPACK4_CURRENT_ITEM_HALF(src_ptr, src_iter); |
| 394 | s1[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 4, 0)); |
| 395 | r[0] = vec4(s1[0].xz, s1[1].xz); |
| 396 | |
| 397 | s2[0] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, int(STRIDE_Y))); |
| 398 | s2[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 4, int(STRIDE_Y))); |
| 399 | r[1] = vec4(s2[0].xz, s2[1].xz); |
| 400 | |
| 401 | s3[0] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, (2 * int(STRIDE_Y)))); |
| 402 | s3[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 4, (2 * int(STRIDE_Y)))); |
| 403 | r[2] = vec4(s3[0].xz, s3[1].xz); |
| 404 | |
| 405 | r[0] *= w; |
| 406 | r[1] *= w; |
| 407 | r[2] *= w; |
| 408 | |
| 409 | return r; |
| 410 | } |
| 411 | |
| 412 | void main() |
| 413 | { |
| 414 | ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); |
| 415 | Tensor3DIterator weights_iter = CONVERT_TO_TENSOR3D_ITERATOR_NO_STEP(weights_attrs, weights_shift); |
| 416 | Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); |
| 417 | |
| 418 | #ifdef BIAS |
| 419 | VectorIterator biases_iter = CONVERT_TO_VECTOR_ITERATOR_NO_STEP(biases_attrs, biases_shift); |
| 420 | #endif /* BIAS */ |
| 421 | |
| 422 | vec4 pixels[3]; |
| 423 | pixels[0] = vec4(0.f); |
| 424 | pixels[1] = vec4(0.f); |
| 425 | pixels[2] = vec4(0.f); |
| 426 | |
| 427 | uint z_index = gl_GlobalInvocationID.z; |
| 428 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_index * weights_stride_w); |
| 429 | |
| 430 | #ifdef WEIGHTS_OPTIMIZATION |
| 431 | float w1, w2; |
| 432 | int nums = (int(weights_depth)) / 2; |
| 433 | for(int d = 0; d < nums; ++d) |
| 434 | { |
| 435 | vec2 vec2_w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter); |
| 436 | |
| 437 | w1 = vec2_w.x; |
| 438 | vec4 r1[3] = CONVOLVE(src_iter, w1); |
| 439 | pixels[0] += r1[0]; |
| 440 | pixels[1] += r1[1]; |
| 441 | pixels[2] += r1[2]; |
| 442 | |
| 443 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 444 | |
| 445 | w2 = vec2_w.y; |
| 446 | vec4 r2[3] = CONVOLVE(src_iter, w2); |
| 447 | pixels[0] += r2[0]; |
| 448 | pixels[1] += r2[1]; |
| 449 | pixels[2] += r2[2]; |
| 450 | |
| 451 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 452 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z * uint(2)); |
| 453 | } |
| 454 | #else /* WEIGHTS_OPTIMIZATION */ |
| 455 | float w; |
| 456 | for(int d = 0; d < int(weights_depth); ++d) |
| 457 | { |
| 458 | w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter).x; |
| 459 | |
| 460 | vec4 r[3] = CONVOLVE(src_iter, w); |
| 461 | pixels[0] += r[0]; |
| 462 | pixels[1] += r[1]; |
| 463 | pixels[2] += r[2]; |
| 464 | |
| 465 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 466 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z); |
| 467 | } |
| 468 | #endif /* WEIGHTS_OPTIMIZATION */ |
| 469 | |
| 470 | #ifdef BIAS |
| 471 | vec2 vec2_b; |
| 472 | float b; |
| 473 | |
| 474 | vec2_b = LOAD_UNPACK2_HALF(biases_ptr, VECTOR_OFFSET(biases_iter, z_index)); |
| 475 | |
| 476 | if(z_index % uint(2) == uint(0)) |
| 477 | { |
| 478 | b = vec2_b.x; |
| 479 | } |
| 480 | else |
| 481 | { |
| 482 | b = vec2_b.y; |
| 483 | } |
| 484 | |
| 485 | pixels[0] += b; |
| 486 | pixels[1] += b; |
| 487 | pixels[2] += b; |
| 488 | #endif /* BIAS */ |
| 489 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 490 | #ifdef FUSED_ACTIVATION |
| 491 | pixels[0] = ACT_OP(pixels[0]); |
| 492 | pixels[1] = ACT_OP(pixels[1]); |
| 493 | pixels[2] = ACT_OP(pixels[2]); |
| 494 | #endif /* FUSED_ACTIVATION */ |
| 495 | |
Xinghang Zhou | 4af62a0 | 2017-11-02 16:37:24 +0800 | [diff] [blame] | 496 | STORE_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, pixels[0]); |
| 497 | STORE_PACK4_HALF(dst_ptr, TENSOR3D_OFFSET(dst_iter, 0, 1, 0), pixels[1]); |
| 498 | STORE_PACK4_HALF(dst_ptr, TENSOR3D_OFFSET(dst_iter, 0, 2, 0), pixels[2]); |
| 499 | } |
| 500 | #elif defined(PROCESS_4X_4Y_1Z) |
| 501 | TENSOR_DECLARATION(1, srcBuffer, uvec2, src_ptr, src_shift, 3, readonly); |
| 502 | TENSOR_DECLARATION(2, dstBuffer, uvec2, dst_ptr, dst_shift, 3, writeonly); |
| 503 | TENSOR_DECLARATION(3, weightsBuffer, uint, weights_ptr, weights_shift, 2, readonly); |
| 504 | #ifdef BIAS |
| 505 | TENSOR_DECLARATION(4, biasesBuffer, uint, biases_ptr, biases_shift, 2, readonly); |
| 506 | #endif /* BIAS */ |
| 507 | |
| 508 | #if STRIDE_X == 2 |
| 509 | #define CONVOLVE(s, w, x1, y1) convolve_stride2(s, w, x1, y1) |
| 510 | #elif STRIDE_X == 1 /* STRIDE_X == 1 */ |
| 511 | #define CONVOLVE(s, w, x1, y1) convolve_stride1(s, w, x1, y1) |
| 512 | #else /* STRIDE_X not equals 1 or 2 */ |
| 513 | #error STRIDE_X larger than 2 is not supported |
| 514 | #endif /* STRIDE_X == 2 */ |
| 515 | |
| 516 | vec4[2] convolve_stride1(ImageIterator src_iter, float w, int x1, int y1) |
| 517 | { |
| 518 | vec4 s[2]; |
| 519 | s[0] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, x1, y1)); |
| 520 | s[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, x1, (y1 + int(STRIDE_Y)))); |
| 521 | |
| 522 | s[0] *= w; |
| 523 | s[1] *= w; |
| 524 | |
| 525 | return s; |
| 526 | } |
| 527 | |
| 528 | vec4[2] convolve_stride2(ImageIterator src_iter, float w, int x1, int y1) |
| 529 | { |
| 530 | vec4 s1[2]; |
| 531 | vec4 s2[2]; |
| 532 | vec4 r[2]; |
| 533 | |
| 534 | s1[0] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, x1, y1)); |
| 535 | s1[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, (4 + x1), y1)); |
| 536 | r[0] = vec4(s1[0].xz, s1[1].xz); |
| 537 | |
| 538 | s2[0] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, x1, (y1 + int(STRIDE_Y)))); |
| 539 | s2[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, (4 + x1), (y1 + int(STRIDE_Y)))); |
| 540 | r[1] = vec4(s2[0].xz, s2[1].xz); |
| 541 | |
| 542 | r[0] *= w; |
| 543 | r[1] *= w; |
| 544 | |
| 545 | return r; |
| 546 | } |
| 547 | |
| 548 | void main() |
| 549 | { |
| 550 | ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); |
| 551 | Tensor3DIterator weights_iter = CONVERT_TO_TENSOR3D_ITERATOR_NO_STEP(weights_attrs, weights_shift); |
| 552 | Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); |
| 553 | |
| 554 | #ifdef BIAS |
| 555 | VectorIterator biases_iter = CONVERT_TO_VECTOR_ITERATOR_NO_STEP(biases_attrs, biases_shift); |
| 556 | #endif /* BIAS */ |
| 557 | |
| 558 | vec4 pixels[2]; |
| 559 | vec4 pixels1[2]; |
| 560 | pixels[0] = vec4(0.f); |
| 561 | pixels[1] = vec4(0.f); |
| 562 | pixels1[0] = vec4(0.f); |
| 563 | pixels1[1] = vec4(0.f); |
| 564 | |
| 565 | uint z_index = gl_GlobalInvocationID.z; |
| 566 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_index * weights_stride_w); |
| 567 | |
| 568 | #ifdef WEIGHTS_OPTIMIZATION |
| 569 | float w1, w2; |
| 570 | int nums = (int(weights_depth)) / 2; |
| 571 | for(int d = 0; d < nums; ++d) |
| 572 | { |
| 573 | vec2 vec2_w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter); |
| 574 | |
| 575 | w1 = vec2_w.x; |
| 576 | vec4 r1[2] = CONVOLVE(src_iter, w1, 0, 0); |
| 577 | vec4 r2[2] = CONVOLVE(src_iter, w1, 0, (2 * int(STRIDE_Y))); |
| 578 | pixels[0] += r1[0]; |
| 579 | pixels[1] += r1[1]; |
| 580 | pixels1[0] += r2[0]; |
| 581 | pixels1[1] += r2[1]; |
| 582 | |
| 583 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 584 | |
| 585 | w2 = vec2_w.y; |
| 586 | vec4 r3[2] = CONVOLVE(src_iter, w2, 0, 0); |
| 587 | vec4 r4[2] = CONVOLVE(src_iter, w2, 0, (2 * int(STRIDE_Y))); |
| 588 | pixels[0] += r3[0]; |
| 589 | pixels[1] += r3[1]; |
| 590 | pixels1[0] += r4[0]; |
| 591 | pixels1[1] += r4[1]; |
| 592 | |
| 593 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 594 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z * uint(2)); |
| 595 | } |
| 596 | #else /* WEIGHTS_OPTIMIZATION */ |
| 597 | float w; |
| 598 | for(int d = 0; d < int(weights_depth); ++d) |
| 599 | { |
| 600 | w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter).x; |
| 601 | |
| 602 | vec4 r1[2] = CONVOLVE(src_iter, w, 0, 0); |
| 603 | vec4 r2[2] = CONVOLVE(src_iter, w, 0, (2 * int(STRIDE_Y))); |
| 604 | pixels[0] += r1[0]; |
| 605 | pixels[1] += r1[1]; |
| 606 | pixels1[0] += r2[0]; |
| 607 | pixels1[1] += r2[1]; |
| 608 | |
| 609 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 610 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z); |
| 611 | } |
| 612 | #endif /* WEIGHTS_OPTIMIZATION */ |
| 613 | |
| 614 | #ifdef BIAS |
| 615 | vec2 vec2_b; |
| 616 | float b; |
| 617 | |
| 618 | vec2_b = LOAD_UNPACK2_HALF(biases_ptr, VECTOR_OFFSET(biases_iter, z_index)); |
| 619 | |
| 620 | if(z_index % uint(2) == uint(0)) |
| 621 | { |
| 622 | b = vec2_b.x; |
| 623 | } |
| 624 | else |
| 625 | { |
| 626 | b = vec2_b.y; |
| 627 | } |
| 628 | |
| 629 | pixels[0] += b; |
| 630 | pixels[1] += b; |
| 631 | pixels1[0] += b; |
| 632 | pixels1[1] += b; |
| 633 | #endif /* BIAS */ |
| 634 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 635 | #ifdef FUSED_ACTIVATION |
| 636 | pixels[0] = ACT_OP(pixels[0]); |
| 637 | pixels[1] = ACT_OP(pixels[1]); |
| 638 | pixels1[0] = ACT_OP(pixels1[0]); |
| 639 | pixels1[1] = ACT_OP(pixels1[1]); |
| 640 | #endif /* FUSED_ACTIVATION */ |
| 641 | |
Xinghang Zhou | 4af62a0 | 2017-11-02 16:37:24 +0800 | [diff] [blame] | 642 | STORE_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, pixels[0]); |
| 643 | STORE_PACK4_HALF(dst_ptr, TENSOR3D_OFFSET(dst_iter, 0, 1, 0), pixels[1]); |
| 644 | STORE_PACK4_HALF(dst_ptr, TENSOR3D_OFFSET(dst_iter, 0, 2, 0), pixels1[0]); |
| 645 | STORE_PACK4_HALF(dst_ptr, TENSOR3D_OFFSET(dst_iter, 0, 3, 0), pixels1[1]); |
| 646 | } |
| 647 | #elif defined(PROCESS_4X_2Y_2Z) |
| 648 | TENSOR_DECLARATION(1, srcBuffer, uvec2, src_ptr, src_shift, 3, readonly); |
| 649 | TENSOR_DECLARATION(2, dstBuffer, uvec2, dst_ptr, dst_shift, 3, writeonly); |
| 650 | TENSOR_DECLARATION(3, weightsBuffer, uint, weights_ptr, weights_shift, 2, readonly); |
| 651 | #ifdef BIAS |
| 652 | TENSOR_DECLARATION(4, biasesBuffer, uint, biases_ptr, biases_shift, 2, readonly); |
| 653 | #endif /* BIAS */ |
| 654 | |
| 655 | #if STRIDE_X == 2 |
| 656 | #define CONVOLVE(s, w) convolve_stride2(s, w) |
| 657 | #elif STRIDE_X == 1 /* STRIDE_X == 1 */ |
| 658 | #define CONVOLVE(s, w) convolve_stride1(s, w) |
| 659 | #else /* STRIDE_X not equals 1 or 2 */ |
| 660 | #error STRIDE_X larger than 2 is not supported |
| 661 | #endif /* STRIDE_X == 2 */ |
| 662 | |
| 663 | vec4[2] convolve_stride1(ImageIterator src_iter, float w) |
| 664 | { |
| 665 | vec4 s[2]; |
| 666 | s[0] = LOAD_UNPACK4_CURRENT_ITEM_HALF(src_ptr, src_iter); |
| 667 | s[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, int(STRIDE_Y))); |
| 668 | |
| 669 | s[0] *= w; |
| 670 | s[1] *= w; |
| 671 | |
| 672 | return s; |
| 673 | } |
| 674 | |
| 675 | vec4[2] convolve_stride2(ImageIterator src_iter, float w) |
| 676 | { |
| 677 | vec4 s1[2]; |
| 678 | vec4 s2[2]; |
| 679 | vec4 r[2]; |
| 680 | |
| 681 | s1[0] = LOAD_UNPACK4_CURRENT_ITEM_HALF(src_ptr, src_iter); |
| 682 | s1[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 4, 0)); |
| 683 | r[0] = vec4(s1[0].xz, s1[1].xz); |
| 684 | |
| 685 | s2[0] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 0, int(STRIDE_Y))); |
| 686 | s2[1] = LOAD_UNPACK4_HALF(src_ptr, IMAGE_OFFSET(src_iter, 4, int(STRIDE_Y))); |
| 687 | r[1] = vec4(s2[0].xz, s2[1].xz); |
| 688 | |
| 689 | r[0] *= w; |
| 690 | r[1] *= w; |
| 691 | |
| 692 | return r; |
| 693 | } |
| 694 | |
| 695 | void main() |
| 696 | { |
| 697 | ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); |
| 698 | Tensor3DIterator weights_iter = CONVERT_TO_TENSOR3D_ITERATOR_NO_STEP(weights_attrs, weights_shift); |
| 699 | Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); |
| 700 | |
| 701 | #ifdef BIAS |
| 702 | VectorIterator biases_iter = CONVERT_TO_VECTOR_ITERATOR_NO_STEP(biases_attrs, biases_shift); |
| 703 | #endif /* BIAS */ |
| 704 | |
| 705 | uint z_base_index = uint(gl_GlobalInvocationID.z) << uint(1); |
| 706 | |
| 707 | // store orginal src current offset |
| 708 | int s_offset_in_bytes = src_iter.current_offset_in_bytes; |
| 709 | |
| 710 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_base_index * weights_stride_w); |
| 711 | |
| 712 | for(int z = 0; z < 2; ++z) |
| 713 | { |
| 714 | uint z_index = z_base_index + uint(z); |
| 715 | |
| 716 | src_iter.current_offset_in_bytes = s_offset_in_bytes; |
| 717 | |
| 718 | vec4 pixels[2]; |
| 719 | pixels[0] = vec4(0.f); |
| 720 | pixels[1] = vec4(0.f); |
| 721 | |
| 722 | #ifdef WEIGHTS_OPTIMIZATION |
| 723 | float w1, w2; |
| 724 | int nums = (int(weights_depth)) / 2; |
| 725 | for(int d = 0; d < nums; ++d) |
| 726 | { |
| 727 | vec2 vec2_w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter); |
| 728 | |
| 729 | w1 = vec2_w.x; |
| 730 | vec4 r1[2] = CONVOLVE(src_iter, w1); |
| 731 | pixels[0] += r1[0]; |
| 732 | pixels[1] += r1[1]; |
| 733 | |
| 734 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 735 | |
| 736 | w2 = vec2_w.y; |
| 737 | vec4 r2[2] = CONVOLVE(src_iter, w2); |
| 738 | pixels[0] += r2[0]; |
| 739 | pixels[1] += r2[1]; |
| 740 | |
| 741 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 742 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z * uint(2)); |
| 743 | } |
| 744 | #else /* WEIGHTS_OPTIMIZATION */ |
| 745 | float w; |
| 746 | for(int d = 0; d < int(weights_depth); ++d) |
| 747 | { |
| 748 | w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter).x; |
| 749 | |
| 750 | vec4 r[2] = CONVOLVE(src_iter, w); |
| 751 | pixels[0] += r[0]; |
| 752 | pixels[1] += r[1]; |
| 753 | |
| 754 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 755 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z); |
| 756 | } |
| 757 | #endif /* WEIGHTS_OPTIMIZATION */ |
| 758 | |
| 759 | #ifdef BIAS |
| 760 | vec2 vec2_b; |
| 761 | float b; |
| 762 | |
| 763 | vec2_b = LOAD_UNPACK2_HALF(biases_ptr, VECTOR_OFFSET(biases_iter, z_index)); |
| 764 | |
| 765 | if(z_index % uint(2) == uint(0)) |
| 766 | { |
| 767 | b = vec2_b.x; |
| 768 | } |
| 769 | else |
| 770 | { |
| 771 | b = vec2_b.y; |
| 772 | } |
| 773 | |
| 774 | pixels[0] += b; |
| 775 | pixels[1] += b; |
| 776 | #endif /* BIAS */ |
| 777 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 778 | #ifdef FUSED_ACTIVATION |
| 779 | pixels[0] = ACT_OP(pixels[0]); |
| 780 | pixels[1] = ACT_OP(pixels[1]); |
| 781 | #endif /* FUSED_ACTIVATION */ |
| 782 | |
Xinghang Zhou | 4af62a0 | 2017-11-02 16:37:24 +0800 | [diff] [blame] | 783 | STORE_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, pixels[0]); |
| 784 | STORE_PACK4_HALF(dst_ptr, TENSOR3D_OFFSET(dst_iter, 0, 1, 0), pixels[1]); |
| 785 | |
| 786 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(dst_iter, dst_attrs.stride_z); |
| 787 | } |
| 788 | } |
| 789 | #elif defined(PROCESS_8X_1Y_1Z) |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 790 | TENSOR_DECLARATION(1, srcBuffer, uvec4, src_ptr, src_shift, 4, readonly); |
| 791 | TENSOR_DECLARATION(2, dstBuffer, uvec4, dst_ptr, dst_shift, 4, writeonly); |
| 792 | TENSOR_DECLARATION(3, weightsBuffer, uint, weights_ptr, weights_shift, 2, readonly); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 793 | #ifdef BIAS |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 794 | TENSOR_DECLARATION(4, biasesBuffer, uint, biases_ptr, biases_shift, 2, readonly); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 795 | #endif /* BIAS */ |
| 796 | |
| 797 | #if STRIDE_X == 2 |
| 798 | #define CONVOLVE(s, w) convolve_stride2(s, w) |
| 799 | #elif STRIDE_X == 1 /* STRIDE_X == 1 */ |
| 800 | #define CONVOLVE(s, w) convolve_stride1(s, w) |
| 801 | #else /* STRIDE_X not equals 1 or 2 */ |
| 802 | #error STRIDE_X larger than 2 is not supported |
| 803 | #endif /* STRIDE_X == 2 */ |
| 804 | |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 805 | vec4[2] convolve_stride1(ImageIterator src_iter, float w) |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 806 | { |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 807 | vec4 s[2]; |
| 808 | s = LOAD_UNPACK8_CURRENT_ITEM_HALF(src_ptr, src_iter); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 809 | |
| 810 | s[0] *= w; |
| 811 | s[1] *= w; |
| 812 | |
| 813 | return s; |
| 814 | } |
| 815 | |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 816 | vec4[2] convolve_stride2(ImageIterator src_iter, float w) |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 817 | { |
Xinghang Zhou | 4af62a0 | 2017-11-02 16:37:24 +0800 | [diff] [blame] | 818 | vec4 s1[2]; |
| 819 | vec4 s2[2]; |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 820 | vec4 r[2]; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 821 | |
Xinghang Zhou | 4af62a0 | 2017-11-02 16:37:24 +0800 | [diff] [blame] | 822 | s1 = LOAD_UNPACK8_CURRENT_ITEM_HALF(src_ptr, src_iter); |
| 823 | r[0] = vec4(s1[0].xz, s1[1].xz); |
| 824 | s2 = LOAD_UNPACK8_HALF(src_ptr, IMAGE_OFFSET(src_iter, 8, 0)); |
| 825 | r[1] = vec4(s2[0].xz, s2[1].xz); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 826 | |
| 827 | r[0] *= w; |
| 828 | r[1] *= w; |
| 829 | |
| 830 | return r; |
| 831 | } |
| 832 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 833 | void main() |
| 834 | { |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 835 | ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); |
| 836 | Tensor3DIterator weights_iter = CONVERT_TO_TENSOR3D_ITERATOR_NO_STEP(weights_attrs, weights_shift); |
| 837 | Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 838 | |
| 839 | #ifdef BIAS |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 840 | VectorIterator biases_iter = CONVERT_TO_VECTOR_ITERATOR_NO_STEP(biases_attrs, biases_shift); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 841 | #endif /* BIAS */ |
| 842 | |
| 843 | vec4 pixels[2]; |
| 844 | pixels[0] = vec4(0.f); |
| 845 | pixels[1] = vec4(0.f); |
| 846 | |
| 847 | uint z_index = gl_GlobalInvocationID.z; |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 848 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_index * weights_stride_w); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 849 | |
zhenglin | 666635c | 2017-12-04 14:38:09 +0800 | [diff] [blame] | 850 | #ifdef WEIGHTS_OPTIMIZATION |
| 851 | float w1, w2; |
| 852 | int nums = (int(weights_depth)) / 2; |
| 853 | for(int d = 0; d < nums; ++d) |
| 854 | { |
| 855 | vec2 vec2_w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter); |
| 856 | |
| 857 | w1 = vec2_w.x; |
| 858 | vec4 r1[2] = CONVOLVE(src_iter, w1); |
| 859 | pixels[0] += r1[0]; |
| 860 | pixels[1] += r1[1]; |
| 861 | |
| 862 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 863 | |
| 864 | w2 = vec2_w.y; |
| 865 | vec4 r2[2] = CONVOLVE(src_iter, w2); |
| 866 | pixels[0] += r2[0]; |
| 867 | pixels[1] += r2[1]; |
| 868 | |
| 869 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 870 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z * uint(2)); |
| 871 | } |
| 872 | #else /* WEIGHTS_OPTIMIZATION */ |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 873 | float w; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 874 | for(int d = 0; d < int(weights_depth); ++d) |
| 875 | { |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 876 | w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter).x; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 877 | |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 878 | vec4 r[2] = CONVOLVE(src_iter, w); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 879 | pixels[0] += r[0]; |
| 880 | pixels[1] += r[1]; |
| 881 | |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 882 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 883 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 884 | } |
zhenglin | 666635c | 2017-12-04 14:38:09 +0800 | [diff] [blame] | 885 | #endif /* WEIGHTS_OPTIMIZATION */ |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 886 | |
| 887 | #ifdef BIAS |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 888 | vec2 vec2_b; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 889 | float b; |
| 890 | |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 891 | vec2_b = LOAD_UNPACK2_HALF(biases_ptr, VECTOR_OFFSET(biases_iter, z_index)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 892 | |
| 893 | if(z_index % uint(2) == uint(0)) |
| 894 | { |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 895 | b = vec2_b.x; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 896 | } |
| 897 | else |
| 898 | { |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 899 | b = vec2_b.y; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 900 | } |
| 901 | |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 902 | pixels[0] += b; |
| 903 | pixels[1] += b; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 904 | #endif /* BIAS */ |
| 905 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 906 | #ifdef FUSED_ACTIVATION |
| 907 | pixels[0] = ACT_OP(pixels[0]); |
| 908 | pixels[1] = ACT_OP(pixels[1]); |
| 909 | #endif /* FUSED_ACTIVATION */ |
| 910 | |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 911 | STORE_PACK8_CURRENT_ITEM_HALF(dst_ptr, dst_iter, pixels); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 912 | } |
Xinghang Zhou | 4af62a0 | 2017-11-02 16:37:24 +0800 | [diff] [blame] | 913 | #elif defined(PROCESS_8X_2Y_1Z) |
| 914 | TENSOR_DECLARATION(1, srcBuffer, uvec4, src_ptr, src_shift, 4, readonly); |
| 915 | TENSOR_DECLARATION(2, dstBuffer, uvec4, dst_ptr, dst_shift, 4, writeonly); |
| 916 | TENSOR_DECLARATION(3, weightsBuffer, uint, weights_ptr, weights_shift, 2, readonly); |
| 917 | #ifdef BIAS |
| 918 | TENSOR_DECLARATION(4, biasesBuffer, uint, biases_ptr, biases_shift, 2, readonly); |
| 919 | #endif /* BIAS */ |
| 920 | |
| 921 | #if STRIDE_X == 2 |
| 922 | #define CONVOLVE(s, w, x1, y1) convolve_stride2(s, w, x1, y1) |
| 923 | #elif STRIDE_X == 1 /* STRIDE_X == 1 */ |
| 924 | #define CONVOLVE(s, w, x1, y1) convolve_stride1(s, w, x1, y1) |
| 925 | #else /* STRIDE_X not equals 1 or 2 */ |
| 926 | #error STRIDE_X larger than 2 is not supported |
| 927 | #endif /* STRIDE_X == 2 */ |
| 928 | |
| 929 | vec4[2] convolve_stride1(ImageIterator src_iter, float w, int x1, int y1) |
| 930 | { |
| 931 | vec4 s[2]; |
| 932 | s = LOAD_UNPACK8_HALF(src_ptr, IMAGE_OFFSET(src_iter, x1, y1)); |
| 933 | |
| 934 | s[0] *= w; |
| 935 | s[1] *= w; |
| 936 | |
| 937 | return s; |
| 938 | } |
| 939 | |
| 940 | vec4[2] convolve_stride2(ImageIterator src_iter, float w, int x1, int y1) |
| 941 | { |
| 942 | vec4 s1[2]; |
| 943 | vec4 s2[2]; |
| 944 | vec4 r[2]; |
| 945 | |
| 946 | s1 = LOAD_UNPACK8_HALF(src_ptr, IMAGE_OFFSET(src_iter, x1, y1)); |
| 947 | r[0] = vec4(s1[0].xz, s1[1].xz); |
| 948 | s2 = LOAD_UNPACK8_HALF(src_ptr, IMAGE_OFFSET(src_iter, (8 + x1), y1)); |
| 949 | r[1] = vec4(s2[0].xz, s2[1].xz); |
| 950 | |
| 951 | r[0] *= w; |
| 952 | r[1] *= w; |
| 953 | |
| 954 | return r; |
| 955 | } |
| 956 | |
| 957 | void main() |
| 958 | { |
| 959 | ImageIterator src_iter = CONVERT_TO_IMAGE_ITERATOR(src_attrs, src_shift); |
| 960 | Tensor3DIterator weights_iter = CONVERT_TO_TENSOR3D_ITERATOR_NO_STEP(weights_attrs, weights_shift); |
| 961 | Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); |
| 962 | |
| 963 | #ifdef BIAS |
| 964 | VectorIterator biases_iter = CONVERT_TO_VECTOR_ITERATOR_NO_STEP(biases_attrs, biases_shift); |
| 965 | #endif /* BIAS */ |
| 966 | |
| 967 | vec4 pixels[2]; |
| 968 | vec4 pixels1[2]; |
| 969 | pixels[0] = vec4(0.f); |
| 970 | pixels[1] = vec4(0.f); |
| 971 | pixels1[0] = vec4(0.f); |
| 972 | pixels1[1] = vec4(0.f); |
| 973 | |
| 974 | uint z_index = gl_GlobalInvocationID.z; |
| 975 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_index * weights_stride_w); |
| 976 | |
| 977 | #ifdef WEIGHTS_OPTIMIZATION |
| 978 | float w1, w2; |
| 979 | int nums = (int(weights_depth)) / 2; |
| 980 | for(int d = 0; d < nums; ++d) |
| 981 | { |
| 982 | vec2 vec2_w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter); |
| 983 | |
| 984 | w1 = vec2_w.x; |
| 985 | vec4 r1[2] = CONVOLVE(src_iter, w1, 0, 0); |
| 986 | vec4 r2[2] = CONVOLVE(src_iter, w1, 0, (int(STRIDE_Y))); |
| 987 | pixels[0] += r1[0]; |
| 988 | pixels[1] += r1[1]; |
| 989 | pixels1[0] += r2[0]; |
| 990 | pixels1[1] += r2[1]; |
| 991 | |
| 992 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 993 | |
| 994 | w2 = vec2_w.y; |
| 995 | vec4 r3[2] = CONVOLVE(src_iter, w2, 0, 0); |
| 996 | vec4 r4[2] = CONVOLVE(src_iter, w2, 0, (int(STRIDE_Y))); |
| 997 | pixels[0] += r3[0]; |
| 998 | pixels[1] += r3[1]; |
| 999 | pixels1[0] += r4[0]; |
| 1000 | pixels1[1] += r4[1]; |
| 1001 | |
| 1002 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 1003 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z * uint(2)); |
| 1004 | } |
| 1005 | #else /* WEIGHTS_OPTIMIZATION */ |
| 1006 | float w; |
| 1007 | for(int d = 0; d < int(weights_depth); ++d) |
| 1008 | { |
| 1009 | w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter).x; |
| 1010 | |
| 1011 | vec4 r1[2] = CONVOLVE(src_iter, w, 0, 0); |
| 1012 | vec4 r2[2] = CONVOLVE(src_iter, w, 0, (int(STRIDE_Y))); |
| 1013 | pixels[0] += r1[0]; |
| 1014 | pixels[1] += r1[1]; |
| 1015 | pixels1[0] += r2[0]; |
| 1016 | pixels1[1] += r2[1]; |
| 1017 | |
| 1018 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); |
| 1019 | TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z); |
| 1020 | } |
| 1021 | #endif /* WEIGHTS_OPTIMIZATION */ |
| 1022 | |
| 1023 | #ifdef BIAS |
| 1024 | vec2 vec2_b; |
| 1025 | float b; |
| 1026 | |
| 1027 | vec2_b = LOAD_UNPACK2_HALF(biases_ptr, VECTOR_OFFSET(biases_iter, z_index)); |
| 1028 | |
| 1029 | if(z_index % uint(2) == uint(0)) |
| 1030 | { |
| 1031 | b = vec2_b.x; |
| 1032 | } |
| 1033 | else |
| 1034 | { |
| 1035 | b = vec2_b.y; |
| 1036 | } |
| 1037 | |
| 1038 | pixels[0] += b; |
| 1039 | pixels[1] += b; |
| 1040 | pixels1[0] += b; |
| 1041 | pixels1[1] += b; |
| 1042 | #endif /* BIAS */ |
| 1043 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 1044 | #ifdef FUSED_ACTIVATION |
| 1045 | pixels[0] = ACT_OP(pixels[0]); |
| 1046 | pixels[1] = ACT_OP(pixels[1]); |
| 1047 | pixels1[0] = ACT_OP(pixels1[0]); |
| 1048 | pixels1[1] = ACT_OP(pixels1[1]); |
| 1049 | #endif /* FUSED_ACTIVATION */ |
| 1050 | |
Xinghang Zhou | 4af62a0 | 2017-11-02 16:37:24 +0800 | [diff] [blame] | 1051 | STORE_PACK8_CURRENT_ITEM_HALF(dst_ptr, dst_iter, pixels); |
| 1052 | STORE_PACK8_HALF(dst_ptr, TENSOR3D_OFFSET(dst_iter, 0, 1, 0), pixels1); |
| 1053 | } |
| 1054 | #endif /* PROCESS_4X_1Y_1Z */ |
Joel Liang | 6387543 | 2018-01-02 14:05:06 +0800 | [diff] [blame] | 1055 | #else /* DATA_TYPE_FP32 */ |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 1056 | #error Data type not supported |
| 1057 | #endif /* DATA_TYPE_FP32 */ |