Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | f6f7876 | 2020-07-06 11:27:21 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2020 Arm Limited. |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +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 | */ |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 24 | #include "helpers.h" |
| 25 | |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 26 | #include "activation_float_helpers.h" |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 27 | |
| 28 | /** Get the pointer position at a certain offset in x and y direction. |
| 29 | * |
| 30 | * @param[in] ptr Pointer to the starting position of the buffer |
| 31 | * @param[in] x Relative X position |
| 32 | * @param[in] y Relative Y position |
| 33 | * @param[in] stride_x Stride of the source tensor in X dimension (in bytes) |
| 34 | * @param[in] stride_y Stride of the source tensor in Y dimension (in bytes) |
| 35 | * |
| 36 | * @return a uchar |
| 37 | */ |
| 38 | inline __global uchar *ptr_offset(__global uchar *ptr, const int x, const int y, const int stride_x, const int stride_y) |
| 39 | { |
| 40 | return ptr + x * stride_x + y * stride_y; |
| 41 | } |
| 42 | |
| 43 | #if(DILATION_X == 1 && DILATION_Y == 1) |
| 44 | |
| 45 | #define CONVOLUTION1x3_BIFROST2X1_STRIDE1(acc, src0, weights_row0) \ |
| 46 | ({ \ |
| 47 | acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \ |
| 48 | acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \ |
| 49 | acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \ |
| 50 | acc.s1 = fma(src0.s1, weights_row0.s0, acc.s1); \ |
| 51 | acc.s1 = fma(src0.s2, weights_row0.s1, acc.s1); \ |
| 52 | acc.s1 = fma(src0.s3, weights_row0.s2, acc.s1); \ |
| 53 | }) |
| 54 | |
| 55 | #define CONVOLUTION1x3_BIFROST4X1_STRIDE1(acc, src0, weights_row0) \ |
| 56 | ({ \ |
| 57 | acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \ |
| 58 | acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \ |
| 59 | acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \ |
| 60 | acc.s1 = fma(src0.s1, weights_row0.s0, acc.s1); \ |
| 61 | acc.s1 = fma(src0.s2, weights_row0.s1, acc.s1); \ |
| 62 | acc.s1 = fma(src0.s3, weights_row0.s2, acc.s1); \ |
| 63 | acc.s2 = fma(src0.s2, weights_row0.s0, acc.s2); \ |
| 64 | acc.s2 = fma(src0.s3, weights_row0.s1, acc.s2); \ |
| 65 | acc.s2 = fma(src0.s4, weights_row0.s2, acc.s2); \ |
| 66 | acc.s3 = fma(src0.s3, weights_row0.s0, acc.s3); \ |
| 67 | acc.s3 = fma(src0.s4, weights_row0.s1, acc.s3); \ |
| 68 | acc.s3 = fma(src0.s5, weights_row0.s2, acc.s3); \ |
| 69 | }) |
| 70 | |
| 71 | #define CONVOLUTION1x3_BIFROST2X1_STRIDE2(acc, src0, src1, weights_row0) \ |
| 72 | ({ \ |
| 73 | acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \ |
| 74 | acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \ |
| 75 | acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \ |
| 76 | acc.s1 = fma(src0.s2, weights_row0.s0, acc.s1); \ |
| 77 | acc.s1 = fma(src0.s3, weights_row0.s1, acc.s1); \ |
| 78 | acc.s1 = fma(src1.s0, weights_row0.s2, acc.s1); \ |
| 79 | }) |
| 80 | |
| 81 | #define CONVOLUTION1x3_BIFROST4X1_STRIDE2(acc, src0, src1, weights_row0) \ |
| 82 | ({ \ |
| 83 | acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \ |
| 84 | acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \ |
| 85 | acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \ |
| 86 | acc.s1 = fma(src0.s2, weights_row0.s0, acc.s1); \ |
| 87 | acc.s1 = fma(src0.s3, weights_row0.s1, acc.s1); \ |
| 88 | acc.s1 = fma(src0.s4, weights_row0.s2, acc.s1); \ |
| 89 | acc.s2 = fma(src0.s4, weights_row0.s0, acc.s2); \ |
| 90 | acc.s2 = fma(src0.s5, weights_row0.s1, acc.s2); \ |
| 91 | acc.s2 = fma(src0.s6, weights_row0.s2, acc.s2); \ |
| 92 | acc.s3 = fma(src0.s6, weights_row0.s0, acc.s3); \ |
| 93 | acc.s3 = fma(src0.s7, weights_row0.s1, acc.s3); \ |
| 94 | acc.s3 = fma(src1.s0, weights_row0.s2, acc.s3); \ |
| 95 | }) |
| 96 | |
| 97 | #else /* DILATION_X==1 && DILATION_Y==1 */ |
| 98 | |
| 99 | #define CONVOLUTION1x3_BIFROST2X1_STRIDE1(acc, src0_left, src0_mid, src0_right, weights_row0) \ |
| 100 | ({ \ |
| 101 | acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0); \ |
| 102 | acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0); \ |
| 103 | acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0); \ |
| 104 | acc.s1 = fma(src0_left.s1, weights_row0.s0, acc.s1); \ |
| 105 | acc.s1 = fma(src0_mid.s1, weights_row0.s1, acc.s1); \ |
| 106 | acc.s1 = fma(src0_right.s1, weights_row0.s2, acc.s1); \ |
| 107 | }) |
| 108 | |
| 109 | #define CONVOLUTION1x3_BIFROST2X1_STRIDE2(acc, src0_left, src0_mid, src0_right, weights_row0) \ |
| 110 | ({ \ |
| 111 | acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0); \ |
| 112 | acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0); \ |
| 113 | acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0); \ |
| 114 | acc.s1 = fma(src0_left.s2, weights_row0.s0, acc.s1); \ |
| 115 | acc.s1 = fma(src0_mid.s2, weights_row0.s1, acc.s1); \ |
| 116 | acc.s1 = fma(src0_right.s2, weights_row0.s2, acc.s1); \ |
| 117 | }) |
| 118 | |
| 119 | #define CONVOLUTION1x3_BIFROST4X1_STRIDE1(acc, src0_left, src0_mid, src0_right, weights_row0) \ |
| 120 | ({ \ |
| 121 | acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0); \ |
| 122 | acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0); \ |
| 123 | acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0); \ |
| 124 | acc.s1 = fma(src0_left.s1, weights_row0.s0, acc.s1); \ |
| 125 | acc.s1 = fma(src0_mid.s1, weights_row0.s1, acc.s1); \ |
| 126 | acc.s1 = fma(src0_right.s1, weights_row0.s2, acc.s1); \ |
| 127 | acc.s2 = fma(src0_left.s2, weights_row0.s0, acc.s2); \ |
| 128 | acc.s2 = fma(src0_mid.s2, weights_row0.s1, acc.s2); \ |
| 129 | acc.s2 = fma(src0_right.s2, weights_row0.s2, acc.s2); \ |
| 130 | acc.s3 = fma(src0_left.s3, weights_row0.s0, acc.s3); \ |
| 131 | acc.s3 = fma(src0_mid.s3, weights_row0.s1, acc.s3); \ |
| 132 | acc.s3 = fma(src0_right.s3, weights_row0.s2, acc.s3); \ |
| 133 | }) |
| 134 | |
| 135 | #define CONVOLUTION1x3_BIFROST4X1_STRIDE2(acc, src0_left, src0_mid, src0_right, weights_row0) \ |
| 136 | ({ \ |
| 137 | acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0); \ |
| 138 | acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0); \ |
| 139 | acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0); \ |
| 140 | acc.s1 = fma(src0_left.s2, weights_row0.s0, acc.s1); \ |
| 141 | acc.s1 = fma(src0_mid.s2, weights_row0.s1, acc.s1); \ |
| 142 | acc.s1 = fma(src0_right.s2, weights_row0.s2, acc.s1); \ |
| 143 | acc.s2 = fma(src0_left.s4, weights_row0.s0, acc.s2); \ |
| 144 | acc.s2 = fma(src0_mid.s4, weights_row0.s1, acc.s2); \ |
| 145 | acc.s2 = fma(src0_right.s4, weights_row0.s2, acc.s2); \ |
| 146 | acc.s3 = fma(src0_left.s6, weights_row0.s0, acc.s3); \ |
| 147 | acc.s3 = fma(src0_mid.s6, weights_row0.s1, acc.s3); \ |
| 148 | acc.s3 = fma(src0_right.s6, weights_row0.s2, acc.s3); \ |
| 149 | }) |
| 150 | |
| 151 | #endif /* DILATION_X==1 && DILATION_Y==1 */ |
| 152 | |
| 153 | #if defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS) && defined(IS_F32) |
Giorgio Arena | 9fe4144 | 2017-08-23 16:36:24 +0100 | [diff] [blame] | 154 | #if defined(CONV_STRIDE_X) |
| 155 | |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 156 | #if CONV_STRIDE_X == 1 |
| 157 | #define convolution1x3 convolution1x3_stride_1 |
| 158 | #elif CONV_STRIDE_X == 2 |
| 159 | #define convolution1x3 convolution1x3_stride_2 |
| 160 | #elif CONV_STRIDE_X == 3 |
| 161 | #define convolution1x3 convolution1x3_stride_3 |
| 162 | #else /* CONV_STRIDE_X */ |
| 163 | #error "Stride not supported" |
| 164 | #endif /* CONV_STRIDE_X */ |
| 165 | |
| 166 | /** Compute a 1D horizontal convolution of size 3 and stride 1 for floating point type. |
| 167 | * |
| 168 | * @param[in] left_pixel Pointer to the left pixel. |
| 169 | * @param[in] left_coeff Weight of the left pixel |
| 170 | * @param[in] middle_coeff Weight of the middle pixel |
| 171 | * @param[in] right_coeff Weight of the right pixel |
| 172 | * |
| 173 | * @return a float2 containing 2 convoluted values. |
| 174 | */ |
| 175 | inline float2 convolution1x3_stride_1(__global const uchar *left_pixel, |
| 176 | const float left_coeff, |
| 177 | const float middle_coeff, |
| 178 | const float right_coeff) |
| 179 | { |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 180 | #if(DILATION_X == 1 && DILATION_Y == 1) |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 181 | float4 temp = vload4(0, (__global float *)left_pixel); |
| 182 | |
| 183 | float2 left = CONVERT(temp.s01, float2); |
| 184 | float2 middle = CONVERT(temp.s12, float2); |
| 185 | float2 right = CONVERT(temp.s23, float2); |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 186 | return left * (float2)left_coeff + middle * (float2)middle_coeff + right * (float2)right_coeff; |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 187 | #else /* DILATION_X==1 && DILATION_Y==1 */ |
| 188 | return vload2(0, (__global float *)left_pixel) * (float2)left_coeff |
| 189 | + vload2(0, (__global float *)(left_pixel) + DILATION_X) * (float2)middle_coeff |
| 190 | + vload2(0, (__global float *)(left_pixel) + 2 * DILATION_X) * (float2)right_coeff; |
| 191 | #endif /* DILATION_X==1 && DILATION_Y==1 */ |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 192 | } |
| 193 | |
| 194 | /** Compute a 1D horizontal convolution of size 3 and stride 2 for floating point type. |
| 195 | * |
| 196 | * @param[in] left_pixel Pointer to the left pixel. |
| 197 | * @param[in] left_coeff Weight of the left pixel |
| 198 | * @param[in] middle_coeff Weight of the middle pixel |
| 199 | * @param[in] right_coeff Weight of the right pixel |
| 200 | * |
| 201 | * @return a float2 containing 2 convoluted values. |
| 202 | */ |
| 203 | inline float2 convolution1x3_stride_2(__global const uchar *left_pixel, |
| 204 | const float left_coeff, |
| 205 | const float middle_coeff, |
| 206 | const float right_coeff) |
| 207 | { |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 208 | #if(DILATION_X == 1 && DILATION_Y == 1) |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 209 | float4 temp0 = vload4(0, (__global float *)left_pixel); |
| 210 | float temp1 = *((__global float *)(left_pixel + 4 * sizeof(float))); |
| 211 | |
| 212 | float2 left = CONVERT(temp0.s02, float2); |
| 213 | float2 middle = CONVERT(temp0.s13, float2); |
| 214 | float2 right = CONVERT((float2)(temp0.s2, temp1), float2); |
| 215 | |
| 216 | return left * (float2)left_coeff + middle * (float2)middle_coeff + right * (float2)right_coeff; |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 217 | #else /* DILATION_X==1 && DILATION_Y==1 */ |
| 218 | __global float *left_pixel_float = (__global float *)left_pixel; |
| 219 | |
| 220 | return vload4(0, left_pixel_float).s02 * (float2)left_coeff |
| 221 | + vload4(0, left_pixel_float + DILATION_X).s02 * (float2)middle_coeff |
| 222 | + vload4(0, left_pixel_float + DILATION_X * 2).s02 * (float2)right_coeff; |
| 223 | |
| 224 | #endif /* DILATION_X==1 && DILATION_Y==1 */ |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 225 | } |
| 226 | |
| 227 | /** Compute a 1D horizontal convolution of size 3 and stride 3 for floating point type. |
| 228 | * |
| 229 | * @param[in] left_pixel Pointer to the left pixel. |
| 230 | * @param[in] left_coeff Weight of the left pixel |
| 231 | * @param[in] middle_coeff Weight of the middle pixel |
| 232 | * @param[in] right_coeff Weight of the right pixel |
| 233 | * |
| 234 | * @return a float2 containing 2 convoluted values. |
| 235 | */ |
| 236 | inline float2 convolution1x3_stride_3(__global const uchar *left_pixel, |
| 237 | const float left_coeff, |
| 238 | const float middle_coeff, |
| 239 | const float right_coeff) |
| 240 | { |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 241 | #if(DILATION_X == 1 && DILATION_Y == 1) |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 242 | float4 temp0 = vload4(0, (__global float *)left_pixel); |
| 243 | float2 temp1 = vload2(0, (__global float *)(left_pixel + 4 * sizeof(float))); |
| 244 | |
| 245 | float2 left = CONVERT(temp0.s03, float2); |
| 246 | float2 middle = CONVERT((float2)(temp0.s1, temp1.s0), float2); |
| 247 | float2 right = CONVERT((float2)(temp0.s2, temp1.s1), float2); |
| 248 | |
| 249 | return left * (float2)left_coeff + middle * (float2)middle_coeff + right * (float2)right_coeff; |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 250 | #else /* DILATION_X==1 && DILATION_Y==1 */ |
| 251 | __global float *left_pixel_float = (__global float *)left_pixel; |
| 252 | |
| 253 | return (float2)(*left_pixel_float, *(left_pixel_float + 3)) * (float2)left_coeff |
| 254 | + (float2)(*(left_pixel_float + DILATION_X), *(left_pixel_float + DILATION_X + 3)) * (float2)middle_coeff |
| 255 | + (float2)(*(left_pixel_float + DILATION_X * 2), *(left_pixel_float + DILATION_X * 2 + 3)) * (float2)right_coeff; |
| 256 | #endif /* DILATION_X==1 && DILATION_Y==1 */ |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 257 | } |
| 258 | |
| 259 | /** Apply a 3x3 convolution matrix to a single channel F32 input image and return the result. |
| 260 | * |
| 261 | * Convolution matrix layout: |
| 262 | * |
| 263 | * [ mat0, mat1, mat2 ]\n |
| 264 | * [ mat3, mat4, mat5 ]\n |
| 265 | * [ mat6, mat7, mat8 ]\n |
| 266 | * |
| 267 | * @param[in] src A pointer to source Image structure |
| 268 | * @param[in] mat0 Coefficient from the convolution matrix |
| 269 | * @param[in] mat1 Coefficient from the convolution matrix |
| 270 | * @param[in] mat2 Coefficient from the convolution matrix |
| 271 | * @param[in] mat3 Coefficient from the convolution matrix |
| 272 | * @param[in] mat4 Coefficient from the convolution matrix |
| 273 | * @param[in] mat5 Coefficient from the convolution matrix |
| 274 | * @param[in] mat6 Coefficient from the convolution matrix |
| 275 | * @param[in] mat0 Coefficient from the convolution matrix |
| 276 | * @param[in] mat7 Coefficient from the convolution matrix |
| 277 | * @param[in] mat8 Coefficient from the convolution matrix |
| 278 | * |
| 279 | * @return a float2 containing 2 convoluted values. |
| 280 | */ |
| 281 | inline float2 convolution3x3( |
Gian Marco Iodice | 3ae323f | 2019-08-13 14:56:50 +0100 | [diff] [blame] | 282 | __global const uchar *src, |
| 283 | unsigned int src_stride_y, |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 284 | const float mat0, const float mat1, const float mat2, |
| 285 | const float mat3, const float mat4, const float mat5, |
| 286 | const float mat6, const float mat7, const float mat8) |
| 287 | { |
| 288 | float2 pixels; |
| 289 | |
Gian Marco Iodice | 3ae323f | 2019-08-13 14:56:50 +0100 | [diff] [blame] | 290 | pixels = convolution1x3((src + 0 * DILATION_Y * src_stride_y), mat0, mat1, mat2); |
| 291 | pixels += convolution1x3((src + 1 * DILATION_Y * src_stride_y), mat3, mat4, mat5); |
| 292 | pixels += convolution1x3((src + 2 * DILATION_Y * src_stride_y), mat6, mat7, mat8); |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 293 | |
| 294 | return pixels; |
| 295 | } |
| 296 | |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 297 | /** This OpenCL kernel computes the depthwise convolution 3x3 |
Anthony Barbier | f202e50 | 2017-11-23 18:02:04 +0000 | [diff] [blame] | 298 | * |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 299 | * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu |
| 300 | * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively |
| 301 | * |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 302 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 303 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
Anthony Barbier | f202e50 | 2017-11-23 18:02:04 +0000 | [diff] [blame] | 304 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 305 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
Anthony Barbier | f202e50 | 2017-11-23 18:02:04 +0000 | [diff] [blame] | 306 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 307 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
Anthony Barbier | f202e50 | 2017-11-23 18:02:04 +0000 | [diff] [blame] | 308 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 309 | * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 310 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: F32 |
Anthony Barbier | f202e50 | 2017-11-23 18:02:04 +0000 | [diff] [blame] | 311 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 312 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 313 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 314 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 315 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 316 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
| 317 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 318 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: F32 |
Anthony Barbier | f202e50 | 2017-11-23 18:02:04 +0000 | [diff] [blame] | 319 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 320 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 321 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 322 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 323 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 324 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 325 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the biases vector |
| 326 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: F16/F32 |
| 327 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 328 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 329 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
| 330 | */ |
Georgios Pinitas | 81a26ad | 2017-10-23 20:29:30 +0100 | [diff] [blame] | 331 | __kernel void depthwise_convolution_3x3( |
| 332 | TENSOR3D_DECLARATION(src), |
| 333 | TENSOR3D_DECLARATION(dst), |
| 334 | TENSOR3D_DECLARATION(weights) |
| 335 | #if defined(HAS_BIAS) |
| 336 | , |
| 337 | VECTOR_DECLARATION(biases) |
| 338 | #endif //defined(HAS_BIAS) |
| 339 | ) |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 340 | { |
| 341 | Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src); |
| 342 | Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 343 | Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
Gian Marco Iodice | 3ae323f | 2019-08-13 14:56:50 +0100 | [diff] [blame] | 344 | |
| 345 | float2 pixels = 0.0f; |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 346 | |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 347 | // Extract channel and linearized batch indices |
| 348 | const int channel = get_global_id(2) % DST_CHANNELS; |
| 349 | const int batch = get_global_id(2) / DST_CHANNELS; |
| 350 | // Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER) |
Gian Marco Iodice | 3ae323f | 2019-08-13 14:56:50 +0100 | [diff] [blame] | 351 | |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 352 | __global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z; |
Giorgio Arena | 7657224 | 2018-04-04 17:44:26 +0100 | [diff] [blame] | 353 | |
Gian Marco Iodice | 3ae323f | 2019-08-13 14:56:50 +0100 | [diff] [blame] | 354 | __global uchar *src_addr = src.ptr - batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z - (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z; |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 355 | |
Gian Marco Iodice | 3ae323f | 2019-08-13 14:56:50 +0100 | [diff] [blame] | 356 | // Load the weights |
| 357 | float3 weights_values0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y)); |
| 358 | float3 weights_values1 = vload3(0, (__global float *)(weights_addr + 1 * weights_stride_y)); |
| 359 | float3 weights_values2 = vload3(0, (__global float *)(weights_addr + 2 * weights_stride_y)); |
| 360 | |
| 361 | pixels = convolution3x3(src_addr, src_stride_y, |
| 362 | weights_values0.s0, weights_values0.s1, weights_values0.s2, |
| 363 | weights_values1.s0, weights_values1.s1, weights_values1.s2, |
| 364 | weights_values2.s0, weights_values2.s1, weights_values2.s2); |
Georgios Pinitas | 81a26ad | 2017-10-23 20:29:30 +0100 | [diff] [blame] | 365 | #if defined(HAS_BIAS) |
Gian Marco Iodice | 3ae323f | 2019-08-13 14:56:50 +0100 | [diff] [blame] | 366 | Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| 367 | |
| 368 | float bias = *((__global float *)(vector_offset(&biases, channel))); |
| 369 | |
| 370 | pixels += (float2)bias; |
Georgios Pinitas | 81a26ad | 2017-10-23 20:29:30 +0100 | [diff] [blame] | 371 | #endif //defined(HAS_BIAS) |
Giorgio Arena | 93a690e | 2017-08-01 16:09:33 +0100 | [diff] [blame] | 372 | |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 373 | vstore2(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels, A_VAL, B_VAL), 0, (__global float *)dst.ptr); |
Giorgio Arena | 9fe4144 | 2017-08-23 16:36:24 +0100 | [diff] [blame] | 374 | } |
Giorgio Arena | 9fe4144 | 2017-08-23 16:36:24 +0100 | [diff] [blame] | 375 | #endif //defined(CONV_STRIDE_X) |
| 376 | |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 377 | #if(DILATION_X > 1 || DILATION_Y > 1) |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 378 | |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 379 | /** Perform 3x3 convolution for stride_x=1 and stride_y=1 when DILATION_X>1 or DILATION_Y>1 for F32 |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 380 | * |
| 381 | * @param[in] src_addr Pointer to the starting position of where to perform the convolution |
| 382 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 383 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 384 | * @param[in] y_offset Offset from the source tensor from which to start convolution |
| 385 | * @param[in] weights_addr Pointer from where to get weights |
| 386 | * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension |
| 387 | */ |
| 388 | inline float2 convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes, |
| 389 | const int y_offset, __global uchar *weights_addr, const int weights_stride_y) |
| 390 | { |
| 391 | // Load the weights |
| 392 | float3 weights_row0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y)); |
| 393 | float3 weights_row1 = vload3(0, (__global float *)(weights_addr + 1 * weights_stride_y)); |
| 394 | float3 weights_row2 = vload3(0, (__global float *)(weights_addr + 2 * weights_stride_y)); |
| 395 | |
| 396 | float2 pixels0 = 0.0f; |
| 397 | |
| 398 | float2 src00_left = vload2(0, (__global float *)ptr_offset(src_addr, 0, y_offset, stride_x_bytes, stride_y_bytes)); // Row0 |
| 399 | float2 src00_mid = vload2(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset, stride_x_bytes, stride_y_bytes)); |
| 400 | float2 src00_right = vload2(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset, stride_x_bytes, stride_y_bytes)); |
| 401 | |
| 402 | float2 src10_left = vload2(0, (__global float *)ptr_offset(src_addr, 0, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); // Row1 |
| 403 | float2 src10_mid = vload2(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); |
| 404 | float2 src10_right = vload2(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); |
| 405 | |
| 406 | float2 src20_left = vload2(0, (__global float *)ptr_offset(src_addr, 0, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); // Row2 |
| 407 | float2 src20_mid = vload2(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); |
| 408 | float2 src20_right = vload2(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); |
| 409 | |
| 410 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src00_left, src00_mid, src00_right, weights_row0); |
| 411 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src10_left, src10_mid, src10_right, weights_row1); |
| 412 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src20_left, src20_mid, src20_right, weights_row2); |
| 413 | |
| 414 | return pixels0; |
| 415 | } |
| 416 | |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 417 | /** Perform 3x3 convolution for stride_x=2 and stride_y=2 when DILATION_X>1 or DILATION_Y>1 for F32 |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 418 | * |
| 419 | * @param[in] src_addr Pointer to the starting position of where to perform the convolution |
| 420 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 421 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 422 | * @param[in] y_offset Offset from the source tensor from which to start convolution |
| 423 | * @param[in] weights_addr Pointer from where to get weights |
| 424 | * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension |
| 425 | */ |
| 426 | inline float2 convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes, |
| 427 | const int y_offset, __global uchar *weights_addr, const int weights_stride_y) |
| 428 | { |
| 429 | // Load the weights |
| 430 | float3 weights_row0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y)); |
| 431 | float3 weights_row1 = vload3(0, (__global float *)(weights_addr + 1 * weights_stride_y)); |
| 432 | float3 weights_row2 = vload3(0, (__global float *)(weights_addr + 2 * weights_stride_y)); |
| 433 | |
| 434 | float2 pixels0 = 0.0f; |
| 435 | |
| 436 | float3 src00_left = vload3(0, (__global float *)ptr_offset(src_addr, 0, y_offset, stride_x_bytes, stride_y_bytes)); // Row0 |
| 437 | float3 src00_mid = vload3(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset, stride_x_bytes, stride_y_bytes)); |
| 438 | float3 src00_right = vload3(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset, stride_x_bytes, stride_y_bytes)); |
| 439 | |
| 440 | float3 src10_left = vload3(0, (__global float *)ptr_offset(src_addr, 0, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); // Row1 |
| 441 | float3 src10_mid = vload3(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); |
| 442 | float3 src10_right = vload3(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); |
| 443 | |
| 444 | float3 src20_left = vload3(0, (__global float *)ptr_offset(src_addr, 0, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); // Row2 |
| 445 | float3 src20_mid = vload3(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); |
| 446 | float3 src20_right = vload3(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); |
| 447 | |
| 448 | CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src00_left, src00_mid, src00_right, weights_row0); |
| 449 | CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src10_left, src10_mid, src10_right, weights_row1); |
| 450 | CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src20_left, src20_mid, src20_right, weights_row2); |
| 451 | |
| 452 | return pixels0; |
| 453 | } |
| 454 | |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 455 | #endif /* (DILATION_X > 1 || DILATION_Y > 1) */ |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 456 | |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 457 | /** This OpenCL kernel is optimized for Bifrost architectures and computes the depthwise convolution 3x3 when both |
| 458 | * stride_x and stride_y are equal to 1 |
| 459 | * |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 460 | * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 461 | * @note If activation function is enabled, the data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float. |
| 462 | * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively |
| 463 | * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 464 | * |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 465 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 466 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 467 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 468 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 469 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 470 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 471 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 472 | * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) |
| 473 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: F32 |
| 474 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 475 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 476 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 477 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 478 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 479 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
| 480 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 481 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: F32 |
| 482 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 483 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 484 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 485 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 486 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 487 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 488 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the biases vector |
| 489 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: F32 |
| 490 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 491 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 492 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
| 493 | */ |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 494 | __kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32( |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 495 | TENSOR3D_DECLARATION(src), |
| 496 | TENSOR3D_DECLARATION(dst), |
| 497 | TENSOR3D_DECLARATION(weights) |
| 498 | #if defined(HAS_BIAS) |
| 499 | , |
| 500 | VECTOR_DECLARATION(biases) |
| 501 | #endif //defined(HAS_BIAS) |
| 502 | ) |
| 503 | { |
| 504 | Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src); |
| 505 | Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 506 | Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 507 | |
| 508 | float2 pixels0 = 0.0f; |
| 509 | float2 pixels1 = 0.0f; |
| 510 | float2 pixels2 = 0.0f; |
| 511 | float2 pixels3 = 0.0f; |
| 512 | |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 513 | // Extract channel and linearized batch indices |
| 514 | const int channel = get_global_id(2) % DST_CHANNELS; |
| 515 | const int batch = get_global_id(2) / DST_CHANNELS; |
| 516 | // Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER) |
| 517 | __global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z; |
| 518 | __global uchar *src_addr = src.ptr - batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z - (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z; |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 519 | |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 520 | #if(DILATION_X == 1 && DILATION_Y == 1) |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 521 | // Load the weights |
| 522 | float3 weights_row0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y)); |
| 523 | float3 weights_row1 = vload3(0, (__global float *)(weights_addr + 1 * weights_stride_y)); |
| 524 | float3 weights_row2 = vload3(0, (__global float *)(weights_addr + 2 * weights_stride_y)); |
| 525 | |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 526 | // Note: Since each work-item computes 4x2 elements, we need to load 6 rows from the input tensor |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 527 | float4 src00 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y)); // Row0 |
| 528 | float4 src10 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y)); // Row1 |
| 529 | float4 src20 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y)); // Row2 |
| 530 | float4 src30 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y)); // Row3 |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 531 | float4 src40 = vload4(0, (__global float *)(src_addr + 4 * src_stride_y)); // Row4 |
| 532 | float4 src50 = vload4(0, (__global float *)(src_addr + 5 * src_stride_y)); // Row5 |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 533 | |
| 534 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src00, weights_row0); |
| 535 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src10, weights_row1); |
| 536 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src20, weights_row2); |
| 537 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels1, src10, weights_row0); |
| 538 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels1, src20, weights_row1); |
| 539 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels1, src30, weights_row2); |
| 540 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels2, src20, weights_row0); |
| 541 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels2, src30, weights_row1); |
| 542 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels2, src40, weights_row2); |
| 543 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src30, weights_row0); |
| 544 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src40, weights_row1); |
| 545 | CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src50, weights_row2); |
| 546 | |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 547 | #else /* DILATION_X==1 && DILATION_Y==1 */ |
| 548 | |
| 549 | //3x3 Convolution of elements starting in 0th row |
| 550 | pixels0 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src.stride_x, src.stride_y, 0, weights_addr, weights_stride_y); |
| 551 | //3x3 Convolution of elements starting in 1st row |
| 552 | pixels1 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src.stride_x, src.stride_y, 1, weights_addr, weights_stride_y); |
| 553 | //3x3 Convolution of elements starting in 2nd row |
| 554 | pixels2 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y); |
| 555 | //3x3 Convolution of elements starting in 3rd row |
| 556 | pixels3 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src.stride_x, src.stride_y, 3, weights_addr, weights_stride_y); |
| 557 | |
| 558 | #endif /* DILATION_X==1 && DILATION_Y==1 */ |
| 559 | |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 560 | #ifdef HAS_BIAS |
| 561 | Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| 562 | |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 563 | float bias = *((__global float *)(vector_offset(&biases, channel))); |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 564 | |
| 565 | pixels0 += (float2)bias; |
| 566 | pixels1 += (float2)bias; |
| 567 | pixels2 += (float2)bias; |
| 568 | pixels3 += (float2)bias; |
| 569 | #endif /* defined(HAS_BIAS) */ |
| 570 | |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 571 | vstore2(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels0, A_VAL, B_VAL), 0, (__global float *)(dst.ptr + 0 * dst_stride_y)); |
| 572 | vstore2(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels1, A_VAL, B_VAL), 0, (__global float *)(dst.ptr + 1 * dst_stride_y)); |
| 573 | vstore2(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels2, A_VAL, B_VAL), 0, (__global float *)(dst.ptr + 2 * dst_stride_y)); |
| 574 | vstore2(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels3, A_VAL, B_VAL), 0, (__global float *)(dst.ptr + 3 * dst_stride_y)); |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 575 | } |
| 576 | |
| 577 | /** This OpenCL kernel is optimized for Bifrost architectures and computes the depthwise convolution 3x3 when both |
| 578 | * stride_x and stride_y are equal to 2 |
| 579 | * |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 580 | * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 581 | * @note If activation function is enabled, the data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float. |
| 582 | * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively |
| 583 | * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 584 | * |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 585 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 586 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 587 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 588 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 589 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 590 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 591 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 592 | * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) |
| 593 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: F32 |
| 594 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 595 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 596 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 597 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 598 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 599 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
| 600 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 601 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: F32 |
| 602 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 603 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 604 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 605 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 606 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 607 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 608 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the biases vector |
| 609 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: F32 |
| 610 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 611 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 612 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
| 613 | */ |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 614 | __kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32( |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 615 | TENSOR3D_DECLARATION(src), |
| 616 | TENSOR3D_DECLARATION(dst), |
| 617 | TENSOR3D_DECLARATION(weights) |
| 618 | #if defined(HAS_BIAS) |
| 619 | , |
| 620 | VECTOR_DECLARATION(biases) |
| 621 | #endif //defined(HAS_BIAS) |
| 622 | ) |
| 623 | { |
| 624 | Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src); |
| 625 | Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 626 | Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 627 | |
| 628 | float2 pixels0 = 0.0f; |
| 629 | float2 pixels1 = 0.0f; |
| 630 | |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 631 | // Extract channel and linearized batch indices |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 632 | const int channel = get_global_id(2) % DST_CHANNELS; |
| 633 | const int batch = get_global_id(2) / DST_CHANNELS; |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 634 | // Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER) |
| 635 | __global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z; |
| 636 | __global uchar *src_addr = src.ptr - batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z - (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z; |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 637 | |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 638 | #if(DILATION_X == 1 && DILATION_Y == 1) |
| 639 | |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 640 | // Load the weights |
| 641 | float3 weights_row0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y)); |
| 642 | float3 weights_row1 = vload3(0, (__global float *)(weights_addr + 1 * weights_stride_y)); |
| 643 | float3 weights_row2 = vload3(0, (__global float *)(weights_addr + 2 * weights_stride_y)); |
| 644 | |
| 645 | // Note: Since each work-item computes 4x2 elements, we need to load 5 rows from the input tensor |
| 646 | float4 src00 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y)); // Row0 |
| 647 | float2 src01 = vload2(2, (__global float *)(src_addr + 0 * src_stride_y)); // Row0 |
| 648 | float4 src10 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y)); // Row1 |
| 649 | float2 src11 = vload2(2, (__global float *)(src_addr + 1 * src_stride_y)); // Row1 |
| 650 | float4 src20 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y)); // Row2 |
| 651 | float2 src21 = vload2(2, (__global float *)(src_addr + 2 * src_stride_y)); // Row2 |
| 652 | float4 src30 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y)); // Row3 |
| 653 | float2 src31 = vload2(2, (__global float *)(src_addr + 3 * src_stride_y)); // Row3 |
| 654 | float4 src40 = vload4(0, (__global float *)(src_addr + 4 * src_stride_y)); // Row4 |
| 655 | float2 src41 = vload2(2, (__global float *)(src_addr + 4 * src_stride_y)); // Row4 |
| 656 | |
| 657 | CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src00, src01, weights_row0); |
| 658 | CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src10, src11, weights_row1); |
| 659 | CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src20, src21, weights_row2); |
| 660 | CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src20, src21, weights_row0); |
| 661 | CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src30, src31, weights_row1); |
| 662 | CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src40, src41, weights_row2); |
| 663 | |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 664 | #else /* DILATION_X==1 && DILATION_Y==1 */ |
| 665 | |
| 666 | //3x3 Convolution of elements starting in 0th row |
| 667 | pixels0 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(src_addr, src.stride_x, src.stride_y, 0, weights_addr, weights_stride_y); |
| 668 | //3x3 Convolution of elements starting in 2nd row |
| 669 | pixels1 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y); |
| 670 | #endif /* DILATION_X==1 && DILATION_Y==1 */ |
| 671 | |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 672 | #ifdef HAS_BIAS |
| 673 | Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| 674 | |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 675 | float bias = *((__global float *)(vector_offset(&biases, channel))); |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 676 | |
| 677 | pixels0 += (float2)bias; |
| 678 | pixels1 += (float2)bias; |
| 679 | #endif /* defined(HAS_BIAS) */ |
| 680 | |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 681 | vstore2(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels0, A_VAL, B_VAL), 0, (__global float *)(dst.ptr + 0 * dst_stride_y)); |
| 682 | vstore2(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels1, A_VAL, B_VAL), 0, (__global float *)(dst.ptr + 1 * dst_stride_y)); |
Gian Marco | c799ed8 | 2018-02-01 16:57:48 +0000 | [diff] [blame] | 683 | } |
| 684 | |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 685 | #endif // defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS) && defined(IS_F32) |
Giorgio Arena | 7657224 | 2018-04-04 17:44:26 +0100 | [diff] [blame] | 686 | |
giuros01 | 6d10996 | 2019-01-07 17:47:19 +0000 | [diff] [blame] | 687 | #if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DST_WIDTH) |
| 688 | /** Reshape the weights for quantized depthwise convolution |
| 689 | * |
| 690 | * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type, e.g. -DDATA_TYPE=uint8 |
| 691 | * @note Output width should be given as a preprocessor argument using -DDST_WIDTH=width, e.g. -DDST_WIDTH=128 |
| 692 | * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=vec_size, e.g., -DVEC_SIZE=4 |
| 693 | * @attention Input's height and width should be 3 |
| 694 | * |
Michele Di Giorgio | f6f7876 | 2020-07-06 11:27:21 +0100 | [diff] [blame] | 695 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: All |
giuros01 | 6d10996 | 2019-01-07 17:47:19 +0000 | [diff] [blame] | 696 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 697 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 698 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 699 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 700 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 701 | * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) |
| 702 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 703 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 704 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 705 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 706 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 707 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 708 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 709 | */ |
| 710 | __kernel void depthwise_convolution_reshape_weights( |
| 711 | TENSOR3D_DECLARATION(src), |
| 712 | IMAGE_DECLARATION(dst)) |
| 713 | { |
| 714 | Vector src = CONVERT_TO_VECTOR_STRUCT(src); |
| 715 | const int x = get_global_id(0); |
| 716 | |
| 717 | // Load 3x3xVEC_SIZE weights |
| 718 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 719 | w0 = VLOAD(VEC_SIZE)(0, src.ptr + 0 * src_stride_y + 0 * src_stride_z); |
| 720 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 721 | w1 = VLOAD(VEC_SIZE)(0, src.ptr + 1 * src_stride_y + 0 * src_stride_z); |
| 722 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 723 | w2 = VLOAD(VEC_SIZE)(0, src.ptr + 2 * src_stride_y + 0 * src_stride_z); |
| 724 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 725 | w3 = VLOAD(VEC_SIZE)(0, src.ptr + 0 * src_stride_y + 1 * src_stride_z); |
| 726 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 727 | w4 = VLOAD(VEC_SIZE)(0, src.ptr + 1 * src_stride_y + 1 * src_stride_z); |
| 728 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 729 | w5 = VLOAD(VEC_SIZE)(0, src.ptr + 2 * src_stride_y + 1 * src_stride_z); |
| 730 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 731 | w6 = VLOAD(VEC_SIZE)(0, src.ptr + 0 * src_stride_y + 2 * src_stride_z); |
| 732 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 733 | w7 = VLOAD(VEC_SIZE)(0, src.ptr + 1 * src_stride_y + 2 * src_stride_z); |
| 734 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 735 | w8 = VLOAD(VEC_SIZE)(0, src.ptr + 2 * src_stride_y + 2 * src_stride_z); |
| 736 | |
| 737 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * DST_WIDTH * sizeof(DATA_TYPE); |
| 738 | |
| 739 | #if defined(TRANSPOSE) |
| 740 | #if VEC_SIZE != 4 |
| 741 | #error "VEC_SIZE not supported" |
| 742 | #else // VEC_SIZE != 4 |
| 743 | VSTORE(VEC_SIZE) |
| 744 | ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w0.s0, w1.s0, w2.s0, w3.s0), 0, dst_addr + 0); |
| 745 | VSTORE(VEC_SIZE) |
| 746 | ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w4.s0, w5.s0, w6.s0, w7.s0), 0, dst_addr + 1 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 747 | VSTORE(VEC_SIZE) |
| 748 | ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w8.s0, w0.s1, w1.s1, w2.s1), 0, dst_addr + 2 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 749 | VSTORE(VEC_SIZE) |
| 750 | ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w3.s1, w4.s1, w5.s1, w6.s1), 0, dst_addr + 3 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 751 | VSTORE(VEC_SIZE) |
| 752 | ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w7.s1, w8.s1, w0.s2, w1.s2), 0, dst_addr + 4 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 753 | VSTORE(VEC_SIZE) |
| 754 | ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w2.s2, w3.s2, w4.s2, w5.s2), 0, dst_addr + 5 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 755 | VSTORE(VEC_SIZE) |
| 756 | ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w6.s2, w7.s2, w8.s2, w0.s3), 0, dst_addr + 6 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 757 | VSTORE(VEC_SIZE) |
| 758 | ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w1.s3, w2.s3, w3.s3, w4.s3), 0, dst_addr + 7 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 759 | VSTORE(VEC_SIZE) |
| 760 | ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w5.s3, w6.s3, w7.s3, w8.s3), 0, dst_addr + 8 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 761 | #endif // VEC_SIZE != 4 |
| 762 | #else // !defined(TRANSPOSE) |
| 763 | VSTORE(VEC_SIZE) |
| 764 | (w0, 0, dst_addr + 0); |
| 765 | VSTORE(VEC_SIZE) |
| 766 | (w1, 0, dst_addr + 1 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 767 | VSTORE(VEC_SIZE) |
| 768 | (w2, 0, dst_addr + 2 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 769 | VSTORE(VEC_SIZE) |
| 770 | (w3, 0, dst_addr + 3 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 771 | VSTORE(VEC_SIZE) |
| 772 | (w4, 0, dst_addr + 4 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 773 | VSTORE(VEC_SIZE) |
| 774 | (w5, 0, dst_addr + 5 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 775 | VSTORE(VEC_SIZE) |
| 776 | (w6, 0, dst_addr + 6 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 777 | VSTORE(VEC_SIZE) |
| 778 | (w7, 0, dst_addr + 7 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 779 | VSTORE(VEC_SIZE) |
| 780 | (w8, 0, dst_addr + 8 * sizeof(DATA_TYPE) * VEC_SIZE); |
| 781 | #endif // defined(TRANSPOSE) |
| 782 | } |
| 783 | #endif // defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DST_WIDTH) |
| 784 | |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 785 | #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS) && defined(IS_F16) |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 786 | #if defined(CONV_STRIDE_X) |
| 787 | #if CONV_STRIDE_X == 1 |
| 788 | #define convolution1x3_f16 convolution1x3_stride_1_f16 |
| 789 | #elif CONV_STRIDE_X == 2 |
| 790 | #define convolution1x3_f16 convolution1x3_stride_2_f16 |
| 791 | #elif CONV_STRIDE_X == 3 |
| 792 | #define convolution1x3_f16 convolution1x3_stride_3_f16 |
| 793 | #else /* CONV_STRIDE_X */ |
| 794 | #error "Stride not supported" |
| 795 | #endif /* CONV_STRIDE_X */ |
| 796 | |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 797 | #if(DILATION_X > 1 || DILATION_Y > 1) |
| 798 | |
| 799 | /** Perform 3x3 convolution for stride_x=1 and stride_y=1 when DILATION_X>1 or DILATION_Y>1 for f16 |
| 800 | * |
| 801 | * @param[in] src_addr Pointer to the starting position of where to perform the convolution |
| 802 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 803 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 804 | * @param[in] y_offset Offset from the source tensor from which to start convolution |
| 805 | * @param[in] weights_addr Pointer from where to get weights |
| 806 | * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension |
| 807 | */ |
| 808 | inline half4 convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes, |
| 809 | const int y_offset, __global uchar *weights_addr, const int weights_stride_y) |
| 810 | { |
| 811 | // Load the weights |
| 812 | half3 weights_row0 = vload3(0, (__global half *)(weights_addr + 0 * weights_stride_y)); |
| 813 | half3 weights_row1 = vload3(0, (__global half *)(weights_addr + 1 * weights_stride_y)); |
| 814 | half3 weights_row2 = vload3(0, (__global half *)(weights_addr + 2 * weights_stride_y)); |
| 815 | |
| 816 | half4 pixels0 = 0.0f; |
| 817 | |
| 818 | half4 src00_left = vload4(0, (__global half *)ptr_offset(src_addr, 0, y_offset, stride_x_bytes, stride_y_bytes)); // Row0 |
| 819 | half4 src00_mid = vload4(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset, stride_x_bytes, stride_y_bytes)); |
| 820 | half4 src00_right = vload4(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset, stride_x_bytes, stride_y_bytes)); |
| 821 | |
| 822 | half4 src10_left = vload4(0, (__global half *)ptr_offset(src_addr, 0, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); // Row1 |
| 823 | half4 src10_mid = vload4(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); |
| 824 | half4 src10_right = vload4(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); |
| 825 | |
| 826 | half4 src20_left = vload4(0, (__global half *)ptr_offset(src_addr, 0, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); // Row2 |
| 827 | half4 src20_mid = vload4(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); |
| 828 | half4 src20_right = vload4(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); |
| 829 | |
| 830 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src00_left, src00_mid, src00_right, weights_row0); |
| 831 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src10_left, src10_mid, src10_right, weights_row1); |
| 832 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src20_left, src20_mid, src20_right, weights_row2); |
| 833 | |
| 834 | return pixels0; |
| 835 | } |
| 836 | |
| 837 | /** Perform 3x3 convolution for stride_x=2 and stride_y=2 when DILATION_X>1 or DILATION_Y>1 for F16 |
| 838 | * |
| 839 | * @param[in] src_addr Pointer to the starting position of where to perform the convolution |
| 840 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 841 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 842 | * @param[in] y_offset Offset from the source tensor from which to start convolution |
| 843 | * @param[in] weights_addr Pointer from where to get weights |
| 844 | * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension |
| 845 | */ |
| 846 | inline half4 convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes, |
| 847 | const int y_offset, __global uchar *weights_addr, const int weights_stride_y) |
| 848 | { |
| 849 | // Load the weights |
| 850 | half3 weights_row0 = vload3(0, (__global half *)(weights_addr + 0 * weights_stride_y)); |
| 851 | half3 weights_row1 = vload3(0, (__global half *)(weights_addr + 1 * weights_stride_y)); |
| 852 | half3 weights_row2 = vload3(0, (__global half *)(weights_addr + 2 * weights_stride_y)); |
| 853 | |
| 854 | half4 pixels0 = 0.0f; |
| 855 | |
| 856 | half8 src00_left = vload8(0, (__global half *)ptr_offset(src_addr, 0, y_offset, stride_x_bytes, stride_y_bytes)); // Row0 |
| 857 | half8 src00_mid = vload8(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset, stride_x_bytes, stride_y_bytes)); |
| 858 | half8 src00_right = vload8(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset, stride_x_bytes, stride_y_bytes)); |
| 859 | |
| 860 | half8 src10_left = vload8(0, (__global half *)ptr_offset(src_addr, 0, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); // Row1 |
| 861 | half8 src10_mid = vload8(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); |
| 862 | half8 src10_right = vload8(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y, stride_x_bytes, stride_y_bytes)); |
| 863 | |
| 864 | half8 src20_left = vload8(0, (__global half *)ptr_offset(src_addr, 0, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); // Row2 |
| 865 | half8 src20_mid = vload8(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); |
| 866 | half8 src20_right = vload8(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes)); |
| 867 | |
| 868 | CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src00_left, src00_mid, src00_right, weights_row0); |
| 869 | CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src10_left, src10_mid, src10_right, weights_row1); |
| 870 | CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src20_left, src20_mid, src20_right, weights_row2); |
| 871 | |
| 872 | return pixels0; |
| 873 | } |
| 874 | |
| 875 | #endif // (DILATION_X > 1 && DILATION_Y > 1) |
| 876 | |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 877 | /** Compute a 1D horizontal convolution of size 3 and stride 1 for 16bit floating point type. |
| 878 | * |
| 879 | * @param[in] left_pixel Pointer to the left pixel. |
| 880 | * @param[in] left_coeff Weight of the left pixel |
| 881 | * @param[in] middle_coeff Weight of the middle pixel |
| 882 | * @param[in] right_coeff Weight of the right pixel |
| 883 | * |
| 884 | * @return a half4 containing 4 convoluted values. |
| 885 | */ |
| 886 | inline half4 convolution1x3_stride_1_f16(__global const uchar *left_pixel, |
| 887 | const half left_coeff, |
| 888 | const half middle_coeff, |
| 889 | const half right_coeff) |
| 890 | { |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 891 | #if(DILATION_X == 1 && DILATION_Y == 1) |
| 892 | |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 893 | half8 temp = vload8(0, (__global half *)left_pixel); |
| 894 | |
| 895 | half4 left = CONVERT(temp.s0123, half4); |
| 896 | half4 middle = CONVERT(temp.s1234, half4); |
| 897 | half4 right = CONVERT(temp.s2345, half4); |
| 898 | |
| 899 | return left * (half4)left_coeff + middle * (half4)middle_coeff + right * (half4)right_coeff; |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 900 | #else /* DILATION_X==1 && DILATION_Y==1 */ |
| 901 | return vload4(0, (__global half *)left_pixel) * (half4)left_coeff |
| 902 | + vload4(0, (__global half *)(left_pixel) + DILATION_X) * (half4)middle_coeff |
| 903 | + vload4(0, (__global half *)(left_pixel) + 2 * DILATION_X) * (half4)right_coeff; |
| 904 | |
| 905 | #endif /* DILATION_X==1 && DILATION_Y==1 */ |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 906 | } |
| 907 | |
| 908 | /** Compute a 1D horizontal convolution of size 3 and stride 2 for 16bit floating point type. |
| 909 | * |
| 910 | * @param[in] left_pixel Pointer to the left pixel. |
| 911 | * @param[in] left_coeff Weight of the left pixel |
| 912 | * @param[in] middle_coeff Weight of the middle pixel |
| 913 | * @param[in] right_coeff Weight of the right pixel |
| 914 | * |
| 915 | * @return a half4 containing 4 convoluted values. |
| 916 | */ |
| 917 | inline half4 convolution1x3_stride_2_f16(__global const uchar *left_pixel, |
| 918 | const half left_coeff, |
| 919 | const half middle_coeff, |
| 920 | const half right_coeff) |
| 921 | { |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 922 | #if(DILATION_X == 1 && DILATION_Y == 1) |
| 923 | |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 924 | half8 temp0 = vload8(0, (__global half *)left_pixel); |
| 925 | half temp1 = *((__global half *)(left_pixel + 8 * sizeof(half))); |
| 926 | |
| 927 | half4 left = CONVERT(temp0.s0246, half4); |
| 928 | half4 middle = CONVERT(temp0.s1357, half4); |
| 929 | half4 right = CONVERT((half4)(temp0.s246, temp1), half4); |
| 930 | |
| 931 | return left * (half4)left_coeff + middle * (half4)middle_coeff + right * (half4)right_coeff; |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 932 | #else /* DILATION_X==1 && DILATION_Y==1 */ |
| 933 | |
| 934 | __global half *left_pixel_float = (__global half *)left_pixel; |
| 935 | |
| 936 | return (half4)(*left_pixel_float, *(left_pixel_float + 2), *(left_pixel_float + 4), *(left_pixel_float + 6)) * (half4)left_coeff |
| 937 | + (half4)(*(left_pixel_float + DILATION_X), *(left_pixel_float + DILATION_X + 2), *(left_pixel_float + DILATION_X + 4), *(left_pixel_float + DILATION_X + 6)) * (half4)middle_coeff |
| 938 | + (half4)(*(left_pixel_float + DILATION_X * 2), *(left_pixel_float + DILATION_X * 2 + 2), *(left_pixel_float + DILATION_X * 2 + 4), *(left_pixel_float + DILATION_X * 2 + 6)) * (half4)right_coeff; |
| 939 | |
| 940 | #endif /* DILATION_X==1 && DILATION_Y==1 */ |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 941 | } |
| 942 | |
| 943 | /** Compute a 1D horizontal convolution of size 3 and stride 3 for 16bit floating point type. |
| 944 | * |
| 945 | * @param[in] left_pixel Pointer to the left pixel. |
| 946 | * @param[in] left_coeff Weight of the left pixel |
| 947 | * @param[in] middle_coeff Weight of the middle pixel |
| 948 | * @param[in] right_coeff Weight of the right pixel |
| 949 | * |
| 950 | * @return a half4 containing 4 convoluted values. |
| 951 | */ |
| 952 | inline half4 convolution1x3_stride_3_f16(__global const uchar *left_pixel, |
| 953 | const half left_coeff, |
| 954 | const half middle_coeff, |
| 955 | const half right_coeff) |
| 956 | { |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 957 | #if(DILATION_X == 1 && DILATION_Y == 1) |
| 958 | |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 959 | half16 temp0 = vload16(0, (__global half *)left_pixel); |
| 960 | |
| 961 | half4 left = CONVERT(temp0.s0369, half4); |
| 962 | half4 middle = CONVERT(temp0.s147A, half4); |
| 963 | half4 right = CONVERT(temp0.s258B, half4); |
| 964 | |
| 965 | return left * (half4)left_coeff + middle * (half4)middle_coeff + right * (half4)right_coeff; |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 966 | #else /* DILATION_X==1 && DILATION_Y==1 */ |
| 967 | |
| 968 | __global half *left_pixel_float = (__global half *)left_pixel; |
| 969 | |
| 970 | return (half4)(*left_pixel_float, *(left_pixel_float + 3), *(left_pixel_float + 6), *(left_pixel_float + 9)) * (half4)left_coeff |
| 971 | + (half4)(*(left_pixel_float + DILATION_X), *(left_pixel_float + DILATION_X + 3), *(left_pixel_float + DILATION_X + 6), *(left_pixel_float + DILATION_X + 9)) * (half4)middle_coeff |
| 972 | + (half4)(*(left_pixel_float + DILATION_X * 2), *(left_pixel_float + DILATION_X * 2 + 3), *(left_pixel_float + DILATION_X * 2 + 6), *(left_pixel_float + DILATION_X * 2 + 9)) * (half4)right_coeff; |
| 973 | |
| 974 | #endif /* DILATION_X==1 && DILATION_Y==1 */ |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 975 | } |
| 976 | |
| 977 | /** Apply a 3x3 convolution matrix to a single channel F16 input image and return the result. |
| 978 | * |
| 979 | * Convolution matrix layout: |
| 980 | * |
| 981 | * [ mat0, mat1, mat2 ]\n |
| 982 | * [ mat3, mat4, mat5 ]\n |
| 983 | * [ mat6, mat7, mat8 ]\n |
| 984 | * |
| 985 | * @param[in] src A pointer to source Image structure |
| 986 | * @param[in] mat0 Coefficient from the convolution matrix |
| 987 | * @param[in] mat1 Coefficient from the convolution matrix |
| 988 | * @param[in] mat2 Coefficient from the convolution matrix |
| 989 | * @param[in] mat3 Coefficient from the convolution matrix |
| 990 | * @param[in] mat4 Coefficient from the convolution matrix |
| 991 | * @param[in] mat5 Coefficient from the convolution matrix |
| 992 | * @param[in] mat6 Coefficient from the convolution matrix |
| 993 | * @param[in] mat0 Coefficient from the convolution matrix |
| 994 | * @param[in] mat7 Coefficient from the convolution matrix |
| 995 | * @param[in] mat8 Coefficient from the convolution matrix |
| 996 | * |
| 997 | * @return a half4 containing 4 convoluted values. |
| 998 | */ |
| 999 | inline half4 convolution3x3_f16( |
| 1000 | Image *src, |
| 1001 | const half mat0, const half mat1, const half mat2, |
| 1002 | const half mat3, const half mat4, const half mat5, |
| 1003 | const half mat6, const half mat7, const half mat8) |
| 1004 | { |
| 1005 | half4 pixels; |
| 1006 | |
| 1007 | pixels = convolution1x3_f16(offset(src, 0, 0), mat0, mat1, mat2); |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 1008 | pixels += convolution1x3_f16(offset(src, 0, DILATION_Y), mat3, mat4, mat5); |
| 1009 | pixels += convolution1x3_f16(offset(src, 0, DILATION_Y * 2), mat6, mat7, mat8); |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1010 | |
| 1011 | return pixels; |
| 1012 | } |
| 1013 | |
Giorgio Arena | 7657224 | 2018-04-04 17:44:26 +0100 | [diff] [blame] | 1014 | #if defined(DEPTH_MULTIPLIER) |
| 1015 | |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1016 | /** This OpenCL kernel computes the depthwise convolution 3x3 |
| 1017 | * |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 1018 | * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1019 | * @note If activation function is enabled, the data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=half. Supported data types: half. |
| 1020 | * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively |
| 1021 | * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1022 | * |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1023 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16 |
| 1024 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1025 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1026 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1027 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1028 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1029 | * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1030 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1031 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1032 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1033 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1034 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1035 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1036 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 1037 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
| 1038 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1039 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1040 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 1041 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 1042 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 1043 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1044 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 1045 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 1046 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the biases vector |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1047 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: F16 |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1048 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 1049 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 1050 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
| 1051 | */ |
| 1052 | __kernel void depthwise_convolution_3x3_f16( |
| 1053 | TENSOR3D_DECLARATION(src), |
| 1054 | TENSOR3D_DECLARATION(dst), |
| 1055 | TENSOR3D_DECLARATION(weights) |
| 1056 | #if defined(HAS_BIAS) |
| 1057 | , |
| 1058 | VECTOR_DECLARATION(biases) |
| 1059 | #endif //defined(HAS_BIAS) |
| 1060 | ) |
| 1061 | { |
| 1062 | Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src); |
| 1063 | Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 1064 | Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1065 | #if defined(HAS_BIAS) |
| 1066 | Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| 1067 | #endif //defined(HAS_BIAS) |
| 1068 | |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 1069 | // Extract channel and linearized batch indices |
| 1070 | const int channel = get_global_id(2) % DST_CHANNELS; |
| 1071 | const int batch = get_global_id(2) / DST_CHANNELS; |
| 1072 | // Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER) |
| 1073 | src.ptr -= batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z + (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z; |
| 1074 | __global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z; |
Giorgio Arena | 7657224 | 2018-04-04 17:44:26 +0100 | [diff] [blame] | 1075 | |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1076 | uchar3 offset = (uchar3)(0, 1, 2) * (uchar3)weights_stride_y; |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 1077 | half3 weights_values0 = vload3(0, (__global half *)(weights_addr + offset.s0)); |
| 1078 | half3 weights_values1 = vload3(0, (__global half *)(weights_addr + offset.s1)); |
| 1079 | half3 weights_values2 = vload3(0, (__global half *)(weights_addr + offset.s2)); |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1080 | |
| 1081 | half4 pixels = convolution3x3_f16(&src, weights_values0.s0, weights_values0.s1, weights_values0.s2, |
| 1082 | weights_values1.s0, weights_values1.s1, weights_values1.s2, |
| 1083 | weights_values2.s0, weights_values2.s1, weights_values2.s2); |
| 1084 | #if defined(HAS_BIAS) |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 1085 | pixels += (half4)(*((__global half *)(biases.ptr + channel * biases_stride_x))); |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1086 | #endif //defined(HAS_BIAS) |
| 1087 | |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 1088 | vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels, A_VAL, B_VAL), 0, (__global half *)dst.ptr); |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1089 | } |
Giorgio Arena | 7657224 | 2018-04-04 17:44:26 +0100 | [diff] [blame] | 1090 | #endif // defined(DEPTH_MULTIPLIER) |
Michele Di Giorgio | 933fe86 | 2018-02-19 15:42:12 +0000 | [diff] [blame] | 1091 | #endif // defined(CONV_STRIDE_X) |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1092 | |
| 1093 | /** This OpenCL kernel is optimized for Bifrost architectures and computes the 16bit floating point depthwise convolution 3x3 |
| 1094 | * when both stride_x and stride_y are equal to 1 |
| 1095 | * |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 1096 | * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1097 | * @note If activation function is enabled, the data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=half. Supported data types: half. |
| 1098 | * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively |
| 1099 | * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1100 | * |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1101 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16 |
| 1102 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1103 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1104 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1105 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1106 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1107 | * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1108 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1109 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 1110 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1111 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1112 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1113 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1114 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 1115 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
| 1116 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 1117 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr |
| 1118 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 1119 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 1120 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 1121 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1122 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 1123 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 1124 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the biases vector |
| 1125 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: same as @p src_ptr |
| 1126 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 1127 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 1128 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
| 1129 | */ |
| 1130 | __kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16( |
| 1131 | TENSOR3D_DECLARATION(src), |
| 1132 | TENSOR3D_DECLARATION(dst), |
| 1133 | TENSOR3D_DECLARATION(weights) |
| 1134 | #if defined(HAS_BIAS) |
| 1135 | , |
| 1136 | VECTOR_DECLARATION(biases) |
| 1137 | #endif //defined(HAS_BIAS) |
| 1138 | ) |
| 1139 | { |
| 1140 | Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src); |
| 1141 | Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 1142 | Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
| 1143 | |
| 1144 | // Extract channel and linearized batch indices |
| 1145 | const int channel = get_global_id(2) % DST_CHANNELS; |
| 1146 | const int batch = get_global_id(2) / DST_CHANNELS; |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1147 | |
| 1148 | #ifdef HAS_BIAS |
| 1149 | Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| 1150 | |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 1151 | half bias = *((__global half *)(vector_offset(&biases, channel))); |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1152 | #endif /* defined(HAS_BIAS) */ |
| 1153 | |
| 1154 | half4 pixels0 = 0.0f; |
| 1155 | half4 pixels1 = 0.0f; |
| 1156 | half4 pixels2 = 0.0f; |
| 1157 | half4 pixels3 = 0.0f; |
| 1158 | |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 1159 | // Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER) |
| 1160 | __global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z; |
| 1161 | __global uchar *src_addr = src.ptr - batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z - (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z; |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1162 | |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 1163 | #if(DILATION_X == 1 && DILATION_Y == 1) |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1164 | // Load the weights |
| 1165 | half3 weights_row0 = vload3(0, (__global half *)(weights_addr + 0 * weights_stride_y)); |
| 1166 | half3 weights_row1 = vload3(0, (__global half *)(weights_addr + 1 * weights_stride_y)); |
| 1167 | half3 weights_row2 = vload3(0, (__global half *)(weights_addr + 2 * weights_stride_y)); |
| 1168 | |
| 1169 | // Note: Since each work-item computes 4x4 elements, we need to load 6 rows from the input tensor |
| 1170 | half8 src00 = vload8(0, (__global half *)(src_addr + 0 * src_stride_y)); // Row0 |
| 1171 | half8 src10 = vload8(0, (__global half *)(src_addr + 1 * src_stride_y)); // Row1 |
| 1172 | half8 src20 = vload8(0, (__global half *)(src_addr + 2 * src_stride_y)); // Row2 |
| 1173 | half8 src30 = vload8(0, (__global half *)(src_addr + 3 * src_stride_y)); // Row3 |
| 1174 | half8 src40 = vload8(0, (__global half *)(src_addr + 4 * src_stride_y)); // Row4 |
| 1175 | half8 src50 = vload8(0, (__global half *)(src_addr + 5 * src_stride_y)); // Row5 |
| 1176 | |
| 1177 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src00, weights_row0); |
| 1178 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src10, weights_row1); |
| 1179 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src20, weights_row2); |
| 1180 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels1, src10, weights_row0); |
| 1181 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels1, src20, weights_row1); |
| 1182 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels1, src30, weights_row2); |
| 1183 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels2, src20, weights_row0); |
| 1184 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels2, src30, weights_row1); |
| 1185 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels2, src40, weights_row2); |
| 1186 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src30, weights_row0); |
| 1187 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src40, weights_row1); |
| 1188 | CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src50, weights_row2); |
| 1189 | |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 1190 | #else /* DILATION_X==1 && DILATION_Y==1 */ |
| 1191 | |
| 1192 | //3x3 Convolution of elements starting in 0th row |
| 1193 | pixels0 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src.stride_x, src.stride_y, 0, weights_addr, weights_stride_y); |
| 1194 | //3x3 Convolution of elements starting in 1st row |
| 1195 | pixels1 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src.stride_x, src.stride_y, 1, weights_addr, weights_stride_y); |
| 1196 | //3x3 Convolution of elements starting in 2nd row |
| 1197 | pixels2 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y); |
| 1198 | //3x3 Convolution of elements starting in 3rd row |
| 1199 | pixels3 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src.stride_x, src.stride_y, 3, weights_addr, weights_stride_y); |
| 1200 | |
| 1201 | #endif /* DILATION_X==1 && DILATION_Y==1 */ |
| 1202 | |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1203 | #ifdef HAS_BIAS |
| 1204 | pixels0 += (half4)bias; |
| 1205 | pixels1 += (half4)bias; |
| 1206 | pixels2 += (half4)bias; |
| 1207 | pixels3 += (half4)bias; |
| 1208 | #endif /* defined(HAS_BIAS) */ |
| 1209 | |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 1210 | vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels0, A_VAL, B_VAL), 0, (__global half *)(dst.ptr + 0 * dst_stride_y)); |
| 1211 | vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels1, A_VAL, B_VAL), 0, (__global half *)(dst.ptr + 1 * dst_stride_y)); |
| 1212 | vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels2, A_VAL, B_VAL), 0, (__global half *)(dst.ptr + 2 * dst_stride_y)); |
| 1213 | vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels3, A_VAL, B_VAL), 0, (__global half *)(dst.ptr + 3 * dst_stride_y)); |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1214 | } |
| 1215 | |
| 1216 | /** This OpenCL kernel is optimized for Bifrost architectures and computes 16bit floating point the depthwise convolution 3x3 |
| 1217 | * when both stride_x and stride_y are equal to 2 |
| 1218 | * |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 1219 | * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1220 | * @note If activation function is enabled, the data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=half. Supported data types: half. |
| 1221 | * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively |
| 1222 | * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1223 | * |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1224 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16 |
| 1225 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1226 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1227 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1228 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1229 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1230 | * @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes) |
| 1231 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1232 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 1233 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1234 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1235 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1236 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1237 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 1238 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
| 1239 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 1240 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr |
| 1241 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 1242 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 1243 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 1244 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1245 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 1246 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 1247 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the biases vector |
| 1248 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: same as @p src_ptr |
| 1249 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 1250 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 1251 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
| 1252 | */ |
| 1253 | __kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16( |
| 1254 | TENSOR3D_DECLARATION(src), |
| 1255 | TENSOR3D_DECLARATION(dst), |
| 1256 | TENSOR3D_DECLARATION(weights) |
| 1257 | #if defined(HAS_BIAS) |
| 1258 | , |
| 1259 | VECTOR_DECLARATION(biases) |
| 1260 | #endif //defined(HAS_BIAS) |
| 1261 | ) |
| 1262 | { |
| 1263 | Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src); |
| 1264 | Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 1265 | Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
| 1266 | |
| 1267 | // Extract channel and linearized batch indices |
| 1268 | const int channel = get_global_id(2) % DST_CHANNELS; |
| 1269 | const int batch = get_global_id(2) / DST_CHANNELS; |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1270 | |
| 1271 | #ifdef HAS_BIAS |
| 1272 | Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| 1273 | |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 1274 | half bias = *((__global half *)(vector_offset(&biases, channel))); |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1275 | #endif /* defined(HAS_BIAS) */ |
| 1276 | |
| 1277 | half4 pixels0 = 0.0f; |
| 1278 | half4 pixels1 = 0.0f; |
| 1279 | |
Georgios Pinitas | e55b40a | 2018-09-13 17:20:04 +0100 | [diff] [blame] | 1280 | // Load relevant input and weights data ( Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER) |
| 1281 | __global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z; |
| 1282 | __global uchar *src_addr = src.ptr - batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z - (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z; |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1283 | |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 1284 | #if(DILATION_X == 1 && DILATION_Y == 1) |
| 1285 | |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1286 | // Load the weights |
| 1287 | half3 weights_row0 = vload3(0, (__global half *)(weights_addr + 0 * weights_stride_y)); |
| 1288 | half3 weights_row1 = vload3(0, (__global half *)(weights_addr + 1 * weights_stride_y)); |
| 1289 | half3 weights_row2 = vload3(0, (__global half *)(weights_addr + 2 * weights_stride_y)); |
| 1290 | |
| 1291 | // Note: Since each work-item computes 2x4 elements, we need to load 5 rows from the input tensor |
| 1292 | half8 src00 = vload8(0, (__global half *)(src_addr + 0 * src_stride_y)); // Row0 |
| 1293 | half2 src01 = vload2(4, (__global half *)(src_addr + 0 * src_stride_y)); // Row0 |
| 1294 | half8 src10 = vload8(0, (__global half *)(src_addr + 1 * src_stride_y)); // Row1 |
| 1295 | half2 src11 = vload2(4, (__global half *)(src_addr + 1 * src_stride_y)); // Row1 |
| 1296 | half8 src20 = vload8(0, (__global half *)(src_addr + 2 * src_stride_y)); // Row2 |
| 1297 | half2 src21 = vload2(4, (__global half *)(src_addr + 2 * src_stride_y)); // Row2 |
| 1298 | half8 src30 = vload8(0, (__global half *)(src_addr + 3 * src_stride_y)); // Row3 |
| 1299 | half2 src31 = vload2(4, (__global half *)(src_addr + 3 * src_stride_y)); // Row3 |
| 1300 | half8 src40 = vload8(0, (__global half *)(src_addr + 4 * src_stride_y)); // Row4 |
| 1301 | half2 src41 = vload2(4, (__global half *)(src_addr + 4 * src_stride_y)); // Row4 |
| 1302 | |
| 1303 | CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src00, src01, weights_row0); |
| 1304 | CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src10, src11, weights_row1); |
| 1305 | CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src20, src21, weights_row2); |
| 1306 | CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src20, src21, weights_row0); |
| 1307 | CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src30, src31, weights_row1); |
| 1308 | CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src40, src41, weights_row2); |
| 1309 | |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 1310 | #else /* DILATION_X==1 && DILATION_Y==1 */ |
| 1311 | //3x3 Convolution of elements starting in 0th row |
| 1312 | pixels0 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src.stride_x, src.stride_y, 0, weights_addr, weights_stride_y); |
| 1313 | //3x3 Convolution of elements starting in 2nd row |
Michele Di Giorgio | a046e16 | 2019-10-08 09:36:26 +0100 | [diff] [blame] | 1314 | pixels1 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y); |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 1315 | #endif /* DILATION_X==1 && DILATION_Y==1 */ |
| 1316 | |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1317 | #ifdef HAS_BIAS |
| 1318 | pixels0 += (half4)bias; |
| 1319 | pixels1 += (half4)bias; |
| 1320 | #endif /* defined(HAS_BIAS) */ |
| 1321 | |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 1322 | vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels0, A_VAL, B_VAL), 0, (__global half *)(dst.ptr + 0 * dst_stride_y)); |
| 1323 | vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, pixels1, A_VAL, B_VAL), 0, (__global half *)(dst.ptr + 1 * dst_stride_y)); |
Michele Di Giorgio | 3ebef32 | 2018-02-21 10:02:58 +0000 | [diff] [blame] | 1324 | } |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1325 | #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS) && defined(IS_F16) |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1326 | |
Gian Marco Iodice | 9285adb | 2019-09-05 16:10:27 +0100 | [diff] [blame] | 1327 | #if defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(N0) && defined(DATA_TYPE) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP) |
| 1328 | /** This function computes the depthwise convolution for NHWC data layout. This kernel assumes that the weights tensor is NOT reshaped |
| 1329 | * |
| 1330 | * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float |
| 1331 | * @note The number of elements processed must be passed at compile time using -DN0 (e.g. -DN0=2) |
| 1332 | * @note The depth multiplier must be passed at compile time using -DDEPTH_MULTIPLIER (e.g. -DDEPTH_MULTIPLIER=1) |
| 1333 | * @note The first dimension of the input tensor must be passed at compile time using -DSRC_DIM1 (e.g. -DSRC_DIM1=112) |
| 1334 | * @note The second dimension of the input tensor must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM2=80) |
| 1335 | * @note The kernel width must be passed at compile time using -DKERNEL_WIDTH (e.g. -DKERNEL_WIDTH=5) |
| 1336 | * @note The kernel height must be passed at compile time using -DKERNEL_HEIGHT (e.g. -DKERNEL_HEIGHT=5) |
| 1337 | * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1) |
| 1338 | * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1) |
| 1339 | * @note The convolution stride along the width must be passed at compile time using -DCONV_STRIDE_X (e.g. -DCONV_STRIDE_Y=X) |
| 1340 | * @note The convolution stride along the height must be passed at compile time using -DCONV_STRIDE_Y (e.g. -DCONV_STRIDE_Y=1) |
| 1341 | * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu |
| 1342 | * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively |
| 1343 | * |
| 1344 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 |
| 1345 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 1346 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1347 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 1348 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1349 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1350 | * @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes) |
| 1351 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 1352 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 1353 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 1354 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as src_ptr |
| 1355 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1356 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1357 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1358 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1359 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 1360 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
| 1361 | * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) |
| 1362 | * @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes) |
| 1363 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 1364 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: F16/F32 |
| 1365 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 1366 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 1367 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 1368 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1369 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 1370 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 1371 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 1372 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: same as src_ptr |
| 1373 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 1374 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 1375 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
| 1376 | */ |
| 1377 | __kernel void dwc_MxN_native_fp_nhwc( |
| 1378 | TENSOR4D_DECLARATION(src), |
| 1379 | TENSOR4D_DECLARATION(dst), |
Michele Di Giorgio | 1dce310 | 2019-10-22 10:29:03 +0100 | [diff] [blame] | 1380 | TENSOR3D_DECLARATION(weights) |
Gian Marco Iodice | 9285adb | 2019-09-05 16:10:27 +0100 | [diff] [blame] | 1381 | #if defined(HAS_BIAS) |
Michele Di Giorgio | 1dce310 | 2019-10-22 10:29:03 +0100 | [diff] [blame] | 1382 | , |
Gian Marco Iodice | 9285adb | 2019-09-05 16:10:27 +0100 | [diff] [blame] | 1383 | VECTOR_DECLARATION(biases) |
| 1384 | #endif // defined(HAS_BIAS) |
| 1385 | ) |
| 1386 | { |
| 1387 | int x = get_global_id(0); // channels |
| 1388 | int y = get_global_id(1); // spatial coordinate x |
| 1389 | #if defined(DST_DEPTH) |
| 1390 | int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y |
| 1391 | int b = get_global_id(2) / (int)DST_DEPTH; // batch |
| 1392 | #else // defined(DST_DEPTH) |
Michele Di Giorgio | a046e16 | 2019-10-08 09:36:26 +0100 | [diff] [blame] | 1393 | int z = get_global_id(2); // spatial coordinate y |
Gian Marco Iodice | 9285adb | 2019-09-05 16:10:27 +0100 | [diff] [blame] | 1394 | #endif // defined(DST_DEPTH) |
| 1395 | |
Michele Di Giorgio | a046e16 | 2019-10-08 09:36:26 +0100 | [diff] [blame] | 1396 | __global uchar *s_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)N0; |
Gian Marco Iodice | 9285adb | 2019-09-05 16:10:27 +0100 | [diff] [blame] | 1397 | |
Michele Di Giorgio | a046e16 | 2019-10-08 09:36:26 +0100 | [diff] [blame] | 1398 | __global uchar *d_addr = dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0 + y * dst_stride_y + z * dst_stride_z; |
Gian Marco Iodice | 9285adb | 2019-09-05 16:10:27 +0100 | [diff] [blame] | 1399 | |
Michele Di Giorgio | a046e16 | 2019-10-08 09:36:26 +0100 | [diff] [blame] | 1400 | __global uchar *w_addr = weights_ptr + weights_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0; |
Gian Marco Iodice | 9285adb | 2019-09-05 16:10:27 +0100 | [diff] [blame] | 1401 | |
| 1402 | #if defined(HAS_BIAS) |
Michele Di Giorgio | a046e16 | 2019-10-08 09:36:26 +0100 | [diff] [blame] | 1403 | __global uchar *b_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0; |
Gian Marco Iodice | 9285adb | 2019-09-05 16:10:27 +0100 | [diff] [blame] | 1404 | #endif // defined(HAS_BIAS) |
| 1405 | |
| 1406 | #if defined(DST_DEPTH) |
| 1407 | s_addr += b * src_stride_w; |
| 1408 | d_addr += b * dst_stride_w; |
| 1409 | #endif // defined(DST_DEPTH) |
| 1410 | |
| 1411 | for(int d = 0; d < (int)DEPTH_MULTIPLIER; ++d) |
| 1412 | { |
| 1413 | // Each work-item computes N0x1x1 elements |
| 1414 | VEC_DATA_TYPE(DATA_TYPE, N0) |
| 1415 | res = 0; |
| 1416 | |
| 1417 | int x_coord = y * CONV_STRIDE_X - (int)CONV_PAD_LEFT; |
| 1418 | int y_coord = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP; |
| 1419 | |
| 1420 | for(int yk = 0; yk < KERNEL_HEIGHT; ++yk) |
| 1421 | { |
| 1422 | if(y_coord >= 0 && y_coord < SRC_DIM2) |
| 1423 | { |
| 1424 | int x_coord_tmp = x_coord; |
| 1425 | |
| 1426 | for(int xk = 0; xk < KERNEL_WIDTH; ++xk) |
| 1427 | { |
| 1428 | if(x_coord_tmp >= 0 && x_coord_tmp < SRC_DIM1) |
| 1429 | { |
| 1430 | int s_offset = x_coord_tmp * (int)src_stride_y + y_coord * (int)src_stride_z; |
| 1431 | int w_offset = xk * weights_stride_y + yk * weights_stride_z; |
| 1432 | |
| 1433 | // Load input and weights values |
| 1434 | VEC_DATA_TYPE(DATA_TYPE, N0) |
| 1435 | i = VLOAD(N0)(0, (__global DATA_TYPE *)(s_addr + s_offset)); |
| 1436 | VEC_DATA_TYPE(DATA_TYPE, N0) |
| 1437 | w = VLOAD(N0)(0, (__global DATA_TYPE *)(w_addr + w_offset)); |
| 1438 | |
| 1439 | #if GPU_ARCH == GPU_ARCH_MIDGARD |
| 1440 | res += i * w; |
| 1441 | #else // GPU_ARCH == GPU_ARCH_MIDGARD |
Michele Di Giorgio | a046e16 | 2019-10-08 09:36:26 +0100 | [diff] [blame] | 1442 | res = fma(i, w, res); |
Gian Marco Iodice | 9285adb | 2019-09-05 16:10:27 +0100 | [diff] [blame] | 1443 | #endif // GPU_ARCH == GPU_ARCH_MIDGARD |
| 1444 | } |
| 1445 | x_coord_tmp += DILATION_X; |
| 1446 | } |
| 1447 | } |
| 1448 | y_coord += DILATION_Y; |
| 1449 | } |
| 1450 | |
| 1451 | #if defined(HAS_BIAS) |
| 1452 | res += VLOAD(N0)(0, (__global DATA_TYPE *)(b_addr)); |
| 1453 | #endif // defined(HAS_BIAS) |
| 1454 | |
| 1455 | res = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, res, A_VAL, B_VAL); |
| 1456 | |
| 1457 | VSTORE(N0) |
| 1458 | (res, 0, (__global DATA_TYPE *)(d_addr)); |
| 1459 | |
| 1460 | w_addr += sizeof(DATA_TYPE); |
| 1461 | d_addr += sizeof(DATA_TYPE); |
| 1462 | #if defined(HAS_BIAS) |
| 1463 | b_addr += sizeof(DATA_TYPE); |
| 1464 | #endif // defined(HAS_BIAS) |
| 1465 | } |
| 1466 | } |
| 1467 | #endif // defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defiend(N0) && defined(DATA_TYPE) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP) |
| 1468 | |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1469 | #if defined(VEC_SIZE) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT) && defined(DATA_TYPE) |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1470 | |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1471 | #if DATA_TYPE != float || DATA_TYPE != half |
| 1472 | #error "Unsupported data type" |
| 1473 | #endif // DATA_TYPE != float || DATA_TYPE != half |
| 1474 | |
| 1475 | #define VEC_FLOAT VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1476 | |
| 1477 | #if defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) |
Gian Marco Iodice | 9285adb | 2019-09-05 16:10:27 +0100 | [diff] [blame] | 1478 | |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1479 | /** This function computes the depthwise convolution for NHWC data layout when the stride along the width or height is not 1. |
| 1480 | * |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1481 | * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1482 | * @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2) |
| 1483 | * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112) |
| 1484 | * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1) |
| 1485 | * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1) |
| 1486 | * @note The convolution stride along the width must be passed at compile time using -DCONV_STRIDE_X (e.g. -DCONV_STRIDE_Y=X) |
| 1487 | * @note The convolution stride along the height must be passed at compile time using -DCONV_STRIDE_Y (e.g. -DCONV_STRIDE_Y=1) |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 1488 | * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1489 | * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively |
| 1490 | * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1491 | * |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1492 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1493 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1494 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1495 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1496 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1497 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1498 | * @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes) |
| 1499 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 1500 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 1501 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1502 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as src_ptr |
| 1503 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1504 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1505 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1506 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1507 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 1508 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1509 | * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) |
| 1510 | * @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes) |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1511 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1512 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: F16/F32 |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1513 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 1514 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 1515 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 1516 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1517 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 1518 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 1519 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 1520 | * @param[in] max_offset Max offset for the input tensor |
| 1521 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: same as src_ptr |
| 1522 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 1523 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 1524 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
| 1525 | */ |
| 1526 | __kernel void depthwise_convolution_3x3_nhwc( |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1527 | TENSOR4D_DECLARATION(src), |
| 1528 | TENSOR4D_DECLARATION(dst), |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1529 | TENSOR3D_DECLARATION(weights), |
| 1530 | #if defined(HAS_BIAS) |
| 1531 | VECTOR_DECLARATION(biases), |
| 1532 | #endif /* defined(HAS_BIAS) */ |
| 1533 | int max_offset) |
| 1534 | { |
| 1535 | int x = get_global_id(0); // channels |
| 1536 | int y = get_global_id(1); // spatial coordinate x |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1537 | #if defined(DST_DEPTH) |
| 1538 | int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y |
| 1539 | int b = get_global_id(2) / (int)DST_DEPTH; // batch |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 1540 | #else // defined(DST_DEPTH) |
| 1541 | int z = get_global_id(2); // spatial coordinate y |
| 1542 | #endif // defined(DST_DEPTH) |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1543 | |
| 1544 | Vector weights = CONVERT_TO_VECTOR_STRUCT(weights); |
| 1545 | |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1546 | #if defined(DST_DEPTH) |
| 1547 | __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * VEC_SIZE + b * src_stride_w; |
| 1548 | #else /* defined(DST_DEPTH) */ |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1549 | __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * VEC_SIZE; |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1550 | #endif /* defined(DST_DEPTH) */ |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1551 | |
| 1552 | int z_coord = 0; |
| 1553 | int4 offset = 0; |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 1554 | int4 y_offset = ((int4)(y * CONV_STRIDE_X) + (int4)(0, DILATION_X * 1, DILATION_X * 2, DILATION_X * 3) - CONV_PAD_LEFT) * (int4)src_stride_y; |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1555 | |
| 1556 | // We compute 2x1x1 [C,W,H] elements |
| 1557 | VEC_FLOAT acc = 0; |
| 1558 | |
| 1559 | // Load weights |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1560 | VEC_FLOAT w0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 0 * weights_stride_y + 0 * weights_stride_z)); |
| 1561 | VEC_FLOAT w1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 1 * weights_stride_y + 0 * weights_stride_z)); |
| 1562 | VEC_FLOAT w2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 2 * weights_stride_y + 0 * weights_stride_z)); |
| 1563 | VEC_FLOAT w3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 0 * weights_stride_y + 1 * weights_stride_z)); |
| 1564 | VEC_FLOAT w4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 1 * weights_stride_y + 1 * weights_stride_z)); |
| 1565 | VEC_FLOAT w5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 2 * weights_stride_y + 1 * weights_stride_z)); |
| 1566 | VEC_FLOAT w6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 0 * weights_stride_y + 2 * weights_stride_z)); |
| 1567 | VEC_FLOAT w7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 1 * weights_stride_y + 2 * weights_stride_z)); |
| 1568 | VEC_FLOAT w8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 2 * weights_stride_y + 2 * weights_stride_z)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1569 | |
| 1570 | // Load input values |
| 1571 | // z == 0 |
| 1572 | // Clamp z_coord as for z = 0, it can be negative |
| 1573 | // z_coord is casted to unsigned int in order to use just a min() operation |
| 1574 | // A "-1" 32 bit signed variable converted to unsigned gives 4294967295 |
| 1575 | z_coord = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP; |
| 1576 | z_coord = min((uint)z_coord, (uint)SRC_DIM_2); |
| 1577 | offset = y_offset + (int4)(z_coord * src_stride_z); |
Georgios Pinitas | ed32f43 | 2018-07-10 17:03:11 +0100 | [diff] [blame] | 1578 | offset = min(offset, (int4)max_offset); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1579 | |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1580 | VEC_FLOAT values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); |
| 1581 | VEC_FLOAT values1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); |
| 1582 | VEC_FLOAT values2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1583 | |
| 1584 | // z == 1 |
| 1585 | // z_coord can be only negative for z = 0 so we do not need to clamp it |
| 1586 | // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 1587 | z_coord = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y; |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1588 | offset = y_offset + (int4)(z_coord * src_stride_z); |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1589 | VEC_FLOAT values3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); |
| 1590 | VEC_FLOAT values4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); |
| 1591 | VEC_FLOAT values5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1592 | |
| 1593 | // z == 2 |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 1594 | // Offset can be out-of-bound so we need to check if it is greater than max_offset |
| 1595 | z_coord = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y * 2; |
| 1596 | offset = y_offset + (int4)(z_coord * src_stride_z); |
Georgios Pinitas | ed32f43 | 2018-07-10 17:03:11 +0100 | [diff] [blame] | 1597 | offset = min(offset, (int4)max_offset); |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1598 | VEC_FLOAT values6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); |
| 1599 | VEC_FLOAT values7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); |
| 1600 | VEC_FLOAT values8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1601 | |
| 1602 | acc = fma(values0, w0, acc); |
| 1603 | acc = fma(values1, w1, acc); |
| 1604 | acc = fma(values2, w2, acc); |
| 1605 | |
| 1606 | acc = fma(values3, w3, acc); |
| 1607 | acc = fma(values4, w4, acc); |
| 1608 | acc = fma(values5, w5, acc); |
| 1609 | |
| 1610 | acc = fma(values6, w6, acc); |
| 1611 | acc = fma(values7, w7, acc); |
| 1612 | acc = fma(values8, w8, acc); |
| 1613 | |
| 1614 | #if defined(HAS_BIAS) |
| 1615 | Vector biases = CONVERT_TO_VECTOR_STRUCT(biases); |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1616 | VEC_FLOAT bias_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)biases.ptr); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1617 | acc += bias_values; |
| 1618 | #endif // defined(HAS_BIAS) |
| 1619 | |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1620 | #if defined(DST_DEPTH) |
| 1621 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z + b * dst_stride_w; |
| 1622 | #else /* defined(DST_DEPTH) */ |
| 1623 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z; |
| 1624 | #endif /* defined(DST_DEPTH) */ |
| 1625 | |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1626 | VSTORE(VEC_SIZE) |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 1627 | (ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, acc, A_VAL, B_VAL), 0, (__global DATA_TYPE *)(dst_addr)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1628 | } |
| 1629 | #endif // defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) |
| 1630 | |
| 1631 | #if defined(NUM_ROWS_PROCESSED) && defined(NUM_PLANES_PROCESSED) |
| 1632 | /** This function computes the depthwise convolution for NHWC data layout when the stride along the width and height is 1. |
| 1633 | * |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1634 | * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1635 | * @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2) |
| 1636 | * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112) |
| 1637 | * @note The number of rows processed per thread must be passed at compile time using -DNUM_ROWS_PROCESSED (i.e. -DNUM_ROWS_PROCESSED=2) |
| 1638 | * @note The number of planes processed per thread must be passed at compile time using -DNUM_PLANES_PROCESSED (i.e. -DNUM_PLANES_PROCESSED=2) |
| 1639 | * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1) |
| 1640 | * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1) |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 1641 | * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1642 | * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively |
| 1643 | * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1644 | * |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1645 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1646 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1647 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1648 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1649 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1650 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1651 | * @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes) |
| 1652 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 1653 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 1654 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1655 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as src_ptr |
| 1656 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1657 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1658 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1659 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1660 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 1661 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1662 | * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) |
| 1663 | * @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes) |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1664 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
Manuel Bottini | a788c2f | 2019-04-08 13:18:00 +0100 | [diff] [blame] | 1665 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: F16/F32 |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1666 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 1667 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 1668 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 1669 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1670 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 1671 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 1672 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 1673 | * @param[in] max_offset Max offset for the input tensor |
| 1674 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: same as src_ptr |
| 1675 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 1676 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 1677 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
| 1678 | */ |
| 1679 | __kernel void depthwise_convolution_3x3_nhwc_stride1( |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1680 | TENSOR4D_DECLARATION(src), |
| 1681 | TENSOR4D_DECLARATION(dst), |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1682 | TENSOR3D_DECLARATION(weights), |
| 1683 | #if defined(HAS_BIAS) |
| 1684 | VECTOR_DECLARATION(biases), |
| 1685 | #endif /* defined(HAS_BIAS) */ |
| 1686 | int max_offset) |
| 1687 | { |
| 1688 | int x = get_global_id(0); // channels |
| 1689 | int y = get_global_id(1); // spatial coordinate x |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1690 | #if defined(DST_DEPTH) |
| 1691 | int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y |
| 1692 | int b = get_global_id(2) / (int)DST_DEPTH; // batch |
Usama Arif | e73686a | 2019-04-08 17:30:48 +0100 | [diff] [blame] | 1693 | #else // defined(DST_DEPTH) |
| 1694 | int z = get_global_id(2); // spatial coordinate y |
| 1695 | #endif // defined(DST_DEPTH) |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1696 | |
| 1697 | Vector weights = CONVERT_TO_VECTOR_STRUCT(weights); |
| 1698 | |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1699 | #if defined(DST_DEPTH) |
| 1700 | __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * VEC_SIZE + b * src_stride_w; |
| 1701 | #else /* defined(DST_DEPTH) */ |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1702 | __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * VEC_SIZE; |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1703 | #endif /* defined(DST_DEPTH) */ |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1704 | |
| 1705 | int z_coord = 0; |
| 1706 | int4 offset = 0; |
Georgios Pinitas | ed32f43 | 2018-07-10 17:03:11 +0100 | [diff] [blame] | 1707 | int4 y_offset = ((int4)(y * NUM_ROWS_PROCESSED) + (int4)(0, 1, 2, 3) - (int)CONV_PAD_LEFT) * (int4)src_stride_y; |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1708 | |
| 1709 | // We compute 2x2x2 [C,W,H] elements |
| 1710 | VEC_FLOAT acc0 = 0; |
| 1711 | VEC_FLOAT acc1 = 0; |
| 1712 | VEC_FLOAT acc2 = 0; |
| 1713 | VEC_FLOAT acc3 = 0; |
| 1714 | |
| 1715 | // Load weights |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1716 | VEC_FLOAT w0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 0 * weights_stride_y + 0 * weights_stride_z)); |
| 1717 | VEC_FLOAT w1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 1 * weights_stride_y + 0 * weights_stride_z)); |
| 1718 | VEC_FLOAT w2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 2 * weights_stride_y + 0 * weights_stride_z)); |
| 1719 | VEC_FLOAT w3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 0 * weights_stride_y + 1 * weights_stride_z)); |
| 1720 | VEC_FLOAT w4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 1 * weights_stride_y + 1 * weights_stride_z)); |
| 1721 | VEC_FLOAT w5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 2 * weights_stride_y + 1 * weights_stride_z)); |
| 1722 | VEC_FLOAT w6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 0 * weights_stride_y + 2 * weights_stride_z)); |
| 1723 | VEC_FLOAT w7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 1 * weights_stride_y + 2 * weights_stride_z)); |
| 1724 | VEC_FLOAT w8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(weights.ptr + 2 * weights_stride_y + 2 * weights_stride_z)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1725 | |
| 1726 | // Load input values |
| 1727 | // z == 0 |
| 1728 | // Clamp z_coord as for z = 0, it can be negative |
| 1729 | // z_coord is casted to unsigned int in order to use just a min() operation |
| 1730 | // A "-1" 32 bit signed variable converted to unsigned gives 4294967295 |
Georgios Pinitas | ed32f43 | 2018-07-10 17:03:11 +0100 | [diff] [blame] | 1731 | z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP; |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1732 | z_coord = min((uint)z_coord, (uint)SRC_DIM_2); |
| 1733 | offset = y_offset + (int4)(z_coord * src_stride_z); |
Georgios Pinitas | ed32f43 | 2018-07-10 17:03:11 +0100 | [diff] [blame] | 1734 | offset = min(offset, (int4)max_offset); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1735 | |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1736 | VEC_FLOAT values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); |
| 1737 | VEC_FLOAT values1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); |
| 1738 | VEC_FLOAT values2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); |
| 1739 | VEC_FLOAT values3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1740 | |
| 1741 | // z == 1 |
| 1742 | // z_coord can be only negative for z = 0 so we do not need to clamp it |
| 1743 | // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset |
Georgios Pinitas | ed32f43 | 2018-07-10 17:03:11 +0100 | [diff] [blame] | 1744 | z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP + 1; |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1745 | offset = y_offset + (int4)(z_coord * src_stride_z); |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1746 | VEC_FLOAT values4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); |
| 1747 | VEC_FLOAT values5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); |
| 1748 | VEC_FLOAT values6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); |
| 1749 | VEC_FLOAT values7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1750 | |
| 1751 | // z == 2 |
| 1752 | // After z = 1 we can simply add src_stride_z to offset without updating z_coord |
| 1753 | // However offset can be out-of-bound so we need to check if it is greater than max_offset |
| 1754 | offset += (int4)src_stride_z; |
Georgios Pinitas | ed32f43 | 2018-07-10 17:03:11 +0100 | [diff] [blame] | 1755 | offset = min(offset, (int4)max_offset); |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1756 | VEC_FLOAT values8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); |
| 1757 | VEC_FLOAT values9 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); |
| 1758 | VEC_FLOAT values10 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); |
| 1759 | VEC_FLOAT values11 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1760 | |
| 1761 | // z == 3 |
| 1762 | // After z = 1 we can simply add src_stride_z to offset without updating z_coord |
| 1763 | // However offset can be out-of-bound so we need to check if it is greater than max_offset |
Georgios Pinitas | ed32f43 | 2018-07-10 17:03:11 +0100 | [diff] [blame] | 1764 | offset += (int4)src_stride_z; |
| 1765 | offset = min(offset, (int4)max_offset); |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1766 | VEC_FLOAT values12 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); |
| 1767 | VEC_FLOAT values13 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1)); |
| 1768 | VEC_FLOAT values14 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2)); |
| 1769 | VEC_FLOAT values15 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1770 | |
| 1771 | acc0 = fma(values0, w0, acc0); |
| 1772 | acc0 = fma(values1, w1, acc0); |
| 1773 | acc0 = fma(values2, w2, acc0); |
| 1774 | acc1 = fma(values1, w0, acc1); |
| 1775 | acc1 = fma(values2, w1, acc1); |
| 1776 | acc1 = fma(values3, w2, acc1); |
| 1777 | |
| 1778 | acc0 = fma(values4, w3, acc0); |
| 1779 | acc0 = fma(values5, w4, acc0); |
| 1780 | acc0 = fma(values6, w5, acc0); |
| 1781 | acc1 = fma(values5, w3, acc1); |
| 1782 | acc1 = fma(values6, w4, acc1); |
| 1783 | acc1 = fma(values7, w5, acc1); |
| 1784 | |
| 1785 | acc0 = fma(values8, w6, acc0); |
| 1786 | acc0 = fma(values9, w7, acc0); |
| 1787 | acc0 = fma(values10, w8, acc0); |
| 1788 | acc1 = fma(values9, w6, acc1); |
| 1789 | acc1 = fma(values10, w7, acc1); |
| 1790 | acc1 = fma(values11, w8, acc1); |
| 1791 | |
| 1792 | acc2 = fma(values4, w0, acc2); |
| 1793 | acc2 = fma(values5, w1, acc2); |
| 1794 | acc2 = fma(values6, w2, acc2); |
| 1795 | acc3 = fma(values5, w0, acc3); |
| 1796 | acc3 = fma(values6, w1, acc3); |
| 1797 | acc3 = fma(values7, w2, acc3); |
| 1798 | |
| 1799 | acc2 = fma(values8, w3, acc2); |
| 1800 | acc2 = fma(values9, w4, acc2); |
| 1801 | acc2 = fma(values10, w5, acc2); |
| 1802 | acc3 = fma(values9, w3, acc3); |
| 1803 | acc3 = fma(values10, w4, acc3); |
| 1804 | acc3 = fma(values11, w5, acc3); |
| 1805 | |
| 1806 | acc2 = fma(values12, w6, acc2); |
| 1807 | acc2 = fma(values13, w7, acc2); |
| 1808 | acc2 = fma(values14, w8, acc2); |
| 1809 | acc3 = fma(values13, w6, acc3); |
| 1810 | acc3 = fma(values14, w7, acc3); |
| 1811 | acc3 = fma(values15, w8, acc3); |
| 1812 | |
| 1813 | #if defined(HAS_BIAS) |
| 1814 | Vector biases = CONVERT_TO_VECTOR_STRUCT(biases); |
| 1815 | |
Giorgio Arena | e6bb3c6 | 2018-08-23 11:19:11 +0100 | [diff] [blame] | 1816 | VEC_FLOAT bias_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)biases.ptr); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1817 | |
| 1818 | acc0 += bias_values; |
| 1819 | acc1 += bias_values; |
| 1820 | acc2 += bias_values; |
| 1821 | acc3 += bias_values; |
| 1822 | #endif // defined(HAS_BIAS) |
| 1823 | |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1824 | #if defined(DST_DEPTH) |
| 1825 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + (z * NUM_PLANES_PROCESSED) * dst_step_z + b * dst_stride_w; |
| 1826 | #else /* defined(DST_DEPTH) */ |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1827 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + (z * NUM_PLANES_PROCESSED) * dst_step_z; |
Georgios Pinitas | 3704464 | 2018-10-30 14:53:25 +0000 | [diff] [blame] | 1828 | #endif /* defined(DST_DEPTH) */ |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1829 | |
| 1830 | VSTORE(VEC_SIZE) |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 1831 | (ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, acc0, A_VAL, B_VAL), 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1832 | VSTORE(VEC_SIZE) |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 1833 | (ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, acc1, A_VAL, B_VAL), 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1834 | |
| 1835 | #if((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0) |
| 1836 | if((z * NUM_PLANES_PROCESSED + 1) < DST_DIM_2) |
| 1837 | #endif // ((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0) |
| 1838 | { |
| 1839 | VSTORE(VEC_SIZE) |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 1840 | (ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, acc2, A_VAL, B_VAL), 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y + 1 * dst_stride_z)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1841 | VSTORE(VEC_SIZE) |
Usama Arif | 6a98a6e | 2019-05-10 17:07:27 +0100 | [diff] [blame] | 1842 | (ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, acc3, A_VAL, B_VAL), 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y + 1 * dst_stride_z)); |
Giorgio Arena | d051e97 | 2018-06-20 11:46:42 +0100 | [diff] [blame] | 1843 | } |
| 1844 | } |
| 1845 | |
| 1846 | #endif // defined(NUM_ROWS_PROCESSED) && defined(NUM_PLANES_PROCESSED) |
Gian Marco Iodice | 3ae323f | 2019-08-13 14:56:50 +0100 | [diff] [blame] | 1847 | #endif // defined(VEC_SIZE) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT) && defined(DATA_TYPE) |