Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 1 | /* |
Michalis Spyrou | f6402dd | 2018-01-26 15:06:19 +0000 | [diff] [blame] | 2 | * Copyright (c) 2016-2018 ARM Limited. |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "helpers.h" |
| 25 | |
| 26 | /** Calculate square sum of a vector |
| 27 | * |
| 28 | * @param[in] input Pointer to the first pixel. |
| 29 | * |
| 30 | * @return square sum of vector. |
| 31 | */ |
| 32 | inline DATA_TYPE square_sum(__global const DATA_TYPE *input) |
| 33 | { |
| 34 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 35 | in = vload16(0, input); |
| 36 | |
| 37 | in *= in; |
| 38 | |
| 39 | in.s01234567 += in.s89ABCDEF; |
| 40 | in.s0123 += in.s4567; |
| 41 | in.s01 += in.s23; |
| 42 | |
| 43 | return (in.s0 + in.s1); |
| 44 | } |
| 45 | |
| 46 | /** Calculate sum of a vector |
| 47 | * |
| 48 | * @param[in] input Pointer to the first pixel. |
| 49 | * |
| 50 | * @return sum of vector. |
| 51 | */ |
| 52 | inline DATA_TYPE sum(__global const DATA_TYPE *input) |
| 53 | { |
| 54 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 55 | in = vload16(0, input); |
| 56 | |
| 57 | in.s01234567 += in.s89ABCDEF; |
| 58 | in.s0123 += in.s4567; |
| 59 | in.s01 += in.s23; |
| 60 | |
| 61 | return (in.s0 + in.s1); |
| 62 | } |
| 63 | |
Michalis Spyrou | 7e9391b | 2018-10-05 14:49:28 +0100 | [diff] [blame] | 64 | /** This kernel performs parallel reduction given an operation on x-axis. |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 65 | * |
| 66 | * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 67 | * @note The operation we want to perform must be passed at compile time using -DOPERATION e.g. -DOPERATION=square_sum |
Michalis Spyrou | 7e9391b | 2018-10-05 14:49:28 +0100 | [diff] [blame] | 68 | * @note The mean flag must be passed at compile time using -DMEAN if we want to compute the mean value |
| 69 | * @note The width size must be passed at compile time using -DWIDTH e.g. -DWIDTH=128 if we want to compute the mean value |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 70 | * |
Michalis Spyrou | 7e9391b | 2018-10-05 14:49:28 +0100 | [diff] [blame] | 71 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 72 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 73 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
Michalis Spyrou | f6402dd | 2018-01-26 15:06:19 +0000 | [diff] [blame] | 74 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 75 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 76 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 77 | * @param[in] partial_sum_ptr The local buffer to hold sumed values. Supported data types: same as @p src_ptt |
Michalis Spyrou | f6402dd | 2018-01-26 15:06:19 +0000 | [diff] [blame] | 78 | * @param[in] partial_sum_stride_x Stride of the output tensor in X dimension (in bytes) |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 79 | * @param[in] partial_sum_step_x partial_sum_stride_x * number of elements along X processed per workitem(in bytes) |
Michalis Spyrou | f6402dd | 2018-01-26 15:06:19 +0000 | [diff] [blame] | 80 | * @param[in] partial_sum_stride_y Stride of the output tensor in Y dimension (in bytes) |
| 81 | * @param[in] partial_sum_step_y partial_sum_stride_y * number of elements along Y processed per workitem(in bytes) |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 82 | * @param[in] partial_sum_offset_first_element_in_bytes The offset of the first element in the source tensor |
Michalis Spyrou | f6402dd | 2018-01-26 15:06:19 +0000 | [diff] [blame] | 83 | * @param[in] local_sums Local buffer for storing the partial sum |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 84 | */ |
Michalis Spyrou | 7e9391b | 2018-10-05 14:49:28 +0100 | [diff] [blame] | 85 | __kernel void reduction_operation_x( |
Michalis Spyrou | f6402dd | 2018-01-26 15:06:19 +0000 | [diff] [blame] | 86 | IMAGE_DECLARATION(src), |
| 87 | IMAGE_DECLARATION(partial_sum), |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 88 | __local DATA_TYPE *local_sums) |
| 89 | { |
Michalis Spyrou | f6402dd | 2018-01-26 15:06:19 +0000 | [diff] [blame] | 90 | Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| 91 | Image partial_sum = CONVERT_TO_IMAGE_STRUCT(partial_sum); |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 92 | |
| 93 | unsigned int lsize = get_local_size(0); |
| 94 | unsigned int lid = get_local_id(0); |
| 95 | |
Michalis Spyrou | f6402dd | 2018-01-26 15:06:19 +0000 | [diff] [blame] | 96 | for(unsigned int y = 0; y < get_local_size(1); ++y) |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 97 | { |
Michalis Spyrou | f6402dd | 2018-01-26 15:06:19 +0000 | [diff] [blame] | 98 | local_sums[lid] = OPERATION((__global DATA_TYPE *)offset(&src, 0, y)); |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 99 | barrier(CLK_LOCAL_MEM_FENCE); |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 100 | |
Michalis Spyrou | f6402dd | 2018-01-26 15:06:19 +0000 | [diff] [blame] | 101 | // Perform parallel reduction |
| 102 | for(unsigned int i = lsize >> 1; i > 0; i >>= 1) |
| 103 | { |
| 104 | if(lid < i) |
| 105 | { |
| 106 | local_sums[lid] += local_sums[lid + i]; |
| 107 | } |
| 108 | barrier(CLK_LOCAL_MEM_FENCE); |
| 109 | } |
| 110 | |
| 111 | if(lid == 0) |
| 112 | { |
Michalis Spyrou | 7e9391b | 2018-10-05 14:49:28 +0100 | [diff] [blame] | 113 | #if defined(MEAN) && defined(WIDTH) |
| 114 | if(y == get_local_size(1) - 1) |
| 115 | { |
| 116 | local_sums[0] /= WIDTH; |
| 117 | } |
| 118 | #endif /* defined(MEAN) && defined(WIDTH) */ |
Michalis Spyrou | f6402dd | 2018-01-26 15:06:19 +0000 | [diff] [blame] | 119 | ((__global DATA_TYPE *)offset(&partial_sum, get_group_id(0), y))[0] = local_sums[0]; |
| 120 | } |
Michalis Spyrou | 04f089c | 2017-08-08 17:42:38 +0100 | [diff] [blame] | 121 | } |
Michalis Spyrou | 7e9391b | 2018-10-05 14:49:28 +0100 | [diff] [blame] | 122 | } |
| 123 | |
| 124 | #if defined(WIDTH) |
| 125 | /** This kernel performs reduction on x-axis. (QASYMM8) |
| 126 | * |
| 127 | * @note The width size must be passed at compile time using -DWIDTH e.g. -DWIDTH=128 |
| 128 | * |
| 129 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8 |
| 130 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 131 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 132 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 133 | * @param[in] output_ptr The local buffer to hold sumed values. Supported data types: same as @p src_ptt |
| 134 | * @param[in] output_stride_x Stride of the output tensor in X dimension (in bytes) |
| 135 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 136 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 137 | */ |
| 138 | __kernel void reduction_operation_quantized_x( |
| 139 | VECTOR_DECLARATION(src), |
| 140 | VECTOR_DECLARATION(output)) |
| 141 | { |
| 142 | Vector src = CONVERT_TO_VECTOR_STRUCT(src); |
| 143 | Vector output = CONVERT_TO_VECTOR_STRUCT(output); |
| 144 | |
| 145 | uint res = 0; |
| 146 | |
| 147 | for(unsigned int x = 0; x < WIDTH; ++x) |
| 148 | { |
| 149 | res += *((__global uchar *)vector_offset(&src, x)); |
| 150 | } |
| 151 | |
| 152 | #if defined(MEAN) |
| 153 | res /= WIDTH; |
| 154 | #endif /* defined(MEAN) */ |
| 155 | |
| 156 | // Store result |
| 157 | *((__global uchar *)output.ptr) = convert_uchar(res); |
| 158 | } |
| 159 | #endif /* defined(HEIGHT) */ |
| 160 | |
| 161 | #if defined(HEIGHT) |
| 162 | /** This kernel performs reduction on y-axis. |
| 163 | * |
| 164 | * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float |
| 165 | * @note The height size must be passed at compile time using -DHEIGHT e.g. -DHEIGHT=128 |
| 166 | * |
| 167 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/F16/F32 |
| 168 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 169 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 170 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 171 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 172 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 173 | * @param[in] output_ptr The local buffer to hold sumed values. Supported data types: same as @p src_ptt |
| 174 | * @param[in] output_stride_x Stride of the output tensor in X dimension (in bytes) |
| 175 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 176 | * @param[in] output_stride_y Stride of the output tensor in Y dimension (in bytes) |
| 177 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 178 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 179 | */ |
| 180 | __kernel void reduction_operation_y( |
| 181 | IMAGE_DECLARATION(src), |
| 182 | IMAGE_DECLARATION(output)) |
| 183 | { |
| 184 | Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| 185 | Image output = CONVERT_TO_IMAGE_STRUCT(output); |
| 186 | |
| 187 | VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) |
| 188 | res = 0; |
| 189 | |
| 190 | for(unsigned int y = 0; y < HEIGHT; ++y) |
| 191 | { |
Michalis Spyrou | 8aaf93e | 2018-10-11 17:33:32 +0100 | [diff] [blame] | 192 | VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) |
| 193 | in = CONVERT(vload16(0, (__global DATA_TYPE *)offset(&src, 0, y)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)); |
| 194 | #if defined(SUM_SQUARE) |
| 195 | in *= in; |
| 196 | #endif // SQRSUM |
| 197 | res += in; |
Michalis Spyrou | 7e9391b | 2018-10-05 14:49:28 +0100 | [diff] [blame] | 198 | } |
| 199 | |
| 200 | #if defined(MEAN) |
| 201 | res /= HEIGHT; |
| 202 | #endif /* defined(MEAN) */ |
| 203 | |
| 204 | // Store result |
| 205 | vstore16(CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, 16)), 0, (__global DATA_TYPE *)output.ptr); |
| 206 | } |
| 207 | #endif /* defined(HEIGHT) */ |
| 208 | |
| 209 | #if defined(DEPTH) |
| 210 | /** This kernel performs reduction on z-axis. |
| 211 | * |
| 212 | * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float |
| 213 | * @note The depth size must be passed at compile time using -DDEPTH e.g. -DDEPTH=128 |
| 214 | * |
| 215 | * @param[in] input_ptr Pointer to the source tensor. Supported data types: QASYMM8/F16/F32 |
| 216 | * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) |
| 217 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 218 | * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 219 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 220 | * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 221 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 222 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 223 | * @param[in] output_ptr The local buffer to hold sumed values. Supported data types: same as @p input_ptt |
| 224 | * @param[in] output_stride_x Stride of the output tensor in X dimension (in bytes) |
| 225 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 226 | * @param[in] output_stride_y Stride of the output tensor in Y dimension (in bytes) |
| 227 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 228 | * @param[in] output_stride_z Stride of the output tensor in Z dimension (in bytes) |
| 229 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 230 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 231 | */ |
| 232 | __kernel void reduction_operation_z( |
| 233 | TENSOR3D_DECLARATION(input), |
| 234 | TENSOR3D_DECLARATION(output)) |
| 235 | { |
| 236 | Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); |
| 237 | Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); |
| 238 | |
| 239 | VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) |
| 240 | res = 0; |
| 241 | |
| 242 | for(unsigned int z = 0; z < DEPTH; ++z) |
| 243 | { |
Michalis Spyrou | 8aaf93e | 2018-10-11 17:33:32 +0100 | [diff] [blame] | 244 | VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) |
| 245 | in = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, z)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)); |
| 246 | #if defined(SUM_SQUARE) |
| 247 | in *= in; |
| 248 | #endif // SQRSUM |
| 249 | res += in; |
Michalis Spyrou | 7e9391b | 2018-10-05 14:49:28 +0100 | [diff] [blame] | 250 | } |
| 251 | |
| 252 | #if defined(MEAN) |
| 253 | res /= DEPTH; |
| 254 | #endif /* defined(MEAN) */ |
| 255 | |
| 256 | // Store result |
| 257 | vstore16(CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, 16)), 0, (__global DATA_TYPE *)output.ptr); |
| 258 | } |
| 259 | #endif /* defined(DEPTH) */ |
| 260 | |
| 261 | #if defined(BATCH) && defined(DEPTH) |
| 262 | /** This kernel performs reduction on w-axis. |
| 263 | * |
| 264 | * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float |
| 265 | * @note The batch size must be passed at compile time using -DBATCH e.g. -DBATCH=128 |
| 266 | * @note The depth size must be passed at compile time using -DBATCH e.g. -DDEPTH=128 |
| 267 | * |
| 268 | * @param[in] input_ptr Pointer to the source tensor. Supported data types: QASYMM8/F16/F32 |
| 269 | * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) |
| 270 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 271 | * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 272 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 273 | * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 274 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 275 | * @param[in] input_stride_w Stride of the source tensor in W dimension (in bytes) |
| 276 | * @param[in] input_step_w input_stride_w * number of elements along W processed per workitem(in bytes) |
| 277 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 278 | * @param[in] output_ptr The local buffer to hold sumed values. Supported data types: same as @p input_ptt |
| 279 | * @param[in] output_stride_x Stride of the output tensor in X dimension (in bytes) |
| 280 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 281 | * @param[in] output_stride_y Stride of the output tensor in Y dimension (in bytes) |
| 282 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 283 | * @param[in] output_stride_z Stride of the output tensor in Z dimension (in bytes) |
| 284 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 285 | * @param[in] output_stride_w Stride of the output tensor in W dimension (in bytes) |
| 286 | * @param[in] output_step_w output_stride_w * number of elements along W processed per workitem(in bytes) |
| 287 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 288 | */ |
| 289 | __kernel void reduction_operation_w( |
| 290 | TENSOR4D_DECLARATION(input), |
| 291 | TENSOR4D_DECLARATION(output)) |
| 292 | { |
| 293 | Tensor4D input = CONVERT_TO_TENSOR4D_STRUCT(input, DEPTH); |
| 294 | Tensor4D output = CONVERT_TO_TENSOR4D_STRUCT(output, DEPTH); |
| 295 | |
| 296 | VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) |
| 297 | res = 0; |
| 298 | |
| 299 | for(unsigned int w = 0; w < BATCH; ++w) |
| 300 | { |
Michalis Spyrou | 8aaf93e | 2018-10-11 17:33:32 +0100 | [diff] [blame] | 301 | VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16) |
| 302 | in = CONVERT(vload16(0, (__global DATA_TYPE *)tensor4D_offset(&input, 0, 0, 0, w)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)); |
| 303 | #if defined(SUM_SQUARE) |
| 304 | in *= in; |
| 305 | #endif // SQRSUM |
| 306 | res += in; |
Michalis Spyrou | 7e9391b | 2018-10-05 14:49:28 +0100 | [diff] [blame] | 307 | } |
| 308 | |
| 309 | #if defined(MEAN) |
| 310 | res /= BATCH; |
| 311 | #endif /* defined(MEAN) */ |
| 312 | |
| 313 | // Store result |
| 314 | vstore16(CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, 16)), 0, (__global DATA_TYPE *)output.ptr); |
| 315 | } |
| 316 | #endif /* defined(BATCH) && defined(DEPTH) */ |