lrn_mkldnn_op.cc 6.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.

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. */

#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/operators/lrn_op.h"
17
#include "paddle/fluid/platform/mkldnn_reuse.h"
18 19 20 21 22 23 24 25 26 27 28

namespace paddle {
namespace operators {

using paddle::framework::Tensor;
using paddle::platform::MKLDNNDeviceContext;

template <typename T>
class LRNMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
 public:
  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
M
minqiyang 已提交
29
    const bool is_float_type = std::is_same<T, float>::value;
M
minqiyang 已提交
30
    PADDLE_ENFORCE(is_float_type, "MKLDNN LRN must use float data.");
31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
                   "MKLDNN LRN must use CPUPlace.");

    auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
    const auto& mkldnn_engine = dev_ctx.GetEngine();

    auto x = ctx.Input<Tensor>("X");
    auto out = ctx.Output<Tensor>("Out");
    auto mid = ctx.Output<Tensor>("MidOut");

    auto input_data = x->data<T>();
    auto output_data = out->mutable_data<T>(ctx.GetPlace());
    mid->mutable_data<T>(ctx.GetPlace());

    const int n = ctx.Attr<int>("n");
46 47 48 49 50 51
    // MKL-DNN implements LRN in a caffe way:
    // http://caffe.berkeleyvision.org/tutorial/layers/lrn.html
    // Where sum of squares is divided by size of normalization window
    // this is not the case for PaddlePaddle LRN.
    // Hence we need to compensate for this diffrence by
    // multipliing alpha by size of window(n)
52
    const float alpha = ctx.Attr<float>("alpha") * static_cast<float>(n);
53 54 55 56 57 58
    const float beta = ctx.Attr<float>("beta");
    const float k = ctx.Attr<float>("k");

    auto e_mid = framework::EigenTensor<T, 4>::From(*mid);
    e_mid = e_mid.constant(k);

59 60
    auto dims = paddle::framework::vectorize2int(x->dims());

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
    // Format and dims are assumed to be the same for dst and src
    auto md = paddle::platform::MKLDNNMemDesc(
        dims, platform::MKLDNNGetDataType<T>(), x->format());

    const std::string key = platform::LRNMKLDNNHandler::GetHash(
        dims, n, alpha, beta, k, x->format(), ctx.op().Output("Out"));

    platform::LRNMKLDNNHandler handler(ctx.Attr<bool>("is_test"), dev_ctx,
                                       mkldnn_engine, key);
    auto src_memory =
        handler.AcquireSrcMemory(md, platform::to_void_cast<T>(input_data));

    // TODO(jczaja): Hide getting PD inside of handler for all Acquire API
    handler.AcquireLRNPrimitiveDescriptor(md, n, alpha, beta, k);

    auto dst_memory =
        handler.AcquireDstMemory(md, platform::to_void_cast<T>(output_data));

    auto lrn_p = handler.AcquireLRN(dst_memory, src_memory);

    std::vector<mkldnn::primitive> pipeline = {*lrn_p};
    mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();

    auto output_format =
        (mkldnn::memory::format)dst_memory->get_primitive_desc()
            .desc()
            .data.format;

    out->set_layout(framework::DataLayout::kMKLDNN);
    out->set_format(output_format);
91 92 93 94 95 96 97
  }
};

template <typename T>
class LRNMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
 public:
  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
M
minqiyang 已提交
98 99
    const bool is_float_type = std::is_same<T, float>::value;
    PADDLE_ENFORCE(is_float_type, "MKLDNN LRN must use float data.");
100 101
    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
                   "MKLDNN LRN must use CPUPlace.");
102 103 104
    PADDLE_ENFORCE(
        !ctx.Attr<bool>("is_test"),
        "is_test attribute should be set to False in training phase.");
105 106 107 108 109 110 111

    auto x = ctx.Input<Tensor>("X");

    auto out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
    auto x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));

    const int n = ctx.Attr<int>("n");
112
    const float alpha = ctx.Attr<float>("alpha") * static_cast<float>(n);
113 114 115 116 117 118 119 120 121 122 123
    const float beta = ctx.Attr<float>("beta");
    const float k = ctx.Attr<float>("k");

    auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
    const auto& mkldnn_engine = dev_ctx.GetEngine();

    auto x_grad_data = x_grad->mutable_data<T>(ctx.GetPlace());
    auto out_grad_data = out_grad->data<T>();

    auto dims = paddle::framework::vectorize2int(x->dims());

124 125
    const std::string key = platform::LRNMKLDNNHandler::GetHash(
        dims, n, alpha, beta, k, x->format(), ctx.op().Input("Out"));
126

127
    platform::LRNMKLDNNHandler handler(false, dev_ctx, mkldnn_engine, key);
128

129 130
    auto src_md = paddle::platform::MKLDNNMemDesc(
        dims, platform::MKLDNNGetDataType<T>(), x->format());
131

132 133 134
    // diff_dst and diff_src layouts are assumed to be the same
    auto diff_md = paddle::platform::MKLDNNMemDesc(
        dims, platform::MKLDNNGetDataType<T>(), out_grad->format());
135

136
    auto workspace = handler.AcquireWorkspaceMemory();
137

138 139
    auto diff_dst_memory = handler.AcquireDiffDstMemory(
        diff_md, platform::to_void_cast<T>(out_grad_data));
140

141 142
    auto diff_src_memory = handler.AcquireDiffSrcMemory(
        diff_md, platform::to_void_cast<T>(x_grad_data));
143

144 145
    auto src_memory = handler.AcquireSrcMemory(
        src_md, platform::to_void_cast<T>(x->data<T>()));
146

147 148 149
    // TODO(jczaja): Hide this call inside Handler
    handler.AcquireLRNBackwardPrimitiveDescriptor(src_md, diff_md, n, alpha,
                                                  beta, k);
150

151 152
    auto lrn_bwd = handler.AcquireLRNBackward(src_memory, diff_dst_memory,
                                              workspace, diff_src_memory);
153

154
    std::vector<mkldnn::primitive> pipeline = {*lrn_bwd};
155
    mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
156 157 158 159 160 161 162 163

    auto output_format =
        (mkldnn::memory::format)diff_src_memory->get_primitive_desc()
            .desc()
            .data.format;

    x_grad->set_layout(framework::DataLayout::kMKLDNN);
    x_grad->set_format(output_format);
164 165 166 167 168 169 170 171 172 173 174
  }
};
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;

REGISTER_OP_KERNEL(lrn, MKLDNN, paddle::platform::CPUPlace,
                   ops::LRNMKLDNNOpKernel<float>);
REGISTER_OP_KERNEL(lrn_grad, MKLDNN, paddle::platform::CPUPlace,
                   ops::LRNMKLDNNGradOpKernel<float>);
反馈
建议
客服 返回
顶部