| 1 | /******************************************************************************* |
|---|---|
| 2 | * Copyright 2016-2018 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 "c_types_map.hpp" |
| 18 | #include "nstl.hpp" |
| 19 | #include "type_helpers.hpp" |
| 20 | #include "utils.hpp" |
| 21 | |
| 22 | #include "cpu_barrier.hpp" |
| 23 | |
| 24 | #include "jit_avx512_common_conv_kernel.hpp" |
| 25 | |
| 26 | #define GET_OFF(field) offsetof(jit_conv_call_s, field) |
| 27 | #define KNx_L2_EFFECTIVE_CAPACITY ((512-64)*1024) |
| 28 | |
| 29 | namespace mkldnn { |
| 30 | namespace impl { |
| 31 | namespace cpu { |
| 32 | |
| 33 | using namespace mkldnn::impl::format_tag; |
| 34 | using namespace mkldnn::impl::memory_tracking::names; |
| 35 | using namespace mkldnn::impl::utils; |
| 36 | using namespace Xbyak; |
| 37 | |
| 38 | namespace { |
| 39 | |
| 40 | constexpr auto small_spatial = 14; |
| 41 | unsigned int L1_cache_size = get_cache_size(1, true); |
| 42 | |
| 43 | inline void pick_loop_order(jit_conv_conf_t &jcp) { |
| 44 | using namespace prop_kind; |
| 45 | assert(one_of(jcp.prop_kind, |
| 46 | forward_training, forward_inference, backward_data)); |
| 47 | auto w = (jcp.prop_kind == backward_data) ? jcp.iw : jcp.ow; |
| 48 | auto h = (jcp.prop_kind == backward_data) ? jcp.ih : jcp.oh; |
| 49 | |
| 50 | // ow-threading is currently implemented for forward only |
| 51 | // TODO: single code for fwd and bwd after ow-thr for bwd |
| 52 | // meaningless switch was removed |
| 53 | if (jcp.prop_kind == backward_data) { |
| 54 | jcp.loop_order = (w <= small_spatial && h <= small_spatial) |
| 55 | ? loop_cgn : loop_gnc; |
| 56 | } else { |
| 57 | jcp.loop_order = (w <= small_spatial && h <= small_spatial) |
| 58 | ? loop_cwgn : loop_gncw; |
| 59 | } |
| 60 | } |
| 61 | |
| 62 | inline bool is_1stconv(const jit_conv_conf_t &jcp) { |
| 63 | if (mayiuse(avx512_core)) |
| 64 | return (jcp.ic < 16 && jcp.ngroups == 1); |
| 65 | else |
| 66 | return one_of(jcp.ic, 1, 3); |
| 67 | } |
| 68 | |
| 69 | inline bool is_ow_threading_on(const jit_conv_conf_t &jcp) { |
| 70 | return (jcp.nb_ow > 1); |
| 71 | } |
| 72 | |
| 73 | inline bool is_owb_prefetching(const jit_conv_conf_t &jcp) { |
| 74 | return (jcp.ver == ver_4fma && is_ow_threading_on(jcp)); |
| 75 | } |
| 76 | |
| 77 | } |
| 78 | |
| 79 | template<typename Vmm> |
| 80 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::prepare_output(int ur_w) |
| 81 | { |
| 82 | for (int k = 0; k < jcp.nb_oc_blocking; k++) |
| 83 | for (int j = 0; j < ur_w; j++) { |
| 84 | Vmm vmm = vmm_out(j, k); |
| 85 | vpxord(vmm, vmm, vmm); |
| 86 | if (!is_owb_prefetching(jcp)) { |
| 87 | size_t aux_output_offset = get_output_offset(j, k); |
| 88 | mic_prefetcht1(EVEX_compress_addr_safe(reg_out_prf, |
| 89 | aux_output_offset, reg_out_long_offt)); |
| 90 | } |
| 91 | } |
| 92 | } |
| 93 | |
| 94 | template<typename Vmm> |
| 95 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::store_output(int ur_w) |
| 96 | { |
| 97 | Label no_update_label, store_label, eltwise_label; |
| 98 | |
| 99 | mov(reg_channel, ptr[param1 + GET_OFF(channel)]); |
| 100 | if (jcp.with_bias) { |
| 101 | mov(reg_bias, ptr[param1 + GET_OFF(bias)]); |
| 102 | } |
| 103 | |
| 104 | if (!jcp.with_sum) { |
| 105 | cmp(reg_channel, 0); |
| 106 | je(no_update_label, T_NEAR); |
| 107 | } |
| 108 | |
| 109 | for (int k = 0; k < jcp.nb_oc_blocking; k++) |
| 110 | for (int j = 0; j < ur_w; j++) { |
| 111 | Vmm vmm = vmm_out(j, k); |
| 112 | size_t aux_output_offset = get_output_offset(j, k); |
| 113 | vaddps(vmm, |
| 114 | make_safe_addr(reg_out, aux_output_offset, reg_out_long_offt)); |
| 115 | } |
| 116 | |
| 117 | if (!jcp.with_sum) { |
| 118 | jmp(eltwise_label, T_NEAR); |
| 119 | } else { |
| 120 | cmp(reg_channel, 0); |
| 121 | jne(eltwise_label, T_NEAR); |
| 122 | } |
| 123 | |
| 124 | L(no_update_label); |
| 125 | if (jcp.with_bias) { |
| 126 | for (int k = 0; k < jcp.nb_oc_blocking; k++) { |
| 127 | int bias_offset = jcp.typesize_out * k * jcp.oc_block; |
| 128 | for (int j = 0; j < ur_w; j++) { |
| 129 | Vmm vmm = vmm_out(j, k); |
| 130 | vaddps(vmm, EVEX_compress_addr(reg_bias, bias_offset)); |
| 131 | } |
| 132 | mic_prefetcht1(EVEX_compress_addr(reg_bias, bias_offset + 64)); |
| 133 | } |
| 134 | } |
| 135 | |
| 136 | L(eltwise_label); |
| 137 | if (jcp.with_eltwise) { |
| 138 | cmp(reg_channel, jcp.nb_ic - 1); |
| 139 | jl(store_label, T_NEAR); |
| 140 | |
| 141 | if (ur_w == jcp.ur_w) { |
| 142 | eltwise_injector_->compute_vector_range(0, |
| 143 | jcp.nb_oc_blocking * jcp.ur_w); |
| 144 | } else { |
| 145 | for (int k = 0; k < jcp.nb_oc_blocking; k++) |
| 146 | eltwise_injector_->compute_vector_range(k * jcp.ur_w, |
| 147 | k * jcp.ur_w + ur_w); |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | L(store_label); |
| 152 | for (int k = 0; k < jcp.nb_oc_blocking; k++) |
| 153 | for (int j = 0; j < ur_w; j++) { |
| 154 | Vmm vmm = vmm_out(j, k); |
| 155 | size_t aux_output_offset = (size_t)typesize * |
| 156 | ((size_t)k * jcp.od * jcp.oh * jcp.ow + j) * jcp.oc_block; |
| 157 | vmovups(EVEX_compress_addr_safe(reg_out, aux_output_offset, |
| 158 | reg_out_long_offt), vmm); |
| 159 | if (!is_owb_prefetching(jcp)) |
| 160 | mic_prefetcht0(EVEX_compress_addr_safe(reg_out_prf, |
| 161 | aux_output_offset, reg_out_long_offt)); |
| 162 | } |
| 163 | } |
| 164 | |
| 165 | template<typename Vmm> |
| 166 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::compute_loop_4fma_1st(int ur_w, |
| 167 | int pad_l, int pad_r) |
| 168 | { |
| 169 | } |
| 170 | |
| 171 | template<> |
| 172 | void _jit_avx512_common_conv_fwd_kernel<Zmm>::compute_loop_4fma_1st(int ur_w, |
| 173 | int pad_l, int pad_r) |
| 174 | { |
| 175 | assert(jcp.dilate_d == 0 && jcp.dilate_h == 0 && jcp.dilate_w == 0); |
| 176 | |
| 177 | int iw = jcp.iw; |
| 178 | int ih = jcp.ih; |
| 179 | int kw = jcp.kw; |
| 180 | int stride_w = jcp.stride_w; |
| 181 | int ic_block = jcp.ic_block; |
| 182 | int oc_block = jcp.oc_block; |
| 183 | |
| 184 | Label kh_label, kd_label; |
| 185 | |
| 186 | if (one_of(jcp.ndims, 3, 4)) { |
| 187 | mov(aux_reg_inp, reg_inp); |
| 188 | mov(aux_reg_ker, reg_ker); |
| 189 | mov(aux_reg_inp_prf, reg_inp_prf); |
| 190 | } |
| 191 | |
| 192 | size_t max_input_offset = (size_t)jcp.typesize_in |
| 193 | * ((size_t)(kw + ur_w * stride_w - pad_l) |
| 194 | + (size_t)ic_block * iw * ih * jcp.id); |
| 195 | assert(reg_inp_prf == reg_long_offt); |
| 196 | if (max_input_offset > INT_MAX) push(reg_inp_prf); |
| 197 | |
| 198 | if (jcp.ndims == 5) { |
| 199 | push(reg_out_prf); |
| 200 | push(reg_out); |
| 201 | |
| 202 | mov(reg_ki, ptr[param1 + GET_OFF(kd_padding)]); |
| 203 | mov(aux_reg_ker_d, ptr[param1 + GET_OFF(filt)]); |
| 204 | mov(aux_reg_inp_d, reg_inp); |
| 205 | mov(aux_reg_inp_d_prf, reg_inp_prf); |
| 206 | |
| 207 | L(kd_label); |
| 208 | } |
| 209 | mov(reg_kj, reg_kh); |
| 210 | if (jcp.ndims == 5) { |
| 211 | mov(aux_reg_inp, aux_reg_inp_d); |
| 212 | mov(aux_reg_ker, aux_reg_ker_d); |
| 213 | mov(aux_reg_inp_prf, aux_reg_inp_d_prf); |
| 214 | } |
| 215 | |
| 216 | L(kh_label); |
| 217 | for (int ki = 0; ki < kw; ki += 4) { |
| 218 | for (int ic = 0; ic < ic_block; ic++) { |
| 219 | for (int i = 0; i < 4; i++) { |
| 220 | int aux_ker_offset |
| 221 | = jcp.typesize_in |
| 222 | * ((ki + i) * oc_block |
| 223 | + ic * kw * jcp.kh * jcp.kd * oc_block); |
| 224 | if (ki + i < kw) |
| 225 | vmovups(vmm_ker(i), |
| 226 | EVEX_compress_addr(aux_reg_ker, aux_ker_offset)); |
| 227 | else |
| 228 | vpxord(vmm_ker(i), vmm_ker(i), vmm_ker(i)); |
| 229 | } |
| 230 | |
| 231 | int j_start = get_ow_start(ki, pad_l); |
| 232 | int j_end = get_ow_end(ur_w, ki, pad_r); |
| 233 | |
| 234 | for (int j = j_start, prf_count=0; j < j_end; j++) { |
| 235 | size_t aux_input_offset = (size_t)jcp.typesize_in |
| 236 | * ((size_t)(ki + j * stride_w |
| 237 | - pad_l) + (size_t)ic * iw * ih * jcp.id); |
| 238 | v4fmaddps(vmm_out(j, 0), vmm_ker(0), |
| 239 | EVEX_compress_addr_safe(aux_reg_inp, aux_input_offset, |
| 240 | reg_long_offt)); |
| 241 | if (ki + prf_count < kw && prf_count < 4 |
| 242 | && ((ki < 2 && j % 4) || j % 2)) { |
| 243 | int aux_ker_offset = jcp.typesize_in |
| 244 | * ((ki + prf_count) * oc_block |
| 245 | + ic * kw * jcp.kh * jcp.kd * oc_block + kw * oc_block); |
| 246 | mic_prefetcht0(EVEX_compress_addr(aux_reg_ker, |
| 247 | aux_ker_offset)); |
| 248 | prf_count++; |
| 249 | } |
| 250 | if (ki == 0 |
| 251 | && j % (64 / (stride_w * jcp.typesize_in)) == 0) { |
| 252 | mic_prefetcht0(EVEX_compress_addr_safe(aux_reg_inp_prf, |
| 253 | aux_input_offset, reg_long_offt)); |
| 254 | } |
| 255 | if (ki == 1 |
| 256 | && j % (64 / (stride_w * jcp.typesize_in)) == 0) { |
| 257 | mic_prefetcht0(EVEX_compress_addr_safe(aux_reg_inp, |
| 258 | aux_input_offset+jcp.typesize_in * iw, reg_long_offt)); |
| 259 | } |
| 260 | } |
| 261 | } |
| 262 | } |
| 263 | add(aux_reg_ker, jcp.typesize_in * kw * oc_block); |
| 264 | add(aux_reg_inp, jcp.typesize_in * iw); |
| 265 | add(aux_reg_inp_prf, jcp.typesize_in * iw); |
| 266 | |
| 267 | dec(reg_kj); |
| 268 | cmp(reg_kj, 0); |
| 269 | jg(kh_label, T_NEAR); |
| 270 | |
| 271 | if (jcp.ndims == 5) { |
| 272 | add(aux_reg_inp_d, typesize * jcp.ih * jcp.iw); |
| 273 | add(aux_reg_ker_d, typesize * jcp.kw * jcp.kh * oc_block); |
| 274 | add(aux_reg_inp_d_prf, typesize * jcp.ih * jcp.iw); |
| 275 | |
| 276 | dec(reg_ki); |
| 277 | cmp(reg_ki, 0); |
| 278 | jg(kd_label, T_NEAR); |
| 279 | |
| 280 | pop(reg_out); |
| 281 | pop(reg_out_prf); |
| 282 | } |
| 283 | |
| 284 | if (max_input_offset > INT_MAX) pop(reg_inp_prf); |
| 285 | } |
| 286 | |
| 287 | template<typename Vmm> |
| 288 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::compute_loop_4fma(int ur_w, |
| 289 | int pad_l, int pad_r) |
| 290 | { |
| 291 | } |
| 292 | |
| 293 | template<> |
| 294 | void _jit_avx512_common_conv_fwd_kernel<Zmm>::compute_loop_4fma(int ur_w, |
| 295 | int pad_l, int pad_r) |
| 296 | { |
| 297 | int stride_w = jcp.stride_w; |
| 298 | int ic_block = jcp.ic_block; |
| 299 | int oc_block = jcp.oc_block; |
| 300 | Label kh_label, last_iter_label, loop_end_label, kd_label; |
| 301 | int ker_load_number = 4; |
| 302 | int shift_kernel_ptr = typesize * jcp.kw * jcp.oc_block * jcp.ic_block; |
| 303 | int shift_input_ptr = typesize * (jcp.dilate_h + 1) * jcp.iw * jcp.ic_block; |
| 304 | |
| 305 | bool check_last_kh = (jcp.kh > 3); |
| 306 | bool pref_current_inp = (jcp.iw < 14 || jcp.iw > 28); |
| 307 | |
| 308 | int oi_ipref_t0 = get_ow_start(0, pad_l); |
| 309 | int ow_end_ipref = get_ow_end(ur_w, 0, pad_r); |
| 310 | |
| 311 | assert(jcp.oc % jcp.nb_oc_blocking == 0); |
| 312 | |
| 313 | auto kernel_offset = [=](int ocb, int ic, int ki) { |
| 314 | int blk_idx = ocb * jcp.nb_ic * jcp.kh * jcp.kw * jcp.kd + ki; |
| 315 | int blk_offset = blk_idx * jcp.oc_block * jcp.ic_block; |
| 316 | int ic_offset = ic * jcp.oc_block; |
| 317 | return typesize * (blk_offset + ic_offset); |
| 318 | }; |
| 319 | auto kernel_loads = [=](int ki, int ic, int kk) { |
| 320 | for (int ii = 0; ii < ker_load_number; ii++) { |
| 321 | int aux_kernel_offset = kernel_offset(kk, ic + ii, ki); |
| 322 | vmovups(vmm_ker(ii), |
| 323 | EVEX_compress_addr(aux_reg_ker, aux_kernel_offset)); |
| 324 | } |
| 325 | }; |
| 326 | auto prefetch_inp_next_kh = [&](int ki, int ki_start, int cnt0, int cnt1) { |
| 327 | if (cnt1 >= ker_load_number && cnt0 >= ker_load_number |
| 328 | && ki >= ki_start && oi_ipref_t0 < ow_end_ipref) { |
| 329 | int aux_inp_offset |
| 330 | = typesize |
| 331 | * ((oi_ipref_t0 * stride_w - pad_l) * ic_block |
| 332 | + (jcp.dilate_h + 1) * jcp.iw * ic_block); |
| 333 | prefetcht0(EVEX_compress_addr(aux_reg_inp, |
| 334 | aux_inp_offset)); |
| 335 | oi_ipref_t0++; |
| 336 | } |
| 337 | }; |
| 338 | |
| 339 | if (one_of(jcp.ndims, 3, 4)) { |
| 340 | mov(aux_reg_inp, reg_inp); |
| 341 | mov(aux_reg_ker, reg_ker); |
| 342 | mov(aux_reg_ker_prf, reg_ker_prf); |
| 343 | mov(aux_reg_inp_prf, reg_inp_prf); |
| 344 | } |
| 345 | |
| 346 | if (jcp.ndims == 5) { |
| 347 | push(reg_out_prf); |
| 348 | push(reg_out); |
| 349 | |
| 350 | mov(reg_ki, ptr[param1 + GET_OFF(kd_padding)]); |
| 351 | mov(aux_reg_ker_d, ptr[param1 + GET_OFF(filt)]); |
| 352 | mov(aux_reg_inp_d, reg_inp); |
| 353 | mov(aux_reg_inp_d_prf, reg_inp_prf); |
| 354 | mov(aux_reg_ker_d_prf, reg_ker_prf); |
| 355 | L(kd_label); |
| 356 | mov(reg_kj, ptr[param1 + GET_OFF(kh_padding)]); |
| 357 | } else { |
| 358 | mov(reg_kj, reg_kh); |
| 359 | } |
| 360 | if (jcp.ndims == 5) { |
| 361 | mov(aux_reg_inp, aux_reg_inp_d); |
| 362 | mov(aux_reg_ker, aux_reg_ker_d); |
| 363 | mov(aux_reg_ker_prf, aux_reg_ker_d_prf); |
| 364 | mov(aux_reg_inp_prf, aux_reg_inp_d_prf); |
| 365 | } |
| 366 | |
| 367 | align(16); |
| 368 | L(kh_label); |
| 369 | int kw = jcp.kw; |
| 370 | if (check_last_kh) { |
| 371 | for (int ki = 0; ki < kw; ki++) |
| 372 | for (int ic = 0; ic < ic_block; ic += 4) |
| 373 | for (int kk = 0; kk < jcp.nb_oc_blocking; kk++) { |
| 374 | bool last_kernel_loads = (kk == jcp.nb_oc_blocking - 1 |
| 375 | && ki == kw - 1 && (ic + 4) == ic_block); |
| 376 | |
| 377 | if (last_kernel_loads) { |
| 378 | cmp(reg_kj, 1); |
| 379 | je(last_iter_label, T_NEAR); |
| 380 | } |
| 381 | |
| 382 | kernel_loads(ki, ic, kk); |
| 383 | for (int oi = get_ow_start(ki, pad_l), prf_count_t1 = 0, |
| 384 | prf_count_t0 = 0; |
| 385 | oi < get_ow_end(ur_w, ki, pad_r); oi++) { |
| 386 | int aux_input_offset = typesize |
| 387 | * ((ki * (jcp.dilate_w + 1) + oi * stride_w |
| 388 | - pad_l) * ic_block |
| 389 | + ic); |
| 390 | v4fmaddps(vmm_out(oi, kk), vmm_ker(0), |
| 391 | EVEX_compress_addr(aux_reg_inp, aux_input_offset)); |
| 392 | |
| 393 | if (oi % 2) { |
| 394 | if (prf_count_t0 < 4) { |
| 395 | int aux_kernel_prf; |
| 396 | if (last_kernel_loads) |
| 397 | aux_kernel_prf= kernel_offset(0, |
| 398 | prf_count_t0 + ic + 4 |
| 399 | - ic_block, 0) + typesize * kw |
| 400 | * oc_block * ic_block; |
| 401 | else |
| 402 | aux_kernel_prf = kernel_offset(kk, ic + 4 |
| 403 | + prf_count_t0, ki); |
| 404 | mic_prefetcht0(EVEX_compress_addr(aux_reg_ker, |
| 405 | aux_kernel_prf)); |
| 406 | prf_count_t0++; |
| 407 | } else if (prf_count_t1 < 4) { |
| 408 | mic_prefetcht1(EVEX_compress_addr( |
| 409 | aux_reg_ker_prf, kernel_offset(kk, ic |
| 410 | + prf_count_t1, ki))); |
| 411 | prf_count_t1++; |
| 412 | } |
| 413 | } else |
| 414 | prefetch_inp_next_kh(ki, 2, prf_count_t0, |
| 415 | prf_count_t1); |
| 416 | } |
| 417 | |
| 418 | if (last_kernel_loads) { |
| 419 | jmp(loop_end_label, T_NEAR); |
| 420 | |
| 421 | L(last_iter_label); |
| 422 | |
| 423 | kernel_loads(ki, ic, kk); |
| 424 | for (int oi = get_ow_start(ki, pad_l), prf_count_t1 = 0, |
| 425 | prf_count_t0 = 0; |
| 426 | oi < get_ow_end(ur_w, ki, pad_r); oi++) { |
| 427 | int aux_input_offset = typesize |
| 428 | * ((ki * (jcp.dilate_w + 1) + oi * stride_w |
| 429 | - pad_l) * ic_block |
| 430 | + ic); |
| 431 | v4fmaddps(vmm_out(oi, kk), vmm_ker(0), |
| 432 | EVEX_compress_addr(aux_reg_inp, |
| 433 | aux_input_offset)); |
| 434 | if (oi % 2) { |
| 435 | if (prf_count_t0 < 4) { |
| 436 | mic_prefetcht0(EVEX_compress_addr( |
| 437 | aux_reg_ker_prf, kernel_offset(0, |
| 438 | prf_count_t0, 0))); |
| 439 | prf_count_t0++; |
| 440 | } else if (prf_count_t1 < 4) { |
| 441 | mic_prefetcht1(EVEX_compress_addr( |
| 442 | aux_reg_ker_prf, kernel_offset(kk, |
| 443 | ic + prf_count_t1, ki))); |
| 444 | prf_count_t1++; |
| 445 | } |
| 446 | } |
| 447 | } |
| 448 | L(loop_end_label); |
| 449 | } |
| 450 | } |
| 451 | } else { |
| 452 | for (int ki = 0; ki < kw; ki++) |
| 453 | for (int ic = 0; ic < ic_block; ic += 4) |
| 454 | for (int kk = 0; kk < jcp.nb_oc_blocking; kk++) { |
| 455 | kernel_loads(ki, ic, kk); |
| 456 | for (int oi = get_ow_start(ki, pad_l), |
| 457 | prf_count_t1 = 0, prf_count_t0 = 0; |
| 458 | oi < get_ow_end(ur_w, ki, pad_r); oi++) { |
| 459 | int aux_input_offset = typesize |
| 460 | * ((ki * (jcp.dilate_w + 1) + oi * stride_w |
| 461 | - pad_l) * ic_block + ic); |
| 462 | v4fmaddps(vmm_out(oi, kk), vmm_ker(0), |
| 463 | EVEX_compress_addr(aux_reg_inp, |
| 464 | aux_input_offset)); |
| 465 | |
| 466 | if (!is_owb_prefetching(jcp)) { |
| 467 | if ((oi % 2) && (prf_count_t1 < 4)) { |
| 468 | mic_prefetcht1(EVEX_compress_addr( |
| 469 | aux_reg_ker_prf, kernel_offset(kk, |
| 470 | ic + prf_count_t1, ki))); |
| 471 | prf_count_t1++; |
| 472 | } |
| 473 | } else { |
| 474 | if (!(ki == 0 && ic == 0) |
| 475 | && !(ki == kw-1 && ic == 0) && |
| 476 | (oi % 2) && (prf_count_t1 < 4) |
| 477 | ) { |
| 478 | mic_prefetcht0(EVEX_compress_addr( |
| 479 | aux_reg_ker, kernel_offset(kk, |
| 480 | ic + 4 + prf_count_t0, ki))); |
| 481 | prf_count_t0++; |
| 482 | } |
| 483 | } |
| 484 | if (!is_owb_prefetching(jcp)) { |
| 485 | if (pref_current_inp) { |
| 486 | if (ki == 0 && ic == 0 && kk == 0) |
| 487 | mic_prefetcht0(EVEX_compress_addr( |
| 488 | aux_reg_inp, |
| 489 | aux_input_offset + shift_input_ptr)); |
| 490 | } else { |
| 491 | if (ki == 1 && ic == 0 && kk == 0) |
| 492 | mic_prefetcht1(EVEX_compress_addr( |
| 493 | aux_reg_inp_prf, aux_input_offset)); |
| 494 | } |
| 495 | } else { |
| 496 | int inp_mult = jcp.is_1stconv ? 1 : jcp.ic_block; |
| 497 | int inp_shift |
| 498 | = jcp.typesize_in * ur_w * stride_w * inp_mult; |
| 499 | bool kk_pref_slot = kk ? oi % 2 : !(oi % 2); |
| 500 | if (ki == 0 && ic == 0 && kk_pref_slot) |
| 501 | mic_prefetcht1(EVEX_compress_addr( |
| 502 | aux_reg_inp, |
| 503 | aux_input_offset + inp_shift)); |
| 504 | |
| 505 | if (ki == kw - 1 && ic == 0 && kk_pref_slot) |
| 506 | mic_prefetcht0(EVEX_compress_addr( |
| 507 | aux_reg_inp, |
| 508 | aux_input_offset + inp_shift)); |
| 509 | } |
| 510 | } |
| 511 | } |
| 512 | } |
| 513 | |
| 514 | add(aux_reg_ker, shift_kernel_ptr); |
| 515 | add(aux_reg_inp, shift_input_ptr); |
| 516 | add(aux_reg_ker_prf, shift_kernel_ptr); |
| 517 | add(aux_reg_inp_prf, shift_input_ptr); |
| 518 | |
| 519 | dec(reg_kj); |
| 520 | cmp(reg_kj, 0); |
| 521 | jg(kh_label, T_NEAR); |
| 522 | |
| 523 | if (jcp.ndims == 5) { |
| 524 | add(aux_reg_inp_d, |
| 525 | typesize * (jcp.dilate_d + 1) * jcp.ih * jcp.iw * jcp.ic_block); |
| 526 | add(aux_reg_ker_d, typesize * jcp.kw * jcp.kh * jcp.oc_block |
| 527 | * jcp.ic_block); |
| 528 | add(aux_reg_inp_d_prf, |
| 529 | typesize * (jcp.dilate_d + 1) * jcp.ih * jcp.iw * jcp.ic_block); |
| 530 | add(aux_reg_ker_d_prf, typesize * jcp.kw * jcp.kh * jcp.oc_block |
| 531 | * jcp.ic_block); |
| 532 | |
| 533 | dec(reg_ki); |
| 534 | cmp(reg_ki, 0); |
| 535 | jg(kd_label, T_NEAR); |
| 536 | |
| 537 | pop(reg_out); |
| 538 | pop(reg_out_prf); |
| 539 | } |
| 540 | } |
| 541 | |
| 542 | template<typename Vmm> |
| 543 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::compute_loop_fma(int ur_w, |
| 544 | int pad_l, int pad_r) |
| 545 | { |
| 546 | bool prf_ker = true; |
| 547 | bool prf_inp = true; |
| 548 | int ih = jcp.ih; |
| 549 | int stride_w = jcp.stride_w; |
| 550 | int id = jcp.id; |
| 551 | int iw = jcp.iw; |
| 552 | int kw = jcp.kw; |
| 553 | int ic_block = jcp.ic_block; |
| 554 | int oc_block = jcp.oc_block; |
| 555 | int nb_oc_block = jcp.nb_oc_blocking; |
| 556 | Label kh_label, kd_label; |
| 557 | |
| 558 | int ker_pipeline_depth = 4; |
| 559 | assert(ker_reg_base_idx + ker_pipeline_depth <= 32); |
| 560 | assert(oc_block >= ker_pipeline_depth); |
| 561 | |
| 562 | int num_ker_loads = ic_block * nb_oc_block * kw; |
| 563 | int num_ker_prfs = prf_ker ? num_ker_loads : 0; |
| 564 | int num_inp_prfs = prf_inp ? |
| 565 | ur_w * nstl::min(kw, stride_w) + nstl::max(0, kw - stride_w) : |
| 566 | 0; |
| 567 | if (jcp.is_1stconv && prf_inp) { |
| 568 | num_inp_prfs = div_up(num_inp_prfs, jcp.simd_w) * ic_block; |
| 569 | } |
| 570 | int num_prfs = num_ker_prfs + num_inp_prfs; |
| 571 | int num_fmas = num_ker_loads * ur_w; |
| 572 | int prf_inst_spacing |
| 573 | = (prf_ker || prf_inp) ? nstl::max(1, num_fmas / num_prfs) : 1; |
| 574 | int prf_inst_trigger = (num_fmas % prf_inst_spacing) / 2; |
| 575 | int inp_mul = !jcp.is_1stconv ? ic_block : 1; |
| 576 | |
| 577 | if (one_of(jcp.ndims, 3, 4)) { |
| 578 | mov(aux_reg_inp, reg_inp); |
| 579 | mov(aux_reg_ker, reg_ker); |
| 580 | mov(aux_reg_inp_prf, reg_inp_prf); |
| 581 | mov(aux_reg_ker_prf, reg_ker_prf); |
| 582 | } |
| 583 | |
| 584 | size_t max_input_offset = (size_t)jcp.typesize_in * ic_block * iw * ih * id; |
| 585 | assert(reg_inp_prf == reg_long_offt); |
| 586 | if (max_input_offset > INT_MAX) push(reg_inp_prf); |
| 587 | |
| 588 | |
| 589 | if (jcp.ndims == 5) { |
| 590 | push(reg_out_prf); |
| 591 | push(reg_out); |
| 592 | |
| 593 | mov(reg_ki, ptr[param1 + GET_OFF(kd_padding)]); |
| 594 | mov(aux_reg_ker_d, ptr[param1 + GET_OFF(filt)]); |
| 595 | mov(aux_reg_inp_d, reg_inp); |
| 596 | mov(aux_reg_inp_d_prf, reg_inp_prf); |
| 597 | mov(aux_reg_ker_d_prf, reg_ker_prf); |
| 598 | |
| 599 | L(kd_label); |
| 600 | mov(reg_kj, ptr[param1 + GET_OFF(kh_padding)]); |
| 601 | } else { |
| 602 | mov(reg_kj, reg_kh); |
| 603 | } |
| 604 | |
| 605 | if (jcp.ndims == 5) { |
| 606 | mov(aux_reg_inp, aux_reg_inp_d); |
| 607 | mov(aux_reg_ker, aux_reg_ker_d); |
| 608 | mov(aux_reg_ker_prf, aux_reg_ker_d_prf); |
| 609 | mov(aux_reg_inp_prf, aux_reg_inp_d_prf); |
| 610 | } |
| 611 | |
| 612 | align(16); |
| 613 | L(kh_label); |
| 614 | { |
| 615 | int step = 0; |
| 616 | int ker_prfs = 0; |
| 617 | for (int ki = 0; ki < kw; ki++) { |
| 618 | for (int ic = 0; ic < ic_block; ic++) { |
| 619 | int aux_kernel_offset = 0; |
| 620 | if (step == 0) { |
| 621 | for (int i = 0; i < ker_pipeline_depth; i++) { |
| 622 | aux_kernel_offset = get_kernel_offset(ki, ic, 0, i); |
| 623 | vmovups(vmm_ker(i), EVEX_compress_addr( |
| 624 | aux_reg_ker, aux_kernel_offset)); |
| 625 | } |
| 626 | } else if (step < num_ker_loads - ker_pipeline_depth + 1) { |
| 627 | int load_offset = ker_pipeline_depth - 1; |
| 628 | int ker_load_reg_idx |
| 629 | = (step + load_offset) % ker_pipeline_depth; |
| 630 | aux_kernel_offset |
| 631 | = get_kernel_offset(ki, ic, 0, load_offset); |
| 632 | vmovups(vmm_ker(ker_load_reg_idx), |
| 633 | EVEX_compress_addr(aux_reg_ker, aux_kernel_offset)); |
| 634 | } |
| 635 | |
| 636 | bool ker_prf_inserted = false; |
| 637 | Vmm vmm_kernel = vmm_ker(step % ker_pipeline_depth); |
| 638 | int j_start = get_ow_start(ki, pad_l); |
| 639 | int j_end = get_ow_end(ur_w, ki, pad_r); |
| 640 | for (int j = j_start; j < j_end; j++) { |
| 641 | size_t aux_input_offset = get_input_offset(ki, ic, j, pad_l); |
| 642 | auto addr = EVEX_compress_addr_safe(aux_reg_inp, |
| 643 | aux_input_offset, reg_long_offt, true); |
| 644 | vfmadd231ps(vmm_out(j, 0), vmm_kernel, addr); |
| 645 | int fma_idx = step * ur_w + j; |
| 646 | int prf_slot_idx = fma_idx / prf_inst_spacing; |
| 647 | if (fma_idx % prf_inst_spacing == prf_inst_trigger) { |
| 648 | if (prf_ker && !ker_prf_inserted |
| 649 | && ker_prfs < num_ker_prfs) { |
| 650 | int ker_prf_offset |
| 651 | = jcp.typesize_in * ker_prfs * jcp.oc_block; |
| 652 | mic_prefetcht2(EVEX_compress_addr( |
| 653 | aux_reg_ker_prf, ker_prf_offset)); |
| 654 | ker_prf_inserted = true; |
| 655 | ker_prfs++; |
| 656 | } else if (prf_inp) { |
| 657 | int inp_prf_idx = prf_slot_idx - ker_prfs; |
| 658 | if (inp_prf_idx < num_inp_prfs) { |
| 659 | size_t inp_prf_stride = nstl::max(kw, stride_w); |
| 660 | size_t inp_prf_offset; |
| 661 | if (!jcp.is_1stconv) { |
| 662 | inp_prf_offset |
| 663 | = ic_block * jcp.typesize_in |
| 664 | * ((inp_prf_idx / kw) |
| 665 | * inp_prf_stride |
| 666 | + (inp_prf_idx % kw)); |
| 667 | } else { |
| 668 | size_t ic_prf_stride = |
| 669 | (size_t)jcp.typesize_in * iw * ih * id; |
| 670 | size_t iw_prf_stride |
| 671 | = jcp.typesize_in * jcp.simd_w; |
| 672 | inp_prf_offset = ((inp_prf_idx / ic_block) |
| 673 | * iw_prf_stride |
| 674 | + (inp_prf_idx % ic_block) |
| 675 | * ic_prf_stride); |
| 676 | } |
| 677 | mic_prefetcht0(EVEX_compress_addr_safe( |
| 678 | aux_reg_inp_prf, inp_prf_offset, |
| 679 | reg_long_offt)); |
| 680 | } |
| 681 | } |
| 682 | } |
| 683 | } |
| 684 | step++; |
| 685 | } |
| 686 | } |
| 687 | add(aux_reg_ker, jcp.typesize_in * kw * oc_block * ic_block); |
| 688 | if (prf_ker) |
| 689 | add(aux_reg_ker_prf, jcp.typesize_in * kw * oc_block * ic_block); |
| 690 | add(aux_reg_inp, jcp.typesize_in * (jcp.dilate_h + 1) * iw * inp_mul); |
| 691 | if (prf_inp) |
| 692 | add(aux_reg_inp_prf, |
| 693 | jcp.typesize_in * (jcp.dilate_h + 1) * iw * inp_mul); |
| 694 | dec(reg_kj); |
| 695 | cmp(reg_kj, 0); |
| 696 | jg(kh_label, T_NEAR); |
| 697 | } |
| 698 | |
| 699 | |
| 700 | if (jcp.ndims == 5) { |
| 701 | add(aux_reg_inp_d, |
| 702 | typesize * (jcp.dilate_d + 1) * jcp.ih * jcp.iw * inp_mul); |
| 703 | add(aux_reg_ker_d, typesize * jcp.kw * jcp.kh * jcp.oc_block |
| 704 | * jcp.ic_block); |
| 705 | add(aux_reg_inp_d_prf, |
| 706 | typesize * (jcp.dilate_d + 1) * jcp.ih * jcp.iw * inp_mul); |
| 707 | add(aux_reg_ker_d_prf, typesize * jcp.kw * jcp.kh * jcp.oc_block |
| 708 | * jcp.ic_block); |
| 709 | |
| 710 | dec(reg_ki); |
| 711 | cmp(reg_ki, 0); |
| 712 | jg(kd_label, T_NEAR); |
| 713 | |
| 714 | pop(reg_out); |
| 715 | pop(reg_out_prf); |
| 716 | } |
| 717 | if (max_input_offset > INT_MAX) pop(reg_inp_prf); |
| 718 | } |
| 719 | |
| 720 | template<typename Vmm> |
| 721 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::compute_loop_fma_core(int ur_w, |
| 722 | int pad_l, int pad_r) |
| 723 | { |
| 724 | int kw = jcp.kw; |
| 725 | int stride_w = jcp.stride_w; |
| 726 | int ic_block = jcp.ic_block; |
| 727 | int oc_block = jcp.oc_block; |
| 728 | int nb_oc_block = jcp.nb_oc_blocking; |
| 729 | Label kh_label, kd_label; |
| 730 | int shift_kernel_ptr = jcp.typesize_in * jcp.kw * jcp.oc_block |
| 731 | * jcp.ic_block; |
| 732 | int inp_mul = !jcp.is_1stconv ? ic_block : 1; |
| 733 | int shift_input_ptr = jcp.typesize_in * (jcp.dilate_h + 1) * jcp.iw |
| 734 | * inp_mul; |
| 735 | |
| 736 | |
| 737 | auto input_offset = [=](int oi, int ic, int ki) { |
| 738 | return (size_t)jcp.typesize_in |
| 739 | * ((size_t)(ki * (jcp.dilate_w + 1) + oi * stride_w - pad_l) |
| 740 | * inp_mul + (size_t)ic |
| 741 | * (!jcp.is_1stconv ? 1 : (size_t)jcp.iw * jcp.ih * jcp.id)); |
| 742 | }; |
| 743 | |
| 744 | if (one_of(jcp.ndims, 3, 4)) { |
| 745 | mov(aux_reg_inp, reg_inp); |
| 746 | mov(aux_reg_ker, reg_ker); |
| 747 | } |
| 748 | |
| 749 | if (jcp.ndims == 5) { |
| 750 | push(reg_out); |
| 751 | |
| 752 | mov(reg_ki, ptr[param1 + GET_OFF(kd_padding)]); |
| 753 | mov(aux_reg_ker_d, ptr[param1 + GET_OFF(filt)]); |
| 754 | mov(aux_reg_inp_d, reg_inp); |
| 755 | |
| 756 | L(kd_label); |
| 757 | mov(reg_kj, ptr[param1 + GET_OFF(kh_padding)]); |
| 758 | } else { |
| 759 | mov(reg_kj, reg_kh); |
| 760 | } |
| 761 | |
| 762 | if (jcp.ndims == 5) { |
| 763 | mov(aux_reg_inp, aux_reg_inp_d); |
| 764 | mov(aux_reg_ker, aux_reg_ker_d); |
| 765 | } |
| 766 | |
| 767 | L(kh_label); |
| 768 | { |
| 769 | for (int ki = 0; ki < kw; ki++) { |
| 770 | int jj_start = get_ow_start(ki, pad_l); |
| 771 | int jj_end = get_ow_end(ur_w, ki, pad_r); |
| 772 | for (int ic = 0; ic < ic_block; ic++) { |
| 773 | if (jcp.kernel_kind == expl_bcast) { |
| 774 | for (int jj = jj_start; jj < jj_end; jj++) { |
| 775 | size_t aux_input_offset = input_offset(jj, ic, ki); |
| 776 | vbroadcastss(vmm_inp(jj, nb_oc_block), |
| 777 | EVEX_compress_addr_safe(aux_reg_inp, |
| 778 | aux_input_offset, reg_long_offt)); |
| 779 | } |
| 780 | } |
| 781 | for (int ii = 0; ii < nb_oc_block; ii++) { |
| 782 | int aux_kernel_offset = jcp.typesize_in |
| 783 | * (ii * jcp.nb_ic * jcp.kh * jcp.kw * jcp.kd * ic_block |
| 784 | * oc_block + ki * ic_block * oc_block + ic * oc_block); |
| 785 | if (jj_end - jj_start > 0) |
| 786 | vmovups(vmm_wei, EVEX_compress_addr(aux_reg_ker, |
| 787 | aux_kernel_offset)); |
| 788 | for (int jj = jj_start; jj < jj_end; jj++) |
| 789 | if (jcp.kernel_kind == expl_bcast) |
| 790 | vfmadd231ps(vmm_out(jj, ii), |
| 791 | vmm_inp(jj, nb_oc_block), vmm_wei); |
| 792 | else { |
| 793 | size_t aux_input_offset = input_offset(jj, ic, ki); |
| 794 | vfmadd231ps(vmm_out(jj, ii), vmm_wei, |
| 795 | EVEX_compress_addr_safe(aux_reg_inp, |
| 796 | aux_input_offset, reg_long_offt, true)); |
| 797 | } |
| 798 | } |
| 799 | } |
| 800 | } |
| 801 | add(aux_reg_ker, shift_kernel_ptr); |
| 802 | add(aux_reg_inp, shift_input_ptr); |
| 803 | dec(reg_kj); |
| 804 | cmp(reg_kj, 0); |
| 805 | jg(kh_label, T_NEAR); |
| 806 | } |
| 807 | |
| 808 | if (jcp.ndims == 5) { |
| 809 | add(aux_reg_inp_d, |
| 810 | typesize * (jcp.dilate_d + 1) * jcp.ih * jcp.iw * inp_mul); |
| 811 | add(aux_reg_ker_d, typesize * jcp.kw * jcp.kh * jcp.oc_block |
| 812 | * jcp.ic_block); |
| 813 | |
| 814 | dec(reg_ki); |
| 815 | cmp(reg_ki, 0); |
| 816 | jg(kd_label, T_NEAR); |
| 817 | |
| 818 | pop(reg_out); |
| 819 | } |
| 820 | } |
| 821 | |
| 822 | template<typename Vmm> |
| 823 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::compute_loop(int ur_w, |
| 824 | int pad_l, int pad_r) |
| 825 | { |
| 826 | if (jcp.ndims == 5) push(reg_oi); |
| 827 | |
| 828 | prepare_output(ur_w); |
| 829 | |
| 830 | Label skip_compute_loop; |
| 831 | if (jcp.ndims == 5) { |
| 832 | if ((jcp.dilate_d >= jcp.id) |
| 833 | || (jcp.kd - 1) * (jcp.dilate_d + 1) < nstl::max(jcp.f_pad, jcp.back_pad)) { |
| 834 | mov(reg_kj, ptr[param1 + GET_OFF(kd_padding)]); |
| 835 | cmp(reg_kj, 0); |
| 836 | je(skip_compute_loop, T_NEAR); |
| 837 | } |
| 838 | } |
| 839 | if ((jcp.dilate_h >= jcp.ih) |
| 840 | || (jcp.kh - 1) * (jcp.dilate_h + 1) < nstl::max(jcp.t_pad, jcp.b_pad)) { |
| 841 | mov(reg_kj, ptr[param1 + GET_OFF(kh_padding)]); |
| 842 | cmp(reg_kj, 0); |
| 843 | je(skip_compute_loop, T_NEAR); |
| 844 | } |
| 845 | |
| 846 | if (jcp.ver == ver_4fma) |
| 847 | if(jcp.is_1stconv) |
| 848 | compute_loop_4fma_1st(ur_w, pad_l, pad_r); |
| 849 | else |
| 850 | compute_loop_4fma(ur_w, pad_l, pad_r); |
| 851 | else if (jcp.ver == ver_fma) |
| 852 | if ((jcp.is_1stconv && jcp.kernel_kind != expl_bcast) |
| 853 | || mayiuse(avx512_mic)) |
| 854 | compute_loop_fma(ur_w, pad_l, pad_r); |
| 855 | else |
| 856 | if (jcp.kernel_kind == embd_bcast && jcp.nb_oc_blocking == 1) |
| 857 | compute_loop_fma(ur_w, pad_l, pad_r); |
| 858 | else |
| 859 | compute_loop_fma_core(ur_w, pad_l, pad_r); |
| 860 | else |
| 861 | assert(!"unknown convolution version"); |
| 862 | |
| 863 | L(skip_compute_loop); |
| 864 | store_output(ur_w); |
| 865 | if (jcp.ndims == 5) pop(reg_oi); |
| 866 | } |
| 867 | |
| 868 | template<typename Vmm> |
| 869 | void _jit_avx512_common_conv_fwd_kernel<Vmm>::generate() |
| 870 | { |
| 871 | int iw = jcp.iw; |
| 872 | int ow = jcp.ow; |
| 873 | int ow_block = jcp.ow_block; |
| 874 | int nb_ow = jcp.nb_ow; |
| 875 | int kw = jcp.kw; |
| 876 | int l_pad = jcp.l_pad; |
| 877 | int ur_w = jcp.ur_w; |
| 878 | int ur_w_tail = jcp.ur_w_tail; |
| 879 | int dilate_w = jcp.dilate_w + 1; |
| 880 | int stride_w = jcp.stride_w; |
| 881 | |
| 882 | int inp_mult = jcp.is_1stconv ? 1 : jcp.ic_block; |
| 883 | int inp_shift_pad = jcp.typesize_in * (ur_w * stride_w - l_pad) * inp_mult; |
| 884 | int inp_shift = jcp.typesize_in * ur_w * stride_w * inp_mult; |
| 885 | int inp_shift_pad_second_block = -1 * jcp.typesize_in * l_pad * inp_mult; |
| 886 | int out_shift = jcp.typesize_out * ur_w * jcp.oc_block; |
| 887 | |
| 888 | preamble(); |
| 889 | mov(reg_inp, ptr[param1 + GET_OFF(src)]); |
| 890 | mov(reg_out, ptr[param1 + GET_OFF(dst)]); |
| 891 | mov(reg_ker, ptr[param1 + GET_OFF(filt)]); |
| 892 | mov(reg_ker_prf, ptr[param1 + GET_OFF(filt_prf)]); |
| 893 | mov(reg_kh, ptr[param1 + GET_OFF(kh_padding)]); |
| 894 | |
| 895 | int r_pad = nstl::max( |
| 896 | 0, (ow - 1) * stride_w + (kw - 1) * dilate_w - (iw + l_pad - 1)); |
| 897 | int n_oi = ow / ur_w; |
| 898 | int r_pad1 = (ur_w * n_oi - 1) * stride_w + (kw - 1) * dilate_w |
| 899 | - (iw + l_pad - 1); |
| 900 | |
| 901 | if (!is_ow_threading_on(jcp)) { |
| 902 | // ow is being processed as a whole - with left and right paddings |
| 903 | if (r_pad1 > 0) n_oi--; |
| 904 | |
| 905 | if (ow == ur_w) { |
| 906 | mov(reg_inp_prf, ptr[param1 + GET_OFF(src_prf)]); |
| 907 | mov(reg_out_prf, ptr[param1 + GET_OFF(dst_prf)]); |
| 908 | compute_loop(ur_w, l_pad, r_pad); |
| 909 | } else { |
| 910 | mov(reg_inp_prf, reg_inp); |
| 911 | mov(reg_out_prf, reg_out); |
| 912 | if (n_oi == 0) { |
| 913 | add(reg_inp_prf, inp_shift_pad); |
| 914 | add(reg_out_prf, out_shift); |
| 915 | compute_loop(ur_w, l_pad, r_pad1); |
| 916 | add(reg_inp, inp_shift_pad); |
| 917 | add(reg_out, out_shift); |
| 918 | if (ur_w_tail != 0) { |
| 919 | add(reg_inp_prf, inp_shift); |
| 920 | add(reg_out_prf, out_shift); |
| 921 | compute_loop(ur_w_tail, 0, r_pad); |
| 922 | } |
| 923 | } else { |
| 924 | xor_(reg_oi, reg_oi); |
| 925 | if (l_pad > 0) { |
| 926 | add(reg_inp_prf, inp_shift_pad); |
| 927 | add(reg_out_prf, out_shift); |
| 928 | compute_loop(ur_w, l_pad, 0); |
| 929 | add(reg_inp, inp_shift_pad); |
| 930 | add(reg_out, out_shift); |
| 931 | inc(reg_oi); |
| 932 | } |
| 933 | if ((l_pad <= 0 && n_oi > 0) || (l_pad > 0 && n_oi > 1)) { |
| 934 | Label ow_loop_label; |
| 935 | L(ow_loop_label); |
| 936 | { |
| 937 | add(reg_inp_prf, inp_shift); |
| 938 | add(reg_out_prf, out_shift); |
| 939 | compute_loop(ur_w, 0, 0); |
| 940 | add(reg_inp, inp_shift); |
| 941 | add(reg_out, out_shift); |
| 942 | inc(reg_oi); |
| 943 | cmp(reg_oi, n_oi); |
| 944 | jl(ow_loop_label, T_NEAR); |
| 945 | } |
| 946 | } |
| 947 | if (r_pad1 > 0) { |
| 948 | add(reg_inp_prf, inp_shift); |
| 949 | add(reg_out_prf, out_shift); |
| 950 | compute_loop(ur_w, 0, r_pad1); |
| 951 | add(reg_inp, inp_shift); |
| 952 | add(reg_out, out_shift); |
| 953 | } |
| 954 | if (ur_w_tail != 0) { |
| 955 | add(reg_inp_prf, inp_shift); |
| 956 | add(reg_out_prf, out_shift); |
| 957 | compute_loop(ur_w_tail, 0, r_pad); |
| 958 | } |
| 959 | } |
| 960 | } |
| 961 | } else { |
| 962 | // ow block is only processed. |
| 963 | // Number of block is passed as parameter owb, |
| 964 | // and padding processing depends on this number. |
| 965 | |
| 966 | Label end_label, last_oi_label, middle_ow_blocks_label, tail_label; |
| 967 | Label oi_loop_label, oi_loop_start_label, oi_loop_end_label; |
| 968 | |
| 969 | assert(ow_block % ur_w == 0); |
| 970 | int n_oi_not_last_ow_block = ow_block / ur_w; |
| 971 | // to simplify code (and general regs usage), |
| 972 | // size of ow block must be >= 2 * ur_w |
| 973 | assert(n_oi_not_last_ow_block > 1); |
| 974 | int n_oi_next_last_ow_block = n_oi_not_last_ow_block; |
| 975 | int n_oi_first_ow_block = n_oi_not_last_ow_block; |
| 976 | |
| 977 | int n_oi_last_ow_block = (ow - ow_block * (nb_ow-1)) / ur_w; |
| 978 | |
| 979 | // prepare right padding |
| 980 | bool next_last_ow_block_padded = r_pad1 > 0 && n_oi_last_ow_block == 0; |
| 981 | bool first_ow_block_padded = next_last_ow_block_padded && jcp.nb_ow == 2; |
| 982 | bool last_ow_block_padded = r_pad1 > 0 && n_oi_last_ow_block > 0; |
| 983 | |
| 984 | if (last_ow_block_padded) n_oi_last_ow_block--; |
| 985 | else if (first_ow_block_padded) n_oi_first_ow_block--; |
| 986 | else if (next_last_ow_block_padded) n_oi_next_last_ow_block--; |
| 987 | |
| 988 | mov(reg_owb, ptr[param1 + GET_OFF(owb)]); |
| 989 | cmp(reg_owb, 0); // is that the first ow-block ? |
| 990 | jg(middle_ow_blocks_label, T_NEAR); |
| 991 | |
| 992 | // the first ow block, compute left padding |
| 993 | |
| 994 | mov(reg_oi, n_oi_first_ow_block); |
| 995 | mov(reg_inp_prf, reg_inp); |
| 996 | mov(reg_out_prf, reg_out); |
| 997 | |
| 998 | if (l_pad > 0) { |
| 999 | mov(reg_ker_prf, ptr[param1 + GET_OFF(filt_prf)]); |
| 1000 | add(reg_inp_prf, inp_shift_pad); |
| 1001 | add(reg_out_prf, out_shift); |
| 1002 | compute_loop(ur_w, l_pad, 0); |
| 1003 | add(reg_inp, inp_shift_pad); |
| 1004 | add(reg_out, out_shift); |
| 1005 | dec(reg_oi); |
| 1006 | } |
| 1007 | jmp(oi_loop_label, T_NEAR); |
| 1008 | |
| 1009 | // middle or last ow block entry |
| 1010 | |
| 1011 | L(middle_ow_blocks_label); |
| 1012 | |
| 1013 | if (l_pad > 0) { |
| 1014 | // just to consider left padding, not compute |
| 1015 | add(reg_inp, inp_shift_pad_second_block); |
| 1016 | add(reg_inp_prf, inp_shift_pad_second_block); |
| 1017 | } |
| 1018 | |
| 1019 | // set number of iteration for oi-loop |
| 1020 | cmp(reg_owb, jcp.nb_ow - 1); // last ow-block ? |
| 1021 | mov(reg_oi, n_oi_last_ow_block); |
| 1022 | je(oi_loop_label, T_NEAR); |
| 1023 | cmp(reg_owb, jcp.nb_ow - 2); // next to last ow-block ? |
| 1024 | mov(reg_oi, n_oi_next_last_ow_block); |
| 1025 | je(oi_loop_label, T_NEAR); |
| 1026 | mov(reg_oi, n_oi_not_last_ow_block); // other middle ow-blocks |
| 1027 | |
| 1028 | // oi loop w/o padding |
| 1029 | L(oi_loop_label); |
| 1030 | mov(reg_ker_prf, ptr[param1 + GET_OFF(filt_prf)]); |
| 1031 | L(oi_loop_start_label); |
| 1032 | cmp(reg_oi, 0); |
| 1033 | jle(oi_loop_end_label, T_NEAR); |
| 1034 | |
| 1035 | add(reg_inp_prf, inp_shift); |
| 1036 | add(reg_out_prf, out_shift); |
| 1037 | compute_loop(ur_w, 0, 0); |
| 1038 | add(reg_inp, inp_shift); |
| 1039 | add(reg_out, out_shift); |
| 1040 | dec(reg_oi); |
| 1041 | jmp(oi_loop_start_label, T_NEAR); |
| 1042 | L(oi_loop_end_label); |
| 1043 | |
| 1044 | mov(reg_owb, ptr[param1 + GET_OFF(owb)]); |
| 1045 | |
| 1046 | cmp(reg_owb, 0); // first ow-block ? |
| 1047 | if (first_ow_block_padded) { |
| 1048 | je(last_oi_label, T_NEAR); |
| 1049 | } else { |
| 1050 | je(end_label, T_NEAR); |
| 1051 | } |
| 1052 | cmp(reg_owb, jcp.nb_ow - 2); // next to last ow-block ? |
| 1053 | jl(end_label, T_NEAR); |
| 1054 | if (next_last_ow_block_padded) { |
| 1055 | je(last_oi_label, T_NEAR); |
| 1056 | } else { |
| 1057 | je(end_label, T_NEAR); |
| 1058 | } |
| 1059 | // that is last block |
| 1060 | if (!last_ow_block_padded) { |
| 1061 | jmp(tail_label, T_NEAR); |
| 1062 | } |
| 1063 | |
| 1064 | // last oi block with right padding |
| 1065 | L(last_oi_label); |
| 1066 | mov(reg_ker_prf, ptr[param1 + GET_OFF(filt_prf)]); |
| 1067 | add(reg_inp_prf, inp_shift); |
| 1068 | add(reg_out_prf, out_shift); |
| 1069 | compute_loop(ur_w, 0, r_pad1); |
| 1070 | add(reg_inp, inp_shift); |
| 1071 | add(reg_out, out_shift); |
| 1072 | |
| 1073 | mov(reg_owb, ptr[param1 + GET_OFF(owb)]); |
| 1074 | cmp(reg_owb, jcp.nb_ow - 1); // last ow_block? |
| 1075 | jl(end_label, T_NEAR); |
| 1076 | |
| 1077 | L(tail_label); |
| 1078 | mov(reg_ker_prf, ptr[param1 + GET_OFF(filt_prf)]); |
| 1079 | if (ur_w_tail != 0) { |
| 1080 | add(reg_inp_prf, inp_shift); |
| 1081 | add(reg_out_prf, out_shift); |
| 1082 | compute_loop(ur_w_tail, 0, r_pad); |
| 1083 | } |
| 1084 | L(end_label); |
| 1085 | } |
| 1086 | postamble(); |
| 1087 | |
| 1088 | if (jcp.with_eltwise) |
| 1089 | eltwise_injector_->prepare_table(); |
| 1090 | } |
| 1091 | |
| 1092 | bool jit_avx512_common_conv_fwd_kernel::post_ops_ok( |
| 1093 | jit_conv_conf_t &jcp, const primitive_attr_t &attr) { |
| 1094 | const auto &p = attr.post_ops_; |
| 1095 | |
| 1096 | auto is_eltwise = [&](int idx) { return p.entry_[idx].is_eltwise(); }; |
| 1097 | auto is_sum = [&](int idx) { return p.entry_[idx].is_sum(); }; |
| 1098 | |
| 1099 | switch (p.len_) { |
| 1100 | case 0: return true; // no post_ops |
| 1101 | case 1: return is_eltwise(0) || is_sum(0); // sum OR eltwise |
| 1102 | case 2: return is_sum(0) && is_eltwise(1); // sum -> eltwise |
| 1103 | default: return false; |
| 1104 | } |
| 1105 | |
| 1106 | return false; |
| 1107 | } |
| 1108 | |
| 1109 | status_t jit_avx512_common_conv_fwd_kernel::init_conf( |
| 1110 | jit_conv_conf_t &jcp, const convolution_desc_t &cd, |
| 1111 | memory_desc_t &src_md, memory_desc_t &weights_md, |
| 1112 | memory_desc_t &dst_md, memory_desc_t &bias_md, |
| 1113 | const primitive_attr_t &attr, int nthreads) |
| 1114 | { |
| 1115 | using namespace prop_kind; |
| 1116 | |
| 1117 | if (!mayiuse(avx512_common)) |
| 1118 | return status::unimplemented; |
| 1119 | |
| 1120 | const memory_desc_wrapper src_d(&src_md); |
| 1121 | const memory_desc_wrapper weights_d(&weights_md); |
| 1122 | const memory_desc_wrapper dst_d(&dst_md); |
| 1123 | const memory_desc_wrapper bias_d(&bias_md); |
| 1124 | |
| 1125 | const int regs = 28; |
| 1126 | const bool with_groups = weights_d.ndims() == src_d.ndims() + 1; |
| 1127 | int ndims = src_d.ndims(); |
| 1128 | |
| 1129 | jcp = zero<decltype(jcp)>(); |
| 1130 | jcp.ndims = ndims; |
| 1131 | jcp.prop_kind = cd.prop_kind; |
| 1132 | jcp.ngroups = with_groups ? weights_d.dims()[0] : 1; |
| 1133 | jcp.mb = src_d.dims()[0]; |
| 1134 | jcp.oc = dst_d.dims()[1] / jcp.ngroups; |
| 1135 | jcp.oc_without_padding = jcp.oc; |
| 1136 | jcp.ic = src_d.dims()[1] / jcp.ngroups; |
| 1137 | jcp.id = (ndims == 5) ? src_d.dims()[2] : 1; |
| 1138 | jcp.ih = (ndims == 3) ? 1 : src_d.dims()[ndims-2]; |
| 1139 | jcp.iw = src_d.dims()[ndims-1]; |
| 1140 | jcp.od = (ndims == 5) ? dst_d.dims()[2] : 1; |
| 1141 | jcp.oh = (ndims == 3) ? 1 : dst_d.dims()[ndims-2]; |
| 1142 | jcp.ow = dst_d.dims()[ndims-1]; |
| 1143 | jcp.kd = (ndims == 5) ? weights_d.dims()[with_groups + 2] : 1; |
| 1144 | jcp.kh = (ndims == 3) ? 1 : weights_d.dims()[with_groups + ndims-2]; |
| 1145 | jcp.kw = weights_d.dims()[with_groups + ndims-1]; |
| 1146 | jcp.f_pad = (ndims == 5) ? cd.padding[0][0] : 0; |
| 1147 | jcp.t_pad = (ndims == 3) ? 0 : cd.padding[0][ndims-4]; |
| 1148 | jcp.l_pad = cd.padding[0][ndims-3]; |
| 1149 | jcp.stride_d = (ndims == 5) ? cd.strides[0] : 1; |
| 1150 | jcp.stride_h = (ndims == 3) ? 1 : cd.strides[ndims-4]; |
| 1151 | jcp.stride_w = cd.strides[ndims-3]; |
| 1152 | |
| 1153 | jcp.dilate_d = (ndims == 5) ? cd.dilates[0] : 0; |
| 1154 | jcp.dilate_h = (ndims == 3) ? 0 : cd.dilates[ndims-4]; |
| 1155 | jcp.dilate_w = cd.dilates[ndims-3]; |
| 1156 | |
| 1157 | jcp.b_pad = (jcp.oh - 1) * jcp.stride_h + (jcp.kh - 1) * (jcp.dilate_h + 1) |
| 1158 | - (jcp.ih + jcp.t_pad - 1); |
| 1159 | jcp.back_pad = (jcp.od - 1) * jcp.stride_d |
| 1160 | + (jcp.kd - 1) * (jcp.dilate_d + 1) - (jcp.id + jcp.f_pad - 1); |
| 1161 | |
| 1162 | jcp.is_1stconv = is_1stconv(jcp); |
| 1163 | |
| 1164 | bool ok_to_pad_channels = true |
| 1165 | && jcp.ngroups == 1 |
| 1166 | && src_d.data_type() == data_type::f32; |
| 1167 | |
| 1168 | const int full_simd_w = cpu_isa_traits<avx512_common>::vlen / sizeof(float); |
| 1169 | jcp.simd_w = full_simd_w; |
| 1170 | bool ok_to_try_xmm = true |
| 1171 | && mayiuse(avx512_core) |
| 1172 | && src_d.data_type() == data_type::f32 |
| 1173 | && !jcp.is_1stconv |
| 1174 | && !ok_to_pad_channels |
| 1175 | && (jcp.ic % jcp.simd_w != 0 || jcp.oc % jcp.simd_w != 0) |
| 1176 | && (jcp.ic % 8 != 0 || jcp.oc % 8 != 0); |
| 1177 | if (ok_to_try_xmm) |
| 1178 | jcp.simd_w = 4; |
| 1179 | |
| 1180 | jcp.oc_block = jcp.simd_w; |
| 1181 | jcp.ic_block = jcp.is_1stconv ? jcp.ic : jcp.simd_w; |
| 1182 | jcp.aligned_threads = 0; |
| 1183 | |
| 1184 | if (ok_to_pad_channels) { |
| 1185 | jcp.oc = rnd_up(jcp.oc, jcp.oc_block); |
| 1186 | jcp.ic = rnd_up(jcp.ic, jcp.ic_block); |
| 1187 | } |
| 1188 | bool args_ok = true |
| 1189 | && jcp.oc % jcp.oc_block == 0 |
| 1190 | && jcp.ic % jcp.ic_block == 0; |
| 1191 | if (!args_ok) |
| 1192 | return status::unimplemented; |
| 1193 | |
| 1194 | if (!post_ops_ok(jcp, attr)) |
| 1195 | return status::unimplemented; |
| 1196 | |
| 1197 | const auto &p = attr.post_ops_; |
| 1198 | jcp.with_sum = p.find(primitive_kind::sum) != -1; |
| 1199 | const int eltwise_ind = p.find(primitive_kind::eltwise); |
| 1200 | jcp.with_eltwise = eltwise_ind != -1; |
| 1201 | if (jcp.with_eltwise) { |
| 1202 | jcp.eltwise = p.entry_[eltwise_ind].eltwise; |
| 1203 | if (dst_d.data_type() == data_type::s32) return status::unimplemented; |
| 1204 | } |
| 1205 | |
| 1206 | auto src_tag = jcp.is_1stconv |
| 1207 | ? pick(ndims - 3, ncw, nchw, ncdhw) |
| 1208 | : ((jcp.simd_w == 4) |
| 1209 | ? pick(ndims - 3, nCw4c, nChw4c, nCdhw4c) |
| 1210 | : pick(ndims - 3, nCw16c, nChw16c, nCdhw16c)); |
| 1211 | auto dst_tag = (jcp.simd_w == 4) |
| 1212 | ? pick(ndims - 3, nCw4c, nChw4c, nCdhw4c) |
| 1213 | : pick(ndims - 3, nCw16c, nChw16c, nCdhw16c); |
| 1214 | auto wei_tag = with_groups |
| 1215 | ? ((jcp.simd_w == 4) |
| 1216 | ? pick(ndims - 3, gOIw4i4o, gOIhw4i4o, gOIdhw4i4o) |
| 1217 | : pick(ndims - 3, gOIw16i16o, gOIhw16i16o, gOIdhw16i16o)) |
| 1218 | : ((jcp.simd_w == 4) |
| 1219 | ? pick(ndims - 3, OIw4i4o, OIhw4i4o, OIdhw4i4o) |
| 1220 | : pick(ndims - 3, OIw16i16o, OIhw16i16o, OIdhw16i16o)); |
| 1221 | |
| 1222 | if (src_d.format_kind() == format_kind::any) { |
| 1223 | CHECK(memory_desc_init_by_tag(src_md, src_tag)); |
| 1224 | jcp.src_tag = src_tag; |
| 1225 | } else { |
| 1226 | jcp.src_tag = src_d.matches_one_of_tag(src_tag); |
| 1227 | } |
| 1228 | if (jcp.src_tag != src_tag) |
| 1229 | return status::unimplemented; |
| 1230 | |
| 1231 | if (dst_d.format_kind() == format_kind::any) { |
| 1232 | CHECK(memory_desc_init_by_tag(dst_md, dst_tag)); |
| 1233 | jcp.dst_tag = dst_tag; |
| 1234 | } else { |
| 1235 | jcp.dst_tag = dst_d.matches_one_of_tag(dst_tag); |
| 1236 | } |
| 1237 | if (jcp.dst_tag != dst_tag) |
| 1238 | return status::unimplemented; |
| 1239 | |
| 1240 | jcp.with_bias = cd.bias_desc.format_kind != format_kind::undef; |
| 1241 | if (jcp.with_bias) { |
| 1242 | if (bias_d.format_kind() == format_kind::any) |
| 1243 | CHECK(memory_desc_init_by_tag(bias_md, x)); |
| 1244 | } |
| 1245 | |
| 1246 | if (mayiuse(avx512_common) && |
| 1247 | src_d.data_type() == data_type::f32 |
| 1248 | && weights_d.data_type() == data_type::f32 |
| 1249 | && dst_d.data_type() == data_type::f32) { |
| 1250 | jcp.ver = ver_fma; |
| 1251 | jcp.typesize_in = sizeof(float); |
| 1252 | jcp.typesize_out = sizeof(float); |
| 1253 | if (mayiuse(avx512_mic_4ops)) |
| 1254 | jcp.ver = ver_4fma; |
| 1255 | |
| 1256 | if (jcp.is_1stconv) { |
| 1257 | // TODO: fix & remove constraints below |
| 1258 | bool not_for_4fma |
| 1259 | = IMPLICATION(everyone_is(0, jcp.l_pad, jcp.t_pad), |
| 1260 | nstl::max(jcp.kw, jcp.kh) < 7); |
| 1261 | bool is_dilated |
| 1262 | = !everyone_is(0, jcp.dilate_d, jcp.dilate_h, jcp.dilate_w); |
| 1263 | if (one_of(true, not_for_4fma, is_dilated)) |
| 1264 | jcp.ver = ver_fma; |
| 1265 | if (jcp.ver == ver_4fma) { |
| 1266 | wei_tag = with_groups |
| 1267 | ? ((jcp.simd_w == 4) |
| 1268 | ? pick(ndims - 3, gOiw4o, gOihw4o, gOidhw4o) |
| 1269 | : pick(ndims - 3, gOiw16o, gOihw16o, gOidhw16o)) |
| 1270 | : ((jcp.simd_w == 4) |
| 1271 | ? pick(ndims - 3, Oiw4o, Oihw4o, Oidhw4o) |
| 1272 | : pick(ndims - 3, Oiw16o, Oihw16o, Oidhw16o)); |
| 1273 | } else { |
| 1274 | wei_tag = with_groups |
| 1275 | ? ((jcp.simd_w == 4) |
| 1276 | ? pick(ndims - 3, gOwi4o, gOhwi4o, gOdhwi4o) |
| 1277 | : pick(ndims - 3, gOwi16o, gOhwi16o, gOdhwi16o)) |
| 1278 | : ((jcp.simd_w == 4) |
| 1279 | ? pick(ndims - 3, Owi4o, Ohwi4o, Odhwi4o) |
| 1280 | : pick(ndims - 3, Owi16o, Ohwi16o, Odhwi16o)); |
| 1281 | } |
| 1282 | } |
| 1283 | } else { |
| 1284 | return status::unimplemented; |
| 1285 | } |
| 1286 | |
| 1287 | if (weights_d.format_kind() == format_kind::any) { |
| 1288 | CHECK(memory_desc_init_by_tag(weights_md, wei_tag)); |
| 1289 | jcp.wei_tag = wei_tag; |
| 1290 | } else { |
| 1291 | jcp.wei_tag = weights_d.matches_one_of_tag(wei_tag); |
| 1292 | } |
| 1293 | if (jcp.wei_tag != wei_tag) |
| 1294 | return status::unimplemented; |
| 1295 | |
| 1296 | if (jcp.is_1stconv) { |
| 1297 | jcp.ur_w = nstl::min(jcp.ow, regs); |
| 1298 | } else { |
| 1299 | // avx512_core guard - just to avoid possible regression for other archs |
| 1300 | if (jcp.ver == ver_fma && mayiuse(avx512_core)) { |
| 1301 | jcp.ur_w = nstl::min(jcp.ow, regs); |
| 1302 | } else { |
| 1303 | for (int ur_w = regs; ur_w > 0; --ur_w) { |
| 1304 | if (jcp.ow % ur_w == 0) { |
| 1305 | jcp.ur_w = ur_w; |
| 1306 | break; |
| 1307 | } |
| 1308 | } |
| 1309 | } |
| 1310 | if ((ndims == 5 && jcp.ur_w <= 8) || (jcp.ur_w <= 1)) { |
| 1311 | jcp.ur_w = nstl::min(jcp.ow, regs); |
| 1312 | } |
| 1313 | } |
| 1314 | // TODO (Tanya): currently applied to Segnet convolutions only. |
| 1315 | // Need to try for other topologies |
| 1316 | if (jcp.ow > 150 && jcp.ur_w < regs/2) |
| 1317 | jcp.ur_w = regs; |
| 1318 | |
| 1319 | int n_oi = (jcp.ow / jcp.ur_w); |
| 1320 | int r_pad = (jcp.ur_w * n_oi - 1) * jcp.stride_w |
| 1321 | + (jcp.kw - 1) * (jcp.dilate_w + 1) - (jcp.iw + jcp.l_pad - 1); |
| 1322 | if (jcp.l_pad > 0 && r_pad > 0) |
| 1323 | n_oi--; |
| 1324 | |
| 1325 | bool large_code_size = jcp.ur_w != jcp.ow && jcp.l_pad > 0 && r_pad > 0 |
| 1326 | && ((jcp.l_pad <= 0 && n_oi > 0) || (jcp.l_pad > 0 && n_oi > 1)); |
| 1327 | if (large_code_size) { |
| 1328 | const int max_code_size = 24 * 1024; |
| 1329 | const int num_ops_per_reg = 6 + jcp.ic_block * jcp.kw; |
| 1330 | int mult = 1; |
| 1331 | if (jcp.l_pad > 0) mult += 1; |
| 1332 | if (r_pad > 0) mult += 1; |
| 1333 | for (int ur_w = jcp.ur_w; ur_w > regs/2; --ur_w) { |
| 1334 | if (ur_w * mult * num_ops_per_reg * 9.0 < max_code_size) { |
| 1335 | jcp.ur_w = ur_w; |
| 1336 | break; |
| 1337 | } |
| 1338 | } |
| 1339 | } |
| 1340 | |
| 1341 | /* Grouped channel offset to support 'non-blocked data' format for |
| 1342 | * convolution sizes with '(input_channel / ngroups) < simd' */ |
| 1343 | jcp.nonblk_group_off |
| 1344 | = (jcp.ngroups > 1 && one_of(jcp.src_tag, ncw, nchw, ncdhw)) ? |
| 1345 | jcp.ic : |
| 1346 | 1; |
| 1347 | |
| 1348 | jcp.nb_ic = jcp.ic / jcp.ic_block; |
| 1349 | jcp.nb_oc = jcp.oc / jcp.oc_block; |
| 1350 | jcp.nb_ic_blocking = jcp.nb_oc_blocking = 1; |
| 1351 | |
| 1352 | auto is_ow_threading_applicable = [=]() { |
| 1353 | return (true && !jcp.is_1stconv && one_of(jcp.ndims, 3, 4) |
| 1354 | && IMPLICATION(mayiuse(avx512_mic), |
| 1355 | jcp.ver == ver_4fma |
| 1356 | && IMPLICATION(jcp.mb != 1, |
| 1357 | jcp.ih == 1 && jcp.kh == 1))); |
| 1358 | }; |
| 1359 | |
| 1360 | if (jcp.ver == ver_4fma && !jcp.is_1stconv) { |
| 1361 | if ((jcp.kw <= 5 && jcp.kh <= 5 && jcp.kw == jcp.kh && jcp.ow <= 8 |
| 1362 | && jcp.oh <= 8 && jcp.ow == jcp.oh) |
| 1363 | || (jcp.stride_h != 1 && jcp.ur_w < jcp.ow)) { |
| 1364 | if (jcp.nb_oc % 2 == 0) { |
| 1365 | jcp.nb_oc_blocking = 2; |
| 1366 | jcp.ur_w = nstl::min(jcp.ow, regs / jcp.nb_oc_blocking); |
| 1367 | } |
| 1368 | } else { |
| 1369 | for (int i = jcp.nb_oc; i > 0; i--) |
| 1370 | if (i * jcp.ur_w <= regs && jcp.nb_oc % i == 0) { |
| 1371 | jcp.nb_oc_blocking = i; |
| 1372 | break; |
| 1373 | } |
| 1374 | } |
| 1375 | if (jcp.ver == ver_4fma && is_ow_threading_applicable()) { |
| 1376 | if (jcp.nb_oc % 2 == 0 && jcp.ur_w < jcp.ow |
| 1377 | && jcp.ow != 2 * jcp.ur_w) { |
| 1378 | jcp.nb_oc_blocking = 2; |
| 1379 | jcp.ur_w = nstl::min(jcp.ow, regs / jcp.nb_oc_blocking); |
| 1380 | } |
| 1381 | } |
| 1382 | } |
| 1383 | |
| 1384 | jcp.ow_block = jcp.ow; |
| 1385 | |
| 1386 | auto get_thr_eff = [=](int nb_oc_blocking, int ow_block) { |
| 1387 | int nb_ow = div_up(jcp.ow, ow_block); |
| 1388 | int nb_oc_chunks = div_up(jcp.nb_oc, nb_oc_blocking); |
| 1389 | int work_amount = jcp.mb * jcp.oh * nb_oc_chunks * nb_ow; |
| 1390 | float disbalance = (float)jcp.ow / rnd_up(jcp.ow, ow_block); |
| 1391 | float thr_eff = disbalance * (float)work_amount |
| 1392 | / rnd_up(work_amount, nthreads); |
| 1393 | return thr_eff; |
| 1394 | }; |
| 1395 | |
| 1396 | auto get_ow_block = [=](int nb_oc_blocking, int ur_w, float &eff) { |
| 1397 | int res_ow_block = jcp.ow; |
| 1398 | eff = get_thr_eff(nb_oc_blocking, res_ow_block); |
| 1399 | if (!is_ow_threading_applicable()) |
| 1400 | return res_ow_block; |
| 1401 | |
| 1402 | int L2_part = (get_cache_size(2) * 7 / 8) / typesize; |
| 1403 | if (jcp.ver == ver_4fma) |
| 1404 | L2_part /= 2; |
| 1405 | int size_src_chunk = jcp.ic_block * ur_w * jcp.kh; |
| 1406 | int size_dst_chunk = jcp.oc_block * nb_oc_blocking * ur_w; |
| 1407 | int size_wei_chunk = jcp.oc_block * nb_oc_blocking * jcp.ic_block |
| 1408 | * jcp.kw * jcp.kh; |
| 1409 | int nurw_cache = (L2_part - 2 * size_wei_chunk) |
| 1410 | / (2 * size_dst_chunk + 2 * size_src_chunk); |
| 1411 | // current design of generate() requires ow_block >= 2 * ur_w |
| 1412 | int ow_block_cache = ur_w * nstl::max(2, nurw_cache); |
| 1413 | |
| 1414 | int ow_block_thr = ow_block_cache; |
| 1415 | eff = get_thr_eff(nb_oc_blocking, ow_block_thr); |
| 1416 | |
| 1417 | int max_nb_ow = div_up(jcp.ow, 2 * ur_w); |
| 1418 | int start_nb_ow = div_up(jcp.ow, ow_block_thr); |
| 1419 | for (int nb_ow = start_nb_ow; nb_ow <= max_nb_ow; nb_ow++) { |
| 1420 | int ow_block |
| 1421 | = nstl::min(rnd_up(div_up(jcp.ow, nb_ow), ur_w), jcp.ow); |
| 1422 | float eff_threshold = (jcp.ver == ver_4fma) ? 0.8f : 0.9f; |
| 1423 | if (ow_block < nb_oc_blocking * jcp.oc_block && eff > eff_threshold) |
| 1424 | break; |
| 1425 | if (div_up(jcp.ow, ow_block) != nb_ow) |
| 1426 | continue; |
| 1427 | float thr_eff = get_thr_eff(nb_oc_blocking, ow_block); |
| 1428 | float eff_step = (jcp.ver == ver_4fma) ? 1.1f : 1.f; |
| 1429 | if (ow_block >= 2 * ur_w && thr_eff > eff_step * eff) { |
| 1430 | ow_block_thr = ow_block; |
| 1431 | eff = thr_eff; |
| 1432 | } |
| 1433 | eff_threshold = (jcp.ver == ver_4fma) ? 0.9f : 0.98f; |
| 1434 | if (eff > eff_threshold) |
| 1435 | break; |
| 1436 | } |
| 1437 | res_ow_block = nstl::min(jcp.ow, nstl::max(2 * ur_w, ow_block_thr)); |
| 1438 | eff = get_thr_eff(nb_oc_blocking, res_ow_block); |
| 1439 | return res_ow_block; |
| 1440 | }; |
| 1441 | |
| 1442 | |
| 1443 | if (jcp.ver == ver_fma && mayiuse(avx512_core)) { |
| 1444 | int try_nb_oc_blocking = 2; |
| 1445 | unsigned int ker_inp_size = typesize * div_up(jcp.iw, jcp.stride_w) |
| 1446 | * jcp.ic_block * jcp.kh * jcp.kd; |
| 1447 | unsigned int ker_out_size = typesize * jcp.ow * jcp.oc_block |
| 1448 | * try_nb_oc_blocking; |
| 1449 | unsigned int ker_wei_size = typesize * jcp.kh * jcp.kw * jcp.ic_block |
| 1450 | * jcp.oc_block * try_nb_oc_blocking * jcp.kd; |
| 1451 | unsigned int ker_total_size = ker_inp_size + ker_out_size |
| 1452 | + ker_wei_size; |
| 1453 | |
| 1454 | bool embd_bcast_condition = true |
| 1455 | && (jcp.kw == 3 && jcp.ow <= 28 && ker_total_size < L1_cache_size) |
| 1456 | && !(jcp.kw == 3 && jcp.ow == 13 && jcp.ic >= 192) |
| 1457 | && !(jcp.kw == 3 && jcp.ow == 28 && jcp.ic >= 512); |
| 1458 | |
| 1459 | if (jcp.mb == 1) { |
| 1460 | unsigned int inp_size = jcp.mb * div_up(jcp.ih, jcp.stride_h) |
| 1461 | * div_up(jcp.iw, jcp.stride_w) * jcp.ic; |
| 1462 | unsigned int wei_size = jcp.ic * jcp.oc * jcp.kh * jcp.kw; |
| 1463 | |
| 1464 | // Estimate whether we need to limit the number of threads |
| 1465 | // and calculate this number. Includes some heuristic. |
| 1466 | int oc_chunks = jcp.nb_oc / jcp.nb_oc_blocking; |
| 1467 | int work_amount = jcp.mb * jcp.ngroups * oc_chunks * jcp.oh; |
| 1468 | int job_size_min = work_amount / nthreads; |
| 1469 | int job_size_max = div_up(work_amount, nthreads); |
| 1470 | int ch_max = rnd_up(jcp.oh, job_size_max); |
| 1471 | int ch_min = (job_size_min == 0) |
| 1472 | ? jcp.oh |
| 1473 | : rnd_up(jcp.oh, job_size_min); |
| 1474 | bool not_aligned_max = ch_max % jcp.oh != 0 && ch_max / jcp.oh < 2 |
| 1475 | && (jcp.oh != 8 || ch_max / jcp.oh > 1); |
| 1476 | bool not_aligned_min = ch_min % jcp.oh != 0 && ch_min / jcp.oh < 2 |
| 1477 | && (jcp.oh != 8 || ch_min / jcp.oh > 1); |
| 1478 | bool eligible_case = (jcp.stride_h == 1 && jcp.stride_w == 1) |
| 1479 | || nthreads > oc_chunks; |
| 1480 | if (jcp.loop_order == loop_cgn && oc_chunks > 1 && nthreads > 1 |
| 1481 | && wei_size / inp_size > 24 |
| 1482 | && (not_aligned_max || not_aligned_min) |
| 1483 | && eligible_case) { |
| 1484 | // Try to find nthreads > mkldnn_get_max_threads() / 2 such |
| 1485 | // that oc_chunks is a multiple of nthreads, or nthreads is a |
| 1486 | // multiple of oc_chunks. Otherwise, keep default value. |
| 1487 | // TODO: implement a task-based alternative without throttling. |
| 1488 | jcp.aligned_threads = nthreads; |
| 1489 | for (int i = nthreads; i > nthreads / 2; i--) { |
| 1490 | if (oc_chunks % i == 0 || i % oc_chunks == 0) { |
| 1491 | jcp.aligned_threads = i; |
| 1492 | break; |
| 1493 | } |
| 1494 | } |
| 1495 | } |
| 1496 | } |
| 1497 | |
| 1498 | if (jcp.kw > 3 |
| 1499 | || (jcp.stride_w == 1 && jcp.stride_h == 1 |
| 1500 | && embd_bcast_condition) |
| 1501 | || ((jcp.stride_w != 1 || jcp.stride_h != 1) |
| 1502 | && ((jcp.mb <= 16 && (jcp.oc <= 192 || jcp.oh <= 10) |
| 1503 | && embd_bcast_condition))) |
| 1504 | || (jcp.mb == 1 |
| 1505 | && (jcp.ur_w >= jcp.ow || jcp.is_1stconv |
| 1506 | || (jcp.ow <= 147 && jcp.oc <= 96)))) { |
| 1507 | jcp.kernel_kind = embd_bcast; |
| 1508 | jcp.ur_w = nstl::min(jcp.ow, regs); |
| 1509 | jcp.nb_ic_blocking = jcp.nb_oc_blocking = 1; |
| 1510 | if (ker_total_size < L1_cache_size && jcp.ow <= 8 && jcp.kh <= 3 |
| 1511 | && jcp.kw <= 3 && jcp.nb_oc % try_nb_oc_blocking == 0 |
| 1512 | && IMPLICATION(jcp.is_1stconv, jcp.mb == 1) |
| 1513 | && IMPLICATION(jcp.mb == 1, jcp.ur_w < jcp.ow)) { |
| 1514 | jcp.nb_oc_blocking = try_nb_oc_blocking; |
| 1515 | jcp.ur_w = nstl::min(jcp.ow, 31 / (jcp.nb_oc_blocking + 1)); |
| 1516 | } |
| 1517 | } else { |
| 1518 | jcp.kernel_kind = expl_bcast; |
| 1519 | jcp.nb_ic_blocking = 1; |
| 1520 | if (IMPLICATION(jcp.is_1stconv, jcp.mb > 1)) { |
| 1521 | float best_thr_eff = 0.f; |
| 1522 | int best_nb_oc_blocking = 1; |
| 1523 | for (int i = nstl::min(jcp.nb_oc, 5); i > 0; i--) { |
| 1524 | if (jcp.nb_oc % i == 0) { |
| 1525 | float thr_eff; |
| 1526 | int ur_w = nstl::min(jcp.ow, 31 / (i + 1)); |
| 1527 | get_ow_block(i, ur_w, thr_eff); |
| 1528 | if (thr_eff > 1.05f * best_thr_eff) { |
| 1529 | best_nb_oc_blocking = i; |
| 1530 | best_thr_eff = thr_eff; |
| 1531 | } |
| 1532 | } |
| 1533 | } |
| 1534 | jcp.nb_oc_blocking = best_nb_oc_blocking; |
| 1535 | jcp.ur_w = nstl::min(jcp.ow, 31 / (jcp.nb_oc_blocking + 1)); |
| 1536 | } |
| 1537 | } |
| 1538 | } |
| 1539 | |
| 1540 | jcp.ur_w_tail = jcp.ow % jcp.ur_w; |
| 1541 | |
| 1542 | args_ok = true |
| 1543 | && jcp.l_pad <= jcp.ur_w |
| 1544 | && jcp.ic <= src_d.padded_dims()[1] |
| 1545 | && jcp.oc <= dst_d.padded_dims()[1] |
| 1546 | && jcp.ic <= weights_d.padded_dims()[with_groups + 1] |
| 1547 | && jcp.oc <= weights_d.padded_dims()[with_groups + 0]; |
| 1548 | if (!args_ok) |
| 1549 | return status::unimplemented; |
| 1550 | |
| 1551 | int r_pad_no_tail = nstl::max(0, (jcp.ow - jcp.ur_w_tail - 1) * jcp.stride_w |
| 1552 | + (jcp.kw - 1) * (jcp.dilate_w + 1) |
| 1553 | - (jcp.iw + jcp.l_pad - 1)); |
| 1554 | if (r_pad_no_tail > jcp.ur_w) |
| 1555 | return status::unimplemented; |
| 1556 | |
| 1557 | pick_loop_order(jcp); |
| 1558 | |
| 1559 | jcp.nb_ic_L2 = jcp.nb_ic; |
| 1560 | |
| 1561 | float thr_eff; |
| 1562 | jcp.ow_block = get_ow_block(jcp.nb_oc_blocking, jcp.ur_w, thr_eff); |
| 1563 | jcp.nb_ow = div_up(jcp.ow, jcp.ow_block); |
| 1564 | |
| 1565 | const int L2_size = get_cache_size(2, true) / sizeof(float); |
| 1566 | // Source and output data needs to fit in L2, |
| 1567 | // leaving some space for weights and prefetching. |
| 1568 | int h_L2 = int(((0.6f * L2_size) / jcp.simd_w |
| 1569 | - nstl::min(0, jcp.kh - jcp.stride_h) * jcp.iw) |
| 1570 | / (jcp.stride_h * jcp.iw + jcp.ow)); |
| 1571 | jcp.h_blocking = nstl::max(1, nstl::min(jcp.oh, h_L2)); |
| 1572 | |
| 1573 | if (jcp.ver == ver_4fma) { |
| 1574 | if (!is_ow_threading_on(jcp)) { |
| 1575 | for (int divf = 2, temp_nb = jcp.nb_ic_L2; divf <= jcp.nb_ic; |
| 1576 | divf++) { |
| 1577 | size_t l2_src |
| 1578 | = (size_t)jcp.iw * jcp.ic_block * jcp.ih * temp_nb * jcp.id; |
| 1579 | size_t l2_dst = (size_t)jcp.ow * jcp.oc_block * jcp.nb_oc_blocking |
| 1580 | * jcp.oh * jcp.od; |
| 1581 | size_t l2_filt = (size_t)jcp.kw * jcp.oc_block * jcp.ic_block |
| 1582 | * jcp.kh * jcp.nb_oc_blocking * temp_nb * jcp.kd; |
| 1583 | if (4 * (l2_src + l2_dst + l2_filt) > KNx_L2_EFFECTIVE_CAPACITY) { |
| 1584 | if (jcp.kh == 3 && jcp.oh == 7) { |
| 1585 | jcp.nb_ic_L2 = 1; |
| 1586 | break; |
| 1587 | } |
| 1588 | temp_nb = (jcp.nb_ic_L2 % divf == 0 ? jcp.nb_ic_L2 / divf |
| 1589 | : jcp.nb_ic_L2); |
| 1590 | } else { |
| 1591 | jcp.nb_ic_L2 = temp_nb; |
| 1592 | break; |
| 1593 | } |
| 1594 | } |
| 1595 | } else if (jcp.ic > 64) { |
| 1596 | jcp.nb_ic_L2 = 2; /* according to performance data*/ |
| 1597 | } |
| 1598 | } |
| 1599 | |
| 1600 | return status::success; |
| 1601 | } |
| 1602 | |
| 1603 | void jit_avx512_common_conv_fwd_kernel::init_scratchpad( |
| 1604 | memory_tracking::registrar_t &scratchpad, const jit_conv_conf_t &jcp) { |
| 1605 | if (jcp.with_bias && jcp.oc != jcp.oc_without_padding) |
| 1606 | scratchpad.book(key_conv_padded_bias, jcp.typesize_out * jcp.oc); |
| 1607 | } |
| 1608 | |
| 1609 | void jit_avx512_common_conv_bwd_data_kernel_f32::prepare_output(int ur_w) |
| 1610 | { |
| 1611 | for (int k = 0; k < jcp.nb_ic_blocking; k++) { |
| 1612 | for (int j = 0; j < ur_w; j++) { |
| 1613 | Zmm zmm = zmm_out(j, k); |
| 1614 | vpxord(zmm, zmm, zmm); |
| 1615 | size_t aux_src_offset |
| 1616 | = (size_t)typesize * ((size_t)k * jcp.ih * jcp.iw * jcp.id + j) |
| 1617 | * jcp.ic_block; |
| 1618 | mic_prefetcht1(EVEX_compress_addr_safe(reg_src_prf, aux_src_offset, |
| 1619 | reg_long_offt)); |
| 1620 | } |
| 1621 | } |
| 1622 | } |
| 1623 | |
| 1624 | void jit_avx512_common_conv_bwd_data_kernel_f32::store_output(int ur_w) |
| 1625 | { |
| 1626 | Label no_update_label; |
| 1627 | |
| 1628 | mov(reg_channel, ptr[param + GET_OFF(channel)]); |
| 1629 | cmp(reg_channel, 0); |
| 1630 | je(no_update_label, T_NEAR); |
| 1631 | for (int k = 0; k < jcp.nb_ic_blocking; k++) { |
| 1632 | for (int j = 0; j < ur_w; j++) { |
| 1633 | Zmm zmm = zmm_out(j, k); |
| 1634 | size_t aux_src_offset = (size_t)typesize |
| 1635 | * ((size_t)k * jcp.ih * jcp.iw * jcp.id + j) * jcp.ic_block; |
| 1636 | vaddps(zmm, EVEX_compress_addr_safe(reg_src, aux_src_offset, |
| 1637 | reg_long_offt)); |
| 1638 | } |
| 1639 | } |
| 1640 | |
| 1641 | L(no_update_label); |
| 1642 | for (int k = 0; k < jcp.nb_ic_blocking; k++) { |
| 1643 | for (int j = 0; j < ur_w; j++) { |
| 1644 | Zmm zmm = zmm_out(j, k); |
| 1645 | size_t aux_src_offset = (size_t)typesize |
| 1646 | * ((size_t)k * jcp.ih * jcp.iw * jcp.id + j) * jcp.ic_block; |
| 1647 | vmovups(EVEX_compress_addr_safe(reg_src, aux_src_offset, |
| 1648 | reg_long_offt), zmm); |
| 1649 | mic_prefetcht0(EVEX_compress_addr_safe(reg_src_prf, aux_src_offset, |
| 1650 | reg_long_offt)); |
| 1651 | } |
| 1652 | } |
| 1653 | } |
| 1654 | |
| 1655 | void jit_avx512_common_conv_bwd_data_kernel_f32::compute_loop_4fma( |
| 1656 | int ur_w, int l_overflow, int r_overflow) |
| 1657 | { |
| 1658 | int ow = jcp.ow; |
| 1659 | int kw = jcp.kw; |
| 1660 | int ic_block = jcp.ic_block; |
| 1661 | int oc_block = jcp.oc_block; |
| 1662 | Label kh_label, last_iter_label, loop_end_label, kd_label; |
| 1663 | int ker_load_number = 4; |
| 1664 | int shift_ker_ptr = typesize * kw * oc_block * ic_block; |
| 1665 | int shift_dst_ptr = typesize * ow * oc_block; |
| 1666 | int ii_dpref_t0 = get_iw_start(0, l_overflow); |
| 1667 | int iw_end_ipref = get_iw_end(ur_w, 0, r_overflow); |
| 1668 | |
| 1669 | bool check_last_kh = (jcp.kh > 3); |
| 1670 | auto kernel_offset = [=](int icb, int oc, int ki) { |
| 1671 | int blk_idx = icb * jcp.kh * jcp.kw * jcp.kd + ki; |
| 1672 | int blk_offset = blk_idx * jcp.oc_block * jcp.ic_block; |
| 1673 | int oc_offset = oc * jcp.oc_block; |
| 1674 | return typesize * (blk_offset + oc_offset); |
| 1675 | }; |
| 1676 | auto kernel_loads = [=](int ki, int oc, int kk) { |
| 1677 | for (int ii = 0; ii < ker_load_number; ii++) { |
| 1678 | int aux_kernel_offset = kernel_offset(kk, oc + ii, ki); |
| 1679 | vmovups(zmm_ker(ii), |
| 1680 | EVEX_compress_addr(aux_reg_ker, aux_kernel_offset)); |
| 1681 | } |
| 1682 | }; |
| 1683 | auto prefetch_dst_next_kh = [&](int ki, int ki_start, int cnt0, int cnt1) { |
| 1684 | if (cnt1 >= ker_load_number && cnt0 >= ker_load_number |
| 1685 | && ki >= ki_start && ii_dpref_t0 < iw_end_ipref) { |
| 1686 | int aux_dst_offset = typesize * ((ii_dpref_t0 |
| 1687 | + jcp.l_pad) * oc_block + jcp.ow * oc_block); |
| 1688 | prefetcht0(EVEX_compress_addr(aux_reg_dst, aux_dst_offset)); |
| 1689 | ii_dpref_t0++; |
| 1690 | } |
| 1691 | }; |
| 1692 | |
| 1693 | if (one_of(jcp.ndims, 3, 4)) { |
| 1694 | mov(aux_reg_dst, reg_dst); |
| 1695 | mov(aux_reg_ker, reg_ker); |
| 1696 | mov(aux_reg_dst_prf, reg_dst_prf); |
| 1697 | mov(aux_reg_ker_prf, reg_ker_prf); |
| 1698 | } |
| 1699 | |
| 1700 | if (jcp.ndims == 5) { |
| 1701 | push(reg_src_prf); |
| 1702 | push(reg_src); |
| 1703 | |
| 1704 | mov(reg_ki, ptr[param + GET_OFF(kd_padding)]); |
| 1705 | mov(aux_reg_dst_d, reg_dst); |
| 1706 | mov(aux_reg_ker_d, ptr[param + GET_OFF(filt)]); |
| 1707 | mov(aux_reg_dst_d_prf, reg_dst_prf); |
| 1708 | mov(aux_reg_ker_d_prf, reg_ker_prf); |
| 1709 | |
| 1710 | L(kd_label); |
| 1711 | mov(reg_kj, ptr[param + GET_OFF(kh_padding)]); |
| 1712 | } else { |
| 1713 | mov(reg_kj, reg_kh); |
| 1714 | } |
| 1715 | |
| 1716 | if (jcp.ndims == 5) { |
| 1717 | mov(aux_reg_dst, aux_reg_dst_d); |
| 1718 | mov(aux_reg_ker, aux_reg_ker_d); |
| 1719 | mov(aux_reg_dst_prf, aux_reg_dst_d_prf); |
| 1720 | mov(aux_reg_ker_prf, aux_reg_ker_d_prf); |
| 1721 | } |
| 1722 | |
| 1723 | align(16); |
| 1724 | L(kh_label); |
| 1725 | if (check_last_kh) { |
| 1726 | for (int ki = 0; ki < kw; ki++) |
| 1727 | for (int oc = 0; oc < oc_block; oc += 4) |
| 1728 | for (int kk = 0; kk < jcp.nb_ic_blocking; kk++) { |
| 1729 | bool last_kernel_loads = (kk == jcp.nb_ic_blocking - 1 |
| 1730 | && ki == kw - 1 && (oc + 4) == oc_block); |
| 1731 | |
| 1732 | if (last_kernel_loads) { |
| 1733 | cmp(reg_kj, 1); |
| 1734 | je(last_iter_label, T_NEAR); |
| 1735 | } |
| 1736 | |
| 1737 | kernel_loads(ki, oc, kk); |
| 1738 | for (int ii = get_iw_start(ki, l_overflow), |
| 1739 | prf_count_t0 = 0, prf_count_t1 = 0; |
| 1740 | ii < get_iw_end(ur_w, ki, r_overflow); ii++) { |
| 1741 | int aux_dst_offset = typesize |
| 1742 | * ((ii + jcp.l_pad - ki) * oc_block + oc); |
| 1743 | v4fmaddps(zmm_out(ii, kk), zmm_ker(0), |
| 1744 | EVEX_compress_addr(aux_reg_dst, aux_dst_offset)); |
| 1745 | |
| 1746 | if (ii % 2) { |
| 1747 | if (prf_count_t0 < 4) { |
| 1748 | int aux_kernel_prf; |
| 1749 | if (last_kernel_loads) |
| 1750 | aux_kernel_prf= kernel_offset(0, prf_count_t0 |
| 1751 | + oc + 4 - oc_block, 0) + typesize * kw |
| 1752 | * oc_block * ic_block; |
| 1753 | else |
| 1754 | aux_kernel_prf = kernel_offset(kk, oc + 4 |
| 1755 | + prf_count_t0, ki); |
| 1756 | mic_prefetcht0(EVEX_compress_addr(aux_reg_ker, |
| 1757 | aux_kernel_prf)); |
| 1758 | prf_count_t0++; |
| 1759 | } else if (prf_count_t1 < 4) { |
| 1760 | mic_prefetcht1(EVEX_compress_addr(aux_reg_ker_prf, |
| 1761 | kernel_offset(kk, oc + prf_count_t1, ki))); |
| 1762 | prf_count_t1++; |
| 1763 | } |
| 1764 | } else |
| 1765 | prefetch_dst_next_kh(ki, 2, prf_count_t0, prf_count_t1); |
| 1766 | } |
| 1767 | if (last_kernel_loads) { |
| 1768 | jmp(loop_end_label, T_NEAR); |
| 1769 | |
| 1770 | L(last_iter_label); |
| 1771 | |
| 1772 | kernel_loads(ki, oc, kk); |
| 1773 | for (int ii = get_iw_start(ki, l_overflow), |
| 1774 | prf_count_t0 = 0, prf_count_t1 = 0; |
| 1775 | ii < get_iw_end(ur_w, ki, r_overflow); ii++) { |
| 1776 | int aux_dst_offset = typesize |
| 1777 | * ((ii + jcp.l_pad - ki) * oc_block + oc); |
| 1778 | v4fmaddps(zmm_out(ii, kk), zmm_ker(0), |
| 1779 | EVEX_compress_addr(aux_reg_dst, aux_dst_offset)); |
| 1780 | if (ii % 2) { |
| 1781 | if (prf_count_t0 < 4) { |
| 1782 | mic_prefetcht0(EVEX_compress_addr(aux_reg_ker_prf, |
| 1783 | kernel_offset(0, prf_count_t0, 0))); |
| 1784 | prf_count_t0++; |
| 1785 | } else if (prf_count_t1 < 4) { |
| 1786 | mic_prefetcht1(EVEX_compress_addr(aux_reg_ker_prf, |
| 1787 | kernel_offset(kk, oc + prf_count_t1, ki))); |
| 1788 | prf_count_t1++; |
| 1789 | } |
| 1790 | } |
| 1791 | } |
| 1792 | L(loop_end_label); |
| 1793 | } |
| 1794 | } |
| 1795 | } else { |
| 1796 | for (int ki = 0; ki < kw; ki++) |
| 1797 | for (int oc = 0; oc < oc_block; oc += 4) |
| 1798 | for (int kk = 0; kk < jcp.nb_ic_blocking; kk++) { |
| 1799 | kernel_loads(ki, oc, kk); |
| 1800 | |
| 1801 | for (int ii = get_iw_start(ki, l_overflow), prf_count_t1 = 0; |
| 1802 | ii < get_iw_end(ur_w, ki, r_overflow); ii++) { |
| 1803 | int aux_dst_offset = typesize |
| 1804 | * ((ii + jcp.l_pad - ki) * oc_block + oc); |
| 1805 | v4fmaddps(zmm_out(ii, kk), zmm_ker(0), |
| 1806 | EVEX_compress_addr(aux_reg_dst, aux_dst_offset)); |
| 1807 | if ((ii % 2) && (prf_count_t1 < 4)) { |
| 1808 | mic_prefetcht1(EVEX_compress_addr( |
| 1809 | aux_reg_ker_prf, kernel_offset(kk, |
| 1810 | oc + prf_count_t1, ki))); |
| 1811 | prf_count_t1++; |
| 1812 | } |
| 1813 | if ( ki == 1 && oc == 0 && kk == 0) |
| 1814 | mic_prefetcht1(EVEX_compress_addr( |
| 1815 | aux_reg_dst_prf, aux_dst_offset)); |
| 1816 | } |
| 1817 | } |
| 1818 | } |
| 1819 | |
| 1820 | add(aux_reg_ker, shift_ker_ptr); |
| 1821 | sub(aux_reg_dst, shift_dst_ptr); |
| 1822 | add(aux_reg_ker_prf, shift_ker_ptr); |
| 1823 | sub(aux_reg_dst_prf, shift_dst_ptr); |
| 1824 | |
| 1825 | dec(reg_kj); |
| 1826 | cmp(reg_kj, 0); |
| 1827 | jg(kh_label, T_NEAR); |
| 1828 | |
| 1829 | if (jcp.ndims == 5) { |
| 1830 | sub(aux_reg_dst_d, typesize * (jcp.oh * ow) * ic_block); |
| 1831 | add(aux_reg_ker_d, typesize * jcp.kw * jcp.kh * oc_block * ic_block); |
| 1832 | sub(aux_reg_dst_d_prf, typesize * (jcp.oh * ow) * ic_block); |
| 1833 | add(aux_reg_ker_d_prf, typesize * jcp.kw * jcp.kh *oc_block * ic_block); |
| 1834 | |
| 1835 | dec(reg_ki); |
| 1836 | cmp(reg_ki, 0); |
| 1837 | jg(kd_label, T_NEAR); |
| 1838 | |
| 1839 | pop(reg_src); |
| 1840 | pop(reg_src_prf); |
| 1841 | } |
| 1842 | } |
| 1843 | |
| 1844 | void jit_avx512_common_conv_bwd_data_kernel_f32::compute_loop_fma( |
| 1845 | int ur_w, int l_overflow, int r_overflow) |
| 1846 | { |
| 1847 | Label kh_label, kd_label; |
| 1848 | int kw = jcp.kw; |
| 1849 | int ow = jcp.ow; |
| 1850 | |
| 1851 | int ic_block = jcp.ic_block; |
| 1852 | int oc_block = jcp.oc_block; |
| 1853 | int l_pad = jcp.l_pad; |
| 1854 | int dilate_w = jcp.dilate_w + 1; |
| 1855 | int stride_w = jcp.stride_w; |
| 1856 | int stride_h = jcp.stride_h; |
| 1857 | |
| 1858 | int ker_pipeline_depth = 4; |
| 1859 | assert(ker_reg_base_idx + ker_pipeline_depth <= 32); |
| 1860 | assert(oc_block >= ker_pipeline_depth); |
| 1861 | |
| 1862 | int num_ker_loads = oc_block * kw; |
| 1863 | int num_inp_prfs = ur_w * nstl::min(kw, stride_w) |
| 1864 | + nstl::max(0, kw - stride_w); |
| 1865 | int num_prfs = num_ker_loads + num_inp_prfs; |
| 1866 | int num_fmas = num_ker_loads * ur_w / stride_w; |
| 1867 | int prf_inst_spacing = nstl::max(1, num_fmas / num_prfs); |
| 1868 | int prf_inst_trigger = (num_fmas % prf_inst_spacing) / 2; |
| 1869 | |
| 1870 | if (one_of(jcp.ndims, 3, 4)) { |
| 1871 | mov(aux_reg_dst, reg_dst); |
| 1872 | mov(aux_reg_ker, reg_ker); |
| 1873 | |
| 1874 | mov(aux_reg_dst_prf, reg_dst_prf); |
| 1875 | mov(aux_reg_ker_prf, reg_ker_prf); |
| 1876 | } |
| 1877 | |
| 1878 | if (jcp.ndims == 5) { |
| 1879 | push(reg_src_prf); |
| 1880 | push(reg_src); |
| 1881 | |
| 1882 | mov(reg_ki, ptr[param + GET_OFF(kd_padding)]); |
| 1883 | mov(aux_reg_dst_d, reg_dst); |
| 1884 | mov(aux_reg_ker_d, ptr[param + GET_OFF(filt)]); |
| 1885 | mov(aux_reg_dst_d_prf, reg_dst_prf); |
| 1886 | mov(aux_reg_ker_d_prf, reg_ker_prf); |
| 1887 | |
| 1888 | L(kd_label); |
| 1889 | mov(reg_kj, ptr[param + GET_OFF(kh_padding)]); |
| 1890 | } else { |
| 1891 | mov(reg_kj, reg_kh); |
| 1892 | } |
| 1893 | |
| 1894 | if (jcp.ndims == 5) { |
| 1895 | mov(aux_reg_dst, aux_reg_dst_d); |
| 1896 | mov(aux_reg_ker, aux_reg_ker_d); |
| 1897 | mov(aux_reg_dst_prf, aux_reg_dst_d_prf); |
| 1898 | mov(aux_reg_ker_prf, aux_reg_ker_d_prf); |
| 1899 | } |
| 1900 | |
| 1901 | L(kh_label); { |
| 1902 | int step = 0; |
| 1903 | int ker_prfs = 0; |
| 1904 | for (int ki = 0; ki < kw; ki++) { |
| 1905 | for (int oc = 0; oc < oc_block; oc++) { |
| 1906 | if (step == 0) { |
| 1907 | for (int i = 0; i < ker_pipeline_depth; i++) { |
| 1908 | int aux_kernel_offset = typesize * ((oc + i) * oc_block |
| 1909 | + ki * ic_block * oc_block); |
| 1910 | vmovups(zmm_ker(i), EVEX_compress_addr( |
| 1911 | aux_reg_ker, aux_kernel_offset)); |
| 1912 | } |
| 1913 | } else if (step < num_ker_loads - ker_pipeline_depth + 1) { |
| 1914 | int load_offset = ker_pipeline_depth - 1; |
| 1915 | int ker_load_reg_idx |
| 1916 | = (step + load_offset) % ker_pipeline_depth; |
| 1917 | int aux_kernel_offset = typesize * ((oc + load_offset) |
| 1918 | * oc_block + ki * ic_block * oc_block); |
| 1919 | vmovups(zmm_ker(ker_load_reg_idx), |
| 1920 | EVEX_compress_addr(aux_reg_ker, aux_kernel_offset)); |
| 1921 | } |
| 1922 | |
| 1923 | bool ker_prf_inserted = false; |
| 1924 | auto zmm_kernel = zmm_ker(step % ker_pipeline_depth); |
| 1925 | |
| 1926 | int jj_start = get_iw_start(ki, l_overflow); |
| 1927 | int jj_end = get_iw_end(ur_w, ki, r_overflow); |
| 1928 | assert(stride_w != 1 |
| 1929 | || jj_start == nstl::max(0, |
| 1930 | l_overflow - (kw - 1 - ki) * dilate_w)); |
| 1931 | assert(stride_w != 1 |
| 1932 | || jj_end == ur_w - nstl::max(0, |
| 1933 | r_overflow - ki * dilate_w)); |
| 1934 | |
| 1935 | for (int jj = jj_start; jj < jj_end; jj += stride_w) { |
| 1936 | assert((jj + l_pad - ki * dilate_w) % stride_w == 0); |
| 1937 | int aux_dst_offset = typesize * |
| 1938 | (((jj + l_pad - ki * dilate_w) |
| 1939 | / stride_w) * jcp.oc_block + oc); |
| 1940 | vfmadd231ps(zmm_out(jj, 0), zmm_kernel, |
| 1941 | EVEX_compress_addr(aux_reg_dst, aux_dst_offset, true)); |
| 1942 | |
| 1943 | int fma_idx = (step * ur_w + jj) / stride_w; |
| 1944 | int prf_slot_idx = fma_idx / prf_inst_spacing; |
| 1945 | if (fma_idx % prf_inst_spacing == prf_inst_trigger) { |
| 1946 | if (!ker_prf_inserted && ker_prfs < num_ker_loads) { |
| 1947 | int ker_prf_offset = typesize |
| 1948 | * ker_prfs * jcp.oc_block; |
| 1949 | mic_prefetcht1(EVEX_compress_addr( |
| 1950 | aux_reg_ker_prf, ker_prf_offset)); |
| 1951 | ker_prf_inserted = true; |
| 1952 | ker_prfs++; |
| 1953 | } else { |
| 1954 | int inp_prf_idx = prf_slot_idx - ker_prfs; |
| 1955 | if (inp_prf_idx < num_inp_prfs) { |
| 1956 | int inp_prf_offset |
| 1957 | = ic_block * typesize |
| 1958 | * ((inp_prf_idx / kw) * kw |
| 1959 | + (inp_prf_idx % kw)); |
| 1960 | mic_prefetcht0(EVEX_compress_addr( |
| 1961 | aux_reg_dst_prf, inp_prf_offset)); |
| 1962 | } |
| 1963 | } |
| 1964 | } |
| 1965 | } |
| 1966 | step++; |
| 1967 | } |
| 1968 | } |
| 1969 | |
| 1970 | add(aux_reg_ker, typesize * stride_h * kw * oc_block * ic_block); |
| 1971 | sub(aux_reg_dst, typesize * (jcp.dilate_h + 1) * ow * oc_block); |
| 1972 | add(aux_reg_ker_prf, typesize * stride_h * kw * oc_block * ic_block); |
| 1973 | sub(aux_reg_dst_prf, typesize * (jcp.dilate_h + 1) * ow * oc_block); |
| 1974 | |
| 1975 | dec(reg_kj); |
| 1976 | cmp(reg_kj, 0); |
| 1977 | jg(kh_label, T_NEAR); |
| 1978 | } |
| 1979 | if (jcp.ndims == 5) { |
| 1980 | sub(aux_reg_dst_d, |
| 1981 | typesize * (jcp.dilate_d + 1) * jcp.oh * ow * ic_block); |
| 1982 | add(aux_reg_ker_d, typesize * jcp.stride_d * jcp.kw * jcp.kh |
| 1983 | * oc_block * ic_block); |
| 1984 | sub(aux_reg_dst_d_prf, |
| 1985 | typesize * (jcp.dilate_d + 1) * jcp.oh * ow * ic_block); |
| 1986 | add(aux_reg_ker_d_prf, typesize * jcp.stride_d * jcp.kw * jcp.kh |
| 1987 | * oc_block * ic_block); |
| 1988 | |
| 1989 | dec(reg_ki); |
| 1990 | cmp(reg_ki, 0); |
| 1991 | jg(kd_label, T_NEAR); |
| 1992 | } |
| 1993 | |
| 1994 | if (jcp.ndims == 5) |
| 1995 | { |
| 1996 | pop(reg_src); |
| 1997 | pop(reg_src_prf); |
| 1998 | } |
| 1999 | } |
| 2000 | |
| 2001 | void jit_avx512_common_conv_bwd_data_kernel_f32::compute_loop_fma_core( |
| 2002 | int ur_w, int l_overflow, int r_overflow) |
| 2003 | { |
| 2004 | int kw = jcp.kw; |
| 2005 | int ow = jcp.ow; |
| 2006 | int dilate_w = jcp.dilate_w + 1; |
| 2007 | int stride_w = jcp.stride_w; |
| 2008 | int ic_block = jcp.ic_block; |
| 2009 | int oc_block = jcp.oc_block; |
| 2010 | int nb_ic_block = jcp.nb_ic_blocking; |
| 2011 | Label kh_label, kd_label; |
| 2012 | |
| 2013 | int shift_ker_ptr = typesize * kw * oc_block * ic_block; |
| 2014 | int shift_dst_ptr = typesize * (jcp.dilate_h + 1) * ow * oc_block; |
| 2015 | |
| 2016 | auto output_offset = [=](int oi, int oc, int ki) { |
| 2017 | return typesize * |
| 2018 | (((oi + jcp.l_pad - ki * dilate_w) / stride_w) * oc_block + oc); |
| 2019 | }; |
| 2020 | auto kernel_offset = [=](int icb, int oc, int ki) { |
| 2021 | int blk_idx = icb * jcp.kh * jcp.kw * jcp.kd + ki; |
| 2022 | int blk_offset = blk_idx * jcp.oc_block * jcp.ic_block; |
| 2023 | int oc_offset = oc * jcp.oc_block; |
| 2024 | return typesize * (blk_offset + oc_offset); |
| 2025 | }; |
| 2026 | |
| 2027 | if (one_of(jcp.ndims, 3, 4)) { |
| 2028 | mov(aux_reg_dst, reg_dst); |
| 2029 | mov(aux_reg_ker, reg_ker); |
| 2030 | } |
| 2031 | |
| 2032 | if (jcp.ndims == 5) { |
| 2033 | push(reg_src_prf); |
| 2034 | push(reg_src); |
| 2035 | |
| 2036 | mov(reg_ki, ptr[param + GET_OFF(kd_padding)]); |
| 2037 | mov(aux_reg_dst_d, reg_dst); |
| 2038 | mov(aux_reg_ker_d, ptr[param + GET_OFF(filt)]); |
| 2039 | |
| 2040 | L(kd_label); |
| 2041 | mov(reg_kj, ptr[param + GET_OFF(kh_padding)]); |
| 2042 | } else { |
| 2043 | mov(reg_kj, reg_kh); |
| 2044 | } |
| 2045 | |
| 2046 | if (jcp.ndims == 5) { |
| 2047 | mov(aux_reg_dst, aux_reg_dst_d); |
| 2048 | mov(aux_reg_ker, aux_reg_ker_d); |
| 2049 | } |
| 2050 | |
| 2051 | L(kh_label); |
| 2052 | { |
| 2053 | for (int ki = 0; ki < kw; ki++) { |
| 2054 | int jj_start = get_iw_start(ki, l_overflow); |
| 2055 | int jj_end = get_iw_end(ur_w, ki, r_overflow); |
| 2056 | for (int oc = 0; oc < oc_block; oc++) { |
| 2057 | if (jcp.kernel_kind == expl_bcast) { |
| 2058 | for (int jj = jj_start; jj < jj_end; jj++) { |
| 2059 | int aux_output_offset = output_offset(jj, oc, ki); |
| 2060 | vbroadcastss(zmm_inp(jj, nb_ic_block), |
| 2061 | ptr[aux_reg_dst + aux_output_offset]); |
| 2062 | } |
| 2063 | } |
| 2064 | for (int ii = 0; ii < nb_ic_block; ii++) { |
| 2065 | int aux_kernel_offset = kernel_offset(ii, oc, ki); |
| 2066 | if (jj_end - jj_start > 0) |
| 2067 | vmovups(zmm_wei, EVEX_compress_addr(aux_reg_ker, |
| 2068 | aux_kernel_offset)); |
| 2069 | for (int jj = jj_start; jj < jj_end; jj += stride_w) |
| 2070 | if (jcp.kernel_kind == expl_bcast) |
| 2071 | vfmadd231ps(zmm_out(jj, ii), |
| 2072 | zmm_inp(jj, nb_ic_block), zmm_wei); |
| 2073 | else |
| 2074 | vfmadd231ps(zmm_out(jj, ii), zmm_wei, |
| 2075 | EVEX_compress_addr(aux_reg_dst, |
| 2076 | output_offset(jj, oc, ki), true)); |
| 2077 | } |
| 2078 | } |
| 2079 | } |
| 2080 | add(aux_reg_ker, shift_ker_ptr); |
| 2081 | sub(aux_reg_dst, shift_dst_ptr); |
| 2082 | dec(reg_kj); |
| 2083 | cmp(reg_kj, 0); |
| 2084 | jg(kh_label, T_NEAR); |
| 2085 | } |
| 2086 | |
| 2087 | if (jcp.ndims == 5) { |
| 2088 | sub(aux_reg_dst_d, |
| 2089 | typesize * (jcp.dilate_d + 1) * jcp.oh * ow * ic_block); |
| 2090 | add(aux_reg_ker_d, typesize * jcp.kw * jcp.kh * oc_block * ic_block); |
| 2091 | |
| 2092 | dec(reg_ki); |
| 2093 | cmp(reg_ki, 0); |
| 2094 | jg(kd_label, T_NEAR); |
| 2095 | |
| 2096 | pop(reg_src); |
| 2097 | pop(reg_src_prf); |
| 2098 | } |
| 2099 | } |
| 2100 | |
| 2101 | inline void jit_avx512_common_conv_bwd_data_kernel_f32::compute_loop( |
| 2102 | int ur_w, int l_overflow, int r_overflow) |
| 2103 | { |
| 2104 | if (jcp.ndims == 5) push(reg_oi); |
| 2105 | |
| 2106 | prepare_output(ur_w); |
| 2107 | |
| 2108 | Label skip_compute_loop; |
| 2109 | if (jcp.ndims == 5) { |
| 2110 | mov(reg_kj, ptr[param + GET_OFF(kd_padding)]); |
| 2111 | cmp(reg_kj, 0); |
| 2112 | je(skip_compute_loop, T_NEAR); |
| 2113 | } |
| 2114 | mov(reg_kj, ptr[param + GET_OFF(kh_padding)]); |
| 2115 | cmp(reg_kj, 0); |
| 2116 | je(skip_compute_loop, T_NEAR); |
| 2117 | |
| 2118 | if (jcp.ver == ver_4fma) |
| 2119 | compute_loop_4fma(ur_w, l_overflow, r_overflow); |
| 2120 | else if (jcp.ver == ver_fma) |
| 2121 | if (mayiuse(avx512_mic)) |
| 2122 | compute_loop_fma(ur_w, l_overflow, r_overflow); |
| 2123 | else |
| 2124 | if (jcp.kernel_kind == embd_bcast && jcp.nb_ic_blocking == 1) |
| 2125 | compute_loop_fma(ur_w, l_overflow, r_overflow); |
| 2126 | else |
| 2127 | compute_loop_fma_core(ur_w, l_overflow, r_overflow); |
| 2128 | else |
| 2129 | assert("!unknown convolution version"); |
| 2130 | |
| 2131 | L(skip_compute_loop); |
| 2132 | store_output(ur_w); |
| 2133 | if (jcp.ndims == 5) pop(reg_oi); |
| 2134 | } |
| 2135 | |
| 2136 | void jit_avx512_common_conv_bwd_data_kernel_f32::generate() |
| 2137 | { |
| 2138 | int iw = jcp.iw; |
| 2139 | int kw = jcp.kw; |
| 2140 | int ur_w = jcp.ur_w; |
| 2141 | int ic_block = jcp.ic_block; |
| 2142 | int oc_block = jcp.oc_block; |
| 2143 | int ur_w_tail = jcp.ur_w_tail; |
| 2144 | int dilate_w = jcp.dilate_w + 1; |
| 2145 | int stride_w = jcp.stride_w; |
| 2146 | |
| 2147 | int dst_shift = jcp.typesize_in * (ur_w / stride_w) * ic_block; |
| 2148 | int src_shift = jcp.typesize_out * ur_w * oc_block; |
| 2149 | |
| 2150 | preamble(); |
| 2151 | |
| 2152 | mov(reg_src, ptr[param + GET_OFF(src)]); |
| 2153 | mov(reg_dst, ptr[param + GET_OFF(dst)]); |
| 2154 | mov(reg_ker, ptr[param + GET_OFF(filt)]); |
| 2155 | |
| 2156 | mov(reg_kh, ptr[param + GET_OFF(kh_padding)]); |
| 2157 | mov(reg_src_prf, ptr[param + GET_OFF(src_prf)]); |
| 2158 | mov(reg_dst_prf, ptr[param + GET_OFF(dst_prf)]); |
| 2159 | mov(reg_ker_prf, ptr[param + GET_OFF(filt_prf)]); |
| 2160 | |
| 2161 | int l_overflow = nstl::max(0, ((kw - 1) * dilate_w - jcp.l_pad) / stride_w); |
| 2162 | int r_overflow = nstl::max(0, ((kw - 1) * dilate_w |
| 2163 | - nstl::max(0, jcp.r_pad)) / stride_w); |
| 2164 | int r_overflow1 = nstl::max(0, ((kw - 1) * dilate_w |
| 2165 | - nstl::max(0, jcp.r_pad) - ur_w_tail) / stride_w); |
| 2166 | |
| 2167 | int n_oi = iw / ur_w; |
| 2168 | if (r_overflow1 > 0) n_oi--; |
| 2169 | |
| 2170 | if (ur_w == iw) { |
| 2171 | compute_loop(ur_w, l_overflow, r_overflow); |
| 2172 | } else if (n_oi == 0) { |
| 2173 | compute_loop(ur_w, l_overflow, r_overflow1); |
| 2174 | add(reg_src, src_shift); |
| 2175 | add(reg_dst, dst_shift); |
| 2176 | add(reg_src_prf, src_shift); |
| 2177 | add(reg_dst_prf, dst_shift); |
| 2178 | if (ur_w_tail != 0) |
| 2179 | compute_loop(ur_w_tail, 0, r_overflow); |
| 2180 | } else { |
| 2181 | xor_(reg_oi, reg_oi); |
| 2182 | if (l_overflow > 0) { |
| 2183 | compute_loop(ur_w, l_overflow, 0); |
| 2184 | add(reg_src, src_shift); |
| 2185 | add(reg_dst, dst_shift); |
| 2186 | add(reg_src_prf, src_shift); |
| 2187 | add(reg_dst_prf, dst_shift); |
| 2188 | |
| 2189 | inc(reg_oi); |
| 2190 | } |
| 2191 | if ((l_overflow <= 0 && n_oi > 0) |
| 2192 | || (l_overflow > 0 && n_oi > 1)) { |
| 2193 | Label ow_loop_label; |
| 2194 | L(ow_loop_label); { |
| 2195 | compute_loop(ur_w, 0, 0); |
| 2196 | add(reg_src, src_shift); |
| 2197 | add(reg_dst, dst_shift); |
| 2198 | add(reg_src_prf, src_shift); |
| 2199 | add(reg_dst_prf, dst_shift); |
| 2200 | |
| 2201 | inc(reg_oi); |
| 2202 | cmp(reg_oi, n_oi); |
| 2203 | jl(ow_loop_label, T_NEAR); |
| 2204 | } |
| 2205 | } |
| 2206 | if (r_overflow1 > 0) { |
| 2207 | compute_loop(ur_w, 0, r_overflow1); |
| 2208 | add(reg_src, src_shift); |
| 2209 | add(reg_dst, dst_shift); |
| 2210 | add(reg_src_prf, src_shift); |
| 2211 | add(reg_dst_prf, dst_shift); |
| 2212 | } |
| 2213 | if (ur_w_tail != 0) { |
| 2214 | compute_loop(ur_w_tail, 0, r_overflow); |
| 2215 | } |
| 2216 | } |
| 2217 | |
| 2218 | postamble(); |
| 2219 | } |
| 2220 | |
| 2221 | status_t jit_avx512_common_conv_bwd_data_kernel_f32::init_conf( |
| 2222 | jit_conv_conf_t &jcp, |
| 2223 | const convolution_desc_t &cd, |
| 2224 | const memory_desc_wrapper &diff_src_d, |
| 2225 | const memory_desc_wrapper &weights_d, |
| 2226 | const memory_desc_wrapper &diff_dst_d) |
| 2227 | { |
| 2228 | if (!mayiuse(avx512_common)) return status::unimplemented; |
| 2229 | |
| 2230 | jcp = zero<decltype(jcp)>(); |
| 2231 | |
| 2232 | jcp.simd_w = cpu_isa_traits<avx512_common>::vlen / sizeof(float); |
| 2233 | const bool with_groups = weights_d.ndims() == diff_src_d.ndims() + 1; |
| 2234 | int ndims = diff_src_d.ndims(); |
| 2235 | |
| 2236 | jcp.ndims = ndims; |
| 2237 | jcp.prop_kind = cd.prop_kind; |
| 2238 | |
| 2239 | jcp.ngroups = with_groups ? weights_d.dims()[0] : 1; |
| 2240 | jcp.mb = diff_src_d.dims()[0]; |
| 2241 | |
| 2242 | jcp.oc = diff_dst_d.dims()[1] / jcp.ngroups; |
| 2243 | jcp.oc_without_padding = jcp.oc; |
| 2244 | jcp.ic = diff_src_d.dims()[1] / jcp.ngroups; |
| 2245 | |
| 2246 | jcp.id = (ndims == 5) ? diff_src_d.dims()[2] : 1; |
| 2247 | jcp.ih = (ndims == 3) ? 1 : diff_src_d.dims()[ndims-2]; |
| 2248 | jcp.iw = diff_src_d.dims()[ndims-1]; |
| 2249 | jcp.od = (ndims == 5) ? diff_dst_d.dims()[2] : 1; |
| 2250 | jcp.oh = (ndims == 3) ? 1 : diff_dst_d.dims()[ndims-2]; |
| 2251 | jcp.ow = diff_dst_d.dims()[ndims-1]; |
| 2252 | |
| 2253 | jcp.kd = (ndims == 5) ? weights_d.dims()[with_groups + 2] : 1; |
| 2254 | jcp.kh = (ndims == 3) ? 1 : weights_d.dims()[with_groups + ndims - 2]; |
| 2255 | jcp.kw = weights_d.dims()[with_groups + ndims - 1]; |
| 2256 | |
| 2257 | jcp.f_pad = (ndims == 5) ? cd.padding[0][0] : 0; |
| 2258 | jcp.t_pad = (ndims == 3) ? 0 : cd.padding[0][ndims-4]; |
| 2259 | jcp.l_pad = cd.padding[0][ndims-3]; |
| 2260 | |
| 2261 | jcp.stride_d = (ndims == 5) ? cd.strides[0] : 1; |
| 2262 | jcp.stride_h = (ndims == 3) ? 1 : cd.strides[ndims-4]; |
| 2263 | jcp.stride_w = cd.strides[ndims-3]; |
| 2264 | |
| 2265 | jcp.dilate_d = (ndims == 5) ? cd.dilates[0] : 0; |
| 2266 | jcp.dilate_h = (ndims == 3) ? 0 : cd.dilates[ndims-4]; |
| 2267 | jcp.dilate_w = cd.dilates[ndims-3]; |
| 2268 | if ((jcp.dilate_w != 0 && jcp.stride_w != 1) |
| 2269 | || (jcp.dilate_d != 0 && jcp.stride_d != 1) |
| 2270 | || (jcp.dilate_h != 0 && jcp.stride_h != 1)) |
| 2271 | return status::unimplemented; |
| 2272 | |
| 2273 | jcp.r_pad = (jcp.ow - 1) * jcp.stride_w + (jcp.kw - 1) * (jcp.dilate_w + 1) |
| 2274 | - (jcp.iw + jcp.l_pad - 1); |
| 2275 | jcp.b_pad = (jcp.oh - 1) * jcp.stride_h + (jcp.kh - 1) * (jcp.dilate_h + 1) |
| 2276 | - (jcp.ih + jcp.t_pad - 1); |
| 2277 | jcp.back_pad = (jcp.od - 1) * jcp.stride_d |
| 2278 | + (jcp.kd - 1) * (jcp.dilate_d + 1) - (jcp.id + jcp.f_pad - 1); |
| 2279 | |
| 2280 | jcp.aligned_threads = 0; |
| 2281 | |
| 2282 | jcp.is_1stconv = false; |
| 2283 | |
| 2284 | jcp.oc_block = jcp.simd_w; |
| 2285 | jcp.ic_block = jcp.is_1stconv ? jcp.ic : jcp.simd_w; |
| 2286 | |
| 2287 | bool ok_to_pad_channels = true |
| 2288 | && jcp.ngroups == 1 |
| 2289 | && diff_src_d.data_type() == data_type::f32; |
| 2290 | |
| 2291 | if (ok_to_pad_channels) { |
| 2292 | jcp.oc = rnd_up(jcp.oc, jcp.oc_block); |
| 2293 | jcp.ic = rnd_up(jcp.ic, jcp.ic_block); |
| 2294 | } |
| 2295 | |
| 2296 | auto dat_tag = pick(ndims - 3, nCw16c, nChw16c, nCdhw16c); |
| 2297 | auto wei_tag = with_groups |
| 2298 | ? pick(ndims - 3, gOIw16o16i, gOIhw16o16i, gOIdhw16o16i) |
| 2299 | : pick(ndims - 3, OIw16o16i, OIhw16o16i, OIdhw16o16i); |
| 2300 | jcp.src_tag = diff_src_d.matches_one_of_tag(dat_tag); |
| 2301 | jcp.dst_tag = diff_dst_d.matches_one_of_tag(dat_tag); |
| 2302 | |
| 2303 | bool args_ok = true |
| 2304 | && jcp.oc % jcp.oc_block == 0 |
| 2305 | && jcp.ic % jcp.ic_block == 0 |
| 2306 | && jcp.src_tag == dat_tag |
| 2307 | && jcp.dst_tag == dat_tag; |
| 2308 | if (!args_ok) |
| 2309 | return status::unimplemented; |
| 2310 | |
| 2311 | jcp.nb_ic = jcp.ic / jcp.ic_block; |
| 2312 | jcp.nb_oc = jcp.oc / jcp.oc_block; |
| 2313 | |
| 2314 | jcp.ur_w = jcp.stride_w; |
| 2315 | |
| 2316 | int regs = 28; |
| 2317 | if (jcp.iw <= regs) |
| 2318 | jcp.ur_w = jcp.iw; |
| 2319 | else { |
| 2320 | for (int ur_w = regs; ur_w > 0; --ur_w) |
| 2321 | if (ur_w % jcp.stride_w == 0) { |
| 2322 | jcp.ur_w = ur_w; |
| 2323 | break; |
| 2324 | } |
| 2325 | } |
| 2326 | int l_overflow = nstl::max(0, ((jcp.kw - 1) * (jcp.dilate_w + 1) |
| 2327 | - jcp.l_pad) / jcp.stride_w); |
| 2328 | int r_overflow1 = nstl::max(0, ((jcp.kw - 1) * (jcp.dilate_w + 1) |
| 2329 | - nstl::max(0, jcp.r_pad) - jcp.iw % jcp.ur_w) / jcp.stride_w); |
| 2330 | int n_oi = jcp.iw / jcp.ur_w; |
| 2331 | if (r_overflow1 > 0) n_oi--; |
| 2332 | |
| 2333 | if (mayiuse(avx512_common) |
| 2334 | && diff_dst_d.data_type() == data_type::f32 |
| 2335 | && weights_d.data_type() == data_type::f32 |
| 2336 | && diff_src_d.data_type() == data_type::f32) { |
| 2337 | jcp.ver = ver_fma; |
| 2338 | jcp.typesize_in = sizeof(float); |
| 2339 | jcp.typesize_out = sizeof(float); |
| 2340 | if (mayiuse(avx512_mic_4ops) |
| 2341 | && jcp.stride_w == 1 && jcp.stride_h == 1 && jcp.stride_d == 1) { |
| 2342 | jcp.ver = ver_4fma; |
| 2343 | } |
| 2344 | } else { |
| 2345 | return status::unimplemented; |
| 2346 | } |
| 2347 | |
| 2348 | jcp.wei_tag = weights_d.matches_one_of_tag(wei_tag); |
| 2349 | if (jcp.wei_tag != wei_tag) |
| 2350 | return status::unimplemented; |
| 2351 | |
| 2352 | if (!utils::everyone_is(0, jcp.dilate_d, jcp.dilate_h, jcp.dilate_w) |
| 2353 | && jcp.ver != ver_fma) |
| 2354 | return status::unimplemented; |
| 2355 | |
| 2356 | jcp.nb_ic_blocking = jcp.nb_oc_blocking = 1; |
| 2357 | if (jcp.ver == ver_4fma) { |
| 2358 | if (jcp.kw == 3 && jcp.kh == 3 && jcp.iw == 7 && jcp.ih == 7) { |
| 2359 | jcp.nb_ic_blocking = 2; |
| 2360 | } else { |
| 2361 | for (int i = jcp.nb_ic; i > 0; i--) |
| 2362 | if (i * jcp.ur_w <= regs && jcp.nb_ic % i == 0) { |
| 2363 | jcp.nb_ic_blocking = i; |
| 2364 | break; |
| 2365 | } |
| 2366 | } |
| 2367 | } |
| 2368 | |
| 2369 | jcp.loop_order = loop_gnc; |
| 2370 | |
| 2371 | bool large_code_size = (jcp.ur_w != jcp.ow) |
| 2372 | && ((l_overflow <= 0 && n_oi > 0) ||(l_overflow > 0 && n_oi > 1)) |
| 2373 | && (r_overflow1 > 0) && (l_overflow > 0); |
| 2374 | if (large_code_size) { |
| 2375 | const int max_code_size = 24 * 1024; |
| 2376 | const int num_ops_per_reg = 6 + jcp.oc_block * jcp.kw; |
| 2377 | int mult = 1; |
| 2378 | if (l_overflow > 0) mult += 1; |
| 2379 | if (r_overflow1 > 0) mult += 1; |
| 2380 | for (int ur_w = jcp.ur_w; ur_w > regs/2; --ur_w) { |
| 2381 | if ((ur_w / jcp.stride_w) * mult * num_ops_per_reg * 9.2 |
| 2382 | < max_code_size) { |
| 2383 | if (ur_w % jcp.stride_w == 0) { |
| 2384 | jcp.ur_w = ur_w; |
| 2385 | break; |
| 2386 | } |
| 2387 | } |
| 2388 | } |
| 2389 | } |
| 2390 | |
| 2391 | if (jcp.ver == ver_fma && mayiuse(avx512_core)) { |
| 2392 | int try_nb_ic_blocking = 2; |
| 2393 | unsigned int ker_inp_size = typesize * jcp.iw * jcp.ic_block |
| 2394 | * try_nb_ic_blocking * jcp.kh; |
| 2395 | unsigned int ker_out_size = typesize * jcp.ow * jcp.oc_block; |
| 2396 | unsigned int ker_wei_size = typesize * jcp.kh * jcp.kw * jcp.ic_block |
| 2397 | * jcp.oc_block * try_nb_ic_blocking; |
| 2398 | unsigned int ker_total_size = ker_inp_size + ker_out_size |
| 2399 | + ker_wei_size; |
| 2400 | if (!(jcp.kw == 1 || (jcp.kw == 5 && jcp.iw < 8) |
| 2401 | || (jcp.kw < 5 && ((jcp.iw <= 5 || (jcp.iw > 8 && jcp.iw <= 13)) |
| 2402 | || ker_total_size > L1_cache_size ))) |
| 2403 | || jcp.stride_h > 1 || jcp.stride_d > 1) { |
| 2404 | jcp.kernel_kind = embd_bcast; |
| 2405 | jcp.ur_w = nstl::min(jcp.iw, regs); |
| 2406 | jcp.nb_ic_blocking = jcp.nb_oc_blocking = 1; |
| 2407 | if (!(jcp.kw > 3 || (jcp.kw == 3 && ker_total_size < L1_cache_size |
| 2408 | && jcp.ow > 8)) && jcp.stride_h == 1) |
| 2409 | if (jcp.nb_ic % try_nb_ic_blocking == 0) { |
| 2410 | jcp.nb_ic_blocking = try_nb_ic_blocking; |
| 2411 | jcp.ur_w = 31 / (jcp.nb_ic_blocking + 1); |
| 2412 | if (jcp.iw < jcp.ur_w) jcp.ur_w = jcp.iw; |
| 2413 | } |
| 2414 | } else { |
| 2415 | jcp.kernel_kind = expl_bcast; |
| 2416 | jcp.nb_oc_blocking = 1; |
| 2417 | jcp.nb_ic_blocking = 4; |
| 2418 | if (jcp.nb_ic < jcp.nb_ic_blocking) jcp.nb_ic_blocking = jcp.nb_ic; |
| 2419 | if (jcp.nb_ic % jcp.nb_ic_blocking != 0) |
| 2420 | for (int i = jcp.nb_ic_blocking; i > 0; i--) |
| 2421 | if (jcp.nb_ic % i == 0) { |
| 2422 | jcp.nb_ic_blocking = i; |
| 2423 | break; |
| 2424 | } |
| 2425 | jcp.ur_w = 31 / (jcp.nb_ic_blocking + 1); |
| 2426 | if (jcp.iw < jcp.ur_w) jcp.ur_w = jcp.iw; |
| 2427 | } |
| 2428 | } |
| 2429 | jcp.ur_w_tail = jcp.iw % jcp.ur_w; |
| 2430 | |
| 2431 | if (l_overflow * jcp.stride_w > jcp.ur_w) |
| 2432 | return status::unimplemented; |
| 2433 | int r_overflow_no_tail = nstl::max(0, ((jcp.kw - 1) * (jcp.dilate_w + 1) |
| 2434 | - nstl::max(0, jcp.r_pad) - jcp.ur_w_tail) / jcp.stride_w); |
| 2435 | if (r_overflow_no_tail * jcp.stride_w > jcp.ur_w) |
| 2436 | return status::unimplemented; |
| 2437 | if ((jcp.iw > jcp.ur_w) && (jcp.ur_w % jcp.stride_w != 0)) |
| 2438 | return status::unimplemented; |
| 2439 | |
| 2440 | pick_loop_order(jcp); |
| 2441 | |
| 2442 | jcp.nb_oc_L2 = jcp.nb_oc; |
| 2443 | if (jcp.ver == ver_4fma && (jcp.kh < 5 && jcp.kw < 5)) { |
| 2444 | for (int divf = 2, temp_nb = jcp.nb_oc_L2; divf <= jcp.nb_oc; |
| 2445 | divf++) { |
| 2446 | size_t l2_src = jcp.iw * jcp.ic_block * jcp.nb_ic_blocking * jcp.ih |
| 2447 | * jcp.id; |
| 2448 | size_t l2_dst = jcp.ow * jcp.oc_block * temp_nb * jcp.oh * jcp.od; |
| 2449 | size_t l2_filt = jcp.kw * jcp.oc_block * jcp.ic_block * jcp.kh |
| 2450 | * jcp.kd * jcp.nb_ic_blocking * temp_nb; |
| 2451 | if (4 * (l2_src + l2_dst + l2_filt) > KNx_L2_EFFECTIVE_CAPACITY) { |
| 2452 | if (jcp.kh == 3 && jcp.ih == 7) { |
| 2453 | jcp.nb_oc_L2 = 1; |
| 2454 | break; |
| 2455 | } |
| 2456 | temp_nb = (jcp.nb_oc_L2 % divf == 0 ? jcp.nb_oc_L2 / divf |
| 2457 | : jcp.nb_oc_L2); |
| 2458 | } else { |
| 2459 | jcp.nb_oc_L2 = temp_nb; |
| 2460 | break; |
| 2461 | } |
| 2462 | } |
| 2463 | } |
| 2464 | |
| 2465 | args_ok = true |
| 2466 | && jcp.ic <= diff_src_d.padded_dims()[1] |
| 2467 | && jcp.oc <= diff_dst_d.padded_dims()[1] |
| 2468 | && jcp.ic <= weights_d.padded_dims()[with_groups + 1] |
| 2469 | && jcp.oc <= weights_d.padded_dims()[with_groups + 0]; |
| 2470 | if (!args_ok) return status::unimplemented; |
| 2471 | |
| 2472 | return status::success; |
| 2473 | } |
| 2474 | |
| 2475 | void jit_avx512_common_conv_bwd_data_kernel_f32::init_scratchpad( |
| 2476 | memory_tracking::registrar_t &scratchpad, const jit_conv_conf_t &jcp) { |
| 2477 | UNUSED(scratchpad); |
| 2478 | UNUSED(jcp); |
| 2479 | } |
| 2480 | |
| 2481 | const int jit_avx512_common_conv_bwd_weights_kernel_f32::max_ur_w = 28; |
| 2482 | |
| 2483 | void jit_avx512_common_conv_bwd_weights_kernel_f32::od_step_comeback_pointers() |
| 2484 | { |
| 2485 | Label kd_comeback_label; |
| 2486 | |
| 2487 | /* 'depth' loop count bound by 'kd_work_size' */ |
| 2488 | mov(kj, reg_kd_count); |
| 2489 | L(kd_comeback_label); { |
| 2490 | int inp_mult = jcp.is_1stconv ? 1 : jcp.ic_block; |
| 2491 | int iw = jcp.ver == ver_4fma ? jcp.tr_iw : jcp.iw; |
| 2492 | sub(reg_input, |
| 2493 | jcp.typesize_in * (jcp.dilate_d + 1) * jcp.ih * iw * inp_mult); |
| 2494 | sub(reg_kernel, |
| 2495 | jcp.typesize_out * jcp.kh * jcp.kw * jcp.ic_block * jcp.oc_block); |
| 2496 | dec(kj); |
| 2497 | cmp(kj, 0); |
| 2498 | jg(kd_comeback_label, T_NEAR); |
| 2499 | } |
| 2500 | } |
| 2501 | |
| 2502 | void jit_avx512_common_conv_bwd_weights_kernel_f32::oh_step_comeback_pointers() |
| 2503 | { |
| 2504 | Label kh_comeback_label, kd_comeback_label; |
| 2505 | mov(kj, reg_kh); |
| 2506 | L(kh_comeback_label); { |
| 2507 | int inp_mult = jcp.is_1stconv ? 1 : jcp.ic_block; |
| 2508 | int iw = jcp.ver == ver_4fma ? jcp.tr_iw : jcp.iw; |
| 2509 | sub(reg_input, jcp.typesize_in * (jcp.dilate_h + 1) * iw * inp_mult); |
| 2510 | sub(reg_kernel, |
| 2511 | jcp.typesize_out * jcp.kw * jcp.ic_block * jcp.oc_block); |
| 2512 | dec(kj); |
| 2513 | cmp(kj, 0); |
| 2514 | jg(kh_comeback_label, T_NEAR); |
| 2515 | } |
| 2516 | } |
| 2517 | |
| 2518 | void jit_avx512_common_conv_bwd_weights_kernel_f32::compute_ic_block_step_fma( |
| 2519 | int ur_w, int pad_l, int pad_r, |
| 2520 | int ic_block_step, int input_offset, int kernel_offset, |
| 2521 | int output_offset, bool input_wraparound) |
| 2522 | { |
| 2523 | |
| 2524 | int kw = jcp.kw; |
| 2525 | int ic_block = jcp.ic_block; |
| 2526 | int oc_block = jcp.oc_block; |
| 2527 | for (int i_kw = 0; i_kw < kw; i_kw++) |
| 2528 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) |
| 2529 | vmovups(Zmm(i_kw * ic_block_step + i_ic), |
| 2530 | EVEX_compress_addr(reg_kernel, typesize * (i_kw * ic_block |
| 2531 | + i_ic) * jcp.oc_block + kernel_offset)); |
| 2532 | |
| 2533 | for (int i_ur = 0; i_ur < ur_w; i_ur++) { |
| 2534 | if (i_ur == 0) { |
| 2535 | vmovups(Zmm(kw * ic_block_step + (i_ur + 0) % 4), |
| 2536 | EVEX_compress_addr(reg_output, typesize * (i_ur + 0) |
| 2537 | * oc_block + output_offset)); |
| 2538 | if (ur_w > 1) vmovups(Zmm(kw * ic_block_step + (i_ur + 1) % 4), |
| 2539 | EVEX_compress_addr(reg_output, typesize * (i_ur + 1) * oc_block |
| 2540 | + output_offset)); |
| 2541 | if (ur_w > 2) vmovups(Zmm(kw * ic_block_step + (i_ur + 2) % 4), |
| 2542 | EVEX_compress_addr(reg_output, typesize * (i_ur + 2) * oc_block |
| 2543 | + output_offset)); |
| 2544 | if (ur_w > 3) vmovups(Zmm(kw * ic_block_step + (i_ur + 3) % 4), |
| 2545 | EVEX_compress_addr(reg_output, typesize * (i_ur + 3) * oc_block |
| 2546 | + output_offset)); |
| 2547 | } else if (i_ur + 3 < ur_w) |
| 2548 | vmovups(Zmm(kw * ic_block_step + (i_ur + 3) % 4), |
| 2549 | EVEX_compress_addr(reg_output, typesize * (i_ur + 3) * oc_block |
| 2550 | + output_offset)); |
| 2551 | |
| 2552 | for (int i_kw = 0; i_kw < kw; i_kw++) { |
| 2553 | int i_iw = i_ur * jcp.stride_w + i_kw * (jcp.dilate_w + 1); |
| 2554 | if (i_iw - pad_l < 0 || i_iw > (ur_w - 1) * jcp.stride_w + |
| 2555 | (kw - 1) * (jcp.dilate_w + 1) - pad_r) continue; |
| 2556 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) { |
| 2557 | const size_t i_offset = (size_t)input_offset |
| 2558 | + (size_t)typesize * (jcp.ver == ver_4fma |
| 2559 | ? (i_iw - pad_l + i_ic * jcp.tr_iw) |
| 2560 | : (jcp.is_1stconv |
| 2561 | ? (i_iw - pad_l) + (size_t)i_ic |
| 2562 | * ((size_t)jcp.ih*jcp.iw*jcp.id) |
| 2563 | : (i_iw - pad_l) * ic_block + i_ic)); |
| 2564 | vfmadd231ps(Zmm(i_kw * ic_block_step + i_ic), |
| 2565 | Zmm(kw * ic_block_step + i_ur % 4), |
| 2566 | EVEX_compress_addr_safe(reg_input, i_offset, reg_long_offt, |
| 2567 | true)); |
| 2568 | } |
| 2569 | } |
| 2570 | } |
| 2571 | |
| 2572 | for (int i_kw = 0; i_kw < kw; i_kw++) |
| 2573 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) |
| 2574 | vmovups(EVEX_compress_addr(reg_kernel, typesize |
| 2575 | * (i_kw * ic_block + i_ic) * jcp.oc_block + kernel_offset), |
| 2576 | Zmm(i_kw * ic_block_step + i_ic)); |
| 2577 | } |
| 2578 | |
| 2579 | void jit_avx512_common_conv_bwd_weights_kernel_f32::compute_ic_block_step_4fma( |
| 2580 | int ur_w, int pad_l, int pad_r, |
| 2581 | int ic_block_step, int input_offset, int kernel_offset, |
| 2582 | int output_offset, bool input_wraparound) |
| 2583 | { |
| 2584 | // TODO: add prefetches to fma version as well |
| 2585 | |
| 2586 | assert(jcp.ver == ver_4fma); |
| 2587 | |
| 2588 | int kw = jcp.kw; |
| 2589 | int ic_block = jcp.ic_block; |
| 2590 | int oc_block = jcp.oc_block; |
| 2591 | |
| 2592 | auto zmm_ker = [=](int i_kw, int i_ic) { |
| 2593 | return Zmm(i_kw * ic_block_step + i_ic); |
| 2594 | }; |
| 2595 | |
| 2596 | auto ker_addr = [=](int i_kw, int i_ic) { |
| 2597 | size_t local_offset |
| 2598 | = jcp.typesize_out * (i_kw * ic_block + i_ic) * jcp.oc_block; |
| 2599 | return EVEX_compress_addr(reg_kernel, local_offset + kernel_offset); |
| 2600 | }; |
| 2601 | |
| 2602 | auto inp_addr = [=](int i_iw, int i_ic, ptrdiff_t extra_offset = 0) { |
| 2603 | int stride = jcp.tr_iw * (jcp.is_1stconv ? jcp.ih : 1); |
| 2604 | int local_offset = jcp.typesize_in * (i_iw + i_ic * stride); |
| 2605 | return EVEX_compress_addr(reg_input, |
| 2606 | local_offset + input_offset + extra_offset); |
| 2607 | }; |
| 2608 | |
| 2609 | auto zmm_out = [=](int i_iw) { |
| 2610 | // TODO: move reg calc to global member funcs |
| 2611 | const int out_zmm_base_idx = 28; |
| 2612 | return Zmm(out_zmm_base_idx + i_iw % 4); |
| 2613 | }; |
| 2614 | |
| 2615 | auto out_addr = [=](int i_ur) { |
| 2616 | return EVEX_compress_addr(reg_output, |
| 2617 | jcp.typesize_in * i_ur * oc_block + output_offset); |
| 2618 | }; |
| 2619 | |
| 2620 | auto pf_callback = [=](int i_ur, int i_kw, int i_ic) { |
| 2621 | assert(i_ur % 4 == 0); |
| 2622 | if (i_ur == 0) |
| 2623 | prefetcht1(ker_addr(i_kw, i_ic)); |
| 2624 | if (i_ur + 4 >= ur_w) |
| 2625 | prefetcht0(ker_addr(i_kw, i_ic)); |
| 2626 | |
| 2627 | const ptrdiff_t next_input_block_offset |
| 2628 | = jcp.typesize_in * ic_block_step * jcp.tr_iw; |
| 2629 | if (i_ur % 16 == 4 && i_kw == 0) { |
| 2630 | if (i_ur + 16 < ur_w) |
| 2631 | prefetcht0(inp_addr(i_ur + 16, i_ic)); |
| 2632 | else |
| 2633 | prefetcht0(inp_addr(0, i_ic, next_input_block_offset)); |
| 2634 | } |
| 2635 | if (i_ur % 16 == 4 && i_kw == 1) { |
| 2636 | if (input_wraparound) |
| 2637 | prefetcht1(inp_addr(i_ur, i_ic, -input_offset)); |
| 2638 | else |
| 2639 | prefetcht1(inp_addr(i_ur, i_ic, next_input_block_offset)); |
| 2640 | } |
| 2641 | }; |
| 2642 | |
| 2643 | for (int i_kw = 0; i_kw < kw; i_kw++) |
| 2644 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) { |
| 2645 | auto zmm = zmm_ker(i_kw, i_ic); |
| 2646 | vpxord(zmm, zmm, zmm); |
| 2647 | } |
| 2648 | |
| 2649 | for (int i_ur = 0; i_ur < ur_w; i_ur += 4) { |
| 2650 | |
| 2651 | for (int i = 0; i < 4; i++) { |
| 2652 | auto zmm = zmm_out(i_ur + i); |
| 2653 | if (i_ur + i < ur_w) |
| 2654 | vmovups(zmm, out_addr(i_ur + i)); |
| 2655 | else |
| 2656 | vpxord(zmm, zmm, zmm); |
| 2657 | prefetcht0(out_addr(i_ur + i + 4)); |
| 2658 | } |
| 2659 | |
| 2660 | for (int i_kw = 0; i_kw < kw; i_kw++) |
| 2661 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) { |
| 2662 | int i_iw = i_ur + i_kw; |
| 2663 | v4fmaddps(zmm_ker(i_kw, i_ic), |
| 2664 | zmm_out(i_ur), inp_addr(i_iw, i_ic)); |
| 2665 | pf_callback(i_ur, i_kw, i_ic); |
| 2666 | } |
| 2667 | } |
| 2668 | |
| 2669 | for (int i_kw = 0; i_kw < kw; i_kw++) |
| 2670 | for (int i_ic = 0; i_ic < ic_block_step; i_ic++) { |
| 2671 | auto addr = ker_addr(i_kw, i_ic); |
| 2672 | auto zmm = zmm_ker(i_kw, i_ic); |
| 2673 | vaddps(zmm, zmm, addr); |
| 2674 | vmovups(addr, zmm); |
| 2675 | } |
| 2676 | } |
| 2677 | |
| 2678 | void jit_avx512_common_conv_bwd_weights_kernel_f32::compute_ic_block_step( |
| 2679 | int ur_w, int pad_l, int pad_r, |
| 2680 | int ic_block_step, int input_offset, int kernel_offset, |
| 2681 | int output_offset, bool input_wraparound) |
| 2682 | { |
| 2683 | if (jcp.ver == ver_4fma) |
| 2684 | compute_ic_block_step_4fma(ur_w, pad_l, pad_r, |
| 2685 | ic_block_step, input_offset, kernel_offset, output_offset, |
| 2686 | input_wraparound); |
| 2687 | else if (jcp.ver == ver_fma) |
| 2688 | compute_ic_block_step_fma(ur_w, pad_l, pad_r, |
| 2689 | ic_block_step, input_offset, kernel_offset, output_offset, |
| 2690 | input_wraparound); |
| 2691 | else |
| 2692 | assert(!"unknown convolution version"); |
| 2693 | } |
| 2694 | |
| 2695 | void jit_avx512_common_conv_bwd_weights_kernel_f32 |
| 2696 | ::compute_oh_step_unroll_ow_icblock( |
| 2697 | int ic_block_step, int max_ur_w) |
| 2698 | { |
| 2699 | UNUSED(max_ur_w); |
| 2700 | |
| 2701 | Label kh_label, kd_label; |
| 2702 | |
| 2703 | int ic_block = jcp.ic_block; |
| 2704 | int oc_block = jcp.oc_block; |
| 2705 | int inp_mul = !jcp.is_1stconv ? ic_block : 1; |
| 2706 | int iw = jcp.ver == ver_4fma ? jcp.tr_iw : jcp.iw; |
| 2707 | int ow = jcp.ow; |
| 2708 | |
| 2709 | int r_pad = nstl::max(0, (ow - 1) * jcp.stride_w |
| 2710 | + (jcp.kw - 1) * (jcp.dilate_w + 1) - (jcp.iw + jcp.l_pad - 1)); |
| 2711 | int l_pad = jcp.l_pad; |
| 2712 | |
| 2713 | if (jcp.ndims == 5) { |
| 2714 | L(kd_label); |
| 2715 | mov(reg_input, aux_reg_input); |
| 2716 | mov(reg_kernel, aux_reg_kernel); |
| 2717 | } |
| 2718 | |
| 2719 | mov(kj, reg_kh); |
| 2720 | L(kh_label); |
| 2721 | { |
| 2722 | for (int i_b_ic = 0; i_b_ic < jcp.ic_block; i_b_ic += ic_block_step) { |
| 2723 | const int input_offset = jcp.typesize_in |
| 2724 | * (jcp.ver == ver_4fma ? i_b_ic * iw : i_b_ic); |
| 2725 | compute_ic_block_step(jcp.ur_w, l_pad, r_pad, ic_block_step, |
| 2726 | input_offset, jcp.typesize_out * i_b_ic * jcp.oc_block, 0, |
| 2727 | i_b_ic + ic_block_step >= jcp.ic_block); |
| 2728 | } |
| 2729 | add(reg_input, jcp.typesize_in * (jcp.dilate_h + 1) * iw * inp_mul); |
| 2730 | add(reg_kernel, jcp.typesize_out * jcp.kw * ic_block * oc_block); |
| 2731 | dec(kj); |
| 2732 | cmp(kj, 0); |
| 2733 | jg(kh_label, T_NEAR); |
| 2734 | } |
| 2735 | |
| 2736 | if (jcp.ndims == 5) { |
| 2737 | add(aux_reg_input, |
| 2738 | jcp.typesize_in * (jcp.dilate_d + 1) * jcp.ih * iw * inp_mul); |
| 2739 | add(aux_reg_kernel, jcp.typesize_out * jcp.kh * jcp.kw * ic_block |
| 2740 | * oc_block); |
| 2741 | dec(ki); |
| 2742 | cmp(ki, 0); |
| 2743 | jg(kd_label, T_NEAR); |
| 2744 | } |
| 2745 | } |
| 2746 | |
| 2747 | void jit_avx512_common_conv_bwd_weights_kernel_f32 |
| 2748 | ::compute_oh_step_unroll_ow( |
| 2749 | int ic_block_step, int max_ur_w) |
| 2750 | { |
| 2751 | Label kh_label, ic_block_label, kd_label; |
| 2752 | |
| 2753 | UNUSED(max_ur_w); |
| 2754 | |
| 2755 | int ic_block = jcp.ic_block; |
| 2756 | int oc_block = jcp.oc_block; |
| 2757 | |
| 2758 | int ow = jcp.ow; |
| 2759 | |
| 2760 | int r_pad = nstl::max(0, |
| 2761 | (ow - 1) * jcp.stride_w + (jcp.kw - 1) * (jcp.dilate_w + 1) |
| 2762 | - (jcp.iw + jcp.l_pad - 1)); |
| 2763 | int l_pad = jcp.l_pad; |
| 2764 | |
| 2765 | if (jcp.ndims == 5) { |
| 2766 | L(kd_label); |
| 2767 | mov(reg_input, aux_reg_input); |
| 2768 | mov(reg_kernel, aux_reg_kernel); |
| 2769 | } |
| 2770 | |
| 2771 | mov(kj, reg_kh); |
| 2772 | L(kh_label); |
| 2773 | { |
| 2774 | xor_(b_ic, b_ic); |
| 2775 | L(ic_block_label); { |
| 2776 | compute_ic_block_step(ow, l_pad, r_pad, ic_block_step, |
| 2777 | 0, 0, 0); |
| 2778 | size_t inp_icblk_stride = jcp.is_1stconv |
| 2779 | ? (size_t)jcp.ih * jcp.iw * jcp.id |
| 2780 | : (jcp.ver == ver_4fma ? jcp.tr_iw : 1); |
| 2781 | size_t input_offset |
| 2782 | = inp_icblk_stride * jcp.typesize_in * ic_block_step; |
| 2783 | safe_add(reg_input, input_offset, reg_long_offt); |
| 2784 | add(reg_kernel, jcp.typesize_out * ic_block_step * oc_block); |
| 2785 | add(b_ic, ic_block_step); |
| 2786 | cmp(b_ic, jcp.ic_block); |
| 2787 | jl(ic_block_label, T_NEAR); |
| 2788 | } |
| 2789 | |
| 2790 | if (jcp.is_1stconv) { |
| 2791 | size_t input_offset |
| 2792 | = (size_t)jcp.typesize_in * jcp.id * jcp.ih * jcp.iw * ic_block; |
| 2793 | safe_sub(reg_input, input_offset, reg_long_offt); |
| 2794 | add(reg_input, jcp.typesize_in * (jcp.dilate_h + 1) * jcp.iw); |
| 2795 | } else if (jcp.ver != ver_4fma) { |
| 2796 | add(reg_input, jcp.typesize_in |
| 2797 | * ((jcp.dilate_h + 1) * jcp.iw - 1) * ic_block); |
| 2798 | } |
| 2799 | add(reg_kernel, jcp.typesize_out * (jcp.kw - 1) * ic_block * oc_block); |
| 2800 | dec(kj); |
| 2801 | cmp(kj, 0); |
| 2802 | jg(kh_label, T_NEAR); |
| 2803 | } |
| 2804 | if (jcp.ndims == 5) { |
| 2805 | add(aux_reg_input, jcp.typesize_in * (jcp.dilate_d + 1) * jcp.ih |
| 2806 | * jcp.iw * (jcp.is_1stconv ? 1 : ic_block)); |
| 2807 | add(aux_reg_kernel, jcp.typesize_out * jcp.kh * jcp.kw * ic_block |
| 2808 | * oc_block); |
| 2809 | dec(ki); |
| 2810 | cmp(ki, 0); |
| 2811 | jg(kd_label, T_NEAR); |
| 2812 | } |
| 2813 | } |
| 2814 | |
| 2815 | void jit_avx512_common_conv_bwd_weights_kernel_f32 |
| 2816 | ::compute_oh_step_common( |
| 2817 | int ic_block_step, int max_ur_w) |
| 2818 | { |
| 2819 | Label kh_label, ic_block_label, ow_block_label, kd_label; |
| 2820 | |
| 2821 | int ic_block = jcp.ic_block; |
| 2822 | int oc_block = jcp.oc_block; |
| 2823 | |
| 2824 | int ow = jcp.ow; |
| 2825 | int r_pad = nstl::max(0, (ow - 1) * jcp.stride_w |
| 2826 | + (jcp.kw - 1) * (jcp.dilate_w + 1) - (jcp.iw + jcp.l_pad - 1)); |
| 2827 | int l_pad = jcp.ver == ver_4fma ? 0 : jcp.l_pad; |
| 2828 | |
| 2829 | int ur_w = nstl::min(ow, max_ur_w); |
| 2830 | int ur_w_trips = ow / ur_w; |
| 2831 | int ur_w_tail = ow % ur_w; |
| 2832 | if ((ur_w_tail == 0 && r_pad != 0) |
| 2833 | || r_pad >= ur_w_tail) { |
| 2834 | if (ur_w_trips > 1) { |
| 2835 | ur_w_tail += ur_w; |
| 2836 | ur_w_trips--; |
| 2837 | } else { |
| 2838 | ur_w_tail += (ur_w - ur_w / 2); |
| 2839 | ur_w = ur_w / 2; |
| 2840 | } |
| 2841 | } |
| 2842 | |
| 2843 | int inp_mult = (jcp.is_1stconv || jcp.ver == ver_4fma) ? 1 : ic_block; |
| 2844 | int input_comeback = (ur_w_trips * ur_w * jcp.stride_w - l_pad) * inp_mult; |
| 2845 | int output_comeback = ur_w_trips * ur_w * oc_block; |
| 2846 | |
| 2847 | if (jcp.ndims == 5) { |
| 2848 | L(kd_label); |
| 2849 | mov(reg_input, aux_reg_input); |
| 2850 | mov(reg_kernel, aux_reg_kernel); |
| 2851 | } |
| 2852 | |
| 2853 | mov(kj, reg_kh); |
| 2854 | L(kh_label); { |
| 2855 | xor_(b_ic, b_ic); |
| 2856 | L(ic_block_label); { |
| 2857 | if (l_pad != 0) { |
| 2858 | ur_w_trips--; |
| 2859 | compute_ic_block_step(ur_w, l_pad, 0, ic_block_step, 0, 0, 0); |
| 2860 | add(reg_input, jcp.typesize_in * (ur_w * jcp.stride_w - l_pad) |
| 2861 | * inp_mult); |
| 2862 | add(reg_output, jcp.typesize_in * ur_w * oc_block); |
| 2863 | } |
| 2864 | |
| 2865 | if (ur_w_trips > 0) { |
| 2866 | xor_(reg_ur_w_trips, reg_ur_w_trips); |
| 2867 | L(ow_block_label); { |
| 2868 | compute_ic_block_step(ur_w, 0, 0, ic_block_step, 0, 0, 0); |
| 2869 | add(reg_input, jcp.typesize_in * ur_w * jcp.stride_w |
| 2870 | * inp_mult); |
| 2871 | add(reg_output, jcp.typesize_in * ur_w * oc_block); |
| 2872 | |
| 2873 | inc(reg_ur_w_trips); |
| 2874 | cmp(reg_ur_w_trips, ur_w_trips); |
| 2875 | jl(ow_block_label, T_NEAR); |
| 2876 | } |
| 2877 | } |
| 2878 | |
| 2879 | if (ur_w_tail > 0) compute_ic_block_step(ur_w_tail, 0, r_pad, |
| 2880 | ic_block_step, 0, 0, 0); |
| 2881 | |
| 2882 | sub(reg_input, jcp.typesize_in * input_comeback); |
| 2883 | sub(reg_output, jcp.typesize_in * output_comeback); |
| 2884 | int inp_icblk_stride = jcp.is_1stconv |
| 2885 | ? jcp.ih * jcp.iw * jcp.id |
| 2886 | : (jcp.ver == ver_4fma ? jcp.tr_iw : 1); |
| 2887 | size_t input_offset |
| 2888 | = inp_icblk_stride * jcp.typesize_in * ic_block_step; |
| 2889 | safe_add(reg_input, input_offset, reg_long_offt); |
| 2890 | add(reg_kernel, jcp.typesize_out * ic_block_step * oc_block); |
| 2891 | |
| 2892 | add(b_ic, ic_block_step); |
| 2893 | cmp(b_ic, jcp.ic_block); |
| 2894 | jl(ic_block_label, T_NEAR); |
| 2895 | } |
| 2896 | if (jcp.is_1stconv) { |
| 2897 | size_t input_offset |
| 2898 | = (size_t)jcp.typesize_in * jcp.id * jcp.ih * jcp.iw * ic_block; |
| 2899 | safe_sub(reg_input, input_offset, reg_long_offt); |
| 2900 | add(reg_input, jcp.typesize_in * (jcp.dilate_h + 1) * jcp.iw); |
| 2901 | } else if (jcp.ver != ver_4fma) { |
| 2902 | add(reg_input, jcp.typesize_in |
| 2903 | * ((jcp.dilate_h + 1 ) * jcp.iw - 1) * ic_block); |
| 2904 | } |
| 2905 | add(reg_kernel, jcp.typesize_out * (jcp.kw - 1) * ic_block * oc_block); |
| 2906 | dec(kj); |
| 2907 | cmp(kj, 0); |
| 2908 | jg(kh_label, T_NEAR); |
| 2909 | } |
| 2910 | if (jcp.ndims == 5) { |
| 2911 | add(aux_reg_input, jcp.typesize_in * (jcp.dilate_d + 1) * jcp.ih |
| 2912 | * jcp.iw * (jcp.is_1stconv ? 1 : ic_block)); |
| 2913 | add(aux_reg_kernel, jcp.typesize_out * jcp.kh * jcp.kw * ic_block |
| 2914 | * oc_block); |
| 2915 | dec(ki); |
| 2916 | cmp(ki, 0); |
| 2917 | jg(kd_label, T_NEAR); |
| 2918 | } |
| 2919 | } |
| 2920 | |
| 2921 | void jit_avx512_common_conv_bwd_weights_kernel_f32 |
| 2922 | ::compute_oh_step_disp() |
| 2923 | { |
| 2924 | int ic_block_step = jcp.kw <= 3 ? 8 : (jcp.kw <= 7 ? 4 : 2); |
| 2925 | if (jcp.is_1stconv) { |
| 2926 | bool large_code = jcp.kw >= 7 && (jcp.l_pad > 0 || jcp.t_pad > 0); |
| 2927 | ic_block_step |
| 2928 | = (jcp.kw * jcp.ic_block <= 28 && !large_code) ? jcp.ic_block : 1; |
| 2929 | } |
| 2930 | |
| 2931 | bool too_large_to_unroll |
| 2932 | = (jcp.kw > 1 || jcp.kh > 1 || jcp.kd > 1) |
| 2933 | && (jcp.stride_w > 1 || jcp.stride_h > 1 || jcp.stride_d > 1); |
| 2934 | |
| 2935 | int ow = jcp.ow; |
| 2936 | if (jcp.ndims == 5) { |
| 2937 | /* NOTE: reg_kd_count = aux_reg_input = r12. The following order of |
| 2938 | * 'movs' must be guaranteed. */ |
| 2939 | mov(ki, reg_kd_count); |
| 2940 | push(reg_kd_count); |
| 2941 | mov(aux_reg_input, reg_input); |
| 2942 | mov(aux_reg_kernel, reg_kernel); |
| 2943 | } |
| 2944 | |
| 2945 | if (jcp.kw <= 3 && ow <= 16 && !too_large_to_unroll) |
| 2946 | compute_oh_step_unroll_ow_icblock(ic_block_step, max_ur_w); |
| 2947 | else if (ow <= max_ur_w) |
| 2948 | compute_oh_step_unroll_ow(ic_block_step, max_ur_w); |
| 2949 | else |
| 2950 | compute_oh_step_common(ic_block_step, max_ur_w); |
| 2951 | |
| 2952 | if (jcp.ndims == 5) { |
| 2953 | mov(reg_input, aux_reg_input); |
| 2954 | mov(reg_kernel, aux_reg_kernel); |
| 2955 | pop(reg_kd_count); |
| 2956 | od_step_comeback_pointers(); |
| 2957 | } else { |
| 2958 | oh_step_comeback_pointers(); |
| 2959 | } |
| 2960 | } |
| 2961 | |
| 2962 | void jit_avx512_common_conv_bwd_weights_kernel_f32::maybe_zero_kernel() |
| 2963 | { |
| 2964 | Label skip_zeroing, zeroing_loop; |
| 2965 | |
| 2966 | mov(reg_tmp, ptr[param + GET_OFF(channel)]); |
| 2967 | cmp(reg_tmp, 0); |
| 2968 | jz(skip_zeroing, T_NEAR); |
| 2969 | |
| 2970 | Zmm zero = Zmm(0); |
| 2971 | vpxord(zero, zero, zero); |
| 2972 | xor_(reg_tmp, reg_tmp); |
| 2973 | L(zeroing_loop); { |
| 2974 | assert(jcp.oc_block * jcp.typesize_out |
| 2975 | == cpu_isa_traits<avx512_common>::vlen); |
| 2976 | for (int ic1 = 0; ic1 < jcp.ic_block; ic1++) |
| 2977 | vmovups(ptr[reg_kernel + reg_tmp + ic1 * jcp.oc_block |
| 2978 | * jcp.typesize_out], zero); |
| 2979 | add(reg_tmp, jcp.ic_block * jcp.oc_block * jcp.typesize_out); |
| 2980 | cmp(reg_tmp, jcp.ic_block * jcp.oc_block * jcp.kw * jcp.kh * jcp.kd |
| 2981 | * jcp.typesize_out); |
| 2982 | jnz(zeroing_loop); |
| 2983 | } |
| 2984 | |
| 2985 | L(skip_zeroing); |
| 2986 | } |
| 2987 | |
| 2988 | void jit_avx512_common_conv_bwd_weights_kernel_f32::bias_kernel() |
| 2989 | { |
| 2990 | Label skip_bias, bias_loop, skip_load_bias; |
| 2991 | |
| 2992 | mov(reg_tmp, ptr[param + GET_OFF(flags)]); |
| 2993 | test(reg_tmp,reg_tmp); |
| 2994 | jne(skip_bias, T_NEAR); |
| 2995 | |
| 2996 | mov(reg_bias, ptr[param + GET_OFF(bias)]); |
| 2997 | mov(reg_output, ptr[param + GET_OFF(dst)]); |
| 2998 | vpxord(Zmm(1), Zmm(1), Zmm(1)); |
| 2999 | |
| 3000 | mov(reg_tmp, ptr[param + GET_OFF(channel)]); |
| 3001 | cmp(reg_tmp, 0); |
| 3002 | jne(skip_load_bias, T_NEAR); |
| 3003 | vmovups(Zmm(1), ptr[reg_bias]); |
| 3004 | |
| 3005 | L(skip_load_bias); |
| 3006 | |
| 3007 | mov(reg_oi, ptr[param + GET_OFF(d_worksize)]); |
| 3008 | sub(reg_oi, ptr[param + GET_OFF(d_index)]); |
| 3009 | mov(reg_tmp, jcp.oc_block * jcp.ow * jcp.oh * jcp.typesize_out); |
| 3010 | imul(reg_oi, reg_tmp); |
| 3011 | |
| 3012 | xor_(reg_tmp, reg_tmp); |
| 3013 | L(bias_loop); { |
| 3014 | vmovups(Zmm(0), ptr[reg_output + reg_tmp]); |
| 3015 | vaddps(Zmm(1), Zmm(1), Zmm(0)); |
| 3016 | add(reg_tmp, jcp.oc_block * jcp.typesize_out); |
| 3017 | cmp(reg_tmp, reg_oi); |
| 3018 | jl(bias_loop); |
| 3019 | } |
| 3020 | vmovups(EVEX_compress_addr(reg_bias,0), Zmm(1)); |
| 3021 | |
| 3022 | L(skip_bias); |
| 3023 | } |
| 3024 | |
| 3025 | void jit_avx512_common_conv_bwd_weights_kernel_f32 |
| 3026 | ::compute_oh_loop_common() |
| 3027 | { |
| 3028 | int b_pad = jcp.b_pad; |
| 3029 | int t_pad = jcp.t_pad; |
| 3030 | bool is_dilated = jcp.dilate_h != 0; |
| 3031 | int dilate_h = jcp.dilate_h + 1; |
| 3032 | int stride_h = jcp.stride_h; |
| 3033 | const int inp_mult = jcp.is_1stconv ? 1 : jcp.ic_block; |
| 3034 | int iw = jcp.ver == ver_4fma ? jcp.tr_iw : jcp.iw; |
| 3035 | Label oh_label, oh_label_end, oh_tpad_label, oh_tpad_tail_label, |
| 3036 | oh_bpad_label, oh_bpad_label_end, od_label, od_label_end, |
| 3037 | oh_dilate_label_shift, oh_dilate_label_noshift, oh_dilate_label_end; |
| 3038 | |
| 3039 | int ow = jcp.ow; |
| 3040 | |
| 3041 | mov(reg_kh, jcp.kh); |
| 3042 | xor_(reg_ih_count, reg_ih_count); |
| 3043 | xor_(reg_oj, reg_oj); |
| 3044 | /* Compute 'top' edge */ |
| 3045 | if (t_pad > 0) { |
| 3046 | const int kh_range = 1 + (jcp.kh - 1) * dilate_h; |
| 3047 | const int overflow |
| 3048 | = nstl::max(0, jcp.kh - div_up(t_pad + jcp.ih, dilate_h)); |
| 3049 | const int underflow = div_up(t_pad, dilate_h); |
| 3050 | const int initial_inp_ker_overlap = jcp.kh - overflow - underflow; |
| 3051 | mov(reg_kh, initial_inp_ker_overlap); |
| 3052 | add(reg_kernel, jcp.typesize_out * underflow * jcp.kw * jcp.ic_block |
| 3053 | * jcp.oc_block); |
| 3054 | // generate loop to process kernel while it remains within t_pad + ih |
| 3055 | if (kh_range < t_pad + jcp.ih) { |
| 3056 | if (is_dilated) { |
| 3057 | const int tail = t_pad % dilate_h; |
| 3058 | const int shift = tail == 0 ? 0 : dilate_h - tail; |
| 3059 | mov(reg_tmp, shift); |
| 3060 | if (tail != 0) |
| 3061 | add(reg_input, jcp.typesize_in * shift * iw * inp_mult); |
| 3062 | } |
| 3063 | L(oh_tpad_label); { |
| 3064 | compute_oh_step_disp(); |
| 3065 | add(reg_output, jcp.typesize_in * ow * jcp.oc_block); |
| 3066 | if (is_dilated) { |
| 3067 | inc(reg_tmp); |
| 3068 | cmp(reg_tmp, dilate_h); |
| 3069 | jl(oh_dilate_label_shift, T_NEAR); |
| 3070 | // unshift input as new kernel element enters |
| 3071 | sub(reg_input, jcp.typesize_in * (dilate_h - 1) * iw * inp_mult); |
| 3072 | xor_(reg_tmp, reg_tmp); |
| 3073 | } |
| 3074 | // kernel overlap only changes when (t_pad + oj) % dilate_h == 0 |
| 3075 | sub(reg_kernel, jcp.typesize_out * stride_h * jcp.kw |
| 3076 | * jcp.ic_block * jcp.oc_block); |
| 3077 | add(reg_kh, stride_h); |
| 3078 | if (is_dilated) { |
| 3079 | jmp(oh_dilate_label_noshift, T_NEAR); |
| 3080 | L(oh_dilate_label_shift); |
| 3081 | // shift input as old kernel element progresses |
| 3082 | add(reg_input, jcp.typesize_in * stride_h * iw * inp_mult); |
| 3083 | L(oh_dilate_label_noshift); |
| 3084 | } |
| 3085 | inc(reg_oj); |
| 3086 | add(reg_ih_count, stride_h); |
| 3087 | |
| 3088 | // final number of kernel elements that overlap with input |
| 3089 | const int final_inp_ker_overlap |
| 3090 | = nstl::min(jcp.kh, div_up(jcp.ih, dilate_h)); |
| 3091 | cmp(reg_kh, final_inp_ker_overlap); |
| 3092 | jl(oh_tpad_label, T_NEAR); |
| 3093 | } |
| 3094 | } |
| 3095 | // need second loop to process kernel if it is larger than the input |
| 3096 | // (does not apply to dilations as they must have unit stride) |
| 3097 | if (kh_range >= jcp.ih + (t_pad % stride_h == 0 ? stride_h : |
| 3098 | t_pad % stride_h)) { |
| 3099 | assert(!is_dilated); |
| 3100 | mov(reg_kh, jcp.ih); |
| 3101 | L(oh_tpad_tail_label); { |
| 3102 | compute_oh_step_disp(); |
| 3103 | add(reg_output, jcp.typesize_in * ow * jcp.oc_block); |
| 3104 | sub(reg_kernel, jcp.typesize_out * stride_h * jcp.kw |
| 3105 | * jcp.ic_block * jcp.oc_block); |
| 3106 | |
| 3107 | inc(reg_oj); |
| 3108 | add(reg_ih_count, stride_h); |
| 3109 | |
| 3110 | cmp(reg_ih_count, nstl::min(t_pad, jcp.oh * stride_h)); |
| 3111 | jl(oh_tpad_tail_label, T_NEAR); |
| 3112 | } |
| 3113 | } |
| 3114 | // correct any excess shifts to kernel and input |
| 3115 | // (does not apply to dilations as they must have unit stride, |
| 3116 | // kernel must fit inside input, and padding is smaller than input) |
| 3117 | if (t_pad <= jcp.oh * stride_h) { |
| 3118 | // kernel has moved beyond padding (adjust for stride effects) |
| 3119 | if (t_pad % stride_h != 0) { |
| 3120 | assert(!is_dilated); |
| 3121 | int inp_corr = stride_h - t_pad % stride_h; |
| 3122 | add(reg_kernel, jcp.typesize_out * inp_corr * jcp.kw |
| 3123 | * jcp.ic_block * jcp.oc_block); |
| 3124 | add(reg_input, jcp.typesize_in * inp_corr * iw * inp_mult); |
| 3125 | } |
| 3126 | } else { |
| 3127 | // kernel still overlaps padding (complete reset) |
| 3128 | assert(!is_dilated); |
| 3129 | sub(reg_kernel, jcp.typesize_out * (t_pad - jcp.oh * stride_h) |
| 3130 | * jcp.kw * jcp.ic_block * jcp.oc_block); |
| 3131 | } |
| 3132 | } |
| 3133 | |
| 3134 | cmp(reg_ih_count, jcp.ihp - b_pad - (jcp.kh - 1) * dilate_h); |
| 3135 | jge(oh_label_end, T_NEAR); |
| 3136 | cmp(reg_oj, jcp.oh); |
| 3137 | jge(oh_label, T_NEAR); |
| 3138 | |
| 3139 | /* Compute middle block(s) */ |
| 3140 | mov(reg_kh, jcp.kh); |
| 3141 | L(oh_label); { |
| 3142 | compute_oh_step_disp(); |
| 3143 | add(reg_input, jcp.typesize_in * stride_h * iw * inp_mult); |
| 3144 | add(reg_output, jcp.typesize_in * ow * jcp.oc_block); |
| 3145 | |
| 3146 | inc(reg_oj); |
| 3147 | add(reg_ih_count, stride_h); |
| 3148 | |
| 3149 | cmp(reg_ih_count, jcp.ihp - b_pad - (jcp.kh - 1) * dilate_h); |
| 3150 | jge(oh_label_end, T_NEAR); |
| 3151 | |
| 3152 | cmp(reg_oj, jcp.oh); |
| 3153 | jl(oh_label, T_NEAR); |
| 3154 | } |
| 3155 | L(oh_label_end); |
| 3156 | |
| 3157 | /* Compute bottom edge */ |
| 3158 | if (b_pad > 0) { |
| 3159 | cmp(reg_oj, jcp.oh); |
| 3160 | jge(oh_bpad_label_end, T_NEAR); |
| 3161 | |
| 3162 | if (is_dilated) { |
| 3163 | mov(reg_kh, jcp.kh - 1); // assumes unit stride for dilations |
| 3164 | mov(reg_tmp, 0); |
| 3165 | } else { |
| 3166 | mov(reg_kh, jcp.ihp - b_pad); |
| 3167 | sub(reg_kh, reg_ih_count); |
| 3168 | } |
| 3169 | L(oh_bpad_label); |
| 3170 | { |
| 3171 | compute_oh_step_disp(); |
| 3172 | add(reg_input, jcp.typesize_in * stride_h * iw * inp_mult); |
| 3173 | add(reg_output, jcp.typesize_in * ow * jcp.oc_block); |
| 3174 | if (is_dilated) { |
| 3175 | inc(reg_tmp); |
| 3176 | cmp(reg_tmp, dilate_h); |
| 3177 | jl(oh_dilate_label_end, T_NEAR); |
| 3178 | xor_(reg_tmp, reg_tmp); |
| 3179 | } |
| 3180 | sub(reg_kh, stride_h); |
| 3181 | cmp(reg_kh, 0); |
| 3182 | jle(oh_bpad_label_end, T_NEAR); |
| 3183 | if (is_dilated) |
| 3184 | L(oh_dilate_label_end); |
| 3185 | |
| 3186 | inc(reg_oj); |
| 3187 | cmp(reg_oj, jcp.oh); |
| 3188 | jl(oh_bpad_label, T_NEAR); |
| 3189 | } |
| 3190 | L(oh_bpad_label_end); |
| 3191 | } |
| 3192 | } |
| 3193 | |
| 3194 | void jit_avx512_common_conv_bwd_weights_kernel_f32::compute_d_loop_common() { |
| 3195 | int ic_block = jcp.ic_block; |
| 3196 | int oc_block = jcp.oc_block; |
| 3197 | const int inp_mult = jcp.is_1stconv ? 1 : jcp.ic_block; |
| 3198 | int iw = jcp.ver == ver_4fma ? jcp.tr_iw : jcp.iw; |
| 3199 | int ow = jcp.ow; |
| 3200 | const int input_backpad_overlap |
| 3201 | = div_up(jcp.id + jcp.f_pad - (jcp.kd - 1), jcp.stride_d); |
| 3202 | |
| 3203 | const size_t filter_shift |
| 3204 | = jcp.typesize_out * jcp.kh * jcp.kw * ic_block * oc_block; |
| 3205 | const size_t input_shift = jcp.typesize_in * jcp.ih * iw * inp_mult; |
| 3206 | const size_t output_shift = jcp.typesize_in * jcp.oh * ow * jcp.oc_block; |
| 3207 | |
| 3208 | Label d_loop_label, loop_end_label, common_block_label, fpad_end_label, |
| 3209 | backpad_end_label, backpad_label; |
| 3210 | |
| 3211 | if (jcp.with_bias) bias_kernel(); |
| 3212 | |
| 3213 | /* initially offset 'kd' by f_pad */ |
| 3214 | add(reg_kernel, ptr[param + GET_OFF(kd_offset)]); |
| 3215 | |
| 3216 | mov(reg_input_d, ptr[param + GET_OFF(src)]); |
| 3217 | mov(reg_output_d, ptr[param + GET_OFF(dst)]); |
| 3218 | mov(reg_d_index, ptr[param + GET_OFF(d_index)]); |
| 3219 | mov(reg_kd_count, ptr[param + GET_OFF(kd_padding)]); |
| 3220 | |
| 3221 | cmp(reg_d_index, ptr[param + GET_OFF(d_worksize)]); |
| 3222 | jge(loop_end_label, T_NEAR); |
| 3223 | |
| 3224 | L(d_loop_label); |
| 3225 | |
| 3226 | mov(reg_input, reg_input_d); |
| 3227 | mov(reg_output, reg_output_d); |
| 3228 | |
| 3229 | push(reg_input_d); |
| 3230 | push(reg_output_d); |
| 3231 | push(reg_d_index); |
| 3232 | |
| 3233 | compute_oh_loop_common(); |
| 3234 | |
| 3235 | pop(reg_d_index); |
| 3236 | pop(reg_output_d); |
| 3237 | pop(reg_input_d); |
| 3238 | |
| 3239 | /* Compute 'front' edge */ |
| 3240 | if (jcp.f_pad > 0) { |
| 3241 | |
| 3242 | /* Check if within fpad region */ |
| 3243 | cmp(reg_d_index, div_up(jcp.f_pad, jcp.stride_d)); |
| 3244 | jge(fpad_end_label, T_NEAR); |
| 3245 | |
| 3246 | /* Fpad steps */ |
| 3247 | sub(reg_kernel, filter_shift * jcp.stride_d); |
| 3248 | add(reg_kd_count, jcp.stride_d); |
| 3249 | |
| 3250 | /* Final number of kernel elements that overlap with input */ |
| 3251 | const int inp_ker_overlap = nstl::min(jcp.kd, jcp.id); |
| 3252 | cmp(reg_kd_count, inp_ker_overlap); |
| 3253 | jl(common_block_label, T_NEAR); |
| 3254 | |
| 3255 | /* Correct any excess shifts to kernel and input */ |
| 3256 | if (jcp.f_pad <= jcp.od * jcp.stride_d) { |
| 3257 | /* Filter has moved beyond padding (adjust for stride effects) */ |
| 3258 | if (jcp.f_pad % jcp.stride_d != 0) { |
| 3259 | int inp_corr = jcp.stride_d - jcp.f_pad % jcp.stride_d; |
| 3260 | add(reg_kernel, filter_shift * inp_corr); |
| 3261 | add(reg_input_d, input_shift * inp_corr); |
| 3262 | } |
| 3263 | } else { |
| 3264 | /* Filter still overlaps padding (complete reset) */ |
| 3265 | sub(reg_kernel, (jcp.f_pad - jcp.od * jcp.stride_d) * filter_shift); |
| 3266 | } |
| 3267 | |
| 3268 | /* Apply correction */ |
| 3269 | mov(reg_kd_count, jcp.kd); |
| 3270 | jmp(common_block_label); |
| 3271 | |
| 3272 | L(fpad_end_label); |
| 3273 | } |
| 3274 | |
| 3275 | /* Compute bottom edge */ |
| 3276 | if (jcp.back_pad > 0) { |
| 3277 | |
| 3278 | /* Check if within back_pad region */ |
| 3279 | cmp(reg_d_index, input_backpad_overlap - 1); |
| 3280 | jl(backpad_end_label, T_NEAR); |
| 3281 | jg(backpad_label, T_NEAR); |
| 3282 | |
| 3283 | /* Execute overlap correction between the filter and the initial |
| 3284 | * back_pad region. */ |
| 3285 | mov(reg_kd_count, |
| 3286 | jcp.id + jcp.f_pad - input_backpad_overlap * jcp.stride_d); |
| 3287 | jmp(backpad_end_label, T_NEAR); |
| 3288 | |
| 3289 | L(backpad_label); |
| 3290 | sub(reg_kd_count, jcp.stride_d); |
| 3291 | cmp(reg_kd_count, 0); |
| 3292 | jle(loop_end_label, T_NEAR); |
| 3293 | |
| 3294 | L(backpad_end_label); |
| 3295 | } |
| 3296 | |
| 3297 | /* Compute middle block */ |
| 3298 | add(reg_input_d, input_shift * jcp.stride_d); |
| 3299 | |
| 3300 | /* Execute common block and loop */ |
| 3301 | L(common_block_label); |
| 3302 | add(reg_output_d, output_shift); |
| 3303 | inc(reg_d_index); |
| 3304 | cmp(reg_d_index, ptr[param + GET_OFF(d_worksize)]); |
| 3305 | jl(d_loop_label, T_NEAR); |
| 3306 | |
| 3307 | L(loop_end_label); |
| 3308 | } |
| 3309 | |
| 3310 | bool jit_avx512_common_conv_bwd_weights_kernel_f32::compute_full_spat_loop() { |
| 3311 | // FIXME: use register mapping from the class declaration |
| 3312 | bool ok = jcp.ver == ver_4fma |
| 3313 | && everyone_is(0, jcp.dilate_h, jcp.dilate_w) |
| 3314 | && everyone_is(1, jcp.stride_h, jcp.stride_w); |
| 3315 | if (!ok) return false; |
| 3316 | if (jcp.l_pad != jcp.kw / 2 || jcp.t_pad != jcp.kh / 2) |
| 3317 | return false; |
| 3318 | |
| 3319 | // General code layout: |
| 3320 | // |
| 3321 | // Blocking over OH -- top level |
| 3322 | // (Reduces L2 pressure; not very useful right now) |
| 3323 | // Loop over all KHxKW kernel -- emit_kh_kw_loop() |
| 3324 | // Loop over OH block -- emit_h_loop() |
| 3325 | // Loop over OW blocks -- emit_fma_block() |
| 3326 | // (Supports both fully unrolled and partially unrolled versions to |
| 3327 | // reduce code size) |
| 3328 | // Loop over OW block -- emit_fma_step() |
| 3329 | |
| 3330 | int max_working_set_size = 128 * 1024; |
| 3331 | int pad_ow = jcp.ow; |
| 3332 | |
| 3333 | int inp_row_size = jcp.ic_block * jcp.tr_iw * jcp.typesize_in; |
| 3334 | int out_row_size = jcp.oc_block * pad_ow * jcp.typesize_in; |
| 3335 | int row_size = inp_row_size + out_row_size; |
| 3336 | |
| 3337 | int h_block_size = jcp.oh; |
| 3338 | int working_set_size = row_size * h_block_size; |
| 3339 | |
| 3340 | if (working_set_size > max_working_set_size) { |
| 3341 | int opt_working_set_size = 48 * 1024; |
| 3342 | assert(opt_working_set_size < max_working_set_size); |
| 3343 | |
| 3344 | while (working_set_size > opt_working_set_size) { |
| 3345 | for (int i = 2; i <= h_block_size; i++) |
| 3346 | if (i == h_block_size) |
| 3347 | h_block_size = h_block_size / 2; |
| 3348 | else if (h_block_size % i == 0) { |
| 3349 | h_block_size = h_block_size / i; |
| 3350 | break; |
| 3351 | } |
| 3352 | working_set_size = row_size * h_block_size; |
| 3353 | |
| 3354 | if (h_block_size == 1 && working_set_size > opt_working_set_size) |
| 3355 | return false; |
| 3356 | } |
| 3357 | } |
| 3358 | |
| 3359 | // NB1: t_pad <= oh_block_size and b_pad <= last_oh_block_size (see below) |
| 3360 | if (h_block_size < nstl::max(1, jcp.t_pad) |
| 3361 | || jcp.b_pad > (jcp.oh % h_block_size == 0 ? h_block_size |
| 3362 | : jcp.oh % h_block_size)) |
| 3363 | return false; |
| 3364 | |
| 3365 | // check that we can use simple arithmetic for prefetch address |
| 3366 | // calculations |
| 3367 | // TODO: we need some traits for this check (Roma) |
| 3368 | int cache_line_size = 64; |
| 3369 | assert(jcp.ic_block * typesize == 64); |
| 3370 | assert(jcp.oc_block * typesize == 64); |
| 3371 | |
| 3372 | int num_inp_l2_pfs = jcp.tr_iw * h_block_size; |
| 3373 | int avg_h_loop_len = h_block_size; |
| 3374 | int num_inp_l2_pfs_per_fma_block |
| 3375 | = div_up(num_inp_l2_pfs, avg_h_loop_len * jcp.kw * jcp.kh); |
| 3376 | int num_out_l2_pfs = pad_ow * h_block_size; |
| 3377 | int num_out_l2_pfs_per_fma_block |
| 3378 | = div_up(num_out_l2_pfs, avg_h_loop_len * jcp.kw * jcp.kh); |
| 3379 | |
| 3380 | Opmask reg_h_block = k1; // 32-bit only on Intel(R) Xeon Phi(TM) processors |
| 3381 | Reg64 reg_kh = rax; |
| 3382 | Reg64 reg_kw = rbx; |
| 3383 | Reg64 reg_tmp = abi_not_param1; |
| 3384 | Reg32 reg_tmp_w = reg_tmp.cvt32(); |
| 3385 | Reg64 reg_ohs = rdx; |
| 3386 | Reg64 reg_ihs = rsi; |
| 3387 | Reg64 reg_h = r8; |
| 3388 | Reg64 reg_i = r9; |
| 3389 | Reg64 reg_j = r10; |
| 3390 | |
| 3391 | Reg64 reg_inp = r13; |
| 3392 | Reg64 reg_out = r14; |
| 3393 | Reg64 reg_ker = r15; |
| 3394 | |
| 3395 | Reg64 reg_inp_pf_l1 = rbp; |
| 3396 | |
| 3397 | Reg64 reg_inp_pf_l2 = r11; |
| 3398 | Reg64 reg_out_pf_l2 = r12; |
| 3399 | |
| 3400 | Xmm reg_inp_pf_save = xmm17; |
| 3401 | Xmm reg_out_pf_save = xmm18; |
| 3402 | |
| 3403 | Reg64 reg_inp_save = abi_param1; |
| 3404 | Reg64 reg_out_save = reg_tmp; |
| 3405 | |
| 3406 | auto zmm_out = [&](int oi) { return Zmm(24 + oi % 8); }; |
| 3407 | auto zmm_ker = [&](int ic1) { return Zmm(ic1); }; |
| 3408 | auto inp_addr = [&](int oi, int ic1) { |
| 3409 | return ptr[reg_inp + (ic1 * jcp.tr_iw + oi) * jcp.typesize_in]; |
| 3410 | }; |
| 3411 | auto out_addr = [&](int oi, int oj = 0) { |
| 3412 | assert(jcp.ver == ver_4fma); |
| 3413 | return ptr[reg_out |
| 3414 | + ((oi + oj * jcp.ow) * jcp.oc_block) * jcp.typesize_in]; |
| 3415 | }; |
| 3416 | auto ker_addr = [&](int ic1) { |
| 3417 | return ptr[reg_ker + ic1 * jcp.oc_block * jcp.typesize_out]; |
| 3418 | }; |
| 3419 | |
| 3420 | auto emit_block = [&](int h_block_size, |
| 3421 | bool is_last_block, bool is_last_kh_kw_iter, bool is_last_row) |
| 3422 | { |
| 3423 | // TODO: add an fma version (Roma) |
| 3424 | auto pad_ow = jcp.ow; |
| 3425 | |
| 3426 | int ow4u = rnd_up(pad_ow, 4); |
| 3427 | int def_step_size = 16; |
| 3428 | |
| 3429 | bool has_w_tail = (pad_ow % def_step_size != 0 |
| 3430 | || pad_ow % 4 != 0); |
| 3431 | bool full_w_unroll = pad_ow / def_step_size < 2 + has_w_tail; |
| 3432 | |
| 3433 | auto emit_step = [&](int ur_ow, |
| 3434 | int num_inp_l1_pfs_per_fma_step, |
| 3435 | int num_inp_l2_pfs_per_fma_step, |
| 3436 | int num_out_l2_pfs_per_fma_step, bool is_w_tail) |
| 3437 | { |
| 3438 | bool block_wraparound = is_w_tail && is_last_row; |
| 3439 | |
| 3440 | assert(ur_ow % 4 == 0); |
| 3441 | int tail_size = ow4u % ur_ow; |
| 3442 | int this_ur_ow |
| 3443 | = (is_w_tail && tail_size) ? tail_size : ur_ow; |
| 3444 | int ow_last_chunk4 = pad_ow % 4; |
| 3445 | int ow_zero_tail4 = ow_last_chunk4 |
| 3446 | ? 4 - ow_last_chunk4 : 0; |
| 3447 | |
| 3448 | auto emit_out_pf = [&](int oi) { |
| 3449 | #if 1 |
| 3450 | if (oi + def_step_size < ur_ow || !block_wraparound) |
| 3451 | mic_prefetcht0(ptr[reg_out |
| 3452 | + ((def_step_size + oi) |
| 3453 | * jcp.oc_block * jcp.typesize_in)]); |
| 3454 | else { |
| 3455 | assert(block_wraparound); |
| 3456 | assert(oi + def_step_size >= ur_ow); |
| 3457 | mic_prefetcht0(ptr[reg_out_save |
| 3458 | + ((oi + def_step_size - ur_ow) |
| 3459 | * jcp.oc_block * jcp.typesize_in)]); |
| 3460 | } |
| 3461 | #else |
| 3462 | // XXX: This is an alternative prefetching strategy that |
| 3463 | // always prefetches the next row. Keeping it here for |
| 3464 | // future experiments (Roma) |
| 3465 | if (!block_wraparound) |
| 3466 | mic_prefetcht0(ptr[reg_out |
| 3467 | + (jcp.ow + oi) * jcp.oc_block * jcp.typesize_in]); |
| 3468 | else |
| 3469 | mic_prefetcht0(ptr[reg_out + reg_ohs |
| 3470 | - ((h_block_size - 1) * jcp.ow |
| 3471 | - oi) * jcp.oc_block * jcp.typesize_in]); |
| 3472 | #endif |
| 3473 | if (oi < num_out_l2_pfs_per_fma_step) |
| 3474 | mic_prefetcht1(ptr[reg_out_pf_l2 |
| 3475 | + oi * jcp.oc_block * jcp.typesize_in]); |
| 3476 | }; |
| 3477 | |
| 3478 | auto emit_inp_pf = [&](int oi4, int ic1) { |
| 3479 | int pf_slot_idx = ic1 + oi4 / 4 * jcp.ic_block; |
| 3480 | int num_pf_slots = jcp.ic_block * ur_ow / 4; |
| 3481 | |
| 3482 | int num_pfs = num_inp_l1_pfs_per_fma_step |
| 3483 | + num_inp_l2_pfs_per_fma_step; |
| 3484 | int pf_freq = nstl::max(1, num_pf_slots / num_pfs); |
| 3485 | |
| 3486 | if (pf_slot_idx % pf_freq) |
| 3487 | return; |
| 3488 | |
| 3489 | int pf_idx = pf_slot_idx / pf_freq; |
| 3490 | |
| 3491 | if (pf_idx < num_inp_l2_pfs_per_fma_step) |
| 3492 | mic_prefetcht1(ptr[reg_inp_pf_l2 |
| 3493 | + pf_idx * jcp.ic_block * jcp.typesize_in]); |
| 3494 | else { |
| 3495 | pf_idx -= num_inp_l2_pfs_per_fma_step; |
| 3496 | // prefetch the 'tail' of the cache line because most of |
| 3497 | // the accesses are not aligned |
| 3498 | mic_prefetcht0(ptr[reg_inp_pf_l1 |
| 3499 | + pf_idx * jcp.ic_block * jcp.typesize_in |
| 3500 | + cache_line_size - jcp.typesize_in]); |
| 3501 | } |
| 3502 | }; |
| 3503 | |
| 3504 | auto numloads = 4; |
| 3505 | |
| 3506 | int steps = this_ur_ow; |
| 3507 | for (int oi4 = 0; oi4 < steps; oi4 += numloads) { |
| 3508 | for (int oi1 = 0; oi1 < numloads; oi1++) { |
| 3509 | int oi = oi4 + oi1; |
| 3510 | if (!is_w_tail || oi < (this_ur_ow - ow_zero_tail4)) { |
| 3511 | vmovups(zmm_out(oi), out_addr(oi)); |
| 3512 | emit_out_pf(oi); |
| 3513 | } else { |
| 3514 | auto zmm = zmm_out(oi); |
| 3515 | vpxord(zmm, zmm, zmm); |
| 3516 | } |
| 3517 | } |
| 3518 | |
| 3519 | for (int ic1 = 0; ic1 < jcp.ic_block; ic1++) { |
| 3520 | if (jcp.ver == ver_4fma) { |
| 3521 | v4fmaddps(zmm_ker(ic1), |
| 3522 | zmm_out(oi4), inp_addr(oi4, ic1)); |
| 3523 | } else { |
| 3524 | assert(!"unknown convolution version"); |
| 3525 | } |
| 3526 | emit_inp_pf(oi4, ic1); |
| 3527 | } |
| 3528 | } |
| 3529 | }; |
| 3530 | |
| 3531 | // Input is transposed and padded but we only access about jcp.iw |
| 3532 | // elements so use that to compute the # of cache lines in each 'row' |
| 3533 | int num_inp_l1_pfs |
| 3534 | = div_up(jcp.iw * jcp.typesize_in, cache_line_size) * jcp.ic_block; |
| 3535 | |
| 3536 | if (full_w_unroll) { |
| 3537 | emit_step(ow4u, num_inp_l1_pfs, |
| 3538 | num_inp_l2_pfs_per_fma_block, |
| 3539 | num_out_l2_pfs_per_fma_block, true); |
| 3540 | add(reg_inp_pf_l2, num_inp_l2_pfs_per_fma_block * cache_line_size); |
| 3541 | add(reg_out_pf_l2, num_out_l2_pfs_per_fma_block * cache_line_size); |
| 3542 | } else { |
| 3543 | Label w_loop; |
| 3544 | int num_w_iters = pad_ow / def_step_size; |
| 3545 | int num_w_iters_full = num_w_iters + has_w_tail; |
| 3546 | int num_inp_l1_pfs_per_fma_step |
| 3547 | = div_up(num_inp_l1_pfs, num_w_iters_full); |
| 3548 | int num_inp_l2_pfs_per_fma_step |
| 3549 | = div_up(num_inp_l2_pfs_per_fma_block, num_w_iters_full); |
| 3550 | int num_out_l2_pfs_per_fma_step |
| 3551 | = div_up(num_out_l2_pfs_per_fma_block, num_w_iters_full); |
| 3552 | mov(reg_i, num_w_iters); |
| 3553 | L(w_loop); { |
| 3554 | emit_step(def_step_size, num_inp_l1_pfs_per_fma_step, |
| 3555 | num_inp_l2_pfs_per_fma_step, |
| 3556 | num_out_l2_pfs_per_fma_step, false); |
| 3557 | add(reg_inp, def_step_size * jcp.typesize_in); |
| 3558 | add(reg_out, def_step_size * jcp.oc_block * jcp.typesize_in); |
| 3559 | add(reg_inp_pf_l1, |
| 3560 | num_inp_l1_pfs_per_fma_step * cache_line_size); |
| 3561 | add(reg_inp_pf_l2, |
| 3562 | num_inp_l2_pfs_per_fma_step * cache_line_size); |
| 3563 | add(reg_out_pf_l2, |
| 3564 | num_out_l2_pfs_per_fma_step * cache_line_size); |
| 3565 | sub(reg_i, 1); |
| 3566 | jnz(w_loop); |
| 3567 | } |
| 3568 | if (has_w_tail) { |
| 3569 | emit_step(def_step_size, num_inp_l1_pfs_per_fma_step, |
| 3570 | num_inp_l2_pfs_per_fma_step, |
| 3571 | num_out_l2_pfs_per_fma_step, true); |
| 3572 | add(reg_inp_pf_l2, |
| 3573 | num_inp_l2_pfs_per_fma_step * cache_line_size); |
| 3574 | add(reg_out_pf_l2, |
| 3575 | num_out_l2_pfs_per_fma_step * cache_line_size); |
| 3576 | } |
| 3577 | // reset reg_inp and reg_out because emit_h_loop expects |
| 3578 | // unmodified pointers |
| 3579 | int w_offset = num_w_iters * def_step_size; |
| 3580 | sub(reg_inp, w_offset * jcp.typesize_in); |
| 3581 | sub(reg_out, w_offset * jcp.oc_block * jcp.typesize_in); |
| 3582 | } |
| 3583 | }; |
| 3584 | |
| 3585 | auto emit_h_loop = [&](int h_block_size, |
| 3586 | bool is_last_block, bool is_last_kh_kw_iter) |
| 3587 | { |
| 3588 | Label h_loop, skip_h_loop; |
| 3589 | mov(reg_j, 1); |
| 3590 | cmp(reg_j, reg_h); |
| 3591 | je(skip_h_loop, T_NEAR); |
| 3592 | L(h_loop); { |
| 3593 | |
| 3594 | lea(reg_inp_pf_l1, |
| 3595 | ptr[reg_inp + jcp.tr_iw * jcp.ic_block * jcp.typesize_in]); |
| 3596 | emit_block(h_block_size, |
| 3597 | is_last_block, is_last_kh_kw_iter, false); |
| 3598 | |
| 3599 | add(reg_inp, jcp.tr_iw * jcp.ic_block * jcp.typesize_in); |
| 3600 | add(reg_out, pad_ow * jcp.oc_block * jcp.typesize_in); |
| 3601 | add(reg_j, 1); |
| 3602 | cmp(reg_j, reg_h); |
| 3603 | jb(h_loop); |
| 3604 | } |
| 3605 | |
| 3606 | L(skip_h_loop); |
| 3607 | |
| 3608 | for (int ic1 = 0; ic1 < jcp.ic_block; ic1++) |
| 3609 | mic_prefetcht0(ker_addr(ic1)); |
| 3610 | |
| 3611 | lea(reg_inp_pf_l1, ptr[reg_inp_save + reg_kw * jcp.typesize_in]); |
| 3612 | emit_block(h_block_size, is_last_block, is_last_kh_kw_iter, true); |
| 3613 | }; |
| 3614 | |
| 3615 | auto emit_kh_kw_loop = [&](bool is_first_block, bool is_last_block, |
| 3616 | int h_block_size) |
| 3617 | { |
| 3618 | xor_(reg_kh, reg_kh); |
| 3619 | Label kh_loop, kh_loop_end; |
| 3620 | |
| 3621 | int last_oh_block_size |
| 3622 | = jcp.oh - rnd_up(jcp.oh - h_block_size, h_block_size); |
| 3623 | int oh_block_size = (is_last_block) ? last_oh_block_size : h_block_size; |
| 3624 | // NB1: t_pad <= oh_block_size and b_pad <= last_oh_block_size |
| 3625 | int ih_block_size = oh_block_size - 1 + jcp.kh |
| 3626 | - is_first_block * jcp.t_pad - is_last_block * jcp.b_pad; |
| 3627 | |
| 3628 | L(kh_loop); { |
| 3629 | // determine starting indices for this block |
| 3630 | if (is_first_block) { |
| 3631 | xor_(reg_tmp, reg_tmp); |
| 3632 | mov(reg_ohs, jcp.t_pad); |
| 3633 | sub(reg_ohs, reg_kh); |
| 3634 | cmovb(reg_ohs, reg_tmp); |
| 3635 | |
| 3636 | mov(reg_ihs, reg_ohs); |
| 3637 | sub(reg_ihs, jcp.t_pad); |
| 3638 | add(reg_ihs, reg_kh); |
| 3639 | } else { |
| 3640 | xor_(reg_ohs, reg_ohs); |
| 3641 | mov(reg_ihs, reg_kh); |
| 3642 | } |
| 3643 | |
| 3644 | // determine effective size of block based on padding |
| 3645 | mov(reg_tmp, oh_block_size); |
| 3646 | sub(reg_tmp, reg_ohs); |
| 3647 | mov(reg_h, ih_block_size); |
| 3648 | sub(reg_h, reg_ihs); |
| 3649 | cmp(reg_tmp, reg_h); |
| 3650 | cmovb(reg_h, reg_tmp); |
| 3651 | |
| 3652 | Label kh_loop_work; |
| 3653 | cmp(reg_h, 0); |
| 3654 | jg(kh_loop_work, T_NEAR); |
| 3655 | |
| 3656 | // empty h loop for this jcp.kh: |
| 3657 | // - set the output to 0 if necessary |
| 3658 | // - move ker pt |
| 3659 | // - jump to the end |
| 3660 | sub(reg_h, 1); |
| 3661 | Label skip_ker_zeroing; |
| 3662 | |
| 3663 | // The reg_ker ptr has highest bit set if the output needs to be |
| 3664 | // zeroed. Those who have byte-aligned their data will suffer the |
| 3665 | // consiquences :( |
| 3666 | // TODO: move the flag to a mask register? (Roma) |
| 3667 | test(reg_ker, 1); |
| 3668 | jz(skip_ker_zeroing, T_NEAR); |
| 3669 | |
| 3670 | Label zeroing_loop; |
| 3671 | vpxord(zmm0, zmm0, zmm0); |
| 3672 | and_(reg_ker, ~1); // temporarily clear the zeroing flag |
| 3673 | mov(reg_tmp, jcp.kw); |
| 3674 | L(zeroing_loop); { |
| 3675 | for (int ic1 = 0; ic1 < jcp.ic_block; ic1++) |
| 3676 | vmovups(ker_addr(ic1), zmm0); |
| 3677 | add(reg_ker, jcp.oc_block * jcp.ic_block * jcp.typesize_out); |
| 3678 | sub(reg_tmp, 1); |
| 3679 | jnz(zeroing_loop, T_NEAR); |
| 3680 | } |
| 3681 | // restore the zeroing flag (it will be cleared after the end of |
| 3682 | // emit_kh_kw_loop, but we may need it until then) |
| 3683 | or_(reg_ker, 1); |
| 3684 | jmp(kh_loop_end, T_NEAR); |
| 3685 | |
| 3686 | L(skip_ker_zeroing); |
| 3687 | add(reg_ker, jcp.oc_block * jcp.ic_block * jcp.kw |
| 3688 | * jcp.typesize_out); |
| 3689 | jmp(kh_loop_end, T_NEAR); |
| 3690 | |
| 3691 | L(kh_loop_work); |
| 3692 | |
| 3693 | mul_by_const(reg_ihs, reg_tmp, |
| 3694 | jcp.tr_iw * jcp.ic_block * jcp.typesize_in); |
| 3695 | mul_by_const(reg_ohs, reg_tmp, |
| 3696 | pad_ow * jcp.oc_block * jcp.typesize_in); |
| 3697 | |
| 3698 | add(reg_inp, reg_ihs); |
| 3699 | add(reg_out, reg_ohs); |
| 3700 | |
| 3701 | Label kw_loop; |
| 3702 | xor_(reg_kw, reg_kw); |
| 3703 | L(kw_loop); { |
| 3704 | for (int ic1 = 0; ic1 < jcp.ic_block; ic1++) { |
| 3705 | auto zmm = zmm_ker(ic1); |
| 3706 | vpxord(zmm, zmm, zmm); |
| 3707 | mic_prefetcht1(ker_addr(ic1)); |
| 3708 | } |
| 3709 | |
| 3710 | mov(reg_out_save, reg_out); |
| 3711 | mov(reg_inp_save, reg_inp); |
| 3712 | lea(reg_inp, ptr[reg_inp + reg_kw * jcp.typesize_in]); |
| 3713 | |
| 3714 | #if 0 |
| 3715 | // XXX: Generate code with special prefetches when switching |
| 3716 | // blocks or at the end of the last block. Disabled to reduce |
| 3717 | // code size and because there's no performance benefit (Roma) |
| 3718 | Label regular_h_loop, end_h_loop; |
| 3719 | cmp(reg_kw, jcp.kw - 1); |
| 3720 | jne(regular_h_loop, T_NEAR); |
| 3721 | cmp(reg_kh, jcp.kh - 1); |
| 3722 | jne(regular_h_loop, T_NEAR); |
| 3723 | |
| 3724 | emit_h_loop(oh_block_size, is_last_block, true); |
| 3725 | jmp(end_h_loop, T_NEAR); |
| 3726 | |
| 3727 | L(regular_h_loop); |
| 3728 | emit_h_loop(oh_block_size, is_last_block, false); |
| 3729 | |
| 3730 | L(end_h_loop); |
| 3731 | #else |
| 3732 | emit_h_loop(oh_block_size, is_last_block, false); |
| 3733 | #endif |
| 3734 | |
| 3735 | mov(reg_out, reg_out_save); |
| 3736 | mov(reg_inp, reg_inp_save); |
| 3737 | |
| 3738 | Label do_store; |
| 3739 | // The reg_ker ptr has highest bit set if the output needs to |
| 3740 | // be zeroed. Those who have byte-aligned their data will |
| 3741 | // suffer the consiquences :( |
| 3742 | mov(reg_tmp, reg_ker); |
| 3743 | and_(reg_ker, ~1); |
| 3744 | test(reg_tmp, 1); |
| 3745 | jnz(do_store, T_NEAR); |
| 3746 | |
| 3747 | for (int ic1 = 0; ic1 < jcp.ic_block; ic1++) { |
| 3748 | auto zmm = zmm_ker(ic1); |
| 3749 | if (jcp.ver == ver_4fma) { |
| 3750 | vaddps(zmm, ker_addr(ic1)); |
| 3751 | } else { |
| 3752 | assert(!"unknown convolution version"); |
| 3753 | } |
| 3754 | } |
| 3755 | |
| 3756 | L(do_store); |
| 3757 | for (int ic1 = 0; ic1 < jcp.ic_block; ic1++) { |
| 3758 | auto zmm = zmm_ker(ic1); |
| 3759 | vmovups(ker_addr(ic1), zmm); |
| 3760 | } |
| 3761 | |
| 3762 | mov(reg_ker, reg_tmp); |
| 3763 | add(reg_ker, jcp.ic_block * jcp.oc_block * jcp.typesize_out); |
| 3764 | add(reg_kw, 1); |
| 3765 | cmp(reg_kw, jcp.kw); |
| 3766 | jl(kw_loop); |
| 3767 | } |
| 3768 | |
| 3769 | sub(reg_inp, reg_ihs); |
| 3770 | sub(reg_out, reg_ohs); |
| 3771 | |
| 3772 | |
| 3773 | L(kh_loop_end); |
| 3774 | add(reg_kh, 1); |
| 3775 | cmp(reg_kh, jcp.kh); |
| 3776 | jl(kh_loop); |
| 3777 | } |
| 3778 | }; |
| 3779 | |
| 3780 | mov(reg_inp, ptr[param + GET_OFF(src)]); |
| 3781 | mov(reg_out, ptr[param + GET_OFF(dst)]); |
| 3782 | mov(reg_ker, ptr[param + GET_OFF(filt)]); |
| 3783 | mov(reg_inp_pf_l2, ptr[param + GET_OFF(src_prf)]); |
| 3784 | mov(reg_out_pf_l2, ptr[param + GET_OFF(dst_prf)]); |
| 3785 | mov(reg_tmp, ptr[param + GET_OFF(channel)]); |
| 3786 | or_(reg_ker, reg_tmp); |
| 3787 | |
| 3788 | bool single_kh_kw_loop = (h_block_size == jcp.oh); |
| 3789 | |
| 3790 | size_t inp_row_step = jcp.tr_iw * jcp.ic_block * jcp.typesize_in; |
| 3791 | size_t first_inp_block_step = inp_row_step * (h_block_size - jcp.t_pad); |
| 3792 | size_t inp_block_step = inp_row_step * h_block_size; |
| 3793 | size_t out_block_step = pad_ow * jcp.oc_block * jcp.typesize_in |
| 3794 | * h_block_size; |
| 3795 | |
| 3796 | if (!single_kh_kw_loop) { |
| 3797 | // Save the original prefetch pointers from the OpenMP driver |
| 3798 | vmovq(reg_inp_pf_save, reg_inp_pf_l2); |
| 3799 | vmovq(reg_out_pf_save, reg_out_pf_l2); |
| 3800 | mov(reg_inp_pf_l2, reg_inp); |
| 3801 | add(reg_inp_pf_l2, first_inp_block_step); |
| 3802 | mov(reg_out_pf_l2, reg_out); |
| 3803 | add(reg_out_pf_l2, out_block_step); |
| 3804 | } |
| 3805 | emit_kh_kw_loop(true, single_kh_kw_loop, h_block_size); |
| 3806 | |
| 3807 | if (!single_kh_kw_loop) { |
| 3808 | size_t ker_reset_offset |
| 3809 | = jcp.oc_block * jcp.ic_block * jcp.typesize_out * jcp.kw * jcp.kh; |
| 3810 | sub(reg_ker, ker_reset_offset); |
| 3811 | and_(reg_ker, ~1); // Clear the zeroing flag for subsequent updates |
| 3812 | |
| 3813 | add(reg_inp, first_inp_block_step); |
| 3814 | add(reg_out, out_block_step); |
| 3815 | mov(reg_inp_pf_l2, reg_inp); |
| 3816 | add(reg_inp_pf_l2, inp_block_step); |
| 3817 | mov(reg_out_pf_l2, reg_out); |
| 3818 | add(reg_out_pf_l2, out_block_step); |
| 3819 | |
| 3820 | int num_innermost_iters = div_up(jcp.oh, h_block_size) - 2; |
| 3821 | if (num_innermost_iters > 0) { |
| 3822 | Label h_block_loop; |
| 3823 | |
| 3824 | mov(reg_tmp_w, num_innermost_iters); |
| 3825 | kmovw(reg_h_block, reg_tmp_w); |
| 3826 | L(h_block_loop); { |
| 3827 | emit_kh_kw_loop(false, false, h_block_size); |
| 3828 | sub(reg_ker, ker_reset_offset); |
| 3829 | add(reg_inp, inp_row_step * h_block_size); |
| 3830 | add(reg_out, out_block_step); |
| 3831 | mov(reg_inp_pf_l2, reg_inp); |
| 3832 | add(reg_inp_pf_l2, inp_block_step); |
| 3833 | mov(reg_out_pf_l2, reg_out); |
| 3834 | add(reg_out_pf_l2, out_block_step); |
| 3835 | kmovw(reg_tmp_w, reg_h_block); |
| 3836 | sub(reg_tmp_w, 1); |
| 3837 | kmovw(reg_h_block, reg_tmp_w); |
| 3838 | jnz(h_block_loop); |
| 3839 | } |
| 3840 | } |
| 3841 | |
| 3842 | // Restore the original prefetch pointers that came from the OpenMP |
| 3843 | // driver |
| 3844 | vmovq(reg_inp_pf_l2, reg_inp_pf_save); |
| 3845 | vmovq(reg_out_pf_l2, reg_out_pf_save); |
| 3846 | emit_kh_kw_loop(false, true, h_block_size); |
| 3847 | } |
| 3848 | |
| 3849 | return true; |
| 3850 | } |
| 3851 | |
| 3852 | bool jit_avx512_common_conv_bwd_weights_kernel_f32 |
| 3853 | ::flat_4ops_compute() { |
| 3854 | const auto &j = jcp; |
| 3855 | const bool ok = j.ver == ver_4fma && j.is_1stconv |
| 3856 | && everyone_is(0, j.dilate_h, j.dilate_w); |
| 3857 | if (!ok) return false; |
| 3858 | |
| 3859 | Reg64 reg_ptr_tr_src = r8; |
| 3860 | Reg64 reg_ptr_dst = r9; |
| 3861 | Reg64 reg_ptr_wei = r10; |
| 3862 | Reg64 reg_ptr_bia = r11; |
| 3863 | |
| 3864 | Reg64 reg_kh_step = rax; |
| 3865 | Reg64 reg_oh = abi_not_param1; |
| 3866 | Reg64 reg_kh = rdx; |
| 3867 | |
| 3868 | Reg32 reg_flag_save = ebx; |
| 3869 | Reg32 reg_flag = esi; |
| 3870 | |
| 3871 | Zmm vbia(31); |
| 3872 | |
| 3873 | auto zmm_wei = [&](int kh, int kw) { |
| 3874 | return Zmm(8 + kh * j.kw + kw); |
| 3875 | }; |
| 3876 | auto zmm_dst = [&](int ow) { |
| 3877 | return Zmm(ow % 8); |
| 3878 | }; |
| 3879 | |
| 3880 | auto addr_tr_src = [&](int kh, int iw) { |
| 3881 | return ptr[reg_ptr_tr_src |
| 3882 | + (kh * j.stride_w * j.tr_ld + iw) * jcp.typesize_in]; |
| 3883 | }; |
| 3884 | auto addr_dst = [&](int ow) { |
| 3885 | return ptr[reg_ptr_dst + ow * jcp.oc_block * jcp.typesize_in]; |
| 3886 | }; |
| 3887 | auto addr_wei = [&](int kh, int kw) { |
| 3888 | return ptr[reg_ptr_wei + (kh * j.kw + kw) * j.oc_block |
| 3889 | * jcp.typesize_out]; |
| 3890 | }; |
| 3891 | |
| 3892 | auto emit_fma_block = [&](int kh_step) { |
| 3893 | for (int kh = 0; kh < kh_step; ++kh) { |
| 3894 | for (int kw = 0; kw < j.kw; ++kw) { |
| 3895 | auto vwei = zmm_wei(kh, kw); |
| 3896 | vpxord(vwei, vwei, vwei); |
| 3897 | } |
| 3898 | } |
| 3899 | |
| 3900 | for (int ow = 0; ow < j.ow; ow += 4) { |
| 3901 | for (int _ow = ow; _ow < ow + 4; ++_ow) { |
| 3902 | auto vdst = zmm_dst(_ow); |
| 3903 | if (_ow < j.ow) |
| 3904 | vmovups(vdst, addr_dst(_ow)); |
| 3905 | else |
| 3906 | vpxord(vdst, vdst, vdst); |
| 3907 | } |
| 3908 | |
| 3909 | for (int kh = 0; kh < kh_step; ++kh) { |
| 3910 | for (int kw = 0; kw < j.kw; ++kw) { |
| 3911 | const int iw = ow + (kw % j.stride_w) * j.tr_ld |
| 3912 | + (kw / j.stride_w); |
| 3913 | v4fmaddps(zmm_wei(kh, kw), zmm_dst(ow), |
| 3914 | addr_tr_src(kh, iw)); |
| 3915 | if (1 && kh == 0 && kw < 4) { |
| 3916 | prefetcht1(ptr[reg_ptr_dst |
| 3917 | + (j.ow + ow + kw) * jcp.oc_block |
| 3918 | * jcp.typesize_in]); |
| 3919 | } |
| 3920 | if (j.with_bias && kh_step == 1) { /* [bwd_w:b:r1] */ |
| 3921 | const int off = kw + 4 - j.kw; |
| 3922 | if (off >= 0 && ow + off < j.ow) |
| 3923 | vaddps(vbia, vbia, zmm_dst(ow + off)); |
| 3924 | } |
| 3925 | } |
| 3926 | } |
| 3927 | } |
| 3928 | |
| 3929 | Label l_store; |
| 3930 | test(reg_flag, FLAG_MB_FIRST); |
| 3931 | jnz(l_store, T_NEAR); |
| 3932 | for (int kh = 0; kh < kh_step; ++kh) { |
| 3933 | for (int kw = 0; kw < j.kw; ++kw) |
| 3934 | vaddps(zmm_wei(kh, kw), addr_wei(kh, kw)); |
| 3935 | } |
| 3936 | L(l_store); |
| 3937 | for (int kh = 0; kh < kh_step; ++kh) { |
| 3938 | for (int kw = 0; kw < j.kw; ++kw) |
| 3939 | vmovups(addr_wei(kh, kw), zmm_wei(kh, kw)); |
| 3940 | } |
| 3941 | }; |
| 3942 | |
| 3943 | auto emit_kh_loop = [&]() { |
| 3944 | const int kh_step_rem = j.kh % j.kh_step; |
| 3945 | xor_(reg_kh, reg_kh); |
| 3946 | mov(reg_kh_step, j.kh_step); |
| 3947 | |
| 3948 | Label l_kh_loop; |
| 3949 | L(l_kh_loop); { |
| 3950 | Label l_done; |
| 3951 | |
| 3952 | if (kh_step_rem != 0) { |
| 3953 | Label l_keep_kh_step; |
| 3954 | cmp(reg_kh, j.kh - j.kh_step); |
| 3955 | jle(l_keep_kh_step, T_NEAR); |
| 3956 | |
| 3957 | mov(reg_kh_step, kh_step_rem); |
| 3958 | emit_fma_block(kh_step_rem); |
| 3959 | jmp(l_done, T_NEAR); |
| 3960 | |
| 3961 | L(l_keep_kh_step); |
| 3962 | } |
| 3963 | |
| 3964 | emit_fma_block(j.kh_step); |
| 3965 | |
| 3966 | L(l_done); |
| 3967 | |
| 3968 | add(reg_ptr_tr_src, j.kh_step * j.stride_w * j.tr_ld |
| 3969 | * jcp.typesize_in); |
| 3970 | add(reg_ptr_wei, j.kh_step * j.kw * j.oc_block * jcp.typesize_out); |
| 3971 | add(reg_kh, j.kh_step); |
| 3972 | |
| 3973 | cmp(reg_kh, j.kh); |
| 3974 | jl(l_kh_loop, T_NEAR); |
| 3975 | } |
| 3976 | |
| 3977 | const int kh_steps = rnd_up(j.kh, j.kh_step); |
| 3978 | sub(reg_ptr_tr_src, kh_steps * j.stride_w * j.tr_ld * jcp.typesize_in); |
| 3979 | sub(reg_ptr_wei, kh_steps * j.kw * j.oc_block * jcp.typesize_out); |
| 3980 | }; |
| 3981 | |
| 3982 | auto emit_oh_loop = [&]() { |
| 3983 | mov(reg_oh, j.oh); |
| 3984 | |
| 3985 | Label l_oh_loop; |
| 3986 | L(l_oh_loop); { |
| 3987 | Label l_restore_mb_flag, l_jump; |
| 3988 | |
| 3989 | cmp(reg_oh, j.oh); |
| 3990 | je(l_restore_mb_flag, T_NEAR); |
| 3991 | |
| 3992 | and_(reg_flag, ~FLAG_MB_FIRST); |
| 3993 | jmp(l_jump, T_NEAR); |
| 3994 | |
| 3995 | L(l_restore_mb_flag); |
| 3996 | mov(reg_flag, reg_flag_save); |
| 3997 | |
| 3998 | L(l_jump); |
| 3999 | |
| 4000 | emit_kh_loop(); |
| 4001 | |
| 4002 | add(reg_ptr_tr_src, j.stride_h * j.stride_w * j.tr_ld |
| 4003 | * jcp.typesize_in); |
| 4004 | add(reg_ptr_dst, j.ow * j.oc_block * jcp.typesize_in); |
| 4005 | |
| 4006 | dec(reg_oh); |
| 4007 | jnz(l_oh_loop, T_NEAR); |
| 4008 | } |
| 4009 | }; |
| 4010 | |
| 4011 | auto emit_bia_store = [&]() { |
| 4012 | if (!j.with_bias) return; |
| 4013 | |
| 4014 | Label l_bia_store, l_bia_skip; |
| 4015 | test(reg_flag, FLAG_IC_FIRST); |
| 4016 | jz(l_bia_skip); |
| 4017 | |
| 4018 | test(reg_flag, FLAG_MB_FIRST); |
| 4019 | jnz(l_bia_store, T_NEAR); |
| 4020 | vaddps(vbia, ptr[reg_ptr_bia]); |
| 4021 | L(l_bia_store); |
| 4022 | vmovups(ptr[reg_ptr_bia], vbia); |
| 4023 | L(l_bia_skip); |
| 4024 | }; |
| 4025 | |
| 4026 | mov(reg_ptr_tr_src, ptr[param + GET_OFF(src)]); |
| 4027 | mov(reg_ptr_dst, ptr[param + GET_OFF(dst)]); |
| 4028 | mov(reg_ptr_wei, ptr[param + GET_OFF(filt)]); |
| 4029 | mov(reg_ptr_bia, ptr[param + GET_OFF(bias)]); |
| 4030 | mov(reg_flag_save, ptr[param + GET_OFF(flags)]); |
| 4031 | |
| 4032 | vpxord(vbia, vbia, vbia); |
| 4033 | emit_oh_loop(); |
| 4034 | emit_bia_store(); |
| 4035 | |
| 4036 | return true; |
| 4037 | } |
| 4038 | |
| 4039 | void jit_avx512_common_conv_bwd_weights_kernel_f32::compute_loop() |
| 4040 | { |
| 4041 | if (flat_4ops_compute()) |
| 4042 | return; |
| 4043 | if (compute_full_spat_loop()) |
| 4044 | return; |
| 4045 | |
| 4046 | maybe_zero_kernel(); |
| 4047 | |
| 4048 | if (jcp.ndims == 5) compute_d_loop_common(); |
| 4049 | else compute_oh_loop_common(); |
| 4050 | } |
| 4051 | |
| 4052 | void jit_avx512_common_conv_bwd_weights_kernel_f32::generate() |
| 4053 | { |
| 4054 | preamble(); |
| 4055 | |
| 4056 | mov(reg_input, ptr[param + GET_OFF(src)]); |
| 4057 | mov(reg_output, ptr[param + GET_OFF(dst)]); |
| 4058 | mov(reg_kernel, ptr[param + GET_OFF(filt)]); |
| 4059 | |
| 4060 | compute_loop(); |
| 4061 | |
| 4062 | postamble(); |
| 4063 | } |
| 4064 | |
| 4065 | status_t jit_avx512_common_conv_bwd_weights_kernel_f32::init_conf( |
| 4066 | jit_conv_conf_t &jcp, const convolution_desc_t &cd, |
| 4067 | memory_desc_t &src_md, memory_desc_t &diff_weights_md, |
| 4068 | memory_desc_t &diff_bias_md, memory_desc_t &diff_dst_md) { |
| 4069 | if (!mayiuse(avx512_common)) |
| 4070 | return status::unimplemented; |
| 4071 | |
| 4072 | const memory_desc_wrapper src_d(&src_md); |
| 4073 | const memory_desc_wrapper diff_weights_d(&diff_weights_md); |
| 4074 | const memory_desc_wrapper diff_bias_d(&diff_bias_md); |
| 4075 | const memory_desc_wrapper diff_dst_d(&diff_dst_md); |
| 4076 | |
| 4077 | const bool with_groups = diff_weights_d.ndims() == src_d.ndims() + 1; |
| 4078 | int ndims = src_d.ndims(); |
| 4079 | |
| 4080 | jcp = zero<decltype(jcp)>(); |
| 4081 | |
| 4082 | jcp.simd_w = cpu_isa_traits<avx512_common>::vlen / sizeof(float); |
| 4083 | jcp.ndims = ndims; |
| 4084 | jcp.prop_kind = cd.prop_kind; |
| 4085 | |
| 4086 | jcp.ngroups = with_groups ? diff_weights_d.dims()[0] : 1; |
| 4087 | jcp.mb = src_d.dims()[0]; |
| 4088 | |
| 4089 | jcp.oc = diff_dst_d.dims()[1] / jcp.ngroups; |
| 4090 | jcp.oc_without_padding = jcp.oc; |
| 4091 | jcp.ic = src_d.dims()[1] / jcp.ngroups; |
| 4092 | |
| 4093 | jcp.id = (ndims == 5) ? src_d.dims()[2] : 1; |
| 4094 | jcp.ih = (ndims == 3) ? 1 : src_d.dims()[ndims-2]; |
| 4095 | jcp.iw = src_d.dims()[ndims-1]; |
| 4096 | jcp.od = (ndims == 5) ? diff_dst_d.dims()[2] : 1; |
| 4097 | jcp.oh = (ndims == 3) ? 1 : diff_dst_d.dims()[ndims-2]; |
| 4098 | jcp.ow = diff_dst_d.dims()[ndims-1]; |
| 4099 | |
| 4100 | jcp.kd = (ndims == 5) ? diff_weights_d.dims()[with_groups + 2] : 1; |
| 4101 | jcp.kh = (ndims == 3) ? 1 : diff_weights_d.dims()[with_groups + ndims-2]; |
| 4102 | jcp.kw = diff_weights_d.dims()[with_groups + ndims-1]; |
| 4103 | |
| 4104 | jcp.f_pad = (ndims == 5) ? cd.padding[0][0] : 0; |
| 4105 | jcp.t_pad = (ndims == 3) ? 0 : cd.padding[0][ndims-4]; |
| 4106 | jcp.l_pad = cd.padding[0][ndims-3]; |
| 4107 | |
| 4108 | jcp.stride_d = (ndims == 5) ? cd.strides[0] : 1; |
| 4109 | jcp.stride_h = (ndims == 3) ? 1 : cd.strides[ndims-4]; |
| 4110 | jcp.stride_w = cd.strides[ndims-3]; |
| 4111 | |
| 4112 | jcp.dilate_d = (ndims == 5) ? cd.dilates[0] : 0; |
| 4113 | jcp.dilate_h = (ndims == 3) ? 0 : cd.dilates[ndims-4]; |
| 4114 | jcp.dilate_w = cd.dilates[ndims-3]; |
| 4115 | |
| 4116 | const int kh_range = 1 + (jcp.kh - 1) * (jcp.dilate_h + 1); |
| 4117 | bool ok = true |
| 4118 | // general condition to simplify dilations |
| 4119 | && IMPLICATION(jcp.dilate_d != 0, jcp.stride_d == 1) |
| 4120 | && IMPLICATION(jcp.dilate_h != 0, jcp.stride_h == 1) |
| 4121 | && IMPLICATION(jcp.dilate_w != 0, jcp.stride_w == 1) |
| 4122 | // special condition to simplify dilations in compute_oh_loop_common |
| 4123 | && IMPLICATION(jcp.dilate_h != 0, kh_range <= jcp.ih); |
| 4124 | if (!ok) |
| 4125 | return status::unimplemented; |
| 4126 | |
| 4127 | jcp.r_pad = nstl::max(0, (jcp.ow - 1) * jcp.stride_w |
| 4128 | + (jcp.kw - 1) * (jcp.dilate_w + 1) - (jcp.iw + jcp.l_pad - 1)); |
| 4129 | jcp.b_pad = nstl::max(0, (jcp.oh - 1) * jcp.stride_h |
| 4130 | + (jcp.kh - 1) * (jcp.dilate_h + 1) - (jcp.ih + jcp.t_pad - 1)); |
| 4131 | jcp.back_pad = nstl::max(0, (jcp.od - 1) * jcp.stride_d |
| 4132 | + (jcp.kd - 1) * (jcp.dilate_d + 1) - (jcp.id + jcp.f_pad - 1)); |
| 4133 | |
| 4134 | /* XXX: currently, does not support dilation_d > 0 */ |
| 4135 | if (ndims == 5) |
| 4136 | if (jcp.dilate_d > 0) |
| 4137 | return status::unimplemented; |
| 4138 | |
| 4139 | jcp.ihp = jcp.ih + jcp.t_pad + jcp.b_pad; |
| 4140 | jcp.iwp = jcp.iw + jcp.l_pad + jcp.r_pad; |
| 4141 | jcp.ohp = jcp.oh; |
| 4142 | jcp.owp = jcp.ow; |
| 4143 | jcp.aligned_threads = 0; |
| 4144 | |
| 4145 | /* check for the 1st convolution */ |
| 4146 | jcp.is_1stconv = is_1stconv(jcp); |
| 4147 | |
| 4148 | jcp.oc_block = jcp.simd_w; |
| 4149 | |
| 4150 | bool ok_to_pad_channels = true |
| 4151 | && jcp.ngroups == 1 |
| 4152 | && src_d.data_type() == data_type::f32; |
| 4153 | |
| 4154 | if (ok_to_pad_channels) |
| 4155 | jcp.oc = rnd_up(jcp.oc, jcp.simd_w); |
| 4156 | |
| 4157 | if (jcp.oc % jcp.oc_block) |
| 4158 | return status::unimplemented; |
| 4159 | |
| 4160 | auto dst_tag = pick(ndims - 3, nCw16c, nChw16c, nCdhw16c); |
| 4161 | auto wei_tag = with_groups |
| 4162 | ? pick(ndims - 3, gOIw16i16o, gOIhw16i16o, gOIdhw16i16o) |
| 4163 | : pick(ndims - 3, OIw16i16o, OIhw16i16o, OIdhw16i16o); |
| 4164 | |
| 4165 | if (diff_dst_d.format_kind() == format_kind::any) { |
| 4166 | CHECK(memory_desc_init_by_tag(diff_dst_md, dst_tag)); |
| 4167 | jcp.dst_tag = dst_tag; |
| 4168 | } else { |
| 4169 | jcp.dst_tag = diff_dst_d.matches_one_of_tag(dst_tag); |
| 4170 | } |
| 4171 | if (jcp.dst_tag != dst_tag) |
| 4172 | return status::unimplemented; |
| 4173 | |
| 4174 | /* conditions on bias memory */ |
| 4175 | jcp.with_bias = cd.diff_bias_desc.format_kind != format_kind::undef; |
| 4176 | if (jcp.with_bias) { |
| 4177 | if (diff_bias_d.format_kind() == format_kind::any) |
| 4178 | CHECK(memory_desc_init_by_tag(diff_bias_md, x)); |
| 4179 | } |
| 4180 | |
| 4181 | jcp.nb_oc = jcp.oc / jcp.oc_block; |
| 4182 | |
| 4183 | /* kernel applicability check wrt boundaries |
| 4184 | * the conditions are quite general across the kernels we have, |
| 4185 | * but ideally the check should belong to a specific kernel... */ |
| 4186 | const int max_pad = ((jcp.kh - 1) * (jcp.dilate_h + 1) + 1) / 2; |
| 4187 | const bool boundaries_ok = true |
| 4188 | && jcp.t_pad <= max_pad |
| 4189 | && jcp.b_pad <= max_pad |
| 4190 | && IMPLICATION(jcp.f_pad > 0, jcp.kd < jcp.id + jcp.f_pad) |
| 4191 | && jcp.f_pad < jcp.kd; |
| 4192 | if (!boundaries_ok) |
| 4193 | return status::unimplemented; |
| 4194 | |
| 4195 | /* yet another common check */ |
| 4196 | if (jcp.kw > 14) |
| 4197 | return status::unimplemented; |
| 4198 | |
| 4199 | /* setting register strategy */ |
| 4200 | for (int ur_w = nstl::min(max_ur_w, jcp.ow); ur_w > 0; --ur_w) { |
| 4201 | if (jcp.ow % ur_w == 0) { jcp.ur_w = ur_w; break; } |
| 4202 | } |
| 4203 | |
| 4204 | if (jcp.is_1stconv) { |
| 4205 | auto src_tag = pick(ndims - 3, ncw, nchw, ncdhw); |
| 4206 | if (src_d.format_kind() == format_kind::any) { |
| 4207 | CHECK(memory_desc_init_by_tag(src_md, src_tag)); |
| 4208 | jcp.src_tag = src_tag; |
| 4209 | } else { |
| 4210 | jcp.src_tag = src_d.matches_one_of_tag(src_tag); |
| 4211 | if (jcp.ic == 1 && jcp.src_tag != src_tag) |
| 4212 | jcp.src_tag = src_d.matches_one_of_tag( |
| 4213 | pick(ndims - 3, nwc, nhwc, ndhwc)); |
| 4214 | } |
| 4215 | if (jcp.src_tag == format_tag::undef) |
| 4216 | return status::unimplemented; |
| 4217 | |
| 4218 | const bool src_ok = true |
| 4219 | && utils::everyone_is(data_type::f32, |
| 4220 | src_d.data_type(), diff_weights_d.data_type(), |
| 4221 | diff_dst_d.data_type()) |
| 4222 | && one_of(jcp.ic, 1, 2, 3) |
| 4223 | && jcp.ngroups == 1; |
| 4224 | if (!src_ok) |
| 4225 | return status::unimplemented; |
| 4226 | |
| 4227 | const int tr_ld = rnd_up(div_up(jcp.iw + jcp.l_pad + jcp.r_pad, |
| 4228 | jcp.stride_w), 16); |
| 4229 | const int kh_step = nstl::max((28 - jcp.with_bias) / jcp.kw, 1); |
| 4230 | const int kh_step_rem = jcp.kh % kh_step; |
| 4231 | |
| 4232 | const auto wei_4fma_tag = with_groups |
| 4233 | ? pick(ndims - 3, gOiw16o, gOihw16o, gOidhw16o) |
| 4234 | : pick(ndims - 3, Oiw16o, Oihw16o, Oidhw16o); |
| 4235 | |
| 4236 | auto current_wei_tag = format_tag::undef; |
| 4237 | if (diff_weights_d.format_kind() != format_kind::any) |
| 4238 | current_wei_tag = diff_weights_d.matches_one_of_tag(wei_4fma_tag); |
| 4239 | |
| 4240 | const bool use_4fma = true |
| 4241 | && one_of(ndims, 3, 4) |
| 4242 | && mayiuse(avx512_mic_4ops) |
| 4243 | && mkldnn_thr_syncable() |
| 4244 | && everyone_is(0, jcp.dilate_d, jcp.dilate_h, jcp.dilate_w) |
| 4245 | && everyone_is(0, jcp.l_pad, jcp.r_pad, jcp.t_pad, jcp.b_pad) |
| 4246 | && jcp.kw <= 28 - jcp.with_bias |
| 4247 | && jcp.stride_w == 4 |
| 4248 | && tr_ld / jcp.simd_w <= 4 /* [bwd_w:tr_src:r1] */ |
| 4249 | && IMPLICATION(jcp.with_bias, kh_step_rem == 1) /* [bwd_w:b:r1] */ |
| 4250 | && IMPLICATION(diff_weights_d.format_kind() != format_kind::any, |
| 4251 | current_wei_tag == wei_4fma_tag); |
| 4252 | |
| 4253 | if (use_4fma) { |
| 4254 | jcp.ver = ver_4fma; |
| 4255 | jcp.kh_step = kh_step; |
| 4256 | jcp.tr_ld = tr_ld; |
| 4257 | jcp.ic_block = 1; |
| 4258 | if (diff_weights_d.format_kind() == format_kind::any) |
| 4259 | CHECK(memory_desc_init_by_tag(diff_weights_md, wei_4fma_tag)); |
| 4260 | jcp.wei_tag = wei_4fma_tag; |
| 4261 | } else { |
| 4262 | jcp.ver = ver_fma; |
| 4263 | jcp.ic_block = jcp.ic; |
| 4264 | |
| 4265 | wei_tag = with_groups |
| 4266 | ? pick(ndims - 3, gOwi16o, gOhwi16o, gOdhwi16o) |
| 4267 | : pick(ndims - 3, Owi16o, Ohwi16o, Odhwi16o); |
| 4268 | |
| 4269 | if (diff_weights_d.format_kind() == format_kind::any) { |
| 4270 | CHECK(memory_desc_init_by_tag(diff_weights_md, wei_tag)); |
| 4271 | jcp.wei_tag = wei_tag; |
| 4272 | } else { |
| 4273 | jcp.wei_tag = diff_weights_d.matches_one_of_tag(wei_tag); |
| 4274 | } |
| 4275 | if (jcp.wei_tag != wei_tag) |
| 4276 | return status::unimplemented; |
| 4277 | } |
| 4278 | |
| 4279 | jcp.nb_ic = jcp.ic / jcp.ic_block; |
| 4280 | } else { |
| 4281 | auto src_tag = pick(ndims - 3, nCw16c, nChw16c, nCdhw16c); |
| 4282 | if (src_d.format_kind() == format_kind::any) { |
| 4283 | CHECK(memory_desc_init_by_tag(src_md, src_tag)); |
| 4284 | jcp.src_tag = src_tag; |
| 4285 | } else { |
| 4286 | jcp.src_tag = src_d.matches_one_of_tag(src_tag); |
| 4287 | } |
| 4288 | if (jcp.src_tag != src_tag) |
| 4289 | return status::unimplemented; |
| 4290 | |
| 4291 | if (diff_weights_d.format_kind() == format_kind::any) { |
| 4292 | CHECK(memory_desc_init_by_tag(diff_weights_md, wei_tag)); |
| 4293 | jcp.wei_tag = wei_tag; |
| 4294 | } else { |
| 4295 | jcp.wei_tag = diff_weights_d.matches_one_of_tag(wei_tag); |
| 4296 | } |
| 4297 | if (jcp.wei_tag != wei_tag) |
| 4298 | return status::unimplemented; |
| 4299 | |
| 4300 | jcp.ic_block = jcp.simd_w; |
| 4301 | if (ok_to_pad_channels) |
| 4302 | jcp.ic = rnd_up(jcp.ic, jcp.ic_block); |
| 4303 | jcp.nb_ic = jcp.ic / jcp.ic_block; |
| 4304 | if ((mayiuse(avx512_mic) || mayiuse(avx512_core)) |
| 4305 | && utils::everyone_is(data_type::f32, |
| 4306 | src_d.data_type(), diff_weights_d.data_type(), |
| 4307 | diff_dst_d.data_type())) { |
| 4308 | jcp.ver = ver_fma; |
| 4309 | if (one_of(ndims, 3, 4) && mayiuse(avx512_mic_4ops) && jcp.stride_w == 1 && |
| 4310 | everyone_is(0, jcp.dilate_d, jcp.dilate_h, jcp.dilate_w) && |
| 4311 | mkldnn_thr_syncable()) { |
| 4312 | jcp.ver = ver_4fma; |
| 4313 | } |
| 4314 | } else { |
| 4315 | return status::unimplemented; |
| 4316 | } |
| 4317 | if (jcp.ver == ver_4fma) { |
| 4318 | jcp.ur_w = jcp.ow; |
| 4319 | // XXX, BUGBUGBUG, but not a FIXME: this assumes that it's OK to |
| 4320 | // cross the right boundary. The only requirement is not to have |
| 4321 | // NaNs there because another multiplicand is always guaranteed to |
| 4322 | // be zero. This also may require the top-level driver to allocate |
| 4323 | // four extra guarding elements at the very end of the buffer. |
| 4324 | // I'm not proud of this hack, but it improves performance by |
| 4325 | // about 5-10% depending on the dimensions (Roma) |
| 4326 | |
| 4327 | const int tr_round = 4; |
| 4328 | |
| 4329 | jcp.tr_iw = rnd_up(jcp.iw + jcp.kw - 1, tr_round); |
| 4330 | jcp.tr_src_num_guard_elems = tr_round; // upper bound |
| 4331 | } |
| 4332 | } |
| 4333 | |
| 4334 | if (utils::one_of(jcp.ver, ver_4fma, ver_fma)) { |
| 4335 | jcp.typesize_in = sizeof(float); |
| 4336 | jcp.typesize_out = sizeof(float); |
| 4337 | } else |
| 4338 | return status::unimplemented; |
| 4339 | |
| 4340 | bool args_ok = true |
| 4341 | && jcp.ic % jcp.ic_block == 0 |
| 4342 | && jcp.oc % jcp.oc_block == 0 |
| 4343 | && jcp.ic <= src_d.padded_dims()[1] |
| 4344 | && jcp.oc <= diff_dst_d.padded_dims()[1] |
| 4345 | && jcp.ic <= diff_weights_d.padded_dims()[with_groups + 1] |
| 4346 | && jcp.oc <= diff_weights_d.padded_dims()[with_groups + 0]; |
| 4347 | if (!args_ok) return status::unimplemented; |
| 4348 | |
| 4349 | { // balancing |
| 4350 | int nthr, nthr_mb, nthr_g, nthr_oc_b, nthr_ic_b; |
| 4351 | balance(jcp, nthr, nthr_mb, nthr_g, nthr_oc_b, nthr_ic_b); |
| 4352 | jcp.nthr = nthr; |
| 4353 | jcp.nthr_mb = nthr_mb; |
| 4354 | jcp.nthr_g = nthr_g; |
| 4355 | jcp.nthr_oc_b = nthr_oc_b; |
| 4356 | jcp.nthr_ic_b = nthr_ic_b; |
| 4357 | } |
| 4358 | |
| 4359 | return status::success; |
| 4360 | } |
| 4361 | |
| 4362 | void jit_avx512_common_conv_bwd_weights_kernel_f32::init_scratchpad( |
| 4363 | memory_tracking::registrar_t &scratchpad, const jit_conv_conf_t &jcp) { |
| 4364 | if (jcp.ver == ver_4fma) { |
| 4365 | if (jcp.is_1stconv) { |
| 4366 | const size_t tr_src_size = |
| 4367 | jcp.nthr / jcp.nthr_oc_b * jcp.ih * jcp.stride_w * jcp.tr_ld; |
| 4368 | scratchpad.book(key_conv_tr_src, jcp.typesize_in * tr_src_size); |
| 4369 | } else { |
| 4370 | // XXX: See the comment about tr_iw and guarding elements in |
| 4371 | // jit_avx512_common_conv_bwd_weights_kernel_f32::init_conf() |
| 4372 | const size_t max_nthr = jcp.nthr_mb * jcp.ngroups * jcp.nb_ic; |
| 4373 | const size_t min_tr_src_size_per_thr |
| 4374 | = jcp.ih * jcp.ic_block * jcp.tr_iw; |
| 4375 | const size_t tr_src_size = max_nthr * min_tr_src_size_per_thr |
| 4376 | + jcp.tr_src_num_guard_elems; |
| 4377 | scratchpad.book(key_conv_tr_src, jcp.typesize_in * tr_src_size); |
| 4378 | } |
| 4379 | |
| 4380 | /* prepare synchronization contexts */ |
| 4381 | if (jcp.nthr_oc_b > 1) { |
| 4382 | const int tr_src_bctx_size = jcp.nthr / jcp.nthr_oc_b; |
| 4383 | scratchpad.book(key_conv_tr_src_bctx, |
| 4384 | sizeof(simple_barrier::ctx_t) * tr_src_bctx_size); |
| 4385 | } |
| 4386 | } |
| 4387 | |
| 4388 | if (jcp.nthr_mb > 1) { |
| 4389 | const int wei_size = jcp.ngroups * jcp.oc * jcp.ic |
| 4390 | * jcp.kh * jcp.kw * jcp.kd; |
| 4391 | const int bia_size = jcp.ngroups * jcp.oc; |
| 4392 | const size_t wei_bia_reduction_size = wei_size + bia_size; |
| 4393 | |
| 4394 | scratchpad.book(key_conv_wei_bia_reduction, |
| 4395 | jcp.typesize_out * wei_bia_reduction_size * (jcp.nthr_mb - 1)); |
| 4396 | scratchpad.book(key_conv_wei_bia_reduction_bctx, |
| 4397 | sizeof(simple_barrier::ctx_t)); |
| 4398 | } |
| 4399 | |
| 4400 | if (jcp.with_bias && jcp.oc != jcp.oc_without_padding) |
| 4401 | scratchpad.book(key_conv_padded_bias, jcp.typesize_out * jcp.oc); |
| 4402 | } |
| 4403 | |
| 4404 | void jit_avx512_common_conv_bwd_weights_kernel_f32::balance( |
| 4405 | const jit_conv_conf_t &j, int &nthr_, int &nthr_mb_, int &nthr_g_, |
| 4406 | int &nthr_oc_b_, int &nthr_ic_b_) |
| 4407 | { |
| 4408 | nthr_ = nthr_mb_ = nthr_g_ = nthr_oc_b_ = nthr_ic_b_ = 1; |
| 4409 | |
| 4410 | const int max_threads = mkldnn_get_max_threads(); |
| 4411 | |
| 4412 | if (max_threads < j.ngroups) { |
| 4413 | /* simplification... fortunately it doesn't hurt much */ |
| 4414 | return; |
| 4415 | } |
| 4416 | |
| 4417 | if (!mkldnn_thr_syncable() && j.ver == ver_4fma) { |
| 4418 | // should not happen -- the driver is not ready |
| 4419 | // for TBB-like non-synchronous threading yet |
| 4420 | return; |
| 4421 | } |
| 4422 | |
| 4423 | if (j.ver == ver_4fma && j.is_1stconv) { |
| 4424 | nthr_g_ = 1; |
| 4425 | nthr_oc_b_ = 1; |
| 4426 | nthr_ic_b_ = nstl::min(j.nb_ic, max_threads); |
| 4427 | nthr_mb_ = nstl::min(max_threads / nthr_ic_b_, j.mb); |
| 4428 | nthr_ = nthr_mb_ * nthr_oc_b_ * nthr_ic_b_ * nthr_g_; |
| 4429 | return; |
| 4430 | } |
| 4431 | |
| 4432 | nthr_g_ = j.ngroups; |
| 4433 | const int nthr = max_threads / nthr_g_; |
| 4434 | |
| 4435 | auto calc_mem_cost = [=](int nthr_mb, int nthr_oc_b, int nthr_ic_b) { |
| 4436 | /* calculate per thread memory cost (read/write). high level optimizer |
| 4437 | * tries to minimize memory consumption. few notes: |
| 4438 | * (n1) unclear why, but that essentially helps first convolution... |
| 4439 | * (n2) assuming the reduction over minibatch is always there: |
| 4440 | * - instead of 8 it should be 5 here (write ~= 2 read): |
| 4441 | * kernel: temporal workspace 1 write |
| 4442 | * reduction: 1 read from workspace and 1 write to the diff_wei |
| 4443 | * - but experiments showed 8 works better than 5 or 6... */ |
| 4444 | |
| 4445 | const int src_coef = j.ver == ver_4fma ? 4 : 1; |
| 4446 | const int dst_coef = 1; |
| 4447 | const int wei_coef = 8; |
| 4448 | |
| 4449 | return 0 |
| 4450 | + src_coef |
| 4451 | * div_up(j.mb, nthr_mb) * div_up(j.ngroups, nthr_g_) |
| 4452 | * div_up(j.nb_ic, nthr_ic_b) * j.ic_block * j.ih * j.iw * j.id |
| 4453 | / j.stride_d / j.stride_h / j.stride_w /* (n1) */ |
| 4454 | + dst_coef |
| 4455 | * div_up(j.mb, nthr_mb) * div_up(j.ngroups, nthr_g_) |
| 4456 | * div_up(j.nb_oc, nthr_oc_b) * j.oc_block * j.oh * j.ow * j.od |
| 4457 | + wei_coef /* (n2) */ |
| 4458 | * div_up(j.ngroups, nthr_g_) |
| 4459 | * div_up(j.nb_oc, nthr_oc_b) * div_up(j.nb_ic, nthr_ic_b) |
| 4460 | * j.kh * j.kw * j.kd * j.ic_block * j.oc_block; |
| 4461 | }; |
| 4462 | |
| 4463 | int best_mem_cost = calc_mem_cost(nthr_mb_, nthr_oc_b_, nthr_ic_b_); |
| 4464 | |
| 4465 | /* step 1: find the best thread distribution with lowest memory cost */ |
| 4466 | const int nthr_mb_max = nstl::min(nthr, j.mb * j.od); |
| 4467 | for (int nthr_mb = 1; nthr_mb <= nthr_mb_max; ++nthr_mb) { |
| 4468 | const int nthr_par = nthr / nthr_mb; |
| 4469 | const int nthr_oc_b_max = nstl::min(nthr_par, j.nb_oc); |
| 4470 | for (int nthr_oc_b = 1; nthr_oc_b <= nthr_oc_b_max; ++nthr_oc_b) { |
| 4471 | int nthr_ic_b = nstl::min(nthr_par / nthr_oc_b, j.nb_ic); |
| 4472 | |
| 4473 | int mem_cost = calc_mem_cost(nthr_mb, nthr_oc_b, nthr_ic_b); |
| 4474 | if (mem_cost <= best_mem_cost) { |
| 4475 | best_mem_cost = mem_cost; |
| 4476 | nthr_mb_ = nthr_mb; |
| 4477 | nthr_oc_b_ = nthr_oc_b; |
| 4478 | nthr_ic_b_ = nthr_ic_b; |
| 4479 | } |
| 4480 | } |
| 4481 | |
| 4482 | if (!mkldnn_thr_syncable()) { assert(nthr_mb == 1); break; } |
| 4483 | } |
| 4484 | |
| 4485 | if (!mayiuse(avx512_mic)) { |
| 4486 | auto calc_comp_cost = [=](int nthr_mb, int nthr_oc_b, int nthr_ic_b) { |
| 4487 | return 1 |
| 4488 | * div_up(j.mb, nthr_mb) |
| 4489 | * div_up(j.ngroups, nthr_g_) |
| 4490 | * div_up(j.nb_oc, nthr_oc_b) |
| 4491 | * div_up(j.nb_ic, nthr_ic_b); |
| 4492 | }; |
| 4493 | |
| 4494 | /* step 2: search for a thread distribution with lower compute cost. |
| 4495 | * the constrains: |
| 4496 | * - memory cost cannot exceed 110% of the best found in the step 1 |
| 4497 | * - unless compute cost is 133% lower than the current best case |
| 4498 | * note: both constants were found empirically */ |
| 4499 | int best_comp_cost = calc_comp_cost(nthr_mb_, nthr_oc_b_, nthr_ic_b_); |
| 4500 | for (int nthr_mb = 1; nthr_mb <= nthr_mb_max; ++nthr_mb) { |
| 4501 | const int nthr_par = nthr / nthr_mb; |
| 4502 | const int nthr_oc_b_max = nstl::min(nthr_par, j.nb_oc); |
| 4503 | for (int nthr_oc_b = 1; nthr_oc_b <= nthr_oc_b_max; ++nthr_oc_b) { |
| 4504 | int nthr_ic_b = nstl::min(nthr_par / nthr_oc_b, j.nb_ic); |
| 4505 | int mem_cost = calc_mem_cost(nthr_mb, nthr_oc_b, nthr_ic_b); |
| 4506 | int comp_cost = calc_comp_cost(nthr_mb, nthr_oc_b, nthr_ic_b); |
| 4507 | |
| 4508 | const bool opt1 = comp_cost <= best_comp_cost |
| 4509 | && mem_cost < 1.1 * best_mem_cost; |
| 4510 | const bool opt2 = 4 * comp_cost <= 3 * best_comp_cost; |
| 4511 | |
| 4512 | if (opt1 || opt2) { |
| 4513 | best_comp_cost = comp_cost; |
| 4514 | nthr_mb_ = nthr_mb; |
| 4515 | nthr_oc_b_ = nthr_oc_b; |
| 4516 | nthr_ic_b_ = nthr_ic_b; |
| 4517 | } |
| 4518 | } |
| 4519 | |
| 4520 | if (!mkldnn_thr_syncable()) { assert(nthr_mb == 1); break; } |
| 4521 | } |
| 4522 | } |
| 4523 | |
| 4524 | if (nthr_mb_ > max_threads/2 && nthr_mb_ < max_threads) |
| 4525 | nthr_mb_ = nstl::min(j.mb * j.od, max_threads); |
| 4526 | nthr_ = nthr_mb_ * nthr_g_ * nthr_oc_b_ * nthr_ic_b_; |
| 4527 | |
| 4528 | assert(nthr_ <= max_threads); |
| 4529 | assert(IMPLICATION(!mkldnn_thr_syncable(), nthr_mb_ == 1)); |
| 4530 | } |
| 4531 | |
| 4532 | template struct _jit_avx512_common_conv_fwd_kernel<Zmm>; |
| 4533 | template struct _jit_avx512_common_conv_fwd_kernel<Xmm>; |
| 4534 | |
| 4535 | } |
| 4536 | } |
| 4537 | } |
| 4538 | |
| 4539 | // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s |
| 4540 |
Generated on 2023-Sep-17 from project Godot revision 4.2
Powered by 
Code Browser 2.1
Generator usage only permitted with license.