Giuseppe Rossini | f01201a | 2019-11-06 14:57:49 +0000 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame^] | 2 | * Copyright (c) 2019 Arm Limited. |
Giuseppe Rossini | f01201a | 2019-11-06 14:57:49 +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 | */ |
| 24 | |
| 25 | /* |
| 26 | * !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! |
| 27 | * |
| 28 | * NOTE: Header to be included by implementation files only. |
| 29 | * |
| 30 | * !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! |
| 31 | */ |
| 32 | |
| 33 | #include <limits> |
| 34 | |
| 35 | #include "arm.hpp" |
| 36 | #include "impl_base.hpp" |
| 37 | #include "depthwise_quantized.hpp" |
| 38 | |
| 39 | #pragma once |
| 40 | |
| 41 | namespace { |
| 42 | |
| 43 | template < |
| 44 | unsigned int OutputTileRows, unsigned int OutputTileCols, |
| 45 | unsigned int KernelRows, unsigned int KernelCols, |
| 46 | unsigned int StrideRows, unsigned int StrideCols, |
| 47 | typename FInput, typename FOutput |
| 48 | > |
| 49 | static inline void tilefn_hybrid( |
| 50 | int n_channels, |
| 51 | const void* packed_params, |
| 52 | FInput &get_input_ptr, |
| 53 | FOutput &get_output_ptr, |
| 54 | int32_t clamp_min, |
| 55 | int32_t clamp_max, |
| 56 | uint8_t input_offset, |
| 57 | uint8_t output_offset |
| 58 | ) |
| 59 | { |
| 60 | constexpr int InnerTileRows = StrideRows * (OutputTileRows - 1) + KernelRows; |
| 61 | constexpr int InnerTileCols = StrideCols * (OutputTileCols - 1) + KernelCols; |
| 62 | |
| 63 | // Offset into channels |
| 64 | int channel = 0; |
| 65 | |
| 66 | // Byte type pointer to weights and biases |
| 67 | const int8_t *wbptr = static_cast<const int8_t *>(packed_params); |
| 68 | |
| 69 | for (; n_channels >= 8; n_channels -= 8, channel += 8) |
| 70 | { |
| 71 | const int32x4_t biases[2] = { |
| 72 | vld1q_s32(reinterpret_cast<const int32_t *>(wbptr)), |
| 73 | vld1q_s32(reinterpret_cast<const int32_t *>(wbptr) + 4), |
| 74 | }; |
| 75 | const int32x4_t multipliers[2] = { |
| 76 | vld1q_s32(reinterpret_cast<const int32_t *>(wbptr) + 8), |
| 77 | vld1q_s32(reinterpret_cast<const int32_t *>(wbptr) + 12), |
| 78 | }; |
| 79 | const int32x4_t shifts[2] = { |
| 80 | vld1q_s32(reinterpret_cast<const int32_t *>(wbptr) + 16), |
| 81 | vld1q_s32(reinterpret_cast<const int32_t *>(wbptr) + 20), |
| 82 | }; |
| 83 | wbptr += 24*sizeof(int32_t); |
| 84 | |
| 85 | int16x8_t weights[KernelRows][KernelCols]; |
| 86 | for (unsigned int i = 0; i < KernelRows; i++) |
| 87 | { |
| 88 | for (unsigned int j = 0; j < KernelCols; j++) |
| 89 | { |
| 90 | const auto w = vld1_s8(wbptr); |
| 91 | weights[i][j] = reinterpret_cast<int16x8_t>(vmovl_s8(w)); |
| 92 | wbptr += 8; |
| 93 | } |
| 94 | } |
| 95 | |
| 96 | int16x8_t inputs[InnerTileRows][InnerTileCols]; |
| 97 | const uint8x8_t ioffset = vdup_n_u8(input_offset); |
| 98 | for (unsigned int i = 0; i < InnerTileRows; i++) |
| 99 | { |
| 100 | for (unsigned int j = 0; j < InnerTileCols; j++) |
| 101 | { |
| 102 | const auto x = vld1_u8(get_input_ptr(i, j, channel)); |
| 103 | inputs[i][j] = reinterpret_cast<int16x8_t>(vsubl_u8(x, ioffset)); |
| 104 | } |
| 105 | } |
| 106 | |
| 107 | for (unsigned int oi = 0; oi < OutputTileRows; oi++) |
| 108 | { |
| 109 | for (unsigned int oj = 0; oj < OutputTileCols; oj++) |
| 110 | { |
| 111 | int32x4_t accs[2]; |
| 112 | for (unsigned int i = 0; i < 2; i++) |
| 113 | { |
| 114 | accs[i] = biases[i]; |
| 115 | } |
| 116 | |
| 117 | for (unsigned int wi = 0; wi < KernelRows; wi++) |
| 118 | { |
| 119 | for (unsigned int wj = 0; wj < KernelCols; wj++) |
| 120 | { |
| 121 | const auto w = weights[wi][wj]; |
| 122 | const auto x = inputs[oi * StrideRows + wi][oj * StrideCols + wj]; |
| 123 | accs[0] = vmlal_s16(accs[0], vget_low_s16(w), vget_low_s16(x)); |
| 124 | accs[1] = vmlal_s16(accs[1], vget_high_s16(w), vget_high_s16(x)); |
| 125 | } |
| 126 | } |
| 127 | |
| 128 | int32x4_t final_accs[2]; |
| 129 | for (unsigned int i = 0; i < 2; i++) |
| 130 | { |
| 131 | const int32x4_t y = rounding_divide_by_exp2( |
| 132 | saturating_doubling_high_mul(accs[i], multipliers[i]), |
| 133 | shifts[i]); |
| 134 | const int32x4_t offset = reinterpret_cast<int32x4_t>(vdupq_n_u32(output_offset)); |
| 135 | final_accs[i] = vaddq_s32(y, offset); |
| 136 | final_accs[i] = vmaxq_s32(final_accs[i], vdupq_n_s32(clamp_min)); |
| 137 | final_accs[i] = vminq_s32(final_accs[i], vdupq_n_s32(clamp_max)); |
| 138 | } |
| 139 | |
| 140 | const auto elems_s16 = vuzpq_s16(vreinterpretq_s16_s32(final_accs[0]), |
| 141 | vreinterpretq_s16_s32(final_accs[1])); |
| 142 | const int8x16_t elems = vreinterpretq_s8_s16(elems_s16.val[0]); |
| 143 | const uint8x8_t output = |
| 144 | vget_low_u8(vreinterpretq_u8_s8(vuzpq_s8(elems, elems).val[0])); |
| 145 | |
| 146 | vst1_u8(get_output_ptr(oi, oj, channel), output); |
| 147 | } |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | for (; n_channels; n_channels--, channel++) |
| 152 | { |
| 153 | // Load bias |
| 154 | const int32_t bias = *reinterpret_cast<const int32_t *>(wbptr); |
| 155 | const int32_t multiplier = *reinterpret_cast<const int32_t *>(wbptr + sizeof(int32_t)); |
| 156 | const int32_t shift = *reinterpret_cast<const int32_t *>(wbptr + 2*sizeof(int32_t)); |
| 157 | |
| 158 | wbptr += 3*sizeof(int32_t); |
| 159 | |
| 160 | // Load weights |
| 161 | int16_t weights[KernelRows][KernelCols]; |
| 162 | for (unsigned int i = 0; i < KernelRows; i++) |
| 163 | { |
| 164 | for (unsigned int j = 0; j < KernelCols; j++) |
| 165 | { |
| 166 | weights[i][j] = *(wbptr++); |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | // Load the input activations |
| 171 | int16_t inputs[InnerTileRows][InnerTileCols]; |
| 172 | for (unsigned int i = 0; i < InnerTileRows; i++) |
| 173 | { |
| 174 | for (unsigned int j = 0; j < InnerTileCols; j++) |
| 175 | { |
| 176 | inputs[i][j] = *(get_input_ptr(i, j, channel)) - input_offset; |
| 177 | } |
| 178 | } |
| 179 | |
| 180 | // Perform the convolution |
| 181 | for (unsigned int oi = 0; oi < OutputTileRows; oi++) |
| 182 | { |
| 183 | for (unsigned int oj = 0; oj < OutputTileCols; oj++) |
| 184 | { |
| 185 | int32_t acc = bias; |
| 186 | |
| 187 | for (unsigned int wi = 0; wi < KernelRows; wi++) |
| 188 | { |
| 189 | for (unsigned int wj = 0; wj < KernelCols; wj++) |
| 190 | { |
| 191 | const auto w = weights[wi][wj], x = inputs[oi*StrideRows + wi][oj*StrideCols + wj]; |
| 192 | acc += w * x; |
| 193 | } |
| 194 | } |
| 195 | |
| 196 | // Requantize |
| 197 | acc = rounding_divide_by_exp2( |
| 198 | saturating_doubling_high_mul(acc, multiplier), |
| 199 | -shift); |
| 200 | acc += output_offset; |
| 201 | acc = std::max(acc, clamp_min); |
| 202 | acc = std::min(acc, clamp_max); |
| 203 | uint8_t output = static_cast<uint8_t>(acc); |
| 204 | *(get_output_ptr(oi, oj, channel)) = output; |
| 205 | } |
| 206 | } |
| 207 | } |
| 208 | } |
| 209 | |
| 210 | template < |
| 211 | unsigned int OutputTileRows, unsigned int OutputTileCols, |
| 212 | unsigned int KernelRows, unsigned int KernelCols, |
| 213 | unsigned int StrideRows, unsigned int StrideCols, |
| 214 | typename FInput, typename FOutput |
| 215 | > |
| 216 | static inline void execute_tilefn_hybrid( |
| 217 | int n_channels, |
| 218 | const void* packed_params, |
| 219 | const ActivationFunction actfn, |
| 220 | const qasymm8::QAsymm8Params &input_quant, |
| 221 | const qasymm8::QAsymm8Params &output_quant, |
| 222 | FInput &get_input_ptr, |
| 223 | FOutput &get_output_ptr) { |
| 224 | |
| 225 | // Compute min/max clamp values |
| 226 | int32_t clamp_min = std::numeric_limits<uint8_t>::min(); |
| 227 | int32_t clamp_max = std::numeric_limits<uint8_t>::max(); |
| 228 | |
| 229 | if (actfn == ActivationFunction::ReLU) { |
| 230 | clamp_min = output_quant.offset; |
| 231 | } |
| 232 | |
| 233 | // Disabling Relu6 for now |
| 234 | if (actfn == ActivationFunction::ReLU6) { |
| 235 | const int32_t top_rail = output_quant.quantize(6.0f); |
| 236 | clamp_max = std::min(clamp_max, top_rail); |
| 237 | } |
| 238 | |
| 239 | // Call the tile execution method |
| 240 | tilefn_hybrid<OutputTileRows, OutputTileCols, KernelRows, KernelCols, StrideRows, |
| 241 | StrideCols>(n_channels, packed_params, get_input_ptr, get_output_ptr, clamp_min, clamp_max, input_quant.offset, output_quant.offset); |
| 242 | } |
| 243 | } |
| 244 | |
| 245 | |
| 246 | |
| 247 | namespace depthwise { |
| 248 | using namespace qsymm8; |
| 249 | template < |
| 250 | unsigned int OutputTileRows, unsigned int OutputTileCols, |
| 251 | unsigned int KernelRows, unsigned int KernelCols, |
| 252 | unsigned int StrideRows, unsigned int StrideCols |
| 253 | > |
| 254 | QSymm8HybridPerChannelDepthwiseConvolution< |
| 255 | OutputTileRows, OutputTileCols, KernelRows, KernelCols, StrideRows, StrideCols |
| 256 | >::QSymm8HybridPerChannelDepthwiseConvolution( |
| 257 | int n_batches, int n_input_rows, int n_input_cols, int n_channels, |
| 258 | const ActivationFunction activation, |
| 259 | const QSymm8PerChannelParams& weight_quantisation, |
| 260 | const qasymm8::QAsymm8Params& input_quantisation, |
| 261 | const qasymm8::QAsymm8Params& output_quantisation, |
| 262 | unsigned int padding_top, |
| 263 | unsigned int padding_left, |
| 264 | unsigned int padding_bottom, |
| 265 | unsigned int padding_right |
| 266 | ) : QSymm8HybridPerChannelDepthwiseConvolution( |
| 267 | n_batches, n_input_rows, n_input_cols, n_channels, |
| 268 | activation, weight_quantisation, input_quantisation, output_quantisation, |
| 269 | QSymm8PerChannelRescaleParams::make_rescale_params(weight_quantisation, input_quantisation, output_quantisation), |
| 270 | padding_top, padding_left, padding_bottom, padding_right |
| 271 | ) |
| 272 | { |
| 273 | } |
| 274 | |
| 275 | template < |
| 276 | unsigned int OutputTileRows, unsigned int OutputTileCols, |
| 277 | unsigned int KernelRows, unsigned int KernelCols, |
| 278 | unsigned int StrideRows, unsigned int StrideCols |
| 279 | > |
| 280 | QSymm8HybridPerChannelDepthwiseConvolution< |
| 281 | OutputTileRows, OutputTileCols, KernelRows, KernelCols, StrideRows, StrideCols |
| 282 | >::QSymm8HybridPerChannelDepthwiseConvolution( |
| 283 | int n_batches, int n_input_rows, int n_input_cols, int n_channels, |
| 284 | const ActivationFunction activation, |
| 285 | const QSymm8PerChannelParams& weight_quantisation, |
| 286 | const qasymm8::QAsymm8Params& input_quantisation, |
| 287 | const qasymm8::QAsymm8Params& output_quantisation, |
| 288 | const QSymm8PerChannelRescaleParams& rescale_params, |
| 289 | unsigned int padding_top, |
| 290 | unsigned int padding_left, |
| 291 | unsigned int padding_bottom, |
| 292 | unsigned int padding_right |
| 293 | ) : Base( |
| 294 | n_batches, n_input_rows, n_input_cols, n_channels, activation, |
| 295 | padding_top, padding_left, padding_bottom, padding_right |
| 296 | ), |
| 297 | _weights_quant(weight_quantisation), |
| 298 | _input_quant(input_quantisation), |
| 299 | _output_quant(output_quantisation), |
| 300 | _rescale_parameters(rescale_params) |
| 301 | { |
| 302 | } |
| 303 | |
| 304 | template < |
| 305 | unsigned int OutputTileRows, unsigned int OutputTileCols, |
| 306 | unsigned int KernelRows, unsigned int KernelCols, |
| 307 | unsigned int StrideRows, unsigned int StrideCols |
| 308 | > |
| 309 | uint8_t QSymm8HybridPerChannelDepthwiseConvolution< |
| 310 | OutputTileRows, OutputTileCols, KernelRows, KernelCols, StrideRows, StrideCols |
| 311 | >::_input_padding_value(void) const |
| 312 | { |
| 313 | return _input_quant.offset; |
| 314 | } |
| 315 | |
| 316 | template < |
| 317 | unsigned int OutputTileRows, unsigned int OutputTileCols, |
| 318 | unsigned int KernelRows, unsigned int KernelCols, |
| 319 | unsigned int StrideRows, unsigned int StrideCols |
| 320 | > |
| 321 | void QSymm8HybridPerChannelDepthwiseConvolution< |
| 322 | OutputTileRows, OutputTileCols, KernelRows, KernelCols, StrideRows, StrideCols |
| 323 | >::_pack_params( |
| 324 | void * const buffer, |
| 325 | const void * const weights, |
| 326 | const unsigned int weight_row_stride, |
| 327 | const unsigned int weight_col_stride, |
| 328 | const void * const biases |
| 329 | ) const |
| 330 | { |
| 331 | const int8_t *wptr = static_cast<const int8_t *>(weights); |
| 332 | const int32_t *bptr = static_cast<const int32_t *>(biases); |
| 333 | const int32_t *mptr = static_cast<const int32_t *>(_rescale_parameters.multipliers.data()); |
| 334 | const int32_t *sptr = static_cast<const int32_t *>(_rescale_parameters.shifts.data()); |
| 335 | int8_t *outptr = static_cast<int8_t *>(buffer); |
| 336 | |
| 337 | // We set the vector length to use doubles on both Aarch64 and Aarch32. NOTE |
| 338 | // For SVE set this to half the vector length. |
| 339 | unsigned int veclen = 8; |
| 340 | |
| 341 | // While there are channels left to process, pack a vector length of them at |
| 342 | // a time and reduce the size of vector used as the size of the tensor |
| 343 | // decreases. |
| 344 | for ( |
| 345 | unsigned int n_channels = this->n_channels(); n_channels; |
| 346 | n_channels -= veclen, |
| 347 | outptr += veclen*(3*sizeof(int32_t) + this->kernel_rows*this->kernel_cols) |
| 348 | ) |
| 349 | { |
| 350 | // NOTE Ignore this section if using SVE, the vector length remains the |
| 351 | // same and we just don't fill a full register for the tail. |
| 352 | while (n_channels < veclen) |
| 353 | { |
| 354 | // Reduce the vector length to either 8 or 1 (scalar) |
| 355 | // TODO Support more vector lengths in `execute_tile`. |
| 356 | veclen = (veclen == 16) ? 8 : 1; |
| 357 | } |
| 358 | |
| 359 | // Get pointers to bias and weight portions of the output structure. |
| 360 | int32_t *out_bptr = reinterpret_cast<int32_t *>(outptr); |
| 361 | int32_t *out_mptr = reinterpret_cast<int32_t *>(outptr + veclen*sizeof(int32_t)); |
| 362 | int32_t *out_sptr = reinterpret_cast<int32_t *>(outptr + 2*veclen*sizeof(int32_t)); |
| 363 | int8_t *out_wptr = outptr + 3*veclen*sizeof(int32_t); |
| 364 | |
| 365 | // Copy a vector length of elements |
| 366 | for (unsigned int n = 0; n < veclen && n < n_channels; n++) |
| 367 | { |
| 368 | const int32_t bias = (bptr != nullptr) ? *(bptr++) : 0; |
| 369 | const int32_t multiplier = (mptr != nullptr) ? *(mptr++) : 0; |
| 370 | const int32_t shift = (sptr != nullptr) ? *(sptr++) : 0; |
| 371 | |
| 372 | out_bptr[n] = bias; |
| 373 | out_mptr[n] = multiplier; |
| 374 | out_sptr[n] = -shift; |
| 375 | |
| 376 | for (unsigned int i = 0; i < KernelRows; i++) |
| 377 | { |
| 378 | int8_t *row_outptr = out_wptr + i*KernelCols*veclen; |
| 379 | for (unsigned int j = 0; j < KernelCols; j++) |
| 380 | { |
| 381 | int8_t w = *(wptr + i*weight_row_stride + j*weight_col_stride); |
| 382 | row_outptr[j*veclen + n] = w; |
| 383 | } |
| 384 | } |
| 385 | wptr++; |
| 386 | } |
| 387 | } |
| 388 | } |
| 389 | |
| 390 | |
| 391 | template < |
| 392 | unsigned int OutputTileRows, unsigned int OutputTileCols, |
| 393 | unsigned int KernelRows, unsigned int KernelCols, |
| 394 | unsigned int StrideRows, unsigned int StrideCols |
| 395 | > |
| 396 | template <ActivationFunction Activation> |
| 397 | void QSymm8HybridPerChannelDepthwiseConvolution< |
| 398 | OutputTileRows, OutputTileCols, KernelRows, KernelCols, StrideRows, StrideCols |
| 399 | >::execute_tile( |
| 400 | int n_channels, |
| 401 | const void* packed_params, |
| 402 | const uint8_t* inptr, |
| 403 | unsigned int in_row_stride, |
| 404 | unsigned int in_col_stride, |
| 405 | uint8_t* outptr, |
| 406 | unsigned int out_row_stride, |
| 407 | unsigned int out_col_stride |
| 408 | ) { |
| 409 | |
| 410 | // Construct methods to get pointers |
| 411 | const auto get_input_ptr = [inptr, in_row_stride, in_col_stride]( |
| 412 | const int i, const int j, const int channel) { |
| 413 | return inptr + i * in_row_stride + j * in_col_stride + channel; |
| 414 | }; |
| 415 | |
| 416 | const auto get_output_ptr = [outptr, out_row_stride, out_col_stride]( |
| 417 | const int i, const int j, const int channel) { |
| 418 | return outptr + i * out_row_stride + j * out_col_stride + channel; |
| 419 | }; |
| 420 | |
| 421 | execute_tilefn_hybrid<OutputTileRows, OutputTileCols, KernelRows, KernelCols, |
| 422 | StrideRows, StrideCols>( |
| 423 | n_channels, packed_params, Activation, _input_quant, _output_quant, get_input_ptr, get_output_ptr); |
| 424 | } |
| 425 | |
| 426 | template < |
| 427 | unsigned int OutputTileRows, unsigned int OutputTileCols, |
| 428 | unsigned int KernelRows, unsigned int KernelCols, |
| 429 | unsigned int StrideRows, unsigned int StrideCols |
| 430 | > |
| 431 | template <ActivationFunction Activation> |
| 432 | void QSymm8HybridPerChannelDepthwiseConvolution< |
| 433 | OutputTileRows, OutputTileCols, KernelRows, KernelCols, StrideRows, StrideCols |
| 434 | >::execute_tile( |
| 435 | int n_channels, |
| 436 | const void* packed_params, |
| 437 | const uint8_t* inptrs[Base::inner_tile_rows][Base::inner_tile_cols], |
| 438 | uint8_t* outptrs[Base::output_tile_rows][Base::output_tile_cols] |
| 439 | ) { |
| 440 | // Construct methods to get pointers |
| 441 | const auto get_input_ptr = [inptrs](const int i, const int j, |
| 442 | const int channel) { |
| 443 | return inptrs[i][j] + channel; |
| 444 | }; |
| 445 | |
| 446 | const auto get_output_ptr = [outptrs](const int i, const int j, |
| 447 | const int channel) { |
| 448 | return outptrs[i][j] + channel; |
| 449 | }; |
| 450 | |
| 451 | // Call the tile execution method |
| 452 | execute_tilefn_hybrid<OutputTileRows, OutputTileCols, KernelRows, KernelCols, |
| 453 | StrideRows, StrideCols>( |
| 454 | n_channels, packed_params, Activation, _input_quant, _output_quant, get_input_ptr, get_output_ptr); |
| 455 | } |
| 456 | |
| 457 | } // namespace depthwise |