Michalis Spyrou | b5a450a | 2021-01-06 17:40:30 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2021 Arm Limited. |
| 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 | #ifndef SRC_CORE_NEON_KERNELS_SOFTMAX_LIST_H |
| 25 | #define SRC_CORE_NEON_KERNELS_SOFTMAX_LIST_H |
| 26 | |
Michalis Spyrou | b5a450a | 2021-01-06 17:40:30 +0000 | [diff] [blame] | 27 | #include "src/core/NEON/NEFixedPoint.h" |
| 28 | #include "src/core/NEON/NEMath.h" |
Michalis Spyrou | 373b407 | 2021-01-20 16:41:12 +0000 | [diff] [blame] | 29 | #include "src/core/NEON/wrapper/wrapper.h" |
| 30 | #include "support/SaturateCast.h" |
Michalis Spyrou | b5a450a | 2021-01-06 17:40:30 +0000 | [diff] [blame] | 31 | |
| 32 | namespace arm_compute |
| 33 | { |
| 34 | namespace cpu |
| 35 | { |
Michalis Spyrou | b5a450a | 2021-01-06 17:40:30 +0000 | [diff] [blame] | 36 | template <typename T> |
| 37 | void neon_logits_1d_max(const ITensor *in, ITensor *out, const Window &window) |
| 38 | { |
Sheri Zhang | ac6499a | 2021-02-10 15:32:38 +0000 | [diff] [blame] | 39 | /** Neon vector tag type. */ |
Michalis Spyrou | b5a450a | 2021-01-06 17:40:30 +0000 | [diff] [blame] | 40 | using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>; |
| 41 | |
| 42 | constexpr int window_step_x = 16 / sizeof(T); |
| 43 | const auto window_start_x = static_cast<int>(window.x().start()); |
| 44 | const auto window_end_x = static_cast<int>(window.x().end()); |
| 45 | |
| 46 | Window win{ window }; |
| 47 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 48 | Iterator input(in, win); |
| 49 | Iterator output(out, win); |
| 50 | |
| 51 | const int sum_stages = log2(window_step_x / 2); |
| 52 | execute_window_loop(win, [&](const Coordinates &) |
| 53 | { |
| 54 | // Get pointers |
| 55 | const auto in_ptr = reinterpret_cast<const T *>(input.ptr()); |
| 56 | const auto out_ptr = reinterpret_cast<T *>(output.ptr()); |
| 57 | |
| 58 | // Init max value |
| 59 | auto vec_max = wrapper::vdup_n(support::cpp11::lowest<T>(), ExactTagType{}); |
| 60 | int x = window_start_x; |
| 61 | |
| 62 | for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| 63 | { |
| 64 | const auto current_value = wrapper::vloadq(in_ptr + x); |
| 65 | vec_max = wrapper::vmax(vec_max, current_value); |
| 66 | } |
| 67 | auto carry_max = wrapper::vpmax(wrapper::vgethigh(vec_max), wrapper::vgetlow(vec_max)); |
| 68 | |
| 69 | for(int i = 0; i < sum_stages; ++i) |
| 70 | { |
| 71 | carry_max = wrapper::vpmax(carry_max, carry_max); |
| 72 | } |
| 73 | T max_val = wrapper::vgetlane(carry_max, 0); |
| 74 | |
| 75 | // Compute left-over elements |
| 76 | for(; x < window_end_x; ++x) |
| 77 | { |
| 78 | max_val = *(in_ptr + x) > max_val ? *(in_ptr + x) : max_val; |
| 79 | } |
| 80 | |
| 81 | *out_ptr = max_val; |
| 82 | }, |
| 83 | input, output); |
| 84 | } |
| 85 | |
| 86 | template <typename T> |
| 87 | void neon_softmax_logits_1d_quantized(const ITensor *in, const ITensor *max, void *const tmp, |
| 88 | ITensor *out, float beta, bool is_log, const Window &window) |
| 89 | { |
| 90 | static_assert(std::is_same<T, qasymm8_t>::value |
| 91 | || std::is_same<T, qasymm8_signed_t>::value, |
| 92 | "quantized type should be either qasymm8_t or qasymm8_signed_t."); |
| 93 | |
| 94 | const int start_x = in->info()->valid_region().anchor.x(); |
| 95 | const int input_width = in->info()->valid_region().shape.x(); |
| 96 | |
| 97 | const float scale_beta = -beta * in->info()->quantization_info().uniform().scale; |
| 98 | const auto scale_beta_vec = vdupq_n_f32(scale_beta); |
| 99 | |
| 100 | Iterator in_it(in, window); |
| 101 | Iterator max_it(max, window); |
| 102 | Iterator out_it(out, window); |
| 103 | constexpr int vec_size = 16; |
| 104 | |
| 105 | execute_window_loop(window, [&](const Coordinates &) |
| 106 | { |
| 107 | /* Get pointers */ |
| 108 | const auto in_ptr = reinterpret_cast<const T *>(in_it.ptr()) + start_x; |
| 109 | const auto out_ptr = reinterpret_cast<T *>(out_it.ptr()) + start_x; |
| 110 | const auto tmp_ptr = reinterpret_cast<float *>(tmp); |
| 111 | |
| 112 | float sum{}; |
| 113 | float sum_inversed{}; |
| 114 | |
| 115 | /* Compute exponentials and sum */ |
| 116 | { |
| 117 | /* Get max value */ |
| 118 | const auto max_val = *reinterpret_cast<const T *>(max_it.ptr()); |
| 119 | const auto vec_max = wrapper::vdup_n(max_val, wrapper::traits::vector_128_tag{}); |
| 120 | |
| 121 | /* Init sum to zero */ |
| 122 | float32x4x4_t vec_sum = |
| 123 | { |
| 124 | vdupq_n_f32(0.f), |
| 125 | vdupq_n_f32(0.f), |
| 126 | vdupq_n_f32(0.f), |
| 127 | vdupq_n_f32(0.f), |
| 128 | }; |
| 129 | |
| 130 | /* Loop over row and compute exponentials and sum */ |
| 131 | int x = 0; |
| 132 | for(; x <= (input_width - vec_size); x += vec_size) |
| 133 | { |
| 134 | auto vec_elements = wrapper::vloadq(in_ptr + x); |
| 135 | vec_elements = wrapper::vqsub(vec_max, vec_elements); |
| 136 | auto vec_elements_flt = convert_int_to_float<float32x4x4_t>(vec_elements); |
| 137 | |
| 138 | if(is_log) |
| 139 | { |
| 140 | vec_elements_flt.val[0] = vmulq_f32(vec_elements_flt.val[0], scale_beta_vec); |
| 141 | vec_elements_flt.val[1] = vmulq_f32(vec_elements_flt.val[1], scale_beta_vec); |
| 142 | vec_elements_flt.val[2] = vmulq_f32(vec_elements_flt.val[2], scale_beta_vec); |
| 143 | vec_elements_flt.val[3] = vmulq_f32(vec_elements_flt.val[3], scale_beta_vec); |
| 144 | vec_sum.val[0] = vaddq_f32(vec_sum.val[0], vexpq_f32(vec_elements_flt.val[0])); |
| 145 | vec_sum.val[1] = vaddq_f32(vec_sum.val[1], vexpq_f32(vec_elements_flt.val[1])); |
| 146 | vec_sum.val[2] = vaddq_f32(vec_sum.val[2], vexpq_f32(vec_elements_flt.val[2])); |
| 147 | vec_sum.val[3] = vaddq_f32(vec_sum.val[3], vexpq_f32(vec_elements_flt.val[3])); |
| 148 | } |
| 149 | else |
| 150 | { |
| 151 | vec_elements_flt.val[0] = vexpq_f32(vmulq_f32(vec_elements_flt.val[0], scale_beta_vec)); |
| 152 | vec_elements_flt.val[1] = vexpq_f32(vmulq_f32(vec_elements_flt.val[1], scale_beta_vec)); |
| 153 | vec_elements_flt.val[2] = vexpq_f32(vmulq_f32(vec_elements_flt.val[2], scale_beta_vec)); |
| 154 | vec_elements_flt.val[3] = vexpq_f32(vmulq_f32(vec_elements_flt.val[3], scale_beta_vec)); |
| 155 | vec_sum.val[0] = vaddq_f32(vec_sum.val[0], vec_elements_flt.val[0]); |
| 156 | vec_sum.val[1] = vaddq_f32(vec_sum.val[1], vec_elements_flt.val[1]); |
| 157 | vec_sum.val[2] = vaddq_f32(vec_sum.val[2], vec_elements_flt.val[2]); |
| 158 | vec_sum.val[3] = vaddq_f32(vec_sum.val[3], vec_elements_flt.val[3]); |
| 159 | } |
| 160 | |
| 161 | vst4q_f32(tmp_ptr + x, vec_elements_flt); |
| 162 | } |
| 163 | |
| 164 | /* Reduce sum */ |
| 165 | const auto sum_16_byte = vaddq_f32(vaddq_f32(vec_sum.val[0], vec_sum.val[1]), vaddq_f32(vec_sum.val[2], vec_sum.val[3])); |
| 166 | auto sum_res = vpadd_f32(vget_high_f32(sum_16_byte), vget_low_f32(sum_16_byte)); |
| 167 | sum_res = vpadd_f32(sum_res, sum_res); |
| 168 | sum = wrapper::vgetlane(sum_res, 0); |
| 169 | |
| 170 | /* Run remaining elements */ |
| 171 | for(; x < input_width; ++x) |
| 172 | { |
| 173 | float element{}; |
| 174 | if(is_log) |
| 175 | { |
| 176 | element = (max_val - in_ptr[x]) * scale_beta; |
| 177 | sum += std::exp(element); |
| 178 | } |
| 179 | else |
| 180 | { |
| 181 | element = std::exp((max_val - in_ptr[x]) * scale_beta); |
| 182 | sum += element; |
| 183 | } |
| 184 | |
| 185 | tmp_ptr[x] = element; |
| 186 | } |
| 187 | |
| 188 | if(!is_log) |
| 189 | { |
| 190 | sum_inversed = 256.f / sum; |
| 191 | } |
| 192 | else |
| 193 | { |
| 194 | sum = std::log(sum); |
| 195 | } |
| 196 | } |
| 197 | |
| 198 | /* Normalize exponentials */ |
| 199 | { |
| 200 | constexpr bool is_qasymm8_signed = std::is_same<T, qasymm8_signed_t>::value; |
| 201 | /* Loop over row and compute softmax */ |
| 202 | int x = 0; |
| 203 | for(; x <= (input_width - vec_size); x += vec_size) |
| 204 | { |
| 205 | using int_vec_type = wrapper::traits::neon_vector_t<T, 16>; |
| 206 | float32x4x4_t vec_in = vld4q_f32(tmp_ptr + x); |
| 207 | int_vec_type normalized_value{}; |
| 208 | if(is_log) |
| 209 | { |
| 210 | const float32x4x4_t sub = |
| 211 | { |
| 212 | vsubq_f32(vec_in.val[0], vdupq_n_f32(sum)), |
| 213 | vsubq_f32(vec_in.val[1], vdupq_n_f32(sum)), |
| 214 | vsubq_f32(vec_in.val[2], vdupq_n_f32(sum)), |
| 215 | vsubq_f32(vec_in.val[3], vdupq_n_f32(sum)), |
| 216 | }; |
| 217 | normalized_value = convert_float_to_int<float32x4x4_t, int_vec_type>(sub); |
| 218 | } |
| 219 | else |
| 220 | { |
| 221 | float32x4x4_t mul = |
| 222 | { |
| 223 | vmulq_f32(vec_in.val[0], vdupq_n_f32(sum_inversed)), |
| 224 | vmulq_f32(vec_in.val[1], vdupq_n_f32(sum_inversed)), |
| 225 | vmulq_f32(vec_in.val[2], vdupq_n_f32(sum_inversed)), |
| 226 | vmulq_f32(vec_in.val[3], vdupq_n_f32(sum_inversed)), |
| 227 | }; |
| 228 | |
| 229 | if(is_qasymm8_signed) |
| 230 | { |
| 231 | const auto offset_vec = wrapper::vdup_n(128.f, wrapper::traits::vector_128_tag{}); |
| 232 | mul.val[0] = wrapper::vsub(mul.val[0], offset_vec); |
| 233 | mul.val[1] = wrapper::vsub(mul.val[1], offset_vec); |
| 234 | mul.val[2] = wrapper::vsub(mul.val[2], offset_vec); |
| 235 | mul.val[3] = wrapper::vsub(mul.val[3], offset_vec); |
| 236 | } |
| 237 | |
| 238 | normalized_value = convert_float_to_int<float32x4x4_t, int_vec_type>(mul); |
| 239 | } |
| 240 | wrapper::vstore(out_ptr + x, normalized_value); |
| 241 | } |
| 242 | /* Run remaining elements */ |
| 243 | for(; x < input_width; ++x) |
| 244 | { |
| 245 | if(is_log) |
| 246 | { |
| 247 | out_ptr[x] = utils::cast::saturate_cast<T>(tmp_ptr[x] - sum); |
| 248 | } |
| 249 | else |
| 250 | { |
| 251 | out_ptr[x] = utils::cast::saturate_cast<T>((tmp_ptr[x] * sum_inversed) - (is_qasymm8_signed ? 128.f : 0)); |
| 252 | } |
| 253 | } |
| 254 | } |
| 255 | }, |
| 256 | in_it, max_it, out_it); |
| 257 | } |
| 258 | |
| 259 | template <typename T> |
| 260 | void neon_softmax_logits_1d_float(const ITensor *in, const ITensor *max, void *const tmp, |
| 261 | ITensor *out, const float beta, bool is_log, const Window &window) |
| 262 | { |
| 263 | const int start_x = in->info()->valid_region().anchor.x(); |
| 264 | const int input_width = in->info()->valid_region().shape.x(); |
| 265 | |
| 266 | Iterator in_it(in, window); |
| 267 | Iterator max_it(max, window); |
| 268 | Iterator out_it(out, window); |
| 269 | |
Sheri Zhang | ac6499a | 2021-02-10 15:32:38 +0000 | [diff] [blame] | 270 | /** Neon vector tag type. */ |
Michalis Spyrou | b5a450a | 2021-01-06 17:40:30 +0000 | [diff] [blame] | 271 | using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>; |
| 272 | |
| 273 | constexpr int vec_size = 16 / sizeof(T); |
| 274 | const int sum_stages = log2(vec_size / 2); |
| 275 | |
| 276 | execute_window_loop(window, [&](const Coordinates &) |
| 277 | { |
| 278 | /* Get pointers */ |
| 279 | const auto in_ptr = reinterpret_cast<const T *>(in_it.ptr()) + start_x; |
| 280 | const auto out_ptr = reinterpret_cast<T *>(out_it.ptr()) + start_x; |
| 281 | const auto tmp_ptr = reinterpret_cast<T *>(tmp); |
| 282 | |
| 283 | T sum{}; |
| 284 | T sum_inversed{}; |
| 285 | |
| 286 | /* Compute exponentials and sum */ |
| 287 | { |
| 288 | /* Get max value */ |
| 289 | const auto max_val = *reinterpret_cast<const T *>(max_it.ptr()); |
| 290 | const auto vec_max = wrapper::vdup_n(max_val, ExactTagType{}); |
| 291 | |
| 292 | /* Init sum to zero */ |
| 293 | auto vec_sum = wrapper::vdup_n(static_cast<T>(0), ExactTagType{}); |
| 294 | |
| 295 | /* Loop over row and compute exponentials and sum */ |
| 296 | int x = 0; |
| 297 | for(; x <= (input_width - vec_size); x += vec_size) |
| 298 | { |
| 299 | auto vec_elements = wrapper::vloadq(in_ptr + x); |
| 300 | vec_elements = wrapper::vsub(vec_elements, vec_max); |
| 301 | if(is_log) |
| 302 | { |
| 303 | vec_elements = wrapper::vmul(vec_elements, wrapper::vdup_n(static_cast<T>(beta), ExactTagType{})); |
| 304 | vec_sum = wrapper::vadd(vec_sum, wrapper::vexpq(vec_elements)); |
| 305 | } |
| 306 | else |
| 307 | { |
| 308 | vec_elements = wrapper::vexpq(wrapper::vmul(vec_elements, wrapper::vdup_n(static_cast<T>(beta), ExactTagType{}))); |
| 309 | vec_sum = wrapper::vadd(vec_sum, vec_elements); |
| 310 | } |
| 311 | wrapper::vstore(tmp_ptr + x, vec_elements); |
| 312 | } |
| 313 | |
| 314 | /* Reduce sum */ |
| 315 | auto sum_res = wrapper::vpadd(wrapper::vgethigh(vec_sum), wrapper::vgetlow(vec_sum)); |
| 316 | for(int i = 0; i < sum_stages; ++i) |
| 317 | { |
| 318 | sum_res = wrapper::vpadd(sum_res, sum_res); |
| 319 | } |
| 320 | sum = wrapper::vgetlane(sum_res, 0); |
| 321 | |
| 322 | /* Run remaining elements */ |
| 323 | for(; x < input_width; ++x) |
| 324 | { |
| 325 | T element{}; |
| 326 | |
| 327 | if(is_log) |
| 328 | { |
| 329 | element = (in_ptr[x] - max_val) * beta; |
| 330 | sum += std::exp(element); |
| 331 | } |
| 332 | else |
| 333 | { |
| 334 | element = std::exp((in_ptr[x] - max_val) * beta); |
| 335 | sum += element; |
| 336 | } |
| 337 | tmp_ptr[x] = element; |
| 338 | } |
| 339 | |
| 340 | if(!is_log) |
| 341 | { |
| 342 | sum_inversed = T(1) / sum; |
| 343 | } |
| 344 | else |
| 345 | { |
| 346 | sum = static_cast<T>(std::log(sum)); |
| 347 | } |
| 348 | } |
| 349 | |
| 350 | /* Normalize exponentials */ |
| 351 | { |
| 352 | /* Loop over row and compute softmax */ |
| 353 | int x = 0; |
| 354 | for(; x <= (input_width - vec_size); x += vec_size) |
| 355 | { |
| 356 | auto vec_in = wrapper::vloadq(tmp_ptr + x); |
| 357 | auto normalized_value = wrapper::vdup_n(static_cast<T>(0), ExactTagType{}); |
| 358 | if(is_log) |
| 359 | { |
| 360 | normalized_value = wrapper::vsub(vec_in, wrapper::vdup_n(static_cast<T>(sum), ExactTagType{})); |
| 361 | } |
| 362 | else |
| 363 | { |
| 364 | normalized_value = wrapper::vmul(vec_in, wrapper::vdup_n(static_cast<T>(sum_inversed), ExactTagType{})); |
| 365 | } |
| 366 | wrapper::vstore(out_ptr + x, normalized_value); |
| 367 | } |
| 368 | /* Run remaining elements */ |
| 369 | for(; x < input_width; ++x) |
| 370 | { |
| 371 | if(is_log) |
| 372 | { |
| 373 | out_ptr[x] = tmp_ptr[x] - sum; |
| 374 | } |
| 375 | else |
| 376 | { |
| 377 | out_ptr[x] = tmp_ptr[x] * sum_inversed; |
| 378 | } |
| 379 | } |
| 380 | } |
| 381 | }, |
| 382 | in_it, max_it, out_it); |
| 383 | } |
| 384 | |
| 385 | } // namespace cpu |
| 386 | } // namespace arm_compute |
| 387 | |
| 388 | #endif /* SRC_CORE_NEON_KERNELS_SOFTMAX_LIST_H */ |