| 1 | /******************************************************************************* |
|---|---|
| 2 | * Copyright 2021-2022 Intel Corporation |
| 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | *******************************************************************************/ |
| 16 | |
| 17 | #include "dnnl_types.h" |
| 18 | |
| 19 | #include "common/bfloat16.hpp" |
| 20 | #include "common/c_types_map.hpp" |
| 21 | #include "common/dnnl_thread.hpp" |
| 22 | #include "common/memory_tracking.hpp" |
| 23 | #include "common/type_helpers.hpp" |
| 24 | #include "common/utils.hpp" |
| 25 | |
| 26 | #include "cpu/platform.hpp" |
| 27 | #include "cpu/scale_utils.hpp" |
| 28 | #include "cpu/x64/brgemm/brgemm_utils.hpp" |
| 29 | #include "cpu/x64/cpu_barrier.hpp" |
| 30 | #include "cpu/x64/cpu_isa_traits.hpp" |
| 31 | #include "cpu/x64/injectors/jit_uni_postops_injector.hpp" |
| 32 | #include "cpu/x64/jit_brgemm_conv_utils.hpp" |
| 33 | #include "cpu/x64/jit_generator.hpp" |
| 34 | |
| 35 | namespace dnnl { |
| 36 | namespace impl { |
| 37 | namespace cpu { |
| 38 | namespace x64 { |
| 39 | |
| 40 | using namespace dnnl::impl::status; |
| 41 | using namespace dnnl::impl::format_tag; |
| 42 | using namespace dnnl::impl::memory_tracking::names; |
| 43 | using namespace dnnl::impl::utils; |
| 44 | |
| 45 | using namespace prop_kind; |
| 46 | using namespace data_type; |
| 47 | |
| 48 | namespace brgemm_convolution_utils { |
| 49 | |
| 50 | bool is_any_eligible(const jit_brgemm_conv_conf_t &jcp) { |
| 51 | return (jcp.prop_kind == prop_kind::forward_inference |
| 52 | || one_of(jcp.wei_dt, data_type::s8, data_type::f16) |
| 53 | || (jcp.isa == avx2_vnni_2) || is_amx(jcp.isa)); |
| 54 | } |
| 55 | |
| 56 | inline status_t init_tag(format_tag_t &tag, memory_desc_t &md, |
| 57 | const memory_desc_wrapper &mdw, const format_tag_t tag_value, |
| 58 | bool any_eligible) { |
| 59 | |
| 60 | if (mdw.format_kind() == format_kind::any) { |
| 61 | if (any_eligible) { |
| 62 | CHECK(memory_desc_init_by_tag(md, tag_value)); |
| 63 | tag = tag_value; |
| 64 | } else { |
| 65 | tag = format_tag::undef; |
| 66 | } |
| 67 | } else { |
| 68 | tag = mdw.matches_one_of_tag(tag_value); |
| 69 | } |
| 70 | |
| 71 | if (tag != tag_value) return status::unimplemented; |
| 72 | |
| 73 | return status::success; |
| 74 | } |
| 75 | |
| 76 | bool is_amx(cpu_isa_t isa) { |
| 77 | return is_superset(isa, avx512_core_amx); |
| 78 | } |
| 79 | |
| 80 | bool post_ops_ok(jit_brgemm_conv_conf_t &jcp, primitive_attr_t &attr, |
| 81 | const memory_desc_wrapper &dst_d) { |
| 82 | using namespace injector; |
| 83 | |
| 84 | const auto &post_ops = attr.post_ops_; |
| 85 | |
| 86 | return injector::post_ops_ok(post_ops_ok_args_t(jcp.isa, |
| 87 | {sum, eltwise, binary}, post_ops, &dst_d, |
| 88 | false /*sum_at_pos_0_only*/, false /*sum_requires_scale_one*/, |
| 89 | false /*sum_requires_zp_zero*/, |
| 90 | {broadcasting_strategy_t::per_oc, broadcasting_strategy_t::scalar, |
| 91 | broadcasting_strategy_t::no_broadcast})); |
| 92 | } |
| 93 | |
| 94 | bool is_groups_ok(jit_brgemm_conv_conf_t &jcp) { |
| 95 | // Enable grouped convs for the shapes not supported in direct convs |
| 96 | // direct approach only supports int8/bf16 grouped conv |
| 97 | // when channels per groups is at least multiple of 4 |
| 98 | // and bf16 grouped conv with layout nxc on jit_bf16 impl |
| 99 | // TODO: remove this condition after the restriction on small ic is removed |
| 100 | return jcp.ngroups > 1 |
| 101 | && IMPLICATION(one_of(jcp.src_dt, u8, s8, bf16), |
| 102 | jcp.ic % 4 == 0 && jcp.oc % 4 == 0); |
| 103 | } |
| 104 | |
| 105 | status_t pick_tags(jit_brgemm_conv_conf_t &jcp, memory_desc_t &src_md, |
| 106 | memory_desc_t &weights_md, memory_desc_t &dst_md, |
| 107 | memory_desc_t &bias_md) { |
| 108 | format_tag_t src_tag, dst_tag, wei_tag; |
| 109 | dst_tag = pick(jcp.ndims - 3, nwc, nhwc, ndhwc); |
| 110 | |
| 111 | const memory_desc_wrapper src_d(&src_md); |
| 112 | const memory_desc_wrapper weights_d(&weights_md); |
| 113 | const memory_desc_wrapper dst_d(&dst_md); |
| 114 | const memory_desc_wrapper bias_d(&bias_md); |
| 115 | const bool with_groups = weights_d.ndims() == src_d.ndims() + 1; |
| 116 | const int vnni_granularity |
| 117 | = (jcp.wei_dt == f16 && jcp.isa == avx512_core_fp16) |
| 118 | ? 1 |
| 119 | : data_type_vnni_granularity(jcp.wei_dt); |
| 120 | |
| 121 | const bool is_1d = jcp.ndims == 3; |
| 122 | const bool is_2d = jcp.ndims == 4; |
| 123 | const bool is_3d = jcp.ndims == 5; |
| 124 | |
| 125 | if (jcp.wei_plain) { |
| 126 | jcp.LDB = jcp.oc; |
| 127 | if (is_3d) { |
| 128 | switch (vnni_granularity) { |
| 129 | case 1: wei_tag = with_groups ? gdhwio : dhwio; break; |
| 130 | case 2: wei_tag = with_groups ? gdhwIo2i : dhwIo2i; break; |
| 131 | case 4: wei_tag = with_groups ? gdhwIo4i : dhwIo4i; break; |
| 132 | default: return status::unimplemented; |
| 133 | } |
| 134 | } else if (is_1d) { |
| 135 | switch (vnni_granularity) { |
| 136 | case 1: wei_tag = with_groups ? gwio : wio; break; |
| 137 | case 2: wei_tag = with_groups ? gwIo2i : wIo2i; break; |
| 138 | case 4: wei_tag = with_groups ? gwIo4i : wIo4i; break; |
| 139 | default: return status::unimplemented; |
| 140 | } |
| 141 | } else { |
| 142 | assert(is_2d); |
| 143 | UNUSED(is_2d); |
| 144 | switch (vnni_granularity) { |
| 145 | case 1: wei_tag = with_groups ? ghwio : hwio; break; |
| 146 | case 2: wei_tag = with_groups ? ghwIo2i : hwIo2i; break; |
| 147 | case 4: wei_tag = with_groups ? ghwIo4i : hwIo4i; break; |
| 148 | default: return status::unimplemented; |
| 149 | } |
| 150 | } |
| 151 | } else { |
| 152 | jcp.LDB = jcp.oc_block; |
| 153 | if (jcp.oc_block == 64) { |
| 154 | if (is_3d) { |
| 155 | switch (vnni_granularity) { |
| 156 | case 1: wei_tag = with_groups ? gOdhwi64o : Odhwi64o; break; |
| 157 | case 2: |
| 158 | if (jcp.is_ic_padded) |
| 159 | wei_tag = with_groups ? gOdhwI16i64o2i |
| 160 | : OdhwI16i64o2i; |
| 161 | else |
| 162 | wei_tag = with_groups ? gOdhwI64o2i : OdhwI64o2i; |
| 163 | break; |
| 164 | case 4: |
| 165 | if (jcp.is_ic_padded) |
| 166 | wei_tag = with_groups ? gOdhwI16i64o4i |
| 167 | : OdhwI16i64o4i; |
| 168 | else |
| 169 | wei_tag = with_groups ? gOdhwI64o4i : OdhwI64o4i; |
| 170 | break; |
| 171 | default: return status::unimplemented; |
| 172 | } |
| 173 | } else if (is_1d) { |
| 174 | switch (vnni_granularity) { |
| 175 | case 1: wei_tag = with_groups ? gOwi64o : Owi64o; break; |
| 176 | case 2: |
| 177 | if (jcp.is_ic_padded) |
| 178 | wei_tag = with_groups ? gOwI16i64o2i : OwI16i64o2i; |
| 179 | else |
| 180 | wei_tag = with_groups ? gOwI64o2i : OwI64o2i; |
| 181 | break; |
| 182 | case 4: |
| 183 | if (jcp.is_ic_padded) |
| 184 | wei_tag = with_groups ? gOwI16i64o4i : OwI16i64o4i; |
| 185 | else |
| 186 | wei_tag = with_groups ? gOwI64o4i : OwI64o4i; |
| 187 | break; |
| 188 | default: return status::unimplemented; |
| 189 | } |
| 190 | } else { |
| 191 | assert(is_2d); |
| 192 | UNUSED(is_2d); |
| 193 | switch (vnni_granularity) { |
| 194 | case 1: wei_tag = with_groups ? gOhwi64o : Ohwi64o; break; |
| 195 | case 2: |
| 196 | if (jcp.is_ic_padded) |
| 197 | wei_tag = with_groups ? gOhwI16i64o2i |
| 198 | : OhwI16i64o2i; |
| 199 | else |
| 200 | wei_tag = with_groups ? gOhwI64o2i : OhwI64o2i; |
| 201 | break; |
| 202 | case 4: |
| 203 | if (jcp.is_ic_padded) |
| 204 | wei_tag = with_groups ? gOhwI16i64o4i |
| 205 | : OhwI16i64o4i; |
| 206 | else |
| 207 | wei_tag = with_groups ? gOhwI64o4i : OhwI64o4i; |
| 208 | break; |
| 209 | default: return status::unimplemented; |
| 210 | } |
| 211 | } |
| 212 | } else if (jcp.oc_block == 48) { |
| 213 | if (is_3d) { |
| 214 | switch (vnni_granularity) { |
| 215 | case 1: wei_tag = with_groups ? gOdhwi48o : Odhwi48o; break; |
| 216 | case 2: |
| 217 | if (jcp.is_ic_padded) |
| 218 | wei_tag = with_groups ? gOdhwI16i48o2i |
| 219 | : OdhwI16i48o2i; |
| 220 | else |
| 221 | wei_tag = with_groups ? gOdhwI48o2i : OdhwI48o2i; |
| 222 | break; |
| 223 | case 4: |
| 224 | if (jcp.is_ic_padded) |
| 225 | wei_tag = with_groups ? gOdhwI16i48o4i |
| 226 | : OdhwI16i48o4i; |
| 227 | else |
| 228 | wei_tag = with_groups ? gOdhwI48o4i : OdhwI48o4i; |
| 229 | break; |
| 230 | default: return status::unimplemented; |
| 231 | } |
| 232 | } else if (is_1d) { |
| 233 | switch (vnni_granularity) { |
| 234 | case 1: wei_tag = with_groups ? gOwi48o : Owi48o; break; |
| 235 | case 2: |
| 236 | if (jcp.is_ic_padded) |
| 237 | wei_tag = with_groups ? gOwI16i48o2i : OwI16i48o2i; |
| 238 | else |
| 239 | wei_tag = with_groups ? gOwI48o2i : OwI48o2i; |
| 240 | break; |
| 241 | case 4: |
| 242 | if (jcp.is_ic_padded) |
| 243 | wei_tag = with_groups ? gOwI16i48o4i : OwI16i48o4i; |
| 244 | else |
| 245 | wei_tag = with_groups ? gOwI48o4i : OwI48o4i; |
| 246 | break; |
| 247 | default: return status::unimplemented; |
| 248 | } |
| 249 | } else { |
| 250 | assert(is_2d); |
| 251 | UNUSED(is_2d); |
| 252 | switch (vnni_granularity) { |
| 253 | case 1: wei_tag = with_groups ? gOhwi48o : Ohwi48o; break; |
| 254 | case 2: |
| 255 | if (jcp.is_ic_padded) |
| 256 | wei_tag = with_groups ? gOhwI16i48o2i |
| 257 | : OhwI16i48o2i; |
| 258 | else |
| 259 | wei_tag = with_groups ? gOhwI48o2i : OhwI48o2i; |
| 260 | break; |
| 261 | case 4: |
| 262 | if (jcp.is_ic_padded) |
| 263 | wei_tag = with_groups ? gOhwI16i48o4i |
| 264 | : OhwI16i48o4i; |
| 265 | else |
| 266 | wei_tag = with_groups ? gOhwI48o4i : OhwI48o4i; |
| 267 | break; |
| 268 | default: return status::unimplemented; |
| 269 | } |
| 270 | } |
| 271 | } else if (jcp.oc_block == 32) { |
| 272 | if (is_3d) { |
| 273 | switch (vnni_granularity) { |
| 274 | case 1: wei_tag = with_groups ? gOdhwi32o : Odhwi32o; break; |
| 275 | case 2: |
| 276 | if (jcp.is_ic_padded) |
| 277 | wei_tag = with_groups ? gOdhwI16i32o2i |
| 278 | : OdhwI16i32o2i; |
| 279 | else |
| 280 | wei_tag = with_groups ? gOdhwI32o2i : OdhwI32o2i; |
| 281 | break; |
| 282 | case 4: |
| 283 | if (jcp.is_ic_padded) |
| 284 | wei_tag = with_groups ? gOdhwI16i32o4i |
| 285 | : OdhwI16i32o4i; |
| 286 | else |
| 287 | wei_tag = with_groups ? gOdhwI32o4i : OdhwI32o4i; |
| 288 | break; |
| 289 | default: return status::unimplemented; |
| 290 | } |
| 291 | } else if (is_1d) { |
| 292 | switch (vnni_granularity) { |
| 293 | case 1: wei_tag = with_groups ? gOwi32o : Owi32o; break; |
| 294 | case 2: |
| 295 | if (jcp.is_ic_padded) |
| 296 | wei_tag = with_groups ? gOwI16i32o2i : OwI16i32o2i; |
| 297 | else |
| 298 | wei_tag = with_groups ? gOwI32o2i : OwI32o2i; |
| 299 | break; |
| 300 | case 4: |
| 301 | if (jcp.is_ic_padded) |
| 302 | wei_tag = with_groups ? gOwI16i32o4i : OwI16i32o4i; |
| 303 | else |
| 304 | wei_tag = with_groups ? gOwI32o4i : OwI32o4i; |
| 305 | break; |
| 306 | default: return status::unimplemented; |
| 307 | } |
| 308 | } else { |
| 309 | assert(is_2d); |
| 310 | UNUSED(is_2d); |
| 311 | switch (vnni_granularity) { |
| 312 | case 1: wei_tag = with_groups ? gOhwi32o : Ohwi32o; break; |
| 313 | case 2: |
| 314 | if (jcp.is_ic_padded) |
| 315 | wei_tag = with_groups ? gOhwI16i32o2i |
| 316 | : OhwI16i32o2i; |
| 317 | else |
| 318 | wei_tag = with_groups ? gOhwI32o2i : OhwI32o2i; |
| 319 | break; |
| 320 | case 4: |
| 321 | if (jcp.is_ic_padded) |
| 322 | wei_tag = with_groups ? gOhwI16i32o4i |
| 323 | : OhwI16i32o4i; |
| 324 | else |
| 325 | wei_tag = with_groups ? gOhwI32o4i : OhwI32o4i; |
| 326 | break; |
| 327 | default: return status::unimplemented; |
| 328 | } |
| 329 | } |
| 330 | } else if (jcp.oc_block == 16) { |
| 331 | if (is_3d) { |
| 332 | switch (vnni_granularity) { |
| 333 | case 1: wei_tag = with_groups ? gOdhwi16o : Odhwi16o; break; |
| 334 | case 2: |
| 335 | if (jcp.is_ic_padded) |
| 336 | wei_tag = with_groups ? gOdhwI16i16o2i |
| 337 | : OdhwI16i16o2i; |
| 338 | else |
| 339 | wei_tag = with_groups ? gOdhwI16o2i : OdhwI16o2i; |
| 340 | break; |
| 341 | case 4: |
| 342 | if (jcp.is_ic_padded) |
| 343 | wei_tag = with_groups ? gOdhwI16i16o4i |
| 344 | : OdhwI16i16o4i; |
| 345 | else |
| 346 | wei_tag = with_groups ? gOdhwI16o4i : OdhwI16o4i; |
| 347 | break; |
| 348 | default: return status::unimplemented; |
| 349 | } |
| 350 | } else if (is_1d) { |
| 351 | switch (vnni_granularity) { |
| 352 | case 1: wei_tag = with_groups ? gOwi16o : Owi16o; break; |
| 353 | case 2: |
| 354 | if (jcp.is_ic_padded) |
| 355 | wei_tag = with_groups ? gOwI16i16o2i : OwI16i16o2i; |
| 356 | else |
| 357 | wei_tag = with_groups ? gOwI16o2i : OwI16o2i; |
| 358 | break; |
| 359 | case 4: |
| 360 | if (jcp.is_ic_padded) |
| 361 | wei_tag = with_groups ? gOwI16i16o4i : OwI16i16o4i; |
| 362 | else |
| 363 | wei_tag = with_groups ? gOwI16o4i : OwI16o4i; |
| 364 | break; |
| 365 | default: return status::unimplemented; |
| 366 | } |
| 367 | } else { |
| 368 | assert(is_2d); |
| 369 | UNUSED(is_2d); |
| 370 | |
| 371 | switch (vnni_granularity) { |
| 372 | case 1: wei_tag = with_groups ? gOhwi16o : Ohwi16o; break; |
| 373 | case 2: |
| 374 | if (jcp.is_ic_padded) |
| 375 | wei_tag = with_groups ? gOhwI16i16o2i |
| 376 | : OhwI16i16o2i; |
| 377 | else |
| 378 | wei_tag = with_groups ? gOhwI16o2i : OhwI16o2i; |
| 379 | break; |
| 380 | case 4: |
| 381 | if (jcp.is_ic_padded) |
| 382 | wei_tag = with_groups ? gOhwI16i16o4i |
| 383 | : OhwI16i16o4i; |
| 384 | else |
| 385 | wei_tag = with_groups ? gOhwI16o4i : OhwI16o4i; |
| 386 | break; |
| 387 | default: return status::unimplemented; |
| 388 | } |
| 389 | } |
| 390 | } else if (jcp.oc_block == 8) { |
| 391 | if (vnni_granularity == 1) |
| 392 | wei_tag = with_groups ? gOhwi8o : Ohwi8o; |
| 393 | else |
| 394 | return status::unimplemented; |
| 395 | } else { |
| 396 | return status::unimplemented; |
| 397 | } |
| 398 | } |
| 399 | |
| 400 | src_tag = dst_tag; |
| 401 | |
| 402 | const bool any_eligible = is_any_eligible(jcp); |
| 403 | CHECK(init_tag(jcp.src_tag, src_md, src_d, src_tag, any_eligible)); |
| 404 | CHECK(init_tag(jcp.dst_tag, dst_md, dst_d, dst_tag, any_eligible)); |
| 405 | CHECK(init_tag(jcp.wei_tag, weights_md, weights_d, wei_tag, true)); |
| 406 | |
| 407 | return status::success; |
| 408 | } |
| 409 | |
| 410 | struct brg_blocking_t : public jit_brgemm_conv_conf_t { |
| 411 | struct array_in_loop_t { |
| 412 | dim_t itersize; |
| 413 | float repeatn; |
| 414 | float overlap; |
| 415 | void set(dim_t iter_s, float rpt, float ovlp = 1.f) { |
| 416 | itersize = iter_s; |
| 417 | repeatn = rpt; |
| 418 | overlap = ovlp; |
| 419 | } |
| 420 | }; |
| 421 | |
| 422 | struct loop_t { |
| 423 | array_in_loop_t src; |
| 424 | array_in_loop_t wei; |
| 425 | array_in_loop_t dst; |
| 426 | }; |
| 427 | |
| 428 | brg_blocking_t() { |
| 429 | // TODO: This is a broken form of initialization for a base class. |
| 430 | // Either set default values in a base class, or provide a proper |
| 431 | // default ctor, or take a `jit_brgemm_conv_conf_t` object to initialize |
| 432 | // a base class object. |
| 433 | jit_brgemm_conv_conf_t *base |
| 434 | = static_cast<jit_brgemm_conv_conf_t *>(this); |
| 435 | *base = jit_brgemm_conv_conf_t(); |
| 436 | init(); |
| 437 | } |
| 438 | brg_blocking_t(const jit_brgemm_conv_conf_t &jcp) |
| 439 | : jit_brgemm_conv_conf_t(jcp) { |
| 440 | init(); |
| 441 | } |
| 442 | void init() { |
| 443 | ur = 0; |
| 444 | ur_block = 0; |
| 445 | ur_block_tail = 0; |
| 446 | eff = 0.f; |
| 447 | nb_kd = 0; |
| 448 | nb_kh = 0; |
| 449 | nb_kw = 0; |
| 450 | sp = 0; |
| 451 | sp_block = 0; |
| 452 | nb_sp = 0; |
| 453 | eff = 0; |
| 454 | max_regs = isa_num_vregs(isa); |
| 455 | bcast_simd = acc_simd_w; |
| 456 | } |
| 457 | |
| 458 | int ur, ur_block, ur_block_tail; |
| 459 | int nb_kd, nb_kh, nb_kw; |
| 460 | int max_regs; |
| 461 | int bcast_simd; |
| 462 | float eff; |
| 463 | static unsigned L1; |
| 464 | static unsigned L2; |
| 465 | static unsigned L3; |
| 466 | // These are rough estimates of the latency (relative) of access to various |
| 467 | // cache levels. This is enough for an estimation of data access cost. |
| 468 | // TODO: Improve memory access estimates |
| 469 | static constexpr float L1_k = 1.f; |
| 470 | static constexpr float L2_k = 3.f; |
| 471 | static constexpr float L3_k = 15.f; |
| 472 | // TODO: At the moment, we are primarily evaluating the fit of the data into |
| 473 | // the L1/L2. Need to take into account the difference between the L3 and |
| 474 | // memory. |
| 475 | static constexpr float mem_k = 15.f; |
| 476 | static constexpr int bench_iterations = 1; |
| 477 | |
| 478 | int sp, sp_block, nb_sp; |
| 479 | static int last_ic_block_size; |
| 480 | |
| 481 | void get_from_jcp(const jit_brgemm_conv_conf_t &jcp) { *this = jcp; } |
| 482 | void save_to_jcp(jit_brgemm_conv_conf_t &jcp) const { jcp = *this; } |
| 483 | |
| 484 | status_t estimate_brgemm_ur(); |
| 485 | status_t get_brgemm_ur( |
| 486 | const primitive_attr_t *attr, const memory_desc_t &dst_md); |
| 487 | |
| 488 | float io_k(dim_t src, dim_t wei, dim_t dst, float n, float pk, |
| 489 | bool is_broadcast, bool is_shared) const; |
| 490 | |
| 491 | float io_k(const loop_t loop, const array_in_loop_t arr, float pk, |
| 492 | bool is_broadcast, bool is_shared) const; |
| 493 | |
| 494 | void select_ic_block(); |
| 495 | |
| 496 | void update_blocks(); |
| 497 | bool fast_check_oc_block() const; |
| 498 | float est_eff(); |
| 499 | void iterate_ker_block(brg_blocking_t &best_brgb, int kd_block, |
| 500 | int kh_block, bool maybe_use_buffer, int max_ow_block_thr); |
| 501 | status_t calc_blocks(); |
| 502 | |
| 503 | bool fast_check_oc_block_1x1() const; |
| 504 | float est_eff_1x1(); |
| 505 | void calc_blocks_1x1(); |
| 506 | |
| 507 | // utils |
| 508 | static int get_inp_size( |
| 509 | int max_src_size, int dst_size, int k, int stride, int dilate) { |
| 510 | auto adj_str = nstl::min(k, stride); |
| 511 | const auto res = nstl::min(max_src_size, |
| 512 | calculate_end_padding(0, dst_size, 0, adj_str, |
| 513 | calculate_extended_filter_size(k, dilate))); |
| 514 | return res; |
| 515 | } |
| 516 | |
| 517 | static float squeeze_val(float eff, float koeff) { |
| 518 | if (koeff <= 0) return 1; |
| 519 | if (koeff == 1) return eff; |
| 520 | const auto k = 1.f / koeff; |
| 521 | return (k > 1.f) ? (k - 1 + eff) / k : eff * koeff; |
| 522 | } |
| 523 | |
| 524 | static int estimate_ur(int oc_block) { |
| 525 | const auto est_ur = (oc_block == 64) |
| 526 | ? 6 |
| 527 | : ((oc_block == 48) ? 9 : ((oc_block == 32) ? 14 : 28)); |
| 528 | return est_ur; |
| 529 | } |
| 530 | |
| 531 | int inp_w(int out_w, int ker_w) const { |
| 532 | return get_inp_size(iw, out_w, ker_w, stride_w, dilate_w); |
| 533 | } |
| 534 | |
| 535 | int rnd_simd(int val) const { return rnd_up(val, simd_w); } |
| 536 | |
| 537 | int rnd_inp_simd(int out_w, int ker_w, int vic) const { |
| 538 | const auto vsp = inp_w(out_w, ker_w); |
| 539 | return ((stride_w == 1 && vic >= ic) ? rnd_up(vsp * vic, simd_w) |
| 540 | : vsp * rnd_up(vic, simd_w)); |
| 541 | } |
| 542 | |
| 543 | static constexpr int MAXNLOOPS = 32; |
| 544 | loop_t loop[MAXNLOOPS]; |
| 545 | }; |
| 546 | |
| 547 | unsigned brg_blocking_t::L1; |
| 548 | unsigned brg_blocking_t::L2; |
| 549 | unsigned brg_blocking_t::L3; |
| 550 | int brg_blocking_t::last_ic_block_size; |
| 551 | |
| 552 | float brg_blocking_t::io_k(dim_t src, dim_t wei, dim_t dst, float n, float pk, |
| 553 | bool is_broadcast, bool is_shared) const { |
| 554 | if (n < 1) return 0; |
| 555 | if (n == 1) return pk; |
| 556 | const auto amount = src * src_dsz + wei * wei_dsz + dst * dst_dsz |
| 557 | + (use_buffer ? dst * acc_dsz : 0); |
| 558 | const auto amount_L1 = is_broadcast ? src * src_dsz : amount; |
| 559 | const auto k = is_broadcast |
| 560 | ? ((amount_L1 < L1) ? L1_k |
| 561 | : ((amount < L2) ? L2_k |
| 562 | : (is_shared ? L3_k : mem_k))) |
| 563 | : ((amount < L2) ? L2_k : (is_shared ? L3_k : mem_k)); |
| 564 | const auto cost = pk + k * (n - 1); |
| 565 | return cost / n; |
| 566 | } |
| 567 | |
| 568 | float brg_blocking_t::io_k(const loop_t loop, const array_in_loop_t arr, |
| 569 | float pk, bool is_broadcast, bool is_shared) const { |
| 570 | return io_k(loop.src.itersize, loop.wei.itersize, loop.dst.itersize, |
| 571 | arr.repeatn * arr.overlap, pk, is_broadcast, is_shared); |
| 572 | } |
| 573 | |
| 574 | void brg_blocking_t::select_ic_block() { |
| 575 | if (is_1x1 && is_amx(isa)) { |
| 576 | // TODO: merge with non-1x1 code block below |
| 577 | const int ic_padded_block = 16 * brg_blocking_t::last_ic_block_size; |
| 578 | assert(IMPLICATION( |
| 579 | !is_bf32, ic < ic_padded_block || ic % ic_padded_block == 0)); |
| 580 | MAYBE_UNUSED(ic_padded_block); |
| 581 | // Note: bf32 requires ic_block be less than 64, otherwise it results |
| 582 | // in incorrect output. |
| 583 | ic_block = is_bf32 && (!is_rtus) ? nstl::min(64, ic) : ic; |
| 584 | nb_ic = utils::div_up(ic, ic_block); // trivially 1 for now |
| 585 | return; |
| 586 | } |
| 587 | auto nb_simd = utils::div_up(ic, simd_w); |
| 588 | auto max_simd_blocks = nstl::min(5 * simd_w, nb_simd); |
| 589 | const auto nb_icb_eff_threshold = 0.5f; |
| 590 | const auto padded_ic = last_ic_block_size * (is_ic_padded ? acc_simd_w : 1); |
| 591 | if (is_amx(isa)) { |
| 592 | if (ic * kw_sets < simd_w) { |
| 593 | // this is current requirement from brgemm kernel |
| 594 | ic_block = rnd_up(ic, last_ic_block_size); |
| 595 | } else if (is_bf32) { |
| 596 | ic_block = simd_w; |
| 597 | } else { |
| 598 | if (exec_type == exec_trans) { |
| 599 | auto simd_blocks = 1; |
| 600 | for (int nb_icb = max_simd_blocks; nb_icb >= 1; nb_icb--) { |
| 601 | auto nb_icb_eff = static_cast<float>(nb_simd) |
| 602 | / rnd_up(nb_simd, nb_icb); |
| 603 | if (nb_icb_eff >= nb_icb_eff_threshold) { |
| 604 | simd_blocks = nb_icb; |
| 605 | break; |
| 606 | } |
| 607 | } |
| 608 | ic_block = simd_blocks * simd_w; |
| 609 | } else |
| 610 | ic_block = simd_w; |
| 611 | } |
| 612 | } else { |
| 613 | const auto est_ur = nstl::min(sp_block, estimate_ur(oc_block)); |
| 614 | const auto inp_ur = is_os_blocking ? est_ur : inp_w(est_ur, kw_block); |
| 615 | |
| 616 | if (kw_block > 1) { |
| 617 | // try to fit src into L1 |
| 618 | const auto inp_per_ic = static_cast<unsigned int>(inp_ur) * src_dsz; |
| 619 | max_simd_blocks = saturate(1, max_simd_blocks, |
| 620 | static_cast<int>(L1 / (inp_per_ic * simd_w))); |
| 621 | } |
| 622 | // try to fit all batch for ur into L2 |
| 623 | const auto wei_per_ic = static_cast<unsigned int>(kd_block) * kh_block |
| 624 | * kw_block * oc_block * wei_dsz; |
| 625 | const auto inp_per_ic = static_cast<unsigned int>(kd_block) * kh_block |
| 626 | * inp_ur * src_dsz; |
| 627 | const auto out_size |
| 628 | = static_cast<unsigned int>(ur) * oc_block * dst_dsz; |
| 629 | |
| 630 | max_simd_blocks = saturate(1, max_simd_blocks, |
| 631 | static_cast<int>((L2 - out_size) |
| 632 | / ((wei_per_ic + inp_per_ic) * simd_w))); |
| 633 | |
| 634 | auto simd_blocks = 1; |
| 635 | for (int nb_icb = nstl::min(max_simd_blocks, nb_simd); nb_icb >= 1; |
| 636 | nb_icb--) { |
| 637 | auto nb_icb_eff |
| 638 | = static_cast<float>(nb_simd) / rnd_up(nb_simd, nb_icb); |
| 639 | if (nb_icb_eff >= nb_icb_eff_threshold) { |
| 640 | simd_blocks = nb_icb; |
| 641 | break; |
| 642 | } |
| 643 | } |
| 644 | |
| 645 | ic_block = nstl::min( |
| 646 | (exec_type == exec_trans) ? rnd_up(ic, padded_ic) : ic, |
| 647 | simd_blocks * simd_w); |
| 648 | } |
| 649 | nb_ic = utils::div_up(ic, ic_block); |
| 650 | } |
| 651 | |
| 652 | status_t brg_blocking_t::estimate_brgemm_ur() { |
| 653 | // Simple simulation of brgemm_desc init |
| 654 | if (sp_block <= 0) return status::invalid_arguments; |
| 655 | LDA = is_rtus |
| 656 | ? (ic_block) |
| 657 | : (kh_sets > 1 ? kh_sets : 1) * (kw_sets > 1 ? kw_sets : stride_w) |
| 658 | * (exec_type == exec_trans ? ic_block |
| 659 | : ngroups * ic_without_padding); |
| 660 | LDB = oc_block; |
| 661 | LDC = use_buffer ? oc_block : oc_without_padding; |
| 662 | |
| 663 | // Configure matrix sizes |
| 664 | // for amx if ic_block != ic then we use exec_trans so K is ic_block |
| 665 | const auto padded_ic = last_ic_block_size * (is_ic_padded ? acc_simd_w : 1); |
| 666 | |
| 667 | icp = rnd_up(ic, padded_ic); |
| 668 | M = brgM = sp >= sp_block ? sp_block : 0; |
| 669 | M_tail = brgM_tail = sp % sp_block; |
| 670 | if (is_os_blocking) { |
| 671 | if (!is_1x1) M_tail = brgM_tail = (oh * ow) % sp_block; |
| 672 | oskip = ((ext_kw - 1) / stride_w) * stride_h + (stride_h - 1) * ow; |
| 673 | |
| 674 | brgM = sp_block + oskip * (div_up(M, ow) - 1); |
| 675 | |
| 676 | // round up brgM to help brgemm kernel use max amx_h as brgemm bd_block |
| 677 | if (use_M_mask == 2) { |
| 678 | int ibrgM = 0; |
| 679 | const auto adj_ow = ow_block + oskip; |
| 680 | while (ibrgM < brgM) { |
| 681 | if (ibrgM % adj_ow < ow_block) |
| 682 | ibrgM += amx_h; |
| 683 | else |
| 684 | ibrgM++; |
| 685 | } |
| 686 | brgM = ibrgM; |
| 687 | } else |
| 688 | brgM = rnd_up(brgM, amx_h); |
| 689 | |
| 690 | brgM_tail = brgM; |
| 691 | } |
| 692 | |
| 693 | N = oc >= oc_block ? oc_block : 0; |
| 694 | N_tail = oc % oc_block; |
| 695 | |
| 696 | K = kh_sets * kw_sets * (ic >= ic_block ? ic_block : 0); |
| 697 | K_tail = kh_sets * kw_sets |
| 698 | * (exec_type == exec_trans && (!is_bf32) |
| 699 | ? ic_block |
| 700 | : rnd_up(ic % ic_block, last_ic_block_size)); |
| 701 | |
| 702 | const auto vK = K > 0 ? K : K_tail; |
| 703 | const auto vM = M > 0 ? M : M_tail; |
| 704 | const auto vN = N > 0 ? N : N_tail; |
| 705 | |
| 706 | const float alpha = 1.0; |
| 707 | const float beta = 0.0; |
| 708 | brgemm_t brg; |
| 709 | brgemm_utils::init_brgemm_conf(&brg, isa, brgemm_addr, src_dt, wei_dt, |
| 710 | brgemm_row_major, alpha, beta, LDA, LDB, LDC, vM, vN, vK, nullptr, |
| 711 | is_bf32); |
| 712 | CHECK(brgemm_utils::brgemm_blocking(&brg)); |
| 713 | ur = brg.bd_block * (is_amx(isa) ? brg.bd_block2 : 1); |
| 714 | ur_block = brg.bd_block; |
| 715 | if (is_1x1 && is_amx(isa) && M > 0 && M_tail > 0) { |
| 716 | brgemm_t brg_sp_tail; |
| 717 | brgemm_utils::init_brgemm_conf(&brg_sp_tail, isa, brgemm_addr, src_dt, |
| 718 | wei_dt, brgemm_row_major, alpha, beta, LDA, LDB, LDC, M_tail, |
| 719 | vN, vK, nullptr, is_bf32); |
| 720 | CHECK(brgemm_utils::brgemm_blocking(&brg_sp_tail)); |
| 721 | ur_block_tail = brg_sp_tail.bd_block; |
| 722 | } else { |
| 723 | ur_block_tail = 0; |
| 724 | } |
| 725 | return status::success; |
| 726 | } |
| 727 | |
| 728 | status_t brg_blocking_t::get_brgemm_ur( |
| 729 | const primitive_attr_t *attr, const memory_desc_t &dst_md) { |
| 730 | // Detailed simulation of brgemm convolution init |
| 731 | if (sp_block <= 0 || ic_block <= 0 || oc_block <= 0) |
| 732 | return status::invalid_arguments; |
| 733 | CHECK(estimate_brgemm_ur()); |
| 734 | |
| 735 | LDD = oc_without_padding; |
| 736 | |
| 737 | const float alpha = 1.0; |
| 738 | const float beta = 1.0; |
| 739 | const float beta_init = 0.0; |
| 740 | |
| 741 | for (int i = 0; i < M; i++) { |
| 742 | auto vM = i + 1; |
| 743 | // init only needed brgemm descriptors |
| 744 | if ((utils::one_of(exec_type, exec_trans, exec_vpad) || is_1x1) |
| 745 | && vM != M && vM != M_tail) |
| 746 | continue; |
| 747 | for (int i_init = 0; i_init < 2; i_init++) { |
| 748 | for (int i_N = 0; i_N < 2; i_N++) { |
| 749 | for (int i_K = 0; i_K < 2; i_K++) { |
| 750 | auto vbeta = (i_init) ? beta_init : beta; |
| 751 | auto vN = (i_N) ? N_tail : N; |
| 752 | auto vK = (i_K) ? K_tail : K; |
| 753 | if (vN == 0 || vK == 0) continue; |
| 754 | brgemm_t brg; |
| 755 | brgemm_strides_t brg_strides; |
| 756 | brg_strides.stride_a = ngroups * ic_without_padding |
| 757 | * (dilate_w + 1) * src_dsz; |
| 758 | //weights are padded by oc_block and last_ic_block |
| 759 | brg_strides.stride_b = rnd_up(ic, last_ic_block_size) |
| 760 | * rnd_up(oc, oc_block) * wei_dsz; |
| 761 | const auto strides_ptr = (brg_type == brgemm_strd) |
| 762 | ? &brg_strides |
| 763 | : nullptr; |
| 764 | brgemm_utils::init_brgemm_conf(&brg, isa, brg_type, src_dt, |
| 765 | wei_dt, brgemm_row_major, alpha, vbeta, LDA, LDB, |
| 766 | LDC, vM, vN, vK, strides_ptr, is_bf32); |
| 767 | CHECK(brgemm_utils::brgemm_blocking(&brg)); |
| 768 | |
| 769 | brgemm_attr_t brgattr; |
| 770 | brgattr.max_bs = max_batch; |
| 771 | const auto max_vpad = (exec_type == exec_vpad) |
| 772 | ? nstl::max(l_pad, r_pad) |
| 773 | : 0; |
| 774 | brgattr.max_top_vpad = max_vpad; |
| 775 | brgattr.max_bottom_vpad = max_vpad; |
| 776 | brgattr.fpmath_mode = attr->fpmath_mode_; |
| 777 | CHECK(brgemm_desc_set_attr(&brg, brgattr)); |
| 778 | |
| 779 | brg.with_sum = with_sum; |
| 780 | CHECK(brgemm_desc_set_postops( |
| 781 | &brg, attr, &dst_md, LDD, bia_dt)); |
| 782 | } |
| 783 | } |
| 784 | } |
| 785 | } |
| 786 | |
| 787 | return status::success; |
| 788 | } |
| 789 | |
| 790 | void brg_blocking_t::update_blocks() { |
| 791 | if (sp_block <= 0 |
| 792 | || utils::one_of(0, od_block, oh_block, ic_block, oc_block, |
| 793 | kd_block, kh_block, kw_block, os_block, ow_block)) |
| 794 | return; |
| 795 | |
| 796 | nb_od = div_up(od, od_block); |
| 797 | nb_oh = div_up(oh, oh_block); |
| 798 | nb_ic = div_up(ic, ic_block); |
| 799 | nb_oc = div_up(oc, oc_block); |
| 800 | nb_kd = div_up(kd, kd_block); |
| 801 | nb_kh = div_up(kh, kh_block); |
| 802 | nb_kw = div_up(kw, kw_block); |
| 803 | nb_ow = div_up(ow, ow_block); |
| 804 | if (is_os_blocking) { |
| 805 | nb_os = div_up(os, os_block); |
| 806 | sp = os; |
| 807 | sp_block = os_block; |
| 808 | nb_sp = nb_os; |
| 809 | } else { |
| 810 | sp = ow; |
| 811 | sp_block = ow_block; |
| 812 | nb_sp = nb_ow; |
| 813 | iw_block = get_inp_size(iwp, ow_block, kw, stride_w, dilate_w); |
| 814 | } |
| 815 | } |
| 816 | |
| 817 | bool brg_blocking_t::fast_check_oc_block() const { |
| 818 | // This function for reducing the number of blocking variants |
| 819 | // TODO: eliminate heuristic in this function |
| 820 | const auto rnd_oc = rnd_up(oc, acc_simd_w); |
| 821 | auto res = false; |
| 822 | if (oc_block == 64) { |
| 823 | res = (rnd_oc % oc_block == 0 && rnd_oc * wei_dsz < 192 * 4); |
| 824 | } else if (oc_block == 48) { |
| 825 | const bool big_spatial |
| 826 | = id * ih * iw > 81 * stride_d * stride_h * stride_w; |
| 827 | res = (rnd_oc % oc_block == 0 && rnd_oc * wei_dsz <= 384 * 4 |
| 828 | && big_spatial); |
| 829 | } else |
| 830 | res = true; |
| 831 | |
| 832 | return res; |
| 833 | } |
| 834 | |
| 835 | float brg_blocking_t::est_eff() { |
| 836 | const auto ocblock = oc_block / acc_simd_w; |
| 837 | |
| 838 | const auto brgemm_microkernel_eff |
| 839 | = (static_cast<float>(ocblock) * ur) / ((ur + ocblock) * max_regs); |
| 840 | |
| 841 | const auto ur_eff = static_cast<float>(sp_block) / rnd_up(sp_block, ur); |
| 842 | const auto brgemm_eff = squeeze_val(ur |
| 843 | * (2.f - nstl::min(1.9f, static_cast<float>(ur) / sp_block)) |
| 844 | / 64, |
| 845 | 0.5f); |
| 846 | |
| 847 | const auto sp_amount = nb_od * nb_oh * nb_sp; |
| 848 | const auto work_amount = mb * ngroups * nb_oc * sp_amount; |
| 849 | const auto sp_eff = (static_cast<float>(sp) / rnd_up(sp, sp_block)); |
| 850 | |
| 851 | const auto thr_eff = static_cast<float>(work_amount) |
| 852 | / utils::rnd_up(work_amount, nthr); |
| 853 | |
| 854 | const auto oc_block_eff = static_cast<float>(oc) / rnd_up(oc, oc_block); |
| 855 | |
| 856 | const auto job = div_up(work_amount, nthr); |
| 857 | |
| 858 | auto job_eff = 1.f; |
| 859 | if (job < nthr) { |
| 860 | std::vector<dim_t> thr_jobs(nthr); |
| 861 | |
| 862 | for (int ithr = 0; ithr < nthr; ithr++) { |
| 863 | thr_jobs[ithr] = 0; |
| 864 | if (ithr >= work_amount) continue; |
| 865 | dim_t thr_job = 0; |
| 866 | int start {0}, end {0}; |
| 867 | balance211(work_amount, nthr, ithr, start, end); |
| 868 | int n {0}, g {0}, ocb {0}, odp {0}, ohp {0}, spb {0}; |
| 869 | if (loop_order == loop_ndhwgc) |
| 870 | nd_iterator_init(start, n, mb, odp, od, ohp, oh, spb, nb_sp, g, |
| 871 | ngroups, ocb, nb_oc); |
| 872 | else if (loop_order == loop_ngcdhw) |
| 873 | nd_iterator_init(start, n, mb, g, ngroups, ocb, nb_oc, odp, od, |
| 874 | ohp, oh, spb, nb_sp); |
| 875 | |
| 876 | for (auto work = start; work < end; work++) { |
| 877 | const int ocp = ocb * oc_block; |
| 878 | const auto oc_sz = nstl::min(oc - ocp, oc_block); |
| 879 | int sp_sz = 0; |
| 880 | const int spp = spb * sp_block; |
| 881 | sp_sz = nstl::min(sp - spp, sp_block); |
| 882 | thr_job += sp_sz * oc_sz; |
| 883 | |
| 884 | if (loop_order == loop_ndhwgc) |
| 885 | nd_iterator_step(n, mb, odp, od, ohp, oh, spb, nb_sp, g, |
| 886 | ngroups, ocb, nb_oc); |
| 887 | else if (loop_order == loop_ngcdhw) |
| 888 | nd_iterator_step(n, mb, g, ngroups, ocb, nb_oc, odp, od, |
| 889 | ohp, oh, spb, nb_sp); |
| 890 | } |
| 891 | thr_jobs[ithr] = thr_job; |
| 892 | } |
| 893 | |
| 894 | dim_t max_job = 0; |
| 895 | dim_t sum_job = 0; |
| 896 | for (int ithr = 0; ithr < nthr; ithr++) { |
| 897 | if (thr_jobs[ithr] > max_job) max_job = thr_jobs[ithr]; |
| 898 | sum_job += thr_jobs[ithr]; |
| 899 | } |
| 900 | job_eff = max_job == 0 ? 1 |
| 901 | : static_cast<float>(sum_job) / (max_job * nthr); |
| 902 | |
| 903 | } else { |
| 904 | job_eff = thr_eff; |
| 905 | } |
| 906 | |
| 907 | const auto ic_blocking_size = ic_block * nb_ic_blocking; |
| 908 | const auto oc_blocking_size = oc_block * ic_blocking_size; |
| 909 | |
| 910 | int l = -1; |
| 911 | |
| 912 | // -- brgemm kernel: loop by simd_w -- |
| 913 | l++; |
| 914 | const auto inp_ur = inp_w(ur, kw_block); |
| 915 | loop[l].src.set(inp_ur * simd_w, 1, bcast_simd); |
| 916 | loop[l].dst.set(0, 1); |
| 917 | loop[l].wei.set(oc_block, 1); |
| 918 | |
| 919 | // -- brgemm kernel: loop by kw in kw_block -- |
| 920 | l++; |
| 921 | auto src_is = rnd_inp_simd(ur, kw_block, ic_blocking_size); |
| 922 | loop[l].src.set(src_is, 1, kw_block); |
| 923 | loop[l].dst.set(0, 1); |
| 924 | loop[l].wei.set(oc_blocking_size, 1); |
| 925 | |
| 926 | // -- brgemm kernel: loop by batch (grouped by kw_block) in ur -- |
| 927 | l++; |
| 928 | loop[l].src.set(src_is, 1); |
| 929 | loop[l].dst.set(0, 1); |
| 930 | auto wei_is = kw_block * oc_blocking_size; |
| 931 | loop[l].wei.set(wei_is, 1); |
| 932 | // -- brgemm kernel: loop by ur in sp_block -- |
| 933 | l++; |
| 934 | const auto nb_ur = div_up(sp_block, ur); |
| 935 | loop[l].src.set(kd_block * kh_block * src_is, 1); |
| 936 | loop[l].dst.set(ur * oc_block, 1); |
| 937 | wei_is = kd_block * kh_block * kw_block * oc_blocking_size; |
| 938 | loop[l].wei.set(wei_is, nb_ur); |
| 939 | |
| 940 | // -- harness: loop by k_blocks in ks -- |
| 941 | l++; |
| 942 | loop[l].src.set(kd_block * kh_block |
| 943 | * rnd_inp_simd(sp_block, kw_block, ic_blocking_size), |
| 944 | 1); |
| 945 | loop[l].dst.set(sp_block * oc_block, nb_kd * nb_kh * nb_kw); |
| 946 | loop[l].wei.set(wei_is, 1); |
| 947 | |
| 948 | // -- brgemm kernel: loop by ic_chunks -- |
| 949 | l++; |
| 950 | const auto ic_chunks = div_up(nb_ic, nb_ic_blocking); |
| 951 | loop[l].src.set(kd * kh * rnd_inp_simd(sp_block, kw, ic_blocking_size), 1); |
| 952 | loop[l].dst.set(sp_block * oc_block, ic_chunks); |
| 953 | wei_is = kd * kh * kw * oc_blocking_size; |
| 954 | loop[l].wei.set(wei_is, 1); |
| 955 | |
| 956 | const auto dim_oc = (loop_order == loop_ndhwgc) ? 1 : sp_amount; |
| 957 | const auto nb_oc_thr = nstl::min(nb_oc, div_up(job, dim_oc)); |
| 958 | const auto oc_thr = nstl::min(oc, nb_oc_thr * oc_block); |
| 959 | const auto nsimd_oc_thr = div_up(oc_thr, simd_w); |
| 960 | |
| 961 | const auto dim_sp = (loop_order == loop_ndhwgc) ? ngroups * nb_oc : 1; |
| 962 | const auto nb_sp_thr = nstl::min(nb_sp, div_up(job, dim_sp)); |
| 963 | const auto sp_thr = nstl::min(sp, nb_sp_thr * sp_block); |
| 964 | |
| 965 | int nb_oh_thr {1}, oh_thr {1}, nb_od_thr {1}, od_thr {1}; |
| 966 | if (!is_os_blocking) { |
| 967 | const auto dim_oh = nb_sp * dim_sp; |
| 968 | nb_oh_thr = nstl::min(nb_oh, div_up(job, dim_oh)); |
| 969 | oh_thr = nstl::min(oh, nb_oh_thr * oh_block); |
| 970 | |
| 971 | const auto dim_od = nb_oh * dim_oh; |
| 972 | nb_od_thr = nstl::min(nb_od, div_up(job, dim_od)); |
| 973 | od_thr = nstl::min(od, nb_od_thr * od_block); |
| 974 | } |
| 975 | |
| 976 | src_is = kd * kh * rnd_inp_simd(sp_block, kw, ic); |
| 977 | |
| 978 | auto wei_op = kd * kh * kw * ocblock * ic; |
| 979 | if (loop_order == loop_ndhwgc) { |
| 980 | // -- harness: loop by oc_block -- |
| 981 | l++; |
| 982 | loop[l].src.set(src_is, nb_oc_thr); |
| 983 | loop[l].dst.set(sp_block * oc_block, 1); |
| 984 | wei_is = kd * kh * kw * oc_block * ic; |
| 985 | wei_op = kd * kh * kw * nsimd_oc_thr * ic; |
| 986 | loop[l].wei.set(wei_is, 1); |
| 987 | } |
| 988 | |
| 989 | // -- harness: loop by sp_blocks -- |
| 990 | l++; |
| 991 | loop[l].src.set(src_is, 1); |
| 992 | const auto rnd_oc_for_sp |
| 993 | = simd_w * ((loop_order == loop_ndhwgc) ? nsimd_oc_thr : ocblock); |
| 994 | loop[l].dst.set(sp_block * rnd_oc_for_sp, 1); |
| 995 | loop[l].wei.set(wei_op * simd_w, nb_sp_thr); |
| 996 | // oh_block almost all is 1. TODO: manage oh_block != 1 |
| 997 | // -- harness: loop by oh_blocks -- |
| 998 | l++; |
| 999 | src_is = kd * kh * rnd_inp_simd(sp_thr, kw, ic); |
| 1000 | loop[l].src.set(oh_block * src_is, 1); |
| 1001 | loop[l].dst.set(sp_thr * rnd_oc_for_sp, 1); |
| 1002 | loop[l].wei.set(wei_op * simd_w, nb_oh_thr); |
| 1003 | // od_block almost all is 1. TODO: manage oh_block != 1 |
| 1004 | // -- harness: loop by od_blocks -- |
| 1005 | l++; |
| 1006 | loop[l].src.set(od_block * oh_thr * src_is, 1); |
| 1007 | loop[l].dst.set(oh_thr * sp_thr * rnd_oc_for_sp, 1); |
| 1008 | loop[l].wei.set(wei_op * simd_w, nb_od_thr); |
| 1009 | |
| 1010 | if (loop_order != loop_ndhwgc) { |
| 1011 | // -- harness: loop by oc_block -- |
| 1012 | l++; |
| 1013 | loop[l].src.set(od_thr * oh_thr * src_is, nb_oc_thr); |
| 1014 | loop[l].dst.set(oc_block * od_thr * oh_thr * sp_thr, 1); |
| 1015 | loop[l].wei.set(kd * kh * kw * oc_block * ic, 1); |
| 1016 | } |
| 1017 | |
| 1018 | // -- harness: loop by mb -- |
| 1019 | l++; |
| 1020 | const auto mb_thr = nstl::min(mb, div_up(job, sp_amount * ngroups * nb_oc)); |
| 1021 | loop[l].src.set(od_thr * oh_thr * src_is, 1); |
| 1022 | loop[l].dst.set(od_thr * oh_thr * sp_thr * nsimd_oc_thr * simd_w, 1); |
| 1023 | loop[l].wei.set(kd * kh * kw * nsimd_oc_thr * simd_w * ic, mb_thr); |
| 1024 | |
| 1025 | const auto src_op = static_cast<dim_t>(mb_thr) * od_thr * oh_thr * sp_thr |
| 1026 | * kd * kh * kw * ic; |
| 1027 | const auto dst_op = static_cast<dim_t>(mb_thr) * od_thr * oh_thr * sp_thr |
| 1028 | * nsimd_oc_thr; |
| 1029 | wei_op = kd * kh * kw * nsimd_oc_thr * ic; |
| 1030 | |
| 1031 | // for "real" application set bench_iterations to 1 |
| 1032 | const auto iterations = bench_iterations; |
| 1033 | l++; |
| 1034 | loop[l].src.set(src_op, iterations); |
| 1035 | loop[l].dst.set(dst_op * simd_w, iterations); |
| 1036 | loop[l].wei.set(wei_op * simd_w, iterations); |
| 1037 | |
| 1038 | auto src_mem_k = mem_k; |
| 1039 | auto dst_mem_k = mem_k; |
| 1040 | auto wei_mem_k = mem_k; |
| 1041 | float src_rp = 1; |
| 1042 | float dst_rp = 1; |
| 1043 | float wei_rp = 1; |
| 1044 | |
| 1045 | for (auto il = l; il >= 0; il--) { |
| 1046 | src_mem_k = io_k(loop[il], loop[il].src, src_mem_k, true, |
| 1047 | loop_order == loop_ndhwgc ? false : true); |
| 1048 | dst_mem_k = io_k(loop[il], loop[il].dst, dst_mem_k, false, false); |
| 1049 | wei_mem_k = io_k(loop[il], loop[il].wei, wei_mem_k, false, |
| 1050 | loop_order == loop_ndhwgc ? true : false); |
| 1051 | src_rp *= loop[il].src.repeatn; |
| 1052 | dst_rp *= loop[il].dst.repeatn; |
| 1053 | wei_rp *= loop[il].wei.repeatn; |
| 1054 | } |
| 1055 | const auto src_ops = (src_op * src_rp) / iterations; |
| 1056 | const auto dst_ops = (dst_op * dst_rp) / iterations; |
| 1057 | const auto wei_ops = (wei_op * wei_rp) / iterations; |
| 1058 | |
| 1059 | const auto src_cost = src_mem_k * src_ops; |
| 1060 | const auto dst_cost = dst_mem_k * dst_ops; |
| 1061 | const auto wei_cost = wei_mem_k * wei_ops; |
| 1062 | const auto call_kernel_cost |
| 1063 | = 1000.f * job * ic_chunks * nb_kd * nb_kh * nb_kw; |
| 1064 | |
| 1065 | const auto cache_eff = (static_cast<dim_t>(mb) * od * oh * sp * ic * oc) |
| 1066 | / (nthr * (src_cost + dst_cost + wei_cost + call_kernel_cost)); |
| 1067 | const auto res_eff = oc_block_eff * brgemm_microkernel_eff * sp_eff |
| 1068 | * job_eff * ur_eff * cache_eff * brgemm_eff; |
| 1069 | return res_eff; |
| 1070 | } |
| 1071 | |
| 1072 | void brg_blocking_t::iterate_ker_block(brg_blocking_t &best_brgb, int kd_block_, |
| 1073 | int kh_block_, bool maybe_use_buffer, int max_ow_block_thr) { |
| 1074 | |
| 1075 | unsigned est_k_amount = ic * oc_block * wei_dsz; |
| 1076 | |
| 1077 | kd_block = kd_block_; |
| 1078 | kh_block = kh_block_; |
| 1079 | if (one_of(exec_type, exec_vpad, exec_trans)) { |
| 1080 | kw_block = kw; |
| 1081 | kd_block_pad = kd_block; |
| 1082 | kh_block_pad = kh_block; |
| 1083 | kw_block_pad = kw_block; |
| 1084 | } else { |
| 1085 | kw_block = (est_k_amount * kw < L2) ? kw : 1; |
| 1086 | kd_block_pad = kh_block >= kd ? kd : 1; |
| 1087 | kh_block_pad = kw_block >= kh ? kh : 1; |
| 1088 | kw_block_pad = kw; |
| 1089 | } |
| 1090 | |
| 1091 | if (exec_type == exec_vpad) { |
| 1092 | od_block = 1; |
| 1093 | oh_block = 1; |
| 1094 | } else if (exec_type == exec_trans) { |
| 1095 | const auto w_block_size |
| 1096 | = 2 * src_dsz * ic * iwp + dst_dsz * ow * oc_block; |
| 1097 | const auto other_size = wei_dsz * kd * kh * kw * ic * oc_block |
| 1098 | + acc_dsz * 2 * amx_h * oc_block; |
| 1099 | const auto L2_available = nstl::min(static_cast<size_t>(div_up(L2, 2)), |
| 1100 | other_size > L2 ? 0 : L2 - other_size); |
| 1101 | if (idp * ihp * w_block_size > L2_available) { |
| 1102 | od_block = utils::saturate( |
| 1103 | 1, od, int(L2_available / (ihp * w_block_size))); |
| 1104 | if (od_block == 1) |
| 1105 | oh_block = utils::saturate( |
| 1106 | 1, oh, int(L2_available / (w_block_size))); |
| 1107 | else |
| 1108 | oh_block = oh; |
| 1109 | } else { |
| 1110 | od_block = 1; |
| 1111 | oh_block = oh; |
| 1112 | } |
| 1113 | if (is_amx(isa)) { |
| 1114 | // try to fit into L1 |
| 1115 | bool L1_fit_res = false; |
| 1116 | auto cur_od_block = od_block; |
| 1117 | auto cur_oh_block = oh_block; |
| 1118 | const auto src_w_block_size |
| 1119 | = src_dsz * ic * iwp + dst_dsz * ow * oc_block; |
| 1120 | if (src_w_block_size < L1) { |
| 1121 | cur_od_block = utils::saturate( |
| 1122 | 1, od, int(L1 / (ihp * src_w_block_size))); |
| 1123 | if (cur_od_block == 1) |
| 1124 | cur_oh_block = utils::saturate( |
| 1125 | 1, oh, int(L1 / (src_w_block_size))); |
| 1126 | } |
| 1127 | for (; cur_od_block > 1; cur_od_block--) { |
| 1128 | const auto sp_size = cur_od_block * cur_oh_block * iwp; |
| 1129 | if ((static_cast<float>(od) / rnd_up(od, cur_od_block)) > 0.9f |
| 1130 | && static_cast<float>(sp_size) / rnd_up(sp, amx_h) |
| 1131 | > 0.8f) { |
| 1132 | L1_fit_res = true; |
| 1133 | break; |
| 1134 | } |
| 1135 | } |
| 1136 | if (cur_od_block == 1) { |
| 1137 | for (; cur_oh_block > 1; cur_oh_block--) { |
| 1138 | const auto sp_size = cur_oh_block * iwp; |
| 1139 | if ((static_cast<float>(oh) / rnd_up(oh, cur_oh_block)) |
| 1140 | > 0.9f |
| 1141 | && sp_size > 128) { |
| 1142 | L1_fit_res = true; |
| 1143 | break; |
| 1144 | } |
| 1145 | } |
| 1146 | } |
| 1147 | if (L1_fit_res) { |
| 1148 | od_block = cur_od_block; |
| 1149 | oh_block = cur_oh_block; |
| 1150 | } |
| 1151 | } |
| 1152 | |
| 1153 | // limit oh_block to have good threading |
| 1154 | const auto thr_oc_block = div_up( |
| 1155 | nthr, mb * div_up((oc > 32 ? ngroups : 1) * oc, oc_block)); |
| 1156 | const auto thr_od_block = div_up(od, thr_oc_block); |
| 1157 | const auto thr_oh_block |
| 1158 | = div_up(oh, thr_oc_block * div_up(od, thr_od_block)); |
| 1159 | od_block = nstl::min(od_block, thr_od_block); |
| 1160 | oh_block = nstl::min(oh_block, thr_oh_block); |
| 1161 | } else { |
| 1162 | od_block = 1; |
| 1163 | oh_block = 1; |
| 1164 | } |
| 1165 | |
| 1166 | // --- Select ow_block ---- |
| 1167 | const auto max_ow_block_L2 = ow; |
| 1168 | auto start_ow_block = nstl::min(max_ow_block_thr, max_ow_block_L2); |
| 1169 | |
| 1170 | sp = ow; |
| 1171 | const auto start_sp_block = is_os_blocking ? ow : start_ow_block; |
| 1172 | auto prev_spb = 0; |
| 1173 | for (auto ns = 1; ns <= sp; ns++) { |
| 1174 | const auto spb = div_up(sp, ns); |
| 1175 | if (spb == prev_spb || spb > start_sp_block) continue; |
| 1176 | if (is_os_blocking && spb != ow) continue; |
| 1177 | prev_spb = spb; |
| 1178 | ow_block = spb; |
| 1179 | sp_block = ow_block; |
| 1180 | |
| 1181 | select_ic_block(); |
| 1182 | |
| 1183 | use_buffer = maybe_use_buffer |
| 1184 | && (ic_block * nb_ic_blocking < ic || kd_block != kd |
| 1185 | || kh_block != kh || kw_block != kw |
| 1186 | || kd_block_pad != kd || kh_block_pad != kh |
| 1187 | || kw_block_pad != kw); |
| 1188 | if (exec_type == exec_base) |
| 1189 | use_buffer = use_buffer || (maybe_use_buffer && iwp != iw); |
| 1190 | |
| 1191 | const status_t st = estimate_brgemm_ur(); |
| 1192 | if (st != status::success) continue; |
| 1193 | os_block = sp_block = ow_block; |
| 1194 | update_blocks(); |
| 1195 | |
| 1196 | eff = est_eff(); |
| 1197 | |
| 1198 | if (eff > best_brgb.eff || best_brgb.eff == 0) best_brgb = *this; |
| 1199 | } |
| 1200 | } |
| 1201 | |
| 1202 | status_t brg_blocking_t::calc_blocks() { |
| 1203 | sp = ow; |
| 1204 | |
| 1205 | nb_ic_blocking = 1; |
| 1206 | // --- Select kernel blocking --- |
| 1207 | // if dst_dt != acc_dt and we need to store intermediate |
| 1208 | // results then we need the out buffer |
| 1209 | const auto maybe_use_buffer = (dst_dt != acc_dt || with_sum); |
| 1210 | |
| 1211 | std::vector<int> kd_blocks(1), kh_blocks(1); |
| 1212 | kd_blocks[0] = kd; |
| 1213 | kh_blocks[0] = kh; |
| 1214 | if (kd != 1) { |
| 1215 | kd_blocks.resize(2); |
| 1216 | kd_blocks[1] = 1; |
| 1217 | } |
| 1218 | if (kh != 1) { |
| 1219 | kh_blocks.resize(2); |
| 1220 | kh_blocks[1] = 1; |
| 1221 | } |
| 1222 | |
| 1223 | const auto thr_eff_threshold = 0.9f; |
| 1224 | const auto max_ow_block_thr = utils::saturate(1, ow, |
| 1225 | static_cast<int>(div_up( |
| 1226 | mb * ngroups * nb_oc * os, thr_eff_threshold * nthr))); |
| 1227 | |
| 1228 | ow_block = os_block = sp_block = -1; |
| 1229 | brg_blocking_t best_brgb = *this; |
| 1230 | for (const auto &kd_block : kd_blocks) { |
| 1231 | for (const auto &kh_block : kh_blocks) { |
| 1232 | iterate_ker_block(best_brgb, kd_block, kh_block, maybe_use_buffer, |
| 1233 | max_ow_block_thr); |
| 1234 | } |
| 1235 | } |
| 1236 | *this = best_brgb; |
| 1237 | if (!IMPLICATION(!is_os_blocking, sp_block > 0)) |
| 1238 | return status::unimplemented; |
| 1239 | |
| 1240 | if (is_os_blocking) { |
| 1241 | ow_block = ow; |
| 1242 | os_block = ow * oh_block; |
| 1243 | sp_block = os_block; |
| 1244 | ow_tail = 0; |
| 1245 | } else { |
| 1246 | ow_block = os_block = sp_block; |
| 1247 | ow_tail = ow % ow_block; |
| 1248 | } |
| 1249 | update_blocks(); |
| 1250 | return status::success; |
| 1251 | } |
| 1252 | |
| 1253 | bool brg_blocking_t::fast_check_oc_block_1x1() const { |
| 1254 | // This function for reducing the number of blocking variants |
| 1255 | // TODO: eliminate heuristic in this function |
| 1256 | if (is_1x1 && is_amx(isa)) return true; |
| 1257 | const auto rnd_oc = rnd_up(oc, acc_simd_w); |
| 1258 | auto res = false; |
| 1259 | if (oc_block == 64) { |
| 1260 | const auto big_spatial |
| 1261 | = od * oh * ow >= 64 * stride_d * stride_h * stride_w; |
| 1262 | res = (rnd_oc % oc_block == 0 && big_spatial); |
| 1263 | } else if (oc_block == 48) { |
| 1264 | const auto oc_block_eff = static_cast<float>(oc) / rnd_up(oc, oc_block); |
| 1265 | res = (oc_block_eff >= 0.95f); |
| 1266 | } else |
| 1267 | res = true; |
| 1268 | |
| 1269 | return res; |
| 1270 | } |
| 1271 | |
| 1272 | float brg_blocking_t::est_eff_1x1() { |
| 1273 | const auto ocblock = oc_block / acc_simd_w; |
| 1274 | |
| 1275 | auto calc_ave_blk = [&](int dim, int block, bool use_ave) -> float { |
| 1276 | const int nb = dim / block; |
| 1277 | constexpr int max_nb = 2; // only consider 2x2 tile blocking |
| 1278 | const int block2 = nstl::min(max_nb, nb); |
| 1279 | const int nb2 = nb / block2; |
| 1280 | const int nb2_tail = nb % block2; |
| 1281 | if (!use_ave) return block2; |
| 1282 | return (float(nb2) * block2 + nb2_tail) / div_up(nb, block2); |
| 1283 | }; |
| 1284 | const bool use_ocb_ave = true; |
| 1285 | const auto ocb_ave = calc_ave_blk(oc_block, acc_simd_w, use_ocb_ave); |
| 1286 | const bool use_spb_ave = false; |
| 1287 | const auto spb_ave = calc_ave_blk(sp_block, ur_block, use_spb_ave); |
| 1288 | const auto M_n_sp_blks = ur_block > 0 ? nstl::max(M, M_tail) / ur_block : 0; |
| 1289 | const auto M_tail_n_sp_blks |
| 1290 | = ur_block_tail > 0 ? M_tail / ur_block_tail : 0; |
| 1291 | |
| 1292 | // heuristic for maskrcnn workaround: use old blocking for some convolutions |
| 1293 | // TODO: remove this condition |
| 1294 | const bool maskrcnn_cond = (ic == 1024 && oc == 2048) |
| 1295 | || (ic == 1024 && oc == 512) || (ic == 256 && oc == 1024) |
| 1296 | || (ic == 512 && oc == 1024) || (ic == 512 && oc == 2048); |
| 1297 | const auto amx_fac = maskrcnn_cond |
| 1298 | ? (div_up(M + M_tail, 16) / (M_n_sp_blks + M_tail_n_sp_blks)) |
| 1299 | : (static_cast<float>(div_up(M + M_tail, 16)) |
| 1300 | / (M_n_sp_blks + M_tail_n_sp_blks)); |
| 1301 | |
| 1302 | const auto brgemm_microkernel_eff = is_amx(isa) |
| 1303 | ? amx_fac * (static_cast<float>(ocb_ave) * spb_ave) |
| 1304 | / (ocb_ave + spb_ave) |
| 1305 | : (static_cast<float>(ocblock) * ur) / ((ur + ocblock) * max_regs); |
| 1306 | const auto ur_eff = static_cast<float>(sp_block) / rnd_up(sp_block, ur); |
| 1307 | const auto brgemm_eff = squeeze_val(ur |
| 1308 | * (2.f - nstl::min(1.9f, static_cast<float>(ur) / sp_block)) |
| 1309 | / 64, |
| 1310 | 0.5f); |
| 1311 | |
| 1312 | const auto sp_amount = is_os_blocking ? div_up(nb_os, nb_os_blocking) |
| 1313 | : nb_od * nb_oh * nb_sp; |
| 1314 | const auto work_amount = mb * ngroups * nb_oc * sp_amount; |
| 1315 | |
| 1316 | const auto sp_eff = static_cast<float>(sp) / rnd_up(sp, sp_block); |
| 1317 | const auto thr_eff = static_cast<float>(work_amount) |
| 1318 | / utils::rnd_up(work_amount, nthr); |
| 1319 | const auto oc_block_eff = static_cast<float>(oc) / rnd_up(oc, oc_block); |
| 1320 | |
| 1321 | const auto job = div_up(work_amount, nthr); |
| 1322 | |
| 1323 | const auto dim_oc = (loop_order == loop_ndhwgc) ? 1 : sp_amount; |
| 1324 | const auto nb_oc_thr = nstl::min(nb_oc, div_up(job, dim_oc)); |
| 1325 | const auto oc_thr = nstl::min(oc, nb_oc_thr * oc_block); |
| 1326 | const auto nsimd_oc_thr = div_up(oc_thr, simd_w); |
| 1327 | |
| 1328 | const auto dim_sp = (loop_order == loop_ndhwgc) ? ngroups * nb_oc : 1; |
| 1329 | const auto nb_sp_thr = nstl::min(nb_sp, div_up(job, dim_sp)); |
| 1330 | const auto sp_thr = nstl::min(sp, nb_sp_thr * sp_block); |
| 1331 | |
| 1332 | int nb_oh_thr {1}, oh_thr {1}, nb_od_thr {1}, od_thr {1}; |
| 1333 | if (!is_os_blocking) { |
| 1334 | const auto dim_oh = nb_sp * dim_sp; |
| 1335 | nb_oh_thr = nstl::min(nb_oh, div_up(job, dim_oh)); |
| 1336 | oh_thr = nstl::min(oh, nb_oh_thr * oh_block); |
| 1337 | |
| 1338 | const auto dim_od = nb_oh * dim_oh; |
| 1339 | nb_od_thr = nstl::min(nb_od, div_up(job, dim_od)); |
| 1340 | od_thr = nstl::min(od, nb_od_thr * od_block); |
| 1341 | } |
| 1342 | |
| 1343 | auto job_eff = 1.f; |
| 1344 | if (job < nthr) { |
| 1345 | std::vector<dim_t> thr_jobs(nthr); |
| 1346 | for (int ithr = 0; ithr < nthr; ithr++) { |
| 1347 | thr_jobs[ithr] = 0; |
| 1348 | if (ithr >= work_amount) continue; |
| 1349 | dim_t thr_job = 0; |
| 1350 | int start {0}, end {0}; |
| 1351 | balance211(work_amount, nthr, ithr, start, end); |
| 1352 | int n {0}, g {0}, ocb {0}, oss {0}, odp {0}, ohp {0}, spb {0}; |
| 1353 | if (loop_order == loop_ndhwgc) { |
| 1354 | if (is_os_blocking) |
| 1355 | nd_iterator_init(start, n, mb, oss, sp_amount, g, ngroups, |
| 1356 | ocb, nb_oc); |
| 1357 | else |
| 1358 | nd_iterator_init(start, n, mb, odp, od, ohp, oh, spb, nb_sp, |
| 1359 | g, ngroups, ocb, nb_oc); |
| 1360 | } else if (loop_order == loop_ngcdhw) { |
| 1361 | if (is_os_blocking) |
| 1362 | nd_iterator_init(start, n, mb, g, ngroups, ocb, nb_oc, oss, |
| 1363 | sp_amount); |
| 1364 | else |
| 1365 | nd_iterator_init(start, n, mb, g, ngroups, ocb, nb_oc, odp, |
| 1366 | od, ohp, oh, spb, nb_sp); |
| 1367 | } |
| 1368 | |
| 1369 | for (auto work = start; work < end; work++) { |
| 1370 | const int ocp = ocb * oc_block; |
| 1371 | const auto oc_sz = nstl::min(oc - ocp, oc_block); |
| 1372 | int sp_sz = 0; |
| 1373 | if (is_os_blocking) { |
| 1374 | const auto osb_start = oss * nb_os_blocking; |
| 1375 | const auto osb_range |
| 1376 | = nstl::min(nb_os - osb_start, nb_os_blocking); |
| 1377 | for (int osb = 0; osb < osb_range; osb++) { |
| 1378 | const int osp = (osb_start + osb) * sp_block; |
| 1379 | sp_sz = nstl::min(os - osp, sp_block); |
| 1380 | } |
| 1381 | } else { |
| 1382 | const int spp = spb * sp_block; |
| 1383 | sp_sz = nstl::min(sp - spp, sp_block); |
| 1384 | } |
| 1385 | thr_job += sp_sz * oc_sz; |
| 1386 | |
| 1387 | if (loop_order == loop_ndhwgc) { |
| 1388 | if (is_os_blocking) |
| 1389 | nd_iterator_step( |
| 1390 | n, mb, oss, sp_amount, g, ngroups, ocb, nb_oc); |
| 1391 | else |
| 1392 | nd_iterator_step(n, mb, odp, od, ohp, oh, spb, nb_sp, g, |
| 1393 | ngroups, ocb, nb_oc); |
| 1394 | } else if (loop_order == loop_ngcdhw) { |
| 1395 | if (is_os_blocking) |
| 1396 | nd_iterator_step( |
| 1397 | n, mb, g, ngroups, ocb, nb_oc, oss, sp_amount); |
| 1398 | else |
| 1399 | nd_iterator_step(n, mb, g, ngroups, ocb, nb_oc, odp, od, |
| 1400 | ohp, oh, spb, nb_sp); |
| 1401 | } |
| 1402 | } |
| 1403 | thr_jobs[ithr] = thr_job; |
| 1404 | } |
| 1405 | |
| 1406 | dim_t max_job = 0; |
| 1407 | dim_t sum_job = 0; |
| 1408 | for (int ithr = 0; ithr < nthr; ithr++) { |
| 1409 | if (thr_jobs[ithr] > max_job) max_job = thr_jobs[ithr]; |
| 1410 | sum_job += thr_jobs[ithr]; |
| 1411 | } |
| 1412 | |
| 1413 | job_eff = max_job == 0 ? 1 |
| 1414 | : static_cast<float>(sum_job) / (max_job * nthr); |
| 1415 | } else { |
| 1416 | job_eff = thr_eff; |
| 1417 | } |
| 1418 | |
| 1419 | const auto ic_blocking_size = ic_block * nb_ic_blocking; |
| 1420 | const auto oc_blocking_size = oc_block * ic_blocking_size; |
| 1421 | |
| 1422 | int l = -1; |
| 1423 | // -- brgemm kernel: loop by simd_w -- |
| 1424 | l++; |
| 1425 | loop[l].src.set(ur * simd_w, 1, bcast_simd); |
| 1426 | loop[l].dst.set(0, 1); |
| 1427 | loop[l].wei.set(oc_block, 1); |
| 1428 | |
| 1429 | // -- brgemm kernel: loop by ur in sp_block -- |
| 1430 | l++; |
| 1431 | const auto nb_ur = div_up(sp_block, ur); |
| 1432 | const auto nb_sp_no_tail = sp / sp_block; |
| 1433 | const auto sp_block_tail = sp % sp_block; |
| 1434 | const auto nb_ur_average |
| 1435 | = (nb_sp_no_tail * nb_ur + div_up(sp_block_tail, ur)) / nb_sp; |
| 1436 | loop[l].src.set(ur * rnd_simd(ic_blocking_size), 1); |
| 1437 | loop[l].dst.set(ur * oc_block, 1); |
| 1438 | loop[l].wei.set(oc_blocking_size, is_amx(isa) ? nb_ur_average : nb_ur); |
| 1439 | // -- brgemm kernel: loop by ic_chunks -- |
| 1440 | l++; |
| 1441 | const auto ic_chunks = div_up(nb_ic, nb_ic_blocking); |
| 1442 | loop[l].src.set(sp_block * ic_blocking_size, 1); |
| 1443 | loop[l].dst.set(sp_block * oc_block, ic_chunks); |
| 1444 | auto wei_is = oc_blocking_size; |
| 1445 | auto wei_op = ocblock * ic; |
| 1446 | loop[l].wei.set(wei_is, 1); |
| 1447 | |
| 1448 | if (loop_order == loop_ndhwgc) { |
| 1449 | // -- harness: loop by oc_block -- |
| 1450 | l++; |
| 1451 | loop[l].src.set(sp_block * rnd_simd(ic), nb_oc_thr); |
| 1452 | loop[l].dst.set(sp_block * oc_block, 1); |
| 1453 | wei_is = oc_block * ic; |
| 1454 | wei_op = nsimd_oc_thr * ic; |
| 1455 | loop[l].wei.set(wei_is, 1); |
| 1456 | } |
| 1457 | |
| 1458 | const auto rnd_oc_for_sp |
| 1459 | = simd_w * ((loop_order == loop_ndhwgc) ? nsimd_oc_thr : ocblock); |
| 1460 | if (is_os_blocking) { |
| 1461 | // -- harness: loop by os_blocks -- |
| 1462 | l++; |
| 1463 | loop[l].src.set(sp_block * ic_blocking_size, 1); |
| 1464 | loop[l].dst.set(sp_block * rnd_oc_for_sp, 1); |
| 1465 | loop[l].wei.set(wei_op * simd_w, nb_sp_thr); |
| 1466 | } else { |
| 1467 | // -- harness: loop by sp_blocks -- |
| 1468 | l++; |
| 1469 | loop[l].src.set(sp_block * ic_blocking_size, 1); |
| 1470 | loop[l].dst.set(sp_block * rnd_oc_for_sp, 1); |
| 1471 | loop[l].wei.set(wei_op * simd_w, nb_sp_thr); |
| 1472 | // -- harness: loop by oh_blocks -- |
| 1473 | l++; |
| 1474 | loop[l].src.set(oh_block * sp_thr * rnd_simd(ic_blocking_size), 1); |
| 1475 | loop[l].dst.set(oh_block * sp_thr * rnd_oc_for_sp, 1); |
| 1476 | loop[l].wei.set(wei_op * simd_w, nb_oh_thr); |
| 1477 | // -- harness: loop by od_blocks -- |
| 1478 | l++; |
| 1479 | loop[l].src.set( |
| 1480 | od_block * oh_thr * sp_thr * rnd_simd(ic_blocking_size), 1); |
| 1481 | loop[l].dst.set(od_block * oh_thr * sp_thr * rnd_oc_for_sp, 1); |
| 1482 | loop[l].wei.set(wei_op * simd_w, nb_od_thr); |
| 1483 | } |
| 1484 | |
| 1485 | if (loop_order != loop_ndhwgc) { |
| 1486 | // -- harness: loop by oc_block -- |
| 1487 | l++; |
| 1488 | loop[l].src.set(od_thr * oh_thr * rnd_simd(sp_thr * ic_blocking_size), |
| 1489 | nb_oc_thr); |
| 1490 | loop[l].dst.set(oc_block * od_thr * oh_thr * sp_thr, 1); |
| 1491 | loop[l].wei.set(oc_block * ic, 1); |
| 1492 | } |
| 1493 | |
| 1494 | // -- harness: loop by mb -- |
| 1495 | l++; |
| 1496 | const auto mb_thr = nstl::min(mb, div_up(job, sp_amount * ngroups * nb_oc)); |
| 1497 | loop[l].src.set(od_thr * oh_thr * sp_thr * rnd_simd(ic_blocking_size), 1); |
| 1498 | loop[l].dst.set(nsimd_oc_thr * simd_w * od_thr * oh_thr * sp_thr, 1); |
| 1499 | loop[l].wei.set(nsimd_oc_thr * ic * simd_w, mb_thr); |
| 1500 | |
| 1501 | const auto src_op = static_cast<dim_t>(mb_thr) * od_thr * oh_thr * sp_thr |
| 1502 | * ic_blocking_size; |
| 1503 | const auto dst_op = static_cast<dim_t>(mb_thr) * nsimd_oc_thr * od_thr |
| 1504 | * oh_thr * sp_thr; |
| 1505 | wei_op = nsimd_oc_thr * ic; |
| 1506 | |
| 1507 | // for "real" application set bench_iterations to 1 |
| 1508 | const auto iterations = bench_iterations; |
| 1509 | l++; |
| 1510 | loop[l].src.set(src_op, iterations); |
| 1511 | loop[l].dst.set(dst_op * simd_w, iterations); |
| 1512 | loop[l].wei.set(wei_op * simd_w, iterations); |
| 1513 | |
| 1514 | auto src_mem_k = mem_k; |
| 1515 | auto dst_mem_k = mem_k; |
| 1516 | auto wei_mem_k = mem_k; |
| 1517 | float src_rp = 1; |
| 1518 | float dst_rp = 1; |
| 1519 | float wei_rp = 1; |
| 1520 | |
| 1521 | for (auto il = l; il >= 0; il--) { |
| 1522 | src_mem_k = io_k(loop[il], loop[il].src, src_mem_k, true, false); |
| 1523 | dst_mem_k = io_k(loop[il], loop[il].dst, dst_mem_k, false, false); |
| 1524 | wei_mem_k = io_k(loop[il], loop[il].wei, wei_mem_k, false, true); |
| 1525 | src_rp *= loop[il].src.repeatn; |
| 1526 | dst_rp *= loop[il].dst.repeatn; |
| 1527 | wei_rp *= loop[il].wei.repeatn; |
| 1528 | } |
| 1529 | const auto src_ops = (src_op * src_rp) / iterations; |
| 1530 | const auto dst_ops = (dst_op * dst_rp) / iterations; |
| 1531 | const auto wei_ops = (wei_op * wei_rp) / iterations; |
| 1532 | |
| 1533 | const auto src_cost = src_mem_k * src_ops; |
| 1534 | const auto dst_cost = dst_mem_k * dst_ops; |
| 1535 | const auto wei_cost = wei_mem_k * wei_ops; |
| 1536 | const auto call_kernel_cost = 1000.f * job * ic_chunks; |
| 1537 | |
| 1538 | const auto up_sp_size = is_os_blocking ? 1 : od * oh; |
| 1539 | |
| 1540 | const auto cache_eff = (static_cast<dim_t>(mb) * up_sp_size * sp * ic * oc) |
| 1541 | / (nthr * (src_cost + dst_cost + wei_cost + call_kernel_cost)); |
| 1542 | |
| 1543 | const auto res_eff = oc_block_eff * brgemm_microkernel_eff * sp_eff |
| 1544 | * job_eff * ur_eff * cache_eff * brgemm_eff; |
| 1545 | return res_eff; |
| 1546 | } |
| 1547 | |
| 1548 | void brg_blocking_t::calc_blocks_1x1() { |
| 1549 | const bool is_os_blocking_ok |
| 1550 | = utils::everyone_is(1, stride_d, stride_h) && iw % stride_w == 0; |
| 1551 | const bool is_ic_zero_padded = ic != ic_without_padding; |
| 1552 | is_rtus = is_ic_zero_padded || (!is_os_blocking_ok && is_amx(isa)); |
| 1553 | if (is_os_blocking_ok || is_rtus) { |
| 1554 | sp = os; |
| 1555 | is_os_blocking = true; |
| 1556 | } else { |
| 1557 | sp = ow; |
| 1558 | is_os_blocking = false; |
| 1559 | } |
| 1560 | |
| 1561 | od_block = 1; |
| 1562 | oh_block = 1; |
| 1563 | kd_block = kh_block = kw_block = 1; |
| 1564 | kd_block_pad = kh_block_pad = kw_block_pad = 1; |
| 1565 | nb_ic_blocking = 1; |
| 1566 | |
| 1567 | const auto thr_eff_threshold = 0.9f; |
| 1568 | |
| 1569 | const auto max_sp_block_L2 = os; |
| 1570 | // TODO: nb_os_blocking always is 1 for now. Update this code |
| 1571 | nb_os_blocking = 1; |
| 1572 | int start_sp_block = 0; |
| 1573 | |
| 1574 | if (is_os_blocking) { |
| 1575 | ow_block = 0; |
| 1576 | |
| 1577 | const auto max_os_block_thr |
| 1578 | = (src_dsz * ic >= 1024 && src_dsz * ic < 4096) |
| 1579 | ? nstl::max(nstl::min(16, os), |
| 1580 | div_up(os, div_up(nthr, mb * div_up(oc, oc_block)))) |
| 1581 | : nstl::max(div_up(2048, oc_block), |
| 1582 | static_cast<int>(div_up(mb * ngroups * os, nthr))); |
| 1583 | const auto max_os_block_L2 = max_sp_block_L2; |
| 1584 | |
| 1585 | auto max_os_block_aliasing = 1000000 / nthr; |
| 1586 | if ((oc_without_padding * os * dst_dsz) % P4K == 0) { |
| 1587 | max_os_block_aliasing /= 1; |
| 1588 | for (auto cur_oc = oc_without_padding; |
| 1589 | max_os_block_aliasing * dst_dsz > 400 && cur_oc % 2 == 0 |
| 1590 | && cur_oc * os * dst_dsz >= P4K; |
| 1591 | cur_oc /= 2) { |
| 1592 | max_os_block_aliasing /= 2; |
| 1593 | } |
| 1594 | max_os_block_aliasing += max_os_block_aliasing % 2 ? 0 : 1; |
| 1595 | } |
| 1596 | max_os_block_aliasing |
| 1597 | = nstl::min(div_up(1001, dst_dsz), max_os_block_aliasing); |
| 1598 | |
| 1599 | start_sp_block = utils::saturate(1, os, |
| 1600 | nstl::min(nstl::min(max_os_block_thr, max_os_block_L2), |
| 1601 | max_os_block_aliasing)); |
| 1602 | |
| 1603 | } else { |
| 1604 | os_block = 0; |
| 1605 | |
| 1606 | const auto max_ow_block_thr = utils::saturate(1, ow, |
| 1607 | static_cast<int>(div_up( |
| 1608 | mb * ngroups * nb_oc * os, thr_eff_threshold * nthr))); |
| 1609 | const auto max_ow_block_L2 = max_sp_block_L2; |
| 1610 | |
| 1611 | start_sp_block = utils::saturate( |
| 1612 | 1, ow, nstl::min(max_ow_block_thr, max_ow_block_L2)); |
| 1613 | } |
| 1614 | os_block = ow_block = sp_block = -1; |
| 1615 | brg_blocking_t best_brgb = *this; |
| 1616 | |
| 1617 | auto prev_spb = 0; |
| 1618 | for (auto ns = 1; ns <= sp; ns++) { |
| 1619 | auto spb = div_up(sp, ns); |
| 1620 | if (is_amx(isa)) { |
| 1621 | auto min_dis = 16; |
| 1622 | auto best_w = 16; |
| 1623 | const auto max_tile_width = nstl::min(16, sp); |
| 1624 | const auto min_tile_width = utils::saturate(1, 11, sp / 2); |
| 1625 | if (spb < min_tile_width) break; |
| 1626 | for (auto w = max_tile_width; w >= min_tile_width; w--) { |
| 1627 | const auto dis = nstl::additive_inverse_modulo(spb, w); |
| 1628 | if (dis < min_dis) { |
| 1629 | min_dis = dis; |
| 1630 | best_w = w; |
| 1631 | } |
| 1632 | } |
| 1633 | spb = nstl::min(sp, rnd_dn(spb, best_w)); |
| 1634 | if (spb == prev_spb) continue; |
| 1635 | } |
| 1636 | if (spb == prev_spb || spb > start_sp_block) continue; |
| 1637 | prev_spb = spb; |
| 1638 | os_block = ow_block = sp_block = spb; |
| 1639 | select_ic_block(); |
| 1640 | const status_t st = estimate_brgemm_ur(); |
| 1641 | if (st != status::success) continue; |
| 1642 | update_blocks(); |
| 1643 | |
| 1644 | use_buffer = (dst_dt != acc_dt || with_sum) |
| 1645 | && (ic_block * nb_ic_blocking < ic); |
| 1646 | |
| 1647 | eff = est_eff_1x1(); |
| 1648 | if (eff > best_brgb.eff || best_brgb.eff == 0) best_brgb = *this; |
| 1649 | } |
| 1650 | *this = best_brgb; |
| 1651 | os_block = ow_block = sp_block; |
| 1652 | update_blocks(); |
| 1653 | } |
| 1654 | |
| 1655 | brgemm_broadcast_t get_zp_type(const primitive_attr_t &attr, int arg) { |
| 1656 | return attr.zero_points_.has_default_values(arg) |
| 1657 | ? brgemm_broadcast_t::none |
| 1658 | : brgemm_broadcast_t::per_tensor; |
| 1659 | } |
| 1660 | status_t init_jcp(jit_brgemm_conv_conf_t &jcp, cpu_isa_t isa, |
| 1661 | const convolution_desc_t &cd, memory_desc_t &src_md, |
| 1662 | memory_desc_t &weights_md, memory_desc_t &dst_md, |
| 1663 | memory_desc_t &bias_md, primitive_attr_t &attr, int nthreads) { |
| 1664 | using namespace prop_kind; |
| 1665 | |
| 1666 | brg_blocking_t::L1 = platform::get_per_core_cache_size(1); |
| 1667 | brg_blocking_t::L2 = platform::get_per_core_cache_size(2); |
| 1668 | brg_blocking_t::L3 = platform::get_per_core_cache_size(2); |
| 1669 | |
| 1670 | const memory_desc_wrapper src_d(&src_md); |
| 1671 | const memory_desc_wrapper weights_d(&weights_md); |
| 1672 | const memory_desc_wrapper dst_d(&dst_md); |
| 1673 | const memory_desc_wrapper bias_d(&bias_md); |
| 1674 | |
| 1675 | const bool with_groups = weights_d.ndims() == src_d.ndims() + 1; |
| 1676 | int ndims = src_d.ndims(); |
| 1677 | |
| 1678 | jcp = zero<decltype(jcp)>(); |
| 1679 | jcp.isa = isa; |
| 1680 | |
| 1681 | if (is_amx(isa)) { |
| 1682 | const int target_palette = amx::get_target_palette(); |
| 1683 | if (amx::get_max_tiles(target_palette) != 8 |
| 1684 | || amx::get_max_rows(target_palette) != 16) |
| 1685 | return status::unimplemented; |
| 1686 | } |
| 1687 | |
| 1688 | jcp.ndims = ndims; |
| 1689 | jcp.prop_kind = cd.prop_kind; |
| 1690 | jcp.ngroups = with_groups ? weights_d.dims()[0] : 1; |
| 1691 | jcp.mb = src_d.dims()[0]; |
| 1692 | jcp.oc_without_padding = dst_d.dims()[1]; |
| 1693 | jcp.oc = jcp.oc_without_padding / jcp.ngroups; |
| 1694 | jcp.ic_without_padding = src_d.dims()[1] / jcp.ngroups; |
| 1695 | jcp.ic = jcp.ic_without_padding; |
| 1696 | jcp.id = (ndims == 5) ? src_d.dims()[2] : 1; |
| 1697 | jcp.ih = (ndims == 3) ? 1 : src_d.dims()[ndims - 2]; |
| 1698 | jcp.iw = src_d.dims()[ndims - 1]; |
| 1699 | jcp.od = (ndims == 5) ? dst_d.dims()[2] : 1; |
| 1700 | jcp.oh = (ndims == 3) ? 1 : dst_d.dims()[ndims - 2]; |
| 1701 | jcp.ow = dst_d.dims()[ndims - 1]; |
| 1702 | jcp.kd = (ndims == 5) ? weights_d.dims()[with_groups + 2] : 1; |
| 1703 | jcp.kh = (ndims == 3) ? 1 : weights_d.dims()[with_groups + ndims - 2]; |
| 1704 | jcp.kw = weights_d.dims()[with_groups + ndims - 1]; |
| 1705 | jcp.f_pad = (ndims == 5) ? cd.padding[0][0] : 0; |
| 1706 | jcp.t_pad = (ndims == 3) ? 0 : cd.padding[0][ndims - 4]; |
| 1707 | jcp.l_pad = cd.padding[0][ndims - 3]; |
| 1708 | jcp.stride_d = (ndims == 5) ? cd.strides[0] : 1; |
| 1709 | jcp.stride_h = (ndims == 3) ? 1 : cd.strides[ndims - 4]; |
| 1710 | jcp.stride_w = cd.strides[ndims - 3]; |
| 1711 | |
| 1712 | jcp.dilate_d = (ndims == 5) ? cd.dilates[0] : 0; |
| 1713 | jcp.dilate_h = (ndims == 3) ? 0 : cd.dilates[ndims - 4]; |
| 1714 | jcp.dilate_w = cd.dilates[ndims - 3]; |
| 1715 | |
| 1716 | jcp.os = jcp.od * jcp.oh * jcp.ow; |
| 1717 | |
| 1718 | jcp.ext_kd = calculate_extended_filter_size(jcp.kd, jcp.dilate_d); |
| 1719 | jcp.ext_kh = calculate_extended_filter_size(jcp.kh, jcp.dilate_h); |
| 1720 | jcp.ext_kw = calculate_extended_filter_size(jcp.kw, jcp.dilate_w); |
| 1721 | |
| 1722 | jcp.back_pad = calculate_end_padding( |
| 1723 | jcp.f_pad, jcp.od, jcp.id, jcp.stride_d, jcp.ext_kd); |
| 1724 | jcp.b_pad = calculate_end_padding( |
| 1725 | jcp.t_pad, jcp.oh, jcp.ih, jcp.stride_h, jcp.ext_kh); |
| 1726 | jcp.r_pad = calculate_end_padding( |
| 1727 | jcp.l_pad, jcp.ow, jcp.iw, jcp.stride_w, jcp.ext_kw); |
| 1728 | |
| 1729 | jcp.is_1x1 = jcp.f_pad <= 0 && jcp.back_pad <= 0 && jcp.t_pad <= 0 |
| 1730 | && jcp.b_pad <= 0 && jcp.l_pad <= 0 && jcp.r_pad <= 0 |
| 1731 | && utils::everyone_is(1, jcp.kd, jcp.kh, jcp.kw); |
| 1732 | |
| 1733 | jcp.with_bias = bias_md.format_kind != format_kind::undef; |
| 1734 | |
| 1735 | jcp.src_dt = src_md.data_type; |
| 1736 | jcp.dst_dt = dst_md.data_type; |
| 1737 | jcp.wei_dt = weights_md.data_type; |
| 1738 | jcp.bia_dt = jcp.with_bias ? bias_md.data_type : data_type::undef; |
| 1739 | |
| 1740 | if (one_of(jcp.src_dt, u8, s8)) { |
| 1741 | jcp.acc_dt = s32; |
| 1742 | } else if (one_of(jcp.src_dt, f32, bf16, f16)) { |
| 1743 | jcp.acc_dt = f32; |
| 1744 | } else |
| 1745 | return status::unimplemented; |
| 1746 | |
| 1747 | jcp.src_dsz = types::data_type_size(jcp.src_dt); |
| 1748 | jcp.wei_dsz = types::data_type_size(jcp.wei_dt); |
| 1749 | jcp.dst_dsz = types::data_type_size(jcp.dst_dt); |
| 1750 | jcp.acc_dsz = types::data_type_size(jcp.acc_dt); |
| 1751 | jcp.bia_dsz = jcp.with_bias ? types::data_type_size(jcp.bia_dt) : 0; |
| 1752 | |
| 1753 | jcp.simd_w = isa_max_vlen(isa) / jcp.src_dsz; |
| 1754 | jcp.acc_simd_w = isa_max_vlen(isa) / jcp.acc_dsz; |
| 1755 | jcp.is_bf32 = everyone_is(f32, jcp.src_dt, jcp.wei_dt) |
| 1756 | && attr.fpmath_mode_ == fpmath_mode::bf16 && isa == avx512_core_amx; |
| 1757 | |
| 1758 | brg_blocking_t::last_ic_block_size |
| 1759 | = (jcp.wei_dt == f16 && isa == avx512_core_fp16) |
| 1760 | ? 1 |
| 1761 | : data_type_vnni_granularity(jcp.wei_dt); |
| 1762 | |
| 1763 | // TODO: optimize depthwise convolutions (for now direct approach is faster) |
| 1764 | const bool is_depthwise |
| 1765 | = with_groups && jcp.ngroups > 1 && everyone_is(1, jcp.ic, jcp.oc); |
| 1766 | if (is_depthwise) return status::unimplemented; |
| 1767 | |
| 1768 | // TODO: optimize grouped convolutions with small ic |
| 1769 | const bool is_grouped_small_ic |
| 1770 | = jcp.prop_kind != prop_kind::backward_weights && with_groups |
| 1771 | && jcp.ngroups > 1 && jcp.ic <= jcp.acc_simd_w |
| 1772 | && IMPLICATION(is_amx(jcp.isa), |
| 1773 | jcp.ic < 16 |
| 1774 | && jcp.oc < 16 |
| 1775 | // already optimized for amx 1x1 convs |
| 1776 | && !jcp.is_1x1) |
| 1777 | // Enable the shapes not supported in direct convs |
| 1778 | && IMPLICATION(with_groups, is_groups_ok(jcp)); |
| 1779 | if (is_grouped_small_ic) return status::unimplemented; |
| 1780 | |
| 1781 | // Dispatch the shapes to VNNI for better performance |
| 1782 | // TODO: optimize the perf of 3d shape with small ic and large spatial |
| 1783 | const auto max_small_shapes_sz = jcp.is_1x1 |
| 1784 | ? static_cast<int32_t>(brg_blocking_t::L1) / 2 |
| 1785 | : static_cast<int32_t>(brg_blocking_t::L1); |
| 1786 | const auto is_small_shape = is_amx(jcp.isa) && jcp.os <= 4 && jcp.ic <= 512 |
| 1787 | && jcp.mb * jcp.ngroups * jcp.ic * jcp.oc <= max_small_shapes_sz; |
| 1788 | const auto is_3d_small_ic = is_amx(jcp.isa) && jcp.ndims == 5 |
| 1789 | && jcp.ic * jcp.oc <= 32 && jcp.od >= 128 && jcp.oh >= 128 |
| 1790 | && jcp.ow >= 128; |
| 1791 | if (one_of(jcp.prop_kind, prop_kind::forward_training, |
| 1792 | prop_kind::forward_inference) |
| 1793 | && (is_small_shape || is_3d_small_ic)) |
| 1794 | return status::unimplemented; |
| 1795 | |
| 1796 | jcp.s8s8_avx512 = jcp.src_dt == s8 && !is_amx(jcp.isa); |
| 1797 | |
| 1798 | if (!IMPLICATION(jcp.wei_dt == s8, mayiuse(avx512_core_vnni))) |
| 1799 | return status::unimplemented; |
| 1800 | if (!IMPLICATION(jcp.wei_dt == bf16, |
| 1801 | mayiuse(avx512_core_bf16) || mayiuse(avx2_vnni_2))) |
| 1802 | return status::unimplemented; |
| 1803 | if (!IMPLICATION(jcp.wei_dt == f16, |
| 1804 | mayiuse(avx512_core_fp16) || mayiuse(avx2_vnni_2))) |
| 1805 | return status::unimplemented; |
| 1806 | const bool is_f32 |
| 1807 | = utils::everyone_is(f32, jcp.src_dt, jcp.wei_dt, jcp.dst_dt); |
| 1808 | if (!IMPLICATION(is_f32, one_of(isa, avx512_core, avx2) || jcp.is_bf32)) |
| 1809 | return status::unimplemented; |
| 1810 | |
| 1811 | if (!post_ops_ok(jcp, attr, dst_d)) return status::unimplemented; |
| 1812 | |
| 1813 | jcp.amx_h = 16; |
| 1814 | jcp.amx_w = 64 / (jcp.is_bf32 ? types::data_type_size(bf16) : jcp.src_dsz); |
| 1815 | |
| 1816 | const auto &p = attr.post_ops_; |
| 1817 | jcp.with_sum = p.find(primitive_kind::sum) != -1; |
| 1818 | const int eltwise_ind = p.find(primitive_kind::eltwise); |
| 1819 | jcp.with_eltwise = eltwise_ind != -1; |
| 1820 | |
| 1821 | const int binary_ind = p.find(primitive_kind::binary); |
| 1822 | jcp.with_binary = binary_ind != -1; |
| 1823 | |
| 1824 | jcp.src_zero_point |
| 1825 | = get_zp_type(attr, DNNL_ARG_SRC) != brgemm_broadcast_t::none; |
| 1826 | jcp.dst_zero_point |
| 1827 | = get_zp_type(attr, DNNL_ARG_DST) != brgemm_broadcast_t::none; |
| 1828 | |
| 1829 | // Only common zero points for the whole output tensor is supported now |
| 1830 | // TODO: Extend zero points support to AMX |
| 1831 | const bool has_zero_points = jcp.src_zero_point || jcp.dst_zero_point; |
| 1832 | if (has_zero_points || jcp.s8s8_avx512) { |
| 1833 | const bool params_ok = IMPLICATION(has_zero_points, !is_amx(jcp.isa)) |
| 1834 | && IMPLICATION( |
| 1835 | has_zero_points, utils::one_of(jcp.src_dt, u8, s8)) |
| 1836 | && IMPLICATION(jcp.src_zero_point, |
| 1837 | attr.zero_points_.common(DNNL_ARG_SRC)) |
| 1838 | && IMPLICATION(jcp.dst_zero_point, |
| 1839 | attr.zero_points_.common(DNNL_ARG_DST)); |
| 1840 | if (!params_ok) return status::unimplemented; |
| 1841 | } |
| 1842 | |
| 1843 | jcp.nthr = nthreads; |
| 1844 | jcp.kh_sets = 1; |
| 1845 | jcp.kw_sets = 1; |
| 1846 | jcp.copy_block_only = false; |
| 1847 | jcp.amx_tile_load_xx = false; |
| 1848 | jcp.use_M_mask = 0; |
| 1849 | jcp.is_os_blocking = false; |
| 1850 | jcp.oskip = 0; |
| 1851 | jcp.use_uker = false; |
| 1852 | jcp.use_interleave_stores = false; |
| 1853 | jcp.hint_prefetching = brgemm_kernel_prefetching_t::brgemm_prf_default; |
| 1854 | jcp.brgemm_bd_loop_innermost = false; |
| 1855 | |
| 1856 | if (jcp.prop_kind != prop_kind::backward_weights) { |
| 1857 | // fast check data layout before spending time for blocking selection |
| 1858 | format_tag_t src_tag = pick(jcp.ndims - 3, nwc, nhwc, ndhwc); |
| 1859 | CHECK(init_tag( |
| 1860 | jcp.src_tag, src_md, src_d, src_tag, is_any_eligible(jcp))); |
| 1861 | } |
| 1862 | if (jcp.with_bias) { |
| 1863 | if (bias_d.format_kind() == format_kind::any) |
| 1864 | CHECK(memory_desc_init_by_tag(bias_md, x)); |
| 1865 | } |
| 1866 | |
| 1867 | const auto ic_padded_block |
| 1868 | = jcp.acc_simd_w * brg_blocking_t::last_ic_block_size; |
| 1869 | jcp.is_ic_padded = !jcp.is_1x1 && one_of(jcp.wei_dt, bf16, f16, s8) |
| 1870 | && jcp.ic * jcp.kw_sets > ic_padded_block && is_amx(isa); |
| 1871 | |
| 1872 | jcp.idp = jcp.id + jcp.f_pad + jcp.back_pad; |
| 1873 | jcp.ihp = jcp.ih + jcp.t_pad + jcp.b_pad; |
| 1874 | jcp.iwp = jcp.iw + jcp.l_pad + jcp.r_pad; |
| 1875 | |
| 1876 | return status::success; |
| 1877 | } |
| 1878 | |
| 1879 | status_t init_conf(jit_brgemm_conv_conf_t &jcp, cpu_isa_t isa, |
| 1880 | const convolution_desc_t &cd, memory_desc_t &src_md, |
| 1881 | memory_desc_t &weights_md, memory_desc_t &dst_md, |
| 1882 | memory_desc_t &bias_md, primitive_attr_t &attr, int nthreads) { |
| 1883 | |
| 1884 | using namespace prop_kind; |
| 1885 | if (!mayiuse(isa)) return status::unimplemented; |
| 1886 | |
| 1887 | CHECK(init_jcp( |
| 1888 | jcp, isa, cd, src_md, weights_md, dst_md, bias_md, attr, nthreads)); |
| 1889 | |
| 1890 | if (jcp.is_1x1) return status::unimplemented; |
| 1891 | const memory_desc_wrapper src_d(&src_md); |
| 1892 | const memory_desc_wrapper weights_d(&weights_md); |
| 1893 | const memory_desc_wrapper dst_d(&dst_md); |
| 1894 | const memory_desc_wrapper bias_d(&bias_md); |
| 1895 | |
| 1896 | const bool with_groups = weights_d.ndims() == src_d.ndims() + 1; |
| 1897 | |
| 1898 | // TODO: check these restrictions |
| 1899 | if (is_amx(isa)) { |
| 1900 | // disabled for two convolutions from ssd_resnet34 |
| 1901 | if ((jcp.ic == jcp.oc) && (jcp.ic == 128 || jcp.ic == 256) |
| 1902 | && (jcp.oh == jcp.ow) && (jcp.oh == 150)) |
| 1903 | return status::unimplemented; |
| 1904 | // disabled for first convolutions excepting 3d |
| 1905 | const bool is_real_3d = (jcp.ndims == 5 |
| 1906 | && (jcp.id > 1 || jcp.od > 1 || jcp.kd > 1 |
| 1907 | || jcp.dilate_d > 0)); |
| 1908 | |
| 1909 | if (jcp.ic <= 4 && !is_real_3d |
| 1910 | && IMPLICATION(with_groups, is_groups_ok(jcp))) |
| 1911 | return status::unimplemented; |
| 1912 | |
| 1913 | if (jcp.f_pad >= jcp.ext_kd || jcp.t_pad >= jcp.ext_kh |
| 1914 | || jcp.r_pad >= jcp.ext_kw) |
| 1915 | return status::unimplemented; |
| 1916 | } |
| 1917 | |
| 1918 | using namespace data_type; |
| 1919 | // ======================= blocking ================================= |
| 1920 | |
| 1921 | auto bcast_amount |
| 1922 | = static_cast<size_t>(jcp.id) * jcp.ih * jcp.iw * jcp.src_dsz; |
| 1923 | auto wei_amount = static_cast<size_t>(jcp.oc) * jcp.kd * jcp.kh * jcp.kw |
| 1924 | * jcp.wei_dsz; |
| 1925 | |
| 1926 | jcp.loop_order = (bcast_amount < wei_amount) ? loop_ngcdhw : loop_ndhwgc; |
| 1927 | |
| 1928 | const int min_oc_block = jcp.acc_simd_w; |
| 1929 | |
| 1930 | int selected_ur = 0; |
| 1931 | MAYBE_UNUSED(selected_ur); |
| 1932 | |
| 1933 | auto try_exec_type = [&]() { |
| 1934 | brg_blocking_t best_brgb = zero<decltype(best_brgb)>(); |
| 1935 | best_brgb.oc_block = min_oc_block; |
| 1936 | brg_blocking_t cur_brgb = zero<decltype(best_brgb)>(); |
| 1937 | cur_brgb.get_from_jcp(jcp); |
| 1938 | auto start_ocb = (is_amx(isa) && jcp.is_os_blocking) ? 2 : 4; |
| 1939 | if (jcp.wei_plain) |
| 1940 | start_ocb = nstl::min(jcp.ic > 128 ? (jcp.ic > 256 ? 8 : 16) : 32, |
| 1941 | div_up(jcp.oc, jcp.acc_simd_w)); |
| 1942 | start_ocb = nstl::min(div_up(jcp.oc, jcp.acc_simd_w), start_ocb); |
| 1943 | |
| 1944 | auto finish_ocb = 1; |
| 1945 | for (auto ocb = start_ocb; ocb >= finish_ocb; ocb--) { |
| 1946 | cur_brgb.oc_block = ocb * jcp.acc_simd_w; |
| 1947 | cur_brgb.nb_oc = utils::div_up(jcp.oc, cur_brgb.oc_block); |
| 1948 | if (!cur_brgb.fast_check_oc_block()) continue; |
| 1949 | |
| 1950 | const status_t blocking_ok = cur_brgb.calc_blocks(); |
| 1951 | if (blocking_ok != status::success) continue; |
| 1952 | |
| 1953 | const status_t st = cur_brgb.get_brgemm_ur(&attr, dst_md); |
| 1954 | if (st != status::success) continue; |
| 1955 | cur_brgb.eff = cur_brgb.est_eff(); |
| 1956 | if (cur_brgb.eff > best_brgb.eff) best_brgb = cur_brgb; |
| 1957 | } |
| 1958 | if (best_brgb.oc_block == 0 || best_brgb.ic_block == 0 |
| 1959 | || best_brgb.ow_block == 0) |
| 1960 | return false; |
| 1961 | best_brgb.save_to_jcp(jcp); |
| 1962 | selected_ur = best_brgb.ur; |
| 1963 | return true; |
| 1964 | }; |
| 1965 | |
| 1966 | //----------------------------------------------------------------------- |
| 1967 | |
| 1968 | jcp.exec_type = exec_base; |
| 1969 | jcp.brg_type = brgemm_addr; // TODO: Choose right type of BRGEMM |
| 1970 | |
| 1971 | bool try_exec_vpad = false; |
| 1972 | bool try_exec_trans = false; |
| 1973 | bool try_exec_base = true; |
| 1974 | |
| 1975 | if (!is_amx(isa) && div_up(jcp.l_pad, jcp.stride_w) < jcp.kw |
| 1976 | && div_up(jcp.r_pad, jcp.stride_w) < jcp.kw) { |
| 1977 | try_exec_vpad = true; |
| 1978 | } |
| 1979 | |
| 1980 | const auto ic_padded_block |
| 1981 | = jcp.acc_simd_w * brg_blocking_t::last_ic_block_size; |
| 1982 | // TODO: remove this restriction |
| 1983 | const auto w_padding = jcp.l_pad > 0 || jcp.r_pad > 0; |
| 1984 | if (is_amx(isa)) { |
| 1985 | try_exec_base = !w_padding |
| 1986 | && IMPLICATION(jcp.ic <= ic_padded_block, |
| 1987 | jcp.ic % brg_blocking_t::last_ic_block_size == 0) |
| 1988 | && IMPLICATION( |
| 1989 | jcp.ic > ic_padded_block, jcp.ic % ic_padded_block == 0) |
| 1990 | && jcp.ow > 50 /*TODO: reinvestigate this heuristic */; |
| 1991 | try_exec_trans = !try_exec_base; |
| 1992 | } |
| 1993 | |
| 1994 | bool must_exec_vpad = false; |
| 1995 | |
| 1996 | // TODO: in future use (kd/kh/kw) and (kd/kh/kw)_pad blocks for more |
| 1997 | // precise calculation of jcp.max_batch |
| 1998 | jcp.max_batch = jcp.kd * jcp.kh * jcp.kw; |
| 1999 | |
| 2000 | //TODO: check wei plain |
| 2001 | jcp.wei_plain = false; |
| 2002 | jcp.wei_plain = jcp.exec_type == exec_vpad ? jcp.wei_plain : false; |
| 2003 | |
| 2004 | bool try_exec_type_res = false; |
| 2005 | |
| 2006 | if (try_exec_vpad) { |
| 2007 | jcp.exec_type = exec_vpad; |
| 2008 | try_exec_type_res = try_exec_type(); |
| 2009 | // to avoid case when both top and bottom virtual padding are non-zero |
| 2010 | // TODO: remove this restriction |
| 2011 | const auto iw_block = (jcp.ow_block - 1) * jcp.stride_w + 1; |
| 2012 | if (!must_exec_vpad && (iw_block > jcp.iw)) try_exec_type_res = false; |
| 2013 | } |
| 2014 | if (try_exec_type_res == false && try_exec_trans) { |
| 2015 | jcp.exec_type = exec_trans; |
| 2016 | |
| 2017 | // try loop_ndhwgc always for exec_trans |
| 2018 | jcp.loop_order = loop_ndhwgc; |
| 2019 | |
| 2020 | // we read input block only once for loop_ndhwgc, so we don't need to |
| 2021 | // keep it memory |
| 2022 | if (jcp.loop_order == loop_ndhwgc) { jcp.copy_block_only = true; } |
| 2023 | |
| 2024 | jcp.is_ic_padded = one_of(jcp.wei_dt, bf16, f16, s8) |
| 2025 | && jcp.ic * jcp.kw_sets > ic_padded_block; |
| 2026 | |
| 2027 | if (is_amx(isa) && (/* heuristic*/ jcp.kw_sets == 1 && jcp.ow < 256)) { |
| 2028 | jcp.is_os_blocking = jcp.f_pad < jcp.kd && jcp.back_pad < jcp.kd |
| 2029 | && jcp.t_pad < jcp.kh && jcp.b_pad < jcp.kh |
| 2030 | && jcp.r_pad < jcp.kw && jcp.l_pad < jcp.kw; |
| 2031 | jcp.use_M_mask = jcp.is_os_blocking ? 2 : 0; |
| 2032 | jcp.use_uker = true; |
| 2033 | jcp.use_interleave_stores = true; |
| 2034 | jcp.hint_prefetching = brgemm_kernel_prefetching_t::brgemm_prf1; |
| 2035 | // assuming 2x2 decomposition in amx brgemm kernel |
| 2036 | // and overlap of input by kw |
| 2037 | const auto bd_blocking = 2 * jcp.amx_h; |
| 2038 | const auto ld_blocking = 2 * 16; |
| 2039 | const auto A_ds |
| 2040 | = jcp.src_dsz * bd_blocking * jcp.ic * jcp.kd * jcp.kh; |
| 2041 | const auto B_ds = jcp.wei_dsz * ld_blocking * jcp.ic * jcp.kd |
| 2042 | * jcp.kh * jcp.kw; |
| 2043 | const auto C_ds = jcp.acc_dsz * bd_blocking * ld_blocking; |
| 2044 | if (A_ds + B_ds + C_ds > brg_blocking_t::L1) |
| 2045 | jcp.amx_tile_load_xx = true; |
| 2046 | } |
| 2047 | |
| 2048 | try_exec_type_res = try_exec_type(); |
| 2049 | } |
| 2050 | if (try_exec_base && try_exec_type_res == false) { |
| 2051 | jcp.exec_type = exec_base; |
| 2052 | try_exec_type_res = try_exec_type(); |
| 2053 | } |
| 2054 | |
| 2055 | if (try_exec_type_res == false) return status::unimplemented; |
| 2056 | |
| 2057 | // ============ end blocking =========================================== |
| 2058 | if (jcp.exec_type == exec_vpad) |
| 2059 | jcp.max_vpad = nstl::max(jcp.l_pad, jcp.r_pad); |
| 2060 | else |
| 2061 | jcp.max_vpad = 0; |
| 2062 | |
| 2063 | if (jcp.ow_block == 0 || jcp.ic_block == 0 || jcp.oc_block == 0) |
| 2064 | return status::unimplemented; |
| 2065 | |
| 2066 | jcp.gemm_batch_size = jcp.nb_ic_blocking |
| 2067 | * nstl::max(jcp.kd_block * jcp.kh_block * jcp.kw_block, |
| 2068 | jcp.kd_block_pad * jcp.kh_block_pad * jcp.kw_block_pad); |
| 2069 | // to avoid cache concurrent write access from different threads |
| 2070 | size_t sc_size = sizeof(brgemm_batch_element_t); |
| 2071 | jcp.adjusted_batch_size |
| 2072 | = div_up(rnd_up(jcp.gemm_batch_size * sc_size, P4K), sc_size); |
| 2073 | |
| 2074 | CHECK(pick_tags(jcp, src_md, weights_md, dst_md, bias_md)); |
| 2075 | CHECK(attr.set_default_formats(&dst_md)); |
| 2076 | |
| 2077 | const auto &src_scales = attr.scales_.get(DNNL_ARG_SRC); |
| 2078 | const auto &wei_scales = attr.scales_.get(DNNL_ARG_WEIGHTS); |
| 2079 | const auto &dst_scales = attr.scales_.get(DNNL_ARG_DST); |
| 2080 | jcp.with_scales = !src_scales.has_default_values() |
| 2081 | || !wei_scales.has_default_values(); |
| 2082 | const int wei_mask_per_oc = 1 << (int)with_groups; |
| 2083 | jcp.is_oc_scale = wei_scales.mask_ == wei_mask_per_oc; |
| 2084 | |
| 2085 | // only common and per-oc-channel scales are supported |
| 2086 | const bool scales_ok = one_of(wei_scales.mask_, 0, wei_mask_per_oc) |
| 2087 | && src_scales.mask_ == 0 && dst_scales.has_default_values(); |
| 2088 | if (!scales_ok) return status::unimplemented; |
| 2089 | |
| 2090 | jcp.buffer_size = jcp.LDC * jcp.M; |
| 2091 | |
| 2092 | jcp.nb_od = div_up(jcp.od, jcp.od_block); |
| 2093 | jcp.nb_oh = div_up(jcp.oh, jcp.oh_block); |
| 2094 | |
| 2095 | if (jcp.exec_type == exec_trans) { |
| 2096 | // TODO: this is rough estimation of buffer for transpose input |
| 2097 | dim_t ds = jcp.copy_block_only |
| 2098 | ? (brg_blocking_t::get_inp_size(jcp.idp, jcp.od_block, jcp.kd, |
| 2099 | jcp.stride_d, jcp.dilate_d) |
| 2100 | + nstl::max(0, jcp.f_pad) + nstl::max(0, jcp.back_pad)) |
| 2101 | : jcp.idp; |
| 2102 | dim_t hs = jcp.copy_block_only |
| 2103 | ? (brg_blocking_t::get_inp_size(jcp.ihp, jcp.oh_block, jcp.kh, |
| 2104 | jcp.stride_h, jcp.dilate_h) |
| 2105 | + nstl::max(0, jcp.t_pad) + nstl::max(0, jcp.b_pad)) |
| 2106 | : jcp.ihp; |
| 2107 | if (jcp.is_os_blocking) |
| 2108 | hs = div_up(rnd_up(hs * jcp.iwp, jcp.brgM), jcp.iwp); |
| 2109 | |
| 2110 | jcp.inp_buffer_size = rnd_up(ds * hs * jcp.iwp * jcp.ngroups * jcp.nb_ic |
| 2111 | * jcp.ic_block * jcp.kh_sets * jcp.kw_sets, |
| 2112 | P4K); |
| 2113 | jcp.inp_buffer_mask_size = rnd_up(static_cast<dim_t>(jcp.nb_od) |
| 2114 | * jcp.nb_oh * jcp.nb_ow * jcp.ngroups * jcp.nb_ic, |
| 2115 | P4K); |
| 2116 | } |
| 2117 | |
| 2118 | const bool with_pad = jcp.f_pad > 0 || jcp.back_pad > 0 || jcp.t_pad > 0 |
| 2119 | || jcp.b_pad > 0; |
| 2120 | |
| 2121 | if (jcp.s8s8_avx512) { |
| 2122 | weights_md.extra.flags = 0 | memory_extra_flags::compensation_conv_s8s8; |
| 2123 | weights_md.extra.compensation_mask = with_groups ? 0x3 : 0x1; |
| 2124 | } |
| 2125 | if (jcp.src_zero_point && !is_amx(jcp.isa)) { |
| 2126 | weights_md.extra.flags |
| 2127 | |= memory_extra_flags::compensation_conv_asymmetric_src; |
| 2128 | weights_md.extra.asymm_compensation_mask = with_groups ? 0x3 : 0x1; |
| 2129 | } |
| 2130 | |
| 2131 | // disables the shape with small ic but large spatial |
| 2132 | // or specific large spatial shapes for int8 conv |
| 2133 | const auto is_ok_large_spatial |
| 2134 | = IMPLICATION(!is_amx(jcp.isa) && jcp.ic <= 128, |
| 2135 | jcp.od * jcp.oh < 100 |
| 2136 | || jcp.ic * jcp.oc_block * jcp.ow_block > 8192) |
| 2137 | && !(is_amx(jcp.isa) && jcp.ic < 16 && jcp.ndims == 4) |
| 2138 | && IMPLICATION(is_amx(jcp.isa) && jcp.ic <= 16, |
| 2139 | jcp.ow < 2048 |
| 2140 | || div_up(jcp.ow_block, selected_ur) * jcp.kd |
| 2141 | * jcp.kh * jcp.kw |
| 2142 | > 8192) |
| 2143 | && !(!is_amx(jcp.isa) && jcp.oc == 1024 |
| 2144 | && utils::everyone_is(1, jcp.od, jcp.oh, jcp.kd, jcp.kh) |
| 2145 | && jcp.ow >= 595 && jcp.kw <= 5); |
| 2146 | if (one_of(jcp.src_dt, u8, s8) && !is_ok_large_spatial) |
| 2147 | return status::unimplemented; |
| 2148 | |
| 2149 | // For padding shapes, we calculate the comp along with the computation |
| 2150 | // inside brgemm kernel when output size is small to get optimal perf |
| 2151 | // Or we calculate the comp using brgemm_coomp_pad kernel |
| 2152 | const auto output_sz = static_cast<dim_t>(jcp.mb) * jcp.ngroups * jcp.oc |
| 2153 | * jcp.od * jcp.oh * jcp.ow; |
| 2154 | const auto comp_with_pads = (jcp.src_zero_point || jcp.s8s8_avx512) |
| 2155 | && IMPLICATION(jcp.exec_type == exec_vpad, with_pad); |
| 2156 | jcp.req_brg_comp_pad = comp_with_pads && output_sz <= 8192 && jcp.oc < 512; |
| 2157 | jcp.req_cal_comp_pad = comp_with_pads && !jcp.req_brg_comp_pad; |
| 2158 | |
| 2159 | // estimate the number of kernel range combination for compensation |
| 2160 | const auto kd_cnt = 1 + utils::div_up(abs(jcp.f_pad), jcp.dilate_d + 1) |
| 2161 | + utils::div_up(abs(jcp.back_pad), jcp.dilate_d + 1); |
| 2162 | const auto kh_cnt = 1 + utils::div_up(abs(jcp.t_pad), jcp.dilate_h + 1) |
| 2163 | + utils::div_up(abs(jcp.b_pad), jcp.dilate_h + 1); |
| 2164 | const auto kw_cnt |
| 2165 | = (1 |
| 2166 | + (utils::div_up(abs(jcp.l_pad), jcp.dilate_w + 1) |
| 2167 | + utils::div_up( |
| 2168 | abs(jcp.r_pad), jcp.dilate_w + 1))) |
| 2169 | * 2; |
| 2170 | |
| 2171 | jcp.ker_ranges_size = jcp.exec_type == exec_base ? kd_cnt * kh_cnt * kw_cnt |
| 2172 | : kd_cnt * kh_cnt; |
| 2173 | jcp.comp_a_buffer_size |
| 2174 | = jcp.ngroups * jcp.nb_oc * jcp.ker_ranges_size * jcp.oc_block; |
| 2175 | jcp.s8s8_comp_buffer_size = jcp.comp_a_buffer_size; |
| 2176 | |
| 2177 | if (!IMPLICATION(jcp.is_bf32, jcp.use_uker)) return status::unimplemented; |
| 2178 | |
| 2179 | return status::success; |
| 2180 | } |
| 2181 | |
| 2182 | status_t init_1x1_conf(jit_brgemm_conv_conf_t &jcp, cpu_isa_t isa, |
| 2183 | const convolution_desc_t &cd, memory_desc_t &src_md, |
| 2184 | memory_desc_t &weights_md, memory_desc_t &dst_md, |
| 2185 | memory_desc_t &bias_md, primitive_attr_t &attr, int nthreads) { |
| 2186 | |
| 2187 | using namespace prop_kind; |
| 2188 | if (!mayiuse(isa)) return status::unimplemented; |
| 2189 | |
| 2190 | CHECK(init_jcp( |
| 2191 | jcp, isa, cd, src_md, weights_md, dst_md, bias_md, attr, nthreads)); |
| 2192 | |
| 2193 | const memory_desc_wrapper src_d(&src_md); |
| 2194 | const memory_desc_wrapper weights_d(&weights_md); |
| 2195 | const memory_desc_wrapper dst_d(&dst_md); |
| 2196 | const memory_desc_wrapper bias_d(&bias_md); |
| 2197 | |
| 2198 | if (!jcp.is_1x1) return status::unimplemented; |
| 2199 | |
| 2200 | using namespace data_type; |
| 2201 | // ===================== blocking ================================= |
| 2202 | |
| 2203 | auto bcast_amount |
| 2204 | = static_cast<size_t>(jcp.id) * jcp.ih * jcp.iw * jcp.src_dsz; |
| 2205 | auto wei_amount = static_cast<size_t>(jcp.oc) * jcp.wei_dsz; |
| 2206 | |
| 2207 | jcp.loop_order = (bcast_amount < wei_amount) ? loop_ngcdhw : loop_ndhwgc; |
| 2208 | |
| 2209 | if (is_amx(isa)) { |
| 2210 | // round up ic if needed |
| 2211 | const int vnni_width = brg_blocking_t::last_ic_block_size; |
| 2212 | const int n_vnni_blocks = utils::div_up(jcp.ic, vnni_width); |
| 2213 | const int ic_block |
| 2214 | = nstl::min(jcp.acc_simd_w, n_vnni_blocks) * vnni_width; |
| 2215 | const bool do_zeropad = (!jcp.is_bf32) |
| 2216 | && (jcp.ic % vnni_width != 0 || jcp.ic > ic_block); |
| 2217 | if (do_zeropad) jcp.ic = utils::rnd_up(jcp.ic, ic_block); |
| 2218 | const auto ic_padded_block = jcp.acc_simd_w * vnni_width; |
| 2219 | jcp.is_ic_padded = jcp.ic > ic_padded_block && !(jcp.is_bf32); |
| 2220 | |
| 2221 | // try to choose optimal loop order |
| 2222 | // TODO: incorporate loop order into smart blocking selection |
| 2223 | auto wei_size = (size_t)jcp.oc * jcp.ic * jcp.wei_dsz; |
| 2224 | auto max_size = 0.75f * brg_blocking_t::L2; |
| 2225 | const dim_t os = jcp.od * jcp.oh * jcp.ow; |
| 2226 | const dim_t os_cutoff = 400; // approximate and empiric |
| 2227 | const bool use_loop_ngcdhw |
| 2228 | = max_size < wei_size || (jcp.mb == 1 && os < os_cutoff); |
| 2229 | jcp.loop_order = use_loop_ngcdhw ? loop_ngcdhw : loop_ndhwgc; |
| 2230 | } |
| 2231 | |
| 2232 | const auto min_oc_block = jcp.acc_simd_w; |
| 2233 | |
| 2234 | jcp.brg_type = brgemm_addr; // TODO: Choose right type of BRGEMM |
| 2235 | |
| 2236 | // max_batch is 1 and max_vpad is 0 for 1x1 convolutions |
| 2237 | jcp.max_batch = 1; |
| 2238 | jcp.max_vpad = 0; |
| 2239 | |
| 2240 | jcp.wei_plain = false; |
| 2241 | |
| 2242 | brg_blocking_t best_brgb = zero<decltype(best_brgb)>(); |
| 2243 | best_brgb.oc_block = min_oc_block; |
| 2244 | brg_blocking_t cur_brgb = zero<decltype(cur_brgb)>(); |
| 2245 | cur_brgb.get_from_jcp(jcp); |
| 2246 | auto start_ocb = 4; |
| 2247 | if (jcp.wei_plain) |
| 2248 | start_ocb = nstl::min(jcp.ic > 128 ? (jcp.ic > 256 ? 8 : 16) : 32, |
| 2249 | div_up(jcp.oc, jcp.acc_simd_w)); |
| 2250 | start_ocb = nstl::min(div_up(jcp.oc, jcp.acc_simd_w), start_ocb); |
| 2251 | |
| 2252 | auto finish_ocb = 1; |
| 2253 | for (auto ocb = start_ocb; ocb >= finish_ocb; ocb--) { |
| 2254 | cur_brgb.oc_block = ocb * min_oc_block; |
| 2255 | cur_brgb.nb_oc = utils::div_up(jcp.oc, cur_brgb.oc_block); |
| 2256 | |
| 2257 | if (!cur_brgb.fast_check_oc_block_1x1()) continue; |
| 2258 | |
| 2259 | cur_brgb.calc_blocks_1x1(); |
| 2260 | const status_t st = cur_brgb.get_brgemm_ur(&attr, dst_md); |
| 2261 | if (st != status::success) continue; |
| 2262 | cur_brgb.eff = cur_brgb.est_eff_1x1(); |
| 2263 | if (cur_brgb.eff > best_brgb.eff) best_brgb = cur_brgb; |
| 2264 | } |
| 2265 | best_brgb.save_to_jcp(jcp); |
| 2266 | |
| 2267 | // =============== end blocking ================================= |
| 2268 | jcp.brg_stride_a = jcp.ic_block * jcp.src_dsz; |
| 2269 | jcp.brg_stride_b = jcp.ic_block * jcp.oc * jcp.wei_dsz; |
| 2270 | |
| 2271 | if (jcp.ic_block == 0 || jcp.oc_block == 0) return status::unimplemented; |
| 2272 | |
| 2273 | // Configure matrix sizes |
| 2274 | |
| 2275 | if (best_brgb.is_os_blocking) { |
| 2276 | if (jcp.os_block == 0) return status::unimplemented; |
| 2277 | jcp.M = jcp.brgM = jcp.os_block; |
| 2278 | jcp.M_tail = jcp.brgM_tail = jcp.os % jcp.os_block; |
| 2279 | } else { |
| 2280 | if (jcp.ow_block == 0) return status::unimplemented; |
| 2281 | jcp.M = jcp.brgM = jcp.ow_block; |
| 2282 | jcp.M_tail = jcp.brgM_tail = jcp.ow % jcp.ow_block; |
| 2283 | } |
| 2284 | |
| 2285 | jcp.K = jcp.ic >= jcp.ic_block ? jcp.ic_block : 0; |
| 2286 | jcp.N = jcp.oc >= jcp.oc_block ? jcp.oc_block : 0; |
| 2287 | jcp.N_tail = jcp.oc % jcp.oc_block; |
| 2288 | jcp.K_tail = jcp.ic % jcp.ic_block; |
| 2289 | |
| 2290 | jcp.gemm_batch_size = jcp.nb_ic_blocking; |
| 2291 | // to avoid cache concurrent access from different threads |
| 2292 | size_t sc_size = sizeof(brgemm_batch_element_t); |
| 2293 | jcp.adjusted_batch_size |
| 2294 | = div_up(rnd_up(jcp.gemm_batch_size * sc_size, P4K), sc_size); |
| 2295 | |
| 2296 | if (is_amx(isa)) { |
| 2297 | // heuristic for small mb |
| 2298 | const bool is_small_mb = jcp.mb == 1 && jcp.ic * jcp.oh <= 28 * 1024 |
| 2299 | && jcp.oc * jcp.oh <= 14 * 1024; |
| 2300 | // non-unrolled kernel does not support bf32, only dispatch unrolled |
| 2301 | // kernel for now |
| 2302 | jcp.use_uker = jcp.is_bf32 || !is_small_mb; |
| 2303 | } |
| 2304 | |
| 2305 | // TODO: heuristic to dispatch BF32 BRGeMM |
| 2306 | // The following condition checks for shapes where down-convert execution |
| 2307 | // in brgemm fails |
| 2308 | if (jcp.is_bf32 && jcp.ic < 64 && jcp.ic % 32 != 0) |
| 2309 | return status::unimplemented; |
| 2310 | |
| 2311 | if (jcp.use_uker) |
| 2312 | jcp.hint_prefetching = brgemm_kernel_prefetching_t::brgemm_prf1; |
| 2313 | CHECK(pick_tags(jcp, src_md, weights_md, dst_md, bias_md)); |
| 2314 | CHECK(attr.set_default_formats(&dst_md)); |
| 2315 | |
| 2316 | const bool with_groups = weights_d.ndims() == src_d.ndims() + 1; |
| 2317 | |
| 2318 | const auto &src_scales = attr.scales_.get(DNNL_ARG_SRC); |
| 2319 | const auto &wei_scales = attr.scales_.get(DNNL_ARG_WEIGHTS); |
| 2320 | const auto &dst_scales = attr.scales_.get(DNNL_ARG_DST); |
| 2321 | jcp.with_scales = !src_scales.has_default_values() |
| 2322 | || !wei_scales.has_default_values(); |
| 2323 | const int wei_mask_per_oc = 1 << (int)with_groups; |
| 2324 | jcp.is_oc_scale = wei_scales.mask_ == wei_mask_per_oc; |
| 2325 | |
| 2326 | // only common and per-oc-channel scales are supported |
| 2327 | const bool scales_ok = one_of(wei_scales.mask_, 0, wei_mask_per_oc) |
| 2328 | && src_scales.mask_ == 0 && dst_scales.has_default_values(); |
| 2329 | if (!scales_ok) return status::unimplemented; |
| 2330 | |
| 2331 | // no inp buffer or brgemm_vpad for 1x1 |
| 2332 | constexpr int align_size = platform::get_cache_line_size(); |
| 2333 | jcp.exec_type = jcp.is_rtus ? exec_trans : exec_base; |
| 2334 | jcp.inp_buffer_size |
| 2335 | = jcp.is_rtus ? rnd_up(jcp.LDA * jcp.os, align_size) : 0; |
| 2336 | jcp.inp_buffer_mask_size = jcp.is_rtus |
| 2337 | ? rnd_up(div_up(jcp.nb_ic, jcp.nb_ic_blocking) * jcp.nb_os, |
| 2338 | align_size) |
| 2339 | : 0; |
| 2340 | jcp.buffer_size = jcp.LDC * jcp.M; |
| 2341 | |
| 2342 | if (jcp.s8s8_avx512) { |
| 2343 | weights_md.extra.flags = 0 | memory_extra_flags::compensation_conv_s8s8; |
| 2344 | weights_md.extra.compensation_mask = with_groups ? 0x3 : 0x1; |
| 2345 | } |
| 2346 | if (jcp.src_zero_point) { |
| 2347 | weights_md.extra.flags |
| 2348 | |= memory_extra_flags::compensation_conv_asymmetric_src; |
| 2349 | weights_md.extra.asymm_compensation_mask = with_groups ? 0x3 : 0x1; |
| 2350 | } |
| 2351 | jcp.req_cal_comp_pad = false; |
| 2352 | jcp.s8s8_comp_buffer_size = jcp.ngroups * jcp.nb_oc * jcp.oc_block; |
| 2353 | jcp.comp_a_buffer_size = jcp.ngroups * jcp.nb_oc * jcp.oc_block; |
| 2354 | |
| 2355 | return status::success; |
| 2356 | } |
| 2357 | |
| 2358 | void set_amx_wsp_per_thread(jit_brgemm_conv_conf_t &jcp) { |
| 2359 | // ensure buffers for individual threads do not lie on same page and also |
| 2360 | // they are not contiguous. |
| 2361 | jcp.amx_buf_size_per_thread |
| 2362 | = utils::rnd_up(jcp.amx_buf_size_per_thread + 1, P4K); |
| 2363 | } |
| 2364 | |
| 2365 | void init_scratchpad(memory_tracking::registrar_t &scratchpad, |
| 2366 | const jit_brgemm_conv_conf_t &jcp) { |
| 2367 | if (jcp.brg_type == brgemm_addr || jcp.brg_type == brgemm_offs |
| 2368 | || (jcp.brg_type == brgemm_strd && jcp.exec_type == exec_vpad)) |
| 2369 | scratchpad.book(key_brgemm_primitive_batch, |
| 2370 | static_cast<size_t>(jcp.nthr) * jcp.adjusted_batch_size, |
| 2371 | sizeof(brgemm_batch_element_t), 64, P4K); |
| 2372 | if (jcp.exec_type == exec_trans) { |
| 2373 | size_t inp_buffer_size |
| 2374 | = static_cast<size_t>(jcp.nthr) * jcp.inp_buffer_size; |
| 2375 | scratchpad.book(key_conv_brgemm_inp_buffer, inp_buffer_size, |
| 2376 | jcp.src_dsz, 0, P4K); |
| 2377 | size_t inp_buffer_mask_size |
| 2378 | = static_cast<size_t>(jcp.nthr) * jcp.inp_buffer_mask_size; |
| 2379 | scratchpad.book(key_conv_brgemm_inp_buffer_mask, inp_buffer_mask_size, |
| 2380 | sizeof(uint8_t), 0, P4K); |
| 2381 | } |
| 2382 | if (jcp.use_buffer) { |
| 2383 | scratchpad.book(key_brgemm_primitive_buffer, jcp.nthr * jcp.buffer_size, |
| 2384 | jcp.acc_dsz, 0, P4K); |
| 2385 | } |
| 2386 | if (is_amx(jcp.isa)) { |
| 2387 | scratchpad.book(key_conv_amx_tile_buffer, |
| 2388 | jcp.nthr * jcp.amx_buf_size_per_thread, sizeof(char), 0, P4K); |
| 2389 | } |
| 2390 | if (jcp.s8s8_avx512 && jcp.req_cal_comp_pad) { |
| 2391 | scratchpad.book(key_brgemm_primitive_buffer_comp, |
| 2392 | jcp.s8s8_comp_buffer_size, sizeof(int32_t), 0, P4K); |
| 2393 | } |
| 2394 | if (jcp.src_zero_point && jcp.req_cal_comp_pad && !is_amx(jcp.isa)) { |
| 2395 | scratchpad.book(key_brgemm_primitive_zp_comp_a, jcp.comp_a_buffer_size, |
| 2396 | sizeof(int32_t), 0, P4K); |
| 2397 | } |
| 2398 | } |
| 2399 | |
| 2400 | void balance_bwd_w(jit_brgemm_conv_conf_t &jcp) { |
| 2401 | |
| 2402 | const auto os_chunks = jcp.nthr_mb_work; |
| 2403 | const auto oc_chunks = div_up(jcp.nb_oc, jcp.nb_oc_blocking); |
| 2404 | const auto ic_chunks = div_up(jcp.nb_ic, jcp.nb_ic_blocking); |
| 2405 | |
| 2406 | auto calc_mem_cost = [=](int nthr_mb, int nthr_g, int nthr_oc_b, |
| 2407 | int nthr_ic_b) { |
| 2408 | /* calculate per thread memory cost (read/write). high level |
| 2409 | * optimizer tries to minimize memory consumption. few notes: |
| 2410 | * (n1) if weights tensor size is less than source and destination |
| 2411 | * tensors we apply the ratio of the source and destination |
| 2412 | * tensor sizes to weights one as compensation coefficient to |
| 2413 | * avoid parallelization across batch size only, otherwise we |
| 2414 | * apply additional coefficient to source component based on |
| 2415 | * performance measurements |
| 2416 | * (n2) use scales based on output vs input channels ratio for |
| 2417 | * source and destination components to improve threading |
| 2418 | * balance across input and output channels */ |
| 2419 | |
| 2420 | const dim_t src_type_size = 2; |
| 2421 | const dim_t wei_type_size = 4; |
| 2422 | const dim_t acc_type_size = wei_type_size; |
| 2423 | |
| 2424 | const auto wei_ks = jcp.kh * jcp.kw * jcp.kd; |
| 2425 | |
| 2426 | const auto src_spatial = (dim_t)jcp.mb * jcp.id * jcp.ih * jcp.tr_iw; |
| 2427 | const auto dst_spatial = (dim_t)jcp.mb * jcp.od * jcp.oh * jcp.tr_ow; |
| 2428 | |
| 2429 | dim_t src_size = src_spatial * jcp.ic * src_type_size; |
| 2430 | dim_t dst_size = dst_spatial * jcp.oc * src_type_size; |
| 2431 | dim_t wei_size = (dim_t)jcp.oc * jcp.ic * wei_ks * wei_type_size; |
| 2432 | |
| 2433 | float wei_compensation_scale = 0.5f * (dst_size + src_size) / wei_size; |
| 2434 | float oi_channels_ratio = (float)(oc_chunks) / ic_chunks; |
| 2435 | |
| 2436 | auto get_src_coef = [=]() { |
| 2437 | float src_coef = nstl::max(1.0f / oi_channels_ratio, 1.0f); |
| 2438 | if (wei_compensation_scale < 1.0f) src_coef *= 4.0f; |
| 2439 | return src_coef; |
| 2440 | }; |
| 2441 | |
| 2442 | auto get_dst_coef |
| 2443 | = [=]() { return nstl::max(oi_channels_ratio, 1.0f); }; |
| 2444 | |
| 2445 | auto get_wei_coef |
| 2446 | = [=]() { return nstl::max(wei_compensation_scale, 1.0f); }; |
| 2447 | |
| 2448 | const float src_coef = get_src_coef(); |
| 2449 | const float dst_coef = get_dst_coef(); |
| 2450 | const float wei_coef = get_wei_coef(); |
| 2451 | |
| 2452 | const auto thr_mb = div_up(os_chunks, nthr_mb); |
| 2453 | const auto nb_oc_job = jcp.oc_block * jcp.nb_oc_blocking; |
| 2454 | const auto nb_ic_job = jcp.ic_block * jcp.nb_ic_blocking; |
| 2455 | |
| 2456 | const auto src_chunk = src_spatial / os_chunks; |
| 2457 | const auto dst_chunk = dst_spatial / os_chunks; |
| 2458 | |
| 2459 | const auto thr_g = div_up(jcp.ngroups, nthr_g); |
| 2460 | const auto thr_ic_b = div_up(ic_chunks, nthr_ic_b); |
| 2461 | const auto thr_src_sp = thr_mb * src_chunk / jcp.stride_d / jcp.stride_h |
| 2462 | / jcp.stride_w; |
| 2463 | const auto thr_dst_sp = thr_mb * dst_chunk; |
| 2464 | const auto thr_ic_amount = thr_ic_b * nb_ic_job; |
| 2465 | |
| 2466 | const auto thr_oc_b = div_up(oc_chunks, nb_oc_job * nthr_oc_b); |
| 2467 | |
| 2468 | const auto thr_oc_amount = thr_oc_b * nb_oc_job; |
| 2469 | float src_v |
| 2470 | = src_type_size * src_coef * thr_g * thr_ic_amount * thr_src_sp; |
| 2471 | float dst_v |
| 2472 | = src_type_size * dst_coef * thr_g * thr_oc_amount * thr_dst_sp; |
| 2473 | float wei_v = acc_type_size * wei_coef * thr_g * thr_oc_amount |
| 2474 | * thr_ic_amount * wei_ks; |
| 2475 | |
| 2476 | return src_v + dst_v + wei_v; |
| 2477 | }; |
| 2478 | |
| 2479 | auto balance = [=](int &nthr_, int &nthr_mb_, int &nthr_g_, int &nthr_oc_b_, |
| 2480 | int &nthr_ic_b_) { |
| 2481 | nthr_ = nthr_mb_ = nthr_g_ = nthr_oc_b_ = nthr_ic_b_ = 1; |
| 2482 | |
| 2483 | if (jcp.nthr < jcp.ngroups) { |
| 2484 | /* simplification... fortunately it doesn't hurt much */ |
| 2485 | nthr_ = nthr_g_ = jcp.nthr; |
| 2486 | return; |
| 2487 | } |
| 2488 | |
| 2489 | nthr_g_ = jcp.ngroups; |
| 2490 | const int nthr = jcp.nthr / nthr_g_; |
| 2491 | |
| 2492 | float best_mem_cost |
| 2493 | = calc_mem_cost(nthr_mb_, nthr_g_, nthr_oc_b_, nthr_ic_b_); |
| 2494 | |
| 2495 | /* find the best thread distribution with lowest memory cost */ |
| 2496 | |
| 2497 | const int nthr_mb_max = nstl::min(nthr, jcp.nthr_mb_work); |
| 2498 | for (int nthr_mb = 1; nthr_mb <= nthr_mb_max; ++nthr_mb) { |
| 2499 | const int nthr_par = nthr / nthr_mb; |
| 2500 | const int nthr_oc_b_max = nstl::min(nthr_par, |
| 2501 | oc_chunks); // Amount of nb_oc_blocks |
| 2502 | for (int nthr_oc_b = 1; nthr_oc_b <= nthr_oc_b_max; ++nthr_oc_b) { |
| 2503 | int nthr_ic_b = nstl::min( |
| 2504 | nthr_par / nthr_oc_b, (jcp.nb_ic / jcp.nb_ic_blocking)); |
| 2505 | |
| 2506 | float mem_cost |
| 2507 | = calc_mem_cost(nthr_mb, nthr_g_, nthr_oc_b, nthr_ic_b); |
| 2508 | if (mem_cost <= best_mem_cost) { |
| 2509 | best_mem_cost = mem_cost; |
| 2510 | nthr_mb_ = nthr_mb; |
| 2511 | nthr_oc_b_ = nthr_oc_b; |
| 2512 | nthr_ic_b_ = nthr_ic_b; |
| 2513 | } |
| 2514 | } |
| 2515 | } |
| 2516 | |
| 2517 | if (nthr_mb_ > nthr / 2 && nthr_mb_ < nthr) |
| 2518 | nthr_mb_ = nstl::min(jcp.nthr_mb_work, nthr); |
| 2519 | nthr_ = nthr_mb_ * nthr_g_ * nthr_oc_b_ * nthr_ic_b_; |
| 2520 | |
| 2521 | assert(nthr_ <= jcp.nthr); |
| 2522 | }; |
| 2523 | |
| 2524 | int nthr, nthr_mb, nthr_g, nthr_oc_b, nthr_ic_b; |
| 2525 | balance(nthr, nthr_mb, nthr_g, nthr_oc_b, nthr_ic_b); |
| 2526 | |
| 2527 | // empiric balancing for some shapes |
| 2528 | const auto sps = (jcp.ih * jcp.iw); |
| 2529 | bool neat_1x1 |
| 2530 | = everyone_is(1, jcp.id, jcp.kh, jcp.kw, jcp.ngroups, jcp.stride_h); |
| 2531 | if (neat_1x1 && jcp.nthr >= 28 && jcp.mb >= jcp.nthr) { |
| 2532 | const bool more_oc = (jcp.ic < jcp.oc); |
| 2533 | if (sps >= 56 * 56 && jcp.ic >= 64 && jcp.oc >= 64) { |
| 2534 | nthr_mb = jcp.nthr; |
| 2535 | nthr_oc_b = 1; |
| 2536 | } else if (sps >= 28 * 28 && jcp.ic >= 128 && jcp.oc >= 128) { |
| 2537 | nthr_mb = jcp.nthr / 4; |
| 2538 | nthr_oc_b = more_oc ? jcp.nthr / nthr_mb : 1; |
| 2539 | } else if (sps >= 14 * 14 && jcp.ic >= 256 && jcp.oc >= 256) { |
| 2540 | nthr_mb = div_up(jcp.nthr, 8); |
| 2541 | nthr_oc_b = more_oc ? jcp.nthr / nthr_mb : 1; |
| 2542 | } else if (sps >= 7 * 7 && jcp.ic >= 512 && jcp.oc >= 512) { |
| 2543 | nthr_mb = div_up(jcp.nthr, 14); |
| 2544 | nthr_oc_b = more_oc ? jcp.nthr / nthr_mb : 1; |
| 2545 | } |
| 2546 | nthr_ic_b = jcp.nthr / (nthr_mb * nthr_oc_b); |
| 2547 | nthr = nthr_mb * nthr_g * nthr_oc_b * nthr_ic_b; |
| 2548 | } |
| 2549 | |
| 2550 | jcp.nthr = nthr; |
| 2551 | jcp.nthr_mb = nthr_mb; |
| 2552 | jcp.nthr_g = nthr_g; |
| 2553 | jcp.nthr_oc_b = nthr_oc_b; |
| 2554 | jcp.nthr_ic_b = nthr_ic_b; |
| 2555 | |
| 2556 | // TODO: Optimize memory allocation when threaded on height and depth |
| 2557 | jcp.tr_src_buf_size = jcp.tr_iw * jcp.ic_block * jcp.ih * jcp.id; |
| 2558 | jcp.tr_diff_dst_buf_size = jcp.tr_ow * jcp.oc_block * jcp.oh * jcp.od; |
| 2559 | jcp.tr_src_buf_count = jcp.global_transpose |
| 2560 | ? jcp.nthr_mb * jcp.nb_ic * jcp.ngroups |
| 2561 | : jcp.nthr; |
| 2562 | jcp.tr_diff_dst_buf_count = jcp.global_transpose |
| 2563 | ? jcp.nthr_mb * jcp.nb_oc * jcp.ngroups |
| 2564 | : jcp.nthr; |
| 2565 | } |
| 2566 | |
| 2567 | status_t init_conf_bwd_w(jit_brgemm_conv_conf_t &jcp, |
| 2568 | const convolution_desc_t &cd, memory_desc_t &src_md, |
| 2569 | memory_desc_t &diff_weights_md, memory_desc_t &diff_bias_md, |
| 2570 | memory_desc_t &diff_dst_md, primitive_attr_t &attr, int nthreads) { |
| 2571 | |
| 2572 | const memory_desc_wrapper src_d(&src_md); |
| 2573 | const memory_desc_wrapper diff_weights_d(&diff_weights_md); |
| 2574 | const memory_desc_wrapper diff_dst_d(&diff_dst_md); |
| 2575 | const memory_desc_wrapper diff_bias_d(&diff_bias_md); |
| 2576 | |
| 2577 | const bool is_f16 = src_d.data_type() == data_type::f16; |
| 2578 | |
| 2579 | jcp.isa = is_f16 ? avx512_core_amx_fp16 : avx512_core_amx; |
| 2580 | if (!mayiuse(jcp.isa)) return status::unimplemented; |
| 2581 | |
| 2582 | const bool with_groups = diff_weights_d.ndims() == src_d.ndims() + 1; |
| 2583 | int ndims = src_d.ndims(); |
| 2584 | |
| 2585 | CHECK(init_jcp(jcp, jcp.isa, cd, src_md, diff_weights_md, diff_dst_md, |
| 2586 | diff_bias_md, attr, nthreads)); |
| 2587 | |
| 2588 | jcp.max_batch = jcp.od * jcp.oh; |
| 2589 | jcp.brg_type = brgemm_addr; // TODO: Choose right type of BRGEMM |
| 2590 | jcp.use_uker = true; |
| 2591 | jcp.var_bs = true; |
| 2592 | |
| 2593 | // Process some 1x1 convolutions with small iw as 1d (h=1, w = h*w) |
| 2594 | // convolutions to make brgemm K dimension bigger for better utilization of |
| 2595 | // AMX tiles |
| 2596 | bool neat_1x1_2d = (everyone_is( |
| 2597 | 1, jcp.kh, jcp.kw, jcp.stride_h, jcp.stride_w) |
| 2598 | && everyone_is(0, jcp.t_pad, jcp.b_pad, jcp.l_pad, jcp.r_pad)); |
| 2599 | bool make_1d = neat_1x1_2d && jcp.iw <= 28; |
| 2600 | if (make_1d) { |
| 2601 | jcp.iw *= jcp.ih; |
| 2602 | jcp.ih = 1; |
| 2603 | jcp.ow *= jcp.oh; |
| 2604 | jcp.oh = 1; |
| 2605 | jcp.max_batch = jcp.od; |
| 2606 | } |
| 2607 | // TODO: sometimes we can call brgemm kernel with bs = 0 to do initialization |
| 2608 | // review this condition |
| 2609 | if (jcp.max_batch == 1 |
| 2610 | && everyone_is(0, jcp.f_pad, jcp.back_pad, jcp.t_pad, jcp.b_pad)) |
| 2611 | jcp.var_bs = false; |
| 2612 | |
| 2613 | jcp.has_vnni = true; // Needed for transpose routines |
| 2614 | |
| 2615 | jcp.typesize_in = sizeof(bfloat16_t); |
| 2616 | jcp.typesize_out = sizeof(float); |
| 2617 | |
| 2618 | bool ok = true |
| 2619 | // general condition to simplify dilations |
| 2620 | && IMPLICATION(jcp.dilate_d != 0, jcp.stride_d == 1) |
| 2621 | && IMPLICATION(jcp.dilate_h != 0, jcp.stride_h == 1) |
| 2622 | && IMPLICATION(jcp.dilate_w != 0, jcp.stride_w == 1) |
| 2623 | // special condition to simplify dilations in compute_oh_loop_common |
| 2624 | && IMPLICATION(jcp.dilate_h != 0, jcp.ext_kh <= jcp.ih); |
| 2625 | if (!ok) return status::unimplemented; |
| 2626 | |
| 2627 | jcp.transform_to_vnni = diff_weights_d.data_type() != data_type::f32; |
| 2628 | |
| 2629 | /* XXX: no support for padding when dilation_d > 0 */ |
| 2630 | if (!IMPLICATION(jcp.dilate_d > 0, everyone_is(0, jcp.back_pad, jcp.f_pad))) |
| 2631 | return status::unimplemented; |
| 2632 | |
| 2633 | const bool is_depthwise = true && with_groups && jcp.ngroups > 1 |
| 2634 | && everyone_is(1, jcp.ic, jcp.oc); |
| 2635 | if (is_depthwise) |
| 2636 | return status::unimplemented; // TODO: add support of DW convolution |
| 2637 | |
| 2638 | const int dat_format_tag = ndims - 3; |
| 2639 | format_tag_t dat_tag_nspc = utils::pick(dat_format_tag, format_tag::nwc, |
| 2640 | format_tag::nhwc, format_tag::ndhwc); |
| 2641 | format_tag_t dat_tag_opt = dat_tag_nspc; |
| 2642 | |
| 2643 | if (src_d.format_kind() == format_kind::any) { |
| 2644 | CHECK(memory_desc_init_by_tag(src_md, dat_tag_opt)); |
| 2645 | jcp.src_tag = dat_tag_opt; |
| 2646 | } else |
| 2647 | jcp.src_tag = src_d.matches_one_of_tag(dat_tag_opt); |
| 2648 | if (!one_of(jcp.src_tag, dat_tag_opt)) return status::unimplemented; |
| 2649 | |
| 2650 | const bool is_nspc = jcp.src_tag == dat_tag_nspc; |
| 2651 | if (!is_nspc) return status::unimplemented; |
| 2652 | |
| 2653 | if (diff_dst_d.format_kind() == format_kind::any) { |
| 2654 | CHECK(memory_desc_init_by_tag(diff_dst_md, jcp.src_tag)); |
| 2655 | jcp.dst_tag = jcp.src_tag; |
| 2656 | } else |
| 2657 | jcp.dst_tag = diff_dst_d.matches_one_of_tag(jcp.src_tag); |
| 2658 | if (jcp.dst_tag != jcp.src_tag) return status::unimplemented; |
| 2659 | |
| 2660 | const int wei_format_tag = 2 * ndims - 6 + with_groups; |
| 2661 | format_tag_t wei_tag; |
| 2662 | if (jcp.transform_to_vnni) |
| 2663 | wei_tag = pick(wei_format_tag, format_tag::OIw16i16o2i, |
| 2664 | format_tag::gOIw16i16o2i, format_tag::OIhw16i16o2i, |
| 2665 | format_tag::gOIhw16i16o2i, format_tag::OIdhw16i16o2i, |
| 2666 | format_tag::gOIdhw16i16o2i); |
| 2667 | else |
| 2668 | wei_tag = pick(wei_format_tag, format_tag::OIw16i16o, |
| 2669 | format_tag::gOIw16i16o, format_tag::OIhw16i16o, |
| 2670 | format_tag::gOIhw16i16o, format_tag::OIdhw16i16o, |
| 2671 | format_tag::gOIdhw16i16o); |
| 2672 | if (diff_weights_md.format_kind == format_kind::any) { |
| 2673 | CHECK(memory_desc_init_by_tag(diff_weights_md, wei_tag)); |
| 2674 | jcp.wei_tag = wei_tag; |
| 2675 | } else { |
| 2676 | jcp.wei_tag = diff_weights_d.matches_one_of_tag(wei_tag); |
| 2677 | if (jcp.wei_tag != wei_tag) return status::unimplemented; |
| 2678 | } |
| 2679 | jcp.wei_dt = diff_weights_d.data_type(); |
| 2680 | |
| 2681 | /* kernel applicability check wrt boundaries |
| 2682 | * the conditions are quite general across the kernels we have, |
| 2683 | * but ideally the check should belong to a specific kernel... */ |
| 2684 | const int max_pad_h = jcp.ext_kh / 2; |
| 2685 | const bool boundaries_ok = true && jcp.l_pad < jcp.ext_kw |
| 2686 | && jcp.r_pad < jcp.ext_kw && jcp.t_pad <= max_pad_h |
| 2687 | && jcp.b_pad <= max_pad_h && jcp.f_pad < jcp.ext_kd |
| 2688 | && jcp.back_pad < jcp.ext_kd; |
| 2689 | if (!boundaries_ok) return status::unimplemented; |
| 2690 | |
| 2691 | jcp.ic_block = 16; |
| 2692 | jcp.oc_block = 16; |
| 2693 | |
| 2694 | jcp.nb_ic = utils::div_up(jcp.ic, jcp.ic_block); |
| 2695 | jcp.nb_oc = utils::div_up(jcp.oc, jcp.oc_block); |
| 2696 | |
| 2697 | jcp.ic_tail = jcp.ic % jcp.ic_block; |
| 2698 | jcp.oc_tail = jcp.oc % jcp.oc_block; |
| 2699 | |
| 2700 | jcp.nb_oc_blocking = (jcp.nb_oc > 1) ? 2 : 1; |
| 2701 | jcp.nb_ic_blocking = (jcp.nb_ic > 1) ? 2 : 1; |
| 2702 | |
| 2703 | const bool is_2d = (ndims == 4); |
| 2704 | const bool is_3d = (ndims == 5); |
| 2705 | |
| 2706 | // TODO: Find more shapes (especially 3D with large spatials) for which |
| 2707 | // local transposition will be beneficial. Furthermore, for TBB threads |
| 2708 | // more shapes can potentially benefit from spatial blocking |
| 2709 | int optimal_blk_size = is_3d ? jcp.od : is_2d ? jcp.oh : jcp.ow; |
| 2710 | |
| 2711 | jcp.global_transpose = dnnl_thr_syncable(); |
| 2712 | jcp.spatial_blk_size = optimal_blk_size; |
| 2713 | |
| 2714 | const int tr_round = 32; // To load full tile register |
| 2715 | int tr_pad |
| 2716 | = rnd_up(nstl::max(jcp.l_pad, jcp.r_pad + 1), tr_round); //!!! why? |
| 2717 | jcp.tr_iw = rnd_up(div_up(jcp.iw + jcp.l_pad + jcp.r_pad, jcp.stride_w), |
| 2718 | tr_round) |
| 2719 | * jcp.stride_w; |
| 2720 | |
| 2721 | // TODO: xf16 training is supported only |
| 2722 | const auto rnd_val = data_type_vnni_granularity(bf16); |
| 2723 | jcp.tr_src_num_guard_elems = tr_pad; // upper bound |
| 2724 | jcp.tr_ow = rnd_up(jcp.ow, rnd_val); |
| 2725 | if (jcp.tr_ow > tr_round) { |
| 2726 | // we may increase tr_ow to have better bd_block in brgemm kernel |
| 2727 | int best_bdb = jcp.tr_ow / rnd_val; |
| 2728 | int best_tr_ow = jcp.tr_ow; |
| 2729 | for (int tr_ow = jcp.tr_ow; tr_ow <= rnd_up(jcp.tr_ow, tr_round); |
| 2730 | tr_ow += rnd_val) { |
| 2731 | for (int i = tr_round; i > 0; i -= rnd_val) { |
| 2732 | if (tr_ow % i == 0) { |
| 2733 | const auto cbdb = tr_ow / i; |
| 2734 | if (cbdb < best_bdb) { |
| 2735 | best_bdb = cbdb; |
| 2736 | best_tr_ow = tr_ow; |
| 2737 | } |
| 2738 | break; |
| 2739 | } |
| 2740 | } |
| 2741 | } |
| 2742 | jcp.tr_ow = best_tr_ow; |
| 2743 | } |
| 2744 | |
| 2745 | bool args_ok = true && jcp.ic <= src_d.padded_dims()[1] |
| 2746 | && jcp.oc <= diff_dst_d.padded_dims()[1] |
| 2747 | && jcp.ic <= diff_weights_d.padded_dims()[with_groups + 1] |
| 2748 | && jcp.oc <= diff_weights_d.padded_dims()[with_groups + 0]; |
| 2749 | if (!args_ok) return status::unimplemented; |
| 2750 | |
| 2751 | jcp.harness = ndims == 5 ? harness_3d_reduction : harness_2d_reduction; |
| 2752 | |
| 2753 | if (!one_of(jcp.harness, harness_2d_reduction, harness_3d_reduction)) { |
| 2754 | return status::unimplemented; |
| 2755 | } |
| 2756 | |
| 2757 | switch (jcp.harness) { |
| 2758 | case harness_2d_reduction: jcp.nthr_mb_work = jcp.mb * jcp.oh; break; |
| 2759 | case harness_3d_reduction: jcp.nthr_mb_work = jcp.mb * jcp.od; break; |
| 2760 | default: assert(!"Invalid harness"); jcp.nthr_mb_work = jcp.mb; |
| 2761 | } |
| 2762 | |
| 2763 | balance_bwd_w(jcp); |
| 2764 | |
| 2765 | if (one_of(jcp.harness, harness_2d_reduction, harness_3d_reduction)) { |
| 2766 | jcp.K = jcp.tr_ow; |
| 2767 | } |
| 2768 | jcp.K_tail = 0; |
| 2769 | |
| 2770 | jcp.M = jcp.ic_block * jcp.nb_ic_blocking; |
| 2771 | // assumption that jcp.nb_ic_blocking is always 2 |
| 2772 | if (jcp.nb_ic % jcp.nthr_ic_b == 0 |
| 2773 | && (jcp.nb_ic / jcp.nthr_ic_b) % jcp.nb_ic_blocking == 0) |
| 2774 | jcp.M_tail = 0; |
| 2775 | else |
| 2776 | jcp.M_tail = jcp.ic_block; |
| 2777 | |
| 2778 | jcp.N = jcp.oc_block * jcp.nb_oc_blocking; |
| 2779 | // assumption that jcp.nb_oc_blocking is always 2 |
| 2780 | if (jcp.nb_oc % jcp.nthr_oc_b == 0 |
| 2781 | && (jcp.nb_oc / jcp.nthr_oc_b) % jcp.nb_oc_blocking == 0) |
| 2782 | jcp.N_tail = 0; |
| 2783 | else |
| 2784 | jcp.N_tail = jcp.oc_block; |
| 2785 | |
| 2786 | // for convolutions with big spatial: transpose only chunk |
| 2787 | // (oc_block * nb_oc_blocking) of diff_dst on each iteration by oc blocks |
| 2788 | // for better cache utilization |
| 2789 | // the equal number of channel blocks per thread is required to use this |
| 2790 | // approach to avoid hangs |
| 2791 | bool tr_ocb_chunk_allowed = (jcp.nb_oc % jcp.nthr_oc_b == 0); |
| 2792 | jcp.tr_ocb_chunk = tr_ocb_chunk_allowed && (jcp.oh * jcp.ow > 38 * 38); |
| 2793 | jcp.tr_icb_chunk = false; |
| 2794 | |
| 2795 | const int irow_size = jcp.src_dsz * jcp.tr_iw * jcp.ic_block |
| 2796 | * div_up(jcp.nb_ic, jcp.nthr_ic_b) |
| 2797 | * 2 /*we have real and transposed input */; |
| 2798 | const int orow_size = jcp.dst_dsz * jcp.tr_ow * jcp.oc_block |
| 2799 | * div_up(jcp.nb_oc, jcp.nthr_oc_b) |
| 2800 | * 2 /*we have real and transposed diff_dst*/; |
| 2801 | int oh_block_limit = nstl::max(1.f, |
| 2802 | nstl::max(0.f, 0.8f * brg_blocking_t::L2 - jcp.kh * irow_size) |
| 2803 | / (irow_size + orow_size)); |
| 2804 | // try to split oh by equal oh blocks |
| 2805 | oh_block_limit = div_up(jcp.oh, div_up(jcp.oh, oh_block_limit)); |
| 2806 | jcp.oh_block = utils::saturate(1, jcp.oh, oh_block_limit); |
| 2807 | |
| 2808 | const int iframe_size = irow_size * jcp.id; |
| 2809 | const int oframe_size = orow_size * jcp.od; |
| 2810 | int od_block_limit = nstl::max(1.f, |
| 2811 | nstl::max(0.f, 0.8f * brg_blocking_t::L2 - jcp.kd * iframe_size) |
| 2812 | / (iframe_size + oframe_size)); |
| 2813 | // try to split od by equal od blocks |
| 2814 | od_block_limit = div_up(jcp.od, div_up(jcp.od, od_block_limit)); |
| 2815 | jcp.od_block = utils::saturate(1, jcp.od, od_block_limit); |
| 2816 | |
| 2817 | jcp.use_interleave_stores = false; |
| 2818 | jcp.hint_prefetching = brgemm_kernel_prefetching_t::brgemm_prf1; |
| 2819 | jcp.amx_tile_load_xx = false; |
| 2820 | |
| 2821 | if (one_of(jcp.harness, harness_2d_reduction, harness_3d_reduction)) { |
| 2822 | jcp.LDA = jcp.tr_iw; |
| 2823 | jcp.LDB = jcp.oc_block; |
| 2824 | jcp.LDC = jcp.LDD = jcp.oc_block; |
| 2825 | } |
| 2826 | |
| 2827 | jcp.gemm_batch_size = jcp.max_batch; |
| 2828 | // to avoid cache concurrent access from different threads |
| 2829 | size_t sc_size = sizeof(brgemm_batch_element_t); |
| 2830 | jcp.adjusted_batch_size |
| 2831 | = div_up(rnd_up(jcp.gemm_batch_size * sc_size, P4K), sc_size); |
| 2832 | |
| 2833 | return status::success; |
| 2834 | } |
| 2835 | |
| 2836 | status_t init_scratchpad_bwd_w(memory_tracking::registrar_t &scratchpad, |
| 2837 | const jit_brgemm_conv_conf_t &jcp, memory_desc_t &src_md, |
| 2838 | memory_desc_t &diff_weights_md, memory_desc_t &diff_dst_md) { |
| 2839 | const memory_desc_wrapper src_d(&src_md); |
| 2840 | const memory_desc_wrapper diff_weights_d(&diff_weights_md); |
| 2841 | const memory_desc_wrapper diff_dst_d(&diff_dst_md); |
| 2842 | |
| 2843 | // XXX: See the comment about tr_iw and guarding elements in |
| 2844 | // jit_avx512_core_amx_bwd_weights_kernel_t::init_conf() |
| 2845 | const size_t tr_src_size |
| 2846 | = (jcp.tr_src_buf_count * jcp.tr_src_buf_size * jcp.nb_ic_blocking) |
| 2847 | + jcp.tr_src_num_guard_elems; |
| 2848 | scratchpad.book(key_conv_tr_src, tr_src_size, jcp.src_dsz); |
| 2849 | |
| 2850 | /* prepare synchronization contexts */ |
| 2851 | if (jcp.global_transpose && jcp.nthr_oc_b > 1) { |
| 2852 | const int tr_src_bctx_size = jcp.nthr / jcp.nthr_oc_b; |
| 2853 | scratchpad.book<simple_barrier::ctx_t>( |
| 2854 | key_conv_tr_src_bctx, tr_src_bctx_size); |
| 2855 | } |
| 2856 | |
| 2857 | // The tr_ow <= tr_iw, so we need some guarding at the end of diff_dst |
| 2858 | // TODO: update this guarding: |
| 2859 | // (jcp.tr_diff_dst_buf_size + jcp.tr_iw * jcp.oc_block) |
| 2860 | const auto tr_diff_dst_size = jcp.tr_diff_dst_buf_count |
| 2861 | |
| 2862 | * (jcp.tr_diff_dst_buf_size + jcp.tr_iw * jcp.oc_block) |
| 2863 | * jcp.nb_oc_blocking; |
| 2864 | |
| 2865 | const size_t min_align = 64; |
| 2866 | scratchpad.book( |
| 2867 | key_conv_tr_diff_dst, tr_diff_dst_size, jcp.src_dsz, min_align); |
| 2868 | |
| 2869 | /* prepare synchronization contexts */ |
| 2870 | if (jcp.global_transpose && jcp.nthr_ic_b > 1) { |
| 2871 | const size_t tr_diff_dst_bctx_size = jcp.nthr / jcp.nthr_ic_b; |
| 2872 | scratchpad.book<simple_barrier::ctx_t>( |
| 2873 | key_conv_tr_diff_dst_bctx, tr_diff_dst_bctx_size); |
| 2874 | } |
| 2875 | |
| 2876 | if (IMPLICATION(jcp.nthr_mb == 1, |
| 2877 | (jcp.with_bias && jcp.bia_dt != data_type::f32) |
| 2878 | || jcp.wei_dt != data_type::f32)) { |
| 2879 | const size_t wei_size = jcp.ngroups * jcp.nb_oc * jcp.oc_block |
| 2880 | * jcp.nb_ic * jcp.ic_block * jcp.kh * jcp.kw * jcp.kd; |
| 2881 | const size_t bia_size |
| 2882 | = jcp.with_bias * jcp.ngroups * jcp.nb_oc * jcp.oc_block; |
| 2883 | |
| 2884 | const int num_wei_buffers |
| 2885 | = jcp.wei_dt != data_type::f32 ? jcp.nthr_mb : jcp.nthr_mb - 1; |
| 2886 | const int num_bia_buffers = jcp.with_bias |
| 2887 | ? (jcp.bia_dt != data_type::f32 ? jcp.nthr_mb : jcp.nthr_mb - 1) |
| 2888 | : 0; |
| 2889 | |
| 2890 | const size_t wei_bia_reduction_size |
| 2891 | = wei_size * num_wei_buffers + bia_size * num_bia_buffers; |
| 2892 | |
| 2893 | scratchpad.book<float>( |
| 2894 | key_conv_wei_bia_reduction, wei_bia_reduction_size); |
| 2895 | |
| 2896 | scratchpad.book<simple_barrier::ctx_t>( |
| 2897 | key_conv_wei_bia_reduction_bctx, 1); |
| 2898 | } |
| 2899 | |
| 2900 | if (jcp.with_bias |
| 2901 | && ((jcp.oc % jcp.oc_block != 0) && jcp.bia_dt == data_type::f32)) { |
| 2902 | scratchpad.book(key_conv_padded_bias, |
| 2903 | jcp.ngroups * jcp.nb_oc * jcp.oc_block, jcp.bia_dsz); |
| 2904 | } |
| 2905 | scratchpad.book(key_conv_amx_tilecfg, 1, 64); // 1 whole cacheline |
| 2906 | |
| 2907 | constexpr size_t scratchpad_limit_by_absolute_value = (size_t)32 |
| 2908 | << 30; // 32Gb - TODO: may it's too large? |
| 2909 | const size_t scratchpad_limit_by_tensor_sizes = (size_t)64 * jcp.nthr |
| 2910 | * (src_d.size() + diff_weights_d.size() + diff_dst_d.size()); |
| 2911 | const size_t scratchpad_limit |
| 2912 | = nstl::min(scratchpad_limit_by_absolute_value, |
| 2913 | scratchpad_limit_by_tensor_sizes); |
| 2914 | |
| 2915 | scratchpad.book(key_brgemm_primitive_batch, |
| 2916 | static_cast<size_t>(jcp.nthr) * jcp.adjusted_batch_size, |
| 2917 | sizeof(brgemm_batch_element_t), 64, P4K); |
| 2918 | |
| 2919 | scratchpad.book( |
| 2920 | key_conv_amx_tile_buffer, jcp.nthr * 2 * P4K, sizeof(char), 0, P4K); |
| 2921 | |
| 2922 | if (scratchpad.size() > scratchpad_limit) |
| 2923 | return status::unimplemented; |
| 2924 | else |
| 2925 | return status::success; |
| 2926 | } |
| 2927 | |
| 2928 | } // namespace brgemm_convolution_utils |
| 2929 | |
| 2930 | } // namespace x64 |
| 2931 | } // namespace cpu |
| 2932 | } // namespace impl |
| 2933 | } // namespace dnnl |
| 2934 |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.