| 1 | /******************************************************************************* |
|---|---|
| 2 | * Copyright 2020-2022 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 "cpu/x64/jit_brgemm_inner_product_utils.hpp" |
| 18 | |
| 19 | namespace dnnl { |
| 20 | namespace impl { |
| 21 | namespace cpu { |
| 22 | namespace x64 { |
| 23 | |
| 24 | using namespace dnnl::impl::status; |
| 25 | using namespace dnnl::impl::format_tag; |
| 26 | using namespace dnnl::impl::memory_tracking::names; |
| 27 | using namespace dnnl::impl::utils; |
| 28 | |
| 29 | using namespace prop_kind; |
| 30 | using namespace data_type; |
| 31 | |
| 32 | namespace brgemm_inner_product_utils { |
| 33 | |
| 34 | int get_brg_kernel_index(const jit_brgemm_primitive_conf_t &jbgp, |
| 35 | bool is_bs_tail, bool do_initialization, bool is_M_tail, bool is_N_tail, |
| 36 | bool is_K_tail) { |
| 37 | int idx = 16 * (int)is_bs_tail + 8 * (int)do_initialization |
| 38 | + 4 * (int)is_M_tail + 2 * (int)is_N_tail + (int)is_K_tail; |
| 39 | |
| 40 | assert(idx < max_num_brg_kernels_ip); |
| 41 | return idx; |
| 42 | } |
| 43 | |
| 44 | int get_os_block(const jit_brgemm_primitive_conf_t &jbgp, bool try_to_adjust, |
| 45 | bool is_adjustment) { |
| 46 | const bool is_amx_int8 |
| 47 | = jbgp.isa == avx512_core_amx && one_of(jbgp.wei_dt, s8, u8); |
| 48 | const bool is_amx_xf16 = is_superset(jbgp.isa, avx512_core_amx) |
| 49 | && one_of(jbgp.wei_dt, bf16, f16); |
| 50 | const bool is_avx512_bf16 = jbgp.isa == avx512_core_bf16; |
| 51 | const bool is_f32 = everyone_is(f32, jbgp.src_dt, jbgp.wei_dt, jbgp.dst_dt); |
| 52 | const bool is_bf32 = jbgp.is_bf32; |
| 53 | |
| 54 | int max_os_block = 0; |
| 55 | int min_os_block = 0; |
| 56 | |
| 57 | if (try_to_adjust |
| 58 | || one_of(jbgp.prop_kind, forward_training, forward_inference)) { |
| 59 | min_os_block = (is_amx_int8 || is_amx_xf16) ? 16 : 6; |
| 60 | // Currently gigantic flag is used to separate out transformer_lt and |
| 61 | // alexnet shapes for which larger os_block gives better performance. |
| 62 | // TODO: Figure out how much the constraints for `gigantic-ness` can |
| 63 | // be further loosened. |
| 64 | const bool is_gigantic_shape |
| 65 | = jbgp.ic >= 9216 && jbgp.oc >= 4096 && jbgp.os >= 512; |
| 66 | const bool use_128_block_for_amx |
| 67 | = is_amx_xf16 && jbgp.os % 128 == 0 && jbgp.oc > 128; |
| 68 | const bool enable_128_os_blocking |
| 69 | = use_128_block_for_amx || is_gigantic_shape; |
| 70 | max_os_block = enable_128_os_blocking ? 128 : 64; |
| 71 | // Work done by each thread is given by: |
| 72 | // (nb_oc / nb_oc_blocking) * (nb_os / nb_os_blocking) |
| 73 | // As a first approximation we take nb_oc_blocking = nb_os_blocking = 1 |
| 74 | // Furthermore, we recall that |
| 75 | // nb_oc = oc / oc_block |
| 76 | // nb_os = os / os_block |
| 77 | // |
| 78 | // For f32 data type our objective is to determine the optimal value |
| 79 | // of os_block such that the work amount per thread ~ 2 |
| 80 | if (is_f32 && !is_bf32) { |
| 81 | const bool small_work_amt_per_thread |
| 82 | = div_up(jbgp.os, max_os_block) * jbgp.nb_oc |
| 83 | < 1.8f * jbgp.nthr; |
| 84 | if (small_work_amt_per_thread) |
| 85 | max_os_block = saturate(16, max_os_block, |
| 86 | div_up(jbgp.os * jbgp.nb_oc, 2 * jbgp.nthr)); |
| 87 | } |
| 88 | } else if (jbgp.prop_kind == backward_data) { |
| 89 | int plat_max_os_block = 0; |
| 90 | if (is_amx_xf16) { |
| 91 | plat_max_os_block |
| 92 | = (jbgp.ic >= 512 && jbgp.oc / jbgp.ic <= 4) ? 128 : 64; |
| 93 | } else if (is_avx512_bf16) { |
| 94 | plat_max_os_block = (jbgp.ic > 256) ? 128 : 64; |
| 95 | } else { |
| 96 | plat_max_os_block = 64; |
| 97 | } |
| 98 | max_os_block = nstl::min(plat_max_os_block, jbgp.os); |
| 99 | min_os_block = is_amx_xf16 ? 16 : 6; |
| 100 | } else if (jbgp.prop_kind == backward_weights) { |
| 101 | constexpr int amx_xf16_row = 64; |
| 102 | constexpr int amx_xf16_half_row = amx_xf16_row / 2; |
| 103 | // ensure that os_tail <= amx_xf16_half_row |
| 104 | const bool use_large_os_block = (jbgp.os >= amx_xf16_row) |
| 105 | && (jbgp.os % amx_xf16_row) <= amx_xf16_half_row; |
| 106 | return is_amx_xf16 |
| 107 | ? (use_large_os_block ? amx_xf16_row : amx_xf16_half_row) |
| 108 | : 16; |
| 109 | } else |
| 110 | assert(!"unsupported case"); |
| 111 | |
| 112 | if (is_adjustment) max_os_block /= 2; |
| 113 | int os_block = 1; |
| 114 | for (int osb = max_os_block; osb >= min_os_block; osb--) { |
| 115 | if (osb == 0) break; |
| 116 | if (jbgp.os % osb == 0) { |
| 117 | os_block = osb; |
| 118 | break; |
| 119 | } |
| 120 | } |
| 121 | if (os_block == 1) os_block = nstl::min(jbgp.os, max_os_block); |
| 122 | |
| 123 | return os_block; |
| 124 | } |
| 125 | |
| 126 | std::vector<format_tag_t> get_desired_weights_tag( |
| 127 | const jit_brgemm_primitive_conf_t &jbgp) { |
| 128 | using namespace format_tag; |
| 129 | const int n_sp_dims = jbgp.ndims - 2; |
| 130 | const bool is_xf16 = utils::one_of(jbgp.wei_dt, bf16, f16); |
| 131 | const bool is_not_vnni_tag = jbgp.wei_dt == f32 |
| 132 | || (jbgp.wei_dt == f16 && jbgp.isa == avx512_core_fp16); |
| 133 | if (is_not_vnni_tag) { |
| 134 | return {pick(n_sp_dims, OI16i64o, OIw16i64o, OIhw16i64o, OIdhw16i64o), |
| 135 | pick(n_sp_dims, OI16i32o, OIw16i32o, OIhw16i32o, OIdhw16i32o), |
| 136 | pick(n_sp_dims, OI16i16o, OIw16i16o, OIhw16i16o, OIdhw16i16o)}; |
| 137 | } else if (is_xf16) { |
| 138 | if (is_superset(jbgp.isa, avx512_core_amx)) { |
| 139 | return {pick(n_sp_dims, OI16i64o2i, OIw16i64o2i, OIhw16i64o2i, |
| 140 | OIdhw16i64o2i), |
| 141 | pick(n_sp_dims, OI16i32o2i, OIw16i32o2i, OIhw16i32o2i, |
| 142 | OIdhw16i32o2i), |
| 143 | pick(n_sp_dims, OI16i16o2i, OIw16i16o2i, OIhw16i16o2i, |
| 144 | OIdhw16i16o2i)}; |
| 145 | } else { |
| 146 | return {pick(n_sp_dims, OI8i64o2i, OIw8i64o2i, OIhw8i64o2i, |
| 147 | OIdhw8i64o2i), |
| 148 | pick(n_sp_dims, OI8i32o2i, OIw8i32o2i, OIhw8i32o2i, |
| 149 | OIdhw8i32o2i), |
| 150 | pick(n_sp_dims, OI8i16o2i, OIw8i16o2i, OIhw8i16o2i, |
| 151 | OIdhw8i16o2i)}; |
| 152 | } |
| 153 | } else if (jbgp.wei_dt == data_type::s8) { |
| 154 | if (jbgp.isa == avx512_core_amx) { |
| 155 | return {pick(n_sp_dims, OI16i64o4i, OIw16i64o4i, OIhw16i64o4i, |
| 156 | OIdhw16i64o4i), |
| 157 | pick(n_sp_dims, OI16i32o4i, OIw16i32o4i, OIhw16i32o4i, |
| 158 | OIdhw16i32o4i), |
| 159 | pick(n_sp_dims, OI16i16o4i, OIw16i16o4i, OIhw16i16o4i, |
| 160 | OIdhw16i16o4i)}; |
| 161 | } else { |
| 162 | return {pick(n_sp_dims, OI4i64o4i, OIw4i64o4i, OIhw4i64o4i, |
| 163 | OIdhw4i64o4i), |
| 164 | pick(n_sp_dims, OI4i32o4i, OIw4i32o4i, OIhw4i32o4i, |
| 165 | OIdhw4i32o4i), |
| 166 | pick(n_sp_dims, OI4i16o4i, OIw4i16o4i, OIhw4i16o4i, |
| 167 | OIdhw4i16o4i)}; |
| 168 | } |
| 169 | } else { |
| 170 | return std::vector<format_tag_t> {format_tag::undef}; |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | int get_oc_block(const jit_brgemm_primitive_conf_t &jbgp, bool try_to_adjust) { |
| 175 | const bool amx_xf16_bwd_d_noadjust = !try_to_adjust |
| 176 | && jbgp.prop_kind == backward_data |
| 177 | && is_superset(jbgp.isa, avx512_core_amx) && !jbgp.is_bf32; |
| 178 | if (amx_xf16_bwd_d_noadjust) { |
| 179 | constexpr int amx_xf16_row = 64; |
| 180 | return amx_xf16_row; |
| 181 | } else if (!jbgp.is_wei_layout_any) { |
| 182 | std::vector<format_tag_t> weights_tag = get_desired_weights_tag(jbgp); |
| 183 | if (jbgp.wei_tag == weights_tag[0]) |
| 184 | return 64; |
| 185 | else if (jbgp.wei_tag == weights_tag[1]) |
| 186 | return 32; |
| 187 | else |
| 188 | return 16; |
| 189 | } else { |
| 190 | if (jbgp.oc >= 64) { |
| 191 | return 64; |
| 192 | } else if (jbgp.oc >= 32) { |
| 193 | return 32; |
| 194 | } else { |
| 195 | return 16; |
| 196 | } |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | int ip_fwd_get_nb_oc_blocking( |
| 201 | const jit_brgemm_primitive_conf_t &jbgp, bool is_adjustment) { |
| 202 | const int small_oc_threshold = 256; |
| 203 | const int small_os_threshold = 8; |
| 204 | if (jbgp.os <= small_os_threshold && jbgp.oc <= small_oc_threshold) { |
| 205 | // For small problems compute all oc blocks as a single chunk to avoid |
| 206 | // parallel section |
| 207 | return div_up(jbgp.oc, |
| 208 | (is_adjustment) ? ip_fwd_get_adjusted_oc_block(jbgp) |
| 209 | : get_oc_block(jbgp)); |
| 210 | } else |
| 211 | return 1; |
| 212 | } |
| 213 | |
| 214 | bool ip_fwd_adjust_thread_balance(const jit_brgemm_primitive_conf_t &jbgp) { |
| 215 | if (IMPLICATION(jbgp.is_wei_layout_any, |
| 216 | !is_superset(jbgp.isa, avx512_core_amx))) |
| 217 | return false; |
| 218 | |
| 219 | int os_chunks = div_up(jbgp.os, get_os_block(jbgp, true, false)); |
| 220 | |
| 221 | int nb_oc = div_up(jbgp.oc, get_oc_block(jbgp, true)); |
| 222 | int nb_oc_blocking = ip_fwd_get_nb_oc_blocking(jbgp); |
| 223 | int oc_chunks = div_up(nb_oc, nb_oc_blocking); |
| 224 | |
| 225 | int work_amount = oc_chunks * os_chunks; |
| 226 | const auto work_per_thread = work_amount / jbgp.nthr; |
| 227 | float wb_ratio = static_cast<float>(work_amount % jbgp.nthr) / jbgp.nthr; |
| 228 | |
| 229 | // return true if work distribution between threads has significant |
| 230 | // imbalance - amount of work per thread is small and the last iteration |
| 231 | // is able to load less than half of threads available for the current |
| 232 | // block sizes |
| 233 | return (work_per_thread < 3 && wb_ratio > 0.0f && wb_ratio < .5f); |
| 234 | } |
| 235 | |
| 236 | int ip_fwd_get_adjusted_oc_block(const jit_brgemm_primitive_conf_t &jbgp) { |
| 237 | const bool is_amx_xf16 |
| 238 | = is_superset(jbgp.isa, avx512_core_amx) && !jbgp.is_bf32; |
| 239 | |
| 240 | // we can't change block size on forward and weights update (external) |
| 241 | // if layout is set by user, for backward data it can be choosen different |
| 242 | // from external in this case because copy routine |
| 243 | const bool not_adjustable_oc_block_size |
| 244 | = !jbgp.is_wei_layout_any && jbgp.prop_kind != backward_data; |
| 245 | |
| 246 | if (IMPLICATION(is_amx_xf16 || jbgp.is_bf32, not_adjustable_oc_block_size)) |
| 247 | return get_oc_block(jbgp); |
| 248 | |
| 249 | int oc_block = get_oc_block(jbgp, true); |
| 250 | if (ip_fwd_adjust_thread_balance(jbgp)) { |
| 251 | oc_block = (oc_block > 16) ? oc_block / 2 : oc_block; |
| 252 | } |
| 253 | |
| 254 | constexpr int amx_bf16_half_row = 32; |
| 255 | // ensure that oc_tail <= amx_bf16_half_row (requirement for brgemm kernel) |
| 256 | while (jbgp.oc % oc_block > amx_bf16_half_row) |
| 257 | oc_block /= 2; |
| 258 | return oc_block; |
| 259 | } |
| 260 | |
| 261 | format_tag_t get_brgemm_ip_weights_tag(cpu_isa_t isa, |
| 262 | const jit_brgemm_primitive_conf_t &jbgp, |
| 263 | const memory_desc_t &weights_md) { |
| 264 | std::vector<format_tag_t> weights_tag = get_desired_weights_tag(jbgp); |
| 265 | if (!jbgp.is_wei_layout_any) { |
| 266 | return memory_desc_matches_one_of_tag( |
| 267 | weights_md, weights_tag[0], weights_tag[1], weights_tag[2]); |
| 268 | } else { |
| 269 | const int oc_block = ip_fwd_get_adjusted_oc_block(jbgp); |
| 270 | const int idx = (oc_block == 64 ? 0 : (oc_block == 32 ? 1 : 2)); |
| 271 | return weights_tag[idx]; |
| 272 | } |
| 273 | } |
| 274 | |
| 275 | bool post_ops_ok(jit_brgemm_primitive_conf_t &jbgp, |
| 276 | const primitive_attr_t &attr, const memory_desc_wrapper &dst_d) { |
| 277 | using namespace injector; |
| 278 | |
| 279 | const auto &post_ops = attr.post_ops_; |
| 280 | |
| 281 | return injector::post_ops_ok(post_ops_ok_args_t(get_max_cpu_isa(), |
| 282 | {sum, eltwise, binary}, post_ops, &dst_d, |
| 283 | false /*sum_at_pos_0_only*/, false /*sum_requires_scale_one*/, |
| 284 | true /*sum_requires_zp_zero*/, |
| 285 | {broadcasting_strategy_t::per_oc, broadcasting_strategy_t::scalar, |
| 286 | broadcasting_strategy_t::no_broadcast})); |
| 287 | } |
| 288 | |
| 289 | status_t init_ip_conf_fwd(jit_brgemm_primitive_conf_t &jbgp, |
| 290 | const primitive_attr_t &attr, const memory_desc_wrapper &dst_d) { |
| 291 | const bool is_amx_int8 |
| 292 | = jbgp.isa == avx512_core_amx && one_of(jbgp.wei_dt, s8, u8); |
| 293 | const bool is_amx_xf16 = is_superset(jbgp.isa, avx512_core_amx) |
| 294 | && one_of(jbgp.wei_dt, bf16, f16) && !jbgp.is_bf32; |
| 295 | const bool is_int8 = one_of(jbgp.src_dt, u8, s8) && jbgp.wei_dt == s8; |
| 296 | const bool is_f32 = everyone_is(f32, jbgp.src_dt, jbgp.wei_dt, jbgp.dst_dt); |
| 297 | jbgp.is_amx = is_superset(jbgp.isa, avx512_core_amx); |
| 298 | const auto &p = attr.post_ops_; |
| 299 | jbgp.with_sum = p.find(primitive_kind::sum) != -1; |
| 300 | const int eltwise_ind = p.find(primitive_kind::eltwise); |
| 301 | jbgp.with_eltwise = eltwise_ind != -1; |
| 302 | const int binary_ind = p.find(primitive_kind::binary); |
| 303 | jbgp.with_binary = binary_ind != -1; |
| 304 | if (!post_ops_ok(jbgp, attr, dst_d)) return status::unimplemented; |
| 305 | if (jbgp.with_scales) { |
| 306 | const auto &wei_scales = attr.scales_.get(DNNL_ARG_WEIGHTS); |
| 307 | jbgp.is_oc_scale = wei_scales.mask_ == 1 << 0; |
| 308 | |
| 309 | // only common and per-oc-channel scales are supported |
| 310 | const bool wei_scales_ok = one_of(wei_scales.mask_, 0, 1 << 0); |
| 311 | if (!wei_scales_ok) return status::unimplemented; |
| 312 | } |
| 313 | const int min_ic_divisor = is_amx_int8 ? 4 : is_amx_xf16 ? 2 : 1; |
| 314 | |
| 315 | jbgp.use_buffer_a = jbgp.ic % min_ic_divisor != 0; |
| 316 | |
| 317 | constexpr int amx_int8_row = 64; |
| 318 | constexpr int amx_xf16_row = 32; |
| 319 | jbgp.ic_block = (is_amx_int8) ? amx_int8_row |
| 320 | : (is_amx_xf16) ? amx_xf16_row : jbgp.simd_w; |
| 321 | jbgp.nb_ic = div_up(jbgp.ic, jbgp.ic_block); |
| 322 | |
| 323 | // gemm-based inner product performs better when oc = 1 |
| 324 | if (is_f32 && !jbgp.is_bf32 && jbgp.oc == 1) return status::unimplemented; |
| 325 | |
| 326 | jbgp.oc_block = ip_fwd_get_adjusted_oc_block(jbgp); |
| 327 | jbgp.nb_oc = div_up(jbgp.oc, jbgp.oc_block); |
| 328 | jbgp.nb_oc_blocking = ip_fwd_get_nb_oc_blocking(jbgp); |
| 329 | |
| 330 | jbgp.os_block = get_os_block(jbgp, false, false); |
| 331 | jbgp.nb_os = div_up(jbgp.os, jbgp.os_block); |
| 332 | |
| 333 | jbgp.nb_os_blocking = 1; |
| 334 | // Work done by each thread is given by: |
| 335 | // (nb_oc / nb_oc_blocking) * (nb_os / nb_os_blocking) |
| 336 | // For f32 data type we want to increase the nb_os_blocking such that |
| 337 | // * 1 <= nb_os_blocking <= 8 AND nb_os_blocking <= nb_os |
| 338 | // * Work amount per thread ~ 2 |
| 339 | // * NOTE: here nb_oc_blocking = 1 as os is large |
| 340 | if (jbgp.os > 256 && is_f32 && !jbgp.is_bf32) { |
| 341 | jbgp.nb_os_blocking = saturate(1, nstl::min(8, jbgp.nb_os), |
| 342 | nstl::min(nstl::max(jbgp.oc / jbgp.os / 2, 1), |
| 343 | div_up(jbgp.nb_os * jbgp.nb_oc, 2 * jbgp.nthr))); |
| 344 | } |
| 345 | |
| 346 | // NOTE: comment about is_gigantic_shape is in get_os_block() |
| 347 | const bool is_gigantic_shape = jbgp.oc >= 4096 && jbgp.os >= 512; |
| 348 | const int num_work_to_parallel = div_up(jbgp.nb_oc, jbgp.nb_oc_blocking) |
| 349 | * div_up(jbgp.nb_os, jbgp.nb_os_blocking); |
| 350 | |
| 351 | // TODO: although the below heuristic produces good performance for fp32, |
| 352 | // num_work_to_parallel needs to compared with nthr (instead of nb_ic) |
| 353 | // and os_block needs some further tuning. |
| 354 | |
| 355 | // Use parallel IC reduction for f32 if we have: |
| 356 | // * very large input channels |
| 357 | // * work amount in mb and oc dimensions is small compared to nb_ic |
| 358 | // * number of threads > 1 |
| 359 | // * not a "gigantic shape" since it already has a lot of parallelism |
| 360 | // in mb and oc dimensions w/o enabling IC parallelism |
| 361 | const bool use_parallel_ic_reduction = is_f32 && !jbgp.is_bf32 |
| 362 | && jbgp.ic > 1024 && num_work_to_parallel < jbgp.nb_ic |
| 363 | && jbgp.nthr > 1 && !is_gigantic_shape; |
| 364 | |
| 365 | // For os > 256, compute all os blocks as a single chunk when performing |
| 366 | // IC reduction. Note that this condition is empirical |
| 367 | if (use_parallel_ic_reduction && jbgp.os > 256 && jbgp.nb_os_blocking > 1) |
| 368 | jbgp.nb_os_blocking = jbgp.nb_os; |
| 369 | |
| 370 | jbgp.nb_ic_blocking = 1; |
| 371 | jbgp.nthr_ic_b = 1; |
| 372 | const int k_blk = jbgp.is_bf32 ? amx_xf16_row : jbgp.ic_block; |
| 373 | const int max_nb_ic_blocking = nstl::min(64, jbgp.nb_ic); |
| 374 | if (IMPLICATION(!is_int8, jbgp.ic <= max_nb_ic_blocking * jbgp.ic_block) |
| 375 | && everyone_is(1, jbgp.kw, jbgp.kh, jbgp.kd) |
| 376 | && !jbgp.use_buffer_a) { |
| 377 | // Optimization: data & weights layouts allow to generate |
| 378 | // brgemm kernel with K = ic & batch = 1 |
| 379 | // (K = rnd_dn(ic, ic_block), K_tail = ic % ic_block & batch = 1) |
| 380 | // instead of K = ic_block & batch = nb_ic_blocking |
| 381 | jbgp.K = jbgp.ic <= jbgp.ic_block ? jbgp.ic : rnd_dn(jbgp.ic, k_blk); |
| 382 | jbgp.nb_ic_blocking = jbgp.nb_ic; |
| 383 | jbgp.gemm_batch_size = 1; |
| 384 | } else if (!jbgp.use_buffer_a && use_parallel_ic_reduction) { |
| 385 | const int min_chunk_sz = 16; |
| 386 | const int num_min_chunk_sz = div_up(jbgp.nb_ic, min_chunk_sz); |
| 387 | float reduce_work = 0.5f * num_min_chunk_sz * jbgp.nb_os |
| 388 | + (float)num_min_chunk_sz / jbgp.nb_oc + 0.5f; |
| 389 | const int reduce_thr_groups = jbgp.nb_ic >= 1024 ? 8 : 4; |
| 390 | jbgp.nthr_ic_b |
| 391 | = saturate(1, nstl::min(reduce_thr_groups, num_min_chunk_sz), |
| 392 | int(reduce_work)); |
| 393 | jbgp.nthr_ic_b = nstl::min(jbgp.nthr_ic_b, jbgp.nthr); |
| 394 | if (jbgp.nthr_ic_b > 1) { |
| 395 | jbgp.nb_ic_blocking = div_up(jbgp.nb_ic, jbgp.nthr_ic_b); |
| 396 | jbgp.nb_ic_blocking /= div_up(jbgp.nb_ic_blocking, 64); |
| 397 | } |
| 398 | jbgp.gemm_batch_size = jbgp.nb_ic_blocking; |
| 399 | jbgp.K = jbgp.ic_block; |
| 400 | } else { |
| 401 | // Note: Here, ic divided into K_blocks of gemm_batch |
| 402 | const int ic_blks_per_k = div_up(k_blk, jbgp.ic_block); |
| 403 | const int nb_k_blk = div_up(jbgp.ic, k_blk); |
| 404 | const int max_nb_k_blocking = div_up(max_nb_ic_blocking, ic_blks_per_k); |
| 405 | int nb_k_blocking = max_div(nb_k_blk, max_nb_k_blocking); |
| 406 | const bool small_nb_k_blk = nb_k_blk <= max_nb_k_blocking; |
| 407 | if (small_nb_k_blk && nb_k_blocking == 1) |
| 408 | nb_k_blocking = max_nb_k_blocking; |
| 409 | |
| 410 | // For non small_nb_ic [i.e. that has nb_ic > 64] shape that has |
| 411 | // gcd(nb_ic, 64) < 16, we manually set nb_ic_blocking = 64 |
| 412 | // the coefficients 64 [used in max_nb_ic_blocking] and 16 are empirical |
| 413 | const int min_nb_k_blocking = small_nb_k_blk ? 1 : 16; |
| 414 | if (nb_k_blocking < min_nb_k_blocking) |
| 415 | nb_k_blocking = max_nb_k_blocking; |
| 416 | |
| 417 | jbgp.nb_ic_blocking = nb_k_blocking * ic_blks_per_k; |
| 418 | jbgp.K = k_blk; |
| 419 | jbgp.gemm_batch_size = nb_k_blocking; |
| 420 | } |
| 421 | |
| 422 | // to avoid cache concurrent write access from different threads |
| 423 | size_t sc_size = sizeof(brgemm_batch_element_t); |
| 424 | jbgp.adjusted_batch_size |
| 425 | = div_up(rnd_up(jbgp.gemm_batch_size * sc_size, 4096), sc_size); |
| 426 | |
| 427 | if (is_amx_xf16 || jbgp.is_bf32) { |
| 428 | if (ip_fwd_adjust_thread_balance(jbgp)) { |
| 429 | // Adjust oc_block to improve thread balancing |
| 430 | jbgp.oc_block = ip_fwd_get_adjusted_oc_block(jbgp); |
| 431 | jbgp.nb_oc = div_up(jbgp.oc, jbgp.oc_block); |
| 432 | jbgp.nb_oc_blocking = ip_fwd_get_nb_oc_blocking(jbgp, true); |
| 433 | |
| 434 | // Adjust os_block to improve thread balancing |
| 435 | if (jbgp.oc <= 16 |
| 436 | || types::data_type_size(jbgp.src_dt) * jbgp.mb * jbgp.ic |
| 437 | <= (size_t)platform::get_per_core_cache_size(2)) { |
| 438 | jbgp.os_block = get_os_block(jbgp, false, true); |
| 439 | jbgp.nb_os = div_up(jbgp.os, jbgp.os_block); |
| 440 | } |
| 441 | } |
| 442 | } |
| 443 | jbgp.use_buffer = (IMPLICATION(jbgp.dst_dt == jbgp.acc_dt, jbgp.with_sum)) |
| 444 | || (jbgp.nthr_ic_b > 1); |
| 445 | |
| 446 | // Configure matrix sizes |
| 447 | jbgp.M = jbgp.os_block; |
| 448 | jbgp.M_tail = jbgp.os % jbgp.os_block; |
| 449 | |
| 450 | jbgp.N = jbgp.oc_block; |
| 451 | jbgp.N_tail = jbgp.oc % jbgp.oc_block; |
| 452 | jbgp.K_tail = jbgp.use_buffer_a ? 0 : jbgp.ic % jbgp.K; |
| 453 | |
| 454 | jbgp.LDA = jbgp.use_buffer_a ? jbgp.K * jbgp.gemm_batch_size |
| 455 | : jbgp.ic_without_padding; |
| 456 | jbgp.LDB = jbgp.N; |
| 457 | jbgp.LDD = jbgp.oc_without_padding; |
| 458 | jbgp.LDC = (jbgp.use_buffer && jbgp.nthr_ic_b == 1) ? jbgp.N : jbgp.LDD; |
| 459 | |
| 460 | if (jbgp.is_bf32) { |
| 461 | const float M = static_cast<float>(jbgp.M); |
| 462 | const float N = nstl::min<float>(jbgp.N, jbgp.oc); |
| 463 | const float K |
| 464 | = nstl::min<float>(jbgp.K * jbgp.gemm_batch_size, jbgp.ic); |
| 465 | const float tmul_efficiency = (M / 16) * (N / 16) * (K / 32); |
| 466 | // TODO: Adjust blocking such that bigger M, N, K are generated. |
| 467 | if (one_of(true, M <= 8, K <= 8, N < 16, tmul_efficiency <= 2.25)) |
| 468 | return status::unimplemented; |
| 469 | } |
| 470 | |
| 471 | return status::success; |
| 472 | } |
| 473 | |
| 474 | status_t init_ip_conf_bwd_d(jit_brgemm_primitive_conf_t &jbgp) { |
| 475 | const bool is_amx_xf16 |
| 476 | = is_superset(jbgp.isa, avx512_core_amx) && !jbgp.is_bf32; |
| 477 | const bool is_avx512_bf16 = jbgp.isa == avx512_core_bf16; |
| 478 | const bool is_f32 = everyone_is(f32, jbgp.src_dt, jbgp.wei_dt, jbgp.dst_dt); |
| 479 | const bool is_bf16 = everyone_is(bf16, jbgp.wei_dt, jbgp.dst_dt); |
| 480 | jbgp.is_amx = is_amx_xf16; |
| 481 | |
| 482 | constexpr int amx_xf16_granularity = 2; |
| 483 | jbgp.use_buffer_a = is_amx_xf16 && jbgp.oc % amx_xf16_granularity != 0; |
| 484 | jbgp.use_buffer_b = true; |
| 485 | jbgp.ip_bwd_d_global_b_transpose = false; |
| 486 | |
| 487 | jbgp.oc_block = ip_fwd_get_adjusted_oc_block(jbgp); |
| 488 | |
| 489 | // Optimization: for small shape we avoid large ic_block |
| 490 | // Thinking of os, ic, and oc as three dimensions, the boundary for small |
| 491 | // shapes is heuristically chosen via the following constraints: |
| 492 | // os <= 128 && max(ic, oc) <= 2048 && min(ic, oc) <= 1000 |
| 493 | // |
| 494 | // TODO: Will the optimization be useful for bf16 data type |
| 495 | const bool avoid_max_ic_block = is_f32 && !jbgp.is_bf32 && jbgp.os <= 128 |
| 496 | && nstl::max(jbgp.ic, jbgp.oc) <= 2048 |
| 497 | && nstl::min(jbgp.ic, jbgp.oc) <= 1000; |
| 498 | jbgp.ic_block = !avoid_max_ic_block |
| 499 | && jbgp.ic >= (is_f32 && !jbgp.is_bf32 ? 512 : 64) |
| 500 | ? 64 |
| 501 | : jbgp.ic >= 32 ? 32 : 16; |
| 502 | |
| 503 | jbgp.nb_ic = div_up(jbgp.ic, jbgp.ic_block); |
| 504 | jbgp.nb_ic_blocking = 1; |
| 505 | jbgp.nb_oc = div_up(jbgp.oc, jbgp.oc_block); |
| 506 | |
| 507 | jbgp.os_block = get_os_block(jbgp, false, false); |
| 508 | |
| 509 | jbgp.nb_os = div_up(jbgp.os, jbgp.os_block); |
| 510 | jbgp.nb_os_blocking = 1; |
| 511 | int os_blocking_max = 2; |
| 512 | for (int bl = os_blocking_max; bl >= 1; bl--) |
| 513 | if (jbgp.nb_os % bl == 0) { |
| 514 | jbgp.nb_os_blocking = bl; |
| 515 | break; |
| 516 | } |
| 517 | |
| 518 | if (is_amx_xf16 || jbgp.is_bf32) { |
| 519 | const int os_chunks = div_up(jbgp.nb_os, jbgp.nb_os_blocking); |
| 520 | const int work_amount = jbgp.nb_ic * os_chunks; |
| 521 | float wb_ratio = (float)work_amount / (float)jbgp.nthr; |
| 522 | if (wb_ratio != 1.f && wb_ratio < 2.f) { |
| 523 | jbgp.ic_block |
| 524 | = (jbgp.ic_block > 16) ? jbgp.ic_block / 2 : jbgp.ic_block; |
| 525 | jbgp.nb_ic = div_up(jbgp.ic, jbgp.ic_block); |
| 526 | } |
| 527 | } |
| 528 | |
| 529 | jbgp.nb_oc_blocking = 1; |
| 530 | const int oc_chunk_max_size = 64; |
| 531 | for (int bl = oc_chunk_max_size; bl >= 1; bl--) |
| 532 | if (jbgp.nb_oc % bl == 0) { |
| 533 | jbgp.nb_oc_blocking = bl; |
| 534 | break; |
| 535 | } |
| 536 | |
| 537 | jbgp.nthr_oc_b = 1; |
| 538 | const int num_work_to_parallel = div_up(jbgp.nb_ic, jbgp.nb_ic_blocking) |
| 539 | * div_up(jbgp.nb_os, jbgp.nb_os_blocking); |
| 540 | // Use oc reduction if we have |
| 541 | // * very large output channels |
| 542 | // * small work amount available to each thread |
| 543 | if ((num_work_to_parallel < 2 * jbgp.nthr |
| 544 | || jbgp.oc > (is_bf16 || jbgp.is_bf32 ? 4096 : 1024))) { |
| 545 | const int min_chunck_sz = (is_avx512_bf16) ? 32 : 16; |
| 546 | const int num_min_chunk_sz = div_up(jbgp.nb_oc, min_chunck_sz); |
| 547 | float reduce_work = 0.5f * num_min_chunk_sz * jbgp.nb_os |
| 548 | + (float)num_min_chunk_sz / jbgp.nb_ic + 0.5f; |
| 549 | |
| 550 | // optimization for transformer_lt on CPX/SKX |
| 551 | const int max_nthr_oc_b |
| 552 | = (!is_amx_xf16 && !jbgp.is_bf32 && jbgp.oc > 32000) |
| 553 | ? jbgp.nthr / 2 |
| 554 | : 4; |
| 555 | jbgp.nthr_oc_b = saturate(1, nstl::min(max_nthr_oc_b, num_min_chunk_sz), |
| 556 | int(reduce_work)); |
| 557 | jbgp.nthr_oc_b = nstl::min(jbgp.nthr_oc_b, jbgp.nthr); |
| 558 | if (jbgp.nthr_oc_b > 1) { |
| 559 | jbgp.nb_oc_blocking = div_up(jbgp.nb_oc, jbgp.nthr_oc_b); |
| 560 | jbgp.nb_oc_blocking |
| 561 | /= div_up(jbgp.nb_oc_blocking, oc_chunk_max_size); |
| 562 | } |
| 563 | } |
| 564 | jbgp.gemm_batch_size = jbgp.nb_oc_blocking; |
| 565 | // to avoid cache concurrent write access from different threads |
| 566 | size_t sc_size = sizeof(brgemm_batch_element_t); |
| 567 | jbgp.adjusted_batch_size |
| 568 | = div_up(rnd_up(jbgp.gemm_batch_size * sc_size, 4096), sc_size); |
| 569 | |
| 570 | jbgp.use_buffer = jbgp.src_dt != jbgp.acc_dt || jbgp.nthr_oc_b > 1; |
| 571 | |
| 572 | jbgp.M = jbgp.os_block; |
| 573 | jbgp.M_tail = jbgp.os % jbgp.os_block; |
| 574 | |
| 575 | jbgp.K = jbgp.oc_block; |
| 576 | jbgp.N = jbgp.ic_block; |
| 577 | jbgp.N_tail = jbgp.ic % jbgp.ic_block; |
| 578 | jbgp.K_tail = jbgp.use_buffer_a ? 0 : jbgp.oc % jbgp.oc_block; |
| 579 | |
| 580 | jbgp.LDA = jbgp.use_buffer_a ? jbgp.K * jbgp.nb_oc_blocking |
| 581 | : jbgp.oc_without_padding; |
| 582 | jbgp.LDB = jbgp.N; |
| 583 | jbgp.LDD = jbgp.ic_without_padding; |
| 584 | jbgp.LDC = jbgp.use_buffer && jbgp.nthr_oc_b == 1 ? jbgp.N : jbgp.LDD; |
| 585 | |
| 586 | if (jbgp.is_bf32) { |
| 587 | const float M = static_cast<float>(jbgp.M); |
| 588 | const float N = nstl::min<float>(jbgp.N, jbgp.ic); |
| 589 | const float K |
| 590 | = nstl::min<float>(jbgp.K * jbgp.gemm_batch_size, jbgp.oc); |
| 591 | const float tmul_efficiency = (M / 16) * (N / 16) * (K / 32); |
| 592 | // TODO: Adjust blocking such that bigger M, N, K are generated. |
| 593 | if (one_of(true, M <= 8, K <= 8, N < 16, tmul_efficiency <= 2.25)) |
| 594 | return status::unimplemented; |
| 595 | } |
| 596 | |
| 597 | return status::success; |
| 598 | } |
| 599 | |
| 600 | void thread_balance(const jit_brgemm_primitive_conf_t &j, int &nb_os_blocking_, |
| 601 | int &nb_oc_blocking_, int &nb_ic_blocking_, int &nthr_, int &nthr_mb_, |
| 602 | int &nthr_oc_b_, int &nthr_ic_b_) { |
| 603 | nthr_ = nthr_mb_ = nthr_oc_b_ = nthr_ic_b_ = 1; |
| 604 | nb_os_blocking_ = j.nb_os_blocking; |
| 605 | nb_oc_blocking_ = j.nb_oc_blocking; |
| 606 | nb_ic_blocking_ = j.nb_ic_blocking; |
| 607 | |
| 608 | const bool is_f32 = everyone_is(f32, j.src_dt, j.wei_dt, j.dst_dt); |
| 609 | const bool is_xf16 = one_of(j.src_dt, bf16, f16) && (j.src_dt == j.dst_dt); |
| 610 | |
| 611 | const int max_threads = j.nthr; |
| 612 | const int nthr = max_threads; |
| 613 | auto calc_mem_cost = [=](int nb_os_blocking, int nb_oc_blocking, |
| 614 | int nb_ic_blocking, int nthr_mb, int nthr_oc, |
| 615 | int nthr_ic) { |
| 616 | float src_size = static_cast<float>(j.ic) * j.mb; |
| 617 | float dst_size = static_cast<float>(j.oc) * j.mb; |
| 618 | float wei_size = static_cast<float>(j.ic) * j.oc; |
| 619 | int os_chunks = div_up(j.nb_os, nb_os_blocking); |
| 620 | int oc_chunks = div_up(j.nb_oc, nb_oc_blocking); |
| 621 | int ic_chunks = div_up(j.nb_ic, nb_ic_blocking); |
| 622 | |
| 623 | float wei_compensation_scale = 0.5f * (dst_size + src_size) / wei_size; |
| 624 | |
| 625 | float oi_channels_ratio = 0; |
| 626 | if (is_xf16) { |
| 627 | oi_channels_ratio = ((j.oc > 3 * j.ic && os_chunks > 1) |
| 628 | || (os_chunks == 1 && j.ic > j.oc)) |
| 629 | ? src_size / dst_size |
| 630 | : dst_size / src_size; |
| 631 | } else { |
| 632 | oi_channels_ratio = src_size / dst_size; |
| 633 | } |
| 634 | |
| 635 | auto get_src_coef = [=]() { |
| 636 | if (is_f32) { |
| 637 | float src_coef = nstl::max(1.0f / oi_channels_ratio, 1.0f); |
| 638 | src_coef *= types::data_type_size(j.src_dt); |
| 639 | src_coef *= 4 * saturate(1, 4, div_up(j.ic, 1024)); |
| 640 | if (wei_compensation_scale < 2.0f) |
| 641 | src_coef += sqrtf(2.0f / wei_compensation_scale); |
| 642 | return src_coef; |
| 643 | } |
| 644 | float src_coef = nstl::max(1.0f / oi_channels_ratio, 1.0f); |
| 645 | src_coef *= 4 * types::data_type_size(j.src_dt); |
| 646 | if (wei_compensation_scale < 1.0f) src_coef *= 4.0f; |
| 647 | |
| 648 | return src_coef; |
| 649 | }; |
| 650 | |
| 651 | auto get_dst_coef = [=]() { |
| 652 | if (is_f32) { |
| 653 | float dst_coef = types::data_type_size(j.dst_dt) |
| 654 | * nstl::max(oi_channels_ratio, 1.0f); |
| 655 | return dst_coef; |
| 656 | } |
| 657 | |
| 658 | return 2 * types::data_type_size(j.dst_dt) |
| 659 | * nstl::max(oi_channels_ratio, 1.0f); |
| 660 | }; |
| 661 | |
| 662 | auto get_wei_coef = [=]() { |
| 663 | if (is_f32) { |
| 664 | return nstl::max( |
| 665 | 4.0f - j.mb / 2048 * wei_compensation_scale, 1.0f); |
| 666 | } |
| 667 | |
| 668 | // limit the range of coefficient values to have more stable behavior |
| 669 | // for extreme cases |
| 670 | const float low_limit = 1.0f; |
| 671 | const float upper_limit = 1024.0f; |
| 672 | return utils::saturate( |
| 673 | low_limit, upper_limit, wei_compensation_scale); |
| 674 | }; |
| 675 | |
| 676 | float src_tr = 0.0f; |
| 677 | if (j.use_buffer_a && !is_f32) { |
| 678 | int src_tr_oc_par_work = div_up(os_chunks, nthr_mb) |
| 679 | * div_up(ic_chunks, nthr_ic) * nb_ic_blocking; |
| 680 | src_tr = get_src_coef() * div_up(src_tr_oc_par_work, nthr_oc) |
| 681 | * nb_os_blocking * j.os_block * j.ic_block; |
| 682 | } |
| 683 | |
| 684 | float dst_tr = 0.0f; |
| 685 | if (j.use_buffer_b && !is_f32) { |
| 686 | int dst_tr_ic_par_work = div_up(os_chunks, nthr_mb) |
| 687 | * div_up(oc_chunks, nthr_oc) * nb_oc_blocking; |
| 688 | dst_tr = get_dst_coef() * div_up(dst_tr_ic_par_work, nthr_ic) |
| 689 | * nb_os_blocking * j.os_block * j.oc_block; |
| 690 | } |
| 691 | |
| 692 | float src_v = get_src_coef() * div_up(os_chunks, nthr_mb) |
| 693 | * div_up(ic_chunks, nthr_ic) * nb_os_blocking * j.os_block |
| 694 | * nb_ic_blocking * j.ic_block; |
| 695 | float dst_v = get_dst_coef() * div_up(os_chunks, nthr_mb) |
| 696 | * div_up(oc_chunks, nthr_oc) * nb_os_blocking * j.os_block |
| 697 | * nb_oc_blocking * j.oc_block; |
| 698 | |
| 699 | auto acc_dt_sz = types::data_type_size(j.acc_dt); |
| 700 | float wei_v = get_wei_coef() * acc_dt_sz * div_up(oc_chunks, nthr_oc) |
| 701 | * div_up(ic_chunks, nthr_ic) * nb_oc_blocking * j.oc_block |
| 702 | * nb_ic_blocking * j.ic_block; |
| 703 | |
| 704 | float wei_r = 0; |
| 705 | if (nthr_mb > 1) { |
| 706 | auto wei_dt_sz = types::data_type_size(j.wei_dt); |
| 707 | int wei_r_mb_par_work = div_up(oc_chunks, nthr_oc) |
| 708 | * div_up(ic_chunks, nthr_ic) * nb_oc_blocking |
| 709 | * nb_ic_blocking; |
| 710 | wei_r = get_wei_coef() * div_up(wei_r_mb_par_work, nthr_mb) |
| 711 | * j.oc_block * j.ic_block |
| 712 | * (wei_dt_sz |
| 713 | + (is_f32 ? div_up(j.os, 1024) : 1) * nthr_mb |
| 714 | * acc_dt_sz); |
| 715 | } |
| 716 | |
| 717 | return src_tr + dst_tr + src_v + dst_v + wei_v + wei_r; |
| 718 | }; |
| 719 | |
| 720 | float best_mem_cost = calc_mem_cost(nb_os_blocking_, nb_oc_blocking_, |
| 721 | nb_ic_blocking_, nthr_mb_, nthr_oc_b_, nthr_ic_b_); |
| 722 | |
| 723 | /* Set range of values for nb_oc_blocking/nb_ic_blocking parameters to try. |
| 724 | Use powers-of-2 values to avoid potential issues on converting to |
| 725 | blocked weights layout stage |
| 726 | */ |
| 727 | auto get_blk_values |
| 728 | = [](int max_blk_value, int init_blk, int dim_blk_limit) { |
| 729 | int val_1st = rnd_up_pow2(init_blk); |
| 730 | int val_end = nstl::min(max_blk_value, dim_blk_limit); |
| 731 | std::vector<int> values; |
| 732 | for (int val = val_1st; val <= val_end; val <<= 1) |
| 733 | values.push_back(val); |
| 734 | return values; |
| 735 | }; |
| 736 | |
| 737 | const int max_nb_oc_blocking_pow |
| 738 | = j.ip_bwd_w_local_buffers_for_input_tensors ? 4 : j.nb_oc_blocking; |
| 739 | auto nb_oc_blocking_values |
| 740 | = get_blk_values(max_nb_oc_blocking_pow, j.nb_oc_blocking, j.nb_oc); |
| 741 | const int max_nb_ic_blocking_pow |
| 742 | = j.ip_bwd_w_local_buffers_for_input_tensors ? 4 : j.nb_ic_blocking; |
| 743 | auto nb_ic_blocking_values |
| 744 | = get_blk_values(max_nb_ic_blocking_pow, j.nb_ic_blocking, j.nb_ic); |
| 745 | |
| 746 | /* find the best thread distribution with lowest memory cost */ |
| 747 | const int min_osb_chunk = is_f32 ? 32 : is_xf16 ? 8 : 1; |
| 748 | const int nthr_mb_max = nstl::min(nthr, div_up(j.nb_os, min_osb_chunk)); |
| 749 | for (int nthr_mb = 1; nthr_mb <= nthr_mb_max; ++nthr_mb) { |
| 750 | int nb_os_blocking = j.nb_os_blocking; |
| 751 | int os_chunks = div_up(j.nb_os, nb_os_blocking); |
| 752 | if (os_chunks < nthr_mb) { |
| 753 | int coef = saturate(1, 4, 2 * j.mb / (j.oc + j.ic)); |
| 754 | int os_blocking_max = div_up(div_up(j.nb_os, coef), nthr_mb); |
| 755 | for (int bl = os_blocking_max; bl >= 1; bl--) |
| 756 | if (j.nb_os % bl == 0) { |
| 757 | nb_os_blocking = bl; |
| 758 | break; |
| 759 | } |
| 760 | } |
| 761 | |
| 762 | const int nthr_par = nthr / nthr_mb; |
| 763 | |
| 764 | for (auto nb_oc_blocking : nb_oc_blocking_values) { |
| 765 | int num_oc_chunks = div_up(j.nb_oc, nb_oc_blocking); |
| 766 | const int nthr_oc_b_max = nstl::min(nthr_par, num_oc_chunks); |
| 767 | for_(int nthr_oc_b = 1; nthr_oc_b <= nthr_oc_b_max; ++nthr_oc_b) |
| 768 | for (auto nb_ic_blocking : nb_ic_blocking_values) { |
| 769 | int num_ic_chunks = div_up(j.nb_ic, nb_ic_blocking); |
| 770 | |
| 771 | int nthr_ic_b = nstl::min(nthr_par / nthr_oc_b, num_ic_chunks); |
| 772 | float mem_cost = calc_mem_cost(nb_os_blocking, nb_oc_blocking, |
| 773 | nb_ic_blocking, nthr_mb, nthr_oc_b, nthr_ic_b); |
| 774 | if (mem_cost <= best_mem_cost) { |
| 775 | best_mem_cost = mem_cost; |
| 776 | nb_os_blocking_ = nb_os_blocking; |
| 777 | nb_oc_blocking_ = nb_oc_blocking; |
| 778 | nb_ic_blocking_ = nb_ic_blocking; |
| 779 | nthr_mb_ = nthr_mb; |
| 780 | nthr_oc_b_ = nthr_oc_b; |
| 781 | nthr_ic_b_ = nthr_ic_b; |
| 782 | } |
| 783 | } |
| 784 | } |
| 785 | } |
| 786 | |
| 787 | nthr_ = nthr_mb_ * nthr_oc_b_ * nthr_ic_b_; |
| 788 | } |
| 789 | |
| 790 | status_t init_ip_conf_bwd_w(jit_brgemm_primitive_conf_t &jbgp) { |
| 791 | const bool is_amx_xf16 |
| 792 | = is_superset(jbgp.isa, avx512_core_amx) && !jbgp.is_bf32; |
| 793 | const bool is_f32 = everyone_is(f32, jbgp.src_dt, jbgp.wei_dt, jbgp.dst_dt); |
| 794 | const bool has_weights_buffer = jbgp.wei_dt != jbgp.acc_dt; |
| 795 | jbgp.is_amx = is_amx_xf16; |
| 796 | |
| 797 | const int amx_xf16_row = 64; |
| 798 | const bool big_ic_blk_ok |
| 799 | = is_f32 && jbgp.ic % (4 * jbgp.simd_w) == 0 && jbgp.mb <= 128; |
| 800 | jbgp.ic_block = big_ic_blk_ok && !is_amx_xf16 |
| 801 | ? 4 * jbgp.simd_w |
| 802 | : (is_amx_xf16 && has_weights_buffer) ? amx_xf16_row : jbgp.simd_w; |
| 803 | jbgp.ic_block_ext |
| 804 | = is_amx_xf16 || (jbgp.wei_dt == dnnl::impl::data_type::bf16) ? 32 |
| 805 | : 16; |
| 806 | |
| 807 | jbgp.oc_block = has_weights_buffer ? get_oc_block(jbgp) |
| 808 | : ip_fwd_get_adjusted_oc_block(jbgp); |
| 809 | jbgp.oc_block_ext = ip_fwd_get_adjusted_oc_block(jbgp); |
| 810 | |
| 811 | jbgp.os_block = get_os_block(jbgp, false, false); |
| 812 | jbgp.nb_os = div_up(jbgp.os, jbgp.os_block); |
| 813 | |
| 814 | jbgp.nb_ic = div_up(jbgp.ic, jbgp.ic_block); |
| 815 | jbgp.nb_oc = div_up(jbgp.oc, jbgp.oc_block); |
| 816 | jbgp.nb_oc_blocking = 1; |
| 817 | jbgp.nb_ic_blocking = jbgp.nb_ic % 2 ? 1 : 2; |
| 818 | |
| 819 | // Configure matrix sizes |
| 820 | jbgp.M = jbgp.ic_block; |
| 821 | jbgp.M_tail = jbgp.ic % jbgp.ic_block; |
| 822 | |
| 823 | jbgp.N = jbgp.oc_block; |
| 824 | jbgp.N_tail = jbgp.oc % jbgp.oc_block; |
| 825 | |
| 826 | constexpr int amx_xf16_granularity = 2; |
| 827 | // sanity check, must hold for transpose routines to work fine |
| 828 | assert(IMPLICATION(is_amx_xf16, jbgp.os_block % amx_xf16_granularity == 0)); |
| 829 | const bool do_rnd_os = is_amx_xf16 && jbgp.os % amx_xf16_granularity != 0; |
| 830 | |
| 831 | jbgp.K = jbgp.os_block; |
| 832 | jbgp.K_tail = (jbgp.os % jbgp.os_block) + (do_rnd_os ? 1 : 0); |
| 833 | |
| 834 | jbgp.nb_os_blocking = 1; |
| 835 | int os_blocking_max = (is_amx_xf16 && jbgp.nb_os >= 64) |
| 836 | ? (types::data_type_size(jbgp.src_dt) * jbgp.mb * jbgp.ic |
| 837 | < platform::get_per_core_cache_size(2)) |
| 838 | ? 8 |
| 839 | : 4 |
| 840 | : nstl::min(64, jbgp.nb_os); |
| 841 | |
| 842 | for (int bl = os_blocking_max; bl >= 1; bl--) |
| 843 | if (jbgp.nb_os % bl == 0) { |
| 844 | jbgp.nb_os_blocking = bl; |
| 845 | break; |
| 846 | } |
| 847 | |
| 848 | jbgp.use_buffer_a = true; |
| 849 | const bool is_oc_big_2_pow = jbgp.oc >= 512 && math::is_pow2(jbgp.oc); |
| 850 | const bool is_huge_oc = jbgp.oc >= 4 * 1024; |
| 851 | jbgp.use_buffer_b = jbgp.dst_dt != f32 || is_oc_big_2_pow || is_huge_oc; |
| 852 | const bool os_dim_dominating = jbgp.os >= 5 * (jbgp.ic + jbgp.oc); |
| 853 | const int big_nb_os_threshold = is_amx_xf16 ? 64 : 256; |
| 854 | jbgp.ip_bwd_w_local_buffers_for_input_tensors |
| 855 | = is_amx_xf16 && jbgp.nb_os >= big_nb_os_threshold; |
| 856 | jbgp.harness = os_dim_dominating && jbgp.nb_os >= big_nb_os_threshold |
| 857 | ? harness_mb_reduction |
| 858 | : harness_2d_reduction; |
| 859 | |
| 860 | int nb_os_blocking, nb_oc_blocking, nb_ic_blocking, nthr, nthr_mb, nthr_oc, |
| 861 | nthr_ic; |
| 862 | // Caution: thread_balance requires `use_buffer_a` and `use_buffer_b` |
| 863 | // fields of jbgp to be properly set |
| 864 | thread_balance(jbgp, nb_os_blocking, nb_oc_blocking, nb_ic_blocking, nthr, |
| 865 | nthr_mb, nthr_oc, nthr_ic); |
| 866 | |
| 867 | jbgp.nb_os_blocking = nb_os_blocking; |
| 868 | jbgp.nb_oc_blocking = nb_oc_blocking; |
| 869 | jbgp.nb_ic_blocking = nb_ic_blocking; |
| 870 | jbgp.nthr = nthr; |
| 871 | jbgp.nthr_mb = nthr_mb; |
| 872 | jbgp.nthr_oc_b = nthr_oc; |
| 873 | jbgp.nthr_ic_b = nthr_ic; |
| 874 | |
| 875 | jbgp.gemm_batch_size = jbgp.nb_os_blocking; |
| 876 | // to avoid cache concurrent write access from different threads |
| 877 | size_t sc_size = sizeof(brgemm_batch_element_t); |
| 878 | jbgp.adjusted_batch_size |
| 879 | = div_up(rnd_up(jbgp.gemm_batch_size * sc_size, 4096), sc_size); |
| 880 | |
| 881 | jbgp.use_buffer = IMPLICATION(!has_weights_buffer, jbgp.nthr_mb > 1); |
| 882 | |
| 883 | jbgp.LDA = jbgp.K; |
| 884 | jbgp.LDB = (jbgp.use_buffer_b) ? jbgp.N * jbgp.nb_oc_blocking |
| 885 | : jbgp.oc_without_padding; |
| 886 | jbgp.LDC = jbgp.LDD = jbgp.N; |
| 887 | |
| 888 | if (jbgp.is_bf32) { |
| 889 | const float M = static_cast<float>(jbgp.M); |
| 890 | const float N = nstl::min<float>(jbgp.N, jbgp.oc); |
| 891 | const float K |
| 892 | = nstl::min<float>(jbgp.K * jbgp.gemm_batch_size, jbgp.os); |
| 893 | const float tmul_efficiency = (M / 16) * (N / 16) * (K / 32); |
| 894 | // TODO: Adjust blocking such that bigger M, N, K are generated. |
| 895 | if (one_of(true, M <= 8, K <= 8, N < 16, tmul_efficiency <= 2.25)) |
| 896 | return status::unimplemented; |
| 897 | } |
| 898 | |
| 899 | return status::success; |
| 900 | } |
| 901 | |
| 902 | size_t buf_dt_size(data_type_t dt, cpu_isa_t isa) { |
| 903 | const auto buf_dt = isa == avx512_core_fp16 && dt == data_type::f16 |
| 904 | ? data_type::f32 |
| 905 | : dt; |
| 906 | return types::data_type_size(buf_dt); |
| 907 | } |
| 908 | |
| 909 | status_t init_ip_conf(cpu_isa_t isa, jit_brgemm_primitive_conf_t &jbgp, |
| 910 | const inner_product_desc_t &ipd, memory_desc_t &src_md, |
| 911 | memory_desc_t &weights_md, memory_desc_t &dst_md, |
| 912 | memory_desc_t &bias_md, primitive_attr_t &attr, int nthreads) { |
| 913 | const memory_desc_wrapper src_d(&src_md); |
| 914 | const memory_desc_wrapper weights_d(&weights_md); |
| 915 | const memory_desc_wrapper dst_d(&dst_md); |
| 916 | |
| 917 | using namespace prop_kind; |
| 918 | if (!mayiuse(avx512_core) && !mayiuse(avx2_vnni_2)) |
| 919 | return status::unimplemented; |
| 920 | |
| 921 | int ndims = src_d.ndims(); |
| 922 | if (weights_d.ndims() != ndims || dst_d.ndims() != 2) |
| 923 | return status::unimplemented; |
| 924 | |
| 925 | jbgp = zero<decltype(jbgp)>(); |
| 926 | jbgp.ndims = ndims; |
| 927 | jbgp.isa = isa; |
| 928 | jbgp.prop_kind = ipd.prop_kind; |
| 929 | jbgp.ngroups = 1; |
| 930 | jbgp.mb = src_d.dims()[0]; |
| 931 | jbgp.os = jbgp.mb; |
| 932 | jbgp.oc_without_padding = dst_d.dims()[1]; |
| 933 | jbgp.oc = jbgp.oc_without_padding; |
| 934 | jbgp.ic_without_padding = src_d.dims()[1]; |
| 935 | jbgp.ic = jbgp.ic_without_padding; |
| 936 | jbgp.id = (ndims == 5) ? src_d.dims()[2] : 1; |
| 937 | jbgp.ih = (ndims < 4) ? 1 : src_d.dims()[ndims - 2]; |
| 938 | jbgp.iw = (ndims < 3) ? 1 : src_d.dims()[ndims - 1]; |
| 939 | jbgp.od = jbgp.oh = jbgp.ow = 1; |
| 940 | jbgp.kd = (ndims == 5) ? weights_d.dims()[2] : 1; |
| 941 | jbgp.kh = (ndims < 4) ? 1 : weights_d.dims()[ndims - 2]; |
| 942 | jbgp.kw = (ndims < 3) ? 1 : weights_d.dims()[ndims - 1]; |
| 943 | jbgp.stride_d = jbgp.stride_h = jbgp.stride_w = 1; |
| 944 | |
| 945 | if (!everyone_is(1, jbgp.ow, jbgp.oh, jbgp.od)) |
| 946 | return status::unimplemented; |
| 947 | if (jbgp.kw != jbgp.iw || jbgp.kh != jbgp.ih || jbgp.kd != jbgp.id) |
| 948 | return status::unimplemented; |
| 949 | if (!everyone_is(1, jbgp.kw, jbgp.kh, jbgp.kd)) |
| 950 | return status::unimplemented; |
| 951 | |
| 952 | const int full_simd_w = 16; |
| 953 | jbgp.simd_w = full_simd_w; |
| 954 | |
| 955 | jbgp.with_bias |
| 956 | = pick_by_prop_kind(jbgp.prop_kind, ipd.bias_desc.format_kind, |
| 957 | format_kind::undef, ipd.diff_bias_desc.format_kind) |
| 958 | != format_kind::undef; |
| 959 | |
| 960 | jbgp.src_dt = src_d.data_type(); |
| 961 | jbgp.dst_dt = dst_d.data_type(); |
| 962 | jbgp.wei_dt = weights_d.data_type(); |
| 963 | jbgp.bia_dt = jbgp.with_bias |
| 964 | ? pick_by_prop_kind(jbgp.prop_kind, ipd.bias_desc.data_type, |
| 965 | data_type::undef, ipd.diff_bias_desc.data_type) |
| 966 | : data_type::undef; |
| 967 | jbgp.signed_input = one_of(isa, avx512_core_vnni, avx512_core_bf16) |
| 968 | && jbgp.src_dt == s8; |
| 969 | const bool is_int8 = one_of(jbgp.src_dt, u8, s8) && jbgp.wei_dt == s8; |
| 970 | const bool is_bf16 |
| 971 | = everyone_is(bf16, jbgp.src_dt, jbgp.wei_dt, jbgp.dst_dt) |
| 972 | || pick_by_prop_kind(jbgp.prop_kind, |
| 973 | everyone_is(bf16, jbgp.src_dt, jbgp.wei_dt) |
| 974 | && jbgp.dst_dt == f32, |
| 975 | everyone_is(bf16, jbgp.wei_dt, jbgp.dst_dt) |
| 976 | && jbgp.src_dt == f32, |
| 977 | everyone_is(bf16, jbgp.src_dt, jbgp.dst_dt) |
| 978 | && jbgp.wei_dt == f32); |
| 979 | const bool is_f16 = everyone_is(f16, jbgp.src_dt, jbgp.wei_dt, jbgp.dst_dt) |
| 980 | || pick_by_prop_kind(jbgp.prop_kind, |
| 981 | everyone_is(f16, jbgp.src_dt, jbgp.wei_dt) |
| 982 | && jbgp.dst_dt == f32, |
| 983 | everyone_is(f16, jbgp.wei_dt, jbgp.dst_dt) |
| 984 | && jbgp.src_dt == f32, |
| 985 | everyone_is(f16, jbgp.src_dt, jbgp.dst_dt) |
| 986 | && jbgp.wei_dt == f32); |
| 987 | const bool is_f32 = everyone_is(f32, jbgp.src_dt, jbgp.wei_dt, jbgp.dst_dt); |
| 988 | jbgp.is_bf32 = is_f32 && attr.fpmath_mode_ == fpmath_mode::bf16 |
| 989 | && isa == avx512_core_amx; |
| 990 | |
| 991 | if (!IMPLICATION(is_int8, |
| 992 | one_of(isa, avx512_core_vnni, avx512_core_bf16, |
| 993 | avx512_core_amx))) |
| 994 | return status::unimplemented; |
| 995 | if (!IMPLICATION(is_bf16, |
| 996 | one_of(isa, avx2_vnni_2, avx512_core_bf16, avx512_core_amx))) |
| 997 | return status::unimplemented; |
| 998 | if (!IMPLICATION(is_f32, jbgp.is_bf32 || (isa == avx512_core))) |
| 999 | return status::unimplemented; |
| 1000 | if (!IMPLICATION(is_f16, |
| 1001 | one_of(isa, avx2_vnni_2, avx512_core_fp16, |
| 1002 | avx512_core_amx_fp16))) |
| 1003 | return status::unimplemented; |
| 1004 | |
| 1005 | if (!one_of(true, is_int8, is_bf16, is_f16, is_f32)) |
| 1006 | return status::unimplemented; |
| 1007 | if (is_int8) { |
| 1008 | jbgp.acc_dt = s32; |
| 1009 | jbgp.with_scales = true; |
| 1010 | } else |
| 1011 | jbgp.acc_dt = f32; |
| 1012 | |
| 1013 | // Dispatch small shapes to VNNI for better performance |
| 1014 | const bool is_amx_int8 |
| 1015 | = jbgp.isa == avx512_core_amx && one_of(jbgp.wei_dt, s8, u8); |
| 1016 | const auto amx_row |
| 1017 | = static_cast<int32_t>(data_type_vnni_granularity(jbgp.src_dt)) |
| 1018 | * jbgp.simd_w; |
| 1019 | const auto max_size = is_amx_int8 ? 1024 : 512; |
| 1020 | const bool is_small_shapes |
| 1021 | = (jbgp.os <= 16 && jbgp.ic <= amx_row && jbgp.oc <= amx_row) |
| 1022 | || (jbgp.ic <= max_size && jbgp.oc <= max_size && jbgp.mb == 1 |
| 1023 | && jbgp.ic % amx_row != 0); |
| 1024 | if (one_of(jbgp.isa, avx512_core_amx, avx512_core_amx) && is_small_shapes) |
| 1025 | return status::unimplemented; |
| 1026 | |
| 1027 | auto set_or_check_tags = [&]() -> status_t { |
| 1028 | using namespace format_tag; |
| 1029 | format_tag_t desired_src_tag = pick(ndims - 2, nc, ncw, nchw, ncdhw); |
| 1030 | format_tag_t desired_dst_tag = nc; |
| 1031 | |
| 1032 | if (src_d.format_kind() == format_kind::any) { |
| 1033 | CHECK(memory_desc_init_by_tag(src_md, desired_src_tag)); |
| 1034 | jbgp.src_tag = desired_src_tag; |
| 1035 | } else { |
| 1036 | jbgp.src_tag |
| 1037 | = memory_desc_matches_one_of_tag(src_md, desired_src_tag); |
| 1038 | } |
| 1039 | |
| 1040 | if (dst_d.format_kind() == format_kind::any) { |
| 1041 | CHECK(memory_desc_init_by_tag(dst_md, desired_dst_tag)); |
| 1042 | jbgp.dst_tag = desired_dst_tag; |
| 1043 | } else { |
| 1044 | jbgp.dst_tag = memory_desc_matches_one_of_tag(dst_md, nc); |
| 1045 | } |
| 1046 | |
| 1047 | if (one_of(format_tag::undef, jbgp.src_tag, jbgp.dst_tag)) |
| 1048 | return status::unimplemented; |
| 1049 | |
| 1050 | if (jbgp.with_bias && bias_md.format_kind == format_kind::any) |
| 1051 | CHECK(memory_desc_init_by_tag(bias_md, x)); |
| 1052 | |
| 1053 | jbgp.is_wei_layout_any = weights_d.format_kind() == format_kind::any; |
| 1054 | |
| 1055 | memory_desc_t want_wei_md = weights_md; |
| 1056 | jbgp.wei_tag = get_brgemm_ip_weights_tag(isa, jbgp, weights_md); |
| 1057 | if (jbgp.wei_tag == format_tag::undef) return status::unimplemented; |
| 1058 | CHECK(memory_desc_init_by_tag(want_wei_md, jbgp.wei_tag)); |
| 1059 | |
| 1060 | if (jbgp.signed_input) { |
| 1061 | want_wei_md.extra.flags = 0 |
| 1062 | | memory_extra_flags::compensation_conv_s8s8 |
| 1063 | | memory_extra_flags::scale_adjust; |
| 1064 | want_wei_md.extra.compensation_mask = (1 << 0); |
| 1065 | want_wei_md.extra.scale_adjust |
| 1066 | = platform::s8s8_weights_scale_factor(); |
| 1067 | if (weights_md.format_kind != format_kind::any |
| 1068 | && want_wei_md != weights_md) |
| 1069 | return status::unimplemented; |
| 1070 | } |
| 1071 | weights_md = want_wei_md; |
| 1072 | return status::success; |
| 1073 | }; |
| 1074 | |
| 1075 | jbgp.brg_type = brgemm_addr; |
| 1076 | jbgp.nthr = nthreads; |
| 1077 | |
| 1078 | jbgp.use_uker = true; |
| 1079 | jbgp.use_interleave_stores = jbgp.use_uker; |
| 1080 | if (jbgp.use_uker) |
| 1081 | jbgp.hint_prefetching = brgemm_kernel_prefetching_t::brgemm_prf1; |
| 1082 | CHECK(set_or_check_tags()); |
| 1083 | CHECK(attr.set_default_formats(&dst_md)); |
| 1084 | |
| 1085 | switch (jbgp.prop_kind) { |
| 1086 | case forward_training: |
| 1087 | case forward_inference: |
| 1088 | CHECK(init_ip_conf_fwd(jbgp, attr, dst_d)); |
| 1089 | break; |
| 1090 | case backward_data: CHECK(init_ip_conf_bwd_d(jbgp)); break; |
| 1091 | case backward_weights: CHECK(init_ip_conf_bwd_w(jbgp)); break; |
| 1092 | default: assert(!"invalid prop_kind"); return invalid_arguments; |
| 1093 | } |
| 1094 | |
| 1095 | return status::success; |
| 1096 | } |
| 1097 | |
| 1098 | void init_scratchpad(memory_tracking::registrar_t &scratchpad, |
| 1099 | const jit_brgemm_primitive_conf_t &jbgp) { |
| 1100 | |
| 1101 | size_t sc_size = sizeof(brgemm_batch_element_t); |
| 1102 | size_t n_elems = (size_t)jbgp.nthr * jbgp.adjusted_batch_size; |
| 1103 | |
| 1104 | if (jbgp.brg_type == brgemm_addr) { |
| 1105 | scratchpad.book(key_brgemm_primitive_batch, n_elems, sc_size, 64); |
| 1106 | } |
| 1107 | if (jbgp.use_buffer) { |
| 1108 | size_t nelements = (size_t)jbgp.nthr * jbgp.LDC * jbgp.M; |
| 1109 | if (jbgp.prop_kind == dnnl_backward_weights |
| 1110 | && (jbgp.nthr_mb > 1 || jbgp.harness == harness_mb_reduction)) { |
| 1111 | const size_t n_reduction_buffers = jbgp.nthr_mb > 1 |
| 1112 | ? jbgp.nthr_mb - (jbgp.wei_dt == f32) |
| 1113 | : 1; |
| 1114 | const size_t num_ic_chunks |
| 1115 | = div_up(jbgp.nb_ic, jbgp.nb_ic_blocking); |
| 1116 | const size_t num_oc_chunks |
| 1117 | = div_up(jbgp.nb_oc, jbgp.nb_oc_blocking); |
| 1118 | nelements = (size_t)n_reduction_buffers * num_ic_chunks |
| 1119 | * num_oc_chunks * jbgp.nb_ic_blocking * jbgp.nb_oc_blocking |
| 1120 | * jbgp.ic_block * jbgp.oc_block; |
| 1121 | } else if (jbgp.prop_kind == dnnl_backward_weights |
| 1122 | && jbgp.nthr_mb == 1) { |
| 1123 | nelements = (size_t)jbgp.nthr * jbgp.nb_ic_blocking * jbgp.ic_block |
| 1124 | * jbgp.nb_oc_blocking * jbgp.oc_block; |
| 1125 | } else if (jbgp.prop_kind == dnnl_backward_data && jbgp.nthr_oc_b > 1) { |
| 1126 | const int adj_buffers = (jbgp.src_dt == f32) ? 1 : 0; |
| 1127 | int n_reduction_buffers = jbgp.nthr_oc_b - adj_buffers; |
| 1128 | nelements = (size_t)n_reduction_buffers * jbgp.LDC * jbgp.os; |
| 1129 | } else if (one_of(jbgp.prop_kind, forward_training, forward_inference) |
| 1130 | && jbgp.nthr_ic_b > 1) { |
| 1131 | const bool need_extra_buffer |
| 1132 | = (jbgp.dst_dt == f32 && jbgp.with_sum); |
| 1133 | int n_reduction_buffers = jbgp.nthr_ic_b - !need_extra_buffer; |
| 1134 | nelements = (size_t)n_reduction_buffers * jbgp.oc * jbgp.os; |
| 1135 | } |
| 1136 | scratchpad.book(key_brgemm_primitive_buffer, nelements, |
| 1137 | types::data_type_size(jbgp.acc_dt)); |
| 1138 | } |
| 1139 | if (jbgp.use_buffer_a && jbgp.prop_kind == dnnl_backward_weights) { |
| 1140 | const dim_t num_ic_chunks_per_thread |
| 1141 | = jbgp.ip_bwd_w_local_buffers_for_input_tensors |
| 1142 | ? 1 |
| 1143 | : div_up(div_up(jbgp.nb_ic, jbgp.nb_ic_blocking), |
| 1144 | jbgp.nthr_ic_b); |
| 1145 | const dim_t num_os_chunks_per_thread |
| 1146 | = jbgp.ip_bwd_w_local_buffers_for_input_tensors |
| 1147 | ? 1 |
| 1148 | : div_up(div_up(jbgp.nb_os, jbgp.nb_os_blocking), jbgp.nthr_mb); |
| 1149 | const dim_t num_elems_per_thread = num_ic_chunks_per_thread |
| 1150 | * num_os_chunks_per_thread * jbgp.gemm_batch_size |
| 1151 | * jbgp.os_block * jbgp.ic_block * jbgp.nb_ic_blocking; |
| 1152 | scratchpad.book(key_brgemm_primitive_buffer_a, |
| 1153 | jbgp.nthr * num_elems_per_thread, |
| 1154 | buf_dt_size(jbgp.src_dt, jbgp.isa)); |
| 1155 | } else if (jbgp.use_buffer_a && jbgp.prop_kind == dnnl_backward_data) { |
| 1156 | scratchpad.book(key_brgemm_primitive_buffer_a, |
| 1157 | (size_t)jbgp.nthr * jbgp.os_block * jbgp.LDA, |
| 1158 | buf_dt_size(jbgp.dst_dt, jbgp.isa)); |
| 1159 | } else if (jbgp.use_buffer_a) { // FWD |
| 1160 | scratchpad.book(key_brgemm_primitive_buffer_a, |
| 1161 | (size_t)jbgp.nthr * jbgp.LDA * jbgp.os_block |
| 1162 | * jbgp.nb_os_blocking, |
| 1163 | buf_dt_size(jbgp.src_dt, jbgp.isa)); |
| 1164 | } |
| 1165 | |
| 1166 | if (jbgp.use_buffer_b && jbgp.prop_kind == dnnl_backward_weights) { |
| 1167 | int num_os_chunks_per_thread |
| 1168 | = jbgp.ip_bwd_w_local_buffers_for_input_tensors |
| 1169 | ? 1 |
| 1170 | : div_up(div_up(jbgp.nb_os, jbgp.nb_os_blocking), jbgp.nthr_mb); |
| 1171 | const dim_t num_elems_per_thread = num_os_chunks_per_thread |
| 1172 | * jbgp.gemm_batch_size * jbgp.os_block * jbgp.LDB; |
| 1173 | scratchpad.book(key_brgemm_primitive_buffer_b, |
| 1174 | (size_t)jbgp.nthr * num_elems_per_thread, |
| 1175 | buf_dt_size(jbgp.dst_dt, jbgp.isa)); |
| 1176 | } |
| 1177 | |
| 1178 | if (jbgp.use_buffer_b && jbgp.prop_kind == dnnl_backward_data) { |
| 1179 | auto size_B = (size_t)jbgp.LDB * rnd_up(jbgp.K, 2); |
| 1180 | |
| 1181 | if (!jbgp.ip_bwd_d_global_b_transpose) |
| 1182 | scratchpad.book(key_brgemm_primitive_buffer_b, |
| 1183 | (dim_t)jbgp.nthr * jbgp.gemm_batch_size * size_B, |
| 1184 | buf_dt_size(jbgp.wei_dt, jbgp.isa)); |
| 1185 | else |
| 1186 | scratchpad.book(key_brgemm_primitive_buffer_b, |
| 1187 | (dim_t)jbgp.nb_oc * jbgp.nb_ic * size_B, |
| 1188 | buf_dt_size(jbgp.wei_dt, jbgp.isa)); |
| 1189 | } |
| 1190 | |
| 1191 | if (jbgp.prop_kind == dnnl_backward_weights && jbgp.with_bias |
| 1192 | && (jbgp.bia_dt != f32 || jbgp.nthr_mb > 1)) { |
| 1193 | int nbuffers = jbgp.nthr_mb - (jbgp.bia_dt == f32); |
| 1194 | scratchpad.book(key_iprod_bias_bf16_convert_wsp, |
| 1195 | (size_t)nbuffers * jbgp.oc, types::data_type_size(jbgp.acc_dt)); |
| 1196 | } |
| 1197 | |
| 1198 | if (dnnl_thr_syncable() && jbgp.prop_kind == dnnl_backward_weights) |
| 1199 | scratchpad.book<simple_barrier::ctx_t>( |
| 1200 | key_conv_wei_bia_reduction_bctx, 1); |
| 1201 | |
| 1202 | if (jbgp.is_amx) |
| 1203 | scratchpad.book(key_conv_amx_tile_buffer, |
| 1204 | (size_t)jbgp.nthr * jbgp.amx_buf_size_per_thread, sizeof(char)); |
| 1205 | } |
| 1206 | |
| 1207 | } // namespace brgemm_inner_product_utils |
| 1208 | |
| 1209 | } // namespace x64 |
| 1210 | } // namespace cpu |
| 1211 | } // namespace impl |
| 1212 | } // namespace dnnl |
| 1213 |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.