SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 1 | /* |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame] | 2 | * Copyright (c) 2016-2018 ARM Limited. |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +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 | #include "helpers.h" |
| 25 | |
steniu01 | db00668 | 2017-08-09 16:26:22 +0100 | [diff] [blame] | 26 | #undef CONVERT_SAT |
Michalis Spyrou | def665a | 2017-08-14 11:26:37 +0100 | [diff] [blame] | 27 | |
| 28 | #define ADD_OP(a, b) ((a) + (b)) |
| 29 | #define MUL_OP(a, b) ((a) * (b)) |
| 30 | #define CONVERT_SAT(a, b) ((a)) |
| 31 | |
Gian Marco Iodice | 1c8409d | 2017-09-06 17:24:25 +0100 | [diff] [blame] | 32 | #if defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) |
| 33 | |
Pablo Tello | 3d31946 | 2018-06-21 15:13:17 +0100 | [diff] [blame] | 34 | #if defined(DATA_LAYOUT_NHWC) |
| 35 | |
| 36 | #define PTR_TO_VALUE(PTR, DATA_TYPE) *((__global DATA_TYPE *)(PTR)) |
| 37 | |
| 38 | /** This kernel performs a direct convolution to convolve the low three dimensions of a tensor with data layout NHWC |
| 39 | * |
| 40 | * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float |
| 41 | * @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32 |
| 42 | * @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1 |
| 43 | * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH |
| 44 | * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. |
| 45 | * |
| 46 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 |
| 47 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 48 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 49 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 50 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 51 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 52 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 53 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 54 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 55 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 56 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 57 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 58 | * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) |
| 59 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 60 | * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
| 61 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 62 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr |
| 63 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 64 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 65 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 66 | * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) |
| 67 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 68 | * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) |
| 69 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 70 | * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr |
| 71 | * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) |
| 72 | * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 73 | * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor |
| 74 | * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension |
| 75 | */ |
| 76 | __kernel void direct_convolution1x1_nhwc( |
| 77 | TENSOR3D_DECLARATION(src), |
| 78 | TENSOR3D_DECLARATION(dst), |
| 79 | TENSOR3D_DECLARATION(weights), |
| 80 | #ifdef HAS_BIAS |
| 81 | VECTOR_DECLARATION(biases), |
| 82 | #endif /* defined(HAS_BIAS) */ |
| 83 | unsigned int weights_stride_w) |
| 84 | { |
| 85 | Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| 86 | Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
| 87 | Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); |
| 88 | |
| 89 | #ifdef HAS_BIAS |
| 90 | Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| 91 | #endif /* defined(HAS_BIAS) */ |
| 92 | |
| 93 | VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8) |
| 94 | values = 0; |
| 95 | const int id0 = get_global_id(0); |
| 96 | const int id1 = get_global_id(1); |
| 97 | const int id2 = get_global_id(2); |
| 98 | weights.ptr += id0 * weights_stride_w; |
| 99 | __global uchar *src_addr = (__global uchar *)offset(&src, 0, 0) - src_stride_x * id0 + id2 * STRIDE_Y * (int)src_stride_z; |
| 100 | |
| 101 | for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d) |
| 102 | { |
| 103 | DATA_TYPE weight = *(__global DATA_TYPE *)weights.ptr; |
| 104 | #if STRIDE_X == 1 |
| 105 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 106 | col0 = (VEC_DATA_TYPE(DATA_TYPE, 8))( |
| 107 | PTR_TO_VALUE(src_addr + 0 * src_stride_y, DATA_TYPE), |
| 108 | PTR_TO_VALUE(src_addr + 1 * src_stride_y, DATA_TYPE), |
| 109 | PTR_TO_VALUE(src_addr + 2 * src_stride_y, DATA_TYPE), |
| 110 | PTR_TO_VALUE(src_addr + 3 * src_stride_y, DATA_TYPE), |
| 111 | PTR_TO_VALUE(src_addr + 4 * src_stride_y, DATA_TYPE), |
| 112 | PTR_TO_VALUE(src_addr + 5 * src_stride_y, DATA_TYPE), |
| 113 | PTR_TO_VALUE(src_addr + 6 * src_stride_y, DATA_TYPE), |
| 114 | PTR_TO_VALUE(src_addr + 7 * src_stride_y, DATA_TYPE)); |
| 115 | #elif STRIDE_X == 2 /* STRIDE_X == 1 */ |
| 116 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 117 | col0 = (VEC_DATA_TYPE(DATA_TYPE, 8))( |
| 118 | PTR_TO_VALUE(src_addr + 0 * src_stride_y, DATA_TYPE), |
| 119 | PTR_TO_VALUE(src_addr + 2 * src_stride_y, DATA_TYPE), |
| 120 | PTR_TO_VALUE(src_addr + 4 * src_stride_y, DATA_TYPE), |
| 121 | PTR_TO_VALUE(src_addr + 6 * src_stride_y, DATA_TYPE), |
| 122 | PTR_TO_VALUE(src_addr + 8 * src_stride_y, DATA_TYPE), |
| 123 | PTR_TO_VALUE(src_addr + 10 * src_stride_y, DATA_TYPE), |
| 124 | PTR_TO_VALUE(src_addr + 12 * src_stride_y, DATA_TYPE), |
| 125 | PTR_TO_VALUE(src_addr + 14 * src_stride_y, DATA_TYPE)); |
| 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 | values = ADD_OP(values, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))weight, col0)); |
| 130 | |
| 131 | src_addr += src_stride_x; |
| 132 | weights.ptr += weights_stride_x; |
| 133 | } |
| 134 | |
| 135 | #ifdef HAS_BIAS |
| 136 | values = ADD_OP(values, (VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, id0)))); |
| 137 | #endif /* defined(HAS_BIAS) */ |
| 138 | |
| 139 | *((__global DATA_TYPE *)dst.ptr) = values.s0; |
| 140 | *((__global DATA_TYPE *)(dst.ptr + 1 * dst_stride_y)) = values.s1; |
| 141 | *((__global DATA_TYPE *)(dst.ptr + 2 * dst_stride_y)) = values.s2; |
| 142 | *((__global DATA_TYPE *)(dst.ptr + 3 * dst_stride_y)) = values.s3; |
| 143 | *((__global DATA_TYPE *)(dst.ptr + 4 * dst_stride_y)) = values.s4; |
| 144 | *((__global DATA_TYPE *)(dst.ptr + 5 * dst_stride_y)) = values.s5; |
| 145 | *((__global DATA_TYPE *)(dst.ptr + 6 * dst_stride_y)) = values.s6; |
| 146 | *((__global DATA_TYPE *)(dst.ptr + 7 * dst_stride_y)) = values.s7; |
| 147 | } |
| 148 | #endif // defined(DATA_LAYOUT_NHWC) |
| 149 | |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 150 | #if STRIDE_X == 3 |
| 151 | #define INPUT_PIXEL_STR(data_size) extract_input_stride3_##data_size |
| 152 | #define INPUT_PIXEL(data_size) INPUT_PIXEL_STR(data_size) |
| 153 | #elif STRIDE_X == 2 |
| 154 | #define INPUT_PIXEL(data_size) extract_input_stride2 |
| 155 | #elif STRIDE_X == 1 |
| 156 | #define INPUT_PIXEL(data_size) extract_input_stride1 |
| 157 | #else /* STRIDE_X not equals 1, 2 or 3 */ |
| 158 | #error "Only support strides 1, 2 and 3" |
| 159 | #endif /* STRIDE_X == 3 */ |
| 160 | |
| 161 | /** Extracts a 1D horizontal vector from the input tensor with stride as 1. |
| 162 | * |
| 163 | * @param[in] input_pixel Pointer to the first pixel. |
| 164 | * |
Pablo Tello | 3d31946 | 2018-06-21 15:13:17 +0100 | [diff] [blame] | 165 | * @return extracted input values. |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 166 | */ |
| 167 | inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride1(__global const DATA_TYPE *input_pixel) |
| 168 | { |
| 169 | return vload8(0, input_pixel); |
| 170 | } |
| 171 | |
| 172 | /** Extracts a 1D horizontal vector from the input tensor with stride as 2. |
| 173 | * |
| 174 | * @param[in] input_pixel Pointer to the first pixel. |
| 175 | * |
Pablo Tello | 3d31946 | 2018-06-21 15:13:17 +0100 | [diff] [blame] | 176 | * @return extracted input values. |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 177 | */ |
| 178 | inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride2(__global const DATA_TYPE *input_pixel) |
| 179 | { |
| 180 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 181 | temp = vload16(0, input_pixel); |
| 182 | return temp.s02468ace; |
| 183 | } |
| 184 | |
| 185 | /** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 32-bit data size. |
| 186 | * |
| 187 | * @param[in] input_pixel Pointer to the first pixel. |
| 188 | * |
Pablo Tello | 3d31946 | 2018-06-21 15:13:17 +0100 | [diff] [blame] | 189 | * @return extracted input values. |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 190 | */ |
| 191 | inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_32(__global const DATA_TYPE *input_pixel) |
| 192 | { |
| 193 | VEC_DATA_TYPE(DATA_TYPE, 4) |
| 194 | temp1 = vload4(0, input_pixel); |
| 195 | VEC_DATA_TYPE(DATA_TYPE, 4) |
| 196 | temp2 = vload4(0, input_pixel + 6); |
| 197 | VEC_DATA_TYPE(DATA_TYPE, 4) |
| 198 | temp3 = vload4(0, input_pixel + 12); |
| 199 | VEC_DATA_TYPE(DATA_TYPE, 4) |
| 200 | temp4 = vload4(0, input_pixel + 18); |
| 201 | return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s03, temp2.s03, temp3.s03, temp4.s03); |
| 202 | } |
| 203 | |
| 204 | /** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 16-bit data size. |
| 205 | * |
| 206 | * @param[in] input_pixel Pointer to the first pixel. |
| 207 | * |
Pablo Tello | 3d31946 | 2018-06-21 15:13:17 +0100 | [diff] [blame] | 208 | * @return extracted input values. |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 209 | */ |
| 210 | inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_16(__global const DATA_TYPE *input_pixel) |
| 211 | { |
| 212 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 213 | temp1 = vload8(0, input_pixel); |
| 214 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 215 | temp2 = vload8(0, input_pixel + 8); |
| 216 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 217 | temp3 = vload8(0, input_pixel + 16); |
| 218 | return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s036, temp2.s147, temp3.s25); |
| 219 | } |
| 220 | |
| 221 | /** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 8-bit data size. |
| 222 | * |
| 223 | * @param[in] input_pixel Pointer to the first pixel. |
| 224 | * |
Pablo Tello | 3d31946 | 2018-06-21 15:13:17 +0100 | [diff] [blame] | 225 | * @return extracted input values. |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 226 | */ |
| 227 | inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_8(__global const DATA_TYPE *input_pixel) |
| 228 | { |
| 229 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 230 | temp1 = vload16(0, input_pixel); |
| 231 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 232 | temp2 = vload16(0, input_pixel + 12); |
| 233 | return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s0369, temp2.s0369); |
| 234 | } |
| 235 | |
| 236 | /** This kernel performs a direct convolution to convolve the low three dimensions. |
| 237 | * |
| 238 | * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float |
| 239 | * @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32 |
Gian Marco Iodice | 5cb4d6a | 2017-08-08 10:53:00 +0100 | [diff] [blame] | 240 | * @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1 |
| 241 | * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 242 | * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. |
| 243 | * |
Gian Marco Iodice | 5cb4d6a | 2017-08-08 10:53:00 +0100 | [diff] [blame] | 244 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 245 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 246 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 247 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 248 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 249 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 250 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 251 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 252 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 253 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 254 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 255 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 256 | * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) |
| 257 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 258 | * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
| 259 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 260 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 261 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 262 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 263 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 264 | * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) |
| 265 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 266 | * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) |
| 267 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 268 | * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr |
| 269 | * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) |
| 270 | * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 271 | * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor |
Gian Marco Iodice | 5cb4d6a | 2017-08-08 10:53:00 +0100 | [diff] [blame] | 272 | * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 273 | */ |
| 274 | __kernel void direct_convolution1x1( |
| 275 | TENSOR3D_DECLARATION(src), |
| 276 | TENSOR3D_DECLARATION(dst), |
| 277 | TENSOR3D_DECLARATION(weights), |
| 278 | #ifdef HAS_BIAS |
| 279 | VECTOR_DECLARATION(biases), |
| 280 | #endif /* defined(HAS_BIAS) */ |
Gian Marco Iodice | 5cb4d6a | 2017-08-08 10:53:00 +0100 | [diff] [blame] | 281 | unsigned int weights_stride_w) |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 282 | { |
| 283 | Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| 284 | Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
| 285 | Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); |
| 286 | |
| 287 | #ifdef HAS_BIAS |
| 288 | Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| 289 | #endif /* defined(HAS_BIAS) */ |
| 290 | |
Michalis Spyrou | def665a | 2017-08-14 11:26:37 +0100 | [diff] [blame] | 291 | VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8) |
Pablo Tello | 3d31946 | 2018-06-21 15:13:17 +0100 | [diff] [blame] | 292 | values = 0; |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 293 | |
| 294 | const uint z_index = get_global_id(2); |
| 295 | |
| 296 | weights.ptr += z_index * weights_stride_w; |
Gian Marco Iodice | 744b5ed | 2017-10-06 15:44:27 +0100 | [diff] [blame] | 297 | for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d) |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 298 | { |
| 299 | DATA_TYPE weight = *(__global DATA_TYPE *)weights.ptr; |
| 300 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 301 | input_pixel = INPUT_PIXEL(DATA_SIZE)((__global DATA_TYPE *)src.ptr); |
Pablo Tello | 3d31946 | 2018-06-21 15:13:17 +0100 | [diff] [blame] | 302 | values = ADD_OP(values, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))weight, input_pixel)); |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 303 | src.ptr += src_stride_z; |
| 304 | weights.ptr += weights_stride_z; |
| 305 | } |
| 306 | |
| 307 | #ifdef HAS_BIAS |
Pablo Tello | 3d31946 | 2018-06-21 15:13:17 +0100 | [diff] [blame] | 308 | values = ADD_OP(values, (VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, z_index)))); |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 309 | #endif /* defined(HAS_BIAS) */ |
| 310 | |
Pablo Tello | 3d31946 | 2018-06-21 15:13:17 +0100 | [diff] [blame] | 311 | vstore8(CONVERT_SAT(values, VEC_DATA_TYPE(DATA_TYPE, 8)), 0, (__global DATA_TYPE *)dst.ptr); |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 312 | } |
steniu01 | db00668 | 2017-08-09 16:26:22 +0100 | [diff] [blame] | 313 | #endif // defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) |
Gian Marco Iodice | 1c8409d | 2017-09-06 17:24:25 +0100 | [diff] [blame] | 314 | |
| 315 | #if defined(WEIGHTS_DEPTH) |
| 316 | |
| 317 | #define CONVOLUTION1x1_BIFROST(acc, src, weight_value) \ |
| 318 | ({ \ |
| 319 | acc.s0 = mad(src.s0, weight_value, acc.s0); \ |
| 320 | acc.s1 = mad(src.s1, weight_value, acc.s1); \ |
| 321 | acc.s2 = mad(src.s2, weight_value, acc.s2); \ |
| 322 | acc.s3 = mad(src.s3, weight_value, acc.s3); \ |
| 323 | }) |
| 324 | |
| 325 | /** An optimized direct convolution 1x1 OpenCL kernel for Bifrost architectures when the data type is F32 |
| 326 | * |
| 327 | * @note This OpenCL kernel works only with stride_x and stride_y equal to 1 |
| 328 | * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH |
| 329 | * @note In case biases, -DHAS_BIAS must to be passed at compile |
| 330 | * |
| 331 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 332 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 333 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 334 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 335 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 336 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 337 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 338 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 339 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 340 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 341 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 342 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 343 | * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) |
| 344 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 345 | * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
| 346 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
Joel Liang | f1f3ebd | 2017-11-10 09:59:19 +0800 | [diff] [blame] | 347 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr |
Gian Marco Iodice | 1c8409d | 2017-09-06 17:24:25 +0100 | [diff] [blame] | 348 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 349 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 350 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 351 | * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) |
| 352 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 353 | * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) |
| 354 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 355 | * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr |
| 356 | * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) |
| 357 | * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 358 | * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor |
| 359 | * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension |
| 360 | */ |
| 361 | __kernel void direct_convolution1x1_f32_bifrost( |
| 362 | TENSOR3D_DECLARATION(src), |
| 363 | TENSOR3D_DECLARATION(dst), |
| 364 | TENSOR3D_DECLARATION(weights), |
| 365 | #ifdef HAS_BIAS |
| 366 | VECTOR_DECLARATION(biases), |
| 367 | #endif /* defined(HAS_BIAS) */ |
| 368 | unsigned int weights_stride_w) |
| 369 | { |
| 370 | // Get the kernel index |
| 371 | const int kernel_index = get_global_id(2); |
| 372 | |
| 373 | Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| 374 | Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); |
| 375 | |
| 376 | float4 acc0 = 0.0f; |
| 377 | float4 acc1 = 0.0f; |
| 378 | float4 acc2 = 0.0f; |
| 379 | float4 acc3 = 0.0f; |
| 380 | |
| 381 | __global uchar *weights_addr = (__global uchar *)(weights_ptr + weights_offset_first_element_in_bytes + kernel_index * weights_stride_w); |
| 382 | __global uchar *src_addr = (__global uchar *)offset(&src, 0, 0); |
| 383 | |
| 384 | for(ushort d = 0; d < (ushort)WEIGHTS_DEPTH; ++d) |
| 385 | { |
| 386 | // Load the weights |
| 387 | float weight = *((__global float *)weights_addr); |
| 388 | |
| 389 | // Load values from row0 of input tensor |
| 390 | float4 src0 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y)); |
| 391 | float4 src1 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y)); |
| 392 | float4 src2 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y)); |
| 393 | float4 src3 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y)); |
| 394 | |
| 395 | CONVOLUTION1x1_BIFROST(acc0, src0, weight); |
| 396 | CONVOLUTION1x1_BIFROST(acc1, src1, weight); |
| 397 | CONVOLUTION1x1_BIFROST(acc2, src2, weight); |
| 398 | CONVOLUTION1x1_BIFROST(acc3, src3, weight); |
| 399 | |
| 400 | src_addr += src_stride_z; |
| 401 | weights_addr += weights_stride_z; |
| 402 | } |
| 403 | |
| 404 | #ifdef HAS_BIAS |
| 405 | Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| 406 | |
| 407 | float bias = (float) * ((__global float *)(vector_offset(&biases, kernel_index))); |
| 408 | |
| 409 | acc0.s0 += bias; |
| 410 | acc0.s1 += bias; |
| 411 | acc0.s2 += bias; |
| 412 | acc0.s3 += bias; |
| 413 | acc1.s0 += bias; |
| 414 | acc1.s1 += bias; |
| 415 | acc1.s2 += bias; |
| 416 | acc1.s3 += bias; |
| 417 | acc2.s0 += bias; |
| 418 | acc2.s1 += bias; |
| 419 | acc2.s2 += bias; |
| 420 | acc2.s3 += bias; |
| 421 | acc3.s0 += bias; |
| 422 | acc3.s1 += bias; |
| 423 | acc3.s2 += bias; |
| 424 | acc3.s3 += bias; |
| 425 | #endif /* defined(HAS_BIAS) */ |
| 426 | |
| 427 | vstore4(acc0, 0, (__global float *)(dst.ptr + 0 * dst_stride_y)); |
| 428 | vstore4(acc1, 0, (__global float *)(dst.ptr + 1 * dst_stride_y)); |
| 429 | vstore4(acc2, 0, (__global float *)(dst.ptr + 2 * dst_stride_y)); |
| 430 | vstore4(acc3, 0, (__global float *)(dst.ptr + 3 * dst_stride_y)); |
| 431 | } |
Pablo Tello | 3d31946 | 2018-06-21 15:13:17 +0100 | [diff] [blame] | 432 | #endif // defined(WEIGHTS_DEPTH) |