onednn_reuse.h 79.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once

#include <algorithm>
#include <memory>
#include <sstream>
#include <string>
#include <utility>
#include <vector>

23
#include "paddle/fluid/platform/profiler/event_tracing.h"
24
#include "paddle/phi/backends/onednn/onednn_context.h"
25
#include "paddle/phi/backends/onednn/onednn_helper.h"
26
#include "paddle/phi/common/data_type.h"
27
#include "paddle/phi/common/int_array.h"
28
#include "paddle/phi/common/place.h"
29
#include "paddle/phi/common/scalar.h"
30
#include "paddle/phi/core/dense_tensor.h"
31
#include "paddle/phi/kernels/funcs/axis_utils.h"
32
#include "paddle/phi/kernels/funcs/blas/blas.h"
33
#include "paddle/phi/kernels/funcs/data_layout_transform.h"
34
#include "paddle/phi/kernels/funcs/pooling.h"
35 36 37 38 39 40

namespace phi {
namespace funcs {

using memory = dnnl::memory;

41
using OneDNNMemoryFormat = dnnl::memory::format_tag;
42

43 44 45 46 47 48 49
template <typename T>
bool constexpr is_int8() {
  return std::is_same<T, int8_t>::value || std::is_same<T, uint8_t>::value;
}

template <typename T>
constexpr bool is_bfloat16() {
50
  return std::is_same<T, dtype::bfloat16>::value;
51 52
}

53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104
static void AppendActivation(const OneDNNContext& dev_ctx,
                             dnnl::post_ops& post_ops,  // NOLINT
                             float activation_scale = 1.0f) {
  const auto invalid_attribute =
      dev_ctx.HasDnnAttr("fuse_activation")
          ? PADDLE_GET_CONST(std::string, dev_ctx.GetDnnAttr("fuse_activation"))
                .empty()
          : true;
  if (invalid_attribute) return;

  const auto fuse_activation =
      dev_ctx.HasDnnAttr("fuse_activation")
          ? PADDLE_GET_CONST(std::string, dev_ctx.GetDnnAttr("fuse_activation"))
          : "";
  const auto fuse_alpha =
      dev_ctx.HasDnnAttr("fuse_alpha")
          ? PADDLE_GET_CONST(float, dev_ctx.GetDnnAttr("fuse_alpha"))
          : 0.0f;
  const auto fuse_beta =
      dev_ctx.HasDnnAttr("fuse_beta")
          ? PADDLE_GET_CONST(float, dev_ctx.GetDnnAttr("fuse_beta"))
          : 0.0f;

  if (fuse_activation == "hard_sigmoid") {
    post_ops.append_eltwise(activation_scale,
                            dnnl::algorithm::eltwise_linear,
                            fuse_alpha,
                            fuse_beta);
    post_ops.append_eltwise(
        activation_scale, dnnl::algorithm::eltwise_clip, 0.0f, 1.0f);
  } else {
    const std::unordered_map<std::string, dnnl::algorithm> activation_map = {
        {"abs", dnnl::algorithm::eltwise_abs},
        {"clip", dnnl::algorithm::eltwise_clip},
        {"gelu", dnnl::algorithm::eltwise_gelu_erf},
        {"gelu_erf", dnnl::algorithm::eltwise_gelu_erf},
        {"gelu_tanh", dnnl::algorithm::eltwise_gelu_tanh},
        {"hard_swish", dnnl::algorithm::eltwise_hardswish},
        {"leaky_relu", dnnl::algorithm::eltwise_relu},
        {"mish", dnnl::algorithm::eltwise_mish},
        {"relu", dnnl::algorithm::eltwise_relu},
        {"relu6", dnnl::algorithm::eltwise_bounded_relu},
        {"sigmoid", dnnl::algorithm::eltwise_logistic},
        {"sqrt", dnnl::algorithm::eltwise_sqrt},
        {"swish", dnnl::algorithm::eltwise_swish},
        {"tanh", dnnl::algorithm::eltwise_tanh}};

    const auto& activation_type = activation_map.find(fuse_activation);

    PADDLE_ENFORCE_NE(
        activation_type,
        activation_map.end(),
105
        errors::InvalidArgument(
106 107 108 109 110 111 112 113
            "Activation '%s' not found in oneDNN algorithms mapper",
            fuse_activation));

    post_ops.append_eltwise(
        activation_scale, activation_type->second, fuse_alpha, fuse_beta);
  }
}

114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149
static std::unordered_map<std::string, std::string> GetAttributeMap(
    std::string act_type) {
  std::unordered_map<std::string, std::string> attr_map;
  if (act_type == "swish") {
    attr_map.emplace("beta", "fuse_alpha");
  } else if (act_type == "relu6") {
    attr_map.emplace("threshold", "fuse_alpha");
  } else if (act_type == "hard_sigmoid") {
    attr_map.emplace("slope", "fuse_alpha");
    attr_map.emplace("offset", "fuse_beta");
  } else if (act_type == "clip") {
    attr_map.emplace("min", "fuse_alpha");
    attr_map.emplace("max", "fuse_beta");
  } else {
    attr_map.emplace("alpha", "fuse_alpha");
    attr_map.emplace("beta", "fuse_beta");
  }
  return attr_map;
}

static std::vector<std::string> GetSupportedActivations() {
  return std::vector<std::string>{"abs",
                                  "clip",
                                  "gelu",
                                  "hard_sigmoid",
                                  "hard_swish",
                                  "leaky_relu",
                                  "mish",
                                  "relu",
                                  "relu6",
                                  "sigmoid",
                                  "sqrt",
                                  "swish",
                                  "tanh"};
}

150 151
template <typename T,
          typename TForward,
152 153
          typename TBackward = onednn_dummy_primitive,
          typename TBackward_params = onednn_dummy_primitive>
154
class OneDNNHandlerT {
155
 public:
156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428
  OneDNNHandlerT(const OneDNNContext& dev_ctx,
                 dnnl::engine engine,
                 Place cpu_place,
                 const std::string& base_key)
      : dev_ctx_(dev_ctx),
        engine_(engine),
        place_(cpu_place),
        key_common_(base_key),
        key_(ExtendKeyWithThreadInfoIfNeeded(dev_ctx, base_key)),
        fwd_pd_(nullptr),
        bwd_pd_(nullptr) {
    OneDNNContext::tls().log_lib_version();
  }

  std::shared_ptr<TForward> AcquireForwardPrimitive() {
    const std::string key_p = key_ + "@fwd_p";
    auto forward_p =
        std::static_pointer_cast<TForward>(dev_ctx_.GetBlob(key_p));
    if (forward_p == nullptr) {
      forward_p = std::make_shared<TForward>(*fwd_pd_);
      dev_ctx_.SetBlob(key_p, forward_p);
    }
    return forward_p;
  }

  std::shared_ptr<TBackward> AcquireBackwardPrimitive() {
    const std::string key_p = key_ + "@bwd_p";
    auto backward_p =
        std::static_pointer_cast<TBackward>(dev_ctx_.GetBlob(key_p));
    if (backward_p == nullptr) {
      backward_p = std::make_shared<TBackward>(*bwd_pd_);
      dev_ctx_.SetBlob(key_p, backward_p);
    }
    return backward_p;
  }

  std::shared_ptr<TBackward_params> AcquireBackwardWeightsPrimitive() {
    const std::string key_p = key_ + "@bwd_w_p";
    auto backward_p =
        std::static_pointer_cast<TBackward_params>(dev_ctx_.GetBlob(key_p));
    if (backward_p == nullptr) {
      PADDLE_ENFORCE_NOT_NULL(
          bwd_w_pd_,
          errors::Unavailable("BWD_PD should be set when "
                              "getting BWD prim witk key: %s .",
                              key_p));
      backward_p = std::make_shared<TBackward_params>(*bwd_w_pd_);
      dev_ctx_.SetBlob(key_p, backward_p);
    }
    return backward_p;
  }

  std::shared_ptr<dnnl::memory> AcquireSrcMemory(const DenseTensor* input) {
    const T* input_data = input->data<T>();
    return this->AcquireMemoryFromPrimitive(
        fwd_pd_->src_desc(), to_void_cast<T>(input_data), "@src_mem_p");
  }

  template <typename T_out = T>
  std::shared_ptr<dnnl::memory> AcquireDstMemory(DenseTensor* output) {
    T_out* ptr =
        output->mutable_data<T_out>(place_, fwd_pd_->dst_desc().get_size());
    return this->AcquireMemoryFromPrimitive(
        fwd_pd_->dst_desc(), ptr, "@dst_mem_p");
  }

  template <typename T_out = T>
  std::shared_ptr<dnnl::memory> AcquireDstMemory(void) {
    return this->AcquireMemoryFromPrimitive(fwd_pd_->dst_desc(), "@dstt_mem_p");
  }

  template <typename T_out = T>
  std::shared_ptr<dnnl::memory> AcquireDstMemory(const DenseTensor* output) {
    const T_out* output_data = output->data<T_out>();
    return this->AcquireMemoryFromPrimitive(bwd_pd_->dst_desc(),
                                            to_void_cast<T_out>(output_data),
                                            "@bwd-dst_mem_p");
  }

  std::shared_ptr<dnnl::memory> AcquireDiffDstMemory(
      const DenseTensor* diffdst) {
    const T* ptr = diffdst->data<T>();
    return this->AcquireMemoryFromPrimitive(
        bwd_pd_->diff_dst_desc(), to_void_cast<T>(ptr), "@diff_dst_mem_p");
  }

  std::shared_ptr<dnnl::memory> AcquireDiffSrcMemory(DenseTensor* diffsrc) {
    T* ptr =
        diffsrc->mutable_data<T>(place_, bwd_pd_->diff_src_desc().get_size());
    return this->AcquireMemoryFromPrimitive(
        bwd_pd_->diff_src_desc(), ptr, "@diff_src_mem_p");
  }

  // Buffer of given DenseTensor is used for oneDNN computation
  std::shared_ptr<dnnl::memory> AcquireDiffWeightsMemory(
      DenseTensor* diff_weights) {
    PADDLE_ENFORCE_NOT_NULL(
        bwd_w_pd_,
        errors::Unavailable(
            "BWD_W_PD should be set when getting BWD grad of weights."));
    T* ptr = diff_weights->mutable_data<T>(
        place_, bwd_w_pd_->diff_weights_desc().get_size());
    return this->AcquireMemoryFromPrimitive(
        bwd_w_pd_->diff_weights_desc(), ptr, "@diff_wei_mem_p");
  }

  // Buffer is allocated by oneDNN to store computation results
  std::shared_ptr<dnnl::memory> AcquireDiffWeightsMemory(void) {
    PADDLE_ENFORCE_NOT_NULL(
        bwd_w_pd_,
        errors::Unavailable(
            "BWD_W_PD should be set when getting BWD grad of weights."));
    return this->AcquireMemoryFromPrimitive(bwd_w_pd_->diff_weights_desc(),
                                            "@diff_wei_mem_p");
  }

 protected:
  bool isCached() {
    const std::string key_pd = key_ + "@fwd_pd";
    fwd_pd_ = std::static_pointer_cast<typename TForward::primitive_desc>(
        dev_ctx_.GetBlob(key_pd));

    return (fwd_pd_ != nullptr);
  }

  bool isBwdCached() {
    const std::string key_pd = key_ + "@bwd_pd";
    bwd_pd_ = std::static_pointer_cast<typename TBackward::primitive_desc>(
        dev_ctx_.GetBlob(key_pd));

    if (bwd_pd_ == nullptr) {
      return false;
    } else {
      if (std::is_same<TBackward_params, onednn_dummy_primitive>::value ==
          false) {
        const std::string key_bw_w_pd = key_ + "@bwd_w_pd";
        bwd_w_pd_ =
            std::static_pointer_cast<typename TBackward_params::primitive_desc>(
                dev_ctx_.GetBlob(key_bw_w_pd));
      }

      // When BWD is cached then still we need to Get FWD PD
      const std::string key_fpd = key_ + "@fwd_pd";
      fwd_pd_ = std::static_pointer_cast<typename TForward::primitive_desc>(
          dev_ctx_.GetBlob(key_fpd));
      PADDLE_ENFORCE_NOT_NULL(
          fwd_pd_,
          errors::Unavailable(
              "Error: FWD PD should be set when BWD PD is cached."));
      return true;
    }
  }

  // If your primitive descriptor requires attributes, pass them as a
  // first argument and paramters to descriptor constructor in the following
  // arguments. Otherwise, all arguments will be forwarded to descriptor
  // constructor, including the first one.
  template <typename Arg, typename... Args>
  void AcquireForwardPrimitiveDescriptor(Arg&& first_arg, Args&&... args) {
    // This is used when we can recreate FWD PD in BWD so
    // we do not need to pass FWD to BWD
    const std::string key_pd = key_ + "@fwd_pd";
    fwd_pd_ = std::static_pointer_cast<typename TForward::primitive_desc>(
        dev_ctx_.GetBlob(key_pd));
    if (fwd_pd_ == nullptr) {
      CreateForwardPrimitiveDescriptor(first_arg, std::forward<Args>(args)...);
      dev_ctx_.SetBlob(key_pd, fwd_pd_);
    }
  }

  // Using sfinae to specialise variadic function. Workaround for not having
  // if constexpr in C++ 11.
  template <class First, class... Args>
  typename std::enable_if<std::is_same<typename std::decay<First>::type,
                                       dnnl::primitive_attr>::value>::type
  CreateForwardPrimitiveDescriptor(First&& first, Args&&... args) {
    auto fwd_desc = typename TForward::desc(std::forward<Args>(args)...);
    fwd_pd_ = std::make_shared<typename TForward::primitive_desc>(
        fwd_desc, first, engine_);
  }

  template <class First, class... Args>
  typename std::enable_if<!std::is_same<typename std::decay<First>::type,
                                        dnnl::primitive_attr>::value>::type
  CreateForwardPrimitiveDescriptor(First&& first, Args&&... args) {
    auto fwd_desc = typename TForward::desc(std::forward<First>(first),
                                            std::forward<Args>(args)...);
    fwd_pd_ =
        std::make_shared<typename TForward::primitive_desc>(fwd_desc, engine_);
  }

  template <typename... Args>
  void AcquireBackwardPrimitiveDescriptor(Args&&... args) {
    // fwd_pd_ is set during grad by calling
    // AcquireForwardPrimitiveDescriptor
    PADDLE_ENFORCE_NOT_NULL(
        fwd_pd_,
        errors::Unavailable("Get OneDNN Forward primitive %s failed.",
                            key_ + "@fwd_pd"));
    const std::string key_pd = key_ + "@bwd_pd";
    bwd_pd_ = std::static_pointer_cast<typename TBackward::primitive_desc>(
        dev_ctx_.GetBlob(key_pd));
    if (bwd_pd_ == nullptr) {
      auto bwd_desc = typename TBackward::desc(std::forward<Args>(args)...);
      bwd_pd_ = std::make_shared<typename TBackward::primitive_desc>(
          bwd_desc, engine_, *fwd_pd_);
      dev_ctx_.SetBlob(key_pd, bwd_pd_);
    }
  }

  template <typename... Args>
  void AcquireBackwardWeightsPrimitiveDescriptor(Args&&... args) {
    // fwd_pd_ is set during grad by calling
    // AcquireForwardPrimitiveDescriptor
    PADDLE_ENFORCE_NOT_NULL(
        fwd_pd_,
        errors::Unavailable("Get OneDNN Forward primitive %s failed.",
                            key_ + "@fwd_pd"));
    const std::string key_pd = key_ + "@bwd_w_pd";
    bwd_w_pd_ =
        std::static_pointer_cast<typename TBackward_params::primitive_desc>(
            dev_ctx_.GetBlob(key_pd));
    if (bwd_w_pd_ == nullptr) {
      auto bwd_desc =
          typename TBackward_params::desc(std::forward<Args>(args)...);
      bwd_w_pd_ = std::make_shared<typename TBackward_params::primitive_desc>(
          bwd_desc, engine_, *fwd_pd_);
      dev_ctx_.SetBlob(key_pd, bwd_w_pd_);
    }
  }

  std::shared_ptr<dnnl::memory> AcquireMemoryFromPrimitive(
      const std::string& suffix) {
    return std::static_pointer_cast<dnnl::memory>(
        dev_ctx_.GetBlob(key_ + suffix));
  }

  std::shared_ptr<dnnl::memory> AcquireMemoryFromPrimitive(
      dnnl::memory::desc md, void* ptr, const std::string& suffix) {
    const auto local_key = key_ + suffix;
    auto mem_p =
        std::static_pointer_cast<dnnl::memory>(dev_ctx_.GetBlob(local_key));
    if (mem_p == nullptr) {
      mem_p = std::make_shared<dnnl::memory>(md, engine_, ptr);
      dev_ctx_.SetBlob(local_key, mem_p);
    } else {
      mem_p->set_data_handle(ptr);
    }
    return mem_p;
  }

  std::shared_ptr<dnnl::memory> AcquireMemoryFromPrimitive(
      dnnl::memory::desc md, const std::string& suffix) {
    const auto local_key = key_ + suffix;
    auto mem_p =
        std::static_pointer_cast<dnnl::memory>(dev_ctx_.GetBlob(local_key));
    if (mem_p == nullptr) {
      mem_p = std::make_shared<dnnl::memory>(md, engine_);
      dev_ctx_.SetBlob(local_key, mem_p);
    }
    return mem_p;
  }

  void AcquireReorder(const std::shared_ptr<dnnl::memory>& user_memory_p,
                      const std::shared_ptr<dnnl::memory>& target_memory_p) {
    auto reorder_p =
        std::make_shared<dnnl::reorder>(*user_memory_p, *target_memory_p);

    auto& astream = OneDNNContext::tls().get_stream();

    paddle::platform::RecordEvent record_reorder(
        "int_reorder",
        paddle::platform::TracerEventType::UserDefined,
429
        1,
430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481
        paddle::platform::EventRole::kUniqueOp);
    reorder_p->execute(
        astream,
        {{DNNL_ARG_FROM, *user_memory_p}, {DNNL_ARG_TO, *target_memory_p}});
    astream.wait();
  }

  template <typename F = T>
  std::shared_ptr<dnnl::memory> AcquireMemoryWithReorder(
      const dnnl::memory::desc& user_md,
      const dnnl::memory::desc& target_md,
      void* ptr,
      const std::string& suffix,
      bool is_persistent = false,
      std::function<std::shared_ptr<F>(const F*)> custom_reorder_func = {},
      const std::vector<float>& scale_data = {1.0f},
      int mask = 0) {
    const auto target_key = key_ + suffix + "_target";
    const auto key_reorder_p = key_ + suffix + "reorder_p";
    const auto user_key = key_ + suffix + "_user";

    auto target_memory_p =
        std::static_pointer_cast<dnnl::memory>(dev_ctx_.GetBlob(target_key));

    if (target_memory_p == nullptr) {
      if (custom_reorder_func) {
        auto reordered_data =
            custom_reorder_func(reinterpret_cast<const F*>(ptr));
        dev_ctx_.SetBlob(key_reorder_p + "-custom_reorder", reordered_data);
        ptr = reinterpret_cast<void*>(reordered_data.get());
      }
      auto user_memory_p =
          std::make_shared<dnnl::memory>(user_md, engine_, ptr);
      if (user_md != target_md) {
        target_memory_p = std::make_shared<dnnl::memory>(target_md, engine_);
        dnnl::reorder::primitive_desc reorder_pdesc;
        if (is_int8<T>()) {
          dnnl::primitive_attr attr;
          attr.set_output_scales(mask, scale_data);
          reorder_pdesc = dnnl::reorder::primitive_desc(
              *user_memory_p, *target_memory_p, attr);
        } else {
          reorder_pdesc =
              dnnl::reorder::primitive_desc(*user_memory_p, *target_memory_p);
        }
        auto reorder_p = std::make_shared<dnnl::reorder>(reorder_pdesc);
        dev_ctx_.SetBlob(key_reorder_p, reorder_p);

        auto& astream = OneDNNContext::tls().get_stream();
        paddle::platform::RecordEvent record_reorder(
            "int_reorder",
            paddle::platform::TracerEventType::UserDefined,
482
            1,
483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507
            paddle::platform::EventRole::kUniqueOp);
        reorder_p->execute(
            astream,
            {{DNNL_ARG_FROM, *user_memory_p}, {DNNL_ARG_TO, *target_memory_p}});
        astream.wait();
      } else {
        target_memory_p = user_memory_p;
      }
      dev_ctx_.SetBlob(user_key, user_memory_p);
      dev_ctx_.SetBlob(target_key, target_memory_p);
    } else if (!is_persistent) {
      auto& astream = OneDNNContext::tls().get_stream();

      auto user_memory_p =
          std::static_pointer_cast<dnnl::memory>(dev_ctx_.GetBlob(user_key));
      user_memory_p->set_data_handle(ptr);

      // TODO(jczaja): Here we detect if reorder is cached it means it is needed
      // need to change this to get rid of keys
      auto reorder_p = std::static_pointer_cast<dnnl::reorder>(
          dev_ctx_.GetBlob(key_reorder_p));
      if (reorder_p != nullptr) {
        paddle::platform::RecordEvent record_reorder(
            "int_reorder",
            paddle::platform::TracerEventType::UserDefined,
508
            1,
509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540
            paddle::platform::EventRole::kUniqueOp);
        reorder_p->execute(
            astream,
            {{DNNL_ARG_FROM, *user_memory_p}, {DNNL_ARG_TO, *target_memory_p}});
        astream.wait();
      }
    }
    return target_memory_p;
  }

  std::shared_ptr<dnnl::memory> AcquireMemory(const std::string& suffix) {
    const auto local_key = key_ + suffix;
    return std::static_pointer_cast<dnnl::memory>(dev_ctx_.GetBlob(local_key));
  }

  const OneDNNContext& dev_ctx_;
  dnnl::engine engine_;
  Place place_;
  std::string key_common_;
  std::string key_;
  std::shared_ptr<typename TForward::primitive_desc> fwd_pd_;
  std::shared_ptr<typename TBackward::primitive_desc> bwd_pd_;
  std::shared_ptr<typename TBackward_params::primitive_desc> bwd_w_pd_;
};

template <typename T,
          typename TForward,
          typename TBackward = onednn_dummy_primitive,
          typename TBackward_params = onednn_dummy_primitive>
class OneDNNHandlerNoCachingT {
 public:
  OneDNNHandlerNoCachingT(dnnl::engine engine, Place cpu_place)
541
      : engine_(engine), place_(cpu_place), fwd_pd_(nullptr), bwd_pd_(nullptr) {
542
    OneDNNContext::tls().log_lib_version();
543 544 545 546 547 548 549 550 551 552 553
  }

  std::shared_ptr<TForward> AcquireForwardPrimitive() {
    return std::make_shared<TForward>(*fwd_pd_);
  }

  std::shared_ptr<TBackward> AcquireBackwardPrimitive() {
    return std::make_shared<TBackward>(*bwd_pd_);
  }

  std::shared_ptr<TBackward_params> AcquireBackwardWeightsPrimitive() {
554 555 556
    PADDLE_ENFORCE_NOT_NULL(bwd_w_pd_,
                            errors::Unavailable("BWD_PD should be set when "
                                                "getting BWD prim ."));
557 558 559 560 561
    return std::make_shared<TBackward_params>(*bwd_w_pd_);
  }

  std::shared_ptr<dnnl::memory> AcquireSrcMemory(const DenseTensor* input) {
    const T* input_data = input->data<T>();
562 563
    return this->AcquireMemoryFromPrimitive(fwd_pd_->src_desc(),
                                            to_void_cast<T>(input_data));
564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580
  }

  template <typename T_out = T>
  std::shared_ptr<dnnl::memory> AcquireDstMemory(DenseTensor* output) {
    T_out* ptr =
        output->mutable_data<T_out>(place_, fwd_pd_->dst_desc().get_size());
    return this->AcquireMemoryFromPrimitive(fwd_pd_->dst_desc(), ptr);
  }

  template <typename T_out = T>
  std::shared_ptr<dnnl::memory> AcquireDstMemory(void) {
    return this->AcquireMemoryFromPrimitive(fwd_pd_->dst_desc());
  }

  template <typename T_out = T>
  std::shared_ptr<dnnl::memory> AcquireDstMemory(const DenseTensor* output) {
    const T_out* output_data = output->data<T_out>();
581 582
    return this->AcquireMemoryFromPrimitive(bwd_pd_->dst_desc(),
                                            to_void_cast<T_out>(output_data));
583 584 585 586 587
  }

  std::shared_ptr<dnnl::memory> AcquireDiffDstMemory(
      const DenseTensor* diffdst) {
    const T* ptr = diffdst->data<T>();
588 589
    return this->AcquireMemoryFromPrimitive(bwd_pd_->diff_dst_desc(),
                                            to_void_cast<T>(ptr));
590 591 592 593 594 595 596 597
  }

  std::shared_ptr<dnnl::memory> AcquireDiffSrcMemory(DenseTensor* diffsrc) {
    T* ptr =
        diffsrc->mutable_data<T>(place_, bwd_pd_->diff_src_desc().get_size());
    return this->AcquireMemoryFromPrimitive(bwd_pd_->diff_src_desc(), ptr);
  }

598
  // Buffer of given DenseTensor is used for oneDNN computation
599 600 601 602
  std::shared_ptr<dnnl::memory> AcquireDiffWeightsMemory(
      DenseTensor* diff_weights) {
    PADDLE_ENFORCE_NOT_NULL(
        bwd_w_pd_,
603
        errors::Unavailable(
604 605 606 607 608 609 610 611 612 613 614
            "BWD_W_PD should be set when getting BWD grad of weights."));
    T* ptr = diff_weights->mutable_data<T>(
        place_, bwd_w_pd_->diff_weights_desc().get_size());
    return this->AcquireMemoryFromPrimitive(bwd_w_pd_->diff_weights_desc(),
                                            ptr);
  }

  // Buffer is allocated by oneDNN to store computation results
  std::shared_ptr<dnnl::memory> AcquireDiffWeightsMemory(void) {
    PADDLE_ENFORCE_NOT_NULL(
        bwd_w_pd_,
615
        errors::Unavailable(
616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656
            "BWD_W_PD should be set when getting BWD grad of weights."));
    return this->AcquireMemoryFromPrimitive(bwd_w_pd_->diff_weights_desc());
  }

 protected:
  // If your primitive descriptor requires attributes, pass them as a
  // first argument and paramters to descriptor constructor in the following
  // arguments. Otherwise, all arguments will be forwarded to descriptor
  // constructor, including the first one.
  template <typename Arg, typename... Args>
  void AcquireForwardPrimitiveDescriptor(Arg&& first_arg, Args&&... args) {
    CreateForwardPrimitiveDescriptor(first_arg, std::forward<Args>(args)...);
  }

  // Using sfinae to specialise variadic function. Workaround for not having
  // if constexpr in C++ 11.
  template <class First, class... Args>
  typename std::enable_if<std::is_same<typename std::decay<First>::type,
                                       dnnl::primitive_attr>::value>::type
  CreateForwardPrimitiveDescriptor(First&& first, Args&&... args) {
    auto fwd_desc = typename TForward::desc(std::forward<Args>(args)...);
    fwd_pd_ = std::make_shared<typename TForward::primitive_desc>(
        fwd_desc, first, engine_);
  }

  template <class First, class... Args>
  typename std::enable_if<!std::is_same<typename std::decay<First>::type,
                                        dnnl::primitive_attr>::value>::type
  CreateForwardPrimitiveDescriptor(First&& first, Args&&... args) {
    auto fwd_desc = typename TForward::desc(std::forward<First>(first),
                                            std::forward<Args>(args)...);
    fwd_pd_ =
        std::make_shared<typename TForward::primitive_desc>(fwd_desc, engine_);
  }

  template <typename... Args>
  void AcquireBackwardPrimitiveDescriptor(Args&&... args) {
    // fwd_pd_ is set during grad by calling
    // AcquireForwardPrimitiveDescriptor
    PADDLE_ENFORCE_NOT_NULL(
        fwd_pd_,
657
        errors::Unavailable("Get oneDNN Forward primitive %s failed."));
658 659 660 661 662 663 664 665 666 667 668
    auto bwd_desc = typename TBackward::desc(std::forward<Args>(args)...);
    bwd_pd_ = std::make_shared<typename TBackward::primitive_desc>(
        bwd_desc, engine_, *fwd_pd_);
  }

  template <typename... Args>
  void AcquireBackwardWeightsPrimitiveDescriptor(Args&&... args) {
    // fwd_pd_ is set during grad by calling
    // AcquireForwardPrimitiveDescriptor
    PADDLE_ENFORCE_NOT_NULL(
        fwd_pd_,
669
        errors::Unavailable("Get oneDNN Forward primitive %s failed."));
670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690
    auto bwd_desc =
        typename TBackward_params::desc(std::forward<Args>(args)...);
    bwd_w_pd_ = std::make_shared<typename TBackward_params::primitive_desc>(
        bwd_desc, engine_, *fwd_pd_);
  }

  std::shared_ptr<dnnl::memory> AcquireMemoryFromPrimitive(
      dnnl::memory::desc md, void* ptr) {
    return std::make_shared<dnnl::memory>(md, engine_, ptr);
  }

  std::shared_ptr<dnnl::memory> AcquireMemoryFromPrimitive(
      dnnl::memory::desc md) {
    return std::make_shared<dnnl::memory>(md, engine_);
  }

  void AcquireReorder(const std::shared_ptr<dnnl::memory>& user_memory_p,
                      const std::shared_ptr<dnnl::memory>& target_memory_p) {
    auto reorder_p =
        std::make_shared<dnnl::reorder>(*user_memory_p, *target_memory_p);

691
    auto& astream = OneDNNContext::tls().get_stream();
692 693 694 695

    paddle::platform::RecordEvent record_reorder(
        "int_reorder",
        paddle::platform::TracerEventType::UserDefined,
696
        1,
697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722
        paddle::platform::EventRole::kUniqueOp);
    reorder_p->execute(
        astream,
        {{DNNL_ARG_FROM, *user_memory_p}, {DNNL_ARG_TO, *target_memory_p}});
    astream.wait();
  }

  template <typename F = T>
  std::shared_ptr<dnnl::memory> AcquireMemoryWithReorder(
      const dnnl::memory::desc& user_md,
      const dnnl::memory::desc& target_md,
      void* ptr,
      bool is_persistent = false,
      std::function<std::shared_ptr<F>(const F*)> custom_reorder_func = {}) {
    std::shared_ptr<dnnl::memory> target_memory_p;
    if (custom_reorder_func) {
      auto reordered_data =
          custom_reorder_func(reinterpret_cast<const F*>(ptr));
      ptr = reinterpret_cast<void*>(reordered_data.get());
    }
    auto user_memory_p = std::make_shared<dnnl::memory>(user_md, engine_, ptr);
    if (user_md != target_md) {
      target_memory_p = std::make_shared<dnnl::memory>(target_md, engine_);
      auto reorder_p =
          std::make_shared<dnnl::reorder>(*user_memory_p, *target_memory_p);

723
      auto& astream = OneDNNContext::tls().get_stream();
724 725 726
      paddle::platform::RecordEvent record_reorder(
          "int_reorder",
          paddle::platform::TracerEventType::UserDefined,
727
          1,
728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746
          paddle::platform::EventRole::kUniqueOp);
      reorder_p->execute(
          astream,
          {{DNNL_ARG_FROM, *user_memory_p}, {DNNL_ARG_TO, *target_memory_p}});
      astream.wait();
    } else {
      target_memory_p = user_memory_p;
    }
    return target_memory_p;
  }

  dnnl::engine engine_;
  Place place_;
  std::shared_ptr<typename TForward::primitive_desc> fwd_pd_;
  std::shared_ptr<typename TBackward::primitive_desc> bwd_pd_;
  std::shared_ptr<typename TBackward_params::primitive_desc> bwd_w_pd_;
};

template <typename T>
747
class ActivationOneDNNHandler
748
    : public OneDNNHandlerNoCachingT<T,
749 750 751
                                     dnnl::eltwise_forward,
                                     dnnl::eltwise_backward> {
 public:
752
  ActivationOneDNNHandler(dnnl::algorithm algorithm,
753 754 755 756 757
                          float alpha,
                          float beta,
                          const dnnl::engine engine,
                          Place cpu_place,
                          const DenseTensor* x)
758
      : OneDNNHandlerNoCachingT<T,
759 760 761 762 763 764 765 766 767
                                dnnl::eltwise_forward,
                                dnnl::eltwise_backward>(engine, cpu_place) {
    this->AcquireForwardPrimitiveDescriptor(dnnl::prop_kind::forward_training,
                                            algorithm,
                                            x->mem_desc(),
                                            alpha,
                                            beta);
  }

768
  ActivationOneDNNHandler(dnnl::algorithm algorithm,
769 770 771 772 773 774
                          float alpha,
                          float beta,
                          const dnnl::engine engine,
                          Place cpu_place,
                          const DenseTensor* x,
                          const DenseTensor* dout)
775
      : OneDNNHandlerNoCachingT<T,
776 777 778 779 780 781 782 783 784 785 786 787 788 789
                                dnnl::eltwise_forward,
                                dnnl::eltwise_backward>(engine, cpu_place) {
    this->AcquireForwardPrimitiveDescriptor(dnnl::prop_kind::forward_training,
                                            algorithm,
                                            x->mem_desc(),
                                            alpha,
                                            beta);
    this->AcquireBackwardPrimitiveDescriptor(
        algorithm, dout->mem_desc(), x->mem_desc(), alpha, beta);
  }

  std::shared_ptr<dnnl::memory> AcquireBackwardSrcMemory(
      const DenseTensor* input) {
    const T* input_data = input->data<T>();
790 791 792 793 794
    return this->AcquireMemoryFromPrimitive(this->bwd_pd_->src_desc(),
                                            to_void_cast<T>(input_data));
  }
};

795 796 797 798 799 800 801 802
template <typename T>
class SoftmaxOneDNNHandler
    : public OneDNNHandlerNoCachingT<T,
                                     dnnl::softmax_forward,
                                     dnnl::softmax_backward> {
 public:
  SoftmaxOneDNNHandler(const dnnl::engine onednn_engine,
                       Place cpu_place,
803
                       int axis,
804
                       const DenseTensor* x,
805
                       DenseTensor* out)
806 807 808 809
      : OneDNNHandlerNoCachingT<T,
                                dnnl::softmax_forward,
                                dnnl::softmax_backward>(onednn_engine,
                                                        cpu_place) {
810 811 812
    PADDLE_ENFORCE_EQ(
        x->dims(),
        out->dims(),
813
        errors::InvalidArgument(
814 815
            "The shape of input and output tensor must be identical."));

816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838
    const int canonical_axis = funcs::CanonicalAxis(axis, x->dims().size());
    this->AcquireForwardPrimitiveDescriptor(
        dnnl::prop_kind::forward_scoring, x->mem_desc(), canonical_axis);
  }

  SoftmaxOneDNNHandler(const dnnl::engine onednn_engine,
                       Place cpu_place,
                       int axis,
                       const DenseTensor* out,
                       const DenseTensor* out_grad)
      : OneDNNHandlerNoCachingT<T,
                                dnnl::softmax_forward,
                                dnnl::softmax_backward>(onednn_engine,
                                                        cpu_place) {
    const int canonical_axis =
        funcs::CanonicalAxis(axis, out_grad->dims().size());
    this->AcquireForwardPrimitiveDescriptor(
        dnnl::prop_kind::forward_scoring, out->mem_desc(), canonical_axis);
    this->AcquireBackwardPrimitiveDescriptor(
        out_grad->mem_desc(), out->mem_desc(), canonical_axis);
  }
};

839
class ReorderOneDNNHandler {
840
 public:
841
  ReorderOneDNNHandler(std::vector<int64_t>& dims,  // NOLINT
842 843 844 845 846 847 848 849 850 851
                       DataType ptype,
                       dnnl::memory::data_type dtype,
                       dnnl::engine engine)
      : dims_(dims),
        ptype_(ptype),
        ptype_dst_(ptype),
        dtype_(dtype),
        dtype_dst_(dtype),
        engine_(engine) {}

852
  ReorderOneDNNHandler(std::vector<int64_t>& dims,  // NOLINT
853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869
                       DataType ptype,
                       dnnl::memory::data_type dtype,
                       DataType ptype_dst,
                       dnnl::memory::data_type dtype_dst,
                       dnnl::engine engine)
      : dims_(dims),
        ptype_(ptype),
        ptype_dst_(ptype_dst),
        dtype_(dtype),
        dtype_dst_(dtype_dst),
        engine_(engine) {}

  std::shared_ptr<dnnl::memory> AcquireSrcMemory(const dnnl::memory::desc& md,
                                                 void* ptr) {
    return std::make_shared<dnnl::memory>(md, engine_, ptr);
  }

870
  std::shared_ptr<dnnl::memory> AcquireSrcMemory(const OneDNNMemoryFormat& fmt,
871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886
                                                 void* ptr) {
    auto md = dnnl::memory::desc(dims_, dtype_, fmt);
    return std::make_shared<dnnl::memory>(md, engine_, ptr);
  }

  std::shared_ptr<dnnl::memory> AcquireSubmemory(
      const std::vector<int64_t>& dims,
      const std::vector<int64_t>& offset,
      const std::shared_ptr<dnnl::memory>& mem_p) {
    auto sub_md = mem_p->get_desc().submemory_desc(dims, {offset});
    auto sub_mem_p = std::make_shared<dnnl::memory>(
        sub_md, engine_, mem_p->get_data_handle());
    return sub_mem_p;
  }

  std::shared_ptr<dnnl::memory> AcquireDstMemory(DenseTensor* output,
887
                                                 const OneDNNMemoryFormat& fmt,
888
                                                 Place place) {
889
    auto dst_md = OneDNNMemDesc(dims_, dtype_dst_, fmt);
890 891
    auto dst_data = output->mutable_data(place, ptype_dst_, dst_md.get_size());
    return std::make_shared<dnnl::memory>(dst_md, engine_, dst_data);
892
  }
893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908

  std::shared_ptr<dnnl::memory> AcquireDstMemory(
      DenseTensor* output, const dnnl::memory::desc& src_md, Place place) {
    if (ptype_dst_ == ptype_) {
      auto dst_data =
          output->mutable_data(place, ptype_dst_, src_md.get_size());
      return std::make_shared<dnnl::memory>(src_md, engine_, dst_data);
    } else {
      auto dst_md = src_md;
      dst_md.data.data_type = static_cast<dnnl_data_type_t>(dtype_dst_);
      auto dst_data =
          output->mutable_data(place, ptype_dst_, dst_md.get_size());
      return std::make_shared<dnnl::memory>(dst_md, engine_, dst_data);
    }
  }

909 910 911 912 913 914 915 916 917 918
  std::shared_ptr<dnnl::memory> AcquireDstMemory(
      DenseTensor* output,
      const std::vector<int64_t>& dims,
      const std::vector<int64_t>& strides,
      Place place) {
    auto dst_md = dnnl::memory::desc(dims, dtype_dst_, strides);
    auto dst_data = output->mutable_data(place, ptype_dst_, dst_md.get_size());
    return std::make_shared<dnnl::memory>(dst_md, engine_, dst_data);
  }

919 920 921
  std::shared_ptr<dnnl::memory> AcquireDstMemory(
      DenseTensor* output,
      const std::vector<int64_t>& dims,
922
      const OneDNNMemoryFormat& fmt,
923
      Place place) {
924
    auto dst_md = OneDNNMemDesc(dims, dtype_dst_, fmt);
925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947
    auto dst_data = output->mutable_data(place, ptype_dst_, dst_md.get_size());
    return std::make_shared<dnnl::memory>(dst_md, engine_, dst_data);
  }

  std::shared_ptr<dnnl::reorder> AcquireReorder(
      std::shared_ptr<dnnl::memory> dst_memory_p,
      std::shared_ptr<dnnl::memory> src_memory_p) {
    return std::make_shared<dnnl::reorder>(*(src_memory_p), *(dst_memory_p));
  }

  std::shared_ptr<dnnl::reorder> AcquireReorder(
      std::shared_ptr<dnnl::memory> dst_memory_p,
      std::shared_ptr<dnnl::memory> src_memory_p,
      const dnnl::primitive_attr& attrs) {
    return std::make_shared<dnnl::reorder>(
        *(src_memory_p), *(dst_memory_p), attrs);
  }

 private:
  std::vector<int64_t> dims_;
  DataType ptype_, ptype_dst_;
  dnnl::memory::data_type dtype_, dtype_dst_;
  dnnl::engine engine_;
948 949
};

950 951 952
template <typename T>
class BinaryOneDNNHandler : public OneDNNHandlerNoCachingT<T, dnnl::binary> {
 public:
953
  bool use_broadcasting_hack;
954 955 956 957 958 959 960 961 962 963
  BinaryOneDNNHandler(const dnnl::algorithm algo,
                      const int axis,
                      const dnnl::engine engine,
                      Place cpu_place,
                      const DenseTensor* x,
                      const DenseTensor* y,
                      DenseTensor* out,
                      float scale_x,
                      float scale_y,
                      float scale_out,
964
                      bool allow_hack,
965 966
                      const dnnl::post_ops& post_ops = dnnl::post_ops{})
      : OneDNNHandlerNoCachingT<T, dnnl::binary>(engine, cpu_place) {
967
    use_broadcasting_hack = false;
968 969 970 971 972
    const auto src_x_tz = vectorize(x->dims());
    const auto src_y_tz = vectorize(y->dims());
    // if output tensor(z) is nullptr then we are computing into oneDNN
    // managed buffer
    auto rankdiff = x->dims().size() - y->dims().size();
973 974
    auto dst_tz = (out == nullptr) ? (rankdiff > 0 ? src_x_tz : src_y_tz)
                                   : vectorize(out->dims());
975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004

    auto src0_md = x->mem_desc();
    auto src1_md = y->mem_desc();
    if (rankdiff > 0) {  // Second input is of smaller rank than first
      std::vector<int64_t> dims1_ex(rankdiff, 1);
      dims1_ex.insert(next(dims1_ex.begin(), (axis == -1 ? rankdiff : axis)),
                      src_y_tz.begin(),
                      src_y_tz.end());
      // For broadcasting for NHWC we need rotate extended shape
      if (OneDNNContext::tls().get_cur_paddle_data_layout() ==
          DataLayout::kNHWC) {
        std::rotate(dims1_ex.begin() + 1, dims1_ex.end() - 1, dims1_ex.end());
      }
      src1_md = src1_md.reshape(dims1_ex);
    } else if (rankdiff < 0) {  // First input is of smaller than second
      std::vector<int64_t> dims0_ex(-rankdiff, 1);
      dims0_ex.insert(next(dims0_ex.begin(), (axis == -1 ? -rankdiff : axis)),
                      src_x_tz.begin(),
                      src_x_tz.end());
      // For broadcasting for NHWC we need rotate extended shape
      if (OneDNNContext::tls().get_cur_paddle_data_layout() ==
          DataLayout::kNHWC) {
        std::rotate(dims0_ex.begin() + 1, dims0_ex.end() - 1, dims0_ex.end());
      }
      src0_md = src0_md.reshape(dims0_ex);
    }

    auto attributes =
        CreateAttributes(algo, scale_x, scale_y, scale_out, post_ops);

1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042
    // Workaround for U2++ model which deletes first tensor dimensions to enable
    // optimized oneDNNs broadcasting. Output tensor is reshaped back afterwards
    // at the end of the kernel, after the computation
    if (allow_hack && dst_tz.size() == 4 &&
        src0_md.dims()[2] != src1_md.dims()[2]) {
      auto are_strides_plain = [](int64_t* strides, int ndims) {
        for (int i = 0; i < ndims - 1; ++i) {
          if (strides[i] < strides[i + 1]) {
            return false;
          }
        }
        return true;
      };

      auto src0_strides = src0_md.data.format_desc.blocking.strides;
      auto src1_strides = src1_md.data.format_desc.blocking.strides;
      auto src0_dims = src0_md.dims();
      auto src1_dims = src1_md.dims();

      bool can_squeeze = src0_dims[0] == src1_dims[0] &&
                         src0_dims[1] == src1_dims[1] &&
                         src0_dims[3] == src1_dims[3];

      if (can_squeeze && are_strides_plain(src0_strides, 4) &&
          are_strides_plain(src1_strides, 4)) {
        src0_dims[1] *= dst_tz[0];
        src1_dims[1] *= dst_tz[0];
        dst_tz[1] *= dst_tz[0];
        dst_tz.erase(dst_tz.begin());
        src0_md = src0_md.reshape({src0_dims.begin() + 1, src0_dims.end()});
        src1_md = src1_md.reshape({src1_dims.begin() + 1, src1_dims.end()});
        use_broadcasting_hack = true;
      }
    }

    auto dst_md =
        memory::desc(dst_tz, OneDNNGetDataType<T>(), OneDNNMemoryFormat::any);

1043
    if (x->numel() < y->numel()) {
1044 1045 1046 1047
      if (algo == dnnl::algorithm::binary_sub) {
        attributes = CreateAttributes(
            algo, -1.0 * scale_x, -1.0 * scale_y, scale_out, post_ops);
      }
1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112
      this->AcquireForwardPrimitiveDescriptor(
          attributes, algo, src1_md, src0_md, dst_md);
    } else {
      this->AcquireForwardPrimitiveDescriptor(
          attributes, algo, src0_md, src1_md, dst_md);
    }
  }
  std::shared_ptr<dnnl::memory> AcquireSecondSrcMemory(
      const DenseTensor* input) {
    const T* input_data = input->data<T>();
    return this->AcquireMemoryFromPrimitive(this->fwd_pd_->src1_desc(),
                                            to_void_cast<T>(input_data));
  }

 private:
  static inline dnnl::primitive_attr CreateAttributes(
      dnnl::algorithm op,
      float scale_x,
      float scale_y,
      float scale_out,
      dnnl::post_ops post_ops = dnnl::post_ops{}) {
    // Scales set in attributes for inputs contibute to the output equation
    // in the following way (assuming no broadcasting takes place):
    // output_i = scale_0 * x_i <+ or *> scale_1 * y_i;
    // Hence we have to create scales that will:
    // 1. Dequantize both values, by multiplying with (1.0 / scale_x_or_y)
    // 2. Quantize their result to output scale range, by multiplying with
    // (scale_z)
    // If we combine these two, we end up with following equation
    // output = scale_out * (1/scale_x * x <* or +> 1/scale_y * y)
    // Hence, to mimic such behaviour using provided interface,
    // For add operation the equation is equal to:
    // output = (scale_out / scale_x) * x + (scale_out / scale_y) * y
    //                <scale_0>                  <scale_1>
    // For mul operation on the other hand
    // output = (scale_out / scale_x) * x * (1.0 / scale_y) * y
    //                <scale_0>                 <scale_1>
    float scale_0 = scale_out / scale_x;
    float scale_1 =
        op == dnnl::algorithm::binary_add ? scale_out / scale_y : 1.0 / scale_y;
    dnnl::primitive_attr attributes;
    attributes.set_scales(
        /* input_x_id = */ DNNL_ARG_SRC_0, /* mask = */ 0, {scale_0});
    attributes.set_scales(
        /* input_y_id = */ DNNL_ARG_SRC_1, /* mask = */ 0, {scale_1});
    if (post_ops.len() > 0) attributes.set_post_ops(post_ops);
    return attributes;
  }
};

template <typename T>
class BroadcastDataOneDNNHandler
    : public OneDNNHandlerNoCachingT<T, dnnl::binary> {
 public:
  BroadcastDataOneDNNHandler(const dnnl::algorithm algo,
                             const dnnl::engine engine,
                             Place cpu_place,
                             const DenseTensor* x,
                             DenseTensor* out,
                             float scale_x,
                             float scale_y,
                             const std::vector<int64_t>& extended_x_dims)
      : OneDNNHandlerNoCachingT<T, dnnl::binary>(engine, cpu_place) {
    const auto src0_tz = vectorize(out->dims());
    const auto src0_md = dnnl::memory::desc(
1113
        src0_tz, OneDNNGetDataType<T>(), GetPlainOneDNNFormat(src0_tz.size()));
1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132
    const auto src1_md = x->mem_desc().reshape(extended_x_dims);

    dnnl::primitive_attr attributes;
    attributes.set_scales(DNNL_ARG_SRC_0, 0, {scale_x});
    attributes.set_scales(DNNL_ARG_SRC_1, 0, {scale_y});

    this->AcquireForwardPrimitiveDescriptor(
        attributes, algo, src0_md, src1_md, src0_md);
  }

  template <typename T_out = T>
  std::shared_ptr<dnnl::memory> AcquireZeroedDstMemory(DenseTensor* out) {
    T_out* ptr = out->mutable_data<T_out>(this->place_,
                                          this->fwd_pd_->dst_desc().get_size());
    memset(ptr, 0, this->fwd_pd_->dst_desc().get_size());
    return this->AcquireMemoryFromPrimitive(this->fwd_pd_->dst_desc(), ptr);
  }
};

S
Sylwester Fraczek 已提交
1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147
template <typename T>
class PReluOneDNNHandler
    : public OneDNNHandlerNoCachingT<T,
                                     dnnl::prelu_forward,
                                     dnnl::prelu_backward> {
 public:
  PReluOneDNNHandler(const dnnl::engine engine,
                     Place cpu_place,
                     const DenseTensor& x,
                     const DenseTensor& weights,
                     const std::string& mode,
                     const std::string& data_format,
                     const bool is_test)
      : OneDNNHandlerNoCachingT<T, dnnl::prelu_forward, dnnl::prelu_backward>(
            engine, cpu_place) {
1148
    auto weights_dims = vectorize(weights.dims());
S
Sylwester Fraczek 已提交
1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193
    // weights must have same size as X only for "element" case
    if (weights.dims().size() != x.dims().size()) {
      auto new_weights_dims = std::vector<int64_t>(x.dims().size(), 1);
      if (mode == "channel") {
        new_weights_dims[1] =
            *std::max_element(weights_dims.begin(), weights_dims.end());
      }
      weights_dims = std::move(new_weights_dims);
    }
    auto weights_md = memory::desc(
        weights_dims, OneDNNGetDataType<T>(), memory::format_tag::any);

    this->AcquireForwardPrimitiveDescriptor(
        dnnl::prop_kind::forward_training, x.mem_desc(), weights_md);
    if (!is_test) {
      this->AcquireBackwardPrimitiveDescriptor(
          x.mem_desc(), weights_md, x.mem_desc(), weights_md);
    }
  }

  std::shared_ptr<memory> AcquireWeightsMemoryPossiblyWithReorder(
      const DenseTensor* weights, const bool is_test) {
    const T* weights_data = weights->data<T>();

    // if weights are 1D, every format tag is correct, so we accept
    // format_tag::any's output and no reorder is needed
    if (weights->dims().size() == 1) {
      return this->AcquireMemoryFromPrimitive(this->fwd_pd_->weights_desc(),
                                              to_void_cast<T>(weights_data));
    }

    return this->AcquireMemoryWithReorder(weights->mem_desc(),
                                          this->fwd_pd_->weights_desc(),
                                          to_void_cast<T>(weights_data),
                                          is_test);
  }

  std::shared_ptr<memory> AcquireDiffWeightsMemory(DenseTensor* output) {
    T* output_data = output->mutable_data<T>(
        this->place_, this->bwd_pd_->diff_weights_desc().get_size());
    return this->AcquireMemoryFromPrimitive(this->bwd_pd_->diff_weights_desc(),
                                            output_data);
  }
};

1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208
template <typename T>
class ReductionOneDNNHandler
    : public OneDNNHandlerNoCachingT<T, dnnl::reduction> {
 public:
  ReductionOneDNNHandler(const dnnl::algorithm algo,
                         const float p,
                         const float eps,
                         const dnnl::engine engine,
                         Place cpu_place,
                         const DenseTensor* x,
                         const DenseTensor* out,
                         std::vector<int64_t> out_tz,
                         const dnnl::primitive_attr& attrs = NULL)
      : OneDNNHandlerNoCachingT<T, dnnl::reduction>(engine, cpu_place) {
    const auto out_md = memory::desc(
1209
        out_tz, OneDNNGetDataType<T>(), dnnl::memory::format_tag::any);
1210 1211 1212 1213 1214 1215 1216 1217 1218

    if (attrs)
      this->AcquireForwardPrimitiveDescriptor(
          attrs, algo, x->mem_desc(), out_md, p, eps);
    else
      this->AcquireForwardPrimitiveDescriptor(
          algo, x->mem_desc(), out_md, p, eps);
  }
};
1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273

template <typename T>
class ClipOneDNNHandler
    : public OneDNNHandlerNoCachingT<T,
                                     dnnl::eltwise_forward,
                                     dnnl::eltwise_backward> {
 public:
  ClipOneDNNHandler(const Scalar& min,
                    const Scalar& max,
                    const dnnl::engine engine,
                    Place cpu_place,
                    const DenseTensor* x)
      : OneDNNHandlerNoCachingT<T,
                                dnnl::eltwise_forward,
                                dnnl::eltwise_backward>(engine, cpu_place) {
    float alpha = min.to<float>();
    float beta = max.to<float>();

    this->AcquireForwardPrimitiveDescriptor(dnnl::prop_kind::forward_training,
                                            dnnl::algorithm::eltwise_clip_v2,
                                            x->mem_desc(),
                                            alpha,
                                            beta);
  }

  ClipOneDNNHandler(const Scalar& min,
                    const Scalar& max,
                    const dnnl::engine engine,
                    Place cpu_place,
                    const DenseTensor* x,
                    const DenseTensor* dout)
      : OneDNNHandlerNoCachingT<T,
                                dnnl::eltwise_forward,
                                dnnl::eltwise_backward>(engine, cpu_place) {
    float alpha = min.to<float>();
    float beta = max.to<float>();

    this->AcquireForwardPrimitiveDescriptor(dnnl::prop_kind::forward_training,
                                            dnnl::algorithm::eltwise_clip_v2,
                                            x->mem_desc(),
                                            alpha,
                                            beta);
    this->AcquireBackwardPrimitiveDescriptor(dnnl::algorithm::eltwise_clip_v2,
                                             dout->mem_desc(),
                                             x->mem_desc(),
                                             alpha,
                                             beta);
  }
  std::shared_ptr<dnnl::memory> AcquireBackwardSrcMemory(
      const DenseTensor* input) {
    const T* input_data = input->data<T>();
    return this->AcquireMemoryFromPrimitive(this->bwd_pd_->src_desc(),
                                            to_void_cast<T>(input_data));
  }
};
1274

1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306
template <typename T>
class BatchNormOneDNNHandler
    : public OneDNNHandlerNoCachingT<T,
                                     dnnl::batch_normalization_forward,
                                     dnnl::batch_normalization_backward> {
 public:
  BatchNormOneDNNHandler(const dnnl::engine engine,
                         Place cpu_place,
                         const DenseTensor* x,
                         const float epsilon,
                         const bool fuse_with_relu,
                         const bool global_stats,
                         const bool test_mode)
      : OneDNNHandlerNoCachingT<T,
                                dnnl::batch_normalization_forward,
                                dnnl::batch_normalization_backward>(engine,
                                                                    cpu_place) {
    // Flags are added by bitwise OR operation
    auto flags = dnnl::normalization_flags::use_scale_shift;  // 001
    if (global_stats)
      flags |= dnnl::normalization_flags::use_global_stats;  // 010
    if (fuse_with_relu && test_mode)
      flags |= dnnl::normalization_flags::fuse_norm_relu;  // 100

    this->AcquireForwardPrimitiveDescriptor(
        global_stats ? dnnl::prop_kind::forward_scoring
                     : dnnl::prop_kind::forward_training,
        x->mem_desc(),
        epsilon,
        flags);
  }

1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337
  BatchNormOneDNNHandler(const dnnl::engine engine,
                         Place cpu_place,
                         const float epsilon,
                         const DenseTensor* in_x,
                         const DenseTensor* scale,
                         const DenseTensor* out_grad)
      : OneDNNHandlerNoCachingT<T,
                                dnnl::batch_normalization_forward,
                                dnnl::batch_normalization_backward>(engine,
                                                                    cpu_place) {
    auto scale_tz = vectorize<int64_t>(scale->dims());
    PADDLE_ENFORCE_EQ(
        scale_tz.size(),
        1,
        errors::InvalidArgument(
            "Dims of scale tensor must be 1, but received scale's size is %d",
            scale_tz.size()));

    this->AcquireForwardPrimitiveDescriptor(
        dnnl::prop_kind::forward_training,
        in_x->mem_desc(),
        epsilon,
        dnnl::normalization_flags::use_scale_shift);
    this->AcquireBackwardPrimitiveDescriptor(
        dnnl::prop_kind::backward,
        out_grad->mem_desc(),
        in_x->mem_desc(),
        epsilon,
        dnnl::normalization_flags::use_scale_shift);
  }

1338 1339
  std::shared_ptr<dnnl::memory> AcquireScaleShiftMemory(
      const DenseTensor* scale, const DenseTensor* shift) {
1340
    auto scale_tz = vectorize(scale->dims());
1341 1342 1343 1344
    const unsigned int C = scale_tz[0];
    PADDLE_ENFORCE_EQ(
        scale_tz.size(),
        1,
1345
        errors::InvalidArgument(
1346 1347 1348 1349 1350 1351
            "Dims of scale tensor must be 1, but received scale's size is %d",
            scale_tz.size()));

    auto scaleshift_memory =
        this->AcquireMemoryFromPrimitive(this->fwd_pd_->weights_desc());

1352
    // oneDNN requires a single piece of memory for scale and shift/bias data
1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365
    auto mem_data_handle =
        reinterpret_cast<T*>(scaleshift_memory->get_data_handle());
    std::copy(scale->data<T>(), scale->data<T>() + C, mem_data_handle);
    std::copy(shift->data<T>(), shift->data<T>() + C, mem_data_handle + C);
    return scaleshift_memory;
  }

  std::shared_ptr<dnnl::memory> AcquireDiffScaleShiftMemory(
      T* diff_scaleshift_data) {
    return this->AcquireMemoryFromPrimitive(this->bwd_pd_->diff_weights_desc(),
                                            diff_scaleshift_data);
  }

1366
  std::shared_ptr<dnnl::memory> AcquireMeanMemory(const DenseTensor* mean) {
1367 1368 1369 1370 1371
    const T* mean_data = mean->data<T>();
    return this->AcquireMemoryFromPrimitive(this->fwd_pd_->mean_desc(),
                                            to_void_cast<T>(mean_data));
  }

1372
  std::shared_ptr<dnnl::memory> AcquireMeanMemory(DenseTensor* mean) {
1373 1374 1375 1376 1377 1378 1379
    T* mean_data = mean->mutable_data<T>(this->place_,
                                         this->fwd_pd_->mean_desc().get_size());
    return this->AcquireMemoryFromPrimitive(this->fwd_pd_->mean_desc(),
                                            mean_data);
  }

  std::shared_ptr<dnnl::memory> AcquireVarianceMemory(
1380
      const DenseTensor* variance) {
1381 1382 1383 1384 1385
    const T* variance_data = variance->data<T>();
    return this->AcquireMemoryFromPrimitive(this->fwd_pd_->variance_desc(),
                                            to_void_cast<T>(variance_data));
  }

1386
  std::shared_ptr<dnnl::memory> AcquireVarianceMemory(DenseTensor* variance) {
1387 1388 1389 1390 1391 1392 1393
    T* variance_data = variance->mutable_data<T>(
        this->place_, this->fwd_pd_->variance_desc().get_size());
    return this->AcquireMemoryFromPrimitive(this->fwd_pd_->variance_desc(),
                                            variance_data);
  }
};

1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662
template <typename T>
class PoolingOneDNNHandler
    : public OneDNNHandlerNoCachingT<T,
                                     dnnl::pooling_forward,
                                     dnnl::pooling_backward> {
 public:
  PoolingOneDNNHandler(const OneDNNContext& dev_ctx,
                       const std::string& pooling_type,
                       const IntArray& kernel_size,
                       const std::vector<int>& strides,
                       const std::vector<int>& paddings,
                       bool global_pooling,
                       const std::string& padding_algorithm,
                       bool ceil_mode,
                       bool exclusive,
                       bool adaptive,
                       const DenseTensor* input,
                       DenseTensor* output)
      : OneDNNHandlerNoCachingT<T,
                                dnnl::pooling_forward,
                                dnnl::pooling_backward>(dev_ctx.GetEngine(),
                                                        dev_ctx.GetPlace()) {
    std::vector<int64_t> copied_kernel_size(kernel_size.GetData().begin(),
                                            kernel_size.GetData().end());
    std::vector<int64_t> copied_strides(strides.begin(), strides.end());
    std::vector<int64_t> copied_paddings(paddings.begin(), paddings.end());
    // Only 2D pooling is supported now
    PADDLE_ENFORCE_EQ(
        copied_kernel_size.size(),
        2,
        errors::InvalidArgument("The copied_kernel_size must be 2D, i.e. 2D "
                                "pooling, but received %dD.",
                                copied_kernel_size.size()));
    PADDLE_ENFORCE_EQ(
        pooling_type == "max" || pooling_type == "avg",
        true,
        errors::InvalidArgument(
            "The pooling_type must be 'max' or 'avg', but received %s.",
            pooling_type));
    PADDLE_ENFORCE_EQ(
        input->dims().size(),
        4,
        errors::InvalidArgument(
            "Input dim must be with 4, i.e. NCHW, but received %d.",
            input->dims().size()));

    const auto input_dims = input->dims();
    DDim data_dims = slice_ddim(input_dims, 2, input_dims.size());

    if (global_pooling) {
      UpdateKernelSize<int64_t>(&copied_kernel_size, data_dims);
    }

    UpdatePadding<int64_t>(&copied_paddings,
                           global_pooling,
                           0,
                           padding_algorithm,
                           data_dims,
                           copied_strides,
                           copied_kernel_size);

    auto onednn_paddings = ToOneDNNPadding(copied_paddings);

    const auto dt = ToOneDNNDataType(input->dtype());
    const auto src_tz = vectorize(input->dims());
    const auto dst_tz = vectorize(output->dims());
    const auto dst_md = OneDNNMemDesc(dst_tz, dt, OneDNNMemoryFormat::any);

    if (ceil_mode) {
      CorrectOutputSize(src_tz,
                        dst_tz,
                        copied_kernel_size,
                        copied_paddings,
                        copied_strides,
                        onednn_paddings[1]);
    }

    if (adaptive) {
      ComputeAdaptivePoolParameters(
          src_tz, &copied_kernel_size, &copied_strides);
    }

    bool is_test = dev_ctx.HasDnnAttr("is_test")
                       ? PADDLE_GET_CONST(bool, dev_ctx.GetDnnAttr("is_test"))
                       : false;

    this->AcquireForwardPrimitiveDescriptor(
        is_test ? dnnl::prop_kind::forward_inference
                : dnnl::prop_kind::forward_training,
        pooling_type == "max"
            ? dnnl::algorithm::pooling_max
            : (exclusive ? dnnl::algorithm::pooling_avg_exclude_padding
                         : dnnl::algorithm::pooling_avg_include_padding),
        input->mem_desc(),
        dst_md,
        copied_strides,
        copied_kernel_size,
        onednn_paddings[0],
        onednn_paddings[1]);
  }

  PoolingOneDNNHandler(const OneDNNContext& dev_ctx,
                       const std::string& pooling_type,
                       const IntArray& kernel_size,
                       const std::vector<int>& strides,
                       const std::vector<int>& paddings,
                       bool global_pooling,
                       const std::string& padding_algorithm,
                       bool ceil_mode,
                       bool exclusive,
                       bool adaptive,
                       const DenseTensor* in_x,
                       const DenseTensor* out_grad,
                       DenseTensor* in_x_grad)

      : OneDNNHandlerNoCachingT<T,
                                dnnl::pooling_forward,
                                dnnl::pooling_backward>(dev_ctx.GetEngine(),
                                                        dev_ctx.GetPlace()) {
    bool is_test = dev_ctx.HasDnnAttr("is_test")
                       ? PADDLE_GET_CONST(bool, dev_ctx.GetDnnAttr("is_test"))
                       : false;

    PADDLE_ENFORCE_EQ(
        is_test,
        false,
        errors::InvalidArgument(
            "is_test attribute should be set to False in training phase."));

    std::vector<int64_t> copied_kernel_size(kernel_size.GetData().begin(),
                                            kernel_size.GetData().end());
    std::vector<int64_t> copied_strides(strides.begin(), strides.end());
    std::vector<int64_t> copied_paddings(paddings.begin(), paddings.end());
    auto in_x_dims = in_x->dims();
    DDim data_dims = slice_ddim(in_x_dims, 2, in_x_dims.size());
    if (global_pooling) {
      UpdateKernelSize<int64_t>(&copied_kernel_size, data_dims);
    }

    UpdatePadding<int64_t>(&copied_paddings,
                           global_pooling,
                           0,
                           padding_algorithm,
                           data_dims,
                           copied_strides,
                           copied_kernel_size);

    auto src_tz = vectorize<int64_t>(in_x->dims());
    auto diff_src_tz = vectorize<int64_t>(in_x_grad->dims());
    auto diff_dst_tz = vectorize<int64_t>(out_grad->dims());

    const auto dt = ToOneDNNDataType(in_x->dtype());
    auto dst_md = dnnl::memory::desc(diff_dst_tz, dt, OneDNNMemoryFormat::any);
    auto diff_src_md = dnnl::memory::desc(
        diff_src_tz, OneDNNGetDataType<T>(), OneDNNMemoryFormat::any);

    auto onednn_paddings = ToOneDNNPadding(copied_paddings);

    if (ceil_mode) {
      CorrectOutputSize(src_tz,
                        diff_dst_tz,
                        copied_kernel_size,
                        copied_paddings,
                        copied_strides,
                        onednn_paddings[1]);
    }

    if (adaptive) {
      ComputeAdaptivePoolParameters(
          diff_src_tz, &copied_kernel_size, &copied_strides);
    }

    this->AcquireForwardPrimitiveDescriptor(
        dnnl::prop_kind::forward_training,
        pooling_type == "max"
            ? dnnl::algorithm::pooling_max
            : (exclusive ? dnnl::algorithm::pooling_avg_exclude_padding
                         : dnnl::algorithm::pooling_avg_include_padding),
        in_x->mem_desc(),
        dst_md,
        copied_strides,
        copied_kernel_size,
        onednn_paddings[0],
        onednn_paddings[1]);

    this->AcquireBackwardPrimitiveDescriptor(
        pooling_type == "max"
            ? dnnl::algorithm::pooling_max
            : (exclusive ? dnnl::algorithm::pooling_avg_exclude_padding
                         : dnnl::algorithm::pooling_avg_include_padding),
        diff_src_md,
        out_grad->mem_desc(),
        copied_strides,
        copied_kernel_size,
        onednn_paddings[0],
        onednn_paddings[1]);
  }

  std::shared_ptr<dnnl::memory> AcquireWorkspaceMemory(
      const OneDNNContext& dev_ctx, const std::string& unique_name) {
    dnnl::memory::desc workspace_md = this->fwd_pd_->workspace_desc();
    // Pooling Workspace has to be passed to Grad op that
    // may be executed by diffrent thread, hence
    // for that one we use key that does not contain TID
    std::string workspace_key = CreateKey(dev_ctx,
                                          workspace_md.dims(),
                                          workspace_md.data_type(),
                                          unique_name,
                                          "@wrk");
    auto mem_p =
        std::static_pointer_cast<dnnl::memory>(dev_ctx.GetBlob(workspace_key));
    if (mem_p == nullptr) {
      static std::mutex acquire_barrier;
      std::lock_guard<std::mutex> block_threads_until_finish_this_job(
          acquire_barrier);
      mem_p = std::static_pointer_cast<dnnl::memory>(
          dev_ctx.GetBlob(workspace_key));
      if (mem_p == nullptr) {
        mem_p = std::make_shared<dnnl::memory>(workspace_md, this->engine_);
        dev_ctx.SetBlob(workspace_key, mem_p);
      }
    }
    return mem_p;
  }

  static void ComputeAdaptivePoolParameters(const std::vector<int64_t>& src_tz,
                                            std::vector<int64_t>* kernel_size,
                                            std::vector<int64_t>* strides) {
    // https://github.com/oneapi-src/oneDNN/tree/bkocot/adaptive-pooling/rfcs/20200818-adaptive-pooling
    auto IH = static_cast<double>(src_tz[src_tz.size() - 2]);
    auto IW = static_cast<double>(src_tz[src_tz.size() - 1]);
    auto OH = static_cast<double>(kernel_size->at(0));
    auto OW = static_cast<double>(kernel_size->at(1));

    strides->at(0) =
        static_cast<int64_t>(floor((IH * 2.0) / OH) - floor(IH / OH));
    strides->at(1) =
        static_cast<int64_t>(floor((IW * 2.0) / OW) - floor(IW / OW));
    kernel_size->at(0) =
        static_cast<int64_t>(ceil((IH * 2.0) / OH) - floor(IH / OH));
    kernel_size->at(1) =
        static_cast<int64_t>(ceil((IW * 2.0) / OW) - floor(IW / OW));
  }

 private:
  static inline int ComputeCeiledOutput(int input_size,
                                        int kernel_size,
                                        int padding,
                                        int stride) {
    return (input_size - kernel_size + 2 * padding) / stride + 1;
  }

  static inline void CorrectOutputSize(
      const std::vector<int64_t>& src_tz,
      const std::vector<int64_t>& dst_tz,
      const std::vector<int64_t>& kernel_size,
      const std::vector<int64_t>& paddings,
      const std::vector<int64_t>& strides,
      std::vector<int64_t>& right_bot_padding) {  // NOLINT
    for (size_t i = 0; i < right_bot_padding.size(); i++) {
      int desired_size = ComputeCeiledOutput(
          src_tz[i + 2], kernel_size[i], paddings[i], strides[i]);
      if (desired_size != dst_tz[i + 2]) {
        right_bot_padding[i] += strides[i] - 1;
      }
    }
  }
};

1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725
static DDim RowMatrixDimsFromVector(const DDim& x_dim) {
  return x_dim.size() > 1 ? x_dim : make_ddim({1, x_dim[0]});
}

static DDim ColumnMatrixDimsFromVector(const DDim& y_dim) {
  return y_dim.size() > 1 ? y_dim : make_ddim({y_dim[0], 1});
}

static std::vector<int64_t> TransposeAxis(const std::vector<int64_t>& x,
                                          const std::vector<int>& axis) {
  size_t in_rank = x.size();
  size_t axis_size = axis.size();

  auto axis_set = std::set<int>(axis.begin(), axis.end());
  PADDLE_ENFORCE_EQ(axis_set.size(),
                    axis_size,
                    paddle::platform::errors::InvalidArgument(
                        "In an axis array, elements must be unique."));

  PADDLE_ENFORCE_EQ(in_rank,
                    axis_size,
                    paddle::platform::errors::InvalidArgument(
                        "The input dimension's size "
                        "should be equal to the axis's size. "
                        "But received dimension is %d, "
                        "axis's size is %d",
                        in_rank,
                        axis_size));

  PADDLE_ENFORCE_LT(*std::max_element(axis.begin(), axis.end()),
                    axis_size,
                    paddle::platform::errors::InvalidArgument(
                        "Axis values must be ranging from 0 to (dims - 1)."));

  std::vector<int64_t> new_x(x.size());
  for (size_t i = 0; i < x.size(); i++) {
    new_x[i] = x[axis[i]];
  }
  return new_x;
}

static std::vector<int64_t> GetInputStrides(const OneDNNContext& dev_ctx,
                                            const DDim& input_dims,
                                            const std::string input_name,
                                            const bool transpose_input) {
  auto new_dims = input_dims;
  auto shape =
      dev_ctx.HasDnnAttr("fused_reshape_" + input_name)
          ? PADDLE_GET_CONST(std::vector<int>,
                             dev_ctx.GetDnnAttr("fused_reshape_" + input_name))
          : std::vector<int>();
  auto axis = dev_ctx.HasDnnAttr("fused_transpose_" + input_name)
                  ? PADDLE_GET_CONST(
                        std::vector<int>,
                        dev_ctx.GetDnnAttr("fused_transpose_" + input_name))
                  : std::vector<int>();

  if (!shape.empty() && !axis.empty()) {
    new_dims = input_dims.reshape(shape).transpose(axis);
  }

  auto& MatrixDimsFromVector =
      input_name == "X" ? RowMatrixDimsFromVector : ColumnMatrixDimsFromVector;
1726
  MatDescriptor mat_dim = CreateMatrixDescriptor(
1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761
      MatrixDimsFromVector(new_dims), 0, transpose_input);

  std::vector<int64_t> strides;
  if (!shape.empty()) {
    auto shape2 = input_dims.reshape(shape);
    strides.push_back(1);
    for (auto i = shape2.size() - 1; i > 0; --i) {
      strides.insert(strides.begin(),
                     strides.front() * static_cast<int64_t>(shape2[i]));
    }
    strides = TransposeAxis(strides, axis);
    if (shape.size() == 2)
      strides.insert(strides.begin(),
                     static_cast<int64_t>(shape[0] * shape[1]));
    mat_dim.stride_ = strides[0];
    if (mat_dim.trans_) std::swap(*strides.rbegin(), *(++strides.rbegin()));
  }
  return strides;
}

static bool IsOutputFused(const OneDNNContext& dev_ctx) {
  const auto shape =
      dev_ctx.HasDnnAttr("fused_reshape_Out")
          ? PADDLE_GET_CONST(std::vector<int>,
                             dev_ctx.GetDnnAttr("fused_reshape_Out"))
          : std::vector<int>();
  const auto axis =
      dev_ctx.HasDnnAttr("fused_transpose_Out")
          ? PADDLE_GET_CONST(std::vector<int>,
                             dev_ctx.GetDnnAttr("fused_transpose_Out"))
          : std::vector<int>();
  return !shape.empty() && !axis.empty();
}

template <typename XT, typename YT, typename OT>
1762
class MatmulOneDNNHandler : public OneDNNHandlerNoCachingT<XT, dnnl::matmul> {
1763 1764 1765 1766 1767 1768 1769 1770 1771
 public:
  MatmulOneDNNHandler(const OneDNNContext& dev_ctx,
                      const std::vector<int64_t>& x_org_dims,
                      const std::vector<int64_t>& y_org_dims,
                      bool trans_x,
                      bool trans_y,
                      const std::vector<int64_t>& x_strides_override,
                      const std::vector<int64_t>& y_strides_override,
                      bool is_output_fused)
1772 1773
      : OneDNNHandlerNoCachingT<XT, dnnl::matmul>(dev_ctx.GetEngine(),
                                                  dev_ctx.GetPlace()) {
1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944
    // M X K * K X N
    std::vector<int64_t> x_dims(x_org_dims);
    std::vector<int64_t> y_dims(y_org_dims);

    const int MB_idx = x_dims.size() - 3;
    const int H_idx = x_dims.size() - 2;
    const int W_idx = x_dims.size() - 1;

    if (trans_x) std::swap(x_dims[H_idx], x_dims[W_idx]);
    if (trans_y) std::swap(y_dims[H_idx], y_dims[W_idx]);

    const memory::dim M = x_dims[H_idx];
    const memory::dim K = x_dims[W_idx];
    const memory::dim N = y_dims[W_idx];

    std::vector<int64_t> x_strides(x_dims.size() - 3, 1);
    std::vector<int64_t> y_strides(x_dims.size() - 3, 1);
    std::vector<int64_t> out_strides(x_dims.size() - 3, 1);
    std::vector<int64_t> out_ddims(x_dims.size() - 3, 1);

    x_strides.reserve(x_dims.size());
    y_strides.reserve(x_dims.size());
    out_strides.reserve(x_dims.size());

    if (!x_strides_override.empty()) {
      x_strides = x_strides_override;
    } else {
      if (!trans_x) {
        x_strides.insert(x_strides.end(), {M * K, K, 1});
      } else {
        x_strides.insert(x_strides.end(), {M * K, 1, M});
      }
    }

    if (!y_strides_override.empty()) {
      y_strides = y_strides_override;
    } else {
      if (!trans_y) {
        y_strides.insert(y_strides.end(), {N * K, N, 1});
      } else {
        y_strides.insert(y_strides.end(), {N * K, 1, K});
      }
    }

    out_strides.insert(out_strides.end(), {M * N, N, 1});
    out_ddims.insert(out_ddims.end(),
                     {std::max(x_dims[MB_idx], y_dims[MB_idx]), M, N});

    for (int i = x_dims.size() - 4; i >= 0; --i) {
      out_ddims[i] = std::max(x_dims[i], y_dims[i]);
      if (x_strides_override.empty()) {
        x_strides[i] = x_dims[i + 1] * x_strides[i + 1];
      }
      if (y_strides_override.empty()) {
        y_strides[i] = y_dims[i + 1] * y_strides[i + 1];
      }
      out_strides[i] = out_ddims[i + 1] * out_strides[i + 1];
    }

    // TODO(jczaja): Why not for int8??
    if (!is_int8<OT>() && is_output_fused) {
      out_strides = FakeTransposeStrides(out_ddims);
    }

    auto x_md = memory::desc(x_dims, OneDNNGetDataType<XT>(), x_strides);
    auto y_md = memory::desc(y_dims, OneDNNGetDataType<YT>(), y_strides);
    auto out_md = memory::desc(out_ddims, OneDNNGetDataType<OT>(), out_strides);

    const auto matmul_attrs = CreateMatmulAttrs(dev_ctx);

    this->AcquireForwardPrimitiveDescriptor(matmul_attrs, x_md, y_md, out_md);
  }

  float ComputeOutputScale(const OneDNNContext& dev_ctx) {
    float alpha = dev_ctx.HasDnnAttr("alpha")
                      ? PADDLE_GET_CONST(float, dev_ctx.GetDnnAttr("alpha"))
                      : 1.0f;

    if (dev_ctx.HasDnnAttr("Scale_x") && dev_ctx.HasDnnAttr("Scale_y") &&
        dev_ctx.HasDnnAttr("Scale_out")) {
      float scale_x = PADDLE_GET_CONST(float, dev_ctx.GetDnnAttr("Scale_x"));
      float scale_y = PADDLE_GET_CONST(float, dev_ctx.GetDnnAttr("Scale_y"));
      bool force_fp32_out =
          dev_ctx.HasDnnAttr("force_fp32_output")
              ? PADDLE_GET_CONST(bool, dev_ctx.GetDnnAttr("force_fp32_output"))
              : false;
      float scale_out =
          force_fp32_out
              ? 1.f
              : PADDLE_GET_CONST(float, dev_ctx.GetDnnAttr("Scale_out"));
      alpha *= scale_out / (scale_x * scale_y);
    }
    return alpha;
  }

  dnnl::primitive_attr CreateMatmulAttrs(const OneDNNContext& dev_ctx) {
    dnnl::primitive_attr matmul_attrs;
    dnnl::post_ops post_operations;

    float scale_out = ComputeOutputScale(dev_ctx);
    if (scale_out != 1.0f) {
      matmul_attrs.set_output_scales(0, {scale_out});
    }

    if (dev_ctx.HasDnnInput("ResidualData")) {
      auto* residual_data = dev_ctx.GetDnnInput("ResidualData");
      auto residual_data_tz = vectorize(residual_data->dims());
      auto residual_data_md = memory::desc(residual_data_tz,
                                           OneDNNGetDataType<OT>(),
                                           dnnl::memory::format_tag::any);
      post_operations.append_binary(dnnl::algorithm::binary_add,
                                    residual_data_md);
      if (dev_ctx.HasDnnAttr("Scale_in_eltwise")) {
        float scale_in_eltwise =
            PADDLE_GET_CONST(float, dev_ctx.GetDnnAttr("Scale_in_eltwise"));
        float sum_scale = scale_out / scale_in_eltwise;
        post_operations.append_sum(sum_scale);
      }
    }

    AppendActivation(dev_ctx, post_operations);

    if (dev_ctx.HasDnnAttr("fused_output_scale")) {
      float scale_alpha =
          PADDLE_GET_CONST(float, dev_ctx.GetDnnAttr("fused_output_scale"));
      post_operations.append_eltwise(
          1.0, dnnl::algorithm::eltwise_linear, scale_alpha, 0.0f);
    }

    matmul_attrs.set_post_ops(post_operations);
    return matmul_attrs;
  }

  std::vector<int64_t> FakeTransposeStrides(
      const std::vector<int64_t>& matmul_out_dims) const {
    // fuse matmul_v2 + transpose + reshape guarantees that output is 4D and
    // transpose axis are: {0, 2, 1, 3}
    std::vector<int64_t> transpose_axis = {0, 2, 1, 3};
    std::vector<int64_t> fake_strides(transpose_axis.size());
    int ndims = static_cast<int>(transpose_axis.size());

    int total_stride = 1;

    for (int i = ndims - 1; i >= 0; --i) {
      fake_strides[transpose_axis[i]] = total_stride;
      total_stride *= matmul_out_dims[transpose_axis[i]];
    }

    return fake_strides;
  }

  std::shared_ptr<memory> AcquireWeightsMemory(const DenseTensor* input) {
    const YT* input_data = input->data<YT>();
    return this->AcquireMemoryFromPrimitive(this->fwd_pd_->weights_desc(),
                                            to_void_cast<YT>(input_data));
  }

  std::shared_ptr<dnnl::memory> AcquireDstMemory(const OneDNNContext& dev_ctx,
                                                 DenseTensor* output) {
    // We cannot use base AcquireDstMemory as it makes an allocation request
    // base on DST memory primitive size. This is fine in general, but in MatMul
    // we have primitive that covers only one batch of Data and then shift
    // pointer for every new batch. Hence DenseTensor size is bigger that
    // dst memory primitive size. So would we request less memory that is there
    // and it triggers an assertion.  So as there is no 'any' format here we can
    // leave default size of DenseTensor as computed in ComputeInferShape
    OT* ptr = dev_ctx.template Alloc<OT>(output);
    return this->AcquireMemoryFromPrimitive(this->fwd_pd_->dst_desc(), ptr);
  }
};

1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985
template <typename T>
static void ExecuteMul(const OneDNNContext& dev_ctx,
                       const DenseTensor& x,
                       const DenseTensor& y,
                       const std::vector<int64_t>& x_dims,
                       const std::vector<int64_t>& y_dims,
                       bool trans_x,
                       bool trans_y,
                       DenseTensor* out) {
  static const std::vector<int64_t> vec_placeholder;
  MatmulOneDNNHandler<T, T, T> handler(dev_ctx,
                                       x_dims,
                                       y_dims,
                                       trans_x,
                                       trans_y,
                                       vec_placeholder,
                                       vec_placeholder,
                                       false);

  const auto src_memory_p = handler.AcquireSrcMemory(&x);
  const auto weights_memory_p = handler.AcquireWeightsMemory(&y);
  const auto dst_memory_p = handler.AcquireDstMemory(dev_ctx, out);

  auto matmul_p = handler.AcquireForwardPrimitive();

  std::unordered_map<int, dnnl::memory> matmul_args = {
      {DNNL_ARG_SRC, *src_memory_p},
      {DNNL_ARG_WEIGHTS, *weights_memory_p},
      {DNNL_ARG_DST, *dst_memory_p}};

  auto& astream = OneDNNContext::tls().get_stream();
  matmul_p->execute(astream, matmul_args);
  astream.wait();

  // This kernel is flattening dims so then we need to unflattened version
  // that should be set in out reshape require plain layout, but
  // MatmulV2MKLDNNHanlder enforces one so it should work
  out->set_mem_desc(
      dst_memory_p->get_desc().reshape(vectorize<int64_t>(out->dims())));
}

1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043
template <typename T, typename T_out>
void ExecuteMatmul(const OneDNNContext& dev_ctx,
                   const DenseTensor& x,
                   const DenseTensor& y,
                   const std::vector<int64_t>& x_dims,
                   const std::vector<int64_t>& y_dims,
                   bool trans_x,
                   bool trans_y,
                   DenseTensor* out) {
  auto x_strides_override = GetInputStrides(dev_ctx, x.dims(), "X", trans_x);
  auto y_strides_override = GetInputStrides(dev_ctx, y.dims(), "Y", trans_y);
  MatmulOneDNNHandler<T, T, T_out> handler(dev_ctx,
                                           x_dims,
                                           y_dims,
                                           trans_x,
                                           trans_y,
                                           x_strides_override,
                                           y_strides_override,
                                           IsOutputFused(dev_ctx));

  const auto src_memory_p = handler.AcquireSrcMemory(&x);
  const auto weights_memory_p = handler.AcquireWeightsMemory(&y);
  const auto dst_memory_p = handler.AcquireDstMemory(dev_ctx, out);

  auto matmul_p = handler.AcquireForwardPrimitive();

  std::unordered_map<int, memory> matmul_args = {
      {DNNL_ARG_SRC, *src_memory_p},
      {DNNL_ARG_WEIGHTS, *weights_memory_p},
      {DNNL_ARG_DST, *dst_memory_p}};

  if (dev_ctx.HasDnnInput("ResidualData")) {
    auto* residual_data = dev_ctx.GetDnnInput("ResidualData");
    const auto residual_data_memory_p = handler.AcquireSrcMemory(residual_data);
    matmul_args.insert({DNNL_ARG_ATTR_MULTIPLE_POST_OP(0) | DNNL_ARG_SRC_1,
                        *residual_data_memory_p});
  }

  auto& astream = OneDNNContext::tls().get_stream();
  matmul_p->execute(astream, matmul_args);
  astream.wait();

  // TODO(jczaja): Explain why int8 format of dst is ABCD and do not need
  // permute
  if (IsOutputFused(dev_ctx) && !is_int8<T_out>()) {
    const auto axis =
        dev_ctx.HasDnnAttr("fused_transpose_Out")
            ? PADDLE_GET_CONST(std::vector<int>,
                               dev_ctx.GetDnnAttr("fused_transpose_Out"))
            : std::vector<int>();
    auto permuted_md = dst_memory_p->get_desc().permute_axes(axis);
    out->set_mem_desc(permuted_md.reshape(vectorize<int64_t>(out->dims())));
  } else {
    out->set_mem_desc(
        dst_memory_p->get_desc().reshape(vectorize<int64_t>(out->dims())));
  }
}

2044 2045
}  // namespace funcs
}  // namespace phi