| 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 "common/c_types_map.hpp" |
| 18 | #include "common/nstl.hpp" |
| 19 | #include "common/type_helpers.hpp" |
| 20 | #include "common/utils.hpp" |
| 21 | |
| 22 | #include "cpu/platform.hpp" |
| 23 | #include "cpu/x64/cpu_barrier.hpp" |
| 24 | #include "cpu/x64/injectors/injector_utils.hpp" |
| 25 | #include "cpu/x64/injectors/jit_uni_binary_injector.hpp" |
| 26 | #include "cpu/x64/injectors/jit_uni_eltwise_injector.hpp" |
| 27 | #include "cpu/x64/jit_avx512_common_conv_kernel.hpp" |
| 28 | |
| 29 | #define GET_OFF(field) offsetof(jit_conv_call_s, field) |
| 30 | #define KNx_L2_EFFECTIVE_CAPACITY ((512 - 64) * 1024) |
| 31 | |
| 32 | namespace dnnl { |
| 33 | namespace impl { |
| 34 | namespace cpu { |
| 35 | namespace x64 { |
| 36 | |
| 37 | using namespace dnnl::impl::format_tag; |
| 38 | using namespace dnnl::impl::memory_tracking::names; |
| 39 | using namespace dnnl::impl::utils; |
| 40 | using namespace Xbyak; |
| 41 | |
| 42 | namespace { |
| 43 | |
| 44 | constexpr auto small_spatial = 14; |
| 45 | |
| 46 | inline void pick_loop_order(jit_conv_conf_t &jcp) { |
| 47 | using namespace prop_kind; |
| 48 | assert(one_of( |
| 49 | jcp.prop_kind, forward_training, forward_inference, backward_data)); |
| 50 | auto w = (jcp.prop_kind == backward_data) ? jcp.iw : jcp.ow; |
| 51 | auto h = (jcp.prop_kind == backward_data) ? jcp.ih : jcp.oh; |
| 52 | |
| 53 | // The w in the loop order is currently ignored by 3D BWD_D |
| 54 | jcp.loop_order = (w <= small_spatial && h <= small_spatial) ? loop_cwgn |
| 55 | : loop_gncw; |
| 56 | if (utils::one_of(jcp.src_tag, format_tag::ndhwc, format_tag::nhwc, |
| 57 | format_tag::nwc) |
| 58 | && jcp.ngroups > 1 && jcp.oc < 16) |
| 59 | jcp.loop_order = loop_nhwcg; |
| 60 | } |
| 61 | |
| 62 | inline status_t init_tag(format_tag_t &tag, memory_desc_t &md, |
| 63 | const memory_desc_wrapper &mdw, const format_tag_t tag_value) { |
| 64 | if (mdw.format_kind() == format_kind::any) { |
| 65 | CHECK(memory_desc_init_by_tag(md, tag_value)); |
| 66 | tag = tag_value; |
| 67 | } else { |
| 68 | tag = mdw.matches_one_of_tag(tag_value); |
| 69 | } |
| 70 | |
| 71 | if (tag != tag_value) return status::unimplemented; |
| 72 | |
| 73 | return status::success; |
| 74 | } |
| 75 | |
| 76 | inline bool is_1stconv(const jit_conv_conf_t &jcp) { |
| 77 | if (mayiuse(avx512_core)) |
| 78 | return (jcp.ic < 16 && jcp.ngroups == 1); |
| 79 | else |
| 80 | return one_of(jcp.ic, 1, 3); |
| 81 | } |
| 82 | |
| 83 | inline bool is_ow_threading_on(const jit_conv_conf_t &jcp) { |
| 84 | return (jcp.nb_ow > 1); |
| 85 | } |
| 86 | |
| 87 | inline bool is_iw_threading_on(const jit_conv_conf_t &jcp) { |
| 88 | return (jcp.nb_iw > 1); |
| 89 | } |
| 90 | |
| 91 | } // namespace |
| 92 | |
| 93 | template <typename Vmm> |
| 94 | _jit_avx512_common_conv_fwd_kernel<Vmm>::_jit_avx512_common_conv_fwd_kernel( |
| 95 | const jit_conv_conf_t &ajcp, const primitive_attr_t &attr, |
| 96 | const memory_desc_t &dst_md) |
| 97 | : jit_generator(jit_name()), jcp(ajcp), attr_(attr) { |
| 98 | if (jcp.with_eltwise || jcp.with_binary) { |
| 99 | using namespace binary_injector; |
| 100 | static constexpr bool preserve_gpr = true; |
| 101 | static constexpr bool preserve_vmm = false; |
| 102 | static constexpr size_t helper_vmm_idx = 31; |
| 103 | const size_t tail_size = jcp.oc_without_padding % isa_simd_width_; |
| 104 | static constexpr bool use_exact_tail_scalar_bcast = false; |
| 105 | |
| 106 | const binary_injector::rhs_arg_static_params_t rhs_args_static_params { |
| 107 | helper_vmm_idx, reg_tmp, r15, r14, preserve_gpr, preserve_vmm, |
| 108 | GET_OFF(post_ops_binary_rhs_arg_vec), GET_OFF(dst_orig), |
| 109 | memory_desc_wrapper(dst_md), tail_size, postops_mask, |
| 110 | use_exact_tail_scalar_bcast}; |
| 111 | const binary_injector::static_params_t static_params { |
| 112 | this->param1, rhs_args_static_params}; |
| 113 | |
| 114 | postops_injector_ = utils::make_unique< |
| 115 | injector::jit_uni_postops_injector_t<avx512_core>>( |
| 116 | this, jcp.post_ops, static_params); |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | template <typename Vmm> |
| 121 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::prepare_output(int ur_w) { |
| 122 | for (int k = 0; k < jcp.nb_oc_blocking; k++) |
| 123 | for (int j = 0; j < ur_w; j++) { |
| 124 | Vmm vmm = vmm_out(j, k); |
| 125 | vpxord(vmm, vmm, vmm); |
| 126 | } |
| 127 | } |
| 128 | |
| 129 | template <typename F> |
| 130 | static void iterate(const int nb_oc_blocking, const int ur_w, |
| 131 | const bool oc_tail, const bool force_masking, const F &fun) { |
| 132 | for (int i_load = 0; i_load < nb_oc_blocking; i_load++) { |
| 133 | const auto mask_flag |
| 134 | = force_masking || (oc_tail && i_load + 1 == nb_oc_blocking); |
| 135 | for (int i_ur = 0; i_ur < ur_w; i_ur++) |
| 136 | fun(mask_flag, i_load, i_ur); |
| 137 | } |
| 138 | } |
| 139 | template <typename F> |
| 140 | static void iterate(const int nb_oc_blocking, const int ur_w, const F &fun) { |
| 141 | iterate(nb_oc_blocking, ur_w, false, false, fun); |
| 142 | } |
| 143 | |
| 144 | template <typename Vmm> |
| 145 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::apply_postops(int ur_w) { |
| 146 | injector_utils::vmm_index_set_t vmm_idxs; |
| 147 | if (jcp.with_binary) { |
| 148 | binary_injector::rhs_arg_dynamic_params_t rhs_arg_params; |
| 149 | const bool mask_tail = jcp.oc_without_padding % jcp.simd_w; |
| 150 | const bool oc_blk_is_smaller_than_vmm = jcp.oc_block < isa_simd_width_; |
| 151 | iterate(jcp.nb_oc_blocking, ur_w, mask_tail, oc_blk_is_smaller_than_vmm, |
| 152 | [&](const bool mask_flag, const int i_load, const int i_ur) { |
| 153 | const size_t aux_output_l_off |
| 154 | = get_output_offset(i_ur, i_load); |
| 155 | const auto vmm_idx = vmm_out_idx(i_ur, i_load); |
| 156 | vmm_idxs.emplace(vmm_idx); |
| 157 | |
| 158 | rhs_arg_params.vmm_idx_to_out_reg.emplace(vmm_idx, reg_out); |
| 159 | rhs_arg_params.vmm_idx_to_out_elem_off_val.emplace( |
| 160 | vmm_idx, aux_output_l_off); |
| 161 | if (mask_flag) { |
| 162 | rhs_arg_params.vmm_tail_idx_.emplace(vmm_idx); |
| 163 | } |
| 164 | }); |
| 165 | |
| 166 | postops_injector_->compute_vector_range(vmm_idxs, rhs_arg_params); |
| 167 | } else { |
| 168 | iterate(jcp.nb_oc_blocking, ur_w, |
| 169 | [&](const bool, const int i_load, const int i_ur) { |
| 170 | vmm_idxs.emplace(vmm_out_idx(i_ur, i_load)); |
| 171 | }); |
| 172 | postops_injector_->compute_vector_range(vmm_idxs); |
| 173 | } |
| 174 | } |
| 175 | |
| 176 | template <typename Vmm> |
| 177 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::store_output(int ur_w) { |
| 178 | Label no_update_label, store_label, post_ops_label; |
| 179 | |
| 180 | mov(reg_channel, ptr[param1 + GET_OFF(flags)]); |
| 181 | |
| 182 | if (jcp.with_bias) { mov(reg_bias, ptr[param1 + GET_OFF(bias)]); } |
| 183 | const int oc_tail = jcp.oc_tail; |
| 184 | |
| 185 | if (!jcp.with_sum) { |
| 186 | test(reg_channel, FLAG_IC_FIRST); |
| 187 | jnz(no_update_label, T_NEAR); |
| 188 | } |
| 189 | |
| 190 | for (int k = 0; k < jcp.nb_oc_blocking; k++) |
| 191 | for (int j = 0; j < ur_w; j++) { |
| 192 | Vmm vmm = vmm_out(j, k); |
| 193 | // mask only needed for last oc_block |
| 194 | if (oc_tail && k + 1 == jcp.nb_oc_blocking) |
| 195 | vmm = vmm | k_oc_tail_mask | T_z; |
| 196 | size_t aux_output_offset = get_output_offset(j, k); |
| 197 | vaddps(vmm, |
| 198 | make_safe_addr( |
| 199 | reg_out, aux_output_offset, reg_out_long_offt)); |
| 200 | } |
| 201 | |
| 202 | if (!jcp.with_sum) { |
| 203 | jmp(post_ops_label, T_NEAR); |
| 204 | } else { |
| 205 | test(reg_channel, FLAG_IC_FIRST); |
| 206 | jz(post_ops_label, T_NEAR); |
| 207 | } |
| 208 | |
| 209 | L(no_update_label); |
| 210 | if (jcp.with_bias) { |
| 211 | for (int k = 0; k < jcp.nb_oc_blocking; k++) { |
| 212 | int bias_offset = jcp.typesize_out * k * jcp.oc_block; |
| 213 | for (int j = 0; j < ur_w; j++) { |
| 214 | Vmm vmm = vmm_out(j, k); |
| 215 | // mask only needed for last oc_block |
| 216 | if (oc_tail && k + 1 == jcp.nb_oc_blocking) |
| 217 | vmm = vmm | k_oc_tail_mask | T_z; |
| 218 | vaddps(vmm, EVEX_compress_addr(reg_bias, bias_offset)); |
| 219 | } |
| 220 | } |
| 221 | } |
| 222 | |
| 223 | L(post_ops_label); |
| 224 | |
| 225 | if (jcp.with_eltwise || jcp.with_binary) { |
| 226 | test(reg_channel, FLAG_IC_LAST); |
| 227 | jz(store_label, T_NEAR); |
| 228 | |
| 229 | apply_postops(ur_w); |
| 230 | } |
| 231 | |
| 232 | L(store_label); |
| 233 | |
| 234 | const auto is_padding = jcp.oc_without_padding != jcp.oc; |
| 235 | for (int k = 0; k < jcp.nb_oc_blocking; k++) |
| 236 | for (int j = 0; j < ur_w; j++) { |
| 237 | Vmm vmm = vmm_out(j, k); |
| 238 | // mask only needed for last oc_block |
| 239 | if (oc_tail && k + 1 == jcp.nb_oc_blocking) { |
| 240 | if (is_padding) |
| 241 | vmovups(vmm | k_oc_tail_mask | T_z, vmm); |
| 242 | else |
| 243 | vmm = vmm | k_oc_tail_mask; |
| 244 | } |
| 245 | size_t aux_output_offset = get_output_offset(j, k); |
| 246 | |
| 247 | vmovups(EVEX_compress_addr_safe( |
| 248 | reg_out, aux_output_offset, reg_out_long_offt), |
| 249 | vmm); |
| 250 | } |
| 251 | } |
| 252 | |
| 253 | template <typename Vmm> |
| 254 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::compute_loop_fma( |
| 255 | int ur_w, int pad_l, int pad_r) { |
| 256 | const bool is_source_layout_nxc = is_src_layout_nxc(); |
| 257 | const bool icb_loop_in_compute_function = is_source_layout_nxc; |
| 258 | const int ic_tail = jcp.ic_tail; |
| 259 | const int oc_tail = jcp.oc == jcp.oc_without_padding ? jcp.oc_tail : 0; |
| 260 | int iw = jcp.iw; |
| 261 | int kw = jcp.kw; |
| 262 | int ic_block = jcp.ic_block; |
| 263 | int oc_block = jcp.oc_block; |
| 264 | int nb_oc_block = jcp.nb_oc_blocking; |
| 265 | Label kh_label, kd_label; |
| 266 | std::vector<Label> ic_tail_jmp(kw); |
| 267 | |
| 268 | // It seems that this compute_loop currently only handles one block of oc. |
| 269 | // assert if it is extended in future to catch unpadded_oc_tail. |
| 270 | assert(IMPLICATION(oc_tail, nb_oc_block == 1)); |
| 271 | |
| 272 | int num_ker_loads = ic_block * nb_oc_block * kw; |
| 273 | int ker_pipeline_depth |
| 274 | = oc_tail || ic_tail ? 1 : nstl::min(4, num_ker_loads); |
| 275 | assert(ker_reg_base_idx + ker_pipeline_depth <= 32); |
| 276 | assert(oc_block >= ker_pipeline_depth); |
| 277 | |
| 278 | int inp_mul = is_source_layout_nxc ? jcp.ngroups * jcp.ic |
| 279 | : (!jcp.is_1stconv ? ic_block : 1); |
| 280 | |
| 281 | if (one_of(jcp.ndims, 3, 4)) { |
| 282 | mov(aux_reg_inp, reg_inp); |
| 283 | mov(aux_reg_ker, reg_ker); |
| 284 | } |
| 285 | |
| 286 | if (jcp.ndims == 5) { |
| 287 | push(reg_out); |
| 288 | |
| 289 | mov(reg_ki, ptr[param1 + GET_OFF(kd_padding)]); |
| 290 | if (icb_loop_in_compute_function) { |
| 291 | // need to continue with the same kernel pointer, but as |
| 292 | // aux_reg_ker_d == reg_ker we need to save its value and restore |
| 293 | // it after kd loop |
| 294 | assert(aux_reg_ker_d == reg_ker); |
| 295 | push(aux_reg_ker_d); |
| 296 | } else |
| 297 | mov(aux_reg_ker_d, ptr[param1 + GET_OFF(filt)]); |
| 298 | mov(aux_reg_inp_d, reg_inp); |
| 299 | |
| 300 | L(kd_label); |
| 301 | mov(reg_kj, ptr[param1 + GET_OFF(kh_padding)]); |
| 302 | } else { |
| 303 | mov(reg_kj, reg_kh); |
| 304 | } |
| 305 | |
| 306 | if (jcp.ndims == 5) { |
| 307 | mov(aux_reg_inp, aux_reg_inp_d); |
| 308 | mov(aux_reg_ker, aux_reg_ker_d); |
| 309 | } |
| 310 | |
| 311 | align(16); |
| 312 | L(kh_label); |
| 313 | { |
| 314 | int step = 0; |
| 315 | for (int ki = 0; ki < kw; ki++) { |
| 316 | for (int ic = 0; ic < ic_block; ic++) { |
| 317 | if (ic_tail && ic >= ic_tail) { |
| 318 | // if src has only tails to compute, skip early |
| 319 | if (jcp.ic == ic_tail) |
| 320 | break; |
| 321 | else if (ic == ic_tail) { |
| 322 | cmp(reg_channel, ic_tail); |
| 323 | je(ic_tail_jmp[ki], T_NEAR); |
| 324 | } |
| 325 | } |
| 326 | int aux_kernel_offset = 0; |
| 327 | if (step == 0) { |
| 328 | for (int i = 0; i < ker_pipeline_depth; i++) { |
| 329 | aux_kernel_offset = get_kernel_offset(ki, ic, 0, i); |
| 330 | vmovups(vmm_ker(i), |
| 331 | EVEX_compress_addr( |
| 332 | aux_reg_ker, aux_kernel_offset)); |
| 333 | } |
| 334 | } else if (step < num_ker_loads - ker_pipeline_depth + 1) { |
| 335 | int load_offset = ker_pipeline_depth - 1; |
| 336 | int ker_load_reg_idx |
| 337 | = (step + load_offset) % ker_pipeline_depth; |
| 338 | aux_kernel_offset |
| 339 | = get_kernel_offset(ki, ic, 0, load_offset); |
| 340 | vmovups(vmm_ker(ker_load_reg_idx), |
| 341 | EVEX_compress_addr(aux_reg_ker, aux_kernel_offset)); |
| 342 | } |
| 343 | |
| 344 | Vmm vmm_kernel = vmm_ker(step % ker_pipeline_depth); |
| 345 | int j_start = get_ow_start(ki, pad_l); |
| 346 | int j_end = get_ow_end(ur_w, ki, pad_r); |
| 347 | for (int j = j_start; j < j_end; j++) { |
| 348 | size_t aux_input_offset |
| 349 | = get_input_offset(ki, ic, j, pad_l); |
| 350 | auto addr = EVEX_compress_addr_safe( |
| 351 | aux_reg_inp, aux_input_offset, reg_long_offt, true); |
| 352 | vfmadd231ps(vmm_out(j, 0), vmm_kernel, addr); |
| 353 | } |
| 354 | step++; |
| 355 | } |
| 356 | L(ic_tail_jmp[ki]); |
| 357 | } |
| 358 | int ker_shift = jcp.typesize_in * kw * oc_block * ic_block; |
| 359 | add(aux_reg_ker, ker_shift); |
| 360 | int inp_shift = jcp.typesize_in * (jcp.dilate_h + 1) * iw * inp_mul; |
| 361 | add(aux_reg_inp, inp_shift); |
| 362 | dec(reg_kj); |
| 363 | cmp(reg_kj, 0); |
| 364 | jg(kh_label, T_NEAR); |
| 365 | } |
| 366 | |
| 367 | if (jcp.ndims == 5) { |
| 368 | int inp_shift |
| 369 | = typesize * (jcp.dilate_d + 1) * jcp.ih * jcp.iw * inp_mul; |
| 370 | add(aux_reg_inp_d, inp_shift); |
| 371 | int ker_shift |
| 372 | = typesize * jcp.kw * jcp.kh * jcp.oc_block * jcp.ic_block; |
| 373 | add(aux_reg_ker_d, ker_shift); |
| 374 | |
| 375 | dec(reg_ki); |
| 376 | cmp(reg_ki, 0); |
| 377 | jg(kd_label, T_NEAR); |
| 378 | |
| 379 | if (icb_loop_in_compute_function) pop(aux_reg_ker_d); |
| 380 | pop(reg_out); |
| 381 | } |
| 382 | } |
| 383 | |
| 384 | template <typename Vmm> |
| 385 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::compute_loop_fma_core( |
| 386 | int ur_w, int pad_l, int pad_r) { |
| 387 | int kw = jcp.kw; |
| 388 | int ic_block = jcp.ic_block; |
| 389 | int oc_block = jcp.oc_block; |
| 390 | int nb_oc_block = jcp.nb_oc_blocking; |
| 391 | const bool is_source_layout_nxc = is_src_layout_nxc(); |
| 392 | const bool icb_loop_in_compute_function = is_source_layout_nxc; |
| 393 | const int ic_tail = jcp.ic_tail; |
| 394 | |
| 395 | Label kh_label, kd_label; |
| 396 | std::vector<Label> ic_tail_jmp(kw); |
| 397 | int shift_kernel_ptr |
| 398 | = jcp.typesize_in * jcp.kw * jcp.oc_block * jcp.ic_block; |
| 399 | int inp_mul = is_source_layout_nxc ? jcp.ngroups * jcp.ic |
| 400 | : (!jcp.is_1stconv ? ic_block : 1); |
| 401 | |
| 402 | int shift_input_ptr |
| 403 | = jcp.typesize_in * (jcp.dilate_h + 1) * jcp.iw * inp_mul; |
| 404 | |
| 405 | if (one_of(jcp.ndims, 3, 4)) { |
| 406 | mov(aux_reg_inp, reg_inp); |
| 407 | mov(aux_reg_ker, reg_ker); |
| 408 | } |
| 409 | |
| 410 | if (jcp.ndims == 5) { |
| 411 | push(reg_out); |
| 412 | |
| 413 | mov(reg_ki, ptr[param1 + GET_OFF(kd_padding)]); |
| 414 | if (icb_loop_in_compute_function) { |
| 415 | // need to continue with the same kernel pointer, but as |
| 416 | // aux_reg_ker_d == reg_ker we need to save its value and restore |
| 417 | // it after kd loop |
| 418 | assert(aux_reg_ker_d == reg_ker); |
| 419 | push(aux_reg_ker_d); |
| 420 | } else |
| 421 | mov(aux_reg_ker_d, ptr[param1 + GET_OFF(filt)]); |
| 422 | |
| 423 | mov(aux_reg_inp_d, reg_inp); |
| 424 | |
| 425 | L(kd_label); |
| 426 | mov(reg_kj, ptr[param1 + GET_OFF(kh_padding)]); |
| 427 | } else { |
| 428 | mov(reg_kj, reg_kh); |
| 429 | } |
| 430 | |
| 431 | if (jcp.ndims == 5) { |
| 432 | mov(aux_reg_inp, aux_reg_inp_d); |
| 433 | mov(aux_reg_ker, aux_reg_ker_d); |
| 434 | } |
| 435 | |
| 436 | L(kh_label); |
| 437 | { |
| 438 | for (int ki = 0; ki < kw; ki++) { |
| 439 | int jj_start = get_ow_start(ki, pad_l); |
| 440 | int jj_end = get_ow_end(ur_w, ki, pad_r); |
| 441 | for (int ic = 0; ic < ic_block; ic++) { |
| 442 | if (ic_tail && ic >= ic_tail) { |
| 443 | // if src has only tails to compute, skip early |
| 444 | if (jcp.ic == ic_tail) |
| 445 | break; |
| 446 | else if (ic == ic_tail) { |
| 447 | cmp(reg_channel, ic_tail); |
| 448 | je(ic_tail_jmp[ki], T_NEAR); |
| 449 | } |
| 450 | } |
| 451 | if (jcp.kernel_kind == expl_bcast) { |
| 452 | for (int jj = jj_start; jj < jj_end; jj++) { |
| 453 | size_t aux_input_offset |
| 454 | = get_input_offset(ki, ic, jj, pad_l); |
| 455 | vbroadcastss(vmm_inp(jj, nb_oc_block), |
| 456 | EVEX_compress_addr_safe(aux_reg_inp, |
| 457 | aux_input_offset, reg_long_offt)); |
| 458 | } |
| 459 | } |
| 460 | for (int ii = 0; ii < nb_oc_block; ii++) { |
| 461 | int aux_kernel_offset = jcp.typesize_in |
| 462 | * (ii * jcp.nb_ic * jcp.kh * jcp.kw * jcp.kd |
| 463 | * ic_block * oc_block |
| 464 | + ki * ic_block * oc_block + ic * oc_block); |
| 465 | if (jj_end - jj_start > 0) |
| 466 | vmovups(vmm_wei, |
| 467 | EVEX_compress_addr( |
| 468 | aux_reg_ker, aux_kernel_offset)); |
| 469 | for (int jj = jj_start; jj < jj_end; jj++) |
| 470 | if (jcp.kernel_kind == expl_bcast) |
| 471 | vfmadd231ps(vmm_out(jj, ii), |
| 472 | vmm_inp(jj, nb_oc_block), vmm_wei); |
| 473 | else { |
| 474 | size_t aux_input_offset |
| 475 | = get_input_offset(ki, ic, jj, pad_l); |
| 476 | vfmadd231ps(vmm_out(jj, ii), vmm_wei, |
| 477 | EVEX_compress_addr_safe(aux_reg_inp, |
| 478 | aux_input_offset, reg_long_offt, |
| 479 | true)); |
| 480 | } |
| 481 | } |
| 482 | } |
| 483 | L(ic_tail_jmp[ki]); |
| 484 | } |
| 485 | add(aux_reg_ker, shift_kernel_ptr); |
| 486 | add(aux_reg_inp, shift_input_ptr); |
| 487 | dec(reg_kj); |
| 488 | cmp(reg_kj, 0); |
| 489 | jg(kh_label, T_NEAR); |
| 490 | } |
| 491 | |
| 492 | if (jcp.ndims == 5) { |
| 493 | add(aux_reg_inp_d, |
| 494 | typesize * (jcp.dilate_d + 1) * jcp.ih * jcp.iw * inp_mul); |
| 495 | const int ker_shift |
| 496 | = typesize * jcp.kw * jcp.kh * jcp.oc_block * jcp.ic_block; |
| 497 | add(aux_reg_ker_d, ker_shift); |
| 498 | |
| 499 | dec(reg_ki); |
| 500 | cmp(reg_ki, 0); |
| 501 | jg(kd_label, T_NEAR); |
| 502 | |
| 503 | if (icb_loop_in_compute_function) pop(aux_reg_ker_d); |
| 504 | pop(reg_out); |
| 505 | } |
| 506 | } |
| 507 | |
| 508 | template <typename Vmm> |
| 509 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::compute_loop( |
| 510 | int ur_w, int pad_l, int pad_r) { |
| 511 | if (jcp.ndims == 5) push(reg_oi); |
| 512 | |
| 513 | prepare_output(ur_w); |
| 514 | |
| 515 | Label skip_compute_loop; |
| 516 | if (jcp.ndims == 5) { |
| 517 | if ((jcp.dilate_d >= jcp.id) |
| 518 | || (jcp.kd - 1) * (jcp.dilate_d + 1) |
| 519 | < nstl::max(jcp.f_pad, jcp.back_pad)) { |
| 520 | mov(reg_kj, ptr[param1 + GET_OFF(kd_padding)]); |
| 521 | cmp(reg_kj, 0); |
| 522 | jle(skip_compute_loop, T_NEAR); |
| 523 | } |
| 524 | } |
| 525 | if ((jcp.dilate_h >= jcp.ih) |
| 526 | || (jcp.kh - 1) * (jcp.dilate_h + 1) |
| 527 | < nstl::max(jcp.t_pad, jcp.b_pad)) { |
| 528 | mov(reg_kj, ptr[param1 + GET_OFF(kh_padding)]); |
| 529 | cmp(reg_kj, 0); |
| 530 | jle(skip_compute_loop, T_NEAR); |
| 531 | } |
| 532 | |
| 533 | Label ic_loop; |
| 534 | const bool generate_icb_loop = jcp.nb_ic > 1 && is_src_layout_nxc(); |
| 535 | if (generate_icb_loop) { |
| 536 | push(reg_inp); |
| 537 | push(reg_ker); |
| 538 | |
| 539 | mov(reg_channel, ptr[param1 + GET_OFF(reduce_work)]); |
| 540 | L(ic_loop); |
| 541 | } |
| 542 | |
| 543 | if (jcp.is_1stconv && jcp.kernel_kind != expl_bcast) |
| 544 | compute_loop_fma(ur_w, pad_l, pad_r); |
| 545 | else if (jcp.kernel_kind == embd_bcast && jcp.nb_oc_blocking == 1) |
| 546 | compute_loop_fma(ur_w, pad_l, pad_r); |
| 547 | else |
| 548 | compute_loop_fma_core(ur_w, pad_l, pad_r); |
| 549 | |
| 550 | if (generate_icb_loop) { |
| 551 | assert(is_src_layout_nxc()); |
| 552 | const int inp_shift = jcp.ic_block * jcp.typesize_in; |
| 553 | add(reg_inp, inp_shift); |
| 554 | const size_t ker_shift = (size_t)jcp.kd * jcp.kh * jcp.kw * jcp.ic_block |
| 555 | * jcp.oc_block * jcp.typesize_in; |
| 556 | safe_add(reg_ker, ker_shift, reg_ker_long_offt); |
| 557 | sub(reg_channel, jcp.ic_block); |
| 558 | jg(ic_loop, T_NEAR); |
| 559 | |
| 560 | pop(reg_ker); |
| 561 | pop(reg_inp); |
| 562 | } |
| 563 | |
| 564 | L(skip_compute_loop); |
| 565 | store_output(ur_w); |
| 566 | if (jcp.ndims == 5) pop(reg_oi); |
| 567 | } |
| 568 | |
| 569 | template <typename Vmm> |
| 570 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::generate() { |
| 571 | int iw = jcp.iw; |
| 572 | int ow = jcp.ow; |
| 573 | int ow_block = jcp.ow_block; |
| 574 | int nb_ow = jcp.nb_ow; |
| 575 | int kw = jcp.kw; |
| 576 | int l_pad = jcp.l_pad; |
| 577 | int ur_w = jcp.ur_w; |
| 578 | int ur_w_tail = jcp.ur_w_tail; |
| 579 | int stride_w = jcp.stride_w; |
| 580 | |
| 581 | int inp_mult = is_src_layout_nxc() ? jcp.ngroups * jcp.ic |
| 582 | : (jcp.is_1stconv ? 1 : jcp.ic_block); |
| 583 | int inp_shift_pad = jcp.typesize_in * (ur_w * stride_w - l_pad) * inp_mult; |
| 584 | int inp_shift = jcp.typesize_in * ur_w * stride_w * inp_mult; |
| 585 | int inp_shift_pad_second_block = -1 * jcp.typesize_in * l_pad * inp_mult; |
| 586 | int out_shift = jcp.typesize_out * ur_w |
| 587 | * (is_dst_layout_nxc() ? jcp.ngroups * jcp.oc : jcp.oc_block); |
| 588 | |
| 589 | preamble(); |
| 590 | mov(reg_inp, ptr[param1 + GET_OFF(src)]); |
| 591 | mov(reg_out, ptr[param1 + GET_OFF(dst)]); |
| 592 | mov(reg_ker, ptr[param1 + GET_OFF(filt)]); |
| 593 | mov(reg_kh, ptr[param1 + GET_OFF(kh_padding)]); |
| 594 | |
| 595 | const int oc_tail = jcp.oc_tail; |
| 596 | if (oc_tail) { |
| 597 | Label done; |
| 598 | // dummy mask all 1's |
| 599 | kxnorw(k_oc_tail_mask, k_oc_tail_mask, k_oc_tail_mask); |
| 600 | mov(reg_load_work, ptr[param1 + GET_OFF(load_work)]); |
| 601 | cmp(reg_load_work, jcp.nb_oc_blocking * jcp.oc_block); |
| 602 | je(done, T_NEAR); |
| 603 | Reg32 reg_tail_32 = reg_tail.cvt32(); |
| 604 | mov(reg_tail_32, (1 << oc_tail) - 1); |
| 605 | kmovw(k_oc_tail_mask, reg_tail_32); |
| 606 | L(done); |
| 607 | kmovw(postops_mask, k_oc_tail_mask); |
| 608 | } else if (jcp.with_binary) |
| 609 | if (jcp.oc_block != isa_simd_width_) { |
| 610 | const int mask = (1 << jcp.oc_block) - 1; |
| 611 | const Reg32 reg_tail_32 = reg_tail.cvt32(); |
| 612 | mov(reg_tail_32, mask); |
| 613 | kmovw(postops_mask, reg_tail_32); |
| 614 | } |
| 615 | |
| 616 | int r_pad = nstl::max(0, jcp.r_pad); |
| 617 | int n_oi = ow / ur_w; |
| 618 | int r_pad1 = calculate_end_padding(l_pad, ur_w * n_oi, iw, stride_w, |
| 619 | calculate_extended_filter_size(kw, jcp.dilate_w)); |
| 620 | |
| 621 | if (!is_ow_threading_on(jcp)) { |
| 622 | // ow is being processed as a whole - with left and right paddings |
| 623 | if (r_pad1 > 0) n_oi--; |
| 624 | |
| 625 | if (ow == ur_w) { |
| 626 | compute_loop(ur_w, l_pad, r_pad); |
| 627 | } else { |
| 628 | if (n_oi == 0) { |
| 629 | compute_loop(ur_w, l_pad, r_pad1); |
| 630 | add(reg_inp, inp_shift_pad); |
| 631 | add(reg_out, out_shift); |
| 632 | if (ur_w_tail != 0) { compute_loop(ur_w_tail, 0, r_pad); } |
| 633 | } else { |
| 634 | xor_(reg_oi, reg_oi); |
| 635 | if (l_pad > 0) { |
| 636 | compute_loop(ur_w, l_pad, 0); |
| 637 | add(reg_inp, inp_shift_pad); |
| 638 | add(reg_out, out_shift); |
| 639 | inc(reg_oi); |
| 640 | } |
| 641 | if ((l_pad <= 0 && n_oi > 0) || (l_pad > 0 && n_oi > 1)) { |
| 642 | Label ow_loop_label; |
| 643 | L(ow_loop_label); |
| 644 | { |
| 645 | compute_loop(ur_w, 0, 0); |
| 646 | add(reg_inp, inp_shift); |
| 647 | add(reg_out, out_shift); |
| 648 | inc(reg_oi); |
| 649 | cmp(reg_oi, n_oi); |
| 650 | jl(ow_loop_label, T_NEAR); |
| 651 | } |
| 652 | } |
| 653 | if (r_pad1 > 0) { |
| 654 | compute_loop(ur_w, 0, r_pad1); |
| 655 | add(reg_inp, inp_shift); |
| 656 | add(reg_out, out_shift); |
| 657 | } |
| 658 | if (ur_w_tail != 0) { compute_loop(ur_w_tail, 0, r_pad); } |
| 659 | } |
| 660 | } |
| 661 | } else { |
| 662 | // ow block is only processed. |
| 663 | // Number of block is passed as parameter owb, |
| 664 | // and padding processing depends on this number. |
| 665 | |
| 666 | Label end_label, last_oi_label, middle_ow_blocks_label, tail_label; |
| 667 | Label oi_loop_label, oi_loop_start_label, oi_loop_end_label; |
| 668 | |
| 669 | assert(ow_block % ur_w == 0); |
| 670 | int n_oi_not_last_ow_block = ow_block / ur_w; |
| 671 | // to simplify code (and general regs usage), |
| 672 | // size of ow block must be >= 2 * ur_w |
| 673 | assert(n_oi_not_last_ow_block > 1); |
| 674 | int n_oi_next_last_ow_block = n_oi_not_last_ow_block; |
| 675 | int n_oi_first_ow_block = n_oi_not_last_ow_block; |
| 676 | |
| 677 | int n_oi_last_ow_block = (ow - ow_block * (nb_ow - 1)) / ur_w; |
| 678 | |
| 679 | // prepare right padding |
| 680 | bool next_last_ow_block_padded = r_pad1 > 0 && n_oi_last_ow_block == 0; |
| 681 | bool first_ow_block_padded |
| 682 | = next_last_ow_block_padded && jcp.nb_ow == 2; |
| 683 | bool last_ow_block_padded = r_pad1 > 0 && n_oi_last_ow_block > 0; |
| 684 | |
| 685 | if (last_ow_block_padded) |
| 686 | n_oi_last_ow_block--; |
| 687 | else if (first_ow_block_padded) |
| 688 | n_oi_first_ow_block--; |
| 689 | else if (next_last_ow_block_padded) |
| 690 | n_oi_next_last_ow_block--; |
| 691 | |
| 692 | mov(reg_owb, ptr[param1 + GET_OFF(owb)]); |
| 693 | cmp(reg_owb, 0); // is that the first ow-block ? |
| 694 | jg(middle_ow_blocks_label, T_NEAR); |
| 695 | |
| 696 | // the first ow block, compute left padding |
| 697 | |
| 698 | mov(reg_oi, n_oi_first_ow_block); |
| 699 | |
| 700 | if (l_pad > 0) { |
| 701 | compute_loop(ur_w, l_pad, 0); |
| 702 | add(reg_inp, inp_shift_pad); |
| 703 | add(reg_out, out_shift); |
| 704 | dec(reg_oi); |
| 705 | } |
| 706 | jmp(oi_loop_label, T_NEAR); |
| 707 | |
| 708 | // middle or last ow block entry |
| 709 | |
| 710 | L(middle_ow_blocks_label); |
| 711 | |
| 712 | if (l_pad > 0) { |
| 713 | // just to consider left padding, not compute |
| 714 | add(reg_inp, inp_shift_pad_second_block); |
| 715 | } |
| 716 | |
| 717 | // set number of iteration for oi-loop |
| 718 | cmp(reg_owb, jcp.nb_ow - 1); // last ow-block ? |
| 719 | mov(reg_oi, n_oi_last_ow_block); |
| 720 | je(oi_loop_label, T_NEAR); |
| 721 | cmp(reg_owb, jcp.nb_ow - 2); // next to last ow-block ? |
| 722 | mov(reg_oi, n_oi_next_last_ow_block); |
| 723 | je(oi_loop_label, T_NEAR); |
| 724 | mov(reg_oi, n_oi_not_last_ow_block); // other middle ow-blocks |
| 725 | |
| 726 | // oi loop w/o padding |
| 727 | L(oi_loop_label); |
| 728 | L(oi_loop_start_label); |
| 729 | cmp(reg_oi, 0); |
| 730 | jle(oi_loop_end_label, T_NEAR); |
| 731 | |
| 732 | compute_loop(ur_w, 0, 0); |
| 733 | add(reg_inp, inp_shift); |
| 734 | add(reg_out, out_shift); |
| 735 | dec(reg_oi); |
| 736 | jmp(oi_loop_start_label, T_NEAR); |
| 737 | L(oi_loop_end_label); |
| 738 | |
| 739 | mov(reg_owb, ptr[param1 + GET_OFF(owb)]); |
| 740 | |
| 741 | cmp(reg_owb, 0); // first ow-block ? |
| 742 | if (first_ow_block_padded) { |
| 743 | je(last_oi_label, T_NEAR); |
| 744 | } else { |
| 745 | je(end_label, T_NEAR); |
| 746 | } |
| 747 | cmp(reg_owb, jcp.nb_ow - 2); // next to last ow-block ? |
| 748 | jl(end_label, T_NEAR); |
| 749 | if (next_last_ow_block_padded) { |
| 750 | je(last_oi_label, T_NEAR); |
| 751 | } else { |
| 752 | je(end_label, T_NEAR); |
| 753 | } |
| 754 | // that is last block |
| 755 | if (!last_ow_block_padded) { jmp(tail_label, T_NEAR); } |
| 756 | |
| 757 | // last oi block with right padding |
| 758 | L(last_oi_label); |
| 759 | compute_loop(ur_w, 0, r_pad1); |
| 760 | add(reg_inp, inp_shift); |
| 761 | add(reg_out, out_shift); |
| 762 | |
| 763 | mov(reg_owb, ptr[param1 + GET_OFF(owb)]); |
| 764 | cmp(reg_owb, jcp.nb_ow - 1); // last ow_block? |
| 765 | jl(end_label, T_NEAR); |
| 766 | |
| 767 | L(tail_label); |
| 768 | if (ur_w_tail != 0) { compute_loop(ur_w_tail, 0, r_pad); } |
| 769 | L(end_label); |
| 770 | } |
| 771 | postamble(); |
| 772 | |
| 773 | if (jcp.with_eltwise) postops_injector_->prepare_table(); |
| 774 | } |
| 775 | |
| 776 | status_t jit_avx512_common_conv_fwd_kernel::init_conf(jit_conv_conf_t &jcp, |
| 777 | const convolution_desc_t &cd, memory_desc_t &src_md, |
| 778 | memory_desc_t &weights_md, memory_desc_t &dst_md, |
| 779 | memory_desc_t &bias_md, primitive_attr_t &attr, int nthreads) { |
| 780 | using namespace prop_kind; |
| 781 | |
| 782 | if (!mayiuse(avx512_core)) return status::unimplemented; |
| 783 | |
| 784 | const memory_desc_wrapper src_d(&src_md); |
| 785 | const memory_desc_wrapper weights_d(&weights_md); |
| 786 | const memory_desc_wrapper dst_d(&dst_md); |
| 787 | const memory_desc_wrapper bias_d(&bias_md); |
| 788 | |
| 789 | if (!everyone_is(data_type::f32, src_d.data_type(), weights_d.data_type(), |
| 790 | dst_d.data_type())) |
| 791 | return status::unimplemented; |
| 792 | |
| 793 | const int regs = 28; |
| 794 | const bool with_groups = weights_d.ndims() == src_d.ndims() + 1; |
| 795 | int ndims = src_d.ndims(); |
| 796 | |
| 797 | jcp = zero<decltype(jcp)>(); |
| 798 | jcp.nthr = jcp.aligned_threads = nthreads; |
| 799 | jcp.ndims = ndims; |
| 800 | jcp.prop_kind = cd.prop_kind; |
| 801 | jcp.ngroups = with_groups ? weights_d.dims()[0] : 1; |
| 802 | jcp.mb = src_d.dims()[0]; |
| 803 | jcp.oc = dst_d.dims()[1] / jcp.ngroups; |
| 804 | jcp.oc_without_padding = jcp.oc; |
| 805 | jcp.ic = src_d.dims()[1] / jcp.ngroups; |
| 806 | jcp.ic_without_padding = jcp.ic; |
| 807 | jcp.id = (ndims == 5) ? src_d.dims()[2] : 1; |
| 808 | jcp.ih = (ndims == 3) ? 1 : src_d.dims()[ndims - 2]; |
| 809 | jcp.iw = src_d.dims()[ndims - 1]; |
| 810 | jcp.od = (ndims == 5) ? dst_d.dims()[2] : 1; |
| 811 | jcp.oh = (ndims == 3) ? 1 : dst_d.dims()[ndims - 2]; |
| 812 | jcp.ow = dst_d.dims()[ndims - 1]; |
| 813 | jcp.kd = (ndims == 5) ? weights_d.dims()[with_groups + 2] : 1; |
| 814 | jcp.kh = (ndims == 3) ? 1 : weights_d.dims()[with_groups + ndims - 2]; |
| 815 | jcp.kw = weights_d.dims()[with_groups + ndims - 1]; |
| 816 | jcp.f_pad = (ndims == 5) ? cd.padding[0][0] : 0; |
| 817 | jcp.t_pad = (ndims == 3) ? 0 : cd.padding[0][ndims - 4]; |
| 818 | jcp.l_pad = cd.padding[0][ndims - 3]; |
| 819 | jcp.stride_d = (ndims == 5) ? cd.strides[0] : 1; |
| 820 | jcp.stride_h = (ndims == 3) ? 1 : cd.strides[ndims - 4]; |
| 821 | jcp.stride_w = cd.strides[ndims - 3]; |
| 822 | |
| 823 | jcp.dilate_d = (ndims == 5) ? cd.dilates[0] : 0; |
| 824 | jcp.dilate_h = (ndims == 3) ? 0 : cd.dilates[ndims - 4]; |
| 825 | jcp.dilate_w = cd.dilates[ndims - 3]; |
| 826 | |
| 827 | int ext_kw = calculate_extended_filter_size(jcp.kw, jcp.dilate_w); |
| 828 | int ext_kh = calculate_extended_filter_size(jcp.kh, jcp.dilate_h); |
| 829 | int ext_kd = calculate_extended_filter_size(jcp.kd, jcp.dilate_d); |
| 830 | jcp.r_pad = calculate_end_padding( |
| 831 | jcp.l_pad, jcp.ow, jcp.iw, jcp.stride_w, ext_kw); |
| 832 | jcp.b_pad = calculate_end_padding( |
| 833 | jcp.t_pad, jcp.oh, jcp.ih, jcp.stride_h, ext_kh); |
| 834 | jcp.back_pad = calculate_end_padding( |
| 835 | jcp.f_pad, jcp.od, jcp.id, jcp.stride_d, ext_kd); |
| 836 | bool kernel_outside_src = false || ext_kw <= jcp.l_pad |
| 837 | || ext_kw <= jcp.r_pad || ext_kh <= jcp.t_pad || ext_kh <= jcp.b_pad |
| 838 | || ext_kd <= jcp.f_pad || ext_kd <= jcp.back_pad; |
| 839 | if (kernel_outside_src) return status::unimplemented; |
| 840 | |
| 841 | const auto dat_tag_nxc = pick(ndims - 3, nwc, nhwc, ndhwc); |
| 842 | const auto dat_tag_ncx = pick(ndims - 3, ncw, nchw, ncdhw); |
| 843 | const auto dat_tag_nCx4c = pick(ndims - 3, nCw4c, nChw4c, nCdhw4c); |
| 844 | const auto dat_tag_nCx8c = pick(ndims - 3, nCw8c, nChw8c, nCdhw8c); |
| 845 | const auto dat_tag_nCx16c = pick(ndims - 3, nCw16c, nChw16c, nCdhw16c); |
| 846 | auto curr_src_tag = src_d.matches_one_of_tag(dat_tag_nxc, dat_tag_nCx16c, |
| 847 | dat_tag_nCx8c, dat_tag_nCx4c, dat_tag_ncx); |
| 848 | auto curr_dst_tag = dst_d.matches_one_of_tag( |
| 849 | dat_tag_nxc, dat_tag_nCx16c, dat_tag_nCx8c, dat_tag_nCx4c); |
| 850 | bool is_data_layout_nxc = IMPLICATION(curr_src_tag != dat_tag_nxc, |
| 851 | src_d.format_kind() == format_kind::any) |
| 852 | && IMPLICATION(curr_dst_tag != dat_tag_nxc, |
| 853 | dst_d.format_kind() == format_kind::any) |
| 854 | && utils::one_of(dat_tag_nxc, curr_src_tag, curr_dst_tag); |
| 855 | |
| 856 | jcp.is_1stconv = is_1stconv(jcp); |
| 857 | |
| 858 | bool ok_to_pad_channels = true && !is_data_layout_nxc && jcp.ngroups == 1 |
| 859 | && src_d.data_type() == data_type::f32; |
| 860 | |
| 861 | const int full_simd_w = cpu_isa_traits<avx512_core>::vlen / typesize; |
| 862 | jcp.simd_w = full_simd_w; |
| 863 | bool ok_to_try_lower_zmm = true |
| 864 | && IMPLICATION(is_data_layout_nxc, |
| 865 | jcp.oc < full_simd_w && jcp.ic < full_simd_w |
| 866 | && jcp.ngroups > 1) |
| 867 | && mayiuse(avx512_core) && src_d.data_type() == data_type::f32 |
| 868 | && !jcp.is_1stconv && !ok_to_pad_channels |
| 869 | && (jcp.ic % jcp.simd_w != 0 || jcp.oc % jcp.simd_w != 0); |
| 870 | |
| 871 | if (ok_to_try_lower_zmm) { |
| 872 | for (auto simd : {8, 4}) { |
| 873 | if (jcp.ic % simd == 0 && jcp.oc % simd == 0) { |
| 874 | jcp.simd_w = simd; |
| 875 | break; |
| 876 | } |
| 877 | } |
| 878 | } |
| 879 | |
| 880 | jcp.oc_block = jcp.simd_w; |
| 881 | jcp.ic_block = jcp.is_1stconv ? jcp.ic : jcp.simd_w; |
| 882 | |
| 883 | if (ok_to_pad_channels) { |
| 884 | jcp.oc = rnd_up(jcp.oc, jcp.oc_block); |
| 885 | jcp.ic = rnd_up(jcp.ic, jcp.ic_block); |
| 886 | } |
| 887 | if (!IMPLICATION(!is_data_layout_nxc, |
| 888 | jcp.oc % jcp.oc_block == 0 && jcp.ic % jcp.ic_block == 0)) |
| 889 | return status::unimplemented; |
| 890 | |
| 891 | jcp.ic_tail = is_data_layout_nxc ? jcp.ic % jcp.simd_w : 0; |
| 892 | jcp.oc_tail = jcp.oc_without_padding % jcp.simd_w; |
| 893 | |
| 894 | format_tag_t src_tag, dst_tag, wei_tag; |
| 895 | |
| 896 | if (jcp.simd_w == 8) { |
| 897 | assert(with_groups); |
| 898 | src_tag = is_data_layout_nxc ? dat_tag_nxc : dat_tag_nCx8c; |
| 899 | dst_tag = src_tag; |
| 900 | wei_tag = pick(ndims - 3, gOIw8i8o, gOIhw8i8o, gOIdhw8i8o); |
| 901 | } else if (jcp.simd_w == 4) { |
| 902 | assert(with_groups); |
| 903 | src_tag = is_data_layout_nxc ? dat_tag_nxc : dat_tag_nCx4c; |
| 904 | dst_tag = src_tag; |
| 905 | wei_tag = pick(ndims - 3, gOIw4i4o, gOIhw4i4o, gOIdhw4i4o); |
| 906 | } else { |
| 907 | dst_tag = is_data_layout_nxc ? dat_tag_nxc : dat_tag_nCx16c; |
| 908 | src_tag = is_data_layout_nxc |
| 909 | ? dat_tag_nxc |
| 910 | : (jcp.is_1stconv ? dat_tag_ncx : dat_tag_nCx16c); |
| 911 | wei_tag = pick(2 * ndims - 6 + with_groups, OIw16i16o, gOIw16i16o, |
| 912 | OIhw16i16o, gOIhw16i16o, OIdhw16i16o, gOIdhw16i16o); |
| 913 | } |
| 914 | |
| 915 | if (jcp.is_1stconv) { |
| 916 | wei_tag = with_groups |
| 917 | ? ((jcp.simd_w == 4) |
| 918 | ? pick(ndims - 3, gOwi4o, gOhwi4o, gOdhwi4o) |
| 919 | : pick(ndims - 3, gOwi16o, gOhwi16o, gOdhwi16o)) |
| 920 | : pick(ndims - 3, Owi16o, Ohwi16o, Odhwi16o); |
| 921 | } |
| 922 | |
| 923 | if (src_md.format_kind == format_kind::any) |
| 924 | CHECK(memory_desc_init_by_tag(src_md, src_tag)); |
| 925 | else if (curr_src_tag != src_tag) |
| 926 | return status::unimplemented; |
| 927 | jcp.src_tag = src_tag; |
| 928 | |
| 929 | if (dst_md.format_kind == format_kind::any) |
| 930 | CHECK(memory_desc_init_by_tag(dst_md, dst_tag)); |
| 931 | else if (curr_dst_tag != dst_tag) |
| 932 | return status::unimplemented; |
| 933 | jcp.dst_tag = dst_tag; |
| 934 | |
| 935 | if (init_tag(jcp.wei_tag, weights_md, weights_d, wei_tag) |
| 936 | != status::success) |
| 937 | return status::unimplemented; |
| 938 | |
| 939 | jcp.with_bias = cd.bias_desc.format_kind != format_kind::undef; |
| 940 | if (jcp.with_bias) { |
| 941 | if (bias_d.format_kind() == format_kind::any) |
| 942 | CHECK(memory_desc_init_by_tag(bias_md, x)); |
| 943 | } |
| 944 | |
| 945 | CHECK(attr.set_default_formats(&dst_md)); |
| 946 | |
| 947 | const auto &post_ops = attr.post_ops_; |
| 948 | jcp.with_sum = post_ops.find(primitive_kind::sum) != -1; |
| 949 | const int eltwise_ind = post_ops.find(primitive_kind::eltwise); |
| 950 | jcp.with_eltwise = eltwise_ind != -1; |
| 951 | if (jcp.with_eltwise) { |
| 952 | if (dst_d.data_type() == data_type::s32) return status::unimplemented; |
| 953 | } |
| 954 | const int binary_ind = post_ops.find(primitive_kind::binary); |
| 955 | jcp.with_binary = binary_ind != -1; |
| 956 | |
| 957 | jcp.post_ops = post_ops; |
| 958 | |
| 959 | using namespace injector; |
| 960 | static constexpr bool sum_at_pos_0_only = true; |
| 961 | static constexpr bool sum_requires_scale_one = true; |
| 962 | static constexpr bool sum_requires_zp_zero = true; |
| 963 | const bool post_ops_ok_ = post_ops_ok({avx512_core, {eltwise, binary, sum}, |
| 964 | jcp.post_ops, &dst_d, sum_at_pos_0_only, sum_requires_scale_one, |
| 965 | sum_requires_zp_zero}); |
| 966 | if (!post_ops_ok_) return status::unimplemented; |
| 967 | |
| 968 | jcp.typesize_in = typesize; |
| 969 | jcp.typesize_out = typesize; |
| 970 | |
| 971 | if (jcp.is_1stconv) { |
| 972 | jcp.ur_w = nstl::min(jcp.ow, regs); |
| 973 | } else { |
| 974 | // avx512_core guard - just to avoid possible regression for other archs |
| 975 | if (mayiuse(avx512_core)) { |
| 976 | jcp.ur_w = nstl::min(jcp.ow, regs); |
| 977 | } else { |
| 978 | for (int ur_w = regs; ur_w > 0; --ur_w) { |
| 979 | if (jcp.ow % ur_w == 0) { |
| 980 | jcp.ur_w = ur_w; |
| 981 | break; |
| 982 | } |
| 983 | } |
| 984 | } |
| 985 | if ((ndims == 5 && jcp.ur_w <= 8) || (jcp.ur_w <= 1)) { |
| 986 | jcp.ur_w = nstl::min(jcp.ow, regs); |
| 987 | } |
| 988 | } |
| 989 | // TODO (Tanya): currently applied to Segnet convolutions only. |
| 990 | // Need to try for other topologies |
| 991 | if (jcp.ow > 150 && jcp.ur_w < regs / 2) jcp.ur_w = regs; |
| 992 | |
| 993 | int n_oi = (jcp.ow / jcp.ur_w); |
| 994 | int r_pad = calculate_end_padding( |
| 995 | jcp.l_pad, jcp.ur_w * n_oi, jcp.iw, jcp.stride_w, ext_kw); |
| 996 | if (jcp.l_pad > 0 && r_pad > 0) n_oi--; |
| 997 | |
| 998 | // Heuristic to optimize code size on KNX |
| 999 | bool large_code_size = jcp.ur_w != jcp.ow && jcp.l_pad > 0 && r_pad > 0 |
| 1000 | && ((jcp.l_pad <= 0 && n_oi > 0) || (jcp.l_pad > 0 && n_oi > 1)); |
| 1001 | if (large_code_size) { |
| 1002 | const int max_code_size = 24 * 1024; |
| 1003 | const int num_ops_per_reg = 6 + jcp.ic_block * jcp.kw; |
| 1004 | int mult = 1; |
| 1005 | if (jcp.l_pad > 0) mult += 1; |
| 1006 | if (r_pad > 0) mult += 1; |
| 1007 | for (int ur_w = jcp.ur_w; ur_w > regs / 2; --ur_w) { |
| 1008 | if (ur_w * mult * num_ops_per_reg * 9.0 < max_code_size) { |
| 1009 | jcp.ur_w = ur_w; |
| 1010 | break; |
| 1011 | } |
| 1012 | } |
| 1013 | } |
| 1014 | |
| 1015 | /* Grouped channel offset to support 'non-blocked data' format for |
| 1016 | * convolution sizes with '(input_channel / ngroups) < simd' */ |
| 1017 | jcp.nonblk_group_off |
| 1018 | = (jcp.ngroups > 1 && one_of(jcp.src_tag, ncw, nchw, ncdhw)) |
| 1019 | ? jcp.ic |
| 1020 | : 1; |
| 1021 | |
| 1022 | jcp.nb_ic = div_up(jcp.ic, jcp.ic_block); |
| 1023 | jcp.nb_oc = div_up(jcp.oc, jcp.oc_block); |
| 1024 | jcp.nb_ic_blocking = jcp.nb_oc_blocking = 1; |
| 1025 | |
| 1026 | auto is_ow_threading_applicable |
| 1027 | = [=]() { return (!jcp.is_1stconv && one_of(jcp.ndims, 3, 4)); }; |
| 1028 | |
| 1029 | jcp.ow_block = jcp.ow; |
| 1030 | |
| 1031 | auto get_thr_eff = [=](int nb_oc_blocking, int ow_block, int nthr) { |
| 1032 | int nb_ow = div_up(jcp.ow, ow_block); |
| 1033 | int nb_oc_chunks = div_up(jcp.nb_oc, nb_oc_blocking); |
| 1034 | int work_amount = jcp.mb * jcp.oh * nb_oc_chunks * nb_ow; |
| 1035 | float disbalance = (float)jcp.ow / rnd_up(jcp.ow, ow_block); |
| 1036 | float thr_eff |
| 1037 | = disbalance * (float)work_amount / rnd_up(work_amount, nthr); |
| 1038 | return thr_eff; |
| 1039 | }; |
| 1040 | |
| 1041 | auto get_ow_block = [=](int nb_oc_blocking, int ur_w, int nthr) { |
| 1042 | int res_ow_block = jcp.ow; |
| 1043 | float eff = get_thr_eff(nb_oc_blocking, res_ow_block, nthr); |
| 1044 | if (!is_ow_threading_applicable()) return res_ow_block; |
| 1045 | |
| 1046 | int L2_part = (platform::get_per_core_cache_size(2) * 7 / 8) / typesize; |
| 1047 | int size_src_chunk = jcp.ic_block * ur_w * jcp.kh; |
| 1048 | int size_dst_chunk = jcp.oc_block * nb_oc_blocking * ur_w; |
| 1049 | int size_wei_chunk = jcp.oc_block * nb_oc_blocking * jcp.ic_block |
| 1050 | * jcp.kw * jcp.kh; |
| 1051 | int nurw_cache = (L2_part - 2 * size_wei_chunk) |
| 1052 | / (2 * size_dst_chunk + 2 * size_src_chunk); |
| 1053 | // current design of generate() requires ow_block >= 2 * ur_w |
| 1054 | int ow_block_cache = ur_w * nstl::max(2, nurw_cache); |
| 1055 | |
| 1056 | int ow_block_thr = ow_block_cache; |
| 1057 | eff = get_thr_eff(nb_oc_blocking, ow_block_thr, nthr); |
| 1058 | |
| 1059 | int max_nb_ow = div_up(jcp.ow, 2 * ur_w); |
| 1060 | int start_nb_ow = div_up(jcp.ow, ow_block_thr); |
| 1061 | for (int nb_ow = start_nb_ow; nb_ow <= max_nb_ow; nb_ow++) { |
| 1062 | int ow_block |
| 1063 | = nstl::min(rnd_up(div_up(jcp.ow, nb_ow), ur_w), jcp.ow); |
| 1064 | float eff_threshold = 0.9f; |
| 1065 | if (ow_block < nb_oc_blocking * jcp.oc_block && eff > eff_threshold) |
| 1066 | break; |
| 1067 | if (div_up(jcp.ow, ow_block) != nb_ow) continue; |
| 1068 | float thr_eff = get_thr_eff(nb_oc_blocking, ow_block, nthr); |
| 1069 | float eff_step = 1.f; |
| 1070 | if (ow_block >= 2 * ur_w && thr_eff > eff_step * eff) { |
| 1071 | ow_block_thr = ow_block; |
| 1072 | eff = thr_eff; |
| 1073 | } |
| 1074 | eff_threshold = 0.98f; |
| 1075 | if (eff > eff_threshold) break; |
| 1076 | } |
| 1077 | res_ow_block = nstl::min(jcp.ow, nstl::max(2 * ur_w, ow_block_thr)); |
| 1078 | eff = get_thr_eff(nb_oc_blocking, res_ow_block, nthr); |
| 1079 | return res_ow_block; |
| 1080 | }; |
| 1081 | |
| 1082 | const unsigned int L1_cache_size = platform::get_per_core_cache_size(1); |
| 1083 | if (mayiuse(avx512_core)) { |
| 1084 | int try_nb_oc_blocking = 2; |
| 1085 | unsigned int ker_inp_size = typesize * div_up(jcp.iw, jcp.stride_w) |
| 1086 | * jcp.ic_block * jcp.kh * jcp.kd; |
| 1087 | unsigned int ker_out_size |
| 1088 | = typesize * jcp.ow * jcp.oc_block * try_nb_oc_blocking; |
| 1089 | unsigned int ker_wei_size = typesize * jcp.kh * jcp.kw * jcp.ic_block |
| 1090 | * jcp.oc_block * try_nb_oc_blocking * jcp.kd; |
| 1091 | unsigned int ker_total_size |
| 1092 | = ker_inp_size + ker_out_size + ker_wei_size; |
| 1093 | |
| 1094 | bool embd_bcast_condition_base = true |
| 1095 | && (jcp.kw == 3 && jcp.ow <= 28 |
| 1096 | && ker_total_size < L1_cache_size) |
| 1097 | && !(jcp.kw == 3 && jcp.ow == 13 && jcp.ic >= 192) |
| 1098 | && !(jcp.kw == 3 && jcp.ow == 28 && jcp.ic >= 512); |
| 1099 | // These conditions define a set of shapes with 'ow = 1' which |
| 1100 | // have a very limited optimization space for performance. Try |
| 1101 | // to optimize by using a larger 'nb_oc_blocking' size. |
| 1102 | bool expl_bcast_condition |
| 1103 | = everyone_is(1, jcp.ngroups, jcp.mb, jcp.stride_h, jcp.ow, |
| 1104 | jcp.stride_w, jcp.id, jcp.od, jcp.kd, jcp.stride_d) |
| 1105 | && jcp.iw == jcp.kw && jcp.nb_oc > 1 |
| 1106 | && everyone_is(0, jcp.l_pad, jcp.r_pad, jcp.dilate_w, jcp.f_pad, |
| 1107 | jcp.back_pad, jcp.dilate_d) |
| 1108 | && jcp.oh >= 60 && jcp.kh >= 3; |
| 1109 | |
| 1110 | bool embd_bcast_condition = !expl_bcast_condition |
| 1111 | && (jcp.kw > 3 |
| 1112 | || (jcp.stride_w == 1 && jcp.stride_h == 1 |
| 1113 | && embd_bcast_condition_base) |
| 1114 | || ((jcp.stride_w != 1 || jcp.stride_h != 1) |
| 1115 | && ((jcp.mb <= 16 |
| 1116 | && (jcp.oc <= 192 || jcp.oh <= 10) |
| 1117 | && embd_bcast_condition_base))) |
| 1118 | || (jcp.mb == 1 |
| 1119 | && (jcp.ur_w >= jcp.ow || jcp.is_1stconv |
| 1120 | || (jcp.ow <= 147 && jcp.oc <= 96)))); |
| 1121 | |
| 1122 | if (jcp.mb == 1) { |
| 1123 | unsigned int inp_size = jcp.mb * div_up(jcp.ih, jcp.stride_h) |
| 1124 | * div_up(jcp.iw, jcp.stride_w) * jcp.ic; |
| 1125 | unsigned int wei_size = jcp.ic * jcp.oc * jcp.kh * jcp.kw; |
| 1126 | |
| 1127 | // Estimate whether we need to limit the number of threads |
| 1128 | // and calculate this number. Includes some heuristic. |
| 1129 | int oc_chunks = jcp.nb_oc / jcp.nb_oc_blocking; |
| 1130 | int work_amount = jcp.mb * jcp.ngroups * oc_chunks * jcp.oh; |
| 1131 | int job_size_min = work_amount / nthreads; |
| 1132 | int job_size_max = div_up(work_amount, nthreads); |
| 1133 | int ch_max = rnd_up(jcp.oh, job_size_max); |
| 1134 | int ch_min = (job_size_min == 0) ? jcp.oh |
| 1135 | : rnd_up(jcp.oh, job_size_min); |
| 1136 | bool not_aligned_max = ch_max % jcp.oh != 0 && ch_max / jcp.oh < 2 |
| 1137 | && (jcp.oh != 8 || ch_max / jcp.oh > 1); |
| 1138 | bool not_aligned_min = ch_min % jcp.oh != 0 && ch_min / jcp.oh < 2 |
| 1139 | && (jcp.oh != 8 || ch_min / jcp.oh > 1); |
| 1140 | bool eligible_case = (jcp.stride_h == 1 && jcp.stride_w == 1) |
| 1141 | || nthreads > oc_chunks; |
| 1142 | if (jcp.loop_order == loop_cgn && oc_chunks > 1 && nthreads > 1 |
| 1143 | && wei_size / inp_size > 24 |
| 1144 | && (not_aligned_max || not_aligned_min) && eligible_case) { |
| 1145 | // Try to find number of threads > nthreads / 2 such that |
| 1146 | // oc_chunks is a multiple of nthreads, or nthreads is a |
| 1147 | // multiple of oc_chunks. Otherwise, keep default value. |
| 1148 | // TODO: implement a task-based alternative without throttling. |
| 1149 | jcp.aligned_threads = jcp.nthr; |
| 1150 | for (int i = jcp.nthr; i > jcp.nthr / 2; i--) { |
| 1151 | if (oc_chunks % i == 0 || i % oc_chunks == 0) { |
| 1152 | jcp.aligned_threads = i; |
| 1153 | break; |
| 1154 | } |
| 1155 | } |
| 1156 | } |
| 1157 | } |
| 1158 | |
| 1159 | const int max_nb_oc = 5; |
| 1160 | if (embd_bcast_condition) { |
| 1161 | jcp.kernel_kind = embd_bcast; |
| 1162 | jcp.ur_w = nstl::min(jcp.ow, regs); |
| 1163 | jcp.nb_ic_blocking = jcp.nb_oc_blocking = 1; |
| 1164 | const unsigned int L1_cache_size |
| 1165 | = platform::get_per_core_cache_size(1); |
| 1166 | if (ker_total_size < L1_cache_size && jcp.ow <= 8 && jcp.kh <= 3 |
| 1167 | && jcp.kw <= 3 && jcp.nb_oc % try_nb_oc_blocking == 0 |
| 1168 | && IMPLICATION(jcp.is_1stconv, jcp.mb == 1) |
| 1169 | && IMPLICATION(jcp.mb == 1, jcp.ur_w < jcp.ow)) { |
| 1170 | jcp.nb_oc_blocking = try_nb_oc_blocking; |
| 1171 | jcp.ur_w = nstl::min(jcp.ow, 31 / (jcp.nb_oc_blocking + 1)); |
| 1172 | } |
| 1173 | } else { |
| 1174 | jcp.kernel_kind = expl_bcast; |
| 1175 | jcp.nb_ic_blocking = 1; |
| 1176 | if (IMPLICATION(jcp.is_1stconv, jcp.mb > 1) |
| 1177 | || expl_bcast_condition) { |
| 1178 | float best_thr_eff = 0.f; |
| 1179 | int best_nb_oc_blocking = 1; |
| 1180 | for (int i = nstl::min(jcp.nb_oc, max_nb_oc); i > 0; i--) { |
| 1181 | if (jcp.nb_oc % i == 0) { |
| 1182 | if (expl_bcast_condition) { |
| 1183 | best_nb_oc_blocking = i; |
| 1184 | break; |
| 1185 | } else { |
| 1186 | int ur_w = nstl::min(jcp.ow, 31 / (i + 1)); |
| 1187 | int ow_block = get_ow_block(i, ur_w, jcp.nthr); |
| 1188 | float thr_eff = get_thr_eff(i, ow_block, jcp.nthr); |
| 1189 | if (thr_eff > 1.05f * best_thr_eff) { |
| 1190 | best_nb_oc_blocking = i; |
| 1191 | best_thr_eff = thr_eff; |
| 1192 | } |
| 1193 | } |
| 1194 | } |
| 1195 | } |
| 1196 | jcp.nb_oc_blocking = best_nb_oc_blocking; |
| 1197 | jcp.ur_w = nstl::min(jcp.ow, 31 / (jcp.nb_oc_blocking + 1)); |
| 1198 | } |
| 1199 | } |
| 1200 | } |
| 1201 | |
| 1202 | jcp.ur_w_tail = jcp.ow % jcp.ur_w; |
| 1203 | |
| 1204 | bool args_ok = true && jcp.l_pad <= jcp.ur_w |
| 1205 | && jcp.ic <= src_d.padded_dims()[1] |
| 1206 | && jcp.oc <= dst_d.padded_dims()[1] |
| 1207 | && jcp.ic <= weights_d.padded_dims()[with_groups + 1] |
| 1208 | && jcp.oc <= weights_d.padded_dims()[with_groups + 0]; |
| 1209 | if (!args_ok) return status::unimplemented; |
| 1210 | |
| 1211 | int r_pad_no_tail = nstl::max(0, |
| 1212 | calculate_end_padding(jcp.l_pad, jcp.ow - jcp.ur_w_tail, jcp.iw, |
| 1213 | jcp.stride_w, ext_kw)); |
| 1214 | if (r_pad_no_tail > jcp.ur_w) return status::unimplemented; |
| 1215 | |
| 1216 | pick_loop_order(jcp); |
| 1217 | |
| 1218 | jcp.nb_ic_L2 = jcp.nb_ic; |
| 1219 | |
| 1220 | jcp.ow_block = get_ow_block(jcp.nb_oc_blocking, jcp.ur_w, jcp.nthr); |
| 1221 | jcp.nb_ow = div_up(jcp.ow, jcp.ow_block); |
| 1222 | float thr_eff = get_thr_eff(jcp.nb_oc_blocking, jcp.ow_block, jcp.nthr); |
| 1223 | |
| 1224 | /* adjust the thread decomposition |
| 1225 | * to improve the thr_eff for small size problem |
| 1226 | * the threshold L1_cache_size is empirical */ |
| 1227 | size_t wei_size |
| 1228 | = (size_t)typesize * jcp.ic * jcp.oc * jcp.kh * jcp.kw * jcp.kd; |
| 1229 | size_t out_size = (size_t)jcp.mb * jcp.typesize_out * jcp.oc * jcp.oh |
| 1230 | * jcp.ow * jcp.od; |
| 1231 | size_t inp_size = (size_t)jcp.mb * jcp.typesize_in * jcp.ic * jcp.ih |
| 1232 | * jcp.iw * jcp.id; |
| 1233 | size_t total_size = jcp.ngroups * (wei_size + out_size + inp_size); |
| 1234 | float eff_threshold = 0.98f; |
| 1235 | |
| 1236 | if (thr_eff < eff_threshold && jcp.ngroups < jcp.nthr |
| 1237 | && (total_size < L1_cache_size)) { |
| 1238 | int ow_block = jcp.ow_block; |
| 1239 | float best_thr_eff = -1.0f; |
| 1240 | float eff = -1.0f; |
| 1241 | int end_nthr = with_groups ? jcp.ngroups : 1; |
| 1242 | for (int nthr = jcp.nthr / 2; nthr >= end_nthr; nthr--) { |
| 1243 | ow_block = get_ow_block(jcp.nb_oc_blocking, jcp.ur_w, nthr); |
| 1244 | eff = get_thr_eff(jcp.nb_oc_blocking, ow_block, nthr); |
| 1245 | if (eff > 1.1f * best_thr_eff) { |
| 1246 | best_thr_eff = eff; |
| 1247 | jcp.ow_block = ow_block; |
| 1248 | jcp.nb_ow = div_up(jcp.ow, jcp.ow_block); |
| 1249 | jcp.nthr = jcp.aligned_threads = nthr; |
| 1250 | if (best_thr_eff > eff_threshold) break; |
| 1251 | } |
| 1252 | } |
| 1253 | } |
| 1254 | |
| 1255 | const int L2_size = platform::get_per_core_cache_size(2) / typesize; |
| 1256 | // Source and output data needs to fit in L2, |
| 1257 | // leaving some space for weights and prefetching. |
| 1258 | int h_L2 = int(((0.6f * L2_size) / jcp.simd_w |
| 1259 | - nstl::min(0, jcp.kh - jcp.stride_h) * jcp.iw) |
| 1260 | / (jcp.stride_h * jcp.iw + jcp.ow)); |
| 1261 | jcp.h_blocking = nstl::max(1, nstl::min(jcp.oh, h_L2)); |
| 1262 | |
| 1263 | if (is_data_layout_nxc) { |
| 1264 | // TODO: improve L2 blocking for large IC |
| 1265 | const int nb_ic_theshold_L2 = 32; |
| 1266 | if (jcp.nb_ic > nb_ic_theshold_L2 && jcp.nb_ic < 2 * nb_ic_theshold_L2) |
| 1267 | jcp.nb_ic_L2 = div_up(jcp.nb_ic, 2); |
| 1268 | else |
| 1269 | jcp.nb_ic_L2 = nstl::min(nb_ic_theshold_L2, jcp.nb_ic); |
| 1270 | } |
| 1271 | |
| 1272 | // A rough check on code size |
| 1273 | // TODO: come up with a tighter bound |
| 1274 | { |
| 1275 | const int max_code_size = 256 * 1024; // default size of jit generator |
| 1276 | int mult = 1 + (jcp.l_pad > 0) + (r_pad > 0); |
| 1277 | const float max_instruction_size = 15; |
| 1278 | float ur_fac |
| 1279 | = (float)jcp.kw * jcp.ic_block * jcp.nb_oc_blocking * jcp.ur_w; |
| 1280 | float code_size = mult * ur_fac * max_instruction_size; |
| 1281 | if (code_size > max_code_size) return status::unimplemented; |
| 1282 | } |
| 1283 | |
| 1284 | return status::success; |
| 1285 | } |
| 1286 | |
| 1287 | void jit_avx512_common_conv_fwd_kernel::init_scratchpad( |
| 1288 | memory_tracking::registrar_t &scratchpad, const jit_conv_conf_t &jcp) { |
| 1289 | if (jcp.with_bias && jcp.oc != jcp.oc_without_padding) |
| 1290 | scratchpad.book(key_conv_padded_bias, jcp.oc, jcp.typesize_out); |
| 1291 | } |
| 1292 | |
| 1293 | template <typename Vmm> |
| 1294 | void _jit_avx512_common_conv_bwd_data_kernel_f32<Vmm>::prepare_output( |
| 1295 | int ur_w) { |
| 1296 | for (int k = 0; k < jcp.nb_ic_blocking; k++) { |
| 1297 | for (int j = 0; j < ur_w; j++) { |
| 1298 | Vmm vmm = vmm_out(j, k); |
| 1299 | vpxord(vmm, vmm, vmm); |
| 1300 | } |
| 1301 | } |
| 1302 | } |
| 1303 | |
| 1304 | template <typename Vmm> |
| 1305 | void _jit_avx512_common_conv_bwd_data_kernel_f32<Vmm>::store_output(int ur_w) { |
| 1306 | Label no_update_label; |
| 1307 | const int ic_tail = jcp.ic_without_padding % jcp.simd_w; |
| 1308 | const bool dsrc_layout_nxc = is_dsrc_layout_nxc(); |
| 1309 | mov(reg_channel, ptr[param + GET_OFF(channel)]); |
| 1310 | cmp(reg_channel, 0); |
| 1311 | je(no_update_label, T_NEAR); |
| 1312 | for (int k = 0; k < jcp.nb_ic_blocking; k++) { |
| 1313 | for (int j = 0; j < ur_w; j++) { |
| 1314 | Vmm vmm = vmm_out(j, k); |
| 1315 | size_t aux_src_offset = get_diff_src_offset(j, k); |
| 1316 | vaddps(vmm, |
| 1317 | EVEX_compress_addr_safe( |
| 1318 | reg_src, aux_src_offset, reg_long_offt)); |
| 1319 | } |
| 1320 | } |
| 1321 | |
| 1322 | L(no_update_label); |
| 1323 | for (int k = 0; k < jcp.nb_ic_blocking; k++) { |
| 1324 | for (int j = 0; j < ur_w; j++) { |
| 1325 | Vmm vmm = vmm_out(j, k); |
| 1326 | // mask only needed for last oc_block |
| 1327 | if (ic_tail && k + 1 == jcp.nb_ic_blocking && dsrc_layout_nxc) |
| 1328 | vmm = vmm | k_ic_tail_mask; |
| 1329 | size_t aux_src_offset = get_diff_src_offset(j, k); |
| 1330 | vmovups(EVEX_compress_addr_safe( |
| 1331 | reg_src, aux_src_offset, reg_long_offt), |
| 1332 | vmm); |
| 1333 | } |
| 1334 | } |
| 1335 | } |
| 1336 | |
| 1337 | template <typename Vmm> |
| 1338 | void _jit_avx512_common_conv_bwd_data_kernel_f32<Vmm>::compute_loop_fma( |
| 1339 | int ur_w, int l_overflow, int r_overflow) { |
| 1340 | Label kh_label, kd_label; |
| 1341 | int kw = jcp.kw; |
| 1342 | int ow = jcp.ow; |
| 1343 | |
| 1344 | int ic_block = jcp.ic_block; |
| 1345 | int oc_block = jcp.oc_block; |
| 1346 | int stride_w = jcp.stride_w; |
| 1347 | int stride_h = jcp.stride_h; |
| 1348 | |
| 1349 | int ker_pipeline_depth = 4; |
| 1350 | assert(ker_reg_base_idx + ker_pipeline_depth <= 32); |
| 1351 | assert(oc_block >= ker_pipeline_depth); |
| 1352 | |
| 1353 | int num_ker_loads = oc_block * kw; |
| 1354 | const bool ddst_layout_nxc = is_ddst_layout_nxc(); |
| 1355 | int oc_mult = ddst_layout_nxc ? jcp.ngroups * jcp.oc : oc_block; |
| 1356 | const bool ocb_loop_in_compute_function = ddst_layout_nxc; |
| 1357 | |
| 1358 | const int ic_tail = jcp.ic_tail; |
| 1359 | const int oc_tail = jcp.oc_tail; |
| 1360 | std::vector<Label> oc_tail_jmp(kw); |
| 1361 | if (ic_tail || oc_tail) ker_pipeline_depth = 1; |
| 1362 | |
| 1363 | if (one_of(jcp.ndims, 3, 4)) { |
| 1364 | mov(aux_reg_dst, reg_dst); |
| 1365 | mov(aux_reg_ker, reg_ker); |
| 1366 | } |
| 1367 | |
| 1368 | if (jcp.ndims == 5) { |
| 1369 | push(reg_src); |
| 1370 | |
| 1371 | mov(reg_ki, ptr[param + GET_OFF(kd_padding)]); |
| 1372 | mov(aux_reg_dst_d, reg_dst); |
| 1373 | if (ocb_loop_in_compute_function) { |
| 1374 | // need to continue with the same kernel pointer, but as |
| 1375 | // aux_reg_ker_d == reg_ker we need to save its value and restore |
| 1376 | // it after kd loop |
| 1377 | assert(aux_reg_ker_d == reg_ker); |
| 1378 | push(aux_reg_ker_d); |
| 1379 | } else |
| 1380 | mov(aux_reg_ker_d, ptr[param + GET_OFF(filt)]); |
| 1381 | |
| 1382 | L(kd_label); |
| 1383 | mov(reg_kj, ptr[param + GET_OFF(kh_padding)]); |
| 1384 | } else { |
| 1385 | mov(reg_kj, reg_kh); |
| 1386 | } |
| 1387 | |
| 1388 | if (jcp.ndims == 5) { |
| 1389 | mov(aux_reg_dst, aux_reg_dst_d); |
| 1390 | mov(aux_reg_ker, aux_reg_ker_d); |
| 1391 | } |
| 1392 | |
| 1393 | L(kh_label); |
| 1394 | { |
| 1395 | int step = 0; |
| 1396 | for (int ki = 0; ki < kw; ki++) { |
| 1397 | for (int oc = 0; oc < oc_block; oc++) { |
| 1398 | if (oc_tail && oc >= oc_tail) { |
| 1399 | // if src has only tails to compute, skip early |
| 1400 | if (jcp.oc == oc_tail) |
| 1401 | break; |
| 1402 | else if (oc == oc_tail) { |
| 1403 | cmp(reg_channel, oc_tail); |
| 1404 | je(oc_tail_jmp[ki], T_NEAR); |
| 1405 | } |
| 1406 | } |
| 1407 | if (step == 0) { |
| 1408 | for (int i = 0; i < ker_pipeline_depth; i++) { |
| 1409 | int aux_kernel_offset = typesize |
| 1410 | * ((oc + i) * oc_block |
| 1411 | + ki * ic_block * oc_block); |
| 1412 | vmovups(vmm_ker(i), |
| 1413 | EVEX_compress_addr( |
| 1414 | aux_reg_ker, aux_kernel_offset)); |
| 1415 | } |
| 1416 | } else if (step < num_ker_loads - ker_pipeline_depth + 1) { |
| 1417 | int load_offset = ker_pipeline_depth - 1; |
| 1418 | int ker_load_reg_idx |
| 1419 | = (step + load_offset) % ker_pipeline_depth; |
| 1420 | int aux_kernel_offset = typesize |
| 1421 | * ((oc + load_offset) * oc_block |
| 1422 | + ki * ic_block * oc_block); |
| 1423 | vmovups(vmm_ker(ker_load_reg_idx), |
| 1424 | EVEX_compress_addr(aux_reg_ker, aux_kernel_offset)); |
| 1425 | } |
| 1426 | |
| 1427 | auto vmm_kernel = vmm_ker(step % ker_pipeline_depth); |
| 1428 | |
| 1429 | int jj_start = get_iw_start(ki, l_overflow); |
| 1430 | int jj_end = get_iw_end(ur_w, ki, r_overflow); |
| 1431 | const int dil_w = jcp.dilate_w + 1; |
| 1432 | const int ref_jj_start |
| 1433 | = nstl::max(0, l_overflow - (kw - 1 - ki) * dil_w); |
| 1434 | const int ref_jj_end |
| 1435 | = ur_w - nstl::max(0, r_overflow - ki * dil_w); |
| 1436 | assert(IMPLICATION(stride_w == 1, |
| 1437 | jj_start == ref_jj_start && jj_end == ref_jj_end)); |
| 1438 | UNUSED(dil_w); |
| 1439 | UNUSED(ref_jj_start); |
| 1440 | UNUSED(ref_jj_end); |
| 1441 | |
| 1442 | for (int jj = jj_start; jj < jj_end; jj += stride_w) { |
| 1443 | assert((jj + jcp.l_pad - ki * (jcp.dilate_w + 1)) % stride_w |
| 1444 | == 0); |
| 1445 | int aux_dst_offset = get_dst_offset(jj, oc, ki); |
| 1446 | vfmadd231ps(vmm_out(jj, 0), vmm_kernel, |
| 1447 | EVEX_compress_addr( |
| 1448 | aux_reg_dst, aux_dst_offset, true)); |
| 1449 | } |
| 1450 | step++; |
| 1451 | } |
| 1452 | L(oc_tail_jmp[ki]); |
| 1453 | } |
| 1454 | |
| 1455 | const int ker_shift = typesize * stride_h * kw * oc_block * ic_block; |
| 1456 | add(aux_reg_ker, ker_shift); |
| 1457 | const int ddst_shift = typesize * (jcp.dilate_h + 1) * ow * oc_mult; |
| 1458 | sub(aux_reg_dst, ddst_shift); |
| 1459 | |
| 1460 | dec(reg_kj); |
| 1461 | cmp(reg_kj, 0); |
| 1462 | jg(kh_label, T_NEAR); |
| 1463 | } |
| 1464 | if (jcp.ndims == 5) { |
| 1465 | const int depth_ddst_shift |
| 1466 | = typesize * (jcp.dilate_d + 1) * jcp.oh * ow * oc_mult; |
| 1467 | sub(aux_reg_dst_d, depth_ddst_shift); |
| 1468 | const int depth_ker_shift = typesize * jcp.stride_d * jcp.kw * jcp.kh |
| 1469 | * oc_block * ic_block; |
| 1470 | add(aux_reg_ker_d, depth_ker_shift); |
| 1471 | |
| 1472 | dec(reg_ki); |
| 1473 | cmp(reg_ki, 0); |
| 1474 | jg(kd_label, T_NEAR); |
| 1475 | if (ocb_loop_in_compute_function) pop(aux_reg_ker_d); |
| 1476 | } |
| 1477 | |
| 1478 | if (jcp.ndims == 5) { pop(reg_src); } |
| 1479 | } |
| 1480 | |
| 1481 | template <typename Vmm> |
| 1482 | void _jit_avx512_common_conv_bwd_data_kernel_f32<Vmm>::compute_loop_fma_core( |
| 1483 | int ur_w, int l_overflow, int r_overflow, int k_offset) { |
| 1484 | int kw = jcp.kw; |
| 1485 | int ow = jcp.ow; |
| 1486 | int stride_w = jcp.stride_w; |
| 1487 | int ic_block = jcp.ic_block; |
| 1488 | int oc_block = jcp.oc_block; |
| 1489 | int nb_ic_block = jcp.nb_ic_blocking; |
| 1490 | Label kh_label, kd_label; |
| 1491 | |
| 1492 | const bool ddst_layout_nxc = is_ddst_layout_nxc(); |
| 1493 | int shift_ker_ptr = typesize * kw * oc_block * ic_block; |
| 1494 | int oc_mult = ddst_layout_nxc ? jcp.ngroups * jcp.oc : oc_block; |
| 1495 | int shift_dst_ptr = typesize * (jcp.dilate_h + 1) * ow * oc_mult; |
| 1496 | |
| 1497 | const int oc_tail = jcp.oc_tail; |
| 1498 | const int max_filter_size = 20; |
| 1499 | Label oc_tail_jmp[max_filter_size]; |
| 1500 | |
| 1501 | auto kernel_offset = [=](int icb, int oc, int ki) { |
| 1502 | int blk_idx = icb * jcp.kh * jcp.kw * jcp.kd + ki; |
| 1503 | int blk_offset = blk_idx * jcp.oc_block * jcp.ic_block; |
| 1504 | int oc_offset = oc * jcp.oc_block; |
| 1505 | return typesize * (blk_offset + oc_offset); |
| 1506 | }; |
| 1507 | |
| 1508 | if (one_of(jcp.ndims, 3, 4)) { |
| 1509 | mov(aux_reg_dst, reg_dst); |
| 1510 | mov(aux_reg_ker, reg_ker); |
| 1511 | } |
| 1512 | |
| 1513 | const bool ocb_loop_in_compute_function = ddst_layout_nxc; |
| 1514 | if (jcp.ndims == 5) { |
| 1515 | push(reg_src); |
| 1516 | |
| 1517 | mov(reg_ki, ptr[param + GET_OFF(kd_padding)]); |
| 1518 | mov(aux_reg_dst_d, reg_dst); |
| 1519 | if (ocb_loop_in_compute_function) { |
| 1520 | // need to continue with the same kernel pointer, but as |
| 1521 | // aux_reg_ker_d == reg_ker we need to save its value and restore |
| 1522 | // it after kd loop |
| 1523 | assert(aux_reg_ker_d == reg_ker); |
| 1524 | push(aux_reg_ker_d); |
| 1525 | } else |
| 1526 | mov(aux_reg_ker_d, ptr[param + GET_OFF(filt)]); |
| 1527 | |
| 1528 | L(kd_label); |
| 1529 | mov(reg_kj, ptr[param + GET_OFF(kh_padding)]); |
| 1530 | } else { |
| 1531 | mov(reg_kj, reg_kh); |
| 1532 | } |
| 1533 | |
| 1534 | if (jcp.ndims == 5) { |
| 1535 | mov(aux_reg_dst, aux_reg_dst_d); |
| 1536 | mov(aux_reg_ker, aux_reg_ker_d); |
| 1537 | } |
| 1538 | |
| 1539 | L(kh_label); |
| 1540 | { |
| 1541 | for (int ki = 0; ki < kw; ki++) { |
| 1542 | int jj_start = get_iw_start(ki, l_overflow); |
| 1543 | int jj_end = get_iw_end(ur_w, ki, r_overflow); |
| 1544 | for (int oc = 0; oc < oc_block; oc++) { |
| 1545 | if (oc_tail && oc >= oc_tail) { |
| 1546 | // if src has only tails to compute, skip early |
| 1547 | if (jcp.oc == oc_tail) |
| 1548 | break; |
| 1549 | else if (oc == oc_tail) { |
| 1550 | cmp(reg_channel, oc_tail); |
| 1551 | je(oc_tail_jmp[ki], T_NEAR); |
| 1552 | } |
| 1553 | } |
| 1554 | if (jcp.kernel_kind == expl_bcast) { |
| 1555 | for (int jj = jj_start; jj < jj_end; jj++) { |
| 1556 | int aux_output_offset = get_dst_offset(jj, oc, ki); |
| 1557 | vbroadcastss(vmm_inp(jj, nb_ic_block), |
| 1558 | ptr[aux_reg_dst + aux_output_offset]); |
| 1559 | } |
| 1560 | } |
| 1561 | for (int ii = 0; ii < nb_ic_block; ii++) { |
| 1562 | int aux_kernel_offset |
| 1563 | = kernel_offset(ii, oc, ki + k_offset); |
| 1564 | if (jj_end - jj_start > 0) |
| 1565 | vmovups(vmm_wei, |
| 1566 | EVEX_compress_addr( |
| 1567 | aux_reg_ker, aux_kernel_offset)); |
| 1568 | for (int jj = jj_start; jj < jj_end; jj += stride_w) |
| 1569 | if (jcp.kernel_kind == expl_bcast) |
| 1570 | vfmadd231ps(vmm_out(jj, ii), |
| 1571 | vmm_inp(jj, nb_ic_block), vmm_wei); |
| 1572 | else |
| 1573 | vfmadd231ps(vmm_out(jj, ii), vmm_wei, |
| 1574 | EVEX_compress_addr(aux_reg_dst, |
| 1575 | get_dst_offset(jj, oc, ki), true)); |
| 1576 | } |
| 1577 | } |
| 1578 | L(oc_tail_jmp[ki]); |
| 1579 | } |
| 1580 | add(aux_reg_ker, shift_ker_ptr); |
| 1581 | sub(aux_reg_dst, shift_dst_ptr); |
| 1582 | dec(reg_kj); |
| 1583 | cmp(reg_kj, 0); |
| 1584 | jg(kh_label, T_NEAR); |
| 1585 | } |
| 1586 | |
| 1587 | if (jcp.ndims == 5) { |
| 1588 | sub(aux_reg_dst_d, |
| 1589 | typesize * (jcp.dilate_d + 1) * jcp.oh * ow * oc_mult); |
| 1590 | add(aux_reg_ker_d, typesize * jcp.kw * jcp.kh * oc_block * ic_block); |
| 1591 | |
| 1592 | dec(reg_ki); |
| 1593 | cmp(reg_ki, 0); |
| 1594 | jg(kd_label, T_NEAR); |
| 1595 | |
| 1596 | if (ocb_loop_in_compute_function) pop(aux_reg_ker_d); |
| 1597 | pop(reg_src); |
| 1598 | } |
| 1599 | } |
| 1600 | |
| 1601 | template <typename Vmm> |
| 1602 | inline void _jit_avx512_common_conv_bwd_data_kernel_f32<Vmm>::compute_loop( |
| 1603 | int ur_w, int l_overflow, int r_overflow, int k_offset) { |
| 1604 | if (jcp.ndims == 5) push(reg_oi); |
| 1605 | |
| 1606 | prepare_output(ur_w); |
| 1607 | |
| 1608 | Label skip_compute_loop; |
| 1609 | if (jcp.ndims == 5) { |
| 1610 | mov(reg_kj, ptr[param + GET_OFF(kd_padding)]); |
| 1611 | cmp(reg_kj, 0); |
| 1612 | jle(skip_compute_loop, T_NEAR); |
| 1613 | } |
| 1614 | mov(reg_kj, ptr[param + GET_OFF(kh_padding)]); |
| 1615 | cmp(reg_kj, 0); |
| 1616 | jle(skip_compute_loop, T_NEAR); |
| 1617 | |
| 1618 | const bool generate_ocb_loop = jcp.nb_oc > 1 && is_ddst_layout_nxc(); |
| 1619 | Label oc_loop; |
| 1620 | if (generate_ocb_loop) { |
| 1621 | push(reg_dst); |
| 1622 | push(reg_ker); |
| 1623 | |
| 1624 | mov(reg_channel, ptr[param1 + GET_OFF(reduce_work)]); |
| 1625 | L(oc_loop); |
| 1626 | } |
| 1627 | |
| 1628 | if (jcp.kernel_kind == embd_bcast && jcp.nb_ic_blocking == 1) |
| 1629 | compute_loop_fma(ur_w, l_overflow, r_overflow); |
| 1630 | else |
| 1631 | compute_loop_fma_core(ur_w, l_overflow, r_overflow, k_offset); |
| 1632 | |
| 1633 | if (generate_ocb_loop) { |
| 1634 | add(reg_dst, jcp.oc_block * typesize); |
| 1635 | const int ker_shift = jcp.nb_ic * jcp.kd * jcp.kh * jcp.kw |
| 1636 | * jcp.ic_block * jcp.oc_block * typesize; |
| 1637 | add(reg_ker, ker_shift); |
| 1638 | sub(reg_channel, jcp.oc_block); |
| 1639 | jg(oc_loop, T_NEAR); |
| 1640 | |
| 1641 | pop(reg_ker); |
| 1642 | pop(reg_dst); |
| 1643 | } |
| 1644 | |
| 1645 | L(skip_compute_loop); |
| 1646 | store_output(ur_w); |
| 1647 | if (jcp.ndims == 5) pop(reg_oi); |
| 1648 | } |
| 1649 | |
| 1650 | template <typename Vmm> |
| 1651 | void _jit_avx512_common_conv_bwd_data_kernel_f32<Vmm>::generate() { |
| 1652 | int iw = jcp.iw; |
| 1653 | int kw = jcp.kw; |
| 1654 | int ur_w = jcp.ur_w; |
| 1655 | int ic_block = jcp.ic_block; |
| 1656 | int oc_block = jcp.oc_block; |
| 1657 | int nb_iw = jcp.nb_iw; |
| 1658 | int iw_block = jcp.iw_block; |
| 1659 | int ur_w_tail = jcp.ur_w_tail; |
| 1660 | int dilate_w = jcp.dilate_w + 1; |
| 1661 | int stride_w = jcp.stride_w; |
| 1662 | |
| 1663 | int dst_shift = jcp.typesize_in * (ur_w / stride_w) |
| 1664 | * (is_ddst_layout_nxc() ? jcp.ngroups * jcp.oc : oc_block); |
| 1665 | int src_shift = jcp.typesize_out * ur_w |
| 1666 | * (is_dsrc_layout_nxc() ? jcp.ngroups * jcp.ic : ic_block); |
| 1667 | |
| 1668 | preamble(); |
| 1669 | |
| 1670 | mov(reg_src, ptr[param + GET_OFF(src)]); |
| 1671 | mov(reg_dst, ptr[param + GET_OFF(dst)]); |
| 1672 | mov(reg_ker, ptr[param + GET_OFF(filt)]); |
| 1673 | |
| 1674 | mov(reg_kh, ptr[param + GET_OFF(kh_padding)]); |
| 1675 | |
| 1676 | const int ic_tail = jcp.ic_tail; |
| 1677 | if (ic_tail) { |
| 1678 | Label masking_done; |
| 1679 | // dummy mask all 1's |
| 1680 | kxnorw(k_ic_tail_mask, k_ic_tail_mask, k_ic_tail_mask); |
| 1681 | mov(reg_load_work, ptr[param1 + GET_OFF(load_work)]); |
| 1682 | cmp(reg_load_work, jcp.nb_ic_blocking * jcp.ic_block); |
| 1683 | je(masking_done, T_NEAR); |
| 1684 | Reg32 reg_tail_32 = reg_tail.cvt32(); |
| 1685 | mov(reg_tail_32, (1 << ic_tail) - 1); |
| 1686 | kmovw(k_ic_tail_mask, reg_tail_32); |
| 1687 | L(masking_done); |
| 1688 | } |
| 1689 | |
| 1690 | int l_overflow = nstl::max(0, ((kw - 1) * dilate_w - jcp.l_pad) / stride_w); |
| 1691 | int r_overflow = nstl::max( |
| 1692 | 0, ((kw - 1) * dilate_w - nstl::max(0, jcp.r_pad)) / stride_w); |
| 1693 | int r_overflow_no_tail = nstl::max(0, |
| 1694 | ((kw - 1) * dilate_w - nstl::max(0, jcp.r_pad + ur_w_tail)) |
| 1695 | / stride_w); |
| 1696 | |
| 1697 | int body_l_overflow = 0, body_r_overflow = 0; |
| 1698 | int n_oi = iw / ur_w; |
| 1699 | int head_n_oi = 0, body_n_oi = 0, pretail_n_oi = 0, tail_n_oi = 0; |
| 1700 | int head_thread = 0, pretail_thread = 0, tail_thread = 0; |
| 1701 | bool threaded = is_iw_threading_on(jcp); |
| 1702 | Label head_label, body_label, pretail_label, tail_label, end_label; |
| 1703 | assert(n_oi > 0); |
| 1704 | if (r_overflow_no_tail > 0) n_oi--; |
| 1705 | if (l_overflow > 0) n_oi--; |
| 1706 | if (n_oi < 0) { |
| 1707 | // l_overflow and r_overflow_no_tail are handled in the same compute_loop. |
| 1708 | // Perform one iteration of body handling l_overflow and r_overflow_no_tail. |
| 1709 | // TODO: Align other convolution kernels with this kernel. This version |
| 1710 | // now uses r_overflow_no_tail instead of r_overflow in compute loop, this was |
| 1711 | // done since when iw == ur_w, ur_w_tail == 0 and thus |
| 1712 | // r_overflow_no_tail seems more appropriate |
| 1713 | body_l_overflow = l_overflow; |
| 1714 | body_r_overflow = r_overflow_no_tail; |
| 1715 | n_oi = 1; |
| 1716 | l_overflow = 0; |
| 1717 | r_overflow_no_tail = 0; |
| 1718 | } |
| 1719 | |
| 1720 | if (!threaded) { |
| 1721 | if (n_oi > 1) { mov(reg_oi, n_oi); } |
| 1722 | } else { |
| 1723 | // Setup for threaded code generation, and jump into the correct |
| 1724 | // portion of code for execution. |
| 1725 | head_thread = 0; |
| 1726 | tail_thread = nb_iw - 1; |
| 1727 | pretail_thread = tail_thread; |
| 1728 | |
| 1729 | int base_n_oi = iw_block / ur_w; |
| 1730 | head_n_oi = l_overflow > 0 ? base_n_oi - 1 : base_n_oi; |
| 1731 | tail_n_oi = (iw - iw_block * (nb_iw - 1)) / ur_w; |
| 1732 | pretail_n_oi = tail_n_oi; |
| 1733 | if (r_overflow_no_tail > 0) { |
| 1734 | if (tail_n_oi > 0) { |
| 1735 | pretail_n_oi--; |
| 1736 | tail_n_oi = pretail_n_oi; |
| 1737 | } else { |
| 1738 | // pretail_thread and tail_thread are different |
| 1739 | pretail_n_oi = base_n_oi - 1; |
| 1740 | pretail_thread = tail_thread - 1; |
| 1741 | } |
| 1742 | if (head_thread == pretail_thread) { |
| 1743 | head_n_oi--; |
| 1744 | pretail_n_oi = 0; |
| 1745 | tail_n_oi = 0; |
| 1746 | } |
| 1747 | } |
| 1748 | body_n_oi = (head_thread < pretail_thread - 1) ? base_n_oi : 0; |
| 1749 | |
| 1750 | // n_oi is used to determine how much control flow in the body portion |
| 1751 | // of the code needs generated. As such, n_oi needs to be set to the |
| 1752 | // maximum number of iterations it will be used the body code section. |
| 1753 | n_oi = nstl::max(body_n_oi, head_n_oi); |
| 1754 | n_oi = nstl::max(n_oi, pretail_n_oi); |
| 1755 | |
| 1756 | assert(iw_block % ur_w == 0); |
| 1757 | mov(reg_iwb, ptr[param1 + GET_OFF(iwb)]); |
| 1758 | |
| 1759 | if (head_n_oi != 0) mov(reg_oi, head_n_oi); |
| 1760 | cmp(reg_iwb, head_thread); |
| 1761 | je(head_label, T_NEAR); |
| 1762 | |
| 1763 | cmp(reg_iwb, pretail_thread); |
| 1764 | if (pretail_n_oi == 0) { |
| 1765 | je(pretail_label, T_NEAR); |
| 1766 | } else { |
| 1767 | mov(reg_oi, pretail_n_oi); |
| 1768 | je(body_label, T_NEAR); |
| 1769 | } |
| 1770 | if (pretail_thread != tail_thread) { |
| 1771 | cmp(reg_iwb, tail_thread); |
| 1772 | je(tail_label, T_NEAR); |
| 1773 | } |
| 1774 | if (body_n_oi != 0) { |
| 1775 | mov(reg_oi, body_n_oi); |
| 1776 | jmp(body_label, T_NEAR); |
| 1777 | } else { |
| 1778 | jmp(end_label, T_NEAR); |
| 1779 | } |
| 1780 | } |
| 1781 | L(head_label); |
| 1782 | if (l_overflow > 0) { |
| 1783 | compute_loop(ur_w, l_overflow, 0); |
| 1784 | if (threaded && head_n_oi == 0 && head_thread != pretail_thread) |
| 1785 | jmp(end_label, T_NEAR); |
| 1786 | else { |
| 1787 | add(reg_src, src_shift); |
| 1788 | add(reg_dst, dst_shift); |
| 1789 | } |
| 1790 | } |
| 1791 | L(body_label); |
| 1792 | if (n_oi > 0) { |
| 1793 | Label ow_loop_label; |
| 1794 | L(ow_loop_label); |
| 1795 | { |
| 1796 | compute_loop(ur_w, body_l_overflow, body_r_overflow); |
| 1797 | if (n_oi > 1 || r_overflow_no_tail > 0 || ur_w_tail != 0) { |
| 1798 | add(reg_src, src_shift); |
| 1799 | if (!jcp.large_w_filter) { add(reg_dst, dst_shift); } |
| 1800 | } |
| 1801 | if (n_oi > 1) { |
| 1802 | sub(reg_oi, 1); |
| 1803 | jg(ow_loop_label, T_NEAR); |
| 1804 | } |
| 1805 | } |
| 1806 | } |
| 1807 | if (threaded) { |
| 1808 | mov(reg_iwb, ptr[param1 + GET_OFF(iwb)]); |
| 1809 | cmp(reg_iwb, pretail_thread); |
| 1810 | jne(end_label, T_NEAR); |
| 1811 | } |
| 1812 | L(pretail_label); |
| 1813 | if (r_overflow_no_tail > 0) { |
| 1814 | compute_loop(ur_w, 0, r_overflow_no_tail); |
| 1815 | if (ur_w_tail != 0) { |
| 1816 | if (threaded && tail_thread != pretail_thread) |
| 1817 | jmp(end_label, T_NEAR); |
| 1818 | add(reg_src, src_shift); |
| 1819 | add(reg_dst, dst_shift); |
| 1820 | } |
| 1821 | } |
| 1822 | L(tail_label); |
| 1823 | if (ur_w_tail != 0) { |
| 1824 | /* if 'filter-width > ur_w' then the main loop only partially computes |
| 1825 | * width, ur_w_tail needs to offset the initial ur_w from the filter |
| 1826 | * address. */ |
| 1827 | if (jcp.large_w_filter) |
| 1828 | compute_loop(ur_w_tail, body_l_overflow, r_overflow - ur_w, ur_w); |
| 1829 | else |
| 1830 | compute_loop(ur_w_tail, 0, r_overflow); |
| 1831 | } |
| 1832 | L(end_label); |
| 1833 | |
| 1834 | postamble(); |
| 1835 | } |
| 1836 | |
| 1837 | status_t jit_avx512_common_conv_bwd_data_kernel_f32::init_conf( |
| 1838 | jit_conv_conf_t &jcp, const convolution_desc_t &cd, |
| 1839 | memory_desc_t &diff_src_md, memory_desc_t &weights_md, |
| 1840 | memory_desc_t &diff_dst_md, int nthreads) { |
| 1841 | if (!mayiuse(avx512_core)) return status::unimplemented; |
| 1842 | |
| 1843 | const memory_desc_wrapper diff_src_d(&diff_src_md); |
| 1844 | const memory_desc_wrapper weights_d(&weights_md); |
| 1845 | const memory_desc_wrapper diff_dst_d(&diff_dst_md); |
| 1846 | jcp = zero<decltype(jcp)>(); |
| 1847 | |
| 1848 | if (!everyone_is(data_type::f32, diff_dst_d.data_type(), |
| 1849 | weights_d.data_type(), diff_src_d.data_type())) |
| 1850 | return status::unimplemented; |
| 1851 | |
| 1852 | const bool with_groups = weights_d.ndims() == diff_src_d.ndims() + 1; |
| 1853 | int ndims = diff_src_d.ndims(); |
| 1854 | |
| 1855 | jcp.nthr = jcp.aligned_threads = nthreads; |
| 1856 | jcp.ndims = ndims; |
| 1857 | jcp.prop_kind = cd.prop_kind; |
| 1858 | |
| 1859 | jcp.ngroups = with_groups ? weights_d.dims()[0] : 1; |
| 1860 | jcp.mb = diff_src_d.dims()[0]; |
| 1861 | |
| 1862 | jcp.oc = diff_dst_d.dims()[1] / jcp.ngroups; |
| 1863 | jcp.oc_without_padding = jcp.oc; |
| 1864 | jcp.ic = diff_src_d.dims()[1] / jcp.ngroups; |
| 1865 | jcp.ic_without_padding = jcp.ic; |
| 1866 | |
| 1867 | jcp.id = (ndims == 5) ? diff_src_d.dims()[2] : 1; |
| 1868 | jcp.ih = (ndims == 3) ? 1 : diff_src_d.dims()[ndims - 2]; |
| 1869 | jcp.iw = diff_src_d.dims()[ndims - 1]; |
| 1870 | jcp.od = (ndims == 5) ? diff_dst_d.dims()[2] : 1; |
| 1871 | jcp.oh = (ndims == 3) ? 1 : diff_dst_d.dims()[ndims - 2]; |
| 1872 | jcp.ow = diff_dst_d.dims()[ndims - 1]; |
| 1873 | |
| 1874 | jcp.kd = (ndims == 5) ? weights_d.dims()[with_groups + 2] : 1; |
| 1875 | jcp.kh = (ndims == 3) ? 1 : weights_d.dims()[with_groups + ndims - 2]; |
| 1876 | jcp.kw = weights_d.dims()[with_groups + ndims - 1]; |
| 1877 | |
| 1878 | jcp.f_pad = (ndims == 5) ? cd.padding[0][0] : 0; |
| 1879 | jcp.t_pad = (ndims == 3) ? 0 : cd.padding[0][ndims - 4]; |
| 1880 | jcp.l_pad = cd.padding[0][ndims - 3]; |
| 1881 | |
| 1882 | jcp.stride_d = (ndims == 5) ? cd.strides[0] : 1; |
| 1883 | jcp.stride_h = (ndims == 3) ? 1 : cd.strides[ndims - 4]; |
| 1884 | jcp.stride_w = cd.strides[ndims - 3]; |
| 1885 | |
| 1886 | jcp.dilate_d = (ndims == 5) ? cd.dilates[0] : 0; |
| 1887 | jcp.dilate_h = (ndims == 3) ? 0 : cd.dilates[ndims - 4]; |
| 1888 | jcp.dilate_w = cd.dilates[ndims - 3]; |
| 1889 | if ((jcp.dilate_w != 0 && jcp.stride_w != 1) |
| 1890 | || (jcp.dilate_d != 0 && jcp.stride_d != 1) |
| 1891 | || (jcp.dilate_h != 0 && jcp.stride_h != 1)) |
| 1892 | return status::unimplemented; |
| 1893 | |
| 1894 | int ext_kw = calculate_extended_filter_size(jcp.kw, jcp.dilate_w); |
| 1895 | int ext_kh = calculate_extended_filter_size(jcp.kh, jcp.dilate_h); |
| 1896 | int ext_kd = calculate_extended_filter_size(jcp.kd, jcp.dilate_d); |
| 1897 | jcp.r_pad = calculate_end_padding( |
| 1898 | jcp.l_pad, jcp.ow, jcp.iw, jcp.stride_w, ext_kw); |
| 1899 | jcp.b_pad = calculate_end_padding( |
| 1900 | jcp.t_pad, jcp.oh, jcp.ih, jcp.stride_h, ext_kh); |
| 1901 | jcp.back_pad = calculate_end_padding( |
| 1902 | jcp.f_pad, jcp.od, jcp.id, jcp.stride_d, ext_kd); |
| 1903 | bool kernel_outside_src = false || ext_kw <= jcp.l_pad |
| 1904 | || ext_kw <= jcp.r_pad || ext_kh <= jcp.t_pad || ext_kh <= jcp.b_pad |
| 1905 | || ext_kd <= jcp.f_pad || ext_kd <= jcp.back_pad; |
| 1906 | if (kernel_outside_src) return status::unimplemented; |
| 1907 | |
| 1908 | jcp.aligned_threads = 0; |
| 1909 | const auto dat_tag_nxc = pick(ndims - 3, nwc, nhwc, ndhwc); |
| 1910 | const auto dat_tag_nCx4c = pick(ndims - 3, nCw4c, nChw4c, nCdhw4c); |
| 1911 | const auto dat_tag_nCx8c = pick(ndims - 3, nCw8c, nChw8c, nCdhw8c); |
| 1912 | const auto dat_tag_nCx16c = pick(ndims - 3, nCw16c, nChw16c, nCdhw16c); |
| 1913 | auto curr_src_tag = diff_src_d.matches_one_of_tag( |
| 1914 | dat_tag_nxc, dat_tag_nCx16c, dat_tag_nCx8c, dat_tag_nCx4c); |
| 1915 | auto curr_dst_tag = diff_dst_d.matches_one_of_tag( |
| 1916 | dat_tag_nxc, dat_tag_nCx16c, dat_tag_nCx8c, dat_tag_nCx4c); |
| 1917 | bool is_data_layout_nxc |
| 1918 | = IMPLICATION(curr_src_tag != dat_tag_nxc, |
| 1919 | diff_src_d.format_kind() == format_kind::any) |
| 1920 | && IMPLICATION(curr_dst_tag != dat_tag_nxc, |
| 1921 | diff_dst_d.format_kind() == format_kind::any) |
| 1922 | && utils::one_of(dat_tag_nxc, curr_src_tag, curr_dst_tag); |
| 1923 | |
| 1924 | jcp.is_1stconv = false; |
| 1925 | |
| 1926 | bool ok_to_pad_channels = true && !is_data_layout_nxc && jcp.ngroups == 1 |
| 1927 | && diff_src_d.data_type() == data_type::f32; |
| 1928 | |
| 1929 | const int full_simd_w = cpu_isa_traits<avx512_core>::vlen / typesize; |
| 1930 | jcp.simd_w = full_simd_w; |
| 1931 | bool ok_to_try_lower_zmm = true |
| 1932 | && IMPLICATION(is_data_layout_nxc, |
| 1933 | jcp.ic < full_simd_w && jcp.oc < full_simd_w |
| 1934 | && jcp.ngroups > 1) |
| 1935 | && mayiuse(avx512_core) && diff_src_d.data_type() == data_type::f32 |
| 1936 | && !jcp.is_1stconv |
| 1937 | && (jcp.oc % jcp.simd_w != 0 || jcp.ic % jcp.simd_w != 0) |
| 1938 | && !ok_to_pad_channels; |
| 1939 | |
| 1940 | if (ok_to_try_lower_zmm) { |
| 1941 | for (auto simd : {8, 4}) { |
| 1942 | if (jcp.ic % simd == 0 && jcp.oc % simd == 0) { |
| 1943 | jcp.simd_w = simd; |
| 1944 | break; |
| 1945 | } |
| 1946 | } |
| 1947 | } |
| 1948 | |
| 1949 | jcp.oc_block = jcp.simd_w; |
| 1950 | jcp.ic_block = jcp.is_1stconv ? jcp.ic : jcp.simd_w; |
| 1951 | |
| 1952 | if (ok_to_pad_channels) { |
| 1953 | jcp.oc = rnd_up(jcp.oc, jcp.oc_block); |
| 1954 | jcp.ic = rnd_up(jcp.ic, jcp.ic_block); |
| 1955 | } |
| 1956 | |
| 1957 | if (!IMPLICATION(!is_data_layout_nxc, |
| 1958 | jcp.oc % jcp.oc_block == 0 && jcp.ic % jcp.ic_block == 0)) |
| 1959 | return status::unimplemented; |
| 1960 | jcp.ic_tail = is_data_layout_nxc ? jcp.ic % jcp.simd_w : 0; |
| 1961 | jcp.oc_tail = is_data_layout_nxc ? jcp.oc % jcp.simd_w : 0; |
| 1962 | |
| 1963 | format_tag_t dat_tag, wei_tag; |
| 1964 | const auto nxc_tag = pick(ndims - 3, nwc, nhwc, ndhwc); |
| 1965 | |
| 1966 | if (jcp.simd_w == 8) { |
| 1967 | assert(with_groups); |
| 1968 | dat_tag = is_data_layout_nxc ? nxc_tag |
| 1969 | : pick(ndims - 3, nCw8c, nChw8c, nCdhw8c); |
| 1970 | wei_tag = pick(ndims - 3, gOIw8o8i, gOIhw8o8i, gOIdhw8o8i); |
| 1971 | } else if (jcp.simd_w == 4) { |
| 1972 | assert(with_groups); |
| 1973 | dat_tag = is_data_layout_nxc ? nxc_tag |
| 1974 | : pick(ndims - 3, nCw4c, nChw4c, nCdhw4c); |
| 1975 | wei_tag = pick(ndims - 3, gOIw4o4i, gOIhw4o4i, gOIdhw4o4i); |
| 1976 | } else { |
| 1977 | dat_tag = is_data_layout_nxc |
| 1978 | ? pick(ndims - 3, nwc, nhwc, ndhwc) |
| 1979 | : pick(ndims - 3, nCw16c, nChw16c, nCdhw16c); |
| 1980 | wei_tag = pick(2 * ndims - 6 + with_groups, OIw16o16i, gOIw16o16i, |
| 1981 | OIhw16o16i, gOIhw16o16i, OIdhw16o16i, gOIdhw16o16i); |
| 1982 | } |
| 1983 | |
| 1984 | if (diff_src_md.format_kind == format_kind::any) { |
| 1985 | CHECK(memory_desc_init_by_tag(diff_src_md, dat_tag)); |
| 1986 | } else if (curr_src_tag != dat_tag) |
| 1987 | return status::unimplemented; |
| 1988 | jcp.src_tag = dat_tag; |
| 1989 | |
| 1990 | if (diff_dst_md.format_kind == format_kind::any) { |
| 1991 | CHECK(memory_desc_init_by_tag(diff_dst_md, dat_tag)); |
| 1992 | } else if (curr_dst_tag != dat_tag) |
| 1993 | return status::unimplemented; |
| 1994 | jcp.dst_tag = dat_tag; |
| 1995 | |
| 1996 | if (init_tag(jcp.wei_tag, weights_md, weights_d, wei_tag) |
| 1997 | != status::success) |
| 1998 | return status::unimplemented; |
| 1999 | |
| 2000 | jcp.nb_ic = div_up(jcp.ic, jcp.ic_block); |
| 2001 | jcp.nb_oc = div_up(jcp.oc, jcp.oc_block); |
| 2002 | |
| 2003 | jcp.ur_w = jcp.stride_w; |
| 2004 | |
| 2005 | int regs = 28; |
| 2006 | if (jcp.iw <= regs) |
| 2007 | jcp.ur_w = jcp.iw; |
| 2008 | else { |
| 2009 | for (int ur_w = regs; ur_w > 0; --ur_w) |
| 2010 | if (ur_w % jcp.stride_w == 0) { |
| 2011 | jcp.ur_w = ur_w; |
| 2012 | break; |
| 2013 | } |
| 2014 | } |
| 2015 | int l_overflow = nstl::max( |
| 2016 | 0, ((jcp.kw - 1) * (jcp.dilate_w + 1) - jcp.l_pad) / jcp.stride_w); |
| 2017 | int r_overflow_no_tail = nstl::max(0, |
| 2018 | ((jcp.kw - 1) * (jcp.dilate_w + 1) |
| 2019 | - nstl::max(0, jcp.r_pad + jcp.iw % jcp.ur_w)) |
| 2020 | / jcp.stride_w); |
| 2021 | int n_oi = jcp.iw / jcp.ur_w; |
| 2022 | if (r_overflow_no_tail > 0) n_oi--; |
| 2023 | |
| 2024 | jcp.typesize_in = typesize; |
| 2025 | jcp.typesize_out = typesize; |
| 2026 | |
| 2027 | jcp.nb_ic_blocking = jcp.nb_oc_blocking = 1; |
| 2028 | |
| 2029 | // Heuristic to optimize code size on KNX |
| 2030 | bool large_code_size = (jcp.ur_w != jcp.ow) |
| 2031 | && ((l_overflow <= 0 && n_oi > 0) || (l_overflow > 0 && n_oi > 1)) |
| 2032 | && (r_overflow_no_tail > 0) && (l_overflow > 0); |
| 2033 | if (large_code_size) { |
| 2034 | const int max_code_size = 24 * 1024; |
| 2035 | const int num_ops_per_reg = 6 + jcp.oc_block * jcp.kw; |
| 2036 | int mult = 1; |
| 2037 | if (l_overflow > 0) mult += 1; |
| 2038 | if (r_overflow_no_tail > 0) mult += 1; |
| 2039 | for (int ur_w = jcp.ur_w; ur_w > regs / 2; --ur_w) { |
| 2040 | if ((ur_w / jcp.stride_w) * mult * num_ops_per_reg * 9.2 |
| 2041 | < max_code_size) { |
| 2042 | if (ur_w % jcp.stride_w == 0) { |
| 2043 | jcp.ur_w = ur_w; |
| 2044 | break; |
| 2045 | } |
| 2046 | } |
| 2047 | } |
| 2048 | } |
| 2049 | |
| 2050 | /* Support for large filter 'kw > 14' is only possible when ur_w is small |
| 2051 | * (e.g ur_w = 1) because of register allocation (max_reg = 31) */ |
| 2052 | const int min_filter_size = 14; |
| 2053 | /* Don't let JIT generate too big of a code which might result in an |
| 2054 | * out-of-memory crash. */ |
| 2055 | const int max_filter_size = 20; |
| 2056 | |
| 2057 | /* These conditions define a set of shapes with 'ow = 1' which |
| 2058 | * have a very limited optimization space for performance. |
| 2059 | * Optimize by using a targeted 'jcp.nb_ic_blocking' value. */ |
| 2060 | jcp.large_w_filter = jcp.kw >= min_filter_size && jcp.kw < max_filter_size |
| 2061 | && jcp.ow == 1 && jcp.nb_ic > 1 && jcp.kw == jcp.iw |
| 2062 | && jcp.stride_w == 1 |
| 2063 | && utils::everyone_is(0, jcp.dilate_d, jcp.dilate_h, jcp.dilate_w); |
| 2064 | |
| 2065 | const unsigned int L1_cache_size = platform::get_per_core_cache_size(1); |
| 2066 | if (mayiuse(avx512_core)) { |
| 2067 | int try_nb_ic_blocking = 2; |
| 2068 | unsigned int ker_inp_size = typesize * jcp.iw * jcp.ic_block |
| 2069 | * try_nb_ic_blocking * jcp.kh; |
| 2070 | unsigned int ker_out_size = typesize * jcp.ow * jcp.oc_block; |
| 2071 | unsigned int ker_wei_size = typesize * jcp.kh * jcp.kw * jcp.ic_block |
| 2072 | * jcp.oc_block * try_nb_ic_blocking; |
| 2073 | unsigned int ker_total_size |
| 2074 | = ker_inp_size + ker_out_size + ker_wei_size; |
| 2075 | bool use_expl_bcast |
| 2076 | = !(jcp.kw == 1 || (jcp.kw == 5 && jcp.iw < 8) |
| 2077 | || (jcp.kw < 5 |
| 2078 | && ((jcp.iw <= 5 |
| 2079 | || (jcp.iw > 8 && jcp.iw <= 13)) |
| 2080 | || ker_total_size > L1_cache_size))) |
| 2081 | || jcp.stride_h > 1 || jcp.stride_d > 1; |
| 2082 | if (use_expl_bcast && !jcp.large_w_filter) { |
| 2083 | jcp.kernel_kind = embd_bcast; |
| 2084 | jcp.ur_w = nstl::min(jcp.iw, regs); |
| 2085 | jcp.nb_ic_blocking = jcp.nb_oc_blocking = 1; |
| 2086 | if (!(jcp.kw > 3 |
| 2087 | || (jcp.kw == 3 && ker_total_size < L1_cache_size |
| 2088 | && jcp.ow > 8)) |
| 2089 | && jcp.stride_h == 1 && jcp.stride_d == 1) |
| 2090 | if (jcp.nb_ic % try_nb_ic_blocking == 0) { |
| 2091 | jcp.nb_ic_blocking = try_nb_ic_blocking; |
| 2092 | jcp.ur_w = 31 / (jcp.nb_ic_blocking + 1); |
| 2093 | if (jcp.iw < jcp.ur_w) jcp.ur_w = jcp.iw; |
| 2094 | } |
| 2095 | } else { |
| 2096 | jcp.kernel_kind = expl_bcast; |
| 2097 | jcp.nb_oc_blocking = 1; |
| 2098 | jcp.nb_ic_blocking = jcp.large_w_filter ? 2 : 4; |
| 2099 | if (jcp.nb_ic < jcp.nb_ic_blocking) jcp.nb_ic_blocking = jcp.nb_ic; |
| 2100 | if (jcp.nb_ic % jcp.nb_ic_blocking != 0) |
| 2101 | for (int i = jcp.nb_ic_blocking; i > 0; i--) |
| 2102 | if (jcp.nb_ic % i == 0) { |
| 2103 | jcp.nb_ic_blocking = i; |
| 2104 | break; |
| 2105 | } |
| 2106 | jcp.ur_w = 31 / (jcp.nb_ic_blocking + 1); |
| 2107 | if (jcp.iw < jcp.ur_w) jcp.ur_w = jcp.iw; |
| 2108 | } |
| 2109 | } |
| 2110 | jcp.ur_w_tail = jcp.iw % jcp.ur_w; |
| 2111 | |
| 2112 | auto is_iw_threading_applicable = [=]() { return one_of(jcp.ndims, 3, 4); }; |
| 2113 | |
| 2114 | auto get_thr_eff = [=](int nb_ic_blocking, int iw_block, int nthr) { |
| 2115 | // Cost heuristic for threading overhead. Determined using OMP. |
| 2116 | const float iw_block_cost = 32.0; |
| 2117 | |
| 2118 | int nb_iw = div_up(jcp.iw, iw_block); |
| 2119 | int nb_ic_chunks = div_up(jcp.nb_ic, nb_ic_blocking); |
| 2120 | int work_amount = jcp.mb * jcp.ih * nb_ic_chunks * nb_iw; |
| 2121 | float disbalance = (float)jcp.iw / rnd_up(jcp.iw, iw_block); |
| 2122 | float block_overhead = nstl::max(0.0f, 1.0f - iw_block_cost / iw_block); |
| 2123 | float thr_eff = block_overhead * disbalance |
| 2124 | * ((float)work_amount / rnd_up(work_amount, nthr)); |
| 2125 | return thr_eff; |
| 2126 | }; |
| 2127 | |
| 2128 | auto get_iw_block = [=](int nb_ic_blocking, int ur_w, float &eff, |
| 2129 | int nthr) { |
| 2130 | int res_iw_block = jcp.iw; |
| 2131 | if (!is_iw_threading_applicable()) return res_iw_block; |
| 2132 | |
| 2133 | int max_nb_iw = div_up(jcp.iw, 2 * ur_w); |
| 2134 | int iw_block_thr; |
| 2135 | |
| 2136 | if (jcp.ndims == 3) { |
| 2137 | // Blocking optimization to prevent data from leaving cache This |
| 2138 | // blocking optimization does not handle height blocking, so it does |
| 2139 | // not apply to higher dimensions. |
| 2140 | // TODO: Implement a more general optimization taking into account |
| 2141 | // the height dimension. |
| 2142 | int L2_part |
| 2143 | = (platform::get_per_core_cache_size(2) * 7 / 8) / typesize; |
| 2144 | int size_diff_src_chunk = jcp.ic_block * nb_ic_blocking * ur_w; |
| 2145 | int size_diff_dst_chunk = jcp.oc_block * ur_w; |
| 2146 | int size_wei_chunk |
| 2147 | = jcp.ic_block * nb_ic_blocking * jcp.oc_block * jcp.kw; |
| 2148 | int nurw_cache = (L2_part - 2 * size_wei_chunk) |
| 2149 | / (2 * size_diff_dst_chunk + 2 * size_diff_src_chunk); |
| 2150 | // current design of generate() requires iw_block >= 2 * ur_w |
| 2151 | int iw_block_cache = ur_w * nstl::max(2, nurw_cache); |
| 2152 | |
| 2153 | iw_block_thr = iw_block_cache; |
| 2154 | } else |
| 2155 | iw_block_thr = jcp.iw; |
| 2156 | eff = get_thr_eff(nb_ic_blocking, iw_block_thr, nthr); |
| 2157 | |
| 2158 | // Search for most efficient threading over iw_blocks. |
| 2159 | int start_nb_iw = div_up(jcp.iw, iw_block_thr); |
| 2160 | for (int nb_iw = start_nb_iw; nb_iw <= max_nb_iw; nb_iw++) { |
| 2161 | float eff_threshold = 0.98f; |
| 2162 | if (eff > eff_threshold) break; |
| 2163 | int iw_block |
| 2164 | = nstl::min(rnd_up(div_up(jcp.iw, nb_iw), ur_w), jcp.iw); |
| 2165 | if (div_up(jcp.iw, iw_block) != nb_iw) continue; |
| 2166 | float thr_eff = get_thr_eff(nb_ic_blocking, iw_block, nthr); |
| 2167 | if (iw_block >= 2 * ur_w && thr_eff > eff) { |
| 2168 | iw_block_thr = iw_block; |
| 2169 | eff = thr_eff; |
| 2170 | } |
| 2171 | } |
| 2172 | res_iw_block = nstl::min(jcp.iw, nstl::max(2 * ur_w, iw_block_thr)); |
| 2173 | return res_iw_block; |
| 2174 | }; |
| 2175 | |
| 2176 | float thr_eff = -1.0f; |
| 2177 | jcp.iw_block |
| 2178 | = get_iw_block(jcp.nb_ic_blocking, jcp.ur_w, thr_eff, jcp.nthr); |
| 2179 | jcp.nb_iw = div_up(jcp.iw, jcp.iw_block); |
| 2180 | |
| 2181 | /* adjust the thread decomposition |
| 2182 | * to improve the thr_eff for small size problem |
| 2183 | * the threshold L1_cache_size is empirical */ |
| 2184 | size_t wei_size |
| 2185 | = (size_t)typesize * jcp.ic * jcp.oc * jcp.kh * jcp.kw * jcp.kd; |
| 2186 | size_t out_size = (size_t)jcp.mb * jcp.typesize_out * jcp.oc * jcp.oh |
| 2187 | * jcp.ow * jcp.od; |
| 2188 | size_t inp_size = (size_t)jcp.mb * jcp.typesize_in * jcp.ic * jcp.ih |
| 2189 | * jcp.iw * jcp.id; |
| 2190 | size_t total_size = jcp.ngroups * (wei_size + out_size + inp_size); |
| 2191 | |
| 2192 | if (jcp.ngroups < jcp.nthr && (total_size < L1_cache_size)) { |
| 2193 | int iw_block = jcp.iw_block; |
| 2194 | int end_nthr = with_groups ? jcp.ngroups : ndims - 2; |
| 2195 | float eff = -1.0f; |
| 2196 | float best_thr_eff = -1.0f; |
| 2197 | // When thr_eff equals zero (cannot get the proper effciency) |
| 2198 | // simply set the thread as 4 now |
| 2199 | // And update the eff inside get_iw_block to avoid redundant |
| 2200 | // computation when thr_eff != 0 but current eff == 0 |
| 2201 | if (thr_eff == 0.f) { |
| 2202 | jcp.nthr = nstl::min(jcp.nthr, 4); |
| 2203 | } else { |
| 2204 | for (int nthr = jcp.nthr / 2; nthr >= end_nthr; nthr--) { |
| 2205 | iw_block |
| 2206 | = get_iw_block(jcp.nb_ic_blocking, jcp.ur_w, eff, nthr); |
| 2207 | if (eff > 1.1f * best_thr_eff) { |
| 2208 | best_thr_eff = eff; |
| 2209 | jcp.iw_block = iw_block; |
| 2210 | jcp.nb_iw = div_up(jcp.iw, jcp.iw_block); |
| 2211 | jcp.nthr = jcp.aligned_threads = nthr; |
| 2212 | if (best_thr_eff > 0.98f) break; |
| 2213 | } |
| 2214 | } |
| 2215 | } |
| 2216 | } |
| 2217 | |
| 2218 | if (l_overflow * jcp.stride_w > jcp.ur_w && !jcp.large_w_filter) |
| 2219 | return status::unimplemented; |
| 2220 | |
| 2221 | r_overflow_no_tail = nstl::max(0, |
| 2222 | ((jcp.kw - 1) * (jcp.dilate_w + 1) |
| 2223 | - nstl::max(0, jcp.r_pad + jcp.ur_w_tail)) |
| 2224 | / jcp.stride_w); |
| 2225 | bool tails_not_ok = false |
| 2226 | /* maximum 1 ur_w block with r_overflow so far */ |
| 2227 | || r_overflow_no_tail * jcp.stride_w > jcp.ur_w |
| 2228 | /* ur_w must be a multiple of stride */ |
| 2229 | || ((jcp.iw > jcp.ur_w) && (jcp.ur_w % jcp.stride_w != 0)) |
| 2230 | /* r_pad must not extend beyond ur_w_tail */ |
| 2231 | || ((jcp.iw > jcp.ur_w) && (jcp.r_pad + jcp.ur_w_tail < 0)); |
| 2232 | if (tails_not_ok) return status::unimplemented; |
| 2233 | |
| 2234 | pick_loop_order(jcp); |
| 2235 | |
| 2236 | jcp.nb_oc_L2 = jcp.nb_oc; |
| 2237 | if (is_data_layout_nxc) { |
| 2238 | // TODO: improve L2 blocking for large OC |
| 2239 | const int nb_oc_theshold_L2 = 32; |
| 2240 | if (jcp.nb_oc > nb_oc_theshold_L2 && jcp.nb_oc < 2 * nb_oc_theshold_L2) |
| 2241 | jcp.nb_oc_L2 = div_up(jcp.nb_oc, 2); |
| 2242 | else |
| 2243 | jcp.nb_oc_L2 = nstl::min(nb_oc_theshold_L2, jcp.nb_oc); |
| 2244 | } |
| 2245 | |
| 2246 | bool args_ok = true && jcp.ic <= diff_src_d.padded_dims()[1] |
| 2247 | && jcp.oc <= diff_dst_d.padded_dims()[1] |
| 2248 | && jcp.ic <= weights_d.padded_dims()[with_groups + 1] |
| 2249 | && jcp.oc <= weights_d.padded_dims()[with_groups + 0]; |
| 2250 | if (!args_ok) return status::unimplemented; |
| 2251 | |
| 2252 | // A rough check on code size |
| 2253 | // TODO: come up with a tighter bound |
| 2254 | { |
| 2255 | const int max_code_size = 256 * 1024; // default size of jit generator |
| 2256 | int mult = 1 + (l_overflow > 0) + (r_overflow_no_tail > 0); |
| 2257 | const float max_instruction_size = 15; |
| 2258 | float ur_fac |
| 2259 | = (float)jcp.kw * jcp.oc_block * jcp.nb_ic_blocking * jcp.ur_w; |
| 2260 | float code_size = mult * ur_fac * max_instruction_size; |
| 2261 | if (code_size > max_code_size && !jcp.large_w_filter) |
| 2262 | return status::unimplemented; |
| 2263 | } |
| 2264 | |
| 2265 | return status::success; |
| 2266 | } |
| 2267 | |
| 2268 | void jit_avx512_common_conv_bwd_data_kernel_f32::init_scratchpad( |
| 2269 | memory_tracking::registrar_t &scratchpad, const jit_conv_conf_t &jcp) { |
| 2270 | UNUSED(scratchpad); |
| 2271 | UNUSED(jcp); |
| 2272 | } |
| 2273 | |
| 2274 | // Initialize static data members |
| 2275 | const int jit_avx512_common_conv_bwd_weights_kernel_f32::max_ur_w = 28; |
| 2276 | const int jit_avx512_common_conv_bwd_weights_kernel_f32::min_oh_reduce = 9; |
| 2277 | |
| 2278 | void jit_avx512_common_conv_bwd_weights_kernel_f32:: |
| 2279 | od_step_comeback_pointers() { |
| 2280 | Label kd_comeback_label; |
| 2281 | |
| 2282 | /* 'depth' loop count bound by 'kd_work_size' */ |
| 2283 | mov(kj, reg_kd_count); |
| 2284 | L(kd_comeback_label); |
| 2285 | { |
| 2286 | int inp_mult = is_src_layout_nxc() |
| 2287 | ? jcp.ngroups * jcp.ic |
| 2288 | : (jcp.is_1stconv ? 1 : jcp.ic_block); |
| 2289 | int iw = jcp.iw; |
| 2290 | sub(reg_input, |
| 2291 | jcp.typesize_in * (jcp.dilate_d + 1) * jcp.ih * iw * inp_mult); |
| 2292 | sub(reg_kernel, |
| 2293 | jcp.typesize_out * jcp.kh * jcp.kw * jcp.ic_block |
| 2294 | * jcp.oc_block); |
| 2295 | dec(kj); |
| 2296 | cmp(kj, 0); |
| 2297 | jg(kd_comeback_label, T_NEAR); |
| 2298 | } |
| 2299 | } |
| 2300 | |
| 2301 | void jit_avx512_common_conv_bwd_weights_kernel_f32:: |
| 2302 | oh_step_comeback_pointers() { |
| 2303 | Label kh_comeback_label, kd_comeback_label; |
| 2304 | mov(kj, reg_kh); |
| 2305 | L(kh_comeback_label); |
| 2306 | { |
| 2307 | int kw = jcp.is_hw_transp ? 1 : jcp.kw; |
| 2308 | int inp_mult = is_src_layout_nxc() |
| 2309 | ? jcp.ngroups * jcp.ic |
| 2310 | : (jcp.is_1stconv ? 1 : jcp.ic_block); |
| 2311 | int iw = jcp.is_hw_transp ? 1 : jcp.iw; |
| 2312 | sub(reg_input, jcp.typesize_in * (jcp.dilate_h + 1) * iw * inp_mult); |
| 2313 | sub(reg_kernel, jcp.typesize_out * kw * jcp.ic_block * jcp.oc_block); |
| 2314 | dec(kj); |
| 2315 | cmp(kj, 0); |
| 2316 | jg(kh_comeback_label, T_NEAR); |
| 2317 | } |
| 2318 | } |
| 2319 | |
| 2320 | void jit_avx512_common_conv_bwd_weights_kernel_f32::compute_ic_block_step_fma( |
| 2321 | int ur_w, int pad_l, int pad_r, int ic_block_step, int input_offset, |
| 2322 | int kernel_offset, int output_offset, bool input_wraparound) { |
| 2323 | |
| 2324 | int kw = jcp.is_hw_transp ? jcp.tr_kw : jcp.kw; |
| 2325 | int iw = jcp.is_hw_transp ? jcp.tr_iw : jcp.iw; |
| 2326 | int kw_tr_mult = jcp.is_hw_transp ? jcp.kw : 1; |
| 2327 | int ic_block = jcp.ic_block; |
| 2328 | int oc_block = jcp.oc_block; |
| 2329 | auto get_ker_offt = [=](int i_kw, int i_ic) { |
| 2330 | return typesize * (i_kw * kw_tr_mult * ic_block + i_ic) * jcp.oc_block |
| 2331 | + kernel_offset; |
| 2332 | }; |
| 2333 | for (int i_kw = 0; i_kw < kw; i_kw++) |
| 2334 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) |
| 2335 | vmovups(Zmm(i_kw * ic_block_step + i_ic), |
| 2336 | EVEX_compress_addr(reg_kernel, get_ker_offt(i_kw, i_ic))); |
| 2337 | const int out_mult = is_ddst_layout_nxc() ? jcp.ngroups * jcp.oc : oc_block; |
| 2338 | const int oc_tail = jcp.oc_tail; |
| 2339 | |
| 2340 | for (int i_ur = 0; i_ur < ur_w; i_ur++) { |
| 2341 | const int ddst_pipeline_start_idx = ic_block_step * kw; |
| 2342 | const int ddst_pipeline_len = 4; |
| 2343 | auto get_ddst_reg_idx = [=](int ur_idx) { |
| 2344 | return ddst_pipeline_start_idx + (ur_idx) % ddst_pipeline_len; |
| 2345 | }; |
| 2346 | auto get_ddst_offt = [=](int ur_idx) { |
| 2347 | return typesize * ur_idx * out_mult + output_offset; |
| 2348 | }; |
| 2349 | |
| 2350 | if (i_ur == 0) { |
| 2351 | for (int i = 0; i < nstl::min(ddst_pipeline_len, ur_w); i++) { |
| 2352 | int ur_idx = i_ur + i; |
| 2353 | auto zmm_ddst = Zmm(get_ddst_reg_idx(ur_idx)); |
| 2354 | if (oc_tail) zmm_ddst = zmm_ddst | k_oc_mask | T_z; |
| 2355 | vmovups(zmm_ddst, |
| 2356 | EVEX_compress_addr(reg_output, get_ddst_offt(ur_idx))); |
| 2357 | } |
| 2358 | } else if (i_ur + ddst_pipeline_len - 1 < ur_w) { |
| 2359 | |
| 2360 | int ur_idx = i_ur + ddst_pipeline_len - 1; |
| 2361 | |
| 2362 | auto zmm_ddst = Zmm(get_ddst_reg_idx(ur_idx)); |
| 2363 | if (oc_tail) zmm_ddst = zmm_ddst | k_oc_mask | T_z; |
| 2364 | vmovups(zmm_ddst, |
| 2365 | EVEX_compress_addr(reg_output, get_ddst_offt(ur_idx))); |
| 2366 | } |
| 2367 | |
| 2368 | for (int i_kw = 0; i_kw < kw; i_kw++) { |
| 2369 | int i_iw = get_iw_idx(i_ur, i_kw, pad_l); |
| 2370 | if (i_iw < 0 || i_iw > get_iw_idx(ur_w - 1, kw - 1, pad_l) - pad_r |
| 2371 | || get_iw_idx(i_ur, i_kw, jcp.l_pad) >= iw) |
| 2372 | continue; |
| 2373 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) { |
| 2374 | vfmadd231ps(Zmm(i_kw * ic_block_step + i_ic), |
| 2375 | Zmm(get_ddst_reg_idx(i_ur)), |
| 2376 | EVEX_compress_addr_safe(reg_input, |
| 2377 | get_full_src_offset(i_iw, i_ic, input_offset), |
| 2378 | reg_long_offt, true)); |
| 2379 | } |
| 2380 | } |
| 2381 | } |
| 2382 | |
| 2383 | for (int i_kw = 0; i_kw < kw; i_kw++) |
| 2384 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) |
| 2385 | vmovups(EVEX_compress_addr(reg_kernel, get_ker_offt(i_kw, i_ic)), |
| 2386 | Zmm(i_kw * ic_block_step + i_ic)); |
| 2387 | } |
| 2388 | |
| 2389 | void jit_avx512_common_conv_bwd_weights_kernel_f32:: |
| 2390 | compute_ic_block_step_fma_expl(int ur_w, int pad_l, int pad_r, |
| 2391 | int ic_block_step, int input_offset, int kernel_offset, |
| 2392 | int output_offset, bool input_wraparound) { |
| 2393 | int kw = jcp.kw; |
| 2394 | int ic_block = jcp.ic_block; |
| 2395 | int oc_block = jcp.oc_block; |
| 2396 | const int oc_tail = jcp.oc_tail; |
| 2397 | const bool ddst_layout_nxc = is_ddst_layout_nxc(); |
| 2398 | const int max_regs = 32; |
| 2399 | const int ddst_pipeline_start_idx = 2 * ic_block_step * kw; |
| 2400 | const int ddst_pipeline_len |
| 2401 | = ddst_layout_nxc ? 1 : max_regs - ddst_pipeline_start_idx; |
| 2402 | const int iw_last_value = get_iw_idx(ur_w - 1, kw - 1, pad_l) - pad_r; |
| 2403 | assert(jcp.stride_w == 1 && jcp.dilate_w == 0 && ddst_pipeline_len > 0 |
| 2404 | && jcp.kernel_kind == expl_bcast); |
| 2405 | |
| 2406 | const int out_mult = ddst_layout_nxc ? jcp.ngroups * jcp.oc : oc_block; |
| 2407 | auto get_diff_wei_reg_idx |
| 2408 | = [=](int i_kw, int i_ic) { return i_kw * ic_block_step + i_ic; }; |
| 2409 | auto get_src_reg_idx = [=](int i_iw, int i_ic) { |
| 2410 | return kw * ic_block_step + ((i_iw + pad_l) % kw) * ic_block_step |
| 2411 | + i_ic; |
| 2412 | }; |
| 2413 | auto get_diff_dst_reg_idx = [=](int i_ur) { |
| 2414 | return ddst_pipeline_start_idx + i_ur % ddst_pipeline_len; |
| 2415 | }; |
| 2416 | |
| 2417 | for (int i_kw = 0; i_kw < kw; i_kw++) |
| 2418 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) { |
| 2419 | auto zmm_ker = Zmm(get_diff_wei_reg_idx(i_kw, i_ic)); |
| 2420 | vpxord(zmm_ker, zmm_ker, zmm_ker); |
| 2421 | } |
| 2422 | |
| 2423 | for (int i_ur = 0; i_ur < ur_w; i_ur++) { |
| 2424 | if (i_ur == 0) { |
| 2425 | for (int i = 0; i < nstl::min(ddst_pipeline_len, ur_w); i++) { |
| 2426 | auto addr_out = EVEX_compress_addr( |
| 2427 | reg_output, typesize * i * out_mult + output_offset); |
| 2428 | auto zmm_ddst = Zmm(get_diff_dst_reg_idx(i)); |
| 2429 | if (oc_tail) zmm_ddst = zmm_ddst | k_oc_mask | T_z; |
| 2430 | vmovups(zmm_ddst, addr_out); |
| 2431 | } |
| 2432 | |
| 2433 | for (int i_kw = 0; i_kw < kw; i_kw++) { |
| 2434 | int i_iw = get_iw_idx(0, i_kw, pad_l); |
| 2435 | if (i_iw < 0 || i_iw > iw_last_value) continue; |
| 2436 | |
| 2437 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) { |
| 2438 | auto addr_inp = EVEX_compress_addr_safe(reg_input, |
| 2439 | get_full_src_offset(i_iw, i_ic, input_offset), |
| 2440 | reg_long_offt); |
| 2441 | vbroadcastss(Zmm(get_src_reg_idx(i_iw, i_ic)), addr_inp); |
| 2442 | } |
| 2443 | } |
| 2444 | } else { |
| 2445 | int diff_dst_load_idx = i_ur + ddst_pipeline_len - 1; |
| 2446 | if (diff_dst_load_idx < ur_w) { |
| 2447 | auto addr_out = EVEX_compress_addr(reg_output, |
| 2448 | typesize * diff_dst_load_idx * out_mult |
| 2449 | + output_offset); |
| 2450 | auto zmm_ddst = Zmm(get_diff_dst_reg_idx(diff_dst_load_idx)); |
| 2451 | if (oc_tail) zmm_ddst = zmm_ddst | k_oc_mask | T_z; |
| 2452 | vmovups(zmm_ddst, addr_out); |
| 2453 | } |
| 2454 | |
| 2455 | int i_iw = get_iw_idx(i_ur, kw - 1, pad_l); |
| 2456 | if (i_iw >= 0 && i_iw <= iw_last_value) { |
| 2457 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) { |
| 2458 | auto addr_inp = EVEX_compress_addr_safe(reg_input, |
| 2459 | get_full_src_offset(i_iw, i_ic, input_offset), |
| 2460 | reg_long_offt); |
| 2461 | vbroadcastss(Zmm(get_src_reg_idx(i_iw, i_ic)), addr_inp); |
| 2462 | } |
| 2463 | } |
| 2464 | } |
| 2465 | for (int i_kw = 0; i_kw < kw; i_kw++) { |
| 2466 | int i_iw = get_iw_idx(i_ur, i_kw, pad_l); |
| 2467 | if (i_iw < 0 || i_iw > iw_last_value) continue; |
| 2468 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) { |
| 2469 | vfmadd231ps(Zmm(get_diff_wei_reg_idx(i_kw, i_ic)), |
| 2470 | Zmm(get_diff_dst_reg_idx(i_ur)), |
| 2471 | Zmm(get_src_reg_idx(i_iw, i_ic))); |
| 2472 | } |
| 2473 | } |
| 2474 | } |
| 2475 | |
| 2476 | for (int i_kw = 0; i_kw < kw; i_kw++) |
| 2477 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) { |
| 2478 | auto addr_ker = EVEX_compress_addr(reg_kernel, |
| 2479 | typesize * (i_kw * ic_block + i_ic) * jcp.oc_block |
| 2480 | + kernel_offset); |
| 2481 | vaddps(Zmm(get_diff_wei_reg_idx(i_kw, i_ic)), addr_ker); |
| 2482 | } |
| 2483 | |
| 2484 | for (int i_kw = 0; i_kw < kw; i_kw++) |
| 2485 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) { |
| 2486 | auto addr_ker = EVEX_compress_addr(reg_kernel, |
| 2487 | typesize * (i_kw * ic_block + i_ic) * jcp.oc_block |
| 2488 | + kernel_offset); |
| 2489 | vmovups(addr_ker, Zmm(get_diff_wei_reg_idx(i_kw, i_ic))); |
| 2490 | } |
| 2491 | } |
| 2492 | |
| 2493 | void jit_avx512_common_conv_bwd_weights_kernel_f32::compute_ic_block_step( |
| 2494 | int ur_w, int pad_l, int pad_r, int ic_block_step, int input_offset, |
| 2495 | int kernel_offset, int output_offset, bool input_wraparound) { |
| 2496 | if (jcp.kernel_kind == expl_bcast) |
| 2497 | compute_ic_block_step_fma_expl(ur_w, pad_l, pad_r, ic_block_step, |
| 2498 | input_offset, kernel_offset, output_offset, input_wraparound); |
| 2499 | else |
| 2500 | compute_ic_block_step_fma(ur_w, pad_l, pad_r, ic_block_step, |
| 2501 | input_offset, kernel_offset, output_offset, input_wraparound); |
| 2502 | } |
| 2503 | |
| 2504 | void jit_avx512_common_conv_bwd_weights_kernel_f32 :: |
| 2505 | compute_oh_step_unroll_ow_icblock(int ic_block_step, int max_ur_w) { |
| 2506 | UNUSED(max_ur_w); |
| 2507 | |
| 2508 | Label kh_label, kd_label; |
| 2509 | |
| 2510 | int ic_block = jcp.ic_block; |
| 2511 | int oc_block = jcp.oc_block; |
| 2512 | const bool src_layout_nxc = is_src_layout_nxc(); |
| 2513 | int inp_mul = src_layout_nxc ? jcp.ngroups * jcp.ic |
| 2514 | : (!jcp.is_1stconv ? ic_block : 1); |
| 2515 | int iw = jcp.iw; |
| 2516 | |
| 2517 | int r_pad = nstl::max(0, jcp.r_pad); |
| 2518 | int l_pad = jcp.l_pad; |
| 2519 | |
| 2520 | if (jcp.ndims == 5) { |
| 2521 | L(kd_label); |
| 2522 | mov(reg_input, aux_reg_input); |
| 2523 | mov(reg_kernel, aux_reg_kernel); |
| 2524 | } |
| 2525 | |
| 2526 | const int ic_tail = jcp.ic_tail; |
| 2527 | const bool generate_icb_loop = jcp.nb_ic_blocking_max > 1; |
| 2528 | mov(kj, reg_kh); |
| 2529 | L(kh_label); |
| 2530 | { |
| 2531 | Label icb_block_label, icb_block_label_cb, ic_tail_loop, ic_tail_label; |
| 2532 | if (generate_icb_loop || ic_tail) { |
| 2533 | push(reg_input); |
| 2534 | push(reg_kernel); |
| 2535 | mov(reg_icb, ptr[param + GET_OFF(reduce_work)]); |
| 2536 | } |
| 2537 | |
| 2538 | if (ic_tail) { |
| 2539 | cmp(reg_icb, ic_block); |
| 2540 | jl(ic_tail_loop, T_NEAR); |
| 2541 | } |
| 2542 | |
| 2543 | const int ic_tail_loop_work = rnd_dn(ic_tail, ic_block_step); |
| 2544 | Label icb_block_label_end; |
| 2545 | L(icb_block_label); |
| 2546 | for (int i_b_ic = 0; i_b_ic < jcp.ic_block; i_b_ic += ic_block_step) { |
| 2547 | const int input_offset = jcp.typesize_in * i_b_ic; |
| 2548 | compute_ic_block_step(jcp.ur_w, l_pad, r_pad, ic_block_step, |
| 2549 | input_offset, jcp.typesize_out * i_b_ic * jcp.oc_block, 0, |
| 2550 | i_b_ic + ic_block_step >= jcp.ic_block); |
| 2551 | if (generate_icb_loop || ic_tail) sub(reg_icb, ic_block_step); |
| 2552 | if (ic_tail && i_b_ic + ic_block_step == ic_tail_loop_work) { |
| 2553 | cmp(reg_icb, ic_block_step); |
| 2554 | jl(icb_block_label_end, T_NEAR); |
| 2555 | } |
| 2556 | } |
| 2557 | L(icb_block_label_end); |
| 2558 | |
| 2559 | const int input_icb_shift = jcp.typesize_in * ic_block; |
| 2560 | const size_t kernel_icb_shift = (size_t)jcp.typesize_out * jcp.kd |
| 2561 | * jcp.kh * jcp.kw * ic_block * oc_block; |
| 2562 | |
| 2563 | if (generate_icb_loop) { |
| 2564 | // icb loop supported for src in nxc layout only |
| 2565 | assert(src_layout_nxc); |
| 2566 | add(reg_input, input_icb_shift); |
| 2567 | safe_add(reg_kernel, kernel_icb_shift, reg_long_offt); |
| 2568 | cmp(reg_icb, ic_block); |
| 2569 | jge(icb_block_label, T_NEAR); |
| 2570 | } |
| 2571 | |
| 2572 | if (ic_tail) { |
| 2573 | L(ic_tail_loop); |
| 2574 | Label skip_ic_tail; |
| 2575 | cmp(reg_icb, 0); |
| 2576 | jle(skip_ic_tail, T_NEAR); |
| 2577 | if (ic_tail_loop_work) { |
| 2578 | cmp(reg_icb, ic_tail_loop_work); |
| 2579 | jge(icb_block_label, T_NEAR); |
| 2580 | if (generate_icb_loop) { |
| 2581 | // compensate offset added in generate_icb_loop |
| 2582 | sub(reg_input, input_icb_shift); |
| 2583 | safe_sub(reg_kernel, kernel_icb_shift, reg_long_offt); |
| 2584 | } |
| 2585 | } |
| 2586 | |
| 2587 | L(ic_tail_label); |
| 2588 | if (ic_tail % ic_block_step) { |
| 2589 | cmp(reg_icb, 0); |
| 2590 | jle(skip_ic_tail, T_NEAR); |
| 2591 | const int i_b_ic = ic_tail_loop_work; |
| 2592 | const int input_offset = jcp.typesize_in * i_b_ic; |
| 2593 | compute_ic_block_step(jcp.ur_w, l_pad, r_pad, |
| 2594 | ic_tail % ic_block_step, input_offset, |
| 2595 | jcp.typesize_out * i_b_ic * jcp.oc_block, 0); |
| 2596 | } |
| 2597 | L(skip_ic_tail); |
| 2598 | } |
| 2599 | |
| 2600 | if (generate_icb_loop || ic_tail) { |
| 2601 | pop(reg_kernel); |
| 2602 | pop(reg_input); |
| 2603 | } |
| 2604 | |
| 2605 | add(reg_input, jcp.typesize_in * (jcp.dilate_h + 1) * iw * inp_mul); |
| 2606 | add(reg_kernel, jcp.typesize_out * jcp.kw * ic_block * oc_block); |
| 2607 | dec(kj); |
| 2608 | cmp(kj, 0); |
| 2609 | jg(kh_label, T_NEAR); |
| 2610 | } |
| 2611 | |
| 2612 | if (jcp.ndims == 5) { |
| 2613 | add(aux_reg_input, |
| 2614 | jcp.typesize_in * (jcp.dilate_d + 1) * jcp.ih * iw * inp_mul); |
| 2615 | add(aux_reg_kernel, |
| 2616 | jcp.typesize_out * jcp.kh * jcp.kw * ic_block * oc_block); |
| 2617 | dec(ki); |
| 2618 | cmp(ki, 0); |
| 2619 | jg(kd_label, T_NEAR); |
| 2620 | } |
| 2621 | } |
| 2622 | |
| 2623 | void jit_avx512_common_conv_bwd_weights_kernel_f32 ::compute_oh_step_unroll_ow( |
| 2624 | int ic_block_step, int max_ur_w) { |
| 2625 | Label kh_label, ic_block_label, ic_tail_loop_label, ic_tail_label, kd_label; |
| 2626 | const bool src_layout_nxc = is_src_layout_nxc(); |
| 2627 | int inp_mul = src_layout_nxc ? jcp.ngroups * jcp.ic |
| 2628 | : (!jcp.is_1stconv ? jcp.ic_block : 1); |
| 2629 | const int ic_tail = jcp.ic_tail; |
| 2630 | UNUSED(max_ur_w); |
| 2631 | |
| 2632 | int ic_block = jcp.ic_block; |
| 2633 | int oc_block = jcp.oc_block; |
| 2634 | |
| 2635 | int inp_icb_sp_stride = jcp.is_hw_transp ? 1 : jcp.iw; |
| 2636 | int ow = jcp.is_hw_transp ? jcp.oh : jcp.ow; |
| 2637 | |
| 2638 | int r_pad = nstl::max(0, jcp.r_pad); |
| 2639 | int l_pad = jcp.l_pad; |
| 2640 | |
| 2641 | if (jcp.ndims == 5) { |
| 2642 | L(kd_label); |
| 2643 | mov(reg_input, aux_reg_input); |
| 2644 | mov(reg_kernel, aux_reg_kernel); |
| 2645 | } |
| 2646 | |
| 2647 | const bool generate_icb_loop = jcp.nb_ic_blocking_max > 1; |
| 2648 | mov(kj, reg_kh); |
| 2649 | L(kh_label); |
| 2650 | { |
| 2651 | Label icb_block_label; |
| 2652 | if (generate_icb_loop || ic_tail) { |
| 2653 | push(reg_input); |
| 2654 | push(reg_kernel); |
| 2655 | mov(reg_icb, ptr[param + GET_OFF(reduce_work)]); |
| 2656 | } |
| 2657 | |
| 2658 | if (ic_tail) { |
| 2659 | cmp(reg_icb, ic_block); |
| 2660 | jl(ic_tail_loop_label, T_NEAR); |
| 2661 | } |
| 2662 | |
| 2663 | L(icb_block_label); |
| 2664 | Label icb_block_label_end; |
| 2665 | mov(b_ic, ic_block); |
| 2666 | L(ic_block_label); |
| 2667 | { |
| 2668 | compute_ic_block_step(ow, l_pad, r_pad, ic_block_step, 0, 0, 0); |
| 2669 | size_t inp_icblk_stride = jcp.is_1stconv && !src_layout_nxc |
| 2670 | ? (size_t)jcp.ih * jcp.iw * jcp.id |
| 2671 | : 1; |
| 2672 | size_t input_offset |
| 2673 | = inp_icblk_stride * jcp.typesize_in * ic_block_step; |
| 2674 | safe_add(reg_input, input_offset, reg_long_offt); |
| 2675 | add(reg_kernel, jcp.typesize_out * ic_block_step * oc_block); |
| 2676 | sub(b_ic, ic_block_step); |
| 2677 | if (generate_icb_loop || ic_tail) sub(reg_icb, ic_block_step); |
| 2678 | cmp(b_ic, ic_block_step); |
| 2679 | jge(ic_block_label, T_NEAR); |
| 2680 | } |
| 2681 | L(icb_block_label_end); |
| 2682 | |
| 2683 | const int input_shift = jcp.typesize_in * (jcp.dilate_h + 1) |
| 2684 | * inp_icb_sp_stride * inp_mul; |
| 2685 | |
| 2686 | if (generate_icb_loop || ic_tail) { |
| 2687 | const size_t kernel_icb_shift = (size_t)jcp.typesize_out * jcp.kd |
| 2688 | * jcp.kh * jcp.kw * ic_block * oc_block; |
| 2689 | if (generate_icb_loop) { |
| 2690 | // icb loop supported for src in nxc layout only |
| 2691 | assert(src_layout_nxc); |
| 2692 | Label icb_loop_done; |
| 2693 | safe_add(reg_kernel, |
| 2694 | kernel_icb_shift |
| 2695 | - jcp.typesize_out * ic_block * oc_block, |
| 2696 | reg_long_offt); |
| 2697 | cmp(reg_icb, ic_block); |
| 2698 | jge(icb_block_label, T_NEAR); |
| 2699 | L(icb_loop_done); |
| 2700 | } |
| 2701 | |
| 2702 | L(ic_tail_loop_label); |
| 2703 | if (ic_tail) { |
| 2704 | Label skip_ic_tail; |
| 2705 | const int ic_tail_loop_work = rnd_dn(ic_tail, ic_block_step); |
| 2706 | cmp(reg_icb, 0); |
| 2707 | jle(skip_ic_tail, T_NEAR); |
| 2708 | mov(b_ic, reg_icb); |
| 2709 | if (ic_tail_loop_work) { |
| 2710 | cmp(reg_icb, ic_block_step); |
| 2711 | jge(ic_block_label, T_NEAR); |
| 2712 | if (generate_icb_loop) { |
| 2713 | // compensate offset added in generate_icb_loop |
| 2714 | safe_sub(reg_kernel, |
| 2715 | kernel_icb_shift |
| 2716 | - jcp.typesize_out * ic_block |
| 2717 | * oc_block, |
| 2718 | reg_long_offt); |
| 2719 | } |
| 2720 | } |
| 2721 | |
| 2722 | L(ic_tail_label); |
| 2723 | if (ic_tail % ic_block_step) { |
| 2724 | cmp(reg_icb, 0); |
| 2725 | jle(skip_ic_tail, T_NEAR); |
| 2726 | compute_ic_block_step( |
| 2727 | ow, l_pad, r_pad, ic_tail % ic_block_step, 0, 0, 0); |
| 2728 | } |
| 2729 | L(skip_ic_tail); |
| 2730 | } |
| 2731 | |
| 2732 | pop(reg_kernel); |
| 2733 | pop(reg_input); |
| 2734 | |
| 2735 | add(reg_input, input_shift); |
| 2736 | add(reg_kernel, jcp.typesize_out * jcp.kw * ic_block * oc_block); |
| 2737 | |
| 2738 | } else if (jcp.is_1stconv && !src_layout_nxc) { |
| 2739 | size_t input_offset = (size_t)jcp.typesize_in * jcp.id * jcp.ih |
| 2740 | * jcp.iw * ic_block; |
| 2741 | safe_sub(reg_input, input_offset, reg_long_offt); |
| 2742 | add(reg_input, input_shift); |
| 2743 | } else { |
| 2744 | add(reg_input, input_shift - jcp.typesize_in * jcp.ic_block); |
| 2745 | } |
| 2746 | |
| 2747 | if (!jcp.is_hw_transp && !(generate_icb_loop || ic_tail)) |
| 2748 | add(reg_kernel, |
| 2749 | jcp.typesize_out * (jcp.kw - 1) * ic_block * oc_block); |
| 2750 | dec(kj); |
| 2751 | cmp(kj, 0); |
| 2752 | jg(kh_label, T_NEAR); |
| 2753 | } |
| 2754 | if (jcp.ndims == 5) { |
| 2755 | add(aux_reg_input, |
| 2756 | jcp.typesize_in * (jcp.dilate_d + 1) * jcp.ih * jcp.iw |
| 2757 | * inp_mul); |
| 2758 | add(aux_reg_kernel, |
| 2759 | jcp.typesize_out * jcp.kh * jcp.kw * ic_block * oc_block); |
| 2760 | dec(ki); |
| 2761 | cmp(ki, 0); |
| 2762 | jg(kd_label, T_NEAR); |
| 2763 | } |
| 2764 | } |
| 2765 | |
| 2766 | void jit_avx512_common_conv_bwd_weights_kernel_f32 ::compute_oh_step_common( |
| 2767 | int ic_block_step, int max_ur_w) { |
| 2768 | using namespace nstl; |
| 2769 | Label kh_label, ic_block_label, ic_tail_loop_label, ic_tail_label, kd_label; |
| 2770 | |
| 2771 | const bool src_layout_nxc = is_src_layout_nxc(); |
| 2772 | int ic_block = jcp.ic_block; |
| 2773 | int oc_block = jcp.oc_block; |
| 2774 | |
| 2775 | int ow = jcp.is_hw_transp ? jcp.oh : jcp.ow; |
| 2776 | int r_pad = max(0, jcp.r_pad); |
| 2777 | int l_pad = jcp.l_pad; |
| 2778 | |
| 2779 | int ur_w = min(ow, max_ur_w); |
| 2780 | int ur_w_trips = ow / ur_w; |
| 2781 | int ur_w_tail = ow % ur_w; |
| 2782 | if ((ur_w_tail == 0 && r_pad != 0) || (r_pad > 0 && r_pad >= ur_w_tail)) { |
| 2783 | if (ur_w_trips > 1) { |
| 2784 | ur_w_tail += ur_w; |
| 2785 | ur_w_trips--; |
| 2786 | } else { |
| 2787 | ur_w_tail += (ur_w - ur_w / 2); |
| 2788 | ur_w = ur_w / 2; |
| 2789 | } |
| 2790 | } |
| 2791 | |
| 2792 | assert(l_pad <= max_ur_w); |
| 2793 | int inp_mult = src_layout_nxc |
| 2794 | ? jcp.ngroups * jcp.ic |
| 2795 | : (jcp.is_1stconv ? 1 : ic_block * (jcp.is_hw_transp ? jcp.iw : 1)); |
| 2796 | int out_mult = is_ddst_layout_nxc() ? jcp.ngroups * jcp.oc : oc_block; |
| 2797 | int input_comeback |
| 2798 | = max((ur_w_trips * ur_w * jcp.stride_w - l_pad), 0) * inp_mult; |
| 2799 | int output_comeback = ur_w_trips * ur_w * out_mult; |
| 2800 | const int ic_tail = jcp.ic_tail; |
| 2801 | const bool generate_icb_loop = jcp.nb_ic_blocking_max > 1; |
| 2802 | |
| 2803 | auto ic_loop = [=](int ic_block_step) { |
| 2804 | Label ow_block_label, ic_block_inner_label; |
| 2805 | int ur_w_blocks = ur_w_trips; |
| 2806 | |
| 2807 | int l_pad_tail = max(l_pad - ur_w, 0); |
| 2808 | L(ic_block_inner_label); |
| 2809 | if (l_pad != 0) { |
| 2810 | ur_w_blocks--; |
| 2811 | compute_ic_block_step(ur_w, l_pad, 0, ic_block_step, 0, 0, 0); |
| 2812 | int iw_offset = ur_w * jcp.stride_w - l_pad; |
| 2813 | if (iw_offset > 0) |
| 2814 | add(reg_input, jcp.typesize_in * iw_offset * inp_mult); |
| 2815 | add(reg_output, jcp.typesize_in * ur_w * out_mult); |
| 2816 | } |
| 2817 | |
| 2818 | assert(IMPLICATION(l_pad_tail > 0, ur_w_blocks <= 1)); |
| 2819 | if (ur_w_blocks > 0) { |
| 2820 | xor_(reg_ur_w_trips, reg_ur_w_trips); |
| 2821 | L(ow_block_label); |
| 2822 | { |
| 2823 | compute_ic_block_step( |
| 2824 | ur_w, l_pad_tail, 0, ic_block_step, 0, 0, 0); |
| 2825 | add(reg_input, |
| 2826 | jcp.typesize_in * (ur_w * jcp.stride_w - l_pad_tail) |
| 2827 | * inp_mult); |
| 2828 | add(reg_output, jcp.typesize_in * ur_w * out_mult); |
| 2829 | |
| 2830 | inc(reg_ur_w_trips); |
| 2831 | cmp(reg_ur_w_trips, ur_w_blocks); |
| 2832 | jl(ow_block_label, T_NEAR); |
| 2833 | l_pad_tail = max(l_pad_tail - ur_w, 0); |
| 2834 | } |
| 2835 | } |
| 2836 | |
| 2837 | if (ur_w_tail > 0) |
| 2838 | compute_ic_block_step( |
| 2839 | ur_w_tail, l_pad_tail, r_pad, ic_block_step, 0, 0, 0); |
| 2840 | |
| 2841 | sub(reg_output, jcp.typesize_in * output_comeback); |
| 2842 | }; |
| 2843 | |
| 2844 | if (jcp.ndims == 5) { |
| 2845 | L(kd_label); |
| 2846 | mov(reg_input, aux_reg_input); |
| 2847 | mov(reg_kernel, aux_reg_kernel); |
| 2848 | } |
| 2849 | |
| 2850 | mov(kj, reg_kh); |
| 2851 | L(kh_label); |
| 2852 | { |
| 2853 | Label icb_block_label, icb_block_label_cb; |
| 2854 | if (generate_icb_loop || ic_tail) { |
| 2855 | // TODO: May be broadcast work? |
| 2856 | push(reg_input); |
| 2857 | push(reg_kernel); |
| 2858 | mov(reg_icb, ptr[param + GET_OFF(reduce_work)]); |
| 2859 | } |
| 2860 | |
| 2861 | if (ic_tail) { |
| 2862 | cmp(reg_icb, ic_block); |
| 2863 | jl(ic_tail_loop_label, T_NEAR); |
| 2864 | } |
| 2865 | |
| 2866 | L(icb_block_label); |
| 2867 | mov(b_ic, ic_block); |
| 2868 | L(ic_block_label); |
| 2869 | Label ic_block_label_end; |
| 2870 | { |
| 2871 | ic_loop(ic_block_step); |
| 2872 | sub(reg_input, jcp.typesize_in * input_comeback); |
| 2873 | int inp_icblk_stride = jcp.is_1stconv && !src_layout_nxc |
| 2874 | ? jcp.ih * jcp.iw * jcp.id |
| 2875 | : 1; |
| 2876 | size_t input_offset |
| 2877 | = inp_icblk_stride * jcp.typesize_in * ic_block_step; |
| 2878 | safe_add(reg_input, input_offset, reg_long_offt); |
| 2879 | add(reg_kernel, jcp.typesize_out * ic_block_step * oc_block); |
| 2880 | sub(b_ic, ic_block_step); |
| 2881 | if (generate_icb_loop || ic_tail) sub(reg_icb, ic_block_step); |
| 2882 | cmp(b_ic, ic_block_step); |
| 2883 | jge(ic_block_label, T_NEAR); |
| 2884 | } |
| 2885 | L(ic_block_label_end); |
| 2886 | |
| 2887 | const int input_shift |
| 2888 | = jcp.typesize_in * (jcp.dilate_h + 1) * jcp.iw * inp_mult; |
| 2889 | |
| 2890 | if (generate_icb_loop || ic_tail) { |
| 2891 | const size_t kernel_icb_loop_shift_bytes = (size_t)jcp.typesize_out |
| 2892 | * jcp.kd * jcp.kh * jcp.kw * ic_block * oc_block; |
| 2893 | |
| 2894 | if (generate_icb_loop) { |
| 2895 | // icb loop supported for src in nxc layout only |
| 2896 | assert(src_layout_nxc); |
| 2897 | safe_add(reg_kernel, |
| 2898 | kernel_icb_loop_shift_bytes |
| 2899 | - jcp.typesize_out * ic_block * oc_block, |
| 2900 | reg_long_offt); |
| 2901 | |
| 2902 | cmp(reg_icb, ic_block); |
| 2903 | jge(icb_block_label, T_NEAR); |
| 2904 | } |
| 2905 | |
| 2906 | L(ic_tail_loop_label); |
| 2907 | if (ic_tail) { |
| 2908 | Label skip_ic_tail; |
| 2909 | const int ic_tail_loop_work = rnd_dn(ic_tail, ic_block_step); |
| 2910 | cmp(reg_icb, 0); |
| 2911 | jle(skip_ic_tail, T_NEAR); |
| 2912 | mov(b_ic, reg_icb); |
| 2913 | if (ic_tail_loop_work) { |
| 2914 | cmp(reg_icb, ic_block_step); |
| 2915 | jge(ic_block_label, T_NEAR); |
| 2916 | if (generate_icb_loop) { |
| 2917 | // compensate offset added in generate_icb_loop |
| 2918 | safe_sub(reg_kernel, |
| 2919 | kernel_icb_loop_shift_bytes |
| 2920 | - jcp.typesize_out * ic_block |
| 2921 | * oc_block, |
| 2922 | reg_long_offt); |
| 2923 | } |
| 2924 | } |
| 2925 | |
| 2926 | L(ic_tail_label); |
| 2927 | if (ic_tail % ic_block_step) { |
| 2928 | cmp(reg_icb, 0); |
| 2929 | jle(skip_ic_tail, T_NEAR); |
| 2930 | ic_loop(ic_tail % ic_block_step); |
| 2931 | } |
| 2932 | L(skip_ic_tail); |
| 2933 | } |
| 2934 | |
| 2935 | pop(reg_kernel); |
| 2936 | pop(reg_input); |
| 2937 | |
| 2938 | add(reg_input, input_shift); |
| 2939 | add(reg_kernel, jcp.typesize_out * jcp.kw * ic_block * oc_block); |
| 2940 | } else if (jcp.is_1stconv && !src_layout_nxc) { |
| 2941 | size_t input_offset = (size_t)jcp.typesize_in * jcp.id * jcp.ih |
| 2942 | * jcp.iw * ic_block; |
| 2943 | safe_sub(reg_input, input_offset, reg_long_offt); |
| 2944 | add(reg_input, input_shift); |
| 2945 | } else if (!jcp.is_hw_transp) { |
| 2946 | add(reg_input, input_shift - jcp.typesize_in * ic_block); |
| 2947 | } |
| 2948 | if (!jcp.is_hw_transp && !(generate_icb_loop || ic_tail)) |
| 2949 | add(reg_kernel, |
| 2950 | jcp.typesize_out * (jcp.kw - 1) * ic_block * oc_block); |
| 2951 | dec(kj); |
| 2952 | cmp(kj, 0); |
| 2953 | jg(kh_label, T_NEAR); |
| 2954 | } |
| 2955 | if (jcp.ndims == 5) { |
| 2956 | add(aux_reg_input, |
| 2957 | jcp.typesize_in * (jcp.dilate_d + 1) * jcp.ih * jcp.iw |
| 2958 | * inp_mult); |
| 2959 | add(aux_reg_kernel, |
| 2960 | jcp.typesize_out * jcp.kh * jcp.kw * ic_block * oc_block); |
| 2961 | dec(ki); |
| 2962 | cmp(ki, 0); |
| 2963 | jg(kd_label, T_NEAR); |
| 2964 | } |
| 2965 | } |
| 2966 | |
| 2967 | void jit_avx512_common_conv_bwd_weights_kernel_f32 ::compute_oh_step_disp() { |
| 2968 | int ic_block_step; |
| 2969 | if (jcp.kernel_kind == expl_bcast) |
| 2970 | ic_block_step = jcp.kw <= 3 ? 4 : (jcp.kw <= 7 ? 2 : 1); |
| 2971 | else |
| 2972 | ic_block_step = jcp.kw <= 3 ? 8 : (jcp.kw <= 7 ? 4 : 2); |
| 2973 | |
| 2974 | if (jcp.is_1stconv) { |
| 2975 | bool large_code = jcp.kw >= 7 && (jcp.l_pad > 0 || jcp.t_pad > 0); |
| 2976 | ic_block_step = (jcp.kw * jcp.ic_block <= 28 && !large_code) |
| 2977 | ? jcp.ic_block |
| 2978 | : 1; |
| 2979 | } |
| 2980 | |
| 2981 | bool too_large_to_unroll = (jcp.kw > 1 || jcp.kh > 1 || jcp.kd > 1) |
| 2982 | && (jcp.stride_w > 1 || jcp.stride_h > 1 || jcp.stride_d > 1); |
| 2983 | |
| 2984 | int ow = jcp.is_hw_transp ? jcp.oh : jcp.ow; |
| 2985 | if (jcp.ndims == 5) { |
| 2986 | /* NOTE: reg_kd_count = aux_reg_input = r12. The following order of |
| 2987 | * 'movs' must be guaranteed. */ |
| 2988 | mov(ki, reg_kd_count); |
| 2989 | push(reg_kd_count); |
| 2990 | mov(aux_reg_input, reg_input); |
| 2991 | mov(aux_reg_kernel, reg_kernel); |
| 2992 | } |
| 2993 | |
| 2994 | if (jcp.kw <= 3 && ow <= 16 && !too_large_to_unroll) |
| 2995 | compute_oh_step_unroll_ow_icblock(ic_block_step, max_ur_w); |
| 2996 | else if (ow <= max_ur_w) |
| 2997 | compute_oh_step_unroll_ow(ic_block_step, max_ur_w); |
| 2998 | else |
| 2999 | compute_oh_step_common(ic_block_step, max_ur_w); |
| 3000 | |
| 3001 | if (jcp.ndims == 5) { |
| 3002 | mov(reg_input, aux_reg_input); |
| 3003 | mov(reg_kernel, aux_reg_kernel); |
| 3004 | pop(reg_kd_count); |
| 3005 | od_step_comeback_pointers(); |
| 3006 | } else { |
| 3007 | oh_step_comeback_pointers(); |
| 3008 | } |
| 3009 | } |
| 3010 | |
| 3011 | void jit_avx512_common_conv_bwd_weights_kernel_f32::maybe_zero_kernel() { |
| 3012 | Label skip_zeroing, zeroing_loop; |
| 3013 | |
| 3014 | mov(reg_tmp, ptr[param + GET_OFF(channel)]); |
| 3015 | cmp(reg_tmp, 0); |
| 3016 | jz(skip_zeroing, T_NEAR); |
| 3017 | |
| 3018 | Zmm zero = Zmm(0); |
| 3019 | vpxord(zero, zero, zero); |
| 3020 | const bool generate_icb_loop = jcp.nb_ic_blocking_max > 1; |
| 3021 | const size_t kernel_block_bytes = (size_t)jcp.ic_block * jcp.oc_block |
| 3022 | * jcp.kw * jcp.kh * jcp.kd * jcp.typesize_out; |
| 3023 | Label icb_block_label, icb_block_label_cb; |
| 3024 | if (generate_icb_loop) { |
| 3025 | push(reg_kernel); |
| 3026 | |
| 3027 | mov(reg_icb, ptr[param + GET_OFF(reduce_work)]); |
| 3028 | L(icb_block_label); |
| 3029 | } |
| 3030 | |
| 3031 | xor_(reg_tmp, reg_tmp); |
| 3032 | L(zeroing_loop); |
| 3033 | { |
| 3034 | assert(jcp.oc_block * jcp.typesize_out |
| 3035 | == cpu_isa_traits<avx512_core>::vlen); |
| 3036 | for (int ic1 = 0; ic1 < jcp.ic_block; ic1++) |
| 3037 | vmovups(ptr[reg_kernel + reg_tmp |
| 3038 | + ic1 * jcp.oc_block * jcp.typesize_out], |
| 3039 | zero); |
| 3040 | add(reg_tmp, jcp.ic_block * jcp.oc_block * jcp.typesize_out); |
| 3041 | cmp(reg_tmp, kernel_block_bytes); |
| 3042 | jnz(zeroing_loop); |
| 3043 | } |
| 3044 | if (generate_icb_loop) { |
| 3045 | add(reg_kernel, kernel_block_bytes); |
| 3046 | sub(reg_icb, jcp.ic_block); |
| 3047 | cmp(reg_icb, 0); |
| 3048 | jg(icb_block_label, T_NEAR); |
| 3049 | |
| 3050 | pop(reg_kernel); |
| 3051 | } |
| 3052 | |
| 3053 | L(skip_zeroing); |
| 3054 | } |
| 3055 | |
| 3056 | void jit_avx512_common_conv_bwd_weights_kernel_f32::bias_kernel_2d() { |
| 3057 | assert(jcp.ndims == 4); // only supports 2d |
| 3058 | Label skip_bias, bias_loop; |
| 3059 | const int oc_tail = jcp.oc_tail; |
| 3060 | |
| 3061 | mov(reg_tmp, ptr[param1 + GET_OFF(flags)]); |
| 3062 | mov(reg_bias, ptr[param + GET_OFF(bias)]); |
| 3063 | test(reg_tmp, reg_tmp); |
| 3064 | jnz(skip_bias, T_NEAR); |
| 3065 | |
| 3066 | vmovups(Zmm(0), ptr[reg_bias]); |
| 3067 | |
| 3068 | mov(reg_oi, jcp.ow); |
| 3069 | xor_(reg_tmp, reg_tmp); |
| 3070 | L(bias_loop); |
| 3071 | { |
| 3072 | auto zmm_out = Zmm(1); |
| 3073 | if (oc_tail) zmm_out = zmm_out | k_oc_mask | T_z; |
| 3074 | vmovups(zmm_out, ptr[reg_output + reg_tmp]); |
| 3075 | vaddps(Zmm(0), Zmm(0), Zmm(1)); |
| 3076 | const int oc_stride |
| 3077 | = is_ddst_layout_nxc() ? jcp.ngroups * jcp.oc : jcp.oc_block; |
| 3078 | add(reg_tmp, jcp.typesize_out * oc_stride); |
| 3079 | dec(reg_oi); |
| 3080 | jg(bias_loop); |
| 3081 | } |
| 3082 | vmovups(ptr[reg_bias], Zmm(0)); |
| 3083 | |
| 3084 | L(skip_bias); |
| 3085 | } |
| 3086 | |
| 3087 | void jit_avx512_common_conv_bwd_weights_kernel_f32::bias_kernel_3d() { |
| 3088 | assert(jcp.ndims == 5); // only supports 3d |
| 3089 | Label skip_bias, bias_loop, skip_load_bias; |
| 3090 | const bool oc_tail = jcp.oc_tail; |
| 3091 | |
| 3092 | mov(reg_tmp, ptr[param + GET_OFF(flags)]); |
| 3093 | test(reg_tmp, reg_tmp); |
| 3094 | jne(skip_bias, T_NEAR); |
| 3095 | |
| 3096 | mov(reg_bias, ptr[param + GET_OFF(bias)]); |
| 3097 | mov(reg_output, ptr[param + GET_OFF(dst)]); |
| 3098 | vpxord(Zmm(1), Zmm(1), Zmm(1)); |
| 3099 | |
| 3100 | mov(reg_tmp, ptr[param + GET_OFF(channel)]); |
| 3101 | cmp(reg_tmp, 0); |
| 3102 | jne(skip_load_bias, T_NEAR); |
| 3103 | vmovups(Zmm(1), ptr[reg_bias]); |
| 3104 | |
| 3105 | L(skip_load_bias); |
| 3106 | |
| 3107 | mov(reg_oi, ptr[param + GET_OFF(os_index_end)]); |
| 3108 | sub(reg_oi, ptr[param + GET_OFF(os_index_begin)]); |
| 3109 | cmp(reg_oi, 0); |
| 3110 | jle(skip_bias, T_NEAR); // no iterations along depth dimension |
| 3111 | |
| 3112 | const size_t oc_mult |
| 3113 | = is_ddst_layout_nxc() ? jcp.ngroups * jcp.oc : jcp.oc_block; |
| 3114 | mov(reg_tmp, oc_mult * jcp.ow * jcp.oh * jcp.typesize_out); |
| 3115 | imul(reg_oi, reg_tmp); |
| 3116 | |
| 3117 | xor_(reg_tmp, reg_tmp); |
| 3118 | L(bias_loop); |
| 3119 | { |
| 3120 | auto zmm_out = Zmm(0); |
| 3121 | if (oc_tail) zmm_out = zmm_out | k_oc_mask | T_z; |
| 3122 | vmovups(zmm_out, ptr[reg_output + reg_tmp]); |
| 3123 | vaddps(Zmm(1), Zmm(1), Zmm(0)); |
| 3124 | add(reg_tmp, oc_mult * jcp.typesize_out); |
| 3125 | cmp(reg_tmp, reg_oi); |
| 3126 | jl(bias_loop); |
| 3127 | } |
| 3128 | vmovups(ptr[reg_bias], Zmm(1)); |
| 3129 | |
| 3130 | L(skip_bias); |
| 3131 | } |
| 3132 | |
| 3133 | void jit_avx512_common_conv_bwd_weights_kernel_f32 ::compute_oh_loop_common() { |
| 3134 | assert(one_of(jcp.harness, harness_mb_reduction, harness_3d_reduction)); |
| 3135 | int b_pad = jcp.b_pad; |
| 3136 | int t_pad = jcp.t_pad; |
| 3137 | bool is_dilated = jcp.dilate_h != 0; |
| 3138 | int dilate_h = jcp.dilate_h + 1; |
| 3139 | int stride_h = jcp.stride_h; |
| 3140 | const int inp_mult = is_src_layout_nxc() |
| 3141 | ? jcp.ngroups * jcp.ic |
| 3142 | : (jcp.is_1stconv ? 1 : jcp.ic_block); |
| 3143 | const int out_mult |
| 3144 | = is_ddst_layout_nxc() ? jcp.ngroups * jcp.oc : jcp.oc_block; |
| 3145 | int iw = jcp.is_hw_transp ? 1 : jcp.iw; |
| 3146 | Label oh_label, oh_label_end, oh_tpad_label, oh_tpad_tail_label, |
| 3147 | oh_bpad_label, oh_bpad_label_end, oh_dilate_label_shift, |
| 3148 | oh_dilate_label_noshift, oh_dilate_label_end; |
| 3149 | |
| 3150 | int ow = jcp.is_hw_transp ? jcp.oh : jcp.ow; |
| 3151 | int oh = jcp.is_hw_transp ? jcp.ow : jcp.oh; |
| 3152 | int kw = jcp.is_hw_transp ? jcp.tr_kw : jcp.kw; |
| 3153 | int kh = jcp.is_hw_transp ? jcp.tr_kh : jcp.kh; |
| 3154 | int ih = jcp.is_hw_transp ? jcp.tr_ih : jcp.ih; |
| 3155 | int ihp = jcp.is_hw_transp ? jcp.tr_ih : jcp.ihp; |
| 3156 | |
| 3157 | assert(IMPLICATION(jcp.is_hw_transp, |
| 3158 | everyone_is(1, oh, stride_h, dilate_h) |
| 3159 | && everyone_is(0, b_pad, t_pad))); |
| 3160 | |
| 3161 | mov(reg_kh, kh); |
| 3162 | xor_(reg_oj, reg_oj); |
| 3163 | /* Compute 'top' edge */ |
| 3164 | if (t_pad > 0) { |
| 3165 | const int kh_range = 1 + (kh - 1) * dilate_h; |
| 3166 | const int overflow = nstl::max(0, kh - div_up(t_pad + ih, dilate_h)); |
| 3167 | const int underflow = div_up(t_pad, dilate_h); |
| 3168 | const int initial_inp_ker_overlap = kh - overflow - underflow; |
| 3169 | mov(reg_kh, initial_inp_ker_overlap); |
| 3170 | add(reg_kernel, |
| 3171 | jcp.typesize_out * underflow * kw * jcp.ic_block |
| 3172 | * jcp.oc_block); |
| 3173 | // generate loop to process kernel while it remains within t_pad + ih |
| 3174 | if (kh_range < t_pad + ih) { |
| 3175 | if (is_dilated) { |
| 3176 | const int tail = t_pad % dilate_h; |
| 3177 | const int shift = tail == 0 ? 0 : dilate_h - tail; |
| 3178 | mov(reg_tmp, shift); |
| 3179 | if (tail != 0) |
| 3180 | add(reg_input, jcp.typesize_in * shift * iw * inp_mult); |
| 3181 | } |
| 3182 | L(oh_tpad_label); |
| 3183 | { |
| 3184 | cmp(reg_oj, oh); |
| 3185 | jge(oh_label_end, T_NEAR); |
| 3186 | |
| 3187 | compute_oh_step_disp(); |
| 3188 | add(reg_output, jcp.typesize_in * ow * out_mult); |
| 3189 | if (is_dilated) { |
| 3190 | inc(reg_tmp); |
| 3191 | cmp(reg_tmp, dilate_h); |
| 3192 | jl(oh_dilate_label_shift, T_NEAR); |
| 3193 | // unshift input as new kernel element enters |
| 3194 | sub(reg_input, |
| 3195 | jcp.typesize_in * (dilate_h - 1) * iw * inp_mult); |
| 3196 | xor_(reg_tmp, reg_tmp); |
| 3197 | } |
| 3198 | // kernel overlap only changes when (t_pad + oj) % dilate_h == 0 |
| 3199 | sub(reg_kernel, |
| 3200 | jcp.typesize_out * stride_h * kw * jcp.ic_block |
| 3201 | * jcp.oc_block); |
| 3202 | add(reg_kh, stride_h); |
| 3203 | if (is_dilated) { |
| 3204 | jmp(oh_dilate_label_noshift, T_NEAR); |
| 3205 | L(oh_dilate_label_shift); |
| 3206 | // shift input as old kernel element progresses |
| 3207 | add(reg_input, jcp.typesize_in * stride_h * iw * inp_mult); |
| 3208 | L(oh_dilate_label_noshift); |
| 3209 | } |
| 3210 | inc(reg_oj); |
| 3211 | |
| 3212 | // final number of kernel elements that overlap with input |
| 3213 | const int final_inp_ker_overlap |
| 3214 | = nstl::min(kh, div_up(ih, dilate_h)); |
| 3215 | cmp(reg_kh, final_inp_ker_overlap); |
| 3216 | jl(oh_tpad_label, T_NEAR); |
| 3217 | } |
| 3218 | } |
| 3219 | // need second loop to process kernel if it is larger than the input |
| 3220 | // (does not apply to dilations as they must have unit stride) |
| 3221 | if (kh_range |
| 3222 | >= ih + (t_pad % stride_h == 0 ? stride_h : t_pad % stride_h)) { |
| 3223 | assert(!is_dilated); |
| 3224 | mov(reg_kh, ih); |
| 3225 | L(oh_tpad_tail_label); |
| 3226 | { |
| 3227 | cmp(reg_oj, oh); |
| 3228 | jge(oh_label_end, T_NEAR); |
| 3229 | |
| 3230 | compute_oh_step_disp(); |
| 3231 | add(reg_output, jcp.typesize_in * ow * out_mult); |
| 3232 | sub(reg_kernel, |
| 3233 | jcp.typesize_out * stride_h * kw * jcp.ic_block |
| 3234 | * jcp.oc_block); |
| 3235 | |
| 3236 | inc(reg_oj); |
| 3237 | cmp(reg_oj, nstl::min(utils::div_up(t_pad, stride_h), oh)); |
| 3238 | jl(oh_tpad_tail_label, T_NEAR); |
| 3239 | } |
| 3240 | } |
| 3241 | // correct any excess shifts to kernel and input |
| 3242 | // (does not apply to dilations as they must have unit stride, |
| 3243 | // kernel must fit inside input, and padding is smaller than input) |
| 3244 | if (t_pad <= oh * stride_h) { |
| 3245 | // kernel has moved beyond padding (adjust for stride effects) |
| 3246 | if (t_pad % stride_h != 0) { |
| 3247 | assert(!is_dilated); |
| 3248 | int inp_corr = stride_h - t_pad % stride_h; |
| 3249 | add(reg_kernel, |
| 3250 | jcp.typesize_out * inp_corr * kw * jcp.ic_block |
| 3251 | * jcp.oc_block); |
| 3252 | add(reg_input, jcp.typesize_in * inp_corr * iw * inp_mult); |
| 3253 | } |
| 3254 | } else { |
| 3255 | // kernel still overlaps padding (complete reset) |
| 3256 | assert(!is_dilated); |
| 3257 | sub(reg_kernel, |
| 3258 | jcp.typesize_out * (t_pad - oh * stride_h) * kw |
| 3259 | * jcp.ic_block * jcp.oc_block); |
| 3260 | } |
| 3261 | } |
| 3262 | |
| 3263 | const int oj_end_value = nstl::min( |
| 3264 | oh, utils::div_up(ihp - b_pad - (kh - 1) * dilate_h, stride_h)); |
| 3265 | cmp(reg_oj, oj_end_value); |
| 3266 | jge(oh_label_end, T_NEAR); |
| 3267 | |
| 3268 | /* Compute middle block(s) */ |
| 3269 | mov(reg_kh, kh); |
| 3270 | L(oh_label); |
| 3271 | { |
| 3272 | compute_oh_step_disp(); |
| 3273 | add(reg_input, jcp.typesize_in * stride_h * iw * inp_mult); |
| 3274 | add(reg_output, jcp.typesize_in * ow * out_mult); |
| 3275 | |
| 3276 | inc(reg_oj); |
| 3277 | cmp(reg_oj, oj_end_value); |
| 3278 | jl(oh_label, T_NEAR); |
| 3279 | } |
| 3280 | L(oh_label_end); |
| 3281 | |
| 3282 | /* Compute bottom edge */ |
| 3283 | if (b_pad > 0) { |
| 3284 | cmp(reg_oj, oh); |
| 3285 | jge(oh_bpad_label_end, T_NEAR); |
| 3286 | |
| 3287 | if (is_dilated) { |
| 3288 | mov(reg_kh, kh - 1); // assumes unit stride for dilations |
| 3289 | mov(reg_tmp, 0); |
| 3290 | } else { |
| 3291 | mov(reg_kh, ihp - b_pad); |
| 3292 | imul(reg_tmp, reg_oj, stride_h); |
| 3293 | sub(reg_kh, reg_tmp); |
| 3294 | } |
| 3295 | L(oh_bpad_label); |
| 3296 | { |
| 3297 | compute_oh_step_disp(); |
| 3298 | add(reg_input, jcp.typesize_in * stride_h * iw * inp_mult); |
| 3299 | add(reg_output, jcp.typesize_in * ow * out_mult); |
| 3300 | if (is_dilated) { |
| 3301 | inc(reg_tmp); |
| 3302 | cmp(reg_tmp, dilate_h); |
| 3303 | jl(oh_dilate_label_end, T_NEAR); |
| 3304 | xor_(reg_tmp, reg_tmp); |
| 3305 | } |
| 3306 | sub(reg_kh, stride_h); |
| 3307 | cmp(reg_kh, 0); |
| 3308 | jle(oh_bpad_label_end, T_NEAR); |
| 3309 | if (is_dilated) L(oh_dilate_label_end); |
| 3310 | |
| 3311 | inc(reg_oj); |
| 3312 | cmp(reg_oj, oh); |
| 3313 | jl(oh_bpad_label, T_NEAR); |
| 3314 | } |
| 3315 | L(oh_bpad_label_end); |
| 3316 | } |
| 3317 | } |
| 3318 | |
| 3319 | void jit_avx512_common_conv_bwd_weights_kernel_f32::compute_oh_loop_partial() { |
| 3320 | assert(jcp.harness == harness_2d_reduction); |
| 3321 | int ic_block = jcp.ic_block; |
| 3322 | int oc_block = jcp.oc_block; |
| 3323 | const int inp_mult = is_src_layout_nxc() |
| 3324 | ? jcp.ngroups * jcp.ic |
| 3325 | : (jcp.is_1stconv ? 1 : jcp.ic_block); |
| 3326 | const int out_mult |
| 3327 | = is_ddst_layout_nxc() ? jcp.ngroups * jcp.oc : jcp.oc_block; |
| 3328 | const int input_bottom_padding_overlap |
| 3329 | = div_up(jcp.ih + jcp.t_pad - (jcp.kh - 1), jcp.stride_h); |
| 3330 | const int bottom_pad_input_correction |
| 3331 | = jcp.ih + jcp.t_pad - input_bottom_padding_overlap * jcp.stride_h; |
| 3332 | |
| 3333 | const size_t filter_shift = jcp.typesize_out * jcp.kw * ic_block * oc_block; |
| 3334 | const size_t input_shift = jcp.typesize_in * jcp.iw * inp_mult; |
| 3335 | const size_t output_shift = jcp.typesize_out * jcp.ow * out_mult; |
| 3336 | |
| 3337 | Label loop_begin_label, loop_end_label, common_block_label, |
| 3338 | top_padding_end_label, bottom_padding_end_label, |
| 3339 | bottom_padding_label; |
| 3340 | |
| 3341 | if (jcp.with_bias) { |
| 3342 | Label skip_zero_bias; |
| 3343 | mov(reg_bias, ptr[param1 + GET_OFF(bias)]); |
| 3344 | mov(reg_tmp, ptr[param1 + GET_OFF(channel)]); |
| 3345 | test(reg_tmp, reg_tmp); |
| 3346 | jz(skip_zero_bias, T_NEAR); |
| 3347 | mov(reg_tmp, ptr[param1 + GET_OFF(flags)]); |
| 3348 | test(reg_tmp, reg_tmp); |
| 3349 | jnz(skip_zero_bias, T_NEAR); |
| 3350 | vpxord(Zmm(1), Zmm(1), Zmm(1)); |
| 3351 | vmovups(ptr[reg_bias], Zmm(1)); |
| 3352 | L(skip_zero_bias); |
| 3353 | } |
| 3354 | |
| 3355 | /* Offset filter position to adjust for top padding */ |
| 3356 | add(reg_kernel, ptr[param + GET_OFF(kh_offset)]); |
| 3357 | |
| 3358 | mov(reg_oj, ptr[param + GET_OFF(os_index_begin)]); |
| 3359 | mov(reg_kh, ptr[param + GET_OFF(kh_padding)]); |
| 3360 | |
| 3361 | cmp(reg_kh, 0); |
| 3362 | jle(loop_end_label, T_NEAR); // no iterations along kh |
| 3363 | cmp(reg_oj, ptr[param + GET_OFF(os_index_end)]); |
| 3364 | jge(loop_end_label, T_NEAR); // no iterations along height dimension |
| 3365 | |
| 3366 | L(loop_begin_label); |
| 3367 | |
| 3368 | if (jcp.with_bias) bias_kernel_2d(); |
| 3369 | compute_oh_step_disp(); |
| 3370 | |
| 3371 | /* Compute 'top' edge */ |
| 3372 | if (jcp.t_pad > 0) { |
| 3373 | |
| 3374 | /* Check if within top padding region */ |
| 3375 | cmp(reg_oj, div_up(jcp.t_pad, jcp.stride_h)); |
| 3376 | jge(top_padding_end_label, T_NEAR); |
| 3377 | |
| 3378 | /* Increment step counter and adjust filter position */ |
| 3379 | sub(reg_kernel, filter_shift * jcp.stride_h); |
| 3380 | add(reg_kh, jcp.stride_h); |
| 3381 | |
| 3382 | /* Final number of kernel elements that overlap with input */ |
| 3383 | const int inp_ker_overlap = nstl::min(jcp.kh, jcp.ih); |
| 3384 | cmp(reg_kh, inp_ker_overlap); |
| 3385 | jle(common_block_label, T_NEAR); |
| 3386 | |
| 3387 | /* Correct any excess shifts to kernel and input */ |
| 3388 | if (jcp.t_pad <= jcp.oh * jcp.stride_h) { |
| 3389 | /* Filter has moved beyond padding (adjust for stride effects) */ |
| 3390 | if (jcp.t_pad % jcp.stride_h != 0) { |
| 3391 | int inp_corr = jcp.stride_h - jcp.t_pad % jcp.stride_h; |
| 3392 | add(reg_kernel, filter_shift * inp_corr); |
| 3393 | add(reg_input, input_shift * inp_corr); |
| 3394 | } |
| 3395 | } else { |
| 3396 | /* Filter still overlaps padding (complete reset) */ |
| 3397 | sub(reg_kernel, (jcp.t_pad - jcp.oh * jcp.stride_h) * filter_shift); |
| 3398 | } |
| 3399 | |
| 3400 | /* Set filter element count for outside the t_pad region */ |
| 3401 | if (jcp.t_pad + jcp.ih < jcp.kh + jcp.stride_h) { |
| 3402 | // filter now overlaps with b_pad |
| 3403 | mov(reg_kh, bottom_pad_input_correction); |
| 3404 | } else { |
| 3405 | mov(reg_kh, inp_ker_overlap); |
| 3406 | } |
| 3407 | |
| 3408 | jmp(common_block_label); |
| 3409 | |
| 3410 | L(top_padding_end_label); |
| 3411 | } |
| 3412 | |
| 3413 | /* Compute 'bottom' edge */ |
| 3414 | if (jcp.b_pad > 0) { |
| 3415 | |
| 3416 | /* Check if within bottom padding region */ |
| 3417 | cmp(reg_oj, input_bottom_padding_overlap - 1); |
| 3418 | jl(bottom_padding_end_label, T_NEAR); |
| 3419 | jg(bottom_padding_label, T_NEAR); |
| 3420 | |
| 3421 | /* Execute overlap correction between the filter and the initial |
| 3422 | * bottom padding region. */ |
| 3423 | mov(reg_kh, bottom_pad_input_correction); |
| 3424 | jmp(bottom_padding_end_label, T_NEAR); |
| 3425 | |
| 3426 | L(bottom_padding_label); |
| 3427 | sub(reg_kh, jcp.stride_h); |
| 3428 | cmp(reg_kh, 0); |
| 3429 | jle(loop_end_label, T_NEAR); |
| 3430 | |
| 3431 | L(bottom_padding_end_label); |
| 3432 | } |
| 3433 | |
| 3434 | /* Compute middle block */ |
| 3435 | add(reg_input, input_shift * jcp.stride_h); |
| 3436 | |
| 3437 | /* Execute common block and loop */ |
| 3438 | L(common_block_label); |
| 3439 | add(reg_output, output_shift); |
| 3440 | inc(reg_oj); |
| 3441 | cmp(reg_oj, ptr[param + GET_OFF(os_index_end)]); |
| 3442 | jl(loop_begin_label, T_NEAR); |
| 3443 | |
| 3444 | L(loop_end_label); |
| 3445 | } |
| 3446 | |
| 3447 | void jit_avx512_common_conv_bwd_weights_kernel_f32::compute_od_loop_partial() { |
| 3448 | assert(jcp.harness == harness_3d_reduction); |
| 3449 | int ic_block = jcp.ic_block; |
| 3450 | int oc_block = jcp.oc_block; |
| 3451 | const int inp_mult = is_src_layout_nxc() |
| 3452 | ? jcp.ngroups * jcp.ic |
| 3453 | : (jcp.is_1stconv ? 1 : jcp.ic_block); |
| 3454 | const int out_mult |
| 3455 | = is_ddst_layout_nxc() ? jcp.ngroups * jcp.oc : jcp.oc_block; |
| 3456 | |
| 3457 | int iw = jcp.iw; |
| 3458 | int ow = jcp.ow; |
| 3459 | const int input_backpad_overlap |
| 3460 | = div_up(jcp.id + jcp.f_pad - (jcp.kd - 1), jcp.stride_d); |
| 3461 | const int back_pad_input_correction |
| 3462 | = jcp.id + jcp.f_pad - input_backpad_overlap * jcp.stride_d; |
| 3463 | |
| 3464 | const size_t filter_shift |
| 3465 | = jcp.typesize_out * jcp.kh * jcp.kw * ic_block * oc_block; |
| 3466 | const size_t input_shift = jcp.typesize_in * jcp.ih * iw * inp_mult; |
| 3467 | const size_t output_shift = jcp.typesize_in * jcp.oh * ow * out_mult; |
| 3468 | |
| 3469 | Label d_loop_label, loop_end_label, common_block_label, fpad_end_label, |
| 3470 | backpad_end_label, backpad_label; |
| 3471 | |
| 3472 | if (jcp.with_bias) bias_kernel_3d(); |
| 3473 | |
| 3474 | /* initially offset 'kd' by f_pad */ |
| 3475 | add(reg_kernel, ptr[param + GET_OFF(kd_offset)]); |
| 3476 | |
| 3477 | mov(reg_input_d, ptr[param + GET_OFF(src)]); |
| 3478 | mov(reg_output_d, ptr[param + GET_OFF(dst)]); |
| 3479 | mov(reg_d_index, ptr[param + GET_OFF(os_index_begin)]); |
| 3480 | mov(reg_kd_count, ptr[param + GET_OFF(kd_padding)]); |
| 3481 | |
| 3482 | cmp(reg_kd_count, 0); |
| 3483 | jle(loop_end_label, T_NEAR); // no iterations along kd |
| 3484 | cmp(reg_d_index, ptr[param + GET_OFF(os_index_end)]); |
| 3485 | jge(loop_end_label, T_NEAR); // no iterations along depth dimension |
| 3486 | |
| 3487 | L(d_loop_label); |
| 3488 | |
| 3489 | mov(reg_input, reg_input_d); |
| 3490 | mov(reg_output, reg_output_d); |
| 3491 | |
| 3492 | push(reg_input_d); |
| 3493 | push(reg_output_d); |
| 3494 | push(reg_d_index); |
| 3495 | |
| 3496 | compute_oh_loop_common(); |
| 3497 | |
| 3498 | pop(reg_d_index); |
| 3499 | pop(reg_output_d); |
| 3500 | pop(reg_input_d); |
| 3501 | |
| 3502 | /* Compute 'front' edge */ |
| 3503 | if (jcp.f_pad > 0) { |
| 3504 | |
| 3505 | /* Check if within fpad region */ |
| 3506 | cmp(reg_d_index, div_up(jcp.f_pad, jcp.stride_d)); |
| 3507 | jge(fpad_end_label, T_NEAR); |
| 3508 | |
| 3509 | /* Fpad steps */ |
| 3510 | sub(reg_kernel, filter_shift * jcp.stride_d); |
| 3511 | add(reg_kd_count, jcp.stride_d); |
| 3512 | |
| 3513 | /* Final number of kernel elements that overlap with input */ |
| 3514 | const int inp_ker_overlap = nstl::min(jcp.kd, jcp.id); |
| 3515 | cmp(reg_kd_count, inp_ker_overlap); |
| 3516 | jle(common_block_label, T_NEAR); |
| 3517 | |
| 3518 | /* Correct any excess shifts to kernel and input */ |
| 3519 | if (jcp.f_pad <= jcp.od * jcp.stride_d) { |
| 3520 | /* Filter has moved beyond padding (adjust for stride effects) */ |
| 3521 | if (jcp.f_pad % jcp.stride_d != 0) { |
| 3522 | int inp_corr = jcp.stride_d - jcp.f_pad % jcp.stride_d; |
| 3523 | add(reg_kernel, filter_shift * inp_corr); |
| 3524 | add(reg_input_d, input_shift * inp_corr); |
| 3525 | } |
| 3526 | } else { |
| 3527 | /* Filter still overlaps padding (complete reset) */ |
| 3528 | sub(reg_kernel, (jcp.f_pad - jcp.od * jcp.stride_d) * filter_shift); |
| 3529 | } |
| 3530 | |
| 3531 | /* Set filter element count for outside the f_pad region */ |
| 3532 | if (jcp.f_pad + jcp.id < jcp.kd + jcp.stride_d) { |
| 3533 | // filter now overlaps with back_pad |
| 3534 | mov(reg_kd_count, back_pad_input_correction); |
| 3535 | } else { |
| 3536 | mov(reg_kd_count, inp_ker_overlap); |
| 3537 | } |
| 3538 | |
| 3539 | jmp(common_block_label); |
| 3540 | |
| 3541 | L(fpad_end_label); |
| 3542 | } |
| 3543 | |
| 3544 | /* Compute bottom edge */ |
| 3545 | if (jcp.back_pad > 0) { |
| 3546 | |
| 3547 | /* Check if within back_pad region */ |
| 3548 | cmp(reg_d_index, input_backpad_overlap - 1); |
| 3549 | jl(backpad_end_label, T_NEAR); |
| 3550 | jg(backpad_label, T_NEAR); |
| 3551 | |
| 3552 | /* Execute overlap correction between the filter and the initial |
| 3553 | * back_pad region. */ |
| 3554 | mov(reg_kd_count, back_pad_input_correction); |
| 3555 | jmp(backpad_end_label, T_NEAR); |
| 3556 | |
| 3557 | L(backpad_label); |
| 3558 | sub(reg_kd_count, jcp.stride_d); |
| 3559 | cmp(reg_kd_count, 0); |
| 3560 | jle(loop_end_label, T_NEAR); |
| 3561 | |
| 3562 | L(backpad_end_label); |
| 3563 | } |
| 3564 | |
| 3565 | /* Compute middle block */ |
| 3566 | add(reg_input_d, input_shift * jcp.stride_d); |
| 3567 | |
| 3568 | /* Execute common block and loop */ |
| 3569 | L(common_block_label); |
| 3570 | add(reg_output_d, output_shift); |
| 3571 | inc(reg_d_index); |
| 3572 | cmp(reg_d_index, ptr[param + GET_OFF(os_index_end)]); |
| 3573 | jl(d_loop_label, T_NEAR); |
| 3574 | |
| 3575 | L(loop_end_label); |
| 3576 | } |
| 3577 | |
| 3578 | void jit_avx512_common_conv_bwd_weights_kernel_f32::compute_loop() { |
| 3579 | |
| 3580 | maybe_zero_kernel(); |
| 3581 | |
| 3582 | switch (jcp.harness) { |
| 3583 | case harness_2d_reduction: compute_oh_loop_partial(); break; |
| 3584 | case harness_3d_reduction: compute_od_loop_partial(); break; |
| 3585 | case harness_mb_reduction: compute_oh_loop_common(); break; |
| 3586 | case harness_nxc: break; |
| 3587 | default: assert(!"Invalid harness type"); |
| 3588 | } |
| 3589 | } |
| 3590 | |
| 3591 | void jit_avx512_common_conv_bwd_weights_kernel_f32::generate_microkernel() { |
| 3592 | |
| 3593 | reg64_t reg_dwei = abi_param1; |
| 3594 | reg64_t reg_src = abi_param2; |
| 3595 | reg64_t reg_ddst = abi_param3; |
| 3596 | reg64_t reg_iw_base = abi_param4; |
| 3597 | reg64_t aux_reg_icb = r10; |
| 3598 | reg64_t aux_reg_kwb = r11; |
| 3599 | reg64_t reg_src_save = r12; |
| 3600 | reg64_t reg_dwei_save = r13; |
| 3601 | reg64_t reg_iw_base_save = r14; |
| 3602 | reg64_t reg_tmp = r15; |
| 3603 | |
| 3604 | //Currently kernel is small so passing parameters via registers is preferred |
| 3605 | //whenever possible |
| 3606 | #ifdef _WIN32 |
| 3607 | // Must be a scratch register since load is before preamble |
| 3608 | reg64_t reg_owb = rax; |
| 3609 | mov(reg_owb, ptr[get_stack_params_address(false)]); |
| 3610 | #else |
| 3611 | reg64_t reg_owb = abi_param5; |
| 3612 | #endif |
| 3613 | |
| 3614 | preamble(); |
| 3615 | |
| 3616 | const int kw_unroll = jcp.ur_kw; |
| 3617 | const int ow_unroll = jcp.ur_ow; |
| 3618 | const int iw_unroll = ow_unroll + kw_unroll - 1; |
| 3619 | const int ic_unroll = jcp.ur_ic; |
| 3620 | |
| 3621 | const int ker_reg_count = ic_unroll; |
| 3622 | const int src_reg_count = iw_unroll * ic_unroll; |
| 3623 | const int ddst_reg_count = ow_unroll; |
| 3624 | |
| 3625 | MAYBE_UNUSED(ddst_reg_count); |
| 3626 | assert(ker_reg_count + src_reg_count + ddst_reg_count <= 32); |
| 3627 | |
| 3628 | auto dwei_offset = [&](int i_kw, int i_ic) { |
| 3629 | const int oc_block_size = sizeof(float); |
| 3630 | const int ic_block_size = jcp.oc_block * oc_block_size; |
| 3631 | const int kw_block_size = jcp.ic_block * ic_block_size; |
| 3632 | const int kh_block_size = jcp.kw * kw_block_size; |
| 3633 | const int kd_block_size = jcp.kh * kh_block_size; |
| 3634 | const int icb_block_size = jcp.kd * kd_block_size; |
| 3635 | |
| 3636 | int icb = i_ic / jcp.ic_block; |
| 3637 | i_ic = i_ic % jcp.ic_block; |
| 3638 | |
| 3639 | return icb * icb_block_size + i_kw * kw_block_size |
| 3640 | + i_ic * ic_block_size; |
| 3641 | }; |
| 3642 | |
| 3643 | auto src_offset = [&](int i_ic, int i_iw) { |
| 3644 | const int ic_block_size = sizeof(float); |
| 3645 | const int g_block_size = jcp.ic * ic_block_size; |
| 3646 | const int iw_block_size = jcp.ngroups * g_block_size; |
| 3647 | |
| 3648 | return i_iw * iw_block_size + i_ic * ic_block_size; |
| 3649 | }; |
| 3650 | |
| 3651 | auto ddst_offset = [&](int i_ow) { |
| 3652 | const int oc_block_size = sizeof(float); |
| 3653 | const int g_block_size = jcp.oc * oc_block_size; |
| 3654 | const int ow_block_size = jcp.ngroups * g_block_size; |
| 3655 | |
| 3656 | return i_ow * ow_block_size; |
| 3657 | }; |
| 3658 | |
| 3659 | auto get_src_zmm = [=](int iw_index, int i_ic) { |
| 3660 | int zmm_index = iw_index * ic_unroll + i_ic + ker_reg_count; |
| 3661 | return Zmm(zmm_index); |
| 3662 | }; |
| 3663 | |
| 3664 | auto get_ddst_zmm = [=](int i_ow) { |
| 3665 | int zmm_index = i_ow + src_reg_count + ker_reg_count; |
| 3666 | return Zmm(zmm_index); |
| 3667 | }; |
| 3668 | |
| 3669 | auto get_ker_zmm = [=](int i_ic) { return Zmm(i_ic); }; |
| 3670 | |
| 3671 | auto load_ddsts = [=](int ur_ow) { |
| 3672 | for (int i_ow = 0; i_ow < ur_ow; i_ow++) { |
| 3673 | vmovups(get_ddst_zmm(i_ow), zword[reg_ddst + ddst_offset(i_ow)]); |
| 3674 | } |
| 3675 | }; |
| 3676 | |
| 3677 | auto load_srcs = [=](int ur_iw, int ur_ic, bool is_iw_edge) { |
| 3678 | Label iw_load_end; |
| 3679 | if (is_iw_edge) { |
| 3680 | for_(int i_iw_index = 0; i_iw_index < ur_iw; i_iw_index++) |
| 3681 | for (int i_ic = 0; i_ic < ur_ic; i_ic++) { |
| 3682 | vpxord(get_src_zmm(i_iw_index, i_ic), |
| 3683 | get_src_zmm(i_iw_index, i_ic), |
| 3684 | get_src_zmm(i_iw_index, i_ic)); |
| 3685 | } |
| 3686 | } |
| 3687 | |
| 3688 | for (int i_iw_index = 0; i_iw_index < ur_iw; i_iw_index++) { |
| 3689 | Label ic_load_end; |
| 3690 | if (is_iw_edge) { |
| 3691 | cmp(reg_iw_base, jcp.iw - i_iw_index * jcp.stride_w); |
| 3692 | jge(iw_load_end, T_NEAR); |
| 3693 | if (jcp.l_pad > 0) { |
| 3694 | cmp(reg_iw_base, -i_iw_index * jcp.stride_w); |
| 3695 | jl(ic_load_end, T_NEAR); |
| 3696 | } |
| 3697 | } |
| 3698 | for (int i_ic = 0; i_ic < ur_ic; i_ic++) { |
| 3699 | vbroadcastss(get_src_zmm(i_iw_index, i_ic), |
| 3700 | zword[reg_src |
| 3701 | + src_offset(i_ic, jcp.stride_w * i_iw_index)]); |
| 3702 | } |
| 3703 | L(ic_load_end); |
| 3704 | } |
| 3705 | L(iw_load_end); |
| 3706 | }; |
| 3707 | |
| 3708 | auto compute_kernel = [=](int ur_ow, int ur_ic, int ur_kw, int is_iw_edge) { |
| 3709 | Label kw_loop_end; |
| 3710 | load_srcs(ur_ow + ur_kw - 1, ur_ic, is_iw_edge); |
| 3711 | |
| 3712 | for (int i_kw = 0; i_kw < ur_kw; i_kw++) { |
| 3713 | for (int i_ic = 0; i_ic < ur_ic; i_ic++) { |
| 3714 | vpxord(get_ker_zmm(i_ic), get_ker_zmm(i_ic), get_ker_zmm(i_ic)); |
| 3715 | } |
| 3716 | for (int i_ow = 0; i_ow < ur_ow; i_ow++) { |
| 3717 | for (int i_ic = 0; i_ic < ur_ic; i_ic++) { |
| 3718 | vfmadd231ps(get_ker_zmm(i_ic), |
| 3719 | get_src_zmm(i_ow + i_kw, i_ic), get_ddst_zmm(i_ow)); |
| 3720 | } |
| 3721 | } |
| 3722 | for (int i_ic = 0; i_ic < ur_ic; i_ic++) { |
| 3723 | int ker_offset = dwei_offset(i_kw, i_ic); |
| 3724 | vaddps(get_ker_zmm(i_ic), zword[reg_dwei + ker_offset]); |
| 3725 | vmovups(zword[reg_dwei + ker_offset], get_ker_zmm(i_ic)); |
| 3726 | } |
| 3727 | } |
| 3728 | |
| 3729 | L(kw_loop_end); |
| 3730 | }; |
| 3731 | |
| 3732 | auto kw_loop = [=](int ur_ow, int ur_ic, int is_iw_edge) { |
| 3733 | Label kwb_loop_begin, kwb_loop_end; |
| 3734 | int kw_tail = jcp.kw % kw_unroll; |
| 3735 | int kw_iter = jcp.kw / kw_unroll; |
| 3736 | |
| 3737 | if (kw_iter > 0) { |
| 3738 | if (kw_iter > 1) { |
| 3739 | mov(aux_reg_kwb, jcp.kw - kw_tail); |
| 3740 | L(kwb_loop_begin); |
| 3741 | } |
| 3742 | compute_kernel(ur_ow, ur_ic, kw_unroll, is_iw_edge); |
| 3743 | |
| 3744 | if (kw_iter > 1 || kw_tail) { |
| 3745 | add(reg_iw_base, (jcp.dilate_w + 1) * kw_unroll); |
| 3746 | add(reg_src, src_offset(0, (jcp.dilate_w + 1) * kw_unroll)); |
| 3747 | add(reg_dwei, dwei_offset(kw_unroll, 0)); |
| 3748 | } |
| 3749 | |
| 3750 | if (kw_iter > 1) { |
| 3751 | sub(aux_reg_kwb, kw_unroll); |
| 3752 | jg(kwb_loop_begin, T_NEAR); |
| 3753 | } |
| 3754 | } |
| 3755 | |
| 3756 | if (kw_tail) compute_kernel(ur_ow, ur_ic, kw_tail, is_iw_edge); |
| 3757 | |
| 3758 | L(kwb_loop_end); |
| 3759 | }; |
| 3760 | |
| 3761 | auto ic_loop = [=](int ur_ow, int is_iw_edge) { |
| 3762 | Label icb_loop_begin, icb_loop_end; |
| 3763 | int ic_tail = jcp.ic % ic_unroll; |
| 3764 | int ic_iter = jcp.ic / ic_unroll; |
| 3765 | |
| 3766 | if (ic_iter > 0) { |
| 3767 | if (ic_iter > 1 || ic_tail) { |
| 3768 | mov(aux_reg_icb, jcp.ic - ic_tail); |
| 3769 | L(icb_loop_begin); |
| 3770 | // Saving onto the stack here appears to significantly slow down |
| 3771 | // code execution. If this kernel runs out of registers, getting |
| 3772 | // rid of the *_save registers should be possible by using |
| 3773 | // subtracts to restore the value and maintain performance. |
| 3774 | mov(reg_src_save, reg_src); |
| 3775 | mov(reg_dwei_save, reg_dwei); |
| 3776 | mov(reg_iw_base_save, reg_iw_base); |
| 3777 | } |
| 3778 | |
| 3779 | kw_loop(ur_ow, ic_unroll, is_iw_edge); |
| 3780 | |
| 3781 | if (ic_iter > 1 || ic_tail) { |
| 3782 | mov(reg_iw_base, reg_iw_base_save); |
| 3783 | mov(reg_dwei, reg_dwei_save); |
| 3784 | mov(reg_src, reg_src_save); |
| 3785 | |
| 3786 | Label inter_block_increment, increment_finish; |
| 3787 | sub(aux_reg_icb, ic_unroll); |
| 3788 | if (jcp.ic > jcp.ic_block) { |
| 3789 | const int log2_ic_block = 4; |
| 3790 | lea(reg_tmp, ptr[aux_reg_icb - jcp.ic - ic_tail]); |
| 3791 | test(reg_tmp, (1 << log2_ic_block) - 1); |
| 3792 | jnz(inter_block_increment, T_NEAR); |
| 3793 | |
| 3794 | add(reg_dwei, |
| 3795 | dwei_offset(0, jcp.ic_block) |
| 3796 | - dwei_offset(0, jcp.ic_block - ic_unroll)); |
| 3797 | jmp(increment_finish); |
| 3798 | L(inter_block_increment); |
| 3799 | } |
| 3800 | add(reg_dwei, dwei_offset(0, ic_unroll)); |
| 3801 | L(increment_finish); |
| 3802 | |
| 3803 | add(reg_src, src_offset(ic_unroll, 0)); |
| 3804 | } |
| 3805 | if (ic_iter > 1) { |
| 3806 | cmp(aux_reg_icb, 0); |
| 3807 | jg(icb_loop_begin, T_NEAR); |
| 3808 | } |
| 3809 | } |
| 3810 | |
| 3811 | if (ic_tail) kw_loop(ur_ow, ic_tail, is_iw_edge); |
| 3812 | |
| 3813 | L(icb_loop_end); |
| 3814 | }; |
| 3815 | |
| 3816 | auto ic_loop_dispatch = [=](int ur_ow) { |
| 3817 | Label iw_edge_case, ic_end; |
| 3818 | |
| 3819 | const int iw_overflow_bound = jcp.iw - (ur_ow - 1) * jcp.stride_w |
| 3820 | - (jcp.kw - 1) * (jcp.dilate_w + 1); |
| 3821 | cmp(reg_iw_base, iw_overflow_bound); |
| 3822 | jge(iw_edge_case, T_NEAR); |
| 3823 | if (jcp.l_pad > 0) { |
| 3824 | cmp(reg_iw_base, 0); |
| 3825 | jl(iw_edge_case, T_NEAR); |
| 3826 | } |
| 3827 | |
| 3828 | ic_loop(ur_ow, false); |
| 3829 | jmp(ic_end, T_NEAR); |
| 3830 | |
| 3831 | L(iw_edge_case); |
| 3832 | ic_loop(ur_ow, true); |
| 3833 | |
| 3834 | L(ic_end); |
| 3835 | }; |
| 3836 | |
| 3837 | Label ow_end, ow_tail; |
| 3838 | int ow_tail_size = jcp.ow % ow_unroll; |
| 3839 | cmp(reg_owb, jcp.ow - ow_tail_size); |
| 3840 | jge(ow_tail, T_NEAR); |
| 3841 | |
| 3842 | load_ddsts(ow_unroll); |
| 3843 | ic_loop_dispatch(ow_unroll); |
| 3844 | jmp(ow_end, T_NEAR); |
| 3845 | |
| 3846 | L(ow_tail); |
| 3847 | load_ddsts(ow_tail_size); |
| 3848 | ic_loop_dispatch(ow_tail_size); |
| 3849 | |
| 3850 | L(ow_end); |
| 3851 | |
| 3852 | postamble(); |
| 3853 | ret(); |
| 3854 | } |
| 3855 | |
| 3856 | void jit_avx512_common_conv_bwd_weights_kernel_f32::generate_kernel() { |
| 3857 | preamble(); |
| 3858 | |
| 3859 | mov(reg_input, ptr[param + GET_OFF(src)]); |
| 3860 | mov(reg_output, ptr[param + GET_OFF(dst)]); |
| 3861 | mov(reg_kernel, ptr[param + GET_OFF(filt)]); |
| 3862 | |
| 3863 | const int oc_tail = jcp.oc_tail; |
| 3864 | if (oc_tail) { |
| 3865 | Label skip; |
| 3866 | Reg32 reg_tail_32 = reg_oc_tail.cvt32(); |
| 3867 | if (jcp.nb_oc > 1) { |
| 3868 | kxnorw(k_oc_mask, k_oc_mask, k_oc_mask); |
| 3869 | mov(reg_oc_tail, ptr[param + GET_OFF(load_work)]); |
| 3870 | cmp(reg_oc_tail, 16); |
| 3871 | je(skip, T_NEAR); |
| 3872 | } |
| 3873 | mov(reg_tail_32, (1 << oc_tail) - 1); |
| 3874 | kmovw(k_oc_mask, reg_tail_32); |
| 3875 | L(skip); |
| 3876 | } |
| 3877 | compute_loop(); |
| 3878 | |
| 3879 | postamble(); |
| 3880 | } |
| 3881 | |
| 3882 | status_t jit_avx512_common_conv_bwd_weights_kernel_f32::init_conf( |
| 3883 | jit_conv_conf_t &jcp, const convolution_desc_t &cd, |
| 3884 | memory_desc_t &src_md, memory_desc_t &diff_weights_md, |
| 3885 | memory_desc_t &diff_bias_md, memory_desc_t &diff_dst_md, int nthreads) { |
| 3886 | if (!mayiuse(avx512_core)) return status::unimplemented; |
| 3887 | |
| 3888 | const memory_desc_wrapper src_d(&src_md); |
| 3889 | const memory_desc_wrapper diff_weights_d(&diff_weights_md); |
| 3890 | const memory_desc_wrapper diff_bias_d(&diff_bias_md); |
| 3891 | const memory_desc_wrapper diff_dst_d(&diff_dst_md); |
| 3892 | |
| 3893 | if (!utils::everyone_is(data_type::f32, src_d.data_type(), |
| 3894 | diff_weights_d.data_type(), diff_dst_d.data_type())) |
| 3895 | return status::unimplemented; |
| 3896 | |
| 3897 | const bool with_groups = diff_weights_d.ndims() == src_d.ndims() + 1; |
| 3898 | int ndims = src_d.ndims(); |
| 3899 | |
| 3900 | jcp = zero<decltype(jcp)>(); |
| 3901 | |
| 3902 | jcp.simd_w = cpu_isa_traits<avx512_core>::vlen / typesize; |
| 3903 | jcp.nthr = jcp.aligned_threads = nthreads; |
| 3904 | jcp.ndims = ndims; |
| 3905 | jcp.prop_kind = cd.prop_kind; |
| 3906 | |
| 3907 | jcp.ngroups = with_groups ? diff_weights_d.dims()[0] : 1; |
| 3908 | jcp.mb = src_d.dims()[0]; |
| 3909 | |
| 3910 | jcp.oc = diff_dst_d.dims()[1] / jcp.ngroups; |
| 3911 | jcp.oc_without_padding = jcp.oc; |
| 3912 | jcp.ic = src_d.dims()[1] / jcp.ngroups; |
| 3913 | jcp.ic_without_padding = jcp.ic; |
| 3914 | |
| 3915 | jcp.id = (ndims == 5) ? src_d.dims()[2] : 1; |
| 3916 | jcp.ih = (ndims == 3) ? 1 : src_d.dims()[ndims - 2]; |
| 3917 | jcp.iw = src_d.dims()[ndims - 1]; |
| 3918 | jcp.od = (ndims == 5) ? diff_dst_d.dims()[2] : 1; |
| 3919 | jcp.oh = (ndims == 3) ? 1 : diff_dst_d.dims()[ndims - 2]; |
| 3920 | jcp.ow = diff_dst_d.dims()[ndims - 1]; |
| 3921 | |
| 3922 | jcp.kd = (ndims == 5) ? diff_weights_d.dims()[with_groups + 2] : 1; |
| 3923 | jcp.kh = (ndims == 3) ? 1 : diff_weights_d.dims()[with_groups + ndims - 2]; |
| 3924 | jcp.kw = diff_weights_d.dims()[with_groups + ndims - 1]; |
| 3925 | |
| 3926 | jcp.f_pad = (ndims == 5) ? cd.padding[0][0] : 0; |
| 3927 | jcp.t_pad = (ndims == 3) ? 0 : cd.padding[0][ndims - 4]; |
| 3928 | jcp.l_pad = cd.padding[0][ndims - 3]; |
| 3929 | |
| 3930 | jcp.stride_d = (ndims == 5) ? cd.strides[0] : 1; |
| 3931 | jcp.stride_h = (ndims == 3) ? 1 : cd.strides[ndims - 4]; |
| 3932 | jcp.stride_w = cd.strides[ndims - 3]; |
| 3933 | |
| 3934 | jcp.dilate_d = (ndims == 5) ? cd.dilates[0] : 0; |
| 3935 | jcp.dilate_h = (ndims == 3) ? 0 : cd.dilates[ndims - 4]; |
| 3936 | jcp.dilate_w = cd.dilates[ndims - 3]; |
| 3937 | |
| 3938 | int ext_kw = calculate_extended_filter_size(jcp.kw, jcp.dilate_w); |
| 3939 | int ext_kh = calculate_extended_filter_size(jcp.kh, jcp.dilate_h); |
| 3940 | int ext_kd = calculate_extended_filter_size(jcp.kd, jcp.dilate_d); |
| 3941 | |
| 3942 | bool ok = true |
| 3943 | // general condition to simplify dilations |
| 3944 | && IMPLICATION(jcp.dilate_d != 0, jcp.stride_d == 1) |
| 3945 | && IMPLICATION(jcp.dilate_h != 0, jcp.stride_h == 1) |
| 3946 | // special condition to simplify dilations in compute_oh_loop_common |
| 3947 | && IMPLICATION(jcp.dilate_h != 0, ext_kh <= jcp.ih); |
| 3948 | if (!ok) return status::unimplemented; |
| 3949 | |
| 3950 | jcp.r_pad = nstl::max(0, |
| 3951 | calculate_end_padding( |
| 3952 | jcp.l_pad, jcp.ow, jcp.iw, jcp.stride_w, ext_kw)); |
| 3953 | jcp.b_pad = nstl::max(0, |
| 3954 | calculate_end_padding( |
| 3955 | jcp.t_pad, jcp.oh, jcp.ih, jcp.stride_h, ext_kh)); |
| 3956 | jcp.back_pad = nstl::max(0, |
| 3957 | calculate_end_padding( |
| 3958 | jcp.f_pad, jcp.od, jcp.id, jcp.stride_d, ext_kd)); |
| 3959 | |
| 3960 | /* XXX: currently, does not support dilation_d > 0 */ |
| 3961 | if (ndims == 5 && jcp.dilate_d > 0) return status::unimplemented; |
| 3962 | |
| 3963 | /* Set bounds for large filter 'kw > 14' support and optimized JIT |
| 3964 | * implementation for small output-width 'ow = 1' */ |
| 3965 | const int min_filter_size = 14; |
| 3966 | const int max_filter_size = 20; |
| 3967 | const auto dat_tag_nxc = pick(ndims - 3, nwc, nhwc, ndhwc); |
| 3968 | const auto dat_tag_ncx = pick(ndims - 3, ncw, nchw, ncdhw); |
| 3969 | const auto dat_tag_nCx16c = pick(ndims - 3, nCw16c, nChw16c, nCdhw16c); |
| 3970 | auto curr_src_tag = src_d.matches_one_of_tag( |
| 3971 | dat_tag_nxc, dat_tag_nCx16c, dat_tag_ncx); |
| 3972 | auto curr_dst_tag |
| 3973 | = diff_dst_d.matches_one_of_tag(dat_tag_nxc, dat_tag_nCx16c); |
| 3974 | bool is_data_layout_nxc = IMPLICATION(curr_src_tag != dat_tag_nxc, |
| 3975 | src_d.format_kind() == format_kind::any) |
| 3976 | && IMPLICATION(curr_dst_tag != dat_tag_nxc, |
| 3977 | diff_dst_d.format_kind() == format_kind::any) |
| 3978 | && utils::one_of(dat_tag_nxc, curr_src_tag, curr_dst_tag); |
| 3979 | |
| 3980 | /* Optimization: when `output-width == 1' deploy a special case of the |
| 3981 | * JIT-Kernel by unrolling with regards to height instead of width for |
| 3982 | * the source and filter tensors. The JIT-Kernel also transposes the |
| 3983 | * strides for the input and filter memory access. */ |
| 3984 | jcp.is_hw_transp = !is_data_layout_nxc && ndims == 4 |
| 3985 | && jcp.kw >= min_filter_size && jcp.kw < max_filter_size |
| 3986 | && jcp.ow == 1 && jcp.kw == jcp.iw |
| 3987 | && everyone_is(1, jcp.stride_w, jcp.stride_h) |
| 3988 | && everyone_is(0, jcp.dilate_h, jcp.dilate_w) |
| 3989 | && everyone_is(0, jcp.l_pad, jcp.t_pad, jcp.r_pad, jcp.b_pad); |
| 3990 | if (jcp.is_hw_transp) { |
| 3991 | jcp.tr_kw = jcp.kh; |
| 3992 | jcp.tr_kh = jcp.kw; |
| 3993 | jcp.tr_iw = jcp.ih; |
| 3994 | jcp.tr_ih = jcp.iw; |
| 3995 | } |
| 3996 | |
| 3997 | jcp.ihp = jcp.ih + jcp.t_pad + jcp.b_pad; |
| 3998 | jcp.iwp = jcp.iw + jcp.l_pad + jcp.r_pad; |
| 3999 | jcp.ohp = jcp.oh; |
| 4000 | jcp.owp = jcp.ow; |
| 4001 | jcp.aligned_threads = 0; |
| 4002 | |
| 4003 | /* check for the 1st convolution */ |
| 4004 | jcp.is_1stconv = is_1stconv(jcp); |
| 4005 | |
| 4006 | jcp.oc_block = jcp.simd_w; |
| 4007 | |
| 4008 | bool ok_to_pad_channels = true && !is_data_layout_nxc && jcp.ngroups == 1 |
| 4009 | && src_d.data_type() == data_type::f32; |
| 4010 | |
| 4011 | if (ok_to_pad_channels) jcp.oc = rnd_up(jcp.oc, jcp.simd_w); |
| 4012 | |
| 4013 | if (!IMPLICATION(!is_data_layout_nxc, jcp.oc % jcp.oc_block == 0)) |
| 4014 | return status::unimplemented; |
| 4015 | jcp.ic_tail = is_data_layout_nxc ? jcp.ic % jcp.simd_w : 0; |
| 4016 | jcp.oc_tail = is_data_layout_nxc ? jcp.oc % jcp.simd_w : 0; |
| 4017 | |
| 4018 | auto dst_tag = is_data_layout_nxc ? dat_tag_nxc : dat_tag_nCx16c; |
| 4019 | auto wei_tag = with_groups |
| 4020 | ? pick(ndims - 3, gOIw16i16o, gOIhw16i16o, gOIdhw16i16o) |
| 4021 | : pick(ndims - 3, OIw16i16o, OIhw16i16o, OIdhw16i16o); |
| 4022 | |
| 4023 | if (diff_dst_md.format_kind == format_kind::any) { |
| 4024 | CHECK(memory_desc_init_by_tag(diff_dst_md, dst_tag)); |
| 4025 | } else if (curr_dst_tag != dst_tag) |
| 4026 | return status::unimplemented; |
| 4027 | jcp.dst_tag = dst_tag; |
| 4028 | |
| 4029 | /* conditions on bias memory */ |
| 4030 | jcp.with_bias = cd.diff_bias_desc.format_kind != format_kind::undef; |
| 4031 | if (jcp.with_bias) { |
| 4032 | if (diff_bias_d.format_kind() == format_kind::any) |
| 4033 | CHECK(memory_desc_init_by_tag(diff_bias_md, x)); |
| 4034 | } |
| 4035 | |
| 4036 | jcp.nb_oc = div_up(jcp.oc, jcp.oc_block); |
| 4037 | |
| 4038 | /* kernel applicability check wrt boundaries |
| 4039 | * the conditions are quite general across the kernels we have, |
| 4040 | * but ideally the check should belong to a specific kernel... */ |
| 4041 | const int max_pad_h = ext_kh / 2; |
| 4042 | const bool boundaries_ok = true && jcp.l_pad < ext_kw && jcp.r_pad < ext_kw |
| 4043 | && jcp.t_pad <= max_pad_h && jcp.b_pad <= max_pad_h |
| 4044 | && jcp.f_pad < ext_kd && jcp.back_pad < ext_kd |
| 4045 | && IMPLICATION(jcp.f_pad > 0, jcp.kd < jcp.id + jcp.f_pad) |
| 4046 | && jcp.l_pad <= max_ur_w && jcp.r_pad <= max_ur_w; |
| 4047 | if (!boundaries_ok) return status::unimplemented; |
| 4048 | |
| 4049 | /* yet another common check */ |
| 4050 | if (!jcp.is_hw_transp && jcp.kw > 14) return status::unimplemented; |
| 4051 | |
| 4052 | /* setting register strategy */ |
| 4053 | const int unroll_dim = jcp.is_hw_transp ? jcp.oh : jcp.ow; |
| 4054 | for (int ur_w = nstl::min(max_ur_w, unroll_dim); ur_w > 0; --ur_w) { |
| 4055 | if (unroll_dim % ur_w == 0) { |
| 4056 | jcp.ur_w = ur_w; |
| 4057 | break; |
| 4058 | } |
| 4059 | } |
| 4060 | |
| 4061 | if (jcp.is_1stconv) { |
| 4062 | auto src_tag = is_data_layout_nxc ? dat_tag_nxc : dat_tag_ncx; |
| 4063 | if (src_d.format_kind() == format_kind::any) { |
| 4064 | CHECK(memory_desc_init_by_tag(src_md, src_tag)); |
| 4065 | } else { |
| 4066 | // if `ic == 1`, then `nxc` and `ncx` are effectively equivalent |
| 4067 | if (jcp.ic == 1 && one_of(curr_src_tag, dat_tag_nxc, dat_tag_ncx)) |
| 4068 | src_tag = curr_src_tag; |
| 4069 | if (curr_src_tag != src_tag) return status::unimplemented; |
| 4070 | } |
| 4071 | jcp.src_tag = src_tag; |
| 4072 | |
| 4073 | const bool src_ok = IMPLICATION(!is_data_layout_nxc, |
| 4074 | (one_of(jcp.ic, 1, 2, 3) && jcp.ngroups == 1)); |
| 4075 | if (!src_ok) return status::unimplemented; |
| 4076 | |
| 4077 | jcp.ic_block = jcp.ic; |
| 4078 | |
| 4079 | wei_tag = with_groups ? pick(ndims - 3, gOwi16o, gOhwi16o, gOdhwi16o) |
| 4080 | : pick(ndims - 3, Owi16o, Ohwi16o, Odhwi16o); |
| 4081 | |
| 4082 | if (init_tag(jcp.wei_tag, diff_weights_md, diff_weights_d, wei_tag) |
| 4083 | != status::success) |
| 4084 | return status::unimplemented; |
| 4085 | |
| 4086 | jcp.nb_ic = div_up(jcp.ic, jcp.ic_block); |
| 4087 | |
| 4088 | } else { |
| 4089 | auto src_tag = is_data_layout_nxc ? dat_tag_nxc : dat_tag_nCx16c; |
| 4090 | if (src_md.format_kind == format_kind::any) { |
| 4091 | CHECK(memory_desc_init_by_tag(src_md, src_tag)); |
| 4092 | } else if (curr_src_tag != src_tag) |
| 4093 | return status::unimplemented; |
| 4094 | jcp.src_tag = src_tag; |
| 4095 | |
| 4096 | if (init_tag(jcp.wei_tag, diff_weights_md, diff_weights_d, wei_tag) |
| 4097 | != status::success) |
| 4098 | return status::unimplemented; |
| 4099 | |
| 4100 | jcp.ic_block = jcp.simd_w; |
| 4101 | if (ok_to_pad_channels) jcp.ic = rnd_up(jcp.ic, jcp.ic_block); |
| 4102 | jcp.nb_ic = div_up(jcp.ic, jcp.ic_block); |
| 4103 | } |
| 4104 | |
| 4105 | jcp.typesize_in = typesize; |
| 4106 | jcp.typesize_out = typesize; |
| 4107 | |
| 4108 | bool use_nxc_harness = false; |
| 4109 | if (is_data_layout_nxc) { |
| 4110 | dim_t kernel_size |
| 4111 | = jcp.ic * jcp.oc * jcp.kd * jcp.kh * jcp.kw * jcp.typesize_out; |
| 4112 | dim_t src_size |
| 4113 | = jcp.mb * jcp.ic * jcp.id * jcp.ih * jcp.iw * jcp.typesize_in; |
| 4114 | dim_t diff_dst_size |
| 4115 | = jcp.mb * jcp.oc * jcp.id * jcp.ih * jcp.iw * jcp.typesize_in; |
| 4116 | dim_t data_size = src_size + diff_dst_size; |
| 4117 | |
| 4118 | // The advantage of the nxc kernel is cache traversal, this comes at a |
| 4119 | // cost of extra work updating the weights buffers more often. As such, |
| 4120 | // if everything fits in cache, this kernel is at a disadvantage to the |
| 4121 | // inner loop over ow. More optimizing/balancing is required to |
| 4122 | // determine when this is needed for multidimensional kernels because |
| 4123 | // the data reuses within the kernel height/depth dimension make the |
| 4124 | // computation more computationally bound and cache traversal advantage |
| 4125 | // less important. Due to the current blocked weights format, the |
| 4126 | // weights and the data buffers cannot both be traversed optimally, so |
| 4127 | // for performance, the weights must fit in cache. |
| 4128 | const unsigned int L2_cache_size = platform::get_per_core_cache_size(2); |
| 4129 | use_nxc_harness |
| 4130 | = (data_size / nthreads + kernel_size > L2_cache_size / 3) |
| 4131 | && (jcp.oc % jcp.simd_w == 0) && (jcp.ic % jcp.simd_w == 0) |
| 4132 | && jcp.kw > 1 && ndims == 3 |
| 4133 | && (kernel_size < L2_cache_size / 2); |
| 4134 | } |
| 4135 | |
| 4136 | jcp.harness = use_nxc_harness |
| 4137 | ? harness_nxc |
| 4138 | : ndims == 5 ? harness_3d_reduction : harness_mb_reduction; |
| 4139 | if (jcp.dilate_h == 0 && jcp.ndims == 4 && jcp.oh > min_oh_reduce |
| 4140 | && !jcp.is_hw_transp && !is_data_layout_nxc) |
| 4141 | jcp.harness = harness_2d_reduction; // 2d harness with oh reduction |
| 4142 | bool args_ok = true |
| 4143 | && IMPLICATION(!is_data_layout_nxc, |
| 4144 | jcp.ic % jcp.ic_block == 0 && jcp.oc % jcp.oc_block == 0) |
| 4145 | && jcp.ic <= src_d.padded_dims()[1] |
| 4146 | && jcp.oc <= diff_dst_d.padded_dims()[1] |
| 4147 | && jcp.ic <= diff_weights_d.padded_dims()[with_groups + 1] |
| 4148 | && jcp.oc <= diff_weights_d.padded_dims()[with_groups + 0]; |
| 4149 | if (!args_ok) return status::unimplemented; |
| 4150 | |
| 4151 | int nthr, nthr_mb, nthr_g, nthr_oc_b, nthr_ic_b; |
| 4152 | if (jcp.harness == harness_nxc) { |
| 4153 | // The harness_nxc is quite different from the other kernels. The |
| 4154 | // init_conf function should probably be refactored so that it calls |
| 4155 | // functions along the line of tune_nxc, tune_fma which |
| 4156 | // independently tune the kernels for each implementation with tuning |
| 4157 | // common to multiple implementations performed by helper functions. |
| 4158 | // This will help maintainability and help prevent the different |
| 4159 | // implementations from stepping on each other. |
| 4160 | int zmm_regs = 32; |
| 4161 | |
| 4162 | // Block by ic and kw in the compute kernel to decrease loads from the |
| 4163 | // src buffer |
| 4164 | jcp.ur_ic = 2 - jcp.ic % 2; |
| 4165 | jcp.ur_kw = 1; |
| 4166 | if (jcp.stride_w == jcp.dilate_w + 1) { |
| 4167 | jcp.ur_kw = jcp.kw; |
| 4168 | if (jcp.kw > 7) { |
| 4169 | // Blocking by kw is more effective than by ic in the compute |
| 4170 | // kernel since neighbor kw operations share src data |
| 4171 | jcp.ur_ic = 1; |
| 4172 | if (jcp.kw > zmm_regs / (jcp.ur_ic + 1)) |
| 4173 | jcp.ur_kw = jcp.kw % (zmm_regs / (jcp.ur_ic + 1)); |
| 4174 | } |
| 4175 | } |
| 4176 | |
| 4177 | // Unroll by ow to decrease updates to diff_weights. In practice, this |
| 4178 | // should be approximately 1/4 - 1/2 of the zmm registers |
| 4179 | jcp.ur_ow = nstl::min( |
| 4180 | (zmm_regs - jcp.ur_kw * jcp.ur_ic) / (jcp.ur_ic + 1), jcp.ow); |
| 4181 | |
| 4182 | int work_amount_base = jcp.mb * jcp.od * jcp.oh; |
| 4183 | int ow_iter = div_up(jcp.ow, jcp.ur_ow); |
| 4184 | int nthr_ow = nstl::min( |
| 4185 | jcp.nthr / math::gcd(work_amount_base, jcp.nthr), ow_iter); |
| 4186 | int ow_block = div_up(ow_iter, nthr_ow) * jcp.ur_ow; |
| 4187 | |
| 4188 | jcp.ow_block = ow_block; |
| 4189 | jcp.nb_ow = div_up(jcp.ow, jcp.ow_block); |
| 4190 | |
| 4191 | // Choose a simple parallelization method. A more advance may need made |
| 4192 | // later |
| 4193 | int work_amount = jcp.mb * jcp.od * jcp.oh * jcp.nb_ow; |
| 4194 | nthr_mb = nstl::min(jcp.nthr, work_amount); |
| 4195 | nthr_g = 1; |
| 4196 | nthr_oc_b = 1; |
| 4197 | nthr_ic_b = 1; |
| 4198 | nthr = nthr_mb * nthr_g * nthr_oc_b * nthr_ic_b; |
| 4199 | } else { // balancing |
| 4200 | balance(jcp, nthr, nthr_mb, nthr_g, nthr_oc_b, nthr_ic_b, jcp.nthr); |
| 4201 | } |
| 4202 | |
| 4203 | jcp.nthr = nthr; |
| 4204 | jcp.nthr_mb = nthr_mb; |
| 4205 | jcp.nthr_g = nthr_g; |
| 4206 | jcp.nthr_oc_b = nthr_oc_b; |
| 4207 | jcp.nthr_ic_b = nthr_ic_b; |
| 4208 | |
| 4209 | jcp.kernel_kind = embd_bcast; |
| 4210 | if (is_data_layout_nxc && jcp.stride_w == 1 && jcp.dilate_w == 0 |
| 4211 | && !jcp.is_1stconv) { |
| 4212 | jcp.kernel_kind = expl_bcast; |
| 4213 | } |
| 4214 | |
| 4215 | jcp.nb_ic_blocking_max = 1; |
| 4216 | if (is_data_layout_nxc && (jcp.ow > max_ur_w || jcp.ndims == 5)) { |
| 4217 | assert(!jcp.is_hw_transp); |
| 4218 | jcp.nb_ic_blocking_max = nstl::min(8, div_up(jcp.nb_ic, jcp.nthr_ic_b)); |
| 4219 | } |
| 4220 | |
| 4221 | return status::success; |
| 4222 | } |
| 4223 | |
| 4224 | void jit_avx512_common_conv_bwd_weights_kernel_f32::init_scratchpad( |
| 4225 | memory_tracking::registrar_t &scratchpad, const jit_conv_conf_t &jcp) { |
| 4226 | if (jcp.nthr_mb > 1) { |
| 4227 | const auto wei_size = static_cast<size_t>(jcp.ngroups) |
| 4228 | * rnd_up(jcp.oc, jcp.oc_block) * rnd_up(jcp.ic, jcp.ic_block) |
| 4229 | * jcp.kh * jcp.kw * jcp.kd; |
| 4230 | const auto bia_size = jcp.ngroups * rnd_up(jcp.oc, jcp.oc_block); |
| 4231 | const size_t wei_bia_reduction_size = wei_size + bia_size; |
| 4232 | |
| 4233 | scratchpad.book(key_conv_wei_bia_reduction, |
| 4234 | wei_bia_reduction_size * (jcp.nthr_mb - 1), jcp.typesize_out); |
| 4235 | scratchpad.book<simple_barrier::ctx_t>( |
| 4236 | key_conv_wei_bia_reduction_bctx, 1); |
| 4237 | } |
| 4238 | |
| 4239 | if (jcp.with_bias && jcp.oc_without_padding % jcp.oc_block != 0) { |
| 4240 | const size_t nelems_padded_bias |
| 4241 | = jcp.ngroups * utils::rnd_up(jcp.oc, jcp.oc_block); |
| 4242 | scratchpad.book( |
| 4243 | key_conv_padded_bias, nelems_padded_bias, jcp.typesize_out); |
| 4244 | } |
| 4245 | } |
| 4246 | |
| 4247 | void jit_avx512_common_conv_bwd_weights_kernel_f32::balance( |
| 4248 | const jit_conv_conf_t &j, int &nthr_, int &nthr_mb_, int &nthr_g_, |
| 4249 | int &nthr_oc_b_, int &nthr_ic_b_, int nthreads) { |
| 4250 | nthr_ = nthr_mb_ = nthr_g_ = nthr_oc_b_ = nthr_ic_b_ = 1; |
| 4251 | |
| 4252 | if (nthreads < j.ngroups) { |
| 4253 | /* simplification... fortunately it doesn't hurt much */ |
| 4254 | nthr_ = nthr_g_ = nthreads; |
| 4255 | return; |
| 4256 | } |
| 4257 | |
| 4258 | nthr_g_ = j.ngroups; |
| 4259 | const int nthr = nthreads / nthr_g_; |
| 4260 | |
| 4261 | const int ih = j.is_hw_transp ? j.tr_ih : j.ih; |
| 4262 | const int oh = j.is_hw_transp ? j.ow : j.oh; |
| 4263 | |
| 4264 | int ih_reduce = j.harness == harness_2d_reduction ? ih : 1; |
| 4265 | int oh_reduce = j.harness == harness_2d_reduction ? oh : 1; |
| 4266 | int ih_no_reduce = j.harness == harness_2d_reduction ? 1 : ih; |
| 4267 | int oh_no_reduce = j.harness == harness_2d_reduction ? 1 : oh; |
| 4268 | int nthr_oh_reduce = nstl::max(1, oh_reduce / min_oh_reduce); |
| 4269 | |
| 4270 | auto calc_mem_cost = [=](int nthr_mb, int nthr_oc_b, int nthr_ic_b) { |
| 4271 | /* calculate per thread memory cost (read/write). high level optimizer |
| 4272 | * tries to minimize memory consumption. few notes: |
| 4273 | * (n1) unclear why, but that essentially helps first convolution... |
| 4274 | * (n2) assuming the reduction over minibatch is always there: |
| 4275 | * - instead of 8 it should be 5 here (write ~= 2 read): |
| 4276 | * kernel: temporal workspace 1 write |
| 4277 | * reduction: 1 read from workspace and 1 write to the diff_wei |
| 4278 | * - but experiments showed 8 works better than 5 or 6... */ |
| 4279 | |
| 4280 | const dim_t src_coef = 1; |
| 4281 | const dim_t dst_coef = 1; |
| 4282 | const dim_t wei_coef = 8; |
| 4283 | const dim_t iw = j.is_hw_transp ? j.tr_iw : j.iw; |
| 4284 | const dim_t ow = j.is_hw_transp ? j.oh : j.ow; |
| 4285 | |
| 4286 | return 0 |
| 4287 | + src_coef * div_up(j.mb * ih_reduce, nthr_mb) |
| 4288 | * div_up(j.ngroups, nthr_g_) * div_up(j.nb_ic, nthr_ic_b) |
| 4289 | * j.ic_block * ih_no_reduce * iw * j.id / j.stride_d |
| 4290 | / j.stride_h / j.stride_w /* (n1) */ |
| 4291 | + dst_coef * div_up(j.mb * oh_reduce, nthr_mb) |
| 4292 | * div_up(j.ngroups, nthr_g_) * div_up(j.nb_oc, nthr_oc_b) |
| 4293 | * j.oc_block * oh_no_reduce * ow * j.od |
| 4294 | + wei_coef /* (n2) */ |
| 4295 | * div_up(j.ngroups, nthr_g_) * div_up(j.nb_oc, nthr_oc_b) |
| 4296 | * div_up(j.nb_ic, nthr_ic_b) * j.kh * j.kw * j.kd * j.ic_block |
| 4297 | * j.oc_block; |
| 4298 | }; |
| 4299 | |
| 4300 | dim_t best_mem_cost = calc_mem_cost(nthr_mb_, nthr_oc_b_, nthr_ic_b_); |
| 4301 | |
| 4302 | /* step 1: find the best thread distribution with lowest memory cost */ |
| 4303 | const int nthr_mb_max = nstl::min(nthr, j.mb * j.od * nthr_oh_reduce); |
| 4304 | for (int nthr_mb = 1; nthr_mb <= nthr_mb_max; ++nthr_mb) { |
| 4305 | const int nthr_par = nthr / nthr_mb; |
| 4306 | const int nthr_oc_b_max = nstl::min(nthr_par, j.nb_oc); |
| 4307 | for (int nthr_oc_b = 1; nthr_oc_b <= nthr_oc_b_max; ++nthr_oc_b) { |
| 4308 | int nthr_ic_b = nstl::min(nthr_par / nthr_oc_b, j.nb_ic); |
| 4309 | |
| 4310 | dim_t mem_cost = calc_mem_cost(nthr_mb, nthr_oc_b, nthr_ic_b); |
| 4311 | if (mem_cost <= best_mem_cost) { |
| 4312 | best_mem_cost = mem_cost; |
| 4313 | nthr_mb_ = nthr_mb; |
| 4314 | nthr_oc_b_ = nthr_oc_b; |
| 4315 | nthr_ic_b_ = nthr_ic_b; |
| 4316 | } |
| 4317 | } |
| 4318 | } |
| 4319 | |
| 4320 | auto calc_comp_cost = [=](int nthr_mb, int nthr_oc_b, int nthr_ic_b) { |
| 4321 | return (dim_t)div_up(j.mb * oh_reduce, nthr_mb) |
| 4322 | * div_up(j.ngroups, nthr_g_) * div_up(j.nb_oc, nthr_oc_b) |
| 4323 | * div_up(j.nb_ic, nthr_ic_b); |
| 4324 | }; |
| 4325 | |
| 4326 | /* step 2: search for a thread distribution with lower compute cost. |
| 4327 | * the constrains: |
| 4328 | * - memory cost cannot exceed 110% of the best found in the step 1 |
| 4329 | * - unless compute cost is 133% lower than the current best case |
| 4330 | * note: both constants were found empirically */ |
| 4331 | dim_t best_comp_cost = calc_comp_cost(nthr_mb_, nthr_oc_b_, nthr_ic_b_); |
| 4332 | for (int nthr_mb = 1; nthr_mb <= nthr_mb_max; ++nthr_mb) { |
| 4333 | const int nthr_par = nthr / nthr_mb; |
| 4334 | const int nthr_oc_b_max = nstl::min(nthr_par, j.nb_oc); |
| 4335 | for (int nthr_oc_b = 1; nthr_oc_b <= nthr_oc_b_max; ++nthr_oc_b) { |
| 4336 | int nthr_ic_b = nstl::min(nthr_par / nthr_oc_b, j.nb_ic); |
| 4337 | dim_t mem_cost = calc_mem_cost(nthr_mb, nthr_oc_b, nthr_ic_b); |
| 4338 | dim_t comp_cost = calc_comp_cost(nthr_mb, nthr_oc_b, nthr_ic_b); |
| 4339 | |
| 4340 | const bool opt1 = comp_cost <= best_comp_cost |
| 4341 | && IMPLICATION( |
| 4342 | !j.is_hw_transp, mem_cost < 1.1 * best_mem_cost); |
| 4343 | const bool opt2 = 4 * comp_cost <= 3 * best_comp_cost; |
| 4344 | |
| 4345 | if (opt1 || opt2) { |
| 4346 | best_comp_cost = comp_cost; |
| 4347 | nthr_mb_ = nthr_mb; |
| 4348 | nthr_oc_b_ = nthr_oc_b; |
| 4349 | nthr_ic_b_ = nthr_ic_b; |
| 4350 | } |
| 4351 | } |
| 4352 | } |
| 4353 | |
| 4354 | if (nthr_mb_ > nthreads / 2 && nthr_mb_ < nthreads) |
| 4355 | nthr_mb_ = nstl::min(j.mb * j.od * nthr_oh_reduce, nthreads); |
| 4356 | nthr_ = nthr_mb_ * nthr_g_ * nthr_oc_b_ * nthr_ic_b_; |
| 4357 | |
| 4358 | assert(nthr_ <= nthreads); |
| 4359 | } |
| 4360 | |
| 4361 | template struct _jit_avx512_common_conv_fwd_kernel<Zmm>; |
| 4362 | template struct _jit_avx512_common_conv_fwd_kernel<Ymm>; |
| 4363 | template struct _jit_avx512_common_conv_fwd_kernel<Xmm>; |
| 4364 | template struct _jit_avx512_common_conv_bwd_data_kernel_f32<Zmm>; |
| 4365 | template struct _jit_avx512_common_conv_bwd_data_kernel_f32<Ymm>; |
| 4366 | template struct _jit_avx512_common_conv_bwd_data_kernel_f32<Xmm>; |
| 4367 | |
| 4368 | } // namespace x64 |
| 4369 | } // namespace cpu |
| 4370 | } // namespace impl |
| 4371 | } // namespace dnnl |
| 4372 | |
| 4373 | // vim: et ts=4 sw=4 cindent cino+=l0,\:4,N-s |
| 4374 |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.