elementwise_mkldnn_op.h 3.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
// Copyright (c) 2020 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
16
#include <string>
17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49
#include <unordered_map>
#include "paddle/fluid/operators/elementwise/elementwise_add_op.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"

#include "paddle/fluid/framework/data_layout_transform.h"
#include "paddle/fluid/platform/mkldnn_reuse.h"

namespace paddle {
namespace operators {

using framework::DataLayout;
using framework::Tensor;
using mkldnn::memory;
using mkldnn::primitive;
using mkldnn::stream;

template <typename T, dnnl::algorithm BINARY_OP>
class EltwiseMKLDNNKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    const auto& dev_ctx =
        ctx.template device_context<paddle::platform::MKLDNNDeviceContext>();
    const auto& mkldnn_engine = dev_ctx.GetEngine();

    const auto* x = ctx.Input<Tensor>("X");
    const auto* y = ctx.Input<Tensor>("Y");
    auto* z = ctx.Output<Tensor>("Out");

    float scale_x = ctx.Attr<float>("Scale_x");
    float scale_y = ctx.Attr<float>("Scale_y");
    float scale_o = ctx.Attr<float>("Scale_out");
    int axis = ctx.Attr<int>("axis");

50 51 52
    platform::BinaryMKLDNNHandler<T> handler(BINARY_OP, axis, mkldnn_engine,
                                             ctx.GetPlace(), x, y, z, scale_x,
                                             scale_y, scale_o);
53 54 55

    const auto src_x_memory = handler.AcquireSrcMemory(x);
    const auto src_y_memory = handler.AcquireSecondSrcMemory(y);
56 57 58 59 60 61 62 63 64 65 66 67
    // (jczaja) For Inplace src and dst should be the same memory object.
    // So x should share buffer with z. But UT mechanics is testing inplace
    // execution for this op not checking that x can be bradcasted to match in
    // shape y tensor.
    // This is wrong as when x is to be broadcasted then z(out) will match the
    // shape of y which is bigger than x. Hence if x is smaller in shape than z
    // and they share a buffer (of
    // shape x) then this buffer is not big enough to hold result of elementwise
    // operation.
    auto dst_memory = (x->numel() == z->numel() && x->IsSharedBufferWith(*z))
                          ? src_x_memory
                          : handler.AcquireDstMemory(z);
68 69 70

    const auto binary_prim = handler.AcquireForwardPrimitive();

71
    auto& astream = platform::MKLDNNDeviceContext::tls().get_stream();
72 73 74 75 76 77 78 79 80 81 82 83 84

    const std::unordered_map<int, dnnl::memory> args = {
        {DNNL_ARG_SRC_0, *src_x_memory},
        {DNNL_ARG_SRC_1, *src_y_memory},
        {DNNL_ARG_DST, *dst_memory}};

    binary_prim->execute(astream, args);
    astream.wait();

    z->set_layout(DataLayout::kMKLDNN);
    z->set_format(platform::GetMKLDNNFormat(*dst_memory));
  }
};
85 86 87 88 89 90 91 92 93 94 95 96 97 98 99

inline std::vector<int64_t> CalculateBroadcastedDims(const Tensor* x,
                                                     const Tensor* y) {
  const auto src_tz = framework::vectorize(x->dims());
  const auto dst_tz = framework::vectorize(y->dims());

  size_t j = 0;
  std::vector<int64_t> dst_tz_ex(src_tz.size(), 1);
  for (size_t i = 0; i < src_tz.size(); ++i) {
    dst_tz_ex[i] = (src_tz[i] != dst_tz[j]) ? 1 : dst_tz[j++];
    if (j == dst_tz.size()) break;
  }

  return dst_tz_ex;
}
100 101
}  // namespace operators
}  // namespace paddle