George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 1 | /* |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 2 | * Copyright (c) 2019-2021, 2023 Arm Limited. |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [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 | */ |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 24 | #include "src/cpu/kernels/CpuGemmLowpOffsetContributionOutputStageKernel.h" |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 25 | |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 26 | #include "arm_compute/core/Error.h" |
| 27 | #include "arm_compute/core/Helpers.h" |
| 28 | #include "arm_compute/core/ITensor.h" |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 29 | #include "arm_compute/core/TensorInfo.h" |
| 30 | #include "arm_compute/core/Types.h" |
| 31 | #include "arm_compute/core/Utils.h" |
| 32 | #include "arm_compute/core/Validate.h" |
| 33 | #include "arm_compute/core/Window.h" |
Georgios Pinitas | ddb93bb | 2020-10-02 16:38:59 +0100 | [diff] [blame] | 34 | #include "src/core/NEON/NEAsymm.h" |
| 35 | #include "src/core/NEON/wrapper/wrapper.h" |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame] | 36 | #include "src/core/helpers/AutoConfiguration.h" |
| 37 | #include "src/core/helpers/WindowHelpers.h" |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 38 | |
| 39 | #include <arm_neon.h> |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 40 | |
| 41 | namespace arm_compute |
| 42 | { |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 43 | namespace cpu |
| 44 | { |
| 45 | namespace kernels |
| 46 | { |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 47 | namespace |
| 48 | { |
| 49 | inline int32x4x4_t load_results_input(const Iterator &mm_result_it, int32_t x) |
| 50 | { |
| 51 | return |
| 52 | { |
| 53 | { |
| 54 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + x + 0), |
| 55 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + x + 4), |
| 56 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + x + 8), |
| 57 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + x + 12) |
| 58 | } |
| 59 | }; |
| 60 | } |
| 61 | |
| 62 | inline int32x4x4_t load(const int32_t *ptr, int32_t x) |
| 63 | { |
| 64 | return |
| 65 | { |
| 66 | { |
| 67 | vld1q_s32(ptr + x + 0), |
| 68 | vld1q_s32(ptr + x + 4), |
| 69 | vld1q_s32(ptr + x + 8), |
| 70 | vld1q_s32(ptr + x + 12) |
| 71 | } |
| 72 | }; |
| 73 | } |
| 74 | |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 75 | inline int32x4x4_t add_s32(int32x4x4_t a, int32x4_t b) |
| 76 | { |
| 77 | return |
| 78 | { |
| 79 | { |
| 80 | vaddq_s32(a.val[0], b), |
| 81 | vaddq_s32(a.val[1], b), |
| 82 | vaddq_s32(a.val[2], b), |
| 83 | vaddq_s32(a.val[3], b) |
| 84 | } |
| 85 | }; |
| 86 | } |
| 87 | |
| 88 | inline int32x4x4_t add_s32(int32x4x4_t a, int32x4x4_t b) |
| 89 | { |
| 90 | return |
| 91 | { |
| 92 | { |
| 93 | vaddq_s32(a.val[0], b.val[0]), |
| 94 | vaddq_s32(a.val[1], b.val[1]), |
| 95 | vaddq_s32(a.val[2], b.val[2]), |
| 96 | vaddq_s32(a.val[3], b.val[3]) |
| 97 | } |
| 98 | }; |
| 99 | } |
| 100 | |
| 101 | inline int32x4x4_t mul_s32(int32x4x4_t &a, int32_t mul_scalar) |
| 102 | { |
| 103 | return |
| 104 | { |
| 105 | { |
| 106 | vmulq_n_s32(a.val[0], mul_scalar), |
| 107 | vmulq_n_s32(a.val[1], mul_scalar), |
| 108 | vmulq_n_s32(a.val[2], mul_scalar), |
| 109 | vmulq_n_s32(a.val[3], mul_scalar) |
| 110 | } |
| 111 | }; |
| 112 | } |
| 113 | |
| 114 | inline int32x4x4_t mul_s32(int32x4x4_t &a, const int32_t *multilpier) |
| 115 | { |
| 116 | return |
| 117 | { |
| 118 | { |
| 119 | vmulq_s32(a.val[0], vld1q_s32(multilpier)), |
| 120 | vmulq_s32(a.val[1], vld1q_s32(multilpier + 4)), |
| 121 | vmulq_s32(a.val[2], vld1q_s32(multilpier + 8)), |
| 122 | vmulq_s32(a.val[3], vld1q_s32(multilpier + 12)) |
| 123 | } |
| 124 | }; |
| 125 | } |
| 126 | |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 127 | inline int32x4x4_t get_a_offset(const int32_t *vector_sum_col_ptr, int32_t a_offset, int32_t x) |
| 128 | { |
| 129 | int32x4x4_t a_offset_term_s32 = load(vector_sum_col_ptr, x); |
| 130 | |
| 131 | a_offset_term_s32.val[0] = vmulq_n_s32(a_offset_term_s32.val[0], a_offset); |
| 132 | a_offset_term_s32.val[1] = vmulq_n_s32(a_offset_term_s32.val[1], a_offset); |
| 133 | a_offset_term_s32.val[2] = vmulq_n_s32(a_offset_term_s32.val[2], a_offset); |
| 134 | a_offset_term_s32.val[3] = vmulq_n_s32(a_offset_term_s32.val[3], a_offset); |
| 135 | return a_offset_term_s32; |
| 136 | } |
| 137 | |
| 138 | inline int32x4_t get_b_offset(const int32_t *vector_sum_row_ptr, int32_t b_offset) |
| 139 | { |
| 140 | int32x4_t b_offset_term_s32 = vld1q_dup_s32(vector_sum_row_ptr); |
| 141 | b_offset_term_s32 = vmulq_n_s32(b_offset_term_s32, b_offset); |
| 142 | return b_offset_term_s32; |
| 143 | } |
| 144 | |
| 145 | inline int32x4x4_t get_k_offset(int32_t k_offset) |
| 146 | { |
| 147 | return |
| 148 | { |
| 149 | { |
| 150 | vdupq_n_s32(k_offset), |
| 151 | vdupq_n_s32(k_offset), |
| 152 | vdupq_n_s32(k_offset), |
| 153 | vdupq_n_s32(k_offset) |
| 154 | } |
| 155 | }; |
| 156 | } |
| 157 | |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 158 | inline uint8x16_t finalize_quantization_floating_point(int32x4x4_t &in_s32, int32x4_t result_shift_s32, uint8x16_t min_u8, uint8x16_t max_u8, bool is_bounded_relu) |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 159 | { |
| 160 | const static int32x4_t zero_s32 = vdupq_n_s32(0); |
| 161 | |
| 162 | // Shift final result (negative value shift right) |
| 163 | in_s32.val[0] = vshlq_s32(in_s32.val[0], result_shift_s32); |
| 164 | in_s32.val[1] = vshlq_s32(in_s32.val[1], result_shift_s32); |
| 165 | in_s32.val[2] = vshlq_s32(in_s32.val[2], result_shift_s32); |
| 166 | in_s32.val[3] = vshlq_s32(in_s32.val[3], result_shift_s32); |
| 167 | |
| 168 | // Saturate negative values |
| 169 | in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32); |
| 170 | in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32); |
| 171 | in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32); |
| 172 | in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_s32); |
| 173 | |
| 174 | // Convert S32 to S16 |
| 175 | const int16x8x2_t in_s16 = |
| 176 | { |
| 177 | { |
| 178 | vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])), |
| 179 | vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3])) |
| 180 | } |
| 181 | }; |
| 182 | |
| 183 | // Convert S16 to U8 |
| 184 | uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_s16.val[0]), vqmovun_s16(in_s16.val[1])); |
| 185 | |
| 186 | if(is_bounded_relu) |
| 187 | { |
| 188 | out_u8 = vmaxq_u8(out_u8, min_u8); |
| 189 | out_u8 = vminq_u8(out_u8, max_u8); |
| 190 | } |
| 191 | |
| 192 | return out_u8; |
| 193 | } |
| 194 | |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 195 | inline int8x16_t finalize_quantization_floating_point(int32x4x4_t &in_s32, int32x4_t result_shift_s32, int8x16_t min_s8, int8x16_t max_s8, bool is_bounded_relu) |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 196 | { |
| 197 | const static int32x4_t zero_s32 = vdupq_n_s32(0); |
| 198 | |
| 199 | // Shift final result (negative value shift right) |
| 200 | in_s32.val[0] = vshlq_s32(in_s32.val[0], result_shift_s32); |
| 201 | in_s32.val[1] = vshlq_s32(in_s32.val[1], result_shift_s32); |
| 202 | in_s32.val[2] = vshlq_s32(in_s32.val[2], result_shift_s32); |
| 203 | in_s32.val[3] = vshlq_s32(in_s32.val[3], result_shift_s32); |
| 204 | |
| 205 | // Saturate negative values |
| 206 | in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32); |
| 207 | in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32); |
| 208 | in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32); |
| 209 | in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_s32); |
| 210 | |
| 211 | // Convert S32 to S16 |
| 212 | const int16x8x2_t in_s16 = |
| 213 | { |
| 214 | { |
| 215 | vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])), |
| 216 | vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3])) |
| 217 | } |
| 218 | }; |
| 219 | |
| 220 | // Convert S16 to S8 |
| 221 | int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1])); |
| 222 | |
| 223 | if(is_bounded_relu) |
| 224 | { |
| 225 | out_s8 = vmaxq_s8(out_s8, min_s8); |
| 226 | out_s8 = vminq_s8(out_s8, max_s8); |
| 227 | } |
| 228 | |
| 229 | return out_s8; |
| 230 | } |
| 231 | |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 232 | inline int8x16_t finalize_quantization_floating_point(int32x4x4_t &in_s32, int32x4x4_t result_shift_s32, int8x16_t min_s8, int8x16_t max_s8, bool is_bounded_relu) |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 233 | { |
| 234 | const static int32x4_t zero_s32 = vdupq_n_s32(0); |
| 235 | |
| 236 | // Shift final result (negative value shift right) |
| 237 | in_s32.val[0] = vshlq_s32(in_s32.val[0], vnegq_s32(result_shift_s32.val[0])); |
| 238 | in_s32.val[1] = vshlq_s32(in_s32.val[1], vnegq_s32(result_shift_s32.val[1])); |
| 239 | in_s32.val[2] = vshlq_s32(in_s32.val[2], vnegq_s32(result_shift_s32.val[2])); |
| 240 | in_s32.val[3] = vshlq_s32(in_s32.val[3], vnegq_s32(result_shift_s32.val[3])); |
| 241 | |
| 242 | // Saturate negative values |
| 243 | in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32); |
| 244 | in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32); |
| 245 | in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32); |
| 246 | in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_s32); |
| 247 | |
| 248 | // Convert S32 to S16 |
| 249 | const int16x8x2_t in_s16 = |
| 250 | { |
| 251 | { |
| 252 | vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])), |
| 253 | vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3])) |
| 254 | } |
| 255 | }; |
| 256 | |
| 257 | // Convert S16 to S8 |
| 258 | int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1])); |
| 259 | |
| 260 | if(is_bounded_relu) |
| 261 | { |
| 262 | out_s8 = vmaxq_s8(out_s8, min_s8); |
| 263 | out_s8 = vminq_s8(out_s8, max_s8); |
| 264 | } |
| 265 | |
| 266 | return out_s8; |
| 267 | } |
| 268 | |
Georgios Pinitas | 6e1791b | 2019-12-02 19:01:25 +0000 | [diff] [blame] | 269 | template <typename T> |
| 270 | struct VectorTyper |
| 271 | { |
| 272 | using stype = T; |
| 273 | using vtype = typename wrapper::traits::neon_bitvector_t<T, wrapper::traits::BitWidth::W128>; |
| 274 | }; |
| 275 | |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 276 | inline Window get_win_vector_sum(const Window &window) |
| 277 | { |
| 278 | Window win_vector_sum(window); |
| 279 | win_vector_sum.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 280 | win_vector_sum.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 281 | return win_vector_sum; |
| 282 | } |
| 283 | |
| 284 | inline Iterator get_vector_sum_col_it(const Window &window, const ITensor *vector_sum_col) |
| 285 | { |
| 286 | Iterator vector_sum_col_it(vector_sum_col, get_win_vector_sum(window)); |
| 287 | return vector_sum_col_it; |
| 288 | } |
| 289 | |
| 290 | inline Iterator get_vector_sum_row_it(const Window &window, const ITensor *vector_sum_row) |
| 291 | { |
| 292 | Window win_vector_sum_row = get_win_vector_sum(window); |
| 293 | win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 294 | Iterator vector_sum_row_it(vector_sum_row, win_vector_sum_row); |
| 295 | return vector_sum_row_it; |
| 296 | } |
| 297 | |
| 298 | inline Iterator get_bias_it(const Window &window, const ITensor *bias) |
| 299 | { |
| 300 | Window win_bias(window); |
| 301 | win_bias.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 302 | win_bias.set(Window::DimZ, Window::Dimension(0, 1, 1)); |
| 303 | Iterator bias_it(bias, win_bias); |
| 304 | return bias_it; |
| 305 | } |
| 306 | |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 307 | template <typename VT> |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 308 | inline void run_offset_contribution_output_stage_window(const int32_t *vector_sum_col_ptr, const int32_t *vector_sum_row_ptr, const int32_t *bias_ptr, Iterator mm_result_it, Iterator out_it, |
Georgios Pinitas | 6e1791b | 2019-12-02 19:01:25 +0000 | [diff] [blame] | 309 | const int32x4_t result_offset_s32, const int32x4_t result_shift_s32, |
| 310 | typename VT::vtype min_vec, typename VT::vtype max_vec, |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 311 | int32_t a_offset, int32_t b_offset, int32_t k_offset, |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 312 | int32_t multiplier, int32_t shift, int32_t offset, int32_t min_bound, int32_t max_bound, |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 313 | int window_step_x, int window_start_x, int window_end_x, bool has_a_offset, bool has_b_offset, bool has_bias, bool is_bounded_relu, bool is_fixed_point) |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 314 | { |
| 315 | int32x4x4_t offset_term_s32 = { 0, 0, 0, 0 }; |
| 316 | if(!is_fixed_point) |
| 317 | { |
| 318 | // Combine quantization offset with other offsets. |
| 319 | offset_term_s32 = add_s32(offset_term_s32, result_offset_s32); |
| 320 | } |
| 321 | if(has_a_offset && has_b_offset) |
| 322 | { |
| 323 | offset_term_s32 = add_s32(offset_term_s32, get_k_offset(k_offset)); |
| 324 | } |
| 325 | if(has_b_offset) |
| 326 | { |
| 327 | offset_term_s32 = add_s32(offset_term_s32, get_b_offset(vector_sum_row_ptr, b_offset)); |
| 328 | } |
| 329 | |
| 330 | int x = window_start_x; |
| 331 | for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| 332 | { |
| 333 | int32x4x4_t in_s32 = load_results_input(mm_result_it, x); |
| 334 | |
| 335 | if(has_a_offset) |
| 336 | { |
| 337 | in_s32 = add_s32(in_s32, get_a_offset(vector_sum_col_ptr, a_offset, x)); |
| 338 | } |
| 339 | if(has_bias) |
| 340 | { |
| 341 | in_s32 = add_s32(in_s32, load(bias_ptr, x)); |
| 342 | } |
| 343 | if(!is_fixed_point || has_b_offset) |
| 344 | { |
| 345 | in_s32 = add_s32(in_s32, offset_term_s32); |
| 346 | } |
| 347 | if(!is_fixed_point) |
| 348 | { |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 349 | in_s32 = mul_s32(in_s32, multiplier); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 350 | } |
| 351 | |
| 352 | if(is_fixed_point) |
| 353 | { |
Georgios Pinitas | 6e1791b | 2019-12-02 19:01:25 +0000 | [diff] [blame] | 354 | wrapper::vstore(reinterpret_cast<typename VT::stype *>(out_it.ptr() + x), |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 355 | finalize_quantization(in_s32, multiplier, shift, result_offset_s32, min_vec, max_vec, is_bounded_relu)); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 356 | } |
| 357 | else |
| 358 | { |
Georgios Pinitas | 6e1791b | 2019-12-02 19:01:25 +0000 | [diff] [blame] | 359 | wrapper::vstore(reinterpret_cast<typename VT::stype *>(out_it.ptr() + x), |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 360 | finalize_quantization_floating_point(in_s32, result_shift_s32, min_vec, max_vec, is_bounded_relu)); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 361 | } |
| 362 | } |
| 363 | // Compute left-over elements |
| 364 | for(; x < window_end_x; ++x) |
| 365 | { |
| 366 | int32_t in_value = *(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + x) + wrapper::vgetlane(offset_term_s32.val[0], 0); |
| 367 | |
| 368 | if(has_a_offset) |
| 369 | { |
| 370 | in_value += (*(vector_sum_col_ptr + x) * a_offset); |
| 371 | } |
| 372 | if(has_bias) |
| 373 | { |
| 374 | in_value += *(bias_ptr + x); |
| 375 | } |
| 376 | |
| 377 | if(is_fixed_point) |
| 378 | { |
| 379 | // Finalize and store the result |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 380 | *reinterpret_cast<typename VT::stype *>(out_it.ptr() + x) = finalize_quantization(in_value, multiplier, shift, offset, |
| 381 | static_cast<typename VT::stype>(min_bound), |
| 382 | static_cast<typename VT::stype>(max_bound), is_bounded_relu); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 383 | } |
| 384 | else |
| 385 | { |
| 386 | // Finalize quantization |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 387 | in_value = (in_value * multiplier) >> shift; |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 388 | |
| 389 | // Bound and store the result |
| 390 | if(is_bounded_relu) |
| 391 | { |
Georgios Pinitas | 6e1791b | 2019-12-02 19:01:25 +0000 | [diff] [blame] | 392 | in_value = static_cast<typename VT::stype>(std::max<int32_t>(min_bound, std::min<int32_t>(max_bound, in_value))); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 393 | } |
Georgios Pinitas | 6e1791b | 2019-12-02 19:01:25 +0000 | [diff] [blame] | 394 | *reinterpret_cast<typename VT::stype *>(out_it.ptr() + x) = static_cast<typename VT::stype>(std::max<int32_t>(static_cast<int32_t>(std::numeric_limits<typename VT::stype>::lowest()), |
| 395 | std::min<int32_t>(static_cast<int32_t>(std::numeric_limits<typename VT::stype>::max()), in_value))); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 396 | } |
| 397 | } |
| 398 | } |
| 399 | |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 400 | inline void run_offset_contribution_output_stage_window_symm(const int32_t *vector_sum_col_ptr, const int32_t *bias_ptr, Iterator mm_result_it, Iterator out_it, |
| 401 | const int32_t *result_multipliers, const int32_t *result_shifts, |
| 402 | const int32x4_t result_offset, int8x16_t min_s8, int8x16_t max_s8, |
| 403 | int32_t a_offset, int32_t offset, int32_t min_bound, int32_t max_bound, |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 404 | int window_step_x, int window_start_x, int window_end_x, bool has_a_offset, bool has_bias, bool is_bounded_relu, bool is_fixed_point) |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 405 | { |
| 406 | int32x4x4_t offset_term_s32 = { 0, 0, 0, 0 }; |
| 407 | if(!is_fixed_point) |
| 408 | { |
| 409 | // Combine quantization offset with other offsets. |
| 410 | offset_term_s32 = add_s32(offset_term_s32, result_offset); |
| 411 | } |
| 412 | |
| 413 | int x = window_start_x; |
| 414 | for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| 415 | { |
| 416 | int32x4x4_t in_s32 = load_results_input(mm_result_it, x); |
| 417 | |
| 418 | if(has_a_offset) |
| 419 | { |
| 420 | in_s32 = add_s32(in_s32, get_a_offset(vector_sum_col_ptr, a_offset, x)); |
| 421 | } |
| 422 | if(has_bias) |
| 423 | { |
| 424 | in_s32 = add_s32(in_s32, load(bias_ptr, x)); |
| 425 | } |
| 426 | if(!is_fixed_point) |
| 427 | { |
| 428 | in_s32 = add_s32(in_s32, offset_term_s32); |
| 429 | in_s32 = mul_s32(in_s32, result_multipliers + x); |
| 430 | } |
| 431 | |
| 432 | if(is_fixed_point) |
| 433 | { |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 434 | vst1q_s8(reinterpret_cast<int8_t *>(out_it.ptr() + x), finalize_quantization_symm(in_s32, load(result_multipliers, x), load(result_shifts, x), result_offset, min_s8, max_s8, is_bounded_relu)); |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 435 | } |
| 436 | else |
| 437 | { |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 438 | vst1q_s8(reinterpret_cast<int8_t *>(out_it.ptr() + x), finalize_quantization_floating_point(in_s32, load(result_shifts, x), min_s8, max_s8, is_bounded_relu)); |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 439 | } |
| 440 | } |
| 441 | // Compute left-over elements |
| 442 | for(; x < window_end_x; ++x) |
| 443 | { |
| 444 | int32_t in_value = *(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + x) + wrapper::vgetlane(offset_term_s32.val[0], 0); |
| 445 | |
| 446 | if(has_a_offset) |
| 447 | { |
| 448 | in_value += (*(vector_sum_col_ptr + x) * a_offset); |
| 449 | } |
| 450 | if(has_bias) |
| 451 | { |
| 452 | in_value += *(bias_ptr + x); |
| 453 | } |
| 454 | |
| 455 | if(is_fixed_point) |
| 456 | { |
| 457 | // Finalize and store the result |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 458 | *(out_it.ptr() + x) = finalize_quantization(in_value, result_multipliers[x], result_shifts[x], offset, static_cast<int8_t>(min_bound), static_cast<int8_t>(max_bound), is_bounded_relu); |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 459 | } |
| 460 | else |
| 461 | { |
| 462 | // Finalize quantization |
| 463 | in_value = (in_value * result_multipliers[x]) >> (-result_shifts[x]); |
| 464 | |
| 465 | // Bound and store the result |
| 466 | if(is_bounded_relu) |
| 467 | { |
| 468 | in_value = static_cast<int8_t>(std::max<int32_t>(min_bound, std::min<int32_t>(max_bound, in_value))); |
| 469 | } |
| 470 | *(out_it.ptr() + x) = static_cast<int8_t>(std::max<int32_t>(-128, std::min<int32_t>(127, in_value))); |
| 471 | } |
| 472 | } |
| 473 | } |
| 474 | |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 475 | template <typename T> |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 476 | void run_offset_contribution_output_stage(const Window &window, |
| 477 | const ITensor *mm_result, const ITensor *vector_sum_col, const ITensor *vector_sum_row, const ITensor *bias, ITensor *output, |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 478 | int32_t a_offset, int32_t b_offset, int32_t k_offset, bool is_vector_sum_col_batched, |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 479 | GEMMLowpOutputStageInfo output_stage, bool is_gemm3d, bool is_bounded_relu, bool is_fixed_point) |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 480 | { |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 481 | // Semantics of XYZW Explained for each tensor |
| 482 | // |
| 483 | // | Tensor | XYZW when is_gemm3d == false | XYZW when is_gemm3d == true | |
| 484 | // ------------------------------------------------------------------------------------------------------------------- |
| 485 | // | mm_result | x -> width, y -> height, z -> batch | x -> width, y -> height, z -> depth, w -> batch | |
| 486 | // | collapsed window | x -> width, y -> height, z -> batch | x -> width, y -> height, z -> depth * batch | |
| 487 | // | vector_sum_row | x -> height, y -> batch | x -> height * depth, y -> batch | |
| 488 | // | Vector_sum_col | x -> width, y -> batch | x -> width, y -> batch | |
| 489 | |
Georgios Pinitas | 6e1791b | 2019-12-02 19:01:25 +0000 | [diff] [blame] | 490 | using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>; |
| 491 | using Typer = VectorTyper<T>; |
| 492 | |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 493 | const int height_input = is_gemm3d ? mm_result->info()->dimension(1) : 0; |
| 494 | const int depth_input = is_gemm3d ? mm_result->info()->dimension(2) : 1; |
| 495 | |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 496 | const int32_t multiplier = output_stage.gemmlowp_multiplier; |
| 497 | const int32_t shift = output_stage.gemmlowp_shift; |
| 498 | const int32_t offset = output_stage.gemmlowp_offset; |
| 499 | const int32_t min_bound = output_stage.gemmlowp_min_bound; |
| 500 | const int32_t max_bound = output_stage.gemmlowp_max_bound; |
| 501 | |
Georgios Pinitas | 6e1791b | 2019-12-02 19:01:25 +0000 | [diff] [blame] | 502 | const int32x4_t result_offset_s32 = vdupq_n_s32(offset); |
| 503 | const int32x4_t result_shift_s32 = vdupq_n_s32(is_fixed_point ? shift : -shift); |
| 504 | const auto min_vec = wrapper::vdup_n(static_cast<T>(min_bound), ExactTagType{}); |
| 505 | const auto max_vec = wrapper::vdup_n(static_cast<T>(max_bound), ExactTagType{}); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 506 | |
| 507 | const int window_step_x = 16; |
| 508 | const auto window_start_x = static_cast<int>(window.x().start()); |
| 509 | const auto window_end_x = static_cast<int>(window.x().end()); |
| 510 | |
| 511 | Window win(window); |
| 512 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 513 | |
| 514 | Window collapsed_window = win.collapse_if_possible(win, Window::DimZ); |
| 515 | |
| 516 | Iterator mm_result_it(mm_result, win); |
| 517 | Iterator out_it(output, win); |
| 518 | |
| 519 | if((a_offset != 0) && (b_offset != 0)) |
| 520 | { |
| 521 | ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_col); |
| 522 | ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_row); |
| 523 | |
| 524 | Iterator vector_sum_col_it = get_vector_sum_col_it(collapsed_window, vector_sum_col); |
| 525 | Iterator vector_sum_row_it = get_vector_sum_row_it(collapsed_window, vector_sum_row); |
| 526 | |
| 527 | const size_t sum_row_stride_y = vector_sum_row->info()->strides_in_bytes().y(); |
| 528 | |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 529 | // Offset in case vector_sum_col is batched in y dimension |
| 530 | const int vector_sum_col_stride_batch = is_vector_sum_col_batched ? vector_sum_col->info()->strides_in_bytes().y() : 0; |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 531 | |
| 532 | if(bias != nullptr) |
| 533 | { |
| 534 | Iterator bias_it = get_bias_it(collapsed_window, bias); |
| 535 | execute_window_loop(collapsed_window, [&](const Coordinates & id) |
| 536 | { |
| 537 | const int batch_id = id.z() / depth_input; |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 538 | const auto vector_sum_col_ptr = reinterpret_cast<const int32_t *>(vector_sum_col_it.ptr() + batch_id * vector_sum_col_stride_batch); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 539 | const auto vector_sum_row_ptr = reinterpret_cast<const int32_t *>(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) |
| 540 | + id.y() + (id.z() % depth_input) * height_input; |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 541 | run_offset_contribution_output_stage_window<Typer>(vector_sum_col_ptr, vector_sum_row_ptr, reinterpret_cast<const int32_t *>(bias_it.ptr()), |
| 542 | mm_result_it, |
| 543 | out_it, |
| 544 | result_offset_s32, result_shift_s32, |
| 545 | min_vec, max_vec, a_offset, b_offset, k_offset, |
| 546 | multiplier, shift, offset, min_bound, max_bound, |
| 547 | window_step_x, window_start_x, window_end_x, true, true, true, is_bounded_relu, is_fixed_point); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 548 | }, |
| 549 | vector_sum_col_it, vector_sum_row_it, bias_it, mm_result_it, out_it); |
| 550 | } |
| 551 | else |
| 552 | { |
| 553 | execute_window_loop(collapsed_window, [&](const Coordinates & id) |
| 554 | { |
| 555 | const int batch_id = id.z() / depth_input; |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 556 | const auto vector_sum_col_ptr = reinterpret_cast<const int32_t *>(vector_sum_col_it.ptr() + batch_id * vector_sum_col_stride_batch); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 557 | const auto vector_sum_row_ptr = reinterpret_cast<const int32_t *>(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) |
| 558 | + id.y() + (id.z() % depth_input) * height_input; |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 559 | run_offset_contribution_output_stage_window<Typer>(vector_sum_col_ptr, vector_sum_row_ptr, nullptr, mm_result_it, out_it, |
| 560 | result_offset_s32, result_shift_s32, |
| 561 | min_vec, max_vec, a_offset, b_offset, k_offset, |
| 562 | multiplier, shift, offset, min_bound, max_bound, |
| 563 | window_step_x, window_start_x, window_end_x, true, true, false, is_bounded_relu, is_fixed_point); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 564 | }, |
| 565 | vector_sum_col_it, vector_sum_row_it, mm_result_it, out_it); |
| 566 | } |
| 567 | } |
| 568 | else if((a_offset == 0) && (b_offset != 0)) |
| 569 | { |
| 570 | ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_row); |
| 571 | |
| 572 | Iterator vector_sum_row_it = get_vector_sum_row_it(collapsed_window, vector_sum_row); |
| 573 | |
| 574 | const size_t sum_row_stride_y = vector_sum_row->info()->strides_in_bytes().y(); |
| 575 | |
| 576 | if(bias != nullptr) |
| 577 | { |
| 578 | Iterator bias_it = get_bias_it(collapsed_window, bias); |
| 579 | execute_window_loop(collapsed_window, [&](const Coordinates & id) |
| 580 | { |
| 581 | const int batch_id = id.z() / depth_input; |
| 582 | const auto vector_sum_row_ptr = reinterpret_cast<const int32_t *>(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) |
| 583 | + id.y() + (id.z() % depth_input) * height_input; |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 584 | run_offset_contribution_output_stage_window<Typer>(nullptr, vector_sum_row_ptr, reinterpret_cast<const int32_t *>(bias_it.ptr()), mm_result_it, |
| 585 | out_it, |
| 586 | result_offset_s32, result_shift_s32, |
| 587 | min_vec, max_vec, a_offset, b_offset, k_offset, |
| 588 | multiplier, shift, offset, min_bound, max_bound, |
| 589 | window_step_x, window_start_x, window_end_x, false, true, true, is_bounded_relu, is_fixed_point); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 590 | }, |
| 591 | vector_sum_row_it, bias_it, mm_result_it, out_it); |
| 592 | } |
| 593 | else |
| 594 | { |
| 595 | execute_window_loop(collapsed_window, [&](const Coordinates & id) |
| 596 | { |
| 597 | const int batch_id = id.z() / depth_input; |
| 598 | const auto vector_sum_row_ptr = reinterpret_cast<const int32_t *>(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) |
| 599 | + id.y() + (id.z() % depth_input) * height_input; |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 600 | run_offset_contribution_output_stage_window<Typer>(nullptr, vector_sum_row_ptr, nullptr, mm_result_it, out_it, |
| 601 | result_offset_s32, result_shift_s32, |
| 602 | min_vec, max_vec, a_offset, b_offset, k_offset, |
| 603 | multiplier, shift, offset, min_bound, max_bound, |
| 604 | window_step_x, window_start_x, window_end_x, false, true, false, is_bounded_relu, is_fixed_point); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 605 | }, |
| 606 | vector_sum_row_it, mm_result_it, out_it); |
| 607 | } |
| 608 | } |
| 609 | else if((a_offset != 0) && (b_offset == 0)) |
| 610 | { |
| 611 | ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_col); |
| 612 | |
| 613 | Iterator vector_sum_col_it = get_vector_sum_col_it(collapsed_window, vector_sum_col); |
| 614 | |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 615 | // Offset in case vector_sum_col is batched in y dimension |
| 616 | const int vector_sum_col_stride_batch = is_vector_sum_col_batched ? vector_sum_col->info()->strides_in_bytes().y() : 0; |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 617 | |
| 618 | if(bias != nullptr) |
| 619 | { |
| 620 | Iterator bias_it = get_bias_it(collapsed_window, bias); |
| 621 | execute_window_loop(collapsed_window, [&](const Coordinates & id) |
| 622 | { |
| 623 | const int batch_id = id.z() / depth_input; |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 624 | const auto vector_sum_col_ptr = reinterpret_cast<const int32_t *>(vector_sum_col_it.ptr() + batch_id * vector_sum_col_stride_batch); |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 625 | run_offset_contribution_output_stage_window<Typer>(vector_sum_col_ptr, nullptr, reinterpret_cast<const int32_t *>(bias_it.ptr()), mm_result_it, |
| 626 | out_it, |
| 627 | result_offset_s32, result_shift_s32, |
| 628 | min_vec, max_vec, a_offset, b_offset, k_offset, |
| 629 | multiplier, shift, offset, min_bound, max_bound, |
| 630 | window_step_x, window_start_x, window_end_x, true, false, true, is_bounded_relu, is_fixed_point); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 631 | }, |
| 632 | vector_sum_col_it, bias_it, mm_result_it, out_it); |
| 633 | } |
| 634 | else |
| 635 | { |
| 636 | execute_window_loop(collapsed_window, [&](const Coordinates & id) |
| 637 | { |
| 638 | const int batch_id = id.z() / depth_input; |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 639 | const auto vector_sum_col_ptr = reinterpret_cast<const int32_t *>(vector_sum_col_it.ptr() + batch_id * vector_sum_col_stride_batch); |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 640 | run_offset_contribution_output_stage_window<Typer>(vector_sum_col_ptr, nullptr, nullptr, mm_result_it, out_it, |
| 641 | result_offset_s32, result_shift_s32, |
| 642 | min_vec, max_vec, a_offset, b_offset, k_offset, |
| 643 | multiplier, shift, offset, min_bound, max_bound, |
| 644 | window_step_x, window_start_x, window_end_x, true, false, false, is_bounded_relu, is_fixed_point); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 645 | }, |
| 646 | vector_sum_col_it, mm_result_it, out_it); |
| 647 | } |
| 648 | } |
| 649 | else |
| 650 | { |
| 651 | if(bias != nullptr) |
| 652 | { |
| 653 | Iterator bias_it = get_bias_it(collapsed_window, bias); |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 654 | execute_window_loop(collapsed_window, [&](const Coordinates &) |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 655 | { |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 656 | run_offset_contribution_output_stage_window<Typer>(nullptr, nullptr, reinterpret_cast<const int32_t *>(bias_it.ptr()), mm_result_it, out_it, |
| 657 | result_offset_s32, result_shift_s32, |
| 658 | min_vec, max_vec, a_offset, b_offset, k_offset, |
| 659 | multiplier, shift, offset, min_bound, max_bound, |
| 660 | window_step_x, window_start_x, window_end_x, false, false, true, is_bounded_relu, is_fixed_point); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 661 | }, |
| 662 | bias_it, mm_result_it, out_it); |
| 663 | } |
| 664 | else |
| 665 | { |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 666 | execute_window_loop(collapsed_window, [&](const Coordinates &) |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 667 | { |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 668 | run_offset_contribution_output_stage_window<Typer>(nullptr, nullptr, nullptr, mm_result_it, out_it, |
| 669 | result_offset_s32, result_shift_s32, |
| 670 | min_vec, max_vec, a_offset, b_offset, k_offset, |
| 671 | multiplier, shift, offset, min_bound, max_bound, |
| 672 | window_step_x, window_start_x, window_end_x, false, false, false, is_bounded_relu, is_fixed_point); |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 673 | }, |
| 674 | mm_result_it, out_it); |
| 675 | } |
| 676 | return; |
| 677 | } |
| 678 | } |
| 679 | |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 680 | void run_offset_contribution_output_stage_symm(const Window &window, |
| 681 | const ITensor *mm_result, const ITensor *vector_sum_col, const ITensor *vector_sum_row, const ITensor *bias, ITensor *output, |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 682 | int32_t a_offset, int32_t b_offset, int32_t k_offset, bool is_vector_sum_col_batched, |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 683 | GEMMLowpOutputStageInfo output_stage, bool is_gemm3d, bool is_bounded_relu, bool is_fixed_point) |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 684 | { |
| 685 | ARM_COMPUTE_UNUSED(vector_sum_row, b_offset, k_offset); |
| 686 | |
| 687 | const int depth_input = is_gemm3d ? mm_result->info()->dimension(2) : 1; |
| 688 | |
| 689 | const int32_t offset = output_stage.gemmlowp_offset; |
| 690 | const int32_t min_bound = output_stage.gemmlowp_min_bound; |
| 691 | const int32_t max_bound = output_stage.gemmlowp_max_bound; |
| 692 | |
| 693 | const int32_t *result_multipliers = output_stage.gemmlowp_multipliers.data(); |
| 694 | const int32_t *result_shifts = output_stage.gemmlowp_shifts.data(); |
| 695 | const int32x4_t result_offset_s32 = vdupq_n_s32(offset); |
| 696 | const int8x16_t min_s8 = vdupq_n_s8(static_cast<int8_t>(min_bound)); |
| 697 | const int8x16_t max_s8 = vdupq_n_s8(static_cast<int8_t>(max_bound)); |
| 698 | |
| 699 | const int window_step_x = 16; |
| 700 | const auto window_start_x = static_cast<int>(window.x().start()); |
| 701 | const auto window_end_x = static_cast<int>(window.x().end()); |
| 702 | |
| 703 | Window win(window); |
| 704 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 705 | |
| 706 | Window collapsed_window = win.collapse_if_possible(win, Window::DimZ); |
| 707 | |
| 708 | Iterator mm_result_it(mm_result, win); |
| 709 | Iterator out_it(output, win); |
| 710 | |
| 711 | if(a_offset != 0) |
| 712 | { |
| 713 | ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_col); |
| 714 | |
| 715 | Iterator vector_sum_col_it = get_vector_sum_col_it(collapsed_window, vector_sum_col); |
| 716 | |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 717 | // Offset in case vector_sum_col is batched in y dimension |
| 718 | const int vector_sum_col_stride_batch = is_vector_sum_col_batched ? vector_sum_col->info()->strides_in_bytes().y() : 0; |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 719 | |
| 720 | if(bias != nullptr) |
| 721 | { |
| 722 | Iterator bias_it = get_bias_it(collapsed_window, bias); |
| 723 | execute_window_loop(collapsed_window, [&](const Coordinates & id) |
| 724 | { |
| 725 | const int batch_id = id.z() / depth_input; |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 726 | const auto vector_sum_col_ptr = reinterpret_cast<const int32_t *>(vector_sum_col_it.ptr() + batch_id * vector_sum_col_stride_batch); |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 727 | run_offset_contribution_output_stage_window_symm(vector_sum_col_ptr, reinterpret_cast<const int32_t *>(bias_it.ptr()), mm_result_it, out_it, |
| 728 | result_multipliers, result_shifts, |
| 729 | result_offset_s32, min_s8, max_s8, |
| 730 | a_offset, offset, min_bound, max_bound, |
| 731 | window_step_x, window_start_x, window_end_x, true, true, is_bounded_relu, is_fixed_point); |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 732 | }, |
| 733 | vector_sum_col_it, bias_it, mm_result_it, out_it); |
| 734 | } |
| 735 | else |
| 736 | { |
| 737 | execute_window_loop(collapsed_window, [&](const Coordinates & id) |
| 738 | { |
| 739 | const int batch_id = id.z() / depth_input; |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 740 | const auto vector_sum_col_ptr = reinterpret_cast<const int32_t *>(vector_sum_col_it.ptr() + batch_id * vector_sum_col_stride_batch); |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 741 | run_offset_contribution_output_stage_window_symm(vector_sum_col_ptr, nullptr, mm_result_it, out_it, |
| 742 | result_multipliers, result_shifts, |
| 743 | result_offset_s32, min_s8, max_s8, |
| 744 | a_offset, offset, min_bound, max_bound, |
| 745 | window_step_x, window_start_x, window_end_x, true, false, is_bounded_relu, is_fixed_point); |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 746 | }, |
| 747 | vector_sum_col_it, mm_result_it, out_it); |
| 748 | } |
| 749 | } |
| 750 | else |
| 751 | { |
| 752 | if(bias != nullptr) |
| 753 | { |
| 754 | Iterator bias_it = get_bias_it(collapsed_window, bias); |
| 755 | execute_window_loop(collapsed_window, [&](const Coordinates &) |
| 756 | { |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 757 | run_offset_contribution_output_stage_window_symm(nullptr, reinterpret_cast<const int32_t *>(bias_it.ptr()), mm_result_it, out_it, |
| 758 | result_multipliers, result_shifts, |
| 759 | result_offset_s32, min_s8, max_s8, |
| 760 | a_offset, offset, min_bound, max_bound, |
| 761 | window_step_x, window_start_x, window_end_x, false, true, is_bounded_relu, is_fixed_point); |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 762 | }, |
| 763 | bias_it, mm_result_it, out_it); |
| 764 | } |
| 765 | else |
| 766 | { |
| 767 | execute_window_loop(collapsed_window, [&](const Coordinates &) |
| 768 | { |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 769 | run_offset_contribution_output_stage_window_symm(nullptr, nullptr, mm_result_it, out_it, |
| 770 | result_multipliers, result_shifts, |
| 771 | result_offset_s32, min_s8, max_s8, |
| 772 | a_offset, offset, min_bound, max_bound, |
| 773 | window_step_x, window_start_x, window_end_x, false, false, is_bounded_relu, is_fixed_point); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 774 | }, |
| 775 | mm_result_it, out_it); |
| 776 | } |
| 777 | return; |
| 778 | } |
| 779 | } |
| 780 | |
| 781 | Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, const ITensorInfo *output, |
| 782 | int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage) |
| 783 | { |
| 784 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mm_result, 1, DataType::S32); |
Giorgio Arena | 1856ff7 | 2020-02-07 13:46:45 +0000 | [diff] [blame] | 785 | if(output->data_type() != DataType::QASYMM8) |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 786 | { |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 787 | ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) > 1 && output_stage.gemmlowp_multipliers.size() > 1 && b_offset != 0); |
| 788 | } |
| 789 | ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 790 | ARM_COMPUTE_RETURN_ERROR_ON(output_stage.type != GEMMLowpOutputStageType::QUANTIZE_DOWN && output_stage.type != GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT); |
| 791 | |
| 792 | if(bias != nullptr) |
| 793 | { |
| 794 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32); |
| 795 | ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); |
| 796 | ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) != bias->dimension(0)); |
| 797 | } |
| 798 | |
| 799 | // If a_offset == 0, vector_sum_col can be a nullptr |
| 800 | if(a_offset != 0) |
| 801 | { |
| 802 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32); |
| 803 | ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->dimension(0) != mm_result->dimension(0)); |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 804 | ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->num_dimensions() > 2); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 805 | } |
| 806 | |
| 807 | // If b_offset == 0, vector_sum_row can be a nullptr |
| 808 | if(b_offset != 0) |
| 809 | { |
| 810 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32); |
| 811 | |
| 812 | // Check if input is a 3D reinterpretation |
| 813 | const bool reinterpret_as_3d = mm_result->num_dimensions() > 1 && mm_result->tensor_shape().y() != vector_sum_row->tensor_shape().x(); |
| 814 | |
| 815 | // Validate input |
| 816 | ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (mm_result->dimension(1) * mm_result->dimension(2))); |
| 817 | ARM_COMPUTE_RETURN_ERROR_ON(!reinterpret_as_3d && vector_sum_row->dimension(0) != mm_result->dimension(1)); |
| 818 | |
| 819 | TensorShape output_shape = output->tensor_shape(); |
| 820 | if(output_shape.num_dimensions() > 1) |
| 821 | { |
| 822 | const unsigned int output_batch_idx = reinterpret_as_3d ? 3 : 2; |
| 823 | |
| 824 | TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape(); |
| 825 | vector_sum_row_shape.collapse_from(1); |
| 826 | output_shape.collapse_from(output_batch_idx); |
| 827 | |
| 828 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[output_batch_idx], |
| 829 | "mm_result tensor must have the same number of batches of output tensor"); |
| 830 | |
| 831 | if(a_offset != 0) |
| 832 | { |
| 833 | TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape(); |
| 834 | vector_sum_col_shape.collapse_from(1); |
| 835 | |
| 836 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_col_shape[1] != 1 && vector_sum_col_shape[1] != vector_sum_row_shape[1], |
| 837 | "vector_sum_col tensor must have the same number of batches of vector_sum_row_shape or the number of batches must be set to 1"); |
| 838 | } |
| 839 | } |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 840 | |
| 841 | // Check Tensor Rank of vector_sum_row |
Mohammed Suhail Munshi | fbe94da | 2023-02-03 11:30:52 +0000 | [diff] [blame^] | 842 | ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_row->num_dimensions() > 3); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 843 | } |
| 844 | |
| 845 | if(output->total_size() != 0) |
| 846 | { |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 847 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 848 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mm_result, output); |
| 849 | } |
| 850 | |
| 851 | return Status{}; |
| 852 | } |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 853 | } // namespace |
| 854 | |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 855 | void CpuGemmLowpOffsetContributionOutputStageKernel::configure(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, |
| 856 | const ITensorInfo *vector_sum_row, const ITensorInfo *bias, ITensorInfo *dst, |
| 857 | int32_t k, int32_t a_offset, int32_t b_offset, |
| 858 | GEMMLowpOutputStageInfo output_stage) |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 859 | { |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 860 | ARM_COMPUTE_UNUSED(vector_sum_row, bias); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 861 | // Perform validate step |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 862 | ARM_COMPUTE_ERROR_ON_NULLPTR(mm_result, dst); |
| 863 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, dst, a_offset, b_offset, output_stage)); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 864 | |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 865 | _a_offset = a_offset; |
| 866 | _b_offset = b_offset; |
| 867 | _k_offset = a_offset * b_offset * k; |
| 868 | _output_stage = output_stage; |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 869 | |
| 870 | // If a_offset == 0, vector_sum_col can be a nullptr |
| 871 | if(a_offset != 0) |
| 872 | { |
| 873 | // Check if vector_sum_col_shape should be slidden or not |
| 874 | // Don't slide vector_sum_col_shape along the y dimension if vector_sum_col_shape has just 1 dimension and vector_sum_row_shape more than 1 |
| 875 | // This scenario can happen when the the matrix multiplication is used to perform a convolution operation |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 876 | _is_vector_sum_col_batched = vector_sum_col->tensor_shape().num_dimensions() > 1; |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 877 | } |
| 878 | |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 879 | // Output auto inizialitation if not yet initialized |
| 880 | auto_init_if_empty(*dst, mm_result->clone()->set_data_type(DataType::QASYMM8)); |
| 881 | |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 882 | // Configure kernel window |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 883 | Window win = calculate_max_window(*mm_result, Steps()); |
| 884 | |
| 885 | // Note: This kernel performs 16 elements per iteration. |
| 886 | // However, since we use a left-over for loop, we cannot have any read or write out of memory |
| 887 | // For this reason num_elems_processed_per_iteration is 1 and so update_window_and_padding() can be skipped |
| 888 | ICpuKernel::configure(win); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 889 | } |
| 890 | |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 891 | Status CpuGemmLowpOffsetContributionOutputStageKernel::validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, |
| 892 | const ITensorInfo *vector_sum_row, const ITensorInfo *bias, const ITensorInfo *output, |
| 893 | int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage) |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 894 | { |
| 895 | ARM_COMPUTE_ERROR_ON_NULLPTR(mm_result, output); |
| 896 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, output, a_offset, b_offset, output_stage)); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 897 | return Status{}; |
| 898 | } |
| 899 | |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 900 | void CpuGemmLowpOffsetContributionOutputStageKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 901 | { |
| 902 | ARM_COMPUTE_UNUSED(info); |
| 903 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 904 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); |
| 905 | |
| 906 | auto mm_result = tensors.get_const_tensor(TensorType::ACL_SRC_0); |
| 907 | auto vector_sum_col = tensors.get_const_tensor(TensorType::ACL_SRC_1); |
| 908 | auto vector_sum_row = tensors.get_const_tensor(TensorType::ACL_SRC_2); |
| 909 | auto bias = tensors.get_const_tensor(TensorType::ACL_SRC_3); |
| 910 | auto dst = tensors.get_tensor(TensorType::ACL_DST); |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 911 | |
| 912 | PixelValue type_min{}; |
| 913 | PixelValue type_max{}; |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 914 | std::tie(type_min, type_max) = get_min_max(dst->info()->data_type()); |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 915 | int32_t type_min_int = type_min.get<int32_t>(); |
| 916 | int32_t type_max_int = type_max.get<int32_t>(); |
| 917 | |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 918 | const bool reinterpret_as_3d = vector_sum_row != nullptr |
| 919 | && mm_result->info()->num_dimensions() > 1 |
| 920 | && mm_result->info()->tensor_shape().y() != vector_sum_row->info()->tensor_shape().x(); |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 921 | |
| 922 | const bool is_bounded_relu = !(_output_stage.gemmlowp_min_bound <= type_min_int && _output_stage.gemmlowp_max_bound >= type_max_int); |
| 923 | |
| 924 | // Check if we need to perform fixed point requantization |
| 925 | const bool is_fixed_point = _output_stage.type != GEMMLowpOutputStageType::QUANTIZE_DOWN; |
| 926 | |
| 927 | // Check if symmetric per-channel execution |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 928 | const bool is_signed = dst->info()->data_type() == DataType::QASYMM8_SIGNED; |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 929 | |
| 930 | // Check if symmetric per-channel execution |
| 931 | const bool is_symm = _output_stage.is_quantized_per_channel; |
| 932 | |
| 933 | if(is_symm) |
| 934 | { |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 935 | run_offset_contribution_output_stage_symm(window, mm_result, vector_sum_col, vector_sum_row, bias, dst, _a_offset, _b_offset, _k_offset, _is_vector_sum_col_batched, _output_stage, |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 936 | reinterpret_as_3d, is_bounded_relu, is_fixed_point); |
| 937 | } |
| 938 | else |
| 939 | { |
| 940 | if(is_signed) |
| 941 | { |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 942 | run_offset_contribution_output_stage<int8_t>(window, mm_result, vector_sum_col, vector_sum_row, bias, dst, _a_offset, _b_offset, _k_offset, _is_vector_sum_col_batched, _output_stage, |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 943 | reinterpret_as_3d, is_bounded_relu, is_fixed_point); |
| 944 | } |
| 945 | else |
| 946 | { |
Mohammed Suhail Munshi | 5b9d223 | 2023-01-25 11:51:50 +0000 | [diff] [blame] | 947 | run_offset_contribution_output_stage<uint8_t>(window, mm_result, vector_sum_col, vector_sum_row, bias, dst, _a_offset, _b_offset, _k_offset, _is_vector_sum_col_batched, _output_stage, |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 948 | reinterpret_as_3d, is_bounded_relu, is_fixed_point); |
| 949 | } |
| 950 | } |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 951 | } |
| 952 | |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 953 | const char *CpuGemmLowpOffsetContributionOutputStageKernel::name() const |
| 954 | { |
| 955 | return "CpuGemmLowpOffsetContributionOutputStageKernel"; |
| 956 | } |
| 957 | } // namespace kernels |
| 958 | } // namespace cpu |
Michalis Spyrou | 70d43a3 | 2020-06-22 17:05:43 +0100 | [diff] [blame] | 959 | } // namespace arm_compute |