| 1 | /******************************************************************************* |
|---|---|
| 2 | * Copyright 2016-2022 Intel Corporation |
| 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | *******************************************************************************/ |
| 16 | |
| 17 | #include "oneapi/dnnl/dnnl_types.h" |
| 18 | |
| 19 | #include "common/bfloat16.hpp" |
| 20 | #include "common/c_types_map.hpp" |
| 21 | #include "common/dnnl_thread.hpp" |
| 22 | #include "common/type_helpers.hpp" |
| 23 | #include "common/utils.hpp" |
| 24 | #include "cpu/gemm_convolution_utils.hpp" |
| 25 | #include "cpu/scale_utils.hpp" |
| 26 | #if DNNL_X64 |
| 27 | #include "cpu/x64/injectors/jit_uni_postops_injector.hpp" |
| 28 | #endif |
| 29 | |
| 30 | #include "cpu/platform.hpp" |
| 31 | |
| 32 | #if DNNL_X64 |
| 33 | #include "cpu/x64/cpu_isa_traits.hpp" |
| 34 | #endif |
| 35 | |
| 36 | namespace dnnl { |
| 37 | namespace impl { |
| 38 | namespace cpu { |
| 39 | |
| 40 | using namespace dnnl::impl::status; |
| 41 | using namespace dnnl::impl::utils; |
| 42 | using namespace prop_kind; |
| 43 | using namespace data_type; |
| 44 | |
| 45 | single_gemm_conv_chunk_desc_t::single_gemm_conv_chunk_desc_t(dim_t d_off, |
| 46 | dim_t d_size, dim_t h_off, dim_t h_size, dim_t w_off, dim_t w_size) |
| 47 | : d_off_(d_off) |
| 48 | , d_size_(d_size) |
| 49 | , h_off_(h_off) |
| 50 | , h_size_(h_size) |
| 51 | , w_off_(w_off) |
| 52 | , w_size_(w_size) {} |
| 53 | |
| 54 | namespace jit_gemm_convolution_utils { |
| 55 | |
| 56 | template <typename data_type_t> |
| 57 | void im2col_3d(const conv_gemm_conf_t &jcp, const data_type_t *im, |
| 58 | data_type_t *col, dim_t od, int spatial_step, int spatial_block) { |
| 59 | using data_t = |
| 60 | typename conditional<data_traits<data_type_t>::data_type == bf16, |
| 61 | uint16_t, data_type_t>::type; |
| 62 | const data_t *__restrict _im |
| 63 | = reinterpret_cast<const data_t *__restrict>(im); |
| 64 | data_t *__restrict _col = reinterpret_cast<data_t *__restrict>(col); |
| 65 | |
| 66 | const size_t OHW = spatial_block; |
| 67 | const size_t im_step = jcp.ih * jcp.iw * jcp.id; |
| 68 | const size_t col_step = jcp.ks * OHW; |
| 69 | |
| 70 | auto compute_im2col_outer_padding = [&](dim_t ic) { |
| 71 | const data_t *__restrict im_loc = _im + ic * im_step; |
| 72 | data_t *__restrict col_loc = _col + ic * col_step; |
| 73 | dim_t id = od * jcp.stride_d - jcp.f_pad; |
| 74 | for (dim_t kd = 0; kd < jcp.kd; ++kd) { |
| 75 | data_t *__restrict col_ = col_loc + kd * jcp.kh * jcp.kw * OHW; |
| 76 | if (id < 0 || id >= jcp.id) { |
| 77 | dim_t ih_ = -jcp.t_pad; |
| 78 | for (dim_t kh = 0; kh < jcp.kh; ++kh) { |
| 79 | dim_t ih = ih_; |
| 80 | for (dim_t oh = 0; oh < jcp.oh; ++oh) { |
| 81 | if (ih < 0 || ih >= jcp.ih) { |
| 82 | ih += jcp.stride_h; |
| 83 | continue; |
| 84 | } |
| 85 | dim_t iw_ = -jcp.l_pad; |
| 86 | for (dim_t kw = 0; kw < jcp.kw; ++kw) { |
| 87 | dim_t iw = iw_; |
| 88 | for (dim_t ow = 0; ow < jcp.ow; ++ow) { |
| 89 | if (iw < 0 || iw >= jcp.iw) { |
| 90 | iw += jcp.stride_w; |
| 91 | continue; |
| 92 | } |
| 93 | |
| 94 | const size_t col_idx |
| 95 | = kw * OHW + oh * jcp.ow + ow; |
| 96 | |
| 97 | col_[col_idx] = 0; |
| 98 | iw += jcp.stride_w; |
| 99 | } |
| 100 | iw_ += (1 + jcp.dilate_w); |
| 101 | } |
| 102 | ih += jcp.stride_h; |
| 103 | } |
| 104 | ih_ += (1 + jcp.dilate_h); |
| 105 | col_ += jcp.kw * OHW; |
| 106 | } |
| 107 | } else { |
| 108 | const data_t *__restrict im_ = im_loc + id * jcp.ih * jcp.iw; |
| 109 | dim_t ih_ = -jcp.t_pad; |
| 110 | for (dim_t kh = 0; kh < jcp.kh; ++kh) { |
| 111 | dim_t ih = ih_; |
| 112 | for (dim_t oh = 0; oh < jcp.oh; ++oh) { |
| 113 | if (ih < 0 || ih >= jcp.ih) { |
| 114 | ih += jcp.stride_h; |
| 115 | continue; |
| 116 | } |
| 117 | dim_t iw_ = -jcp.l_pad; |
| 118 | for (dim_t kw = 0; kw < jcp.kw; ++kw) { |
| 119 | dim_t iw = iw_; |
| 120 | for (dim_t ow = 0; ow < jcp.ow; ++ow) { |
| 121 | if (iw < 0 || iw >= jcp.iw) { |
| 122 | iw += jcp.stride_w; |
| 123 | continue; |
| 124 | } |
| 125 | |
| 126 | const size_t col_idx |
| 127 | = kw * OHW + oh * jcp.ow + ow; |
| 128 | const size_t im_idx = ih * jcp.iw + iw; |
| 129 | |
| 130 | col_[col_idx] = im_[im_idx]; |
| 131 | iw += jcp.stride_w; |
| 132 | } |
| 133 | iw_ += (1 + jcp.dilate_w); |
| 134 | } |
| 135 | ih += jcp.stride_h; |
| 136 | } |
| 137 | ih_ += (1 + jcp.dilate_h); |
| 138 | col_ += jcp.kw * OHW; |
| 139 | } |
| 140 | } |
| 141 | id += (1 + jcp.dilate_d); |
| 142 | } |
| 143 | }; |
| 144 | auto compute_im2col_padding = [&](dim_t ic) { |
| 145 | const dim_t first_oh = spatial_step / jcp.ow; |
| 146 | const dim_t last_oh = (spatial_step + spatial_block - 1) / jcp.ow; |
| 147 | const dim_t oh_begin = first_oh; |
| 148 | const dim_t oh_end = last_oh + 1; |
| 149 | const dim_t first_ow = spatial_step % jcp.ow; |
| 150 | const dim_t last_ow = (spatial_step + spatial_block - 1) % jcp.ow; |
| 151 | |
| 152 | const data_t *__restrict im_loc = _im + ic * im_step; |
| 153 | data_t *__restrict col_loc = _col + ic * col_step; |
| 154 | dim_t id = od * jcp.stride_d - jcp.f_pad; |
| 155 | for (dim_t kd = 0; kd < jcp.kd; ++kd) { |
| 156 | data_t *__restrict col_ = col_loc + kd * jcp.kh * jcp.kw * OHW; |
| 157 | if (id < 0 || id >= jcp.id) { |
| 158 | for (dim_t kh = 0; kh < jcp.kh; ++kh) { |
| 159 | for (dim_t oh = oh_begin; oh < oh_end; ++oh) { |
| 160 | const dim_t ow_begin = (oh == first_oh) ? first_ow : 0; |
| 161 | const dim_t ow_end |
| 162 | = (oh == last_oh) ? (last_ow + 1) : jcp.ow; |
| 163 | for (dim_t kw = 0; kw < jcp.kw; ++kw) { |
| 164 | for (dim_t ow = ow_begin; ow < ow_end; ++ow) { |
| 165 | const size_t col_idx = kw * OHW + oh * jcp.ow |
| 166 | + ow - spatial_step; |
| 167 | col_[col_idx] = 0; |
| 168 | } |
| 169 | } |
| 170 | } |
| 171 | col_ += jcp.kw * OHW; |
| 172 | } |
| 173 | } else { |
| 174 | const data_t *__restrict im_ = im_loc + id * jcp.ih * jcp.iw; |
| 175 | dim_t ih_ = oh_begin * jcp.stride_h - jcp.t_pad; |
| 176 | for (dim_t kh = 0; kh < jcp.kh; ++kh) { |
| 177 | dim_t ih = ih_; |
| 178 | for (dim_t oh = oh_begin; oh < oh_end; ++oh) { |
| 179 | const dim_t ow_begin = (oh == first_oh) ? first_ow : 0; |
| 180 | const dim_t ow_end |
| 181 | = (oh == last_oh) ? (last_ow + 1) : jcp.ow; |
| 182 | if (ih < 0 || ih >= jcp.ih) { |
| 183 | for (dim_t kw = 0; kw < jcp.kw; ++kw) { |
| 184 | for (dim_t ow = ow_begin; ow < ow_end; ++ow) { |
| 185 | const size_t col_idx = kw * OHW |
| 186 | + oh * jcp.ow + ow - spatial_step; |
| 187 | col_[col_idx] = 0; |
| 188 | } |
| 189 | } |
| 190 | ih += jcp.stride_h; |
| 191 | continue; |
| 192 | } |
| 193 | dim_t iw_ = ow_begin * jcp.stride_w - jcp.l_pad; |
| 194 | for (dim_t kw = 0; kw < jcp.kw; ++kw) { |
| 195 | dim_t iw = iw_; |
| 196 | for (dim_t ow = ow_begin; ow < ow_end; ++ow) { |
| 197 | const size_t col_idx = kw * OHW + oh * jcp.ow |
| 198 | + ow - spatial_step; |
| 199 | if (iw < 0 || iw >= jcp.iw) { |
| 200 | col_[col_idx] = 0; |
| 201 | iw += jcp.stride_w; |
| 202 | continue; |
| 203 | } |
| 204 | const size_t im_idx = ih * jcp.iw + iw; |
| 205 | col_[col_idx] = im_[im_idx]; |
| 206 | iw += jcp.stride_w; |
| 207 | } |
| 208 | iw_ += (1 + jcp.dilate_w); |
| 209 | } |
| 210 | ih += jcp.stride_h; |
| 211 | } |
| 212 | ih_ += (1 + jcp.dilate_h); |
| 213 | col_ += jcp.kw * OHW; |
| 214 | } |
| 215 | } |
| 216 | id += (1 + jcp.dilate_d); |
| 217 | } |
| 218 | }; |
| 219 | |
| 220 | // zero padding is handled outside im2col |
| 221 | const bool outer_padding = jcp.os_nb_block == 1; |
| 222 | if (outer_padding) |
| 223 | parallel_nd(jcp.ic, compute_im2col_outer_padding); |
| 224 | else |
| 225 | parallel_nd(jcp.ic, compute_im2col_padding); |
| 226 | } |
| 227 | |
| 228 | template void im2col_3d(const conv_gemm_conf_t &jcp, const float *im, |
| 229 | float *col, dim_t od, int spatial_step, int spatial_block); |
| 230 | |
| 231 | template void im2col_3d(const conv_gemm_conf_t &jcp, const bfloat16_t *im, |
| 232 | bfloat16_t *col, dim_t od, int spatial_step, int spatial_block); |
| 233 | |
| 234 | /* imtr[ic][od][oh][ow] <-- im[id][ih][iw][ic]*/ |
| 235 | template <typename T> |
| 236 | void transpose_dt(const conv_gemm_conf_t &jcp, const T *__restrict im, |
| 237 | T *__restrict imtr) { |
| 238 | uint8_t shift = jcp.signed_input ? 128 : 0; |
| 239 | const dim_t ic_stride = jcp.id * jcp.ih * jcp.iw; |
| 240 | const dim_t IC = jcp.ngroups * jcp.ic; |
| 241 | const dim_t IHW = jcp.ih * jcp.iw; |
| 242 | constexpr dim_t ic_block = platform::get_cache_line_size(); |
| 243 | const dim_t nb_ic = jcp.ic / ic_block; |
| 244 | const dim_t ic_blocked = nb_ic * ic_block; |
| 245 | parallel_nd(jcp.id, jcp.ih, [&](dim_t id, dim_t ih) { |
| 246 | const T *__restrict im_h = im + id * IHW * IC + ih * jcp.iw * IC; |
| 247 | T *__restrict imtr_h = imtr + id * IHW + ih * jcp.iw; |
| 248 | for (dim_t iw = 0; iw < jcp.iw; iw++) { |
| 249 | const T *__restrict im_w = im_h + iw * IC; |
| 250 | T *__restrict imtr_w = imtr_h + iw; |
| 251 | for (dim_t icb = 0; icb < nb_ic; icb++) { |
| 252 | const T *__restrict im_icb = im_w + icb * ic_block; |
| 253 | T *__restrict imtr_icb = imtr_w + icb * ic_block * ic_stride; |
| 254 | PRAGMA_OMP_SIMD() |
| 255 | for (dim_t ic = 0; ic < ic_block; ic++) { |
| 256 | imtr_icb[ic * ic_stride] = im_icb[ic] + shift; |
| 257 | } |
| 258 | } |
| 259 | for (dim_t ic = ic_blocked; ic < jcp.ic; ic++) { |
| 260 | imtr_w[ic * ic_stride] = im_w[ic] + shift; |
| 261 | } |
| 262 | } |
| 263 | }); |
| 264 | } |
| 265 | |
| 266 | template void transpose_dt(const conv_gemm_conf_t &jcp, |
| 267 | const int8_t *__restrict im, int8_t *__restrict imtr); |
| 268 | template void transpose_dt(const conv_gemm_conf_t &jcp, |
| 269 | const uint8_t *__restrict im, uint8_t *__restrict imtr); |
| 270 | template void transpose_dt(const conv_gemm_conf_t &jcp, |
| 271 | const char *__restrict im, char *__restrict imtr); |
| 272 | template void transpose_dt(const conv_gemm_conf_t &jcp, |
| 273 | const float *__restrict im, float *__restrict imtr); |
| 274 | template void transpose_dt(const conv_gemm_conf_t &jcp, |
| 275 | const bfloat16_t *__restrict im, bfloat16_t *__restrict imtr); |
| 276 | |
| 277 | /* col[kd][kh][kw][g][ic][od][oh][ow] <-- im2col_dt_3d(im[id][ih][iw][g][ic]) */ |
| 278 | template <typename orig_im_dt, typename orig_col_dt> |
| 279 | void im2col_dt_3d(const conv_gemm_conf_t &jcp, const void *__restrict _imtr, |
| 280 | orig_col_dt *__restrict _col, dim_t od) { |
| 281 | // For performance reasons, use uint16_t as a proxy for bfloat16_t |
| 282 | using im_dt = typename utils::conditional<data_traits<orig_im_dt>::data_type |
| 283 | == bf16, |
| 284 | uint16_t, orig_im_dt>::type; |
| 285 | using col_dt = |
| 286 | typename utils::conditional<data_traits<orig_col_dt>::data_type |
| 287 | == bf16, |
| 288 | uint16_t, orig_col_dt>::type; |
| 289 | const im_dt *__restrict imtr |
| 290 | = reinterpret_cast<const im_dt *__restrict>(_imtr); |
| 291 | col_dt *__restrict col = reinterpret_cast<col_dt *__restrict>(_col); |
| 292 | |
| 293 | col_dt shift = static_cast<col_dt>(jcp.signed_input ? 128 : 0); |
| 294 | const dim_t dd = 1 + jcp.dilate_d; |
| 295 | const dim_t dh = 1 + jcp.dilate_h; |
| 296 | const dim_t dw = 1 + jcp.dilate_w; |
| 297 | const dim_t sd = jcp.stride_d; |
| 298 | const dim_t sh = jcp.stride_h; |
| 299 | const dim_t sw = jcp.stride_w; |
| 300 | const dim_t fp = jcp.f_pad; |
| 301 | const dim_t tp = jcp.t_pad; |
| 302 | const dim_t lp = jcp.l_pad; |
| 303 | const dim_t col_ic_s = jcp.oh * jcp.ow; |
| 304 | const dim_t col_kw_s = jcp.ic * col_ic_s; |
| 305 | const dim_t col_kh_s = jcp.kw * col_kw_s; |
| 306 | const dim_t col_kd_s = jcp.kh * col_kh_s; |
| 307 | const dim_t IHW = jcp.ih * jcp.iw; |
| 308 | const dim_t OHW = jcp.oh * jcp.ow; |
| 309 | |
| 310 | if (sd == 1 && sh == 1 && sw == 1 && dd == 1 && dh == 1 && dw == 1) |
| 311 | parallel_nd(jcp.kd, jcp.kh, jcp.kw, jcp.ic, |
| 312 | [&](dim_t kd, dim_t kh, dim_t kw, dim_t ic) { |
| 313 | col_dt *__restrict col_loc = col + kd * col_kd_s |
| 314 | + kh * col_kh_s + kw * col_kw_s + ic * col_ic_s; |
| 315 | const dim_t id = od - fp + kd; |
| 316 | if (id < 0 || id >= jcp.id) { |
| 317 | for (ptrdiff_t i = 0; i < OHW; i++) |
| 318 | col_loc[i] = shift; |
| 319 | return; |
| 320 | } |
| 321 | const im_dt *__restrict imtr_loc |
| 322 | = imtr + (ic * jcp.id + id) * IHW; |
| 323 | const dim_t oh_start = saturate(dim_t(0), jcp.oh, tp - kh); |
| 324 | const dim_t oh_end |
| 325 | = saturate(dim_t(0), jcp.oh, jcp.ih + tp - kh); |
| 326 | const dim_t ow_start = saturate(dim_t(0), jcp.ow, lp - kw); |
| 327 | const dim_t ow_end |
| 328 | = saturate(dim_t(0), jcp.ow, jcp.iw + lp - kw); |
| 329 | for (dim_t oh = oh_start, ih = oh_start - tp + kh; |
| 330 | oh < oh_end; oh++, ih++) { |
| 331 | col_dt *__restrict col_h = col_loc + oh * jcp.ow; |
| 332 | const im_dt *__restrict imtr_h = imtr_loc + ih * jcp.iw; |
| 333 | for (dim_t ow = ow_start, iw = ow_start - lp + kw; |
| 334 | ow < ow_end; ow++, iw++) { |
| 335 | col_h[ow] = imtr_h[iw]; |
| 336 | } |
| 337 | } |
| 338 | }); |
| 339 | else if (sd == 2 && sh == 2 && sw == 2 && dd == 1 && dh == 1 && dw == 1) |
| 340 | parallel_nd(jcp.kd, jcp.kh, jcp.kw, jcp.ic, |
| 341 | [&](dim_t kd, dim_t kh, dim_t kw, dim_t ic) { |
| 342 | col_dt *__restrict col_loc = col + kd * col_kd_s |
| 343 | + kh * col_kh_s + kw * col_kw_s + ic * col_ic_s; |
| 344 | const dim_t id = od * 2 - fp + kd; |
| 345 | if (id < 0 || id >= jcp.id) { |
| 346 | for (ptrdiff_t i = 0; i < OHW; i++) |
| 347 | col_loc[i] = shift; |
| 348 | return; |
| 349 | } |
| 350 | const im_dt *__restrict imtr_loc |
| 351 | = imtr + (ic * jcp.id + id) * IHW; |
| 352 | const dim_t oh_start |
| 353 | = saturate(dim_t(0), jcp.oh, div_up(tp - kh, 2)); |
| 354 | const dim_t oh_end = saturate( |
| 355 | dim_t(0), jcp.oh, div_up(jcp.ih + tp - kh, 2)); |
| 356 | const dim_t ow_start |
| 357 | = saturate(dim_t(0), jcp.ow, div_up(lp - kw, 2)); |
| 358 | const dim_t ow_end = saturate( |
| 359 | dim_t(0), jcp.ow, div_up(jcp.iw + lp - kw, 2)); |
| 360 | for (dim_t oh = oh_start, ih = oh_start * 2 - tp + kh; |
| 361 | oh < oh_end; ++oh, ih += 2) { |
| 362 | col_dt *__restrict col_h = col_loc + oh * jcp.ow; |
| 363 | const im_dt *__restrict imtr_h = imtr_loc + ih * jcp.iw; |
| 364 | for (dim_t ow = ow_start, iw = ow_start * 2 - lp + kw; |
| 365 | ow < ow_end; ++ow, iw += 2) { |
| 366 | col_h[ow] = imtr_h[iw]; |
| 367 | } |
| 368 | } |
| 369 | }); |
| 370 | else |
| 371 | parallel_nd(jcp.kd, jcp.kh, jcp.kw, jcp.ic, |
| 372 | [&](dim_t kd, dim_t kh, dim_t kw, dim_t ic) { |
| 373 | col_dt *__restrict col_loc = col + kd * col_kd_s |
| 374 | + kh * col_kh_s + kw * col_kw_s + ic * col_ic_s; |
| 375 | const dim_t id = od * sd - fp + kd * dd; |
| 376 | if (id < 0 || id >= jcp.id) { |
| 377 | for (ptrdiff_t i = 0; i < OHW; i++) |
| 378 | col_loc[i] = shift; |
| 379 | return; |
| 380 | } |
| 381 | const im_dt *__restrict imtr_loc |
| 382 | = imtr + (ic * jcp.id + id) * IHW; |
| 383 | const dim_t oh_start = saturate( |
| 384 | dim_t(0), jcp.oh, div_up(tp - kh * dh, sh)); |
| 385 | const dim_t oh_end = saturate(dim_t(0), jcp.oh, |
| 386 | div_up(jcp.ih + tp - kh * dh, sh)); |
| 387 | const dim_t ow_start = saturate( |
| 388 | dim_t(0), jcp.ow, div_up(lp - kw * dw, sw)); |
| 389 | const dim_t ow_end = saturate(dim_t(0), jcp.ow, |
| 390 | div_up(jcp.iw + lp - kw * dw, sw)); |
| 391 | for (dim_t oh = oh_start, ih = oh_start * sh - tp + kh * dh; |
| 392 | oh < oh_end; ++oh, ih += sh) { |
| 393 | col_dt *__restrict col_h = col_loc + oh * jcp.ow; |
| 394 | const im_dt *__restrict imtr_h = imtr_loc + ih * jcp.iw; |
| 395 | for (dim_t ow = ow_start, |
| 396 | iw = ow_start * sw - lp + kw * dw; |
| 397 | ow < ow_end; ++ow, iw += sw) { |
| 398 | col_h[ow] = imtr_h[iw]; |
| 399 | } |
| 400 | } |
| 401 | }); |
| 402 | } |
| 403 | |
| 404 | template void im2col_dt_3d<int8_t, uint8_t>(const conv_gemm_conf_t &jcp, |
| 405 | const void *__restrict im, uint8_t *__restrict col, dim_t od); |
| 406 | template void im2col_dt_3d<uint8_t, uint8_t>(const conv_gemm_conf_t &jcp, |
| 407 | const void *__restrict im, uint8_t *__restrict col, dim_t od); |
| 408 | template void im2col_dt_3d<float, float>(const conv_gemm_conf_t &jcp, |
| 409 | const void *__restrict im, float *__restrict col, dim_t od); |
| 410 | template void im2col_dt_3d<bfloat16_t, bfloat16_t>(const conv_gemm_conf_t &jcp, |
| 411 | const void *__restrict im, bfloat16_t *__restrict col, dim_t od); |
| 412 | |
| 413 | /* col[ic][kh][kw][oh][ow] <-- im2col(im[ic][ih][iw]) */ |
| 414 | template <typename data_type_t> |
| 415 | void im2col(const conv_gemm_conf_t &jcp, const data_type_t *__restrict im, |
| 416 | data_type_t *__restrict col, dim_t ss, dim_t sb, dim_t cs, dim_t cb) { |
| 417 | |
| 418 | using data_t = |
| 419 | typename utils::conditional<data_traits<data_type_t>::data_type |
| 420 | == bf16, |
| 421 | uint16_t, data_type_t>::type; |
| 422 | const data_t *__restrict _im |
| 423 | = reinterpret_cast<const data_t *__restrict>(im); |
| 424 | data_t *__restrict _col = reinterpret_cast<data_t *__restrict>(col); |
| 425 | |
| 426 | const size_t im_step = jcp.is; |
| 427 | const size_t col_step = jcp.ks * sb; |
| 428 | const dim_t dh = 1 + jcp.dilate_h; |
| 429 | const dim_t dw = 1 + jcp.dilate_w; |
| 430 | const dim_t sh = jcp.stride_h; |
| 431 | const dim_t sw = jcp.stride_w; |
| 432 | const dim_t tp = jcp.t_pad; |
| 433 | const dim_t lp = jcp.l_pad; |
| 434 | const dim_t first_oh = ss / jcp.ow; |
| 435 | const dim_t last_oh = (ss + sb - 1) / jcp.ow; |
| 436 | const dim_t oh_begin = first_oh; |
| 437 | const dim_t oh_end = last_oh + 1; |
| 438 | const dim_t first_ow = ss % jcp.ow; |
| 439 | const dim_t last_ow = (ss + sb - 1) % jcp.ow; |
| 440 | |
| 441 | const data_t zero_val = 0; |
| 442 | |
| 443 | if (jcp.outer_threading) { |
| 444 | if (sw == 1) { |
| 445 | // Generated code is more optimized for stride_w == 1 |
| 446 | // because innermost loop is by width |
| 447 | for (dim_t ic = 0; ic < cb; ic++) { |
| 448 | const data_t *__restrict im_ic = _im + (ic + cs) * im_step; |
| 449 | for (dim_t kh = 0; kh < jcp.kh; kh++) { |
| 450 | for (dim_t kw = 0; kw < jcp.kw; kw++) { |
| 451 | data_t *__restrict col_k = _col + ic * col_step |
| 452 | + (kh * jcp.kw + kw) * sb; |
| 453 | for (dim_t oh = oh_begin; oh < oh_end; oh++) { |
| 454 | const dim_t ih = oh * sh - tp + kh * dh; |
| 455 | const data_t *__restrict im_ |
| 456 | = im_ic + ih * jcp.iw - lp + kw * dw; |
| 457 | const dim_t ow_begin |
| 458 | = (oh == first_oh) ? first_ow : 0; |
| 459 | const dim_t ow_end |
| 460 | = (oh == last_oh) ? (last_ow + 1) : jcp.ow; |
| 461 | data_t *__restrict col_ = col_k + oh * jcp.ow - ss; |
| 462 | if (ih < 0 || ih >= jcp.ih) |
| 463 | for (dim_t ow = ow_begin; ow < ow_end; ow++) |
| 464 | col_[ow] = zero_val; |
| 465 | else { |
| 466 | for (dim_t ow = ow_begin; ow < ow_end; ++ow) { |
| 467 | const dim_t iw = ow; |
| 468 | if (iw < lp - kw * dw |
| 469 | || iw >= jcp.iw + lp - kw * dw) |
| 470 | col_[ow] = zero_val; |
| 471 | else |
| 472 | col_[ow] = im_[iw]; |
| 473 | } |
| 474 | } |
| 475 | } |
| 476 | } |
| 477 | } |
| 478 | } |
| 479 | } else { |
| 480 | for (dim_t ic = 0; ic < cb; ic++) { |
| 481 | const data_t *__restrict im_ = _im + (ic + cs) * im_step; |
| 482 | for (dim_t kh = 0; kh < jcp.kh; kh++) { |
| 483 | for (dim_t kw = 0; kw < jcp.kw; kw++) { |
| 484 | data_t *__restrict col_k = _col + ic * col_step |
| 485 | + (kh * jcp.kw + kw) * sb; |
| 486 | for (dim_t oh = oh_begin; oh < oh_end; oh++) { |
| 487 | const dim_t ih = oh * sh - tp + kh * dh; |
| 488 | const dim_t ow_begin |
| 489 | = (oh == first_oh) ? first_ow : 0; |
| 490 | const dim_t ow_end |
| 491 | = (oh == last_oh) ? (last_ow + 1) : jcp.ow; |
| 492 | data_t *__restrict col_oh |
| 493 | = col_k + oh * jcp.ow - ss; |
| 494 | if (ih < 0 || ih >= jcp.ih) |
| 495 | for (dim_t ow = ow_begin; ow < ow_end; ow++) |
| 496 | col_oh[ow] = zero_val; |
| 497 | else |
| 498 | for (dim_t ow = ow_begin; ow < ow_end; ow++) { |
| 499 | const dim_t iw = ow * sw - lp + kw * dw; |
| 500 | if (iw < 0 || iw >= jcp.iw) |
| 501 | col_oh[ow] = zero_val; |
| 502 | else { |
| 503 | const ptrdiff_t im_idx |
| 504 | = ih * jcp.iw + iw; |
| 505 | col_oh[ow] = im_[im_idx]; |
| 506 | } |
| 507 | } |
| 508 | } |
| 509 | } |
| 510 | } |
| 511 | } |
| 512 | } |
| 513 | } else { |
| 514 | // TODO: optimize threading if jcp.ic*jcp.kh*jcp.kw*oh_range is small |
| 515 | // comparing to number of threads |
| 516 | const dim_t oh_range = oh_end - oh_begin; |
| 517 | // Generated code is more optimized for stride_w == 1 |
| 518 | // because innermost loop is by width |
| 519 | if (sw == 1) |
| 520 | parallel_nd(cb, jcp.kh, jcp.kw, oh_range, |
| 521 | [&](dim_t ic, dim_t kh, dim_t kw, dim_t ohr) { |
| 522 | const dim_t oh = ohr + oh_begin; |
| 523 | const dim_t ih = oh * sh - tp + kh * dh; |
| 524 | const dim_t ow_start = (oh == first_oh) ? first_ow : 0; |
| 525 | const dim_t ow_end |
| 526 | = (oh == last_oh) ? (last_ow + 1) : jcp.ow; |
| 527 | data_t *__restrict col_oh = _col + ic * col_step |
| 528 | + (kh * jcp.kw + kw) * sb + oh * jcp.ow - ss; |
| 529 | const data_t *__restrict im_ |
| 530 | = _im + (ic + cs) * im_step + ih * jcp.iw; |
| 531 | const dim_t iw_shift = kw * dw - lp; |
| 532 | if (ih < 0 || ih >= jcp.ih) |
| 533 | for (dim_t ow = ow_start; ow < ow_end; ow++) |
| 534 | col_oh[ow] = zero_val; |
| 535 | else |
| 536 | for (dim_t ow = ow_start; ow < ow_end; ow++) { |
| 537 | const dim_t iw = ow + iw_shift; |
| 538 | if (iw < 0 || iw >= jcp.iw) |
| 539 | col_oh[ow] = zero_val; |
| 540 | else |
| 541 | col_oh[ow] = im_[iw]; |
| 542 | } |
| 543 | }); |
| 544 | else |
| 545 | parallel_nd(cb, jcp.kh, jcp.kw, oh_range, |
| 546 | [&](dim_t ic, dim_t kh, dim_t kw, dim_t ohr) { |
| 547 | const dim_t oh = ohr + oh_begin; |
| 548 | const dim_t ih = oh * sh - tp + kh * dh; |
| 549 | const dim_t ow_start = (oh == first_oh) ? first_ow : 0; |
| 550 | const dim_t ow_end |
| 551 | = (oh == last_oh) ? (last_ow + 1) : jcp.ow; |
| 552 | data_t *__restrict col_oh = _col + ic * col_step |
| 553 | + (kh * jcp.kw + kw) * sb + oh * jcp.ow - ss; |
| 554 | const data_t *__restrict im_ |
| 555 | = _im + (ic + cs) * im_step; |
| 556 | if (ih < 0 || ih >= jcp.ih) |
| 557 | for (dim_t ow = ow_start; ow < ow_end; ow++) |
| 558 | col_oh[ow] = zero_val; |
| 559 | else |
| 560 | for (dim_t ow = ow_start; ow < ow_end; ow++) { |
| 561 | const dim_t iw = ow * sw - lp + kw * dw; |
| 562 | if (iw < 0 || iw >= jcp.iw) |
| 563 | col_oh[ow] = zero_val; |
| 564 | else { |
| 565 | const ptrdiff_t im_idx = ih * jcp.iw + iw; |
| 566 | col_oh[ow] = im_[im_idx]; |
| 567 | } |
| 568 | } |
| 569 | }); |
| 570 | } |
| 571 | } |
| 572 | |
| 573 | template void im2col(const conv_gemm_conf_t &jcp, const float *__restrict im, |
| 574 | float *__restrict col, dim_t hs, dim_t hb, dim_t ws, dim_t wb); |
| 575 | |
| 576 | template void im2col(const conv_gemm_conf_t &jcp, |
| 577 | const bfloat16_t *__restrict im, bfloat16_t *__restrict col, dim_t hs, |
| 578 | dim_t hb, dim_t ws, dim_t wb); |
| 579 | |
| 580 | /* col[kh][kw][ic][oh][ow] <-- im2col_dt(im[ih][iw][ic]) */ |
| 581 | template <typename orig_im_dt, typename orig_col_dt> |
| 582 | void im2col_dt(const conv_gemm_conf_t &jcp, const void *__restrict _im, |
| 583 | void *__restrict _imtr, orig_col_dt *__restrict _col, dim_t hs, |
| 584 | dim_t hb, dim_t ws, dim_t wb) { |
| 585 | // For performance reasons, use uint16_t as a proxy for bfloat16_t |
| 586 | using im_dt = typename utils::conditional<data_traits<orig_im_dt>::data_type |
| 587 | == bf16, |
| 588 | uint16_t, orig_im_dt>::type; |
| 589 | using col_dt = |
| 590 | typename utils::conditional<data_traits<orig_col_dt>::data_type |
| 591 | == bf16, |
| 592 | uint16_t, orig_col_dt>::type; |
| 593 | const im_dt *__restrict im = reinterpret_cast<const im_dt *__restrict>(_im); |
| 594 | im_dt *__restrict imtr = reinterpret_cast<im_dt *__restrict>(_imtr); |
| 595 | col_dt *__restrict col = reinterpret_cast<col_dt *__restrict>(_col); |
| 596 | |
| 597 | col_dt shift = static_cast<col_dt>(jcp.signed_input ? 128 : 0); |
| 598 | const dim_t dh = 1 + jcp.dilate_h; |
| 599 | const dim_t dw = 1 + jcp.dilate_w; |
| 600 | const dim_t sh = jcp.stride_h; |
| 601 | const dim_t sw = jcp.stride_w; |
| 602 | const dim_t im_iw_stride = jcp.ic * jcp.ngroups; |
| 603 | const dim_t im_ih_stride = jcp.iw * im_iw_stride; |
| 604 | const dim_t tp = jcp.t_pad; |
| 605 | const dim_t lp = jcp.l_pad; |
| 606 | |
| 607 | if (jcp.outer_threading && sh == 1 && sw == 1 && dh == 1 && dw == 1) { |
| 608 | /* im[ih][iw][ic] --> imtr[ic][ih][iw] --> col[kh][kw][ic][oh][ow] */ |
| 609 | const dim_t hp = hs - tp; |
| 610 | const dim_t wp = ws - lp; |
| 611 | const dim_t ih_start = saturate(dim_t(0), jcp.ih, hp); |
| 612 | const dim_t ih_end = saturate(dim_t(0), jcp.ih, hp + hb + jcp.kh); |
| 613 | const dim_t iw_start = saturate(dim_t(0), jcp.iw, wp); |
| 614 | const dim_t iw_end = saturate(dim_t(0), jcp.iw, wp + wb + jcp.kw); |
| 615 | |
| 616 | const dim_t ihb = ih_end - ih_start; |
| 617 | const dim_t iwb = iw_end - iw_start; |
| 618 | |
| 619 | const dim_t imtr_ic_stride = ihb * iwb; |
| 620 | const ptrdiff_t imtr_idx_shift = ih_start * iwb + iw_start; |
| 621 | for (dim_t ic = 0; ic < jcp.ic; ic++) { |
| 622 | const ptrdiff_t imtr_idx_ic = ic * imtr_ic_stride - imtr_idx_shift; |
| 623 | for (dim_t ih = ih_start; ih < ih_end; ih++) { |
| 624 | const ptrdiff_t im_idx_ih = ic + ih * im_ih_stride; |
| 625 | const ptrdiff_t imtr_idx_ih = imtr_idx_ic + ih * iwb; |
| 626 | for (dim_t iw = iw_start; iw < iw_end; iw++) |
| 627 | imtr[imtr_idx_ih + iw] = im[im_idx_ih + iw * im_iw_stride]; |
| 628 | } |
| 629 | } |
| 630 | |
| 631 | const dim_t col_ic_str = hb * wb; |
| 632 | const dim_t col_kw_stride = jcp.ic * col_ic_str; |
| 633 | const dim_t col_kh_stride = jcp.kw * col_kw_stride; |
| 634 | |
| 635 | const dim_t oh_init = ih_start - hp; |
| 636 | const dim_t ow_init = iw_start - wp; |
| 637 | for (dim_t kh = 0; kh < jcp.kh; kh++) { |
| 638 | const ptrdiff_t col_idx_kh = kh * col_kh_stride; |
| 639 | const dim_t oh_kh = oh_init - kh; |
| 640 | const dim_t oh_start = saturate(dim_t(0), hb, oh_kh); |
| 641 | const dim_t oh_end = saturate(dim_t(0), hb, oh_kh + ihb); |
| 642 | for (dim_t kw = 0; kw < jcp.kw; kw++) { |
| 643 | const ptrdiff_t col_idx_kw |
| 644 | = col_idx_kh + kw * jcp.ic * col_ic_str; |
| 645 | const dim_t ow_kw = ow_init - kw; |
| 646 | const dim_t imtr_shift = oh_kh * iwb + ow_kw; |
| 647 | const dim_t ow_start = saturate(dim_t(0), wb, ow_kw); |
| 648 | const dim_t ow_end = saturate(dim_t(0), wb, ow_kw + iwb); |
| 649 | for (dim_t ic = 0; ic < jcp.ic; ic++) { |
| 650 | const ptrdiff_t col_idx_ic = col_idx_kw + ic * col_ic_str; |
| 651 | const dim_t imtr_idx_ic = ic * imtr_ic_stride - imtr_shift; |
| 652 | for (dim_t oh = 0; oh < oh_start; oh++) { |
| 653 | const ptrdiff_t col_idx_oh = col_idx_ic + oh * wb; |
| 654 | for (dim_t ow = 0; ow < wb; ++ow) |
| 655 | col[col_idx_oh + ow] = shift; |
| 656 | } |
| 657 | for (dim_t oh = oh_start; oh < oh_end; oh++) { |
| 658 | const ptrdiff_t col_idx_oh = col_idx_ic + oh * wb; |
| 659 | const ptrdiff_t imtr_idx_oh = imtr_idx_ic + oh * iwb; |
| 660 | for (dim_t ow = 0; ow < ow_start; ++ow) |
| 661 | col[col_idx_oh + ow] = shift; |
| 662 | for (dim_t ow = ow_start; ow < ow_end; ++ow) |
| 663 | col[col_idx_oh + ow] |
| 664 | = imtr[imtr_idx_oh + ow] + shift; |
| 665 | for (dim_t ow = ow_end; ow < wb; ++ow) |
| 666 | col[col_idx_oh + ow] = shift; |
| 667 | } |
| 668 | for (dim_t oh = oh_end; oh < hb; oh++) { |
| 669 | const ptrdiff_t col_idx_oh = col_idx_ic + oh * wb; |
| 670 | for (dim_t ow = 0; ow < wb; ++ow) |
| 671 | col[col_idx_oh + ow] = shift; |
| 672 | } |
| 673 | } |
| 674 | } |
| 675 | } |
| 676 | } else { |
| 677 | parallel_nd(jcp.kh, jcp.kw, jcp.ic, hb, |
| 678 | [&](dim_t kh, dim_t kw, dim_t ic, dim_t oh) { |
| 679 | const dim_t hp = tp - kh * dh; |
| 680 | const dim_t ih = (oh + hs) * sh - hp; |
| 681 | const ptrdiff_t col_idx_base |
| 682 | = (((kh * jcp.kw + kw) * jcp.ic + ic) * hb + oh) |
| 683 | * wb; |
| 684 | if (ih < 0 || ih >= jcp.ih) |
| 685 | for (dim_t ow = 0; ow < wb; ow++) |
| 686 | col[col_idx_base + ow] = shift; |
| 687 | else { |
| 688 | const dim_t wp = lp - kw * dw; |
| 689 | const dim_t ow_start |
| 690 | = saturate(dim_t(0), wb, div_up(wp, sw) - ws); |
| 691 | const dim_t ow_end = saturate( |
| 692 | dim_t(0), wb, div_up(jcp.iw + wp, sw) - ws); |
| 693 | for (dim_t ow = 0; ow < ow_start; ow++) |
| 694 | col[col_idx_base + ow] = shift; |
| 695 | const dim_t iw_base = ws * sw - wp; |
| 696 | const ptrdiff_t im_idx_base = ih * im_ih_stride + ic; |
| 697 | for (dim_t ow = ow_start; ow < ow_end; ow++) { |
| 698 | const dim_t iw = iw_base + ow * sw; |
| 699 | const ptrdiff_t im_idx |
| 700 | = im_idx_base + iw * im_iw_stride; |
| 701 | col[col_idx_base + ow] = im[im_idx] + shift; |
| 702 | } |
| 703 | for (dim_t ow = ow_end; ow < wb; ow++) |
| 704 | col[col_idx_base + ow] = shift; |
| 705 | } |
| 706 | }); |
| 707 | } |
| 708 | } |
| 709 | |
| 710 | template void im2col_dt<int8_t, uint8_t>(const conv_gemm_conf_t &jcp, |
| 711 | const void *__restrict im, void *__restrict imtr, |
| 712 | uint8_t *__restrict col, dim_t hs, dim_t hb, dim_t ws, dim_t wb); |
| 713 | template void im2col_dt<uint8_t, uint8_t>(const conv_gemm_conf_t &jcp, |
| 714 | const void *__restrict im, void *__restrict imtr, |
| 715 | uint8_t *__restrict col, dim_t hs, dim_t hb, dim_t ws, dim_t wb); |
| 716 | template void im2col_dt<float, float>(const conv_gemm_conf_t &jcp, |
| 717 | const void *__restrict im, void *__restrict imtr, float *__restrict col, |
| 718 | dim_t hs, dim_t hb, dim_t ws, dim_t wb); |
| 719 | |
| 720 | template void im2col_dt<bfloat16_t, bfloat16_t>(const conv_gemm_conf_t &jcp, |
| 721 | const void *__restrict im, void *__restrict imtr, |
| 722 | bfloat16_t *__restrict col, dim_t hs, dim_t hb, dim_t ws, dim_t wb); |
| 723 | |
| 724 | /* im[id][ih][iw][ic] <-- col2im_dt_3d(col[od][oh][ow][kd][kh][kw][ic]) */ |
| 725 | template <typename orig_T> |
| 726 | void col2im_dt(const conv_gemm_conf_t &jcp, const orig_T *__restrict _col, |
| 727 | orig_T *__restrict _im) { |
| 728 | // For performance reasons, use uint16_t as a proxy for bfloat16_t |
| 729 | using T = |
| 730 | typename utils::conditional<data_traits<orig_T>::data_type == bf16, |
| 731 | uint16_t, orig_T>::type; |
| 732 | const T *__restrict col = reinterpret_cast<const T *__restrict>(_col); |
| 733 | T *__restrict im = reinterpret_cast<T *__restrict>(_im); |
| 734 | |
| 735 | parallel(0, [&](const int ithr, const int nthr) { |
| 736 | dim_t d_nthr = nstl::min(jcp.id, dim_t(nthr)); |
| 737 | dim_t h_nthr = nstl::min(jcp.ih, dim_t(nthr) / d_nthr); |
| 738 | dim_t w_nthr = nstl::min(jcp.iw, dim_t(nthr) / (d_nthr * h_nthr)); |
| 739 | dim_t d_ithr = 1, d_s = 0, d_e = 0, h_ithr = 1, h_s = 0, h_e = 0, |
| 740 | w_ithr = 1, w_s = 0, w_e = 0; |
| 741 | if (ithr < d_nthr * h_nthr * w_nthr) { |
| 742 | d_ithr = ithr / (h_nthr * w_nthr); |
| 743 | h_ithr = (ithr % (h_nthr * w_nthr)) / w_nthr; |
| 744 | w_ithr = (ithr % (h_nthr * w_nthr)) % w_nthr; |
| 745 | balance211(jcp.id, d_nthr, d_ithr, d_s, d_e); |
| 746 | balance211(jcp.ih, h_nthr, h_ithr, h_s, h_e); |
| 747 | balance211(jcp.iw, w_nthr, w_ithr, w_s, w_e); |
| 748 | } else { |
| 749 | d_nthr = h_ithr = w_ithr = -ithr; |
| 750 | d_s = d_e = h_s = h_e = w_s = w_e = -1; |
| 751 | } |
| 752 | |
| 753 | for_(dim_t id = d_s; id < d_e; ++id) |
| 754 | for_(dim_t ih = h_s; ih < h_e; ++ih) |
| 755 | for (dim_t iw = w_s; iw < w_e; ++iw) { |
| 756 | PRAGMA_OMP_SIMD() |
| 757 | for (dim_t ic = 0; ic < jcp.ic; ++ic) { |
| 758 | im[((id * jcp.ih + ih) * jcp.iw + iw) * jcp.ic + ic] = 0; |
| 759 | } |
| 760 | } |
| 761 | |
| 762 | // TODO: reduce region: [0.. oh] --> [h_s * sh .. h_e * sh] |
| 763 | for_(dim_t od = 0; od < jcp.od; ++od) |
| 764 | for_(dim_t oh = 0; oh < jcp.oh; ++oh) |
| 765 | for_(dim_t ow = 0; ow < jcp.ow; ++ow) |
| 766 | for (dim_t kd = 0; kd < jcp.kd; ++kd) { |
| 767 | const dim_t id |
| 768 | = od * jcp.stride_d - jcp.f_pad + kd * (1 + jcp.dilate_d); |
| 769 | if (id < d_s || id >= d_e) continue; |
| 770 | |
| 771 | for (dim_t kh = 0; kh < jcp.kh; ++kh) { |
| 772 | const dim_t ih = oh * jcp.stride_h - jcp.t_pad |
| 773 | + kh * (1 + jcp.dilate_h); |
| 774 | if (ih < h_s || ih >= h_e) continue; |
| 775 | |
| 776 | for (dim_t kw = 0; kw < jcp.kw; ++kw) { |
| 777 | const dim_t iw = ow * jcp.stride_w - jcp.l_pad |
| 778 | + kw * (1 + jcp.dilate_w); |
| 779 | if (iw < w_s || iw >= w_e) continue; |
| 780 | |
| 781 | const size_t col_idx |
| 782 | = (((((od * jcp.oh + oh) * jcp.ow + ow) * jcp.kd |
| 783 | + kd) * jcp.kh |
| 784 | + kh) * jcp.kw |
| 785 | + kw) |
| 786 | * jcp.ic; |
| 787 | const size_t im_idx |
| 788 | = ((id * jcp.ih + ih) * jcp.iw + iw) * jcp.ic; |
| 789 | PRAGMA_OMP_SIMD() |
| 790 | for (dim_t ic = 0; ic < jcp.ic; ++ic) { |
| 791 | im[im_idx + ic] += col[col_idx + ic]; |
| 792 | } |
| 793 | } |
| 794 | } |
| 795 | } |
| 796 | }); |
| 797 | } |
| 798 | |
| 799 | template void col2im_dt<int32_t>(const conv_gemm_conf_t &jcp, |
| 800 | const int32_t *__restrict col, int32_t *__restrict im); |
| 801 | |
| 802 | template void col2im_dt<float>(const conv_gemm_conf_t &jcp, |
| 803 | const float *__restrict col, float *__restrict im); |
| 804 | |
| 805 | template void col2im_dt<bfloat16_t>(const conv_gemm_conf_t &jcp, |
| 806 | const bfloat16_t *__restrict col, bfloat16_t *__restrict im); |
| 807 | |
| 808 | void col2im_3d(const conv_gemm_conf_t &jcp, const float *col, float *im, |
| 809 | dim_t od, int spatial_step, int spatial_block) { |
| 810 | |
| 811 | auto sp_blocked_ker = [&](dim_t ic) { |
| 812 | const size_t col_step = jcp.ks * spatial_block; |
| 813 | const float *__restrict col_ = col + ic * col_step; |
| 814 | float *__restrict im_ic = im + ic * jcp.ih * jcp.iw * jcp.id; |
| 815 | |
| 816 | const dim_t first_oh = spatial_step / jcp.ow; |
| 817 | const dim_t last_oh = (spatial_step + spatial_block - 1) / jcp.ow; |
| 818 | const dim_t oh_begin = first_oh; |
| 819 | const dim_t oh_end = last_oh + 1; |
| 820 | const dim_t first_ow = spatial_step % jcp.ow; |
| 821 | const dim_t last_ow = (spatial_step + spatial_block - 1) % jcp.ow; |
| 822 | const dim_t wei_stride |
| 823 | = nstl::min(jcp.ow * jcp.oh, dim_t(spatial_block)); |
| 824 | |
| 825 | dim_t id = od * jcp.stride_d - jcp.f_pad; |
| 826 | for (dim_t kd = 0; kd < jcp.kd; ++kd) { |
| 827 | if (id < 0 || id >= jcp.id) { |
| 828 | col_ += jcp.kh * jcp.kw * wei_stride; |
| 829 | id += (1 + jcp.dilate_d); |
| 830 | continue; |
| 831 | } |
| 832 | |
| 833 | float *__restrict im_ = im_ic + (size_t)id * jcp.ih * jcp.iw; |
| 834 | for_(dim_t kh = 0; kh < jcp.kh; ++kh) |
| 835 | for_(dim_t kw = 0; kw < jcp.kw; ++kw) |
| 836 | for (dim_t oh = oh_begin, col_off = 0; oh < oh_end; ++oh) { |
| 837 | |
| 838 | const dim_t ow_begin = (oh == first_oh) ? first_ow : 0; |
| 839 | const dim_t ow_end = (oh == last_oh) ? (last_ow + 1) : jcp.ow; |
| 840 | const dim_t ow_work = ow_end - ow_begin; |
| 841 | |
| 842 | const dim_t ih = oh * jcp.stride_h - jcp.t_pad |
| 843 | + kh * (1 + jcp.dilate_h); |
| 844 | if (ih < 0 || ih >= jcp.ih) { |
| 845 | col_off += ow_work; |
| 846 | continue; |
| 847 | } |
| 848 | |
| 849 | for (dim_t ow = ow_begin; ow < ow_end; ++ow, ++col_off) { |
| 850 | const dim_t iw = ow * jcp.stride_w - jcp.l_pad |
| 851 | + kw * (1 + jcp.dilate_w); |
| 852 | if (iw < 0 || iw >= jcp.iw) { continue; } |
| 853 | |
| 854 | const size_t col_idx |
| 855 | = (kh * jcp.kw + kw) * wei_stride + col_off; |
| 856 | const size_t im_idx = ih * jcp.iw + iw; |
| 857 | im_[im_idx] += col_[col_idx]; |
| 858 | } |
| 859 | } |
| 860 | col_ += jcp.kh * jcp.kw * wei_stride; |
| 861 | id += (1 + jcp.dilate_d); |
| 862 | } |
| 863 | }; |
| 864 | |
| 865 | auto ker = [&](dim_t ic) { |
| 866 | const float *__restrict col_ = col + (size_t)ic * jcp.ks * jcp.os; |
| 867 | float *__restrict im_ic = im + (size_t)ic * jcp.ih * jcp.iw * jcp.id; |
| 868 | |
| 869 | dim_t id = od * jcp.stride_d - jcp.f_pad; |
| 870 | for (dim_t kd = 0; kd < jcp.kd; ++kd) { |
| 871 | if (id < 0 || id >= jcp.id) { |
| 872 | col_ += jcp.kh * jcp.kw * jcp.os; |
| 873 | id += (1 + jcp.dilate_d); |
| 874 | continue; |
| 875 | } |
| 876 | |
| 877 | float *__restrict im_ = im_ic + (size_t)id * jcp.ih * jcp.iw; |
| 878 | |
| 879 | for_(dim_t oh = 0; oh < jcp.oh; ++oh) |
| 880 | for (dim_t kh = 0; kh < jcp.kh; ++kh) { |
| 881 | const dim_t ih = oh * jcp.stride_h - jcp.t_pad |
| 882 | + kh * (1 + jcp.dilate_h); |
| 883 | if (ih < 0 || ih >= jcp.ih) continue; |
| 884 | |
| 885 | for_(dim_t ow = 0; ow < jcp.ow; ++ow) |
| 886 | for (dim_t kw = 0; kw < jcp.kw; ++kw) { |
| 887 | const dim_t iw = ow * jcp.stride_w - jcp.l_pad |
| 888 | + kw * (1 + jcp.dilate_w); |
| 889 | if (iw < 0 || iw >= jcp.iw) continue; |
| 890 | |
| 891 | const size_t col_idx |
| 892 | = ((kh * jcp.kw + kw) * jcp.oh + oh) * jcp.ow + ow; |
| 893 | const size_t im_idx = ih * jcp.iw + iw; |
| 894 | im_[im_idx] += col_[col_idx]; |
| 895 | } |
| 896 | } |
| 897 | |
| 898 | col_ += jcp.kh * jcp.kw * jcp.os; |
| 899 | id += (1 + jcp.dilate_d); |
| 900 | } |
| 901 | }; |
| 902 | |
| 903 | const bool blocked_kernel = jcp.os_nb_block > 1; |
| 904 | if (blocked_kernel) |
| 905 | parallel_nd(jcp.ic, sp_blocked_ker); |
| 906 | else |
| 907 | parallel_nd(jcp.ic, ker); |
| 908 | } |
| 909 | |
| 910 | void col2im(const conv_gemm_conf_t &jcp, const float *col, float *im, |
| 911 | int spatial_step, int spatial_block) { |
| 912 | const size_t col_step = jcp.ks * spatial_block; |
| 913 | const size_t im_step = jcp.ih * jcp.iw; |
| 914 | const dim_t iS = jcp.ih * jcp.iw; |
| 915 | |
| 916 | auto sp_blocked_ker = [&](dim_t ic) { |
| 917 | const dim_t wei_stride |
| 918 | = nstl::min(jcp.ow * jcp.oh, dim_t(spatial_block)); |
| 919 | const dim_t first_oh = spatial_step / jcp.ow; |
| 920 | const dim_t last_oh = (spatial_step + spatial_block - 1) / jcp.ow; |
| 921 | const dim_t oh_begin = first_oh; |
| 922 | const dim_t oh_end = last_oh + 1; |
| 923 | const dim_t first_ow = spatial_step % jcp.ow; |
| 924 | const dim_t last_ow = (spatial_step + spatial_block - 1) % jcp.ow; |
| 925 | |
| 926 | float *__restrict img_ithr = im + ic * im_step; |
| 927 | const float *__restrict col_icb = col + ic * col_step; |
| 928 | |
| 929 | if (spatial_step == 0) { |
| 930 | PRAGMA_OMP_SIMD() |
| 931 | for (dim_t is = 0; is < iS; ++is) |
| 932 | img_ithr[is] = 0.; |
| 933 | } |
| 934 | |
| 935 | float *__restrict img_kh = img_ithr; |
| 936 | for (dim_t kh = 0; kh < jcp.kh; ++kh) { |
| 937 | float *__restrict im_ = img_kh; |
| 938 | for (dim_t kw = 0; kw < jcp.kw; ++kw) { |
| 939 | const float *__restrict col_ = col_icb; |
| 940 | for (dim_t oh = oh_begin; oh < oh_end; ++oh) { |
| 941 | const dim_t ow_begin = (oh == first_oh) ? first_ow : 0; |
| 942 | const dim_t ow_end |
| 943 | = (oh == last_oh) ? (last_ow + 1) : jcp.ow; |
| 944 | const dim_t ow_work = ow_end - ow_begin; |
| 945 | |
| 946 | const dim_t ih = oh * jcp.stride_h - jcp.t_pad; |
| 947 | const dim_t ih_ = ih + kh * (1 + jcp.dilate_h); |
| 948 | if (ih_ < 0 || ih_ >= jcp.ih) { |
| 949 | col_ += ow_work; |
| 950 | continue; |
| 951 | } |
| 952 | for (dim_t ow = ow_begin; ow < ow_end; ++ow, ++col_) { |
| 953 | const dim_t iw = ow * jcp.stride_w - jcp.l_pad; |
| 954 | const dim_t iw_ = iw + kw * (1 + jcp.dilate_w); |
| 955 | if (iw_ < 0 || iw_ >= jcp.iw) continue; |
| 956 | |
| 957 | const size_t im_idx = ih * jcp.iw + iw; |
| 958 | im_[im_idx] += *col_; |
| 959 | } |
| 960 | } |
| 961 | col_icb += wei_stride; |
| 962 | im_ += (1 + jcp.dilate_w); |
| 963 | } |
| 964 | img_kh += (jcp.iw * (1 + jcp.dilate_h)); |
| 965 | } |
| 966 | }; |
| 967 | |
| 968 | auto ker = [&](dim_t ic) { |
| 969 | float *__restrict im_ = im + ic * im_step; |
| 970 | const float *__restrict col_ = col + ic * col_step; |
| 971 | PRAGMA_OMP_SIMD() |
| 972 | for (dim_t is = 0; is < iS; ++is) |
| 973 | im_[is] = 0.; |
| 974 | |
| 975 | for_(dim_t kh = 0; kh < jcp.kh; ++kh) |
| 976 | for (dim_t oh = 0; oh < jcp.oh; ++oh) { |
| 977 | const dim_t ih |
| 978 | = oh * jcp.stride_h - jcp.t_pad + kh * (1 + jcp.dilate_h); |
| 979 | if (ih < 0 || ih >= jcp.ih) continue; |
| 980 | |
| 981 | for_(dim_t kw = 0; kw < jcp.kw; ++kw) |
| 982 | for (dim_t ow = 0; ow < jcp.ow; ++ow) { |
| 983 | const dim_t iw = ow * jcp.stride_w - jcp.l_pad |
| 984 | + kw * (1 + jcp.dilate_w); |
| 985 | if (iw < 0 || iw >= jcp.iw) continue; |
| 986 | |
| 987 | const size_t col_idx |
| 988 | = ((kh * jcp.kw + kw) * jcp.oh + oh) * jcp.ow + ow; |
| 989 | const size_t im_idx = ih * jcp.iw + iw; |
| 990 | im_[im_idx] += col_[col_idx]; |
| 991 | } |
| 992 | } |
| 993 | }; |
| 994 | |
| 995 | const bool blocked_kernel = jcp.os_nb_block > 1; |
| 996 | if (blocked_kernel) |
| 997 | parallel_nd(jcp.ic, sp_blocked_ker); |
| 998 | else |
| 999 | parallel_nd(jcp.ic, ker); |
| 1000 | } |
| 1001 | |
| 1002 | status_t init_conf(conv_gemm_conf_t &jcp, |
| 1003 | memory_tracking::registrar_t &scratchpad, const convolution_desc_t &cd, |
| 1004 | memory_desc_t &src_md, memory_desc_t &weights_md, memory_desc_t &dst_md, |
| 1005 | memory_desc_t &bias_md, primitive_attr_t &attr, int max_threads) { |
| 1006 | const memory_desc_wrapper src_d(&src_md); |
| 1007 | const memory_desc_wrapper weights_d(&weights_md); |
| 1008 | const memory_desc_wrapper dst_d(&dst_md); |
| 1009 | |
| 1010 | const bool with_groups = weights_d.ndims() == src_d.ndims() + 1; |
| 1011 | const int ndims = src_d.ndims(); |
| 1012 | const int is_1d = ndims == 3; |
| 1013 | const int is_3d = ndims == 5; |
| 1014 | |
| 1015 | jcp.prop_kind = cd.prop_kind; |
| 1016 | |
| 1017 | jcp.ngroups = with_groups ? weights_d.dims()[0] : 1; |
| 1018 | jcp.mb = src_d.dims()[0]; |
| 1019 | |
| 1020 | jcp.oc = dst_d.dims()[1] / jcp.ngroups; |
| 1021 | jcp.ic = src_d.dims()[1] / jcp.ngroups; |
| 1022 | jcp.id = is_3d ? src_d.dims()[2] : 1; |
| 1023 | jcp.ih = is_1d ? 1 : src_d.dims()[ndims - 2]; |
| 1024 | jcp.iw = src_d.dims()[ndims - 1]; |
| 1025 | jcp.od = is_3d ? dst_d.dims()[2] : 1; |
| 1026 | jcp.oh = is_1d ? 1 : dst_d.dims()[ndims - 2]; |
| 1027 | jcp.ow = dst_d.dims()[ndims - 1]; |
| 1028 | |
| 1029 | jcp.kd = is_3d ? weights_d.dims()[with_groups + 2] : 1; |
| 1030 | jcp.kh = is_1d ? 1 : weights_d.dims()[with_groups + ndims - 2]; |
| 1031 | jcp.kw = weights_d.dims()[with_groups + ndims - 1]; |
| 1032 | |
| 1033 | jcp.f_pad = is_3d ? cd.padding[0][0] : 0; |
| 1034 | jcp.t_pad = is_1d ? 0 : cd.padding[0][ndims - 4]; |
| 1035 | jcp.l_pad = cd.padding[0][ndims - 3]; |
| 1036 | |
| 1037 | jcp.stride_d = is_3d ? cd.strides[0] : 1; |
| 1038 | jcp.stride_h = is_1d ? 1 : cd.strides[ndims - 4]; |
| 1039 | jcp.stride_w = cd.strides[ndims - 3]; |
| 1040 | |
| 1041 | jcp.dilate_d = is_3d ? cd.dilates[0] : 0; |
| 1042 | jcp.dilate_h = is_1d ? 0 : cd.dilates[ndims - 4]; |
| 1043 | jcp.dilate_w = cd.dilates[ndims - 3]; |
| 1044 | |
| 1045 | jcp.with_bias = cd.bias_desc.format_kind != format_kind::undef |
| 1046 | || cd.diff_bias_desc.format_kind != format_kind::undef; |
| 1047 | |
| 1048 | jcp.is = jcp.ih * jcp.iw; |
| 1049 | jcp.os = jcp.oh * jcp.ow; |
| 1050 | jcp.ks = jcp.kh * jcp.kw * jcp.kd; |
| 1051 | |
| 1052 | jcp.signed_input = src_d.data_type() == data_type::s8; |
| 1053 | |
| 1054 | jcp.outer_threading = false; |
| 1055 | |
| 1056 | jcp.zp = zero_point_config_t(attr); |
| 1057 | jcp.b_pad = nstl::max((jcp.oh - 1) * jcp.stride_h |
| 1058 | + (jcp.kh - 1) * (jcp.dilate_h + 1) |
| 1059 | - (jcp.ih + jcp.t_pad - 1), |
| 1060 | dim_t(0)); |
| 1061 | jcp.r_pad = nstl::max((jcp.ow - 1) * jcp.stride_w |
| 1062 | + (jcp.kw - 1) * (jcp.dilate_w + 1) |
| 1063 | - (jcp.iw + jcp.l_pad - 1), |
| 1064 | dim_t(0)); |
| 1065 | jcp.e_pad = nstl::max((jcp.od - 1) * jcp.stride_d |
| 1066 | + (jcp.kd - 1) * (jcp.dilate_d + 1) |
| 1067 | - (jcp.id + jcp.f_pad - 1), |
| 1068 | dim_t(0)); |
| 1069 | |
| 1070 | const bool zp_src_with_padding = jcp.zp.src_exists && padding_exists(jcp); |
| 1071 | |
| 1072 | if (zp_src_with_padding) { |
| 1073 | jcp.zp.src_pad_comp = zero_point_pad_comp_config_t(jcp.f_pad, jcp.e_pad, |
| 1074 | jcp.t_pad, jcp.b_pad, jcp.l_pad, jcp.r_pad, jcp.stride_d, |
| 1075 | jcp.stride_h, jcp.stride_w, jcp.od, jcp.oh, jcp.ow); |
| 1076 | } |
| 1077 | |
| 1078 | const auto set_or_check_tags |
| 1079 | = [&](format_tag_t desired_src_tag, format_tag_t desired_dst_tag, |
| 1080 | bool is_src_s8) -> status_t { |
| 1081 | using namespace format_tag; |
| 1082 | auto src_tag = any, dst_tag = any; |
| 1083 | |
| 1084 | if (src_d.format_kind() == format_kind::any) { |
| 1085 | CHECK(memory_desc_init_by_tag(src_md, desired_src_tag)); |
| 1086 | src_tag = desired_src_tag; |
| 1087 | } else { |
| 1088 | src_tag = memory_desc_matches_one_of_tag( |
| 1089 | src_md, nwc, nhwc, ndhwc, ncw, nchw, ncdhw); |
| 1090 | } |
| 1091 | |
| 1092 | if (dst_d.format_kind() == format_kind::any) { |
| 1093 | CHECK(memory_desc_init_by_tag(dst_md, desired_dst_tag)); |
| 1094 | dst_tag = desired_dst_tag; |
| 1095 | } else { |
| 1096 | dst_tag = memory_desc_matches_one_of_tag( |
| 1097 | dst_md, nwc, nhwc, ndhwc, ncw, nchw, ncdhw); |
| 1098 | } |
| 1099 | |
| 1100 | if (src_tag == format_tag::undef || dst_tag == format_tag::undef) |
| 1101 | return status::unimplemented; |
| 1102 | if (src_tag != dst_tag) return status::unimplemented; |
| 1103 | |
| 1104 | if (jcp.with_bias && bias_md.format_kind == format_kind::any) |
| 1105 | CHECK(memory_desc_init_by_tag(bias_md, x)); |
| 1106 | |
| 1107 | const bool is_nspc = utils::one_of(src_tag, nwc, nhwc, ndhwc); |
| 1108 | jcp.is_nspc = is_nspc; |
| 1109 | |
| 1110 | memory_desc_t want_wei_md = weights_md; |
| 1111 | auto wei_tag = is_nspc |
| 1112 | ? (with_groups ? utils::pick(ndims - 3, wigo, hwigo, dhwigo) |
| 1113 | : utils::pick(ndims - 3, wio, hwio, dhwio)) |
| 1114 | : (with_groups ? utils::pick(ndims - 3, goiw, goihw, goidhw) |
| 1115 | : utils::pick(ndims - 3, oiw, oihw, oidhw)); |
| 1116 | CHECK(memory_desc_init_by_tag(want_wei_md, wei_tag)); |
| 1117 | |
| 1118 | if (is_src_s8) { |
| 1119 | want_wei_md.extra.flags = 0 |
| 1120 | | memory_extra_flags::compensation_conv_s8s8 |
| 1121 | | memory_extra_flags::scale_adjust; |
| 1122 | want_wei_md.extra.compensation_mask |
| 1123 | = (1 << 0) + (with_groups ? (1 << 1) : 0); |
| 1124 | want_wei_md.extra.scale_adjust |
| 1125 | = platform::s8s8_weights_scale_factor(); |
| 1126 | } |
| 1127 | |
| 1128 | if (jcp.zp.src_exists) set_zp_src_comp_flags(want_wei_md, with_groups); |
| 1129 | |
| 1130 | if (weights_md.format_kind == format_kind::any) { |
| 1131 | weights_md = want_wei_md; |
| 1132 | return status::success; |
| 1133 | } |
| 1134 | return (want_wei_md == weights_md) ? status::success |
| 1135 | : status::unimplemented; |
| 1136 | }; |
| 1137 | |
| 1138 | const bool is_bwd_d = jcp.prop_kind == backward_data; |
| 1139 | const bool is_bwd_w = jcp.prop_kind == backward_weights; |
| 1140 | const bool is_fwd = !is_bwd_d && !is_bwd_w; |
| 1141 | |
| 1142 | bool is_int8_conv = (is_fwd ? utils::one_of(src_d.data_type(), s8, u8) |
| 1143 | : utils::one_of(dst_d.data_type(), s8, u8)) |
| 1144 | && weights_d.data_type() == s8; |
| 1145 | |
| 1146 | auto default_dat_tag = is_int8_conv |
| 1147 | ? utils::pick(ndims - 3, format_tag::nwc, format_tag::nhwc, |
| 1148 | format_tag::ndhwc) |
| 1149 | : utils::pick(ndims - 3, format_tag::ncw, format_tag::nchw, |
| 1150 | format_tag::ncdhw); |
| 1151 | CHECK(set_or_check_tags(default_dat_tag, default_dat_tag, |
| 1152 | src_md.data_type == data_type::s8)); |
| 1153 | |
| 1154 | // Does int8 conv ever need to support ncsp input format |
| 1155 | if (is_int8_conv && !src_d.matches_one_of_tag(default_dat_tag)) |
| 1156 | return status::unimplemented; |
| 1157 | |
| 1158 | CHECK(attr.set_default_formats(&dst_md)); |
| 1159 | |
| 1160 | jcp.post_ops = attr.post_ops_; |
| 1161 | |
| 1162 | const int eltwise_ind = jcp.post_ops.find(primitive_kind::eltwise); |
| 1163 | jcp.with_eltwise = eltwise_ind != -1; |
| 1164 | const int binary_ind = jcp.post_ops.find(primitive_kind::binary); |
| 1165 | jcp.with_binary = binary_ind != -1; |
| 1166 | const int sum_ind = jcp.post_ops.find(primitive_kind::sum); |
| 1167 | jcp.with_sum = sum_ind != -1; |
| 1168 | |
| 1169 | bool is_bf16_conv = false |
| 1170 | || (is_fwd |
| 1171 | && utils::everyone_is( |
| 1172 | bf16, src_d.data_type(), weights_d.data_type())) |
| 1173 | || (is_bwd_d |
| 1174 | && utils::everyone_is( |
| 1175 | bf16, dst_d.data_type(), weights_d.data_type())) |
| 1176 | || (is_bwd_w |
| 1177 | && utils::everyone_is( |
| 1178 | bf16, src_d.data_type(), dst_d.data_type())); |
| 1179 | if (is_bf16_conv && !platform::has_data_type_support(bf16)) |
| 1180 | return status::unimplemented; |
| 1181 | |
| 1182 | const int vlen = std::max(platform::get_vector_register_size(), 4); |
| 1183 | const int data_size = (is_int8_conv ? 1 : (is_bf16_conv ? 2 : 4)); |
| 1184 | const int simd_w = vlen / data_size; |
| 1185 | |
| 1186 | jcp.os_block = jcp.os; |
| 1187 | jcp.os_nb_block = 1; |
| 1188 | jcp.oc_block = jcp.oc; |
| 1189 | jcp.ic_block = jcp.ic; |
| 1190 | jcp.loop_order = gemm_loop_rlb; |
| 1191 | jcp.nthr_oc = 1; |
| 1192 | |
| 1193 | jcp.oh_block = is_fwd ? jcp.oh : jcp.ih; |
| 1194 | jcp.ow_block = is_fwd ? jcp.ow : jcp.iw; |
| 1195 | |
| 1196 | using namespace memory_tracking::names; |
| 1197 | bool is_depthwise = jcp.ic == 1 && jcp.oc == 1 && jcp.ngroups != 1; |
| 1198 | |
| 1199 | // TODO: maybe mitigate blocking restriction |
| 1200 | const auto L2 = platform::get_per_core_cache_size(2) / data_size; |
| 1201 | const int gemm_thrld = 64 * 1024; |
| 1202 | |
| 1203 | // Heuristic threshold for requested scratchpad memory to avoid |
| 1204 | // possible crash on memory allocation: |
| 1205 | // 1Gb or size of the buffers already used for this convolution proportional |
| 1206 | // to the number of threads and multiplied by a heuristic coefficient (15) |
| 1207 | const size_t zp_src_pad_comp_size = zp_src_with_padding |
| 1208 | ? (jcp.oc * jcp.ngroups * jcp.zp.src_pad_comp.d |
| 1209 | * jcp.zp.src_pad_comp.h * jcp.zp.src_pad_comp.w) |
| 1210 | : 0u; |
| 1211 | const size_t zp_src_comp_size = jcp.zp.src_is_common |
| 1212 | ? utils::rnd_up(jcp.oc * jcp.ngroups, |
| 1213 | platform::get_cache_line_size() / sizeof(int)) |
| 1214 | : 0u; |
| 1215 | |
| 1216 | const size_t weights_size = weights_d.size() |
| 1217 | + (zp_src_comp_size + zp_src_pad_comp_size) * sizeof(int32_t); |
| 1218 | |
| 1219 | static constexpr size_t scratchpad_limit_by_absolute_value = (size_t)1 |
| 1220 | << 30; // 1Gb |
| 1221 | const size_t scratchpad_limit_by_tensor_sizes |
| 1222 | = 15 * max_threads * (src_d.size() + weights_size + dst_d.size()); |
| 1223 | const size_t scratchpad_limit |
| 1224 | = nstl::min(scratchpad_limit_by_absolute_value, |
| 1225 | scratchpad_limit_by_tensor_sizes); |
| 1226 | |
| 1227 | if (is_int8_conv) { |
| 1228 | if (is_fwd) { |
| 1229 | jcp.im2col_sz |
| 1230 | = !everyone_is(true, jcp.ow == jcp.iw, jcp.oh == jcp.ih, |
| 1231 | jcp.od == jcp.id, jcp.stride_w == 1, |
| 1232 | jcp.stride_h == 1, jcp.stride_d == 1, jcp.ks == 1, |
| 1233 | !jcp.signed_input) |
| 1234 | ? (ptrdiff_t)jcp.ic * jcp.ks * jcp.os |
| 1235 | : 0; |
| 1236 | |
| 1237 | dim_t wei_size = jcp.oc * jcp.ic * jcp.kh * jcp.kw; |
| 1238 | bool is_blocking_applicable = true && is_fwd && jcp.im2col_sz |
| 1239 | && !is_3d && jcp.dilate_h == 0 && jcp.dilate_w == 0 |
| 1240 | && !is_depthwise && wei_size < L2 / 2; |
| 1241 | if (is_blocking_applicable) { |
| 1242 | // looking for oh and ow blocking |
| 1243 | dim_t h_block {jcp.oh_block}, w_block {jcp.ow_block}; |
| 1244 | dim_t ic = jcp.ic; |
| 1245 | dim_t oc = jcp.oc; |
| 1246 | dim_t iw = jcp.iw; |
| 1247 | dim_t ow = jcp.ow; |
| 1248 | dim_t oh = jcp.oh; |
| 1249 | dim_t os = oh * ow; |
| 1250 | |
| 1251 | // 1. cache requirement |
| 1252 | dim_t row_size = ic * ow * jcp.ks + 2 * (ic * iw + oc * ow); |
| 1253 | // Heuristic rule: gemm needed a lot of memory for internal |
| 1254 | // usage |
| 1255 | row_size *= 5; |
| 1256 | // memory for accumulators |
| 1257 | row_size += oc * ow * sizeof(uint32_t); |
| 1258 | // memory for transposition |
| 1259 | row_size += ic * iw; |
| 1260 | |
| 1261 | h_block = nstl::max( |
| 1262 | dim_t(1), nstl::min(oh, div_up(dim_t(L2), row_size))); |
| 1263 | if (h_block == 1) { |
| 1264 | dim_t col_size = ic * jcp.ks + 2 * (ic + oc); |
| 1265 | if (is_int8_conv) { |
| 1266 | col_size *= 5; |
| 1267 | col_size += oc * sizeof(uint32_t); |
| 1268 | col_size += ic; |
| 1269 | } |
| 1270 | w_block = nstl::max(dim_t(1), |
| 1271 | nstl::min(ow, div_up(dim_t(L2), col_size))); |
| 1272 | } |
| 1273 | |
| 1274 | // 2. threading requirement |
| 1275 | if (h_block != oh) |
| 1276 | h_block = nstl::max(dim_t(1), rnd_dn(h_block, dim_t(4))); |
| 1277 | if (w_block != ow) |
| 1278 | w_block = nstl::max(dim_t(1), rnd_dn(w_block, simd_w)); |
| 1279 | |
| 1280 | float thr_eff = 0.f; |
| 1281 | float thr_eff_treshold = 0.9f; |
| 1282 | if (w_block == ow) { |
| 1283 | do { |
| 1284 | dim_t nb_h = div_up(oh, h_block); |
| 1285 | dim_t work = jcp.ngroups * jcp.mb * jcp.od * nb_h; |
| 1286 | float disb = (float)oh / rnd_up(oh, h_block); |
| 1287 | thr_eff = (float)work / rnd_up(work, max_threads); |
| 1288 | thr_eff = (thr_eff + disb) / 2.f; |
| 1289 | if (thr_eff >= thr_eff_treshold) break; |
| 1290 | h_block = rnd_dn(h_block - 4, 4); |
| 1291 | } while (h_block > 0); |
| 1292 | } |
| 1293 | if (thr_eff |
| 1294 | < thr_eff_treshold) // we didn't find suitable h_block |
| 1295 | { |
| 1296 | h_block = 1; |
| 1297 | int nb_h = oh; |
| 1298 | do { |
| 1299 | dim_t nb_w = div_up(ow, w_block); |
| 1300 | dim_t work_amount = jcp.ngroups * jcp.mb * nb_h * nb_w; |
| 1301 | float disb = (float)ow / rnd_up(ow, w_block); |
| 1302 | thr_eff = (float)work_amount |
| 1303 | / rnd_up(work_amount, max_threads); |
| 1304 | thr_eff = (thr_eff + disb) / 2.f; |
| 1305 | if (thr_eff > thr_eff_treshold) break; |
| 1306 | w_block = rnd_dn(w_block - simd_w, simd_w); |
| 1307 | } while (w_block > 0); |
| 1308 | } |
| 1309 | h_block = nstl::max(dim_t(1), h_block); |
| 1310 | w_block = nstl::max(dim_t(1), w_block); |
| 1311 | dim_t inner_work = div_up(os, simd_w) * div_up(oc, simd_w); |
| 1312 | const float inner_thr_eff |
| 1313 | = (float)inner_work / rnd_up(inner_work, max_threads); |
| 1314 | if (thr_eff >= inner_thr_eff / 2 && h_block > 0 |
| 1315 | && w_block > 0) { |
| 1316 | jcp.oh_block = h_block; |
| 1317 | jcp.ow_block = w_block; |
| 1318 | jcp.outer_threading = true; |
| 1319 | } |
| 1320 | // updating jcp.im2col_sz |
| 1321 | if (jcp.oh_block != 1) jcp.ow_block = ow; |
| 1322 | jcp.im2col_sz |
| 1323 | = (ptrdiff_t)ic * jcp.ks * jcp.oh_block * jcp.ow_block; |
| 1324 | } |
| 1325 | // For threading selection in bwd_d we do: |
| 1326 | // 1. Rough estimation of efficiency for inner and outer threading. |
| 1327 | // 2. Gemm size estimation in assumption that it does not work |
| 1328 | // so effectively for small sizes. |
| 1329 | // 64K - this is heuristic gemm size per thread threshold. |
| 1330 | const int gemm_thrld = 64 * 1024; |
| 1331 | if (!jcp.outer_threading && !is_3d) { |
| 1332 | bool is_depthwise |
| 1333 | = jcp.ic == 1 && jcp.oc == 1 && jcp.ngroups != 1; |
| 1334 | const dim_t outer_work = jcp.ngroups * jcp.mb; |
| 1335 | const float outer_thr_eff |
| 1336 | = (float)outer_work / rnd_up(outer_work, max_threads); |
| 1337 | const size_t inner_work |
| 1338 | = div_up(jcp.is, simd_w) * div_up(jcp.ic, simd_w); |
| 1339 | const float inner_thr_eff |
| 1340 | = (float)inner_work / rnd_up(inner_work, max_threads); |
| 1341 | jcp.outer_threading |
| 1342 | = (is_depthwise |
| 1343 | || (jcp.is / max_threads < 64 && jcp.mb != 1)) |
| 1344 | && (outer_thr_eff / inner_thr_eff >= 1.f |
| 1345 | || (jcp.os * jcp.ic * jcp.oc) / max_threads |
| 1346 | < gemm_thrld); |
| 1347 | } |
| 1348 | jcp.nthr = jcp.outer_threading ? max_threads : 1; |
| 1349 | scratchpad.book<int8_t>( |
| 1350 | key_conv_gemm_col, jcp.nthr * jcp.im2col_sz); |
| 1351 | scratchpad.book<int32_t>(key_conv_int_dat_in_acc_dt, |
| 1352 | jcp.nthr * jcp.oh_block * jcp.ow_block * jcp.oc); |
| 1353 | scratchpad.book<int8_t>( |
| 1354 | key_conv_gemm_imtr, jcp.nthr * jcp.id * jcp.is * jcp.ic); |
| 1355 | } else if (is_bwd_d) { |
| 1356 | jcp.im2col_sz |
| 1357 | = !everyone_is(true, jcp.ow == jcp.iw, jcp.oh == jcp.ih, |
| 1358 | jcp.od == jcp.id, jcp.stride_w == 1, |
| 1359 | jcp.stride_h == 1, jcp.stride_d == 1, jcp.ks == 1, |
| 1360 | !jcp.signed_input) |
| 1361 | ? (ptrdiff_t)jcp.ic * jcp.ks * jcp.os * jcp.od |
| 1362 | : 0; |
| 1363 | |
| 1364 | bool is_depthwise = jcp.ic == 1 && jcp.oc == 1 && jcp.ngroups != 1; |
| 1365 | const size_t outer_work = jcp.ngroups * jcp.mb; |
| 1366 | const float outer_thr_eff |
| 1367 | = (float)outer_work / rnd_up(outer_work, max_threads); |
| 1368 | const size_t inner_work |
| 1369 | = div_up(jcp.is, simd_w) * div_up(jcp.ic, simd_w); |
| 1370 | const float inner_thr_eff |
| 1371 | = (float)inner_work / rnd_up(inner_work, max_threads); |
| 1372 | jcp.outer_threading = !is_3d |
| 1373 | && (is_depthwise |
| 1374 | || (jcp.is / max_threads < 64 && jcp.mb != 1)) |
| 1375 | && (outer_thr_eff / inner_thr_eff >= 1.f |
| 1376 | || (jcp.is * jcp.ic * jcp.oc) / max_threads |
| 1377 | < gemm_thrld); |
| 1378 | |
| 1379 | jcp.nthr = jcp.outer_threading ? max_threads : 1; |
| 1380 | scratchpad.book<int32_t>( |
| 1381 | key_conv_gemm_col, jcp.nthr * jcp.im2col_sz); |
| 1382 | scratchpad.book<int32_t>(key_conv_int_dat_in_acc_dt, |
| 1383 | jcp.nthr * jcp.is * jcp.id * jcp.ic); |
| 1384 | } else if (is_bwd_w) { |
| 1385 | assert(!"unimplemented prop_kind"); |
| 1386 | return status::unimplemented; |
| 1387 | } |
| 1388 | } else { |
| 1389 | jcp.im2col_sz = !everyone_is(true, jcp.ow == jcp.iw, jcp.oh == jcp.ih, |
| 1390 | jcp.od == jcp.id, jcp.stride_w == 1, |
| 1391 | jcp.stride_h == 1, jcp.stride_d == 1, |
| 1392 | jcp.ks == 1, !jcp.signed_input) |
| 1393 | ? (ptrdiff_t)jcp.ic * jcp.ks * jcp.os |
| 1394 | : 0; |
| 1395 | if (jcp.is_nspc && is_fwd) { |
| 1396 | const size_t wei_size |
| 1397 | = static_cast<size_t>(jcp.oc) * jcp.ic * jcp.kh * jcp.kw; |
| 1398 | bool is_blocking_applicable = true && is_fwd && jcp.im2col_sz |
| 1399 | && !is_3d && jcp.dilate_h == 0 && jcp.dilate_w == 0 |
| 1400 | && !is_depthwise && wei_size < static_cast<size_t>(L2) / 2; |
| 1401 | // Logic for blocking for f32_nspc gemm convolution follows that of |
| 1402 | // int8_nspc gemm convolution. Currently, not optimized for f32 |
| 1403 | // data type. |
| 1404 | if (is_blocking_applicable) { |
| 1405 | // looking for oh and ow blocking |
| 1406 | size_t h_block = jcp.oh_block; |
| 1407 | size_t w_block = jcp.ow_block; |
| 1408 | |
| 1409 | const size_t ic = jcp.ic; |
| 1410 | const size_t oc = jcp.oc; |
| 1411 | const size_t iw = jcp.iw; |
| 1412 | const size_t ow = jcp.ow; |
| 1413 | const size_t oh = jcp.oh; |
| 1414 | const size_t os = oh * ow; |
| 1415 | |
| 1416 | // 1. cache requirement |
| 1417 | size_t row_size = ic * ow * jcp.ks * data_size |
| 1418 | + 2 * (ic * iw + oc * ow) * data_size; |
| 1419 | // Heuristic rule: gemm needed a lot of memory for internal |
| 1420 | // usage |
| 1421 | row_size *= 5; |
| 1422 | // memory for accumulators |
| 1423 | row_size += oc * ow * data_size; |
| 1424 | // memory for transposition |
| 1425 | row_size += ic * iw * data_size; |
| 1426 | |
| 1427 | const size_t L2_rows = div_up(L2, row_size); |
| 1428 | h_block = saturate(size_t {1}, L2_rows, oh); |
| 1429 | if (h_block == 1) { |
| 1430 | size_t col_size = ic * jcp.ks * data_size |
| 1431 | + 2 * (ic + oc) * data_size; |
| 1432 | const size_t L2_cols = div_up(L2, col_size); |
| 1433 | w_block = saturate(size_t {1}, L2_cols, ow); |
| 1434 | } |
| 1435 | |
| 1436 | // 2. threading requirement |
| 1437 | if (h_block != oh) |
| 1438 | h_block = nstl::max(size_t {1}, rnd_dn(h_block, 4)); |
| 1439 | if (w_block != ow) |
| 1440 | w_block = nstl::max(size_t {1}, rnd_dn(w_block, simd_w)); |
| 1441 | |
| 1442 | float thr_eff = 0.f; |
| 1443 | float thr_eff_treshold = 0.9f; |
| 1444 | if (w_block == ow) { |
| 1445 | do { |
| 1446 | size_t nb_h = div_up(oh, h_block); |
| 1447 | size_t work = jcp.ngroups * jcp.mb * jcp.od * nb_h; |
| 1448 | float disb = (float)oh / rnd_up(oh, h_block); |
| 1449 | thr_eff = (float)work / rnd_up(work, max_threads); |
| 1450 | thr_eff = (thr_eff + disb) / 2.f; |
| 1451 | if (thr_eff >= thr_eff_treshold) break; |
| 1452 | |
| 1453 | if (h_block < 4) |
| 1454 | h_block = 0; |
| 1455 | else |
| 1456 | h_block = rnd_dn(h_block - 4, 4); |
| 1457 | } while (h_block > 0); |
| 1458 | } |
| 1459 | if (thr_eff |
| 1460 | < thr_eff_treshold) // we didn't find suitable h_block |
| 1461 | { |
| 1462 | h_block = 1; |
| 1463 | size_t nb_h = oh; |
| 1464 | do { |
| 1465 | size_t nb_w = div_up(ow, w_block); |
| 1466 | size_t work_amount = jcp.ngroups * jcp.mb * nb_h * nb_w; |
| 1467 | float disb = (float)ow / rnd_up(ow, w_block); |
| 1468 | thr_eff = (float)work_amount |
| 1469 | / rnd_up(work_amount, max_threads); |
| 1470 | thr_eff = (thr_eff + disb) / 2.f; |
| 1471 | if (thr_eff > thr_eff_treshold) break; |
| 1472 | |
| 1473 | if (w_block < static_cast<size_t>(simd_w)) |
| 1474 | w_block = 0; |
| 1475 | else |
| 1476 | w_block = rnd_dn(w_block - simd_w, simd_w); |
| 1477 | } while (w_block > 0); |
| 1478 | } |
| 1479 | h_block = nstl::max(size_t {1}, h_block); |
| 1480 | w_block = nstl::max(size_t {1}, w_block); |
| 1481 | const size_t inner_work |
| 1482 | = div_up(os, simd_w) * div_up(oc, simd_w); |
| 1483 | const float inner_thr_eff |
| 1484 | = (float)inner_work / rnd_up(inner_work, max_threads); |
| 1485 | if (thr_eff >= inner_thr_eff / 2 && h_block > 0 |
| 1486 | && w_block > 0) { |
| 1487 | jcp.oh_block = static_cast<int>(h_block); |
| 1488 | jcp.ow_block = static_cast<int>(w_block); |
| 1489 | jcp.outer_threading = true; |
| 1490 | } |
| 1491 | // updating jcp.im2col_sz |
| 1492 | if (jcp.oh_block != 1) jcp.ow_block = static_cast<int>(ow); |
| 1493 | jcp.im2col_sz |
| 1494 | = (ptrdiff_t)ic * jcp.ks * jcp.oh_block * jcp.ow_block; |
| 1495 | } |
| 1496 | // For threading selection in fwd_d we do: |
| 1497 | // 1. Rough estimation of efficiency for inner and outer threading. |
| 1498 | // 2. Gemm size estimation in assumption that it does not work |
| 1499 | // so effectively for small sizes. |
| 1500 | // 64K - this is heuristic gemm size per thread threshold. |
| 1501 | constexpr size_t gemm_thrld = 64 * 1024; |
| 1502 | if (!jcp.outer_threading && !is_3d) { |
| 1503 | bool is_depthwise |
| 1504 | = jcp.ic == 1 && jcp.oc == 1 && jcp.ngroups != 1; |
| 1505 | const size_t outer_work = jcp.ngroups * jcp.mb; |
| 1506 | const float outer_thr_eff |
| 1507 | = (float)outer_work / rnd_up(outer_work, max_threads); |
| 1508 | const size_t inner_work |
| 1509 | = div_up(jcp.is, simd_w) * div_up(jcp.ic, simd_w); |
| 1510 | const float inner_thr_eff |
| 1511 | = (float)inner_work / rnd_up(inner_work, max_threads); |
| 1512 | jcp.outer_threading |
| 1513 | = (is_depthwise |
| 1514 | || (jcp.is / max_threads < 64 && jcp.mb != 1)) |
| 1515 | && (outer_thr_eff / inner_thr_eff >= 1.f |
| 1516 | || (static_cast<size_t>(jcp.os) * jcp.ic |
| 1517 | * jcp.oc) |
| 1518 | / max_threads |
| 1519 | < gemm_thrld); |
| 1520 | } |
| 1521 | jcp.nthr = jcp.outer_threading ? max_threads : 1; |
| 1522 | const size_t gemm_col_datatype_size |
| 1523 | = is_bf16_conv ? sizeof(bfloat16_t) : sizeof(float); |
| 1524 | |
| 1525 | scratchpad.book(key_conv_gemm_col, jcp.nthr * jcp.im2col_sz, |
| 1526 | gemm_col_datatype_size); |
| 1527 | if (is_bf16_conv) { |
| 1528 | scratchpad.book<float>(key_conv_gemm_acc, |
| 1529 | jcp.nthr * static_cast<size_t>(jcp.oh_block) |
| 1530 | * jcp.ow_block * jcp.oc); |
| 1531 | } |
| 1532 | |
| 1533 | scratchpad.book(key_conv_gemm_imtr, |
| 1534 | jcp.nthr * static_cast<size_t>(jcp.id) * jcp.is * jcp.ic, |
| 1535 | gemm_col_datatype_size); |
| 1536 | if (is_bf16_conv && jcp.with_bias |
| 1537 | && one_of(data_type::bf16, cd.diff_bias_desc.data_type, |
| 1538 | cd.bias_desc.data_type)) { |
| 1539 | scratchpad.book<float>( |
| 1540 | key_conv_bias_bf16_convert_wsp, jcp.ngroups * jcp.oc); |
| 1541 | } |
| 1542 | |
| 1543 | } else if (!jcp.is_nspc && is_fwd) { |
| 1544 | const dim_t sh = jcp.stride_h; |
| 1545 | const dim_t sw = jcp.stride_w; |
| 1546 | const dim_t spatial = jcp.mb * jcp.ngroups * jcp.od * jcp.os; |
| 1547 | dim_t K = jcp.ic * jcp.ks; |
| 1548 | |
| 1549 | // There is some heuristics in the definition of |
| 1550 | // inner/outer threading cross point due to the nature of the |
| 1551 | // gemm implementation which we cannot control |
| 1552 | bool is_blocking_applicable = true && !is_3d |
| 1553 | && (!jcp.im2col_sz |
| 1554 | // spatial is small |
| 1555 | || spatial >= max_threads * simd_w |
| 1556 | // inner threading work is greater then outer |
| 1557 | // threading work |
| 1558 | || jcp.os < jcp.mb * jcp.ngroups * jcp.od |
| 1559 | // im2col is big |
| 1560 | || (sw == 1 && K <= 0.05 * jcp.oc)) |
| 1561 | // heuristic condition |
| 1562 | && (jcp.im2col_sz |
| 1563 | || (jcp.ic / jcp.oc < 42 |
| 1564 | && jcp.ic * jcp.oc * jcp.is < 1024)); |
| 1565 | |
| 1566 | if (is_blocking_applicable) { |
| 1567 | const dim_t min_oc_block = 8; |
| 1568 | const dim_t min_os_block = simd_w; |
| 1569 | const float non_cache_access = 20; |
| 1570 | const float strided_im2col_k = 8; |
| 1571 | const float thr_disb_k = 8; |
| 1572 | const float thr_mem_eff_k {1}, oc_disb_k {1}, os_disb_k {1}, |
| 1573 | ic_disb_k {1}, reg_osb_disb_k {1}, gemm_eff_k {0.5}, |
| 1574 | gemm_calc_eff_k {1}; |
| 1575 | const float k_sum = thr_disb_k + oc_disb_k + os_disb_k |
| 1576 | + ic_disb_k + reg_osb_disb_k + thr_mem_eff_k |
| 1577 | + gemm_eff_k + gemm_calc_eff_k; |
| 1578 | |
| 1579 | auto calc_max_icb = [=](dim_t nthr_oc, dim_t ocb, dim_t osb, |
| 1580 | dim_t oc_per_thr, |
| 1581 | dim_t os_per_thr) { |
| 1582 | const dim_t block_out_size = ocb * osb; |
| 1583 | // TODO: need more precise calculation if stride more than |
| 1584 | // kernel size |
| 1585 | const dim_t inp_row_size = sh * sw * osb; |
| 1586 | dim_t max_icb = 1; |
| 1587 | if (jcp.im2col_sz) { |
| 1588 | const dim_t col_row_size = jcp.ks * osb; |
| 1589 | if (osb >= os_per_thr) { // one pass by os |
| 1590 | const dim_t wei_col_size = jcp.ks * ocb; |
| 1591 | max_icb = L2 / (inp_row_size + col_row_size); |
| 1592 | if (ocb < oc_per_thr) { |
| 1593 | max_icb = nstl::min(max_icb, |
| 1594 | (L2 - block_out_size) |
| 1595 | / (col_row_size |
| 1596 | + wei_col_size)); |
| 1597 | } |
| 1598 | } else { |
| 1599 | const dim_t wei_col_size = jcp.ks * oc_per_thr; |
| 1600 | max_icb = (L2 - block_out_size) |
| 1601 | / (inp_row_size + col_row_size |
| 1602 | + wei_col_size); |
| 1603 | } |
| 1604 | } else { |
| 1605 | if (osb >= os_per_thr) |
| 1606 | max_icb = L2 / inp_row_size; |
| 1607 | else { |
| 1608 | const dim_t wei_col_size = jcp.ks * oc_per_thr; |
| 1609 | max_icb = L2 / (inp_row_size + wei_col_size); |
| 1610 | } |
| 1611 | } |
| 1612 | if (max_icb < jcp.ic) { |
| 1613 | if (jcp.im2col_sz) { |
| 1614 | const dim_t col_row_size = jcp.ks * osb; |
| 1615 | const dim_t wei_col_size = jcp.ks * oc_per_thr; |
| 1616 | max_icb = (L2 - block_out_size) |
| 1617 | / (inp_row_size + col_row_size |
| 1618 | + wei_col_size); |
| 1619 | } |
| 1620 | } |
| 1621 | return max_icb; |
| 1622 | }; |
| 1623 | |
| 1624 | dim_t best_ocb {1}, best_osb {1}; |
| 1625 | dim_t best_nthr_oc {1}; |
| 1626 | dim_t best_icb {jcp.ic}; |
| 1627 | float best_thr_eff = 0; |
| 1628 | |
| 1629 | auto try_cfg = [&](dim_t nthr_oc, dim_t ocb, dim_t osb) { |
| 1630 | // for given nthr_oc, oc block: |
| 1631 | // 1. find ic block to fit into cache |
| 1632 | // 2. estimate efficiency basing on rules and heuristic: |
| 1633 | // - Minimize im2col cost |
| 1634 | // - ratio of FMA number to data size |
| 1635 | // - gemm works better if M divided by 48 and N divided by 8 |
| 1636 | |
| 1637 | const dim_t max_oc = div_up(jcp.oc, nthr_oc); |
| 1638 | const dim_t min_oc = nstl::max(dim_t(1), jcp.oc / nthr_oc); |
| 1639 | const dim_t max_os |
| 1640 | = div_up(spatial, (dim_t)(max_threads / nthr_oc)); |
| 1641 | ocb = utils::saturate(min_oc_block, max_oc, ocb); |
| 1642 | osb = utils::saturate(min_os_block, max_os, osb); |
| 1643 | |
| 1644 | // The computation of max_thr_size and min_thr_size is |
| 1645 | // based on work balance using: |
| 1646 | // balance2D(max_threads, i, spatial, sp_start, sp_end, |
| 1647 | // jcp.oc, oc_start, oc_end, nthr_oc); |
| 1648 | size_t max_thr_size = 1; |
| 1649 | { |
| 1650 | const dim_t min_os = div_up( |
| 1651 | spatial, (dim_t)div_up(max_threads, nthr_oc)); |
| 1652 | /* --- compute max_thr_size ------------ |
| 1653 | may not necessarily be (max_oc * max_os) |
| 1654 | thr_size = thr_oc * (spatial /nthrs_in_slice); |
| 1655 | with spatial as const, thr_size has maxima when |
| 1656 | (A: thr_oc is max) and (B: nthrs_in_slice is min) |
| 1657 | */ |
| 1658 | if (jcp.oc % nthr_oc > max_threads % nthr_oc) { |
| 1659 | // If (A) and (B) are true together, then it is the |
| 1660 | // global max |
| 1661 | max_thr_size = max_oc * max_os; |
| 1662 | } else { |
| 1663 | const size_t oc_max_os_min = max_oc * min_os; |
| 1664 | const size_t oc_min_os_max = min_oc * max_os; |
| 1665 | max_thr_size |
| 1666 | = nstl::max(oc_max_os_min, oc_min_os_max); |
| 1667 | } |
| 1668 | } |
| 1669 | |
| 1670 | size_t min_thr_size {1}; |
| 1671 | { |
| 1672 | const dim_t min_os = nstl::max(dim_t(1), |
| 1673 | spatial / div_up(max_threads, nthr_oc)); |
| 1674 | /* --- compute min_thr_size ------------ |
| 1675 | may not necessarily be (min_oc * min_y) |
| 1676 | thr_size = thr_oc * (spatial /nthrs_in_slice); |
| 1677 | with spatial as const, thr_size has minima when |
| 1678 | (A: thr_oc is min) and (B: nthrs_in_slice is max) |
| 1679 | */ |
| 1680 | if (max_threads % nthr_oc > jcp.oc % nthr_oc) { |
| 1681 | // If (A) and (B) are true together, then it is the |
| 1682 | // global min |
| 1683 | min_thr_size = min_oc * min_os; |
| 1684 | } else { |
| 1685 | const size_t oc_max_os_min = max_oc * min_os; |
| 1686 | const size_t oc_min_os_max = min_oc |
| 1687 | * (size_t)(spatial |
| 1688 | / (dim_t)(max_threads / nthr_oc)); |
| 1689 | min_thr_size |
| 1690 | = nstl::min(oc_max_os_min, oc_min_os_max); |
| 1691 | } |
| 1692 | } |
| 1693 | auto thr_disb = (float)min_thr_size / max_thr_size; |
| 1694 | |
| 1695 | const dim_t oc_per_thr = max_oc; |
| 1696 | const dim_t os_per_thr = max_os; |
| 1697 | ocb = nstl::min(oc_per_thr, ocb); |
| 1698 | const dim_t os_max = nstl::min(jcp.os, os_per_thr); |
| 1699 | osb = nstl::min(os_max, osb); |
| 1700 | |
| 1701 | // -- selecting icb --------------------- |
| 1702 | dim_t max_ic_block = calc_max_icb( |
| 1703 | nthr_oc, ocb, osb, oc_per_thr, os_per_thr); |
| 1704 | // if we don't fit into cache then access to memory is |
| 1705 | // expensive |
| 1706 | dim_t mem_access_cost |
| 1707 | = (max_ic_block < 1) ? non_cache_access : 1; |
| 1708 | max_ic_block = nstl::max(dim_t(1), max_ic_block); |
| 1709 | dim_t icb = nstl::max( |
| 1710 | dim_t(1), jcp.ic / div_up(jcp.ic, max_ic_block)); |
| 1711 | dim_t nb_ic = div_up(jcp.ic, icb); |
| 1712 | dim_t kb = icb * jcp.ks; |
| 1713 | dim_t kb_caligned = rnd_up(kb, simd_w); |
| 1714 | |
| 1715 | // -- mem efficiency ------------ |
| 1716 | const size_t out_size |
| 1717 | = oc_per_thr * rnd_up(os_per_thr, simd_w); |
| 1718 | const size_t out_ops = mem_access_cost * out_size |
| 1719 | * ((icb == jcp.ic) ? 1 : (2 * nb_ic - 1)); |
| 1720 | const dim_t osb_caligned = rnd_up(osb, simd_w); |
| 1721 | const size_t inp_size |
| 1722 | = jcp.ic * rnd_up(os_per_thr * sh * sw, simd_w); |
| 1723 | size_t inp_ops = 0; |
| 1724 | size_t col_ops = 0; |
| 1725 | // TODO: simplify calculations |
| 1726 | if (jcp.im2col_sz) { |
| 1727 | inp_ops = mem_access_cost * jcp.ks * inp_size; |
| 1728 | const float col_tail_koeff = (float)osb_caligned / osb; |
| 1729 | col_ops = mem_access_cost |
| 1730 | * (jcp.ks * inp_size * col_tail_koeff |
| 1731 | + jcp.ks * inp_size * col_tail_koeff); |
| 1732 | if (sw != 1) // im2col with strides is much slower |
| 1733 | col_ops *= strided_im2col_k; |
| 1734 | } else { |
| 1735 | inp_ops = mem_access_cost * jcp.ks * inp_size; |
| 1736 | } |
| 1737 | // TODO: what about groups? |
| 1738 | const size_t wei_size = oc_per_thr * rnd_up(K, simd_w); |
| 1739 | const size_t wei_ops = mem_access_cost * wei_size; |
| 1740 | // ratio of real FMA to number of memory ops |
| 1741 | const float thr_mem_eff |
| 1742 | = (((float)os_per_thr / simd_w) * oc_per_thr * K) |
| 1743 | / (inp_ops + col_ops + wei_ops + out_ops); |
| 1744 | |
| 1745 | auto oc_disb = (float)oc_per_thr / rnd_up(oc_per_thr, ocb); |
| 1746 | auto os_disb = (float)os_max / rnd_up(os_max, osb); |
| 1747 | auto ic_disb = (float)jcp.ic / rnd_up(jcp.ic, icb); |
| 1748 | |
| 1749 | auto reg_osb_disb = (float)osb / rnd_up(osb, 3 * simd_w); |
| 1750 | |
| 1751 | // Heuristics |
| 1752 | const float gemm_eff = ((float)osb * ocb * kb) |
| 1753 | / ((float)oc_per_thr * os_per_thr * K); |
| 1754 | |
| 1755 | // number of FMA to memory size |
| 1756 | const float gemm_calc_eff |
| 1757 | = (((float)osb / simd_w) * ocb * kb) |
| 1758 | / (osb_caligned * kb + ocb * kb_caligned |
| 1759 | + ocb * osb_caligned); |
| 1760 | // optimization: remove pow, when corresponding weight is 1 |
| 1761 | const float res_eff = pow(pow(thr_disb, thr_disb_k) |
| 1762 | * oc_disb // pow(oc_disb, oc_disb_k) |
| 1763 | * os_disb // pow(os_disb, os_disb_k) |
| 1764 | * ic_disb // pow(ic_disb, ic_disb_k) |
| 1765 | // pow(reg_osb_disb, reg_osb_disb_k) |
| 1766 | * reg_osb_disb |
| 1767 | //pow(thr_mem_eff, thr_mem_eff_k) |
| 1768 | * thr_mem_eff |
| 1769 | //pow(gemm_calc_eff, gemm_calc_eff_k) |
| 1770 | * pow(gemm_eff, gemm_eff_k) * gemm_calc_eff, |
| 1771 | 1.f / k_sum); |
| 1772 | |
| 1773 | if (res_eff > best_thr_eff) { |
| 1774 | best_thr_eff = res_eff; |
| 1775 | best_nthr_oc = nthr_oc; |
| 1776 | best_ocb = ocb; |
| 1777 | best_osb = osb; |
| 1778 | best_icb = icb; |
| 1779 | } |
| 1780 | }; |
| 1781 | |
| 1782 | auto explore_cfg = [&](dim_t nthr_oc, dim_t ocb, dim_t osb) { |
| 1783 | try_cfg(nthr_oc, ocb, osb); |
| 1784 | // few combinations to try, as the eff is better when ocb is |
| 1785 | // multiple of 8 and osb is multiple of 48 or min_os_block. |
| 1786 | try_cfg(nthr_oc, rnd_dn(ocb, 8), rnd_dn(osb, 48)); |
| 1787 | try_cfg(nthr_oc, rnd_up(ocb, 8), rnd_dn(osb, 48)); |
| 1788 | try_cfg(nthr_oc, rnd_up(ocb, 8), rnd_up(osb, min_os_block)); |
| 1789 | try_cfg(nthr_oc, rnd_up(ocb, 8), rnd_up(osb, 48)); |
| 1790 | }; |
| 1791 | |
| 1792 | for (dim_t nthr_oc = 1; nthr_oc <= max_threads; ++nthr_oc) { |
| 1793 | const dim_t max_oc_per_thr = div_up(jcp.oc, nthr_oc); |
| 1794 | dim_t max_os_per_thr |
| 1795 | = div_up(spatial, max_threads / nthr_oc); |
| 1796 | dim_t ocb {1}, osb {1}, icb {1}; |
| 1797 | if (jcp.im2col_sz) { |
| 1798 | try_cfg(nthr_oc, max_oc_per_thr, max_os_per_thr); |
| 1799 | if ((best_ocb == max_oc_per_thr) |
| 1800 | && (best_osb == max_os_per_thr) |
| 1801 | && (best_icb == jcp.ic)) { |
| 1802 | // best case scenario |
| 1803 | continue; |
| 1804 | } |
| 1805 | |
| 1806 | /* |
| 1807 | memory eq from calc_max_icb(): |
| 1808 | max_icb = (L2 - block_out_size) |
| 1809 | / (inp_row_size + col_row_size |
| 1810 | + wei_col_size); |
| 1811 | icb*sh*sw*osb + icb*jcp.ks*osb + |
| 1812 | jcp.ks*max_oc_per_thr*icb + osb *ocb = L2 |
| 1813 | |
| 1814 | a_k*icb*osb + b_k*icb + osb*ocb = L2 |
| 1815 | We would like to maximize icb*osb*ocb (FMA). |
| 1816 | |
| 1817 | Unfortunately, above eq and constraint doesn't have |
| 1818 | a single solution. So, based on experiments we try |
| 1819 | few scenarios. |
| 1820 | 1. icb = jcp.ic |
| 1821 | 2. Solving the constraint eq we get |
| 1822 | osb = (L2 - 2*b_k*icb)/(2*a_k*icb) >= min_oc_block |
| 1823 | => icb <= (L2)/(2* min_oc_block * a_k + 2 * b_k) |
| 1824 | 3. Maximize channel compute: |
| 1825 | ocb = max_oc_per_thr; |
| 1826 | icb = jcp.ic; |
| 1827 | */ |
| 1828 | dim_t a_k = sh * sw + jcp.ks; |
| 1829 | dim_t b_k = jcp.ks * max_oc_per_thr; |
| 1830 | |
| 1831 | // Note 1: |
| 1832 | icb = jcp.ic; |
| 1833 | ocb = utils::saturate(min_oc_block, max_oc_per_thr, |
| 1834 | (L2 - a_k * icb * min_os_block - b_k * icb) |
| 1835 | / min_os_block); |
| 1836 | osb = utils::saturate(min_os_block, max_os_per_thr, |
| 1837 | (L2 - b_k * icb) / (a_k * icb + ocb)); |
| 1838 | explore_cfg(nthr_oc, ocb, osb); |
| 1839 | |
| 1840 | // Note 2: |
| 1841 | const dim_t icb_max = nstl::max(dim_t(1), |
| 1842 | L2 / (2 * min_oc_block * a_k + 2 * b_k)); |
| 1843 | if (icb_max < jcp.ic) { |
| 1844 | // adjust icb, such that it is evenly distributed. |
| 1845 | icb = jcp.ic |
| 1846 | / nstl::max(dim_t(1), jcp.ic / icb_max); |
| 1847 | osb = nstl::max(dim_t(1), |
| 1848 | (L2 - 2 * b_k * icb) / (2 * icb * a_k)); |
| 1849 | ocb = L2 / 2 / osb; |
| 1850 | |
| 1851 | if (ocb > max_oc_per_thr) { |
| 1852 | ocb = max_oc_per_thr; |
| 1853 | // reduce mem eq by making ocb constant. we get |
| 1854 | osb = utils::saturate(min_os_block, |
| 1855 | max_os_per_thr, |
| 1856 | (L2 - b_k * icb) / (a_k * icb + ocb)); |
| 1857 | } else if (osb > max_os_per_thr) { |
| 1858 | // reduce mem eq by making osb constant. we get |
| 1859 | osb = max_os_per_thr; |
| 1860 | ocb = utils::saturate(min_oc_block, |
| 1861 | max_oc_per_thr, |
| 1862 | (L2 - a_k * icb * osb - b_k * icb) |
| 1863 | / (osb)); |
| 1864 | } |
| 1865 | |
| 1866 | explore_cfg(nthr_oc, ocb, osb); |
| 1867 | } |
| 1868 | |
| 1869 | // Note 3: |
| 1870 | ocb = max_oc_per_thr; |
| 1871 | icb = jcp.ic; |
| 1872 | osb = nstl::max(min_os_block, |
| 1873 | rnd_dn((L2 - b_k * icb) / (a_k * icb + ocb), |
| 1874 | min_os_block)); |
| 1875 | explore_cfg(nthr_oc, ocb, osb); |
| 1876 | |
| 1877 | } else { |
| 1878 | // from calc_max_icb, memory eq is independent of ocb. |
| 1879 | // So, set it to maximum. |
| 1880 | ocb = max_oc_per_thr; |
| 1881 | osb = (L2 - jcp.ks * jcp.ic) / (sh * sw * jcp.ic); |
| 1882 | explore_cfg(nthr_oc, ocb, osb); |
| 1883 | } |
| 1884 | } |
| 1885 | jcp.outer_threading = true; |
| 1886 | jcp.nthr_oc = best_nthr_oc; |
| 1887 | jcp.oc_block = best_ocb; |
| 1888 | jcp.os_block = best_osb; |
| 1889 | jcp.ic_block = best_icb; |
| 1890 | |
| 1891 | // TODO: define loop order |
| 1892 | // if im2col then gemm_loop_rlb and gemm_loop_lrb looks |
| 1893 | // preferable otherwise other loop orders are possible |
| 1894 | jcp.loop_order = gemm_loop_rlb; |
| 1895 | } else { |
| 1896 | const size_t outer_work_amount = jcp.ngroups * jcp.mb * jcp.od; |
| 1897 | const float outer_thr_eff = (float)outer_work_amount |
| 1898 | / rnd_up(outer_work_amount, max_threads); |
| 1899 | const size_t inner_work_amount |
| 1900 | = div_up(jcp.os, simd_w) * div_up(jcp.oc, simd_w); |
| 1901 | const float inner_thr_eff = (float)inner_work_amount |
| 1902 | / rnd_up(inner_work_amount, max_threads); |
| 1903 | jcp.outer_threading = jcp.os / max_threads < 512 |
| 1904 | && IMPLICATION( |
| 1905 | jcp.od == 1, jcp.mb != 1 || jcp.ngroups > 2) |
| 1906 | && (outer_thr_eff / inner_thr_eff >= 1.f |
| 1907 | || (jcp.os * jcp.ic * jcp.oc) / max_threads |
| 1908 | < gemm_thrld); |
| 1909 | } |
| 1910 | jcp.os_nb_block = div_up(jcp.os, jcp.os_block); |
| 1911 | |
| 1912 | // BF16: other loops should be explored for potential |
| 1913 | // performance speedup, but BF16-dst post-processing implementation |
| 1914 | // would require enabling this support. |
| 1915 | if (is_bf16_conv) jcp.loop_order = gemm_loop_lbr; |
| 1916 | |
| 1917 | if (jcp.im2col_sz) |
| 1918 | jcp.im2col_sz = (ptrdiff_t)jcp.ic_block * jcp.ks * jcp.os_block; |
| 1919 | } else if (jcp.is_nspc && is_bwd_d) { |
| 1920 | jcp.im2col_sz |
| 1921 | = !everyone_is(true, jcp.ow == jcp.iw, jcp.oh == jcp.ih, |
| 1922 | jcp.od == jcp.id, jcp.stride_w == 1, |
| 1923 | jcp.stride_h == 1, jcp.stride_d == 1, jcp.ks == 1, |
| 1924 | !jcp.signed_input) |
| 1925 | ? (ptrdiff_t)jcp.ic * jcp.ks * jcp.os * jcp.od |
| 1926 | : 0; |
| 1927 | |
| 1928 | bool is_depthwise = jcp.ic == 1 && jcp.oc == 1 && jcp.ngroups != 1; |
| 1929 | const size_t outer_work = jcp.ngroups * jcp.mb; |
| 1930 | const float outer_thr_eff |
| 1931 | = (float)outer_work / rnd_up(outer_work, max_threads); |
| 1932 | const size_t inner_work |
| 1933 | = div_up(jcp.is, simd_w) * div_up(jcp.ic, simd_w); |
| 1934 | const float inner_thr_eff |
| 1935 | = (float)inner_work / rnd_up(inner_work, max_threads); |
| 1936 | jcp.outer_threading = !is_3d |
| 1937 | && (is_depthwise |
| 1938 | || (jcp.is / max_threads < 64 && jcp.mb != 1)) |
| 1939 | && (outer_thr_eff / inner_thr_eff >= 1.f |
| 1940 | || (static_cast<size_t>(jcp.is) * jcp.ic * jcp.oc) |
| 1941 | / max_threads |
| 1942 | < gemm_thrld); |
| 1943 | |
| 1944 | jcp.nthr = jcp.outer_threading ? max_threads : 1; |
| 1945 | scratchpad.book<float>(key_conv_gemm_col, jcp.nthr * jcp.im2col_sz); |
| 1946 | if (jcp.ngroups > 1 || is_bf16_conv) |
| 1947 | scratchpad.book<float>(key_conv_gemm_acc, |
| 1948 | jcp.nthr * static_cast<size_t>(jcp.is) * jcp.id |
| 1949 | * jcp.ic); |
| 1950 | } else if (!jcp.is_nspc && is_bwd_d) { |
| 1951 | const size_t outer_work_amount = jcp.ngroups * jcp.mb; |
| 1952 | const float outer_thr_eff = (float)outer_work_amount |
| 1953 | / rnd_up(outer_work_amount, max_threads); |
| 1954 | const size_t inner_work |
| 1955 | = div_up(jcp.is, simd_w) * div_up(jcp.ic, simd_w); |
| 1956 | const float inner_thr_eff |
| 1957 | = (float)inner_work / rnd_up(inner_work, max_threads); |
| 1958 | jcp.outer_threading = (jcp.os / max_threads < 512 || jcp.ks < 64) |
| 1959 | && (jcp.mb != 1 || jcp.ngroups > 2) |
| 1960 | && (outer_thr_eff / inner_thr_eff >= 1.f |
| 1961 | || (jcp.is * jcp.ic * jcp.oc) / max_threads |
| 1962 | < gemm_thrld); |
| 1963 | } else if (jcp.is_nspc && is_bwd_w) { |
| 1964 | jcp.im2col_sz |
| 1965 | = !everyone_is(true, jcp.ow == jcp.iw, jcp.oh == jcp.ih, |
| 1966 | jcp.od == jcp.id, jcp.stride_w == 1, |
| 1967 | jcp.stride_h == 1, jcp.stride_d == 1, jcp.ks == 1, |
| 1968 | !jcp.signed_input) |
| 1969 | ? (ptrdiff_t)jcp.ic * jcp.ks * jcp.os |
| 1970 | : 0; |
| 1971 | const size_t gemm_col_datatype_size |
| 1972 | = is_bf16_conv ? sizeof(bfloat16_t) : sizeof(float); |
| 1973 | |
| 1974 | // Potential scratchpad memory requirement when outer threading is |
| 1975 | // enabled during f32/bf16 BWD_W nspc convolution |
| 1976 | size_t thr_mem_estimate = max_threads |
| 1977 | * (gemm_col_datatype_size * jcp.im2col_sz |
| 1978 | + gemm_col_datatype_size * jcp.id * jcp.is * jcp.ic |
| 1979 | + sizeof(float) * weights_d.size()); |
| 1980 | if (is_bf16_conv) { |
| 1981 | thr_mem_estimate += sizeof(float) * weights_d.size(); |
| 1982 | if (jcp.with_bias |
| 1983 | && one_of(data_type::bf16, cd.diff_bias_desc.data_type, |
| 1984 | cd.bias_desc.data_type)) |
| 1985 | thr_mem_estimate += sizeof(float) * jcp.ngroups * jcp.oc; |
| 1986 | } |
| 1987 | const bool outer_threading_mem_ok |
| 1988 | = thr_mem_estimate < scratchpad_limit; |
| 1989 | |
| 1990 | jcp.outer_threading = outer_threading_mem_ok |
| 1991 | && jcp.os / max_threads < 256 |
| 1992 | && (jcp.mb != 1 || jcp.ngroups > 2); |
| 1993 | jcp.nthr = jcp.outer_threading ? max_threads : 1; |
| 1994 | |
| 1995 | scratchpad.book(key_conv_gemm_col, jcp.nthr * jcp.im2col_sz, |
| 1996 | gemm_col_datatype_size); |
| 1997 | |
| 1998 | jcp.need_wei_reduction = jcp.mb != 1 && jcp.nthr != 1; |
| 1999 | scratchpad.book<float>( |
| 2000 | key_conv_wei_reduction, jcp.nthr * weights_d.size()); |
| 2001 | scratchpad.book(key_conv_gemm_imtr, |
| 2002 | static_cast<size_t>(jcp.nthr) * jcp.id * jcp.is * jcp.ic, |
| 2003 | gemm_col_datatype_size); |
| 2004 | if (is_bf16_conv) { |
| 2005 | size_t conv_acc_buffer_size = weights_d.size(); |
| 2006 | scratchpad.book<float>( |
| 2007 | key_conv_int_dat_in_acc_dt, conv_acc_buffer_size); |
| 2008 | } |
| 2009 | if ((is_bf16_conv) && jcp.with_bias |
| 2010 | && one_of(data_type::bf16, cd.diff_bias_desc.data_type, |
| 2011 | cd.bias_desc.data_type)) |
| 2012 | scratchpad.book<float>( |
| 2013 | key_conv_bias_bf16_convert_wsp, jcp.ngroups * jcp.oc); |
| 2014 | } else if (!jcp.is_nspc && is_bwd_w) { |
| 2015 | // Potential scratchpad memory requirement when outer threading is |
| 2016 | // enabled during f32/bf16 BWD_W blocked convolution |
| 2017 | size_t thr_mem_estimate |
| 2018 | = sizeof(float) * max_threads * weights_d.size(); |
| 2019 | if (is_bf16_conv) { |
| 2020 | thr_mem_estimate += sizeof(float) * weights_d.size(); |
| 2021 | if (jcp.with_bias |
| 2022 | && one_of(data_type::bf16, cd.diff_bias_desc.data_type, |
| 2023 | cd.bias_desc.data_type)) |
| 2024 | thr_mem_estimate += sizeof(float) * jcp.ngroups * jcp.oc; |
| 2025 | } |
| 2026 | const size_t gemm_col_datatype_size |
| 2027 | = is_bf16_conv ? sizeof(bfloat16_t) : sizeof(float); |
| 2028 | // Minimum memory requirement as os_block >= simd_w |
| 2029 | thr_mem_estimate += gemm_col_datatype_size * max_threads * jcp.ic |
| 2030 | * jcp.ks * simd_w; |
| 2031 | |
| 2032 | const bool outer_threading_mem_ok |
| 2033 | = thr_mem_estimate < scratchpad_limit; |
| 2034 | jcp.outer_threading = outer_threading_mem_ok |
| 2035 | && jcp.os / max_threads < 256 |
| 2036 | && (jcp.mb != 1 || jcp.ngroups > 2); |
| 2037 | } |
| 2038 | |
| 2039 | if (!jcp.is_nspc) { |
| 2040 | jcp.nthr = jcp.outer_threading ? max_threads : 1; |
| 2041 | const int sizeof_cacheline_float = 16; |
| 2042 | if (is_bwd_w) { |
| 2043 | jcp.need_wei_reduction = jcp.mb != 1 && jcp.nthr != 1; |
| 2044 | scratchpad.book<float>( |
| 2045 | key_conv_wei_reduction, jcp.nthr * weights_d.size()); |
| 2046 | } |
| 2047 | |
| 2048 | if (is_bf16_conv) { |
| 2049 | size_t conv_acc_buffer_size = 0; |
| 2050 | if (is_fwd) |
| 2051 | conv_acc_buffer_size = jcp.nthr |
| 2052 | * rnd_up(jcp.oc_block * jcp.os_block, |
| 2053 | sizeof_cacheline_float); |
| 2054 | else if (is_bwd_d) |
| 2055 | conv_acc_buffer_size = jcp.nthr |
| 2056 | * rnd_up(jcp.ic * jcp.ih * jcp.iw * jcp.id, |
| 2057 | sizeof_cacheline_float); |
| 2058 | else if (is_bwd_w) |
| 2059 | conv_acc_buffer_size = weights_d.size(); |
| 2060 | scratchpad.book<float>( |
| 2061 | key_conv_int_dat_in_acc_dt, conv_acc_buffer_size); |
| 2062 | if ((is_fwd || is_bwd_w) && jcp.with_bias |
| 2063 | && one_of(data_type::bf16, cd.diff_bias_desc.data_type, |
| 2064 | cd.bias_desc.data_type)) |
| 2065 | scratchpad.book<float>(key_conv_bias_bf16_convert_wsp, |
| 2066 | jcp.ngroups * jcp.oc); |
| 2067 | } |
| 2068 | |
| 2069 | const size_t gemm_col_datatype_size = is_bf16_conv && !is_bwd_d |
| 2070 | ? sizeof(bfloat16_t) |
| 2071 | : sizeof(float); |
| 2072 | size_t gemm_col_memory_sz = jcp.nthr * jcp.im2col_sz; |
| 2073 | |
| 2074 | if (is_bwd_d || is_bwd_w) { |
| 2075 | // check available memory |
| 2076 | if (scratchpad_limit < scratchpad.size()) |
| 2077 | return status::unimplemented; |
| 2078 | const size_t available_mem |
| 2079 | = scratchpad_limit - scratchpad.size(); |
| 2080 | if (available_mem |
| 2081 | < gemm_col_memory_sz * gemm_col_datatype_size) { |
| 2082 | // Required memory in this scenario overflows the |
| 2083 | // available memory due to the large dimensions. |
| 2084 | const int min_os_block = simd_w; |
| 2085 | const int max_os_block = (int)available_mem |
| 2086 | / ((int)gemm_col_datatype_size * jcp.nthr |
| 2087 | * (jcp.im2col_sz / jcp.os)); |
| 2088 | // Choose an arbitrary small coeficient reduce spatial |
| 2089 | // dimensions. |
| 2090 | // TODO: better heuristic to determine os_block based |
| 2091 | // on cache efficiency |
| 2092 | float _coef = is_bwd_w ? 0.05 : 0.1; |
| 2093 | jcp.os_block = nstl::max( |
| 2094 | min_os_block, (int)(max_os_block * _coef)); |
| 2095 | jcp.os_nb_block = div_up(jcp.os, jcp.os_block); |
| 2096 | jcp.im2col_sz = (ptrdiff_t)jcp.ic * jcp.ks * jcp.os_block; |
| 2097 | gemm_col_memory_sz = jcp.nthr * jcp.im2col_sz; |
| 2098 | } |
| 2099 | } |
| 2100 | scratchpad.book(key_conv_gemm_col, gemm_col_memory_sz, |
| 2101 | gemm_col_datatype_size); |
| 2102 | } |
| 2103 | } |
| 2104 | |
| 2105 | jcp.bias_data_type = cd.bias_desc.data_type; |
| 2106 | jcp.dst_data_type = dst_md.data_type; |
| 2107 | jcp.sum_data_type = jcp.post_ops.get_sum_dt(jcp.dst_data_type); |
| 2108 | jcp.dst_os_stride = dst_d.is_blocking_desc() |
| 2109 | ? dst_d.blocking_desc().strides[ndims - 1] |
| 2110 | : 0; |
| 2111 | jcp.scale_idx_mult = (attr.scales_.get(DNNL_ARG_WEIGHTS).mask_ |
| 2112 | == (1 << (int)with_groups)); |
| 2113 | jcp.with_dst_scale = !attr.scales_.get(DNNL_ARG_DST).has_default_values(); |
| 2114 | book_precomputed_scales(scratchpad, attr.scales_, jcp.ngroups * jcp.oc); |
| 2115 | |
| 2116 | if (jcp.zp.src_exists) { |
| 2117 | const auto size = zp_src_comp_size + zp_src_pad_comp_size; |
| 2118 | if (size) scratchpad.book<int32_t>(key_conv_gemm_zp_src_comp, size); |
| 2119 | } |
| 2120 | |
| 2121 | if (scratchpad.size() > scratchpad_limit) return status::unimplemented; |
| 2122 | return status::success; |
| 2123 | } |
| 2124 | |
| 2125 | void bwd_weights_balance(int ithr, int nthr, int ngroups, int mb, int &ithr_g, |
| 2126 | int &nthr_g, int &ithr_mb, int &nthr_mb) { |
| 2127 | nthr_g = nstl::min(ngroups, nthr); |
| 2128 | nthr_mb = nstl::min(mb, nthr / nthr_g); |
| 2129 | if (ithr / nthr_mb >= ngroups) { |
| 2130 | ithr_g = ithr_mb = -1; |
| 2131 | } else { |
| 2132 | ithr_g = ithr / nthr_mb; |
| 2133 | ithr_mb = ithr % nthr_mb; |
| 2134 | } |
| 2135 | } |
| 2136 | |
| 2137 | void bwd_weights_reduction_par_ncsp(int ithr, int nthr, |
| 2138 | const conv_gemm_conf_t &jcp, const float *weights_reduce_ws, |
| 2139 | float *weights) { |
| 2140 | const size_t weights_g_size = jcp.ic * jcp.oc * jcp.ks; |
| 2141 | |
| 2142 | size_t weights_start {0}, weights_end {0}; |
| 2143 | balance211(weights_g_size, nthr, ithr, weights_start, weights_end); |
| 2144 | |
| 2145 | for (int i = 0; i < nthr; ++i) { |
| 2146 | const float *ws_i = weights_reduce_ws + i * weights_g_size; |
| 2147 | for (size_t s = weights_start; s < weights_end; ++s) |
| 2148 | weights[s] = (i == 0 ? 0 : weights[s]) + ws_i[s]; |
| 2149 | } |
| 2150 | } |
| 2151 | |
| 2152 | void bwd_weights_reduction_par_nspc(int ithr, int nthr, size_t g_start, |
| 2153 | size_t g_end, const conv_gemm_conf_t &jcp, |
| 2154 | const float *weights_reduce_base, float *diff_weights) { |
| 2155 | const dim_t weights_g_size = jcp.oc; |
| 2156 | dim_t weights_start {0}, weights_end {0}; |
| 2157 | balance211(jcp.ks * jcp.ic, nthr, ithr, weights_start, weights_end); |
| 2158 | |
| 2159 | // Threads divide work w.r.t. min-batch and groups, therefore |
| 2160 | // - weights_reduce_base format: spatial-input_channels-output_channels |
| 2161 | // - diff_weights format: spatial-input_channels-groups-output_channels |
| 2162 | for (auto tidx = 0; tidx < nthr; ++tidx) { |
| 2163 | const float *ws_base |
| 2164 | = weights_reduce_base + tidx * weights_g_size * jcp.ks * jcp.ic; |
| 2165 | for_(auto w = weights_start; w < weights_end; ++w) |
| 2166 | for (auto g = g_start; g < g_end; ++g) { |
| 2167 | float *__restrict dwei_ptr |
| 2168 | = diff_weights + (w * jcp.ngroups + g) * jcp.oc; |
| 2169 | const float *__restrict ws_ptr = ws_base + w * jcp.oc; |
| 2170 | if (tidx == 0) { |
| 2171 | PRAGMA_OMP_SIMD() |
| 2172 | for (auto oc = 0; oc < jcp.oc; ++oc) { |
| 2173 | dwei_ptr[oc] = ws_ptr[oc]; |
| 2174 | } |
| 2175 | } else { |
| 2176 | PRAGMA_OMP_SIMD() |
| 2177 | for (auto oc = 0; oc < jcp.oc; ++oc) { |
| 2178 | dwei_ptr[oc] += ws_ptr[oc]; |
| 2179 | } |
| 2180 | } |
| 2181 | } |
| 2182 | } |
| 2183 | } |
| 2184 | |
| 2185 | bool padding_exists(const conv_gemm_conf_t &jcp) noexcept { |
| 2186 | return jcp.l_pad || jcp.t_pad || jcp.f_pad || jcp.e_pad || jcp.b_pad |
| 2187 | || jcp.r_pad; |
| 2188 | } |
| 2189 | |
| 2190 | } // namespace jit_gemm_convolution_utils |
| 2191 | } // namespace cpu |
| 2192 | } // namespace impl |
| 2193 | } // namespace dnnl |
| 2194 |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.