提交 d4d4580d 编写于 作者: T Tao Luo 提交者: GitHub

Merge pull request #4140 from tensor-tang/mkldnn_pool

Add MKLDNN pool
/* Copyright (c) 2017 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 "MKLDNNPoolLayer.h"
#include "paddle/math/MathUtils.h"
#include "paddle/utils/Logging.h"
using namespace mkldnn; // NOLINT
typedef memory::format format;
namespace paddle {
REGISTER_LAYER(mkldnn_pool, MKLDNNPoolLayer);
bool MKLDNNPoolLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
if (!MKLDNNLayer::init(layerMap, parameterMap)) {
return false;
}
/* the size of inputs for pool-layer is 1 */
CHECK_EQ(config_.inputs_size(), 1);
const PoolConfig& conf = config_.inputs(0).pool_conf();
ic_ = conf.channels();
ih_ = conf.img_size_y();
iw_ = conf.img_size();
oc_ = ic_;
oh_ = conf.output_y();
ow_ = conf.output_x();
fh_ = conf.size_y();
fw_ = conf.size_x();
ph_ = conf.padding_y();
pw_ = conf.padding();
sh_ = conf.stride_y();
sw_ = conf.stride();
const std::string& type = conf.pool_type();
if (type == "max-projection") {
poolAlgo_ = algorithm::pooling_max;
} else if (type == "avg-projection") {
// paddle only use exclude_padding
poolAlgo_ = algorithm::pooling_avg_exclude_padding;
} else {
LOG(FATAL) << "unknow pooling type!";
}
return true;
}
void MKLDNNPoolLayer::reshape(
int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
reshapeInput(bs, ih, iw);
// ic_ and oc can not be changed
CHECK_EQ(inputElemenCnt_ / bs / ih / iw, (size_t)ic)
<< "Input channel can not be changed";
// cal output sizes
// paddle used false caffeMode for pooling
oh = outputSize(ih, fh_, ph_, sh_, false);
ow = outputSize(iw, fw_, pw_, sw_, false);
reshapeOutput(oh, ow);
resizeOutput(bs, oc * oh * ow);
printSizeInfo();
}
void MKLDNNPoolLayer::resetFwd(std::vector<primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
resetFwdBuffers(in, out);
resetFwdPD(fwdPD_, in, out);
resetFwdPipeline(pipeline, fwdPD_, in, out);
printValueFormatFlow();
}
void MKLDNNPoolLayer::resetBwd(std::vector<primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
std::shared_ptr<pool_bwd::primitive_desc> pd;
resetBwdBuffers(in, out);
resetBwdPD(pd, in, out);
resetBwdPipeline(pipeline, pd, in, out);
printGradFormatFlow();
}
void MKLDNNPoolLayer::updateInputData() {
inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
}
void MKLDNNPoolLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
resetInValue(in);
resetOutValue(out);
}
void MKLDNNPoolLayer::resetInValue(MKLDNNMatrixPtr& in) {
if (inputIsOnlyMKLDNN()) {
const MatrixPtr& dnnIn = getInputValue(0);
in = std::dynamic_pointer_cast<MKLDNNMatrix>(dnnIn);
CHECK(in) << "Input should be MKLDNNMatrix";
} else {
CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
in = MKLDNNMatrix::create(
cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
}
}
void MKLDNNPoolLayer::resetOutValue(MKLDNNMatrixPtr& out) {
CHECK(inVal_) << "Should reset input value first";
memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
out = MKLDNNMatrix::create(
output_.value, outDims, inVal_->getFormat(), engine_);
output_.value = std::dynamic_pointer_cast<Matrix>(out);
// create reorder if output value has cpu device and pd do not match
cpuOutVal_ = nullptr;
cvtOutVal_ = nullptr;
if (!outputIsOnlyMKLDNN()) {
const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
cpuOutVal_ = MKLDNNMatrix::create(cpuOut, outDims, format::nchw, engine_);
if (cpuOutVal_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
cvtOutVal_ = MKLDNNMatrix::createReorder(out, cpuOutVal_);
CHECK(cvtOutVal_) << "should not be emptry";
} else {
// CPU output share the same data of MKLDNN output
cpuOut->setData(out->getData());
cpuOutVal_ = out;
}
}
}
void MKLDNNPoolLayer::resetFwdPD(std::shared_ptr<pool_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr in,
MKLDNNMatrixPtr out) {
memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
memory::dims kernels = memory::dims{fh_, fw_};
memory::dims strides = memory::dims{sh_, sw_};
memory::dims padL = memory::dims{ph_, pw_};
memory::dims padR = getPaddingR();
padding_kind padKind = padding_kind::zero;
prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring
: prop_kind::forward_training;
auto fwdDesc = pool_fwd::desc(pk,
poolAlgo_,
in->getMemoryDesc(),
out->getMemoryDesc(),
strides,
kernels,
padL,
padR,
padKind);
pd.reset(new pool_fwd::primitive_desc(fwdDesc, engine_));
// prepare workspace if necessary
workspace_ =
(passType_ != PASS_TEST && poolAlgo_ == algorithm::pooling_max)
? std::make_shared<memory>(memory(pd->workspace_primitive_desc()))
: nullptr;
}
void MKLDNNPoolLayer::resetFwdPipeline(
std::vector<primitive>& pipeline,
std::shared_ptr<pool_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
pipeline.clear();
fwd_ = workspace_
? std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out, *workspace_))
: std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out));
pipeline.push_back(*fwd_);
if (cvtOutVal_) {
pipeline.push_back(*cvtOutVal_);
}
}
void MKLDNNPoolLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
resetOutGrad(out);
resetInGrad(in);
}
void MKLDNNPoolLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
CHECK(outVal_) << "Should have output value";
out = MKLDNNMatrix::create(output_.grad, outVal_->getPrimitiveDesc());
// create reorder if output value has cpu device and pd do not match
cpuOutGrad_ = nullptr;
cvtOutGrad_ = nullptr;
if (!outputIsOnlyMKLDNN()) {
const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
cpuOutGrad_ = MKLDNNMatrix::create(
cpuOut, memory::dims{bs_, oc_, oh_, ow_}, format::nchw, engine_);
if (cpuOutGrad_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
CHECK(cvtOutGrad_) << "should not be emptry";
} else {
// share the same data of CPU output
output_.grad->setData(cpuOut->getData());
out = cpuOutGrad_;
}
}
}
void MKLDNNPoolLayer::resetInGrad(MKLDNNMatrixPtr& in) {
in = nullptr;
const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
if (inGrad == nullptr) {
return;
}
CHECK(inVal_);
in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
}
void MKLDNNPoolLayer::resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
memory::dims kernels = memory::dims{fh_, fw_};
memory::dims strides = memory::dims{sh_, sw_};
memory::dims padL = memory::dims{ph_, pw_};
memory::dims padR = getPaddingR();
CHECK(in);
CHECK(out);
auto bwdDesc = pool_bwd::desc(poolAlgo_,
in->getMemoryDesc(),
out->getMemoryDesc(),
strides,
kernels,
padL,
padR,
padding_kind::zero);
pd.reset(new pool_bwd::primitive_desc(bwdDesc, engine_, *fwdPD_));
}
void MKLDNNPoolLayer::resetBwdPipeline(
std::vector<primitive>& pipeline,
std::shared_ptr<pool_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
pipeline.clear();
if (cvtOutGrad_) {
pipeline.push_back(*cvtOutGrad_);
}
bwdData_ =
workspace_
? std::make_shared<pool_bwd>(pool_bwd(*pd, *out, *workspace_, *in))
: std::make_shared<pool_bwd>(pool_bwd(*pd, *out, *in));
pipeline.push_back(*bwdData_);
}
} // namespace paddle
/* Copyright (c) 2017 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. */
#pragma once
#include "MKLDNNLayer.h"
#include "mkldnn.hpp"
namespace paddle {
typedef mkldnn::pooling_forward pool_fwd;
typedef mkldnn::pooling_backward pool_bwd;
/**
* @brief A subclass of MKLDNNLayer pool layer.
*
* The config file api is mkldnn_pool
*/
class MKLDNNPoolLayer : public MKLDNNLayer {
protected:
// padding height and width
int ph_, pw_;
// stride height and width
int sh_, sw_;
// filter(kenerl) height and width
int fh_, fw_;
// pooling_avg or pooling_max
mkldnn::algorithm poolAlgo_;
// MKLDNNMatrixPtr which should be created from CPU Device
MKLDNNMatrixPtr cpuOutVal_;
MKLDNNMatrixPtr cpuOutGrad_;
// convert handle between CPU device and MKLDNN device
std::shared_ptr<mkldnn::reorder> cvtOutVal_;
std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
// save forward primitive_desc, which can be used backward
std::shared_ptr<pool_fwd::primitive_desc> fwdPD_;
// according to https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
// test_pooling_forward.cpp, pool need workspace for backward
std::shared_ptr<mkldnn::memory> workspace_;
public:
explicit MKLDNNPoolLayer(const LayerConfig& config) : MKLDNNLayer(config) {}
~MKLDNNPoolLayer() {}
bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) override;
void reshape(
int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
void resetFwd(std::vector<mkldnn::primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) override;
void resetBwd(std::vector<mkldnn::primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) override;
void updateInputData() override;
void printSizeInfo() override {
MKLDNNLayer::printSizeInfo();
VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
<< ": ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
<< ", sw: " << sw_;
}
protected:
/**
* Forward functions: reset buffers(input, output),
* reset primitive descriptor,
* reset pipeline.
*/
void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
void resetInValue(MKLDNNMatrixPtr& in);
void resetOutValue(MKLDNNMatrixPtr& out);
void resetFwdPD(std::shared_ptr<pool_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr in,
MKLDNNMatrixPtr out);
void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
std::shared_ptr<pool_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out);
/**
* Backward functions: reset buffers(input, output),
* reset primitive descriptor,
* reset pipeline.
*/
void resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
void resetOutGrad(MKLDNNMatrixPtr& out);
void resetInGrad(MKLDNNMatrixPtr& in);
void resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out);
void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
std::shared_ptr<pool_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out);
/**
* get padding_r according to
* https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
* test_pooling_forward.cpp
*/
mkldnn::memory::dims getPaddingR() const {
mkldnn::memory::dims padR = {ph_, pw_};
for (int i = 0; i < 2; ++i) {
if ((ih_ + ph_ + padR[0] - fh_) / sh_ + 1 < oh_) {
++padR[0];
}
if ((iw_ + pw_ + padR[1] - fw_) / sw_ + 1 < ow_) {
++padR[1];
}
}
return padR;
}
};
} // namespace paddle
...@@ -141,6 +141,68 @@ TEST(MKLDNNLayer, ConvLayer) { ...@@ -141,6 +141,68 @@ TEST(MKLDNNLayer, ConvLayer) {
testConvLayer({4, 4, 16, 3, 3, 16, 3, 3, 3, 3, 1, 1, 1, 1, 1, 1}); testConvLayer({4, 4, 16, 3, 3, 16, 3, 3, 3, 3, 1, 1, 1, 1, 1, 1});
} }
struct testPoolDesc {
int bs, ch; // input channel and output channel are the same
int ih, iw;
int oh, ow;
int fh, fw;
int ph, pw;
int sh, sw;
};
void testPoolLayer(const testPoolDesc& pm) {
const std::string compareTypes[] = {"mkldnn_pool", "pool"};
TestConfig cfg;
cfg.layerConfig.set_type(compareTypes[0]);
cfg.layerConfig.set_size(pm.ch * pm.oh * pm.ow);
cfg.inputDefs.push_back(
{INPUT_DATA,
"layer_0",
/* size of input layer= */ size_t(pm.ch * pm.ih * pm.iw),
0});
LayerInputConfig* input = cfg.layerConfig.add_inputs();
PoolConfig* pool = input->mutable_pool_conf();
// pool->set_pool_type(poolType);
pool->set_channels(pm.ch);
pool->set_img_size(pm.iw);
pool->set_img_size_y(pm.ih);
pool->set_output_x(pm.ow);
pool->set_output_y(pm.oh);
pool->set_size_x(pm.fw);
pool->set_size_y(pm.fh);
pool->set_padding(pm.pw);
pool->set_padding_y(pm.ph);
pool->set_stride(pm.sw);
pool->set_stride_y(pm.sh);
int oh = outputSize(pm.ih, pm.fh, pm.ph, pm.sh, false);
int ow = outputSize(pm.iw, pm.fw, pm.pw, pm.sw, false);
CHECK_EQ(ow, pm.ow) << "output size check failed";
CHECK_EQ(oh, pm.oh) << "output size check failed";
MKLDNNTester tester;
for (auto type : {"max-projection", "avg-projection"}) {
pool->set_pool_type(type);
TestConfig ref = cfg;
ref.layerConfig.set_type(compareTypes[1]);
for (auto bs : {pm.bs, 1}) {
tester.run(cfg, ref, bs, pm.ih, pm.iw);
}
}
}
TEST(MkldnnLayer, PoolLayer) {
/* bs, ch, ih, iw, oh, ow, fh, fw, ph, pw, sh, sw*/
testPoolLayer({2, 1, 4, 4, 2, 2, 3, 3, 0, 0, 2, 2});
testPoolLayer({10, 8, 16, 16, 8, 8, 2, 2, 0, 0, 2, 2});
testPoolLayer({4, 2, 5, 5, 3, 3, 3, 3, 1, 1, 2, 2});
testPoolLayer({8, 16, 56, 56, 28, 28, 3, 3, 0, 0, 2, 2});
testPoolLayer({8, 16, 14, 14, 7, 7, 3, 3, 0, 0, 2, 2});
testPoolLayer({4, 16, 7, 7, 1, 1, 7, 7, 0, 0, 1, 1});
testPoolLayer({4, 2, 5, 5, 3, 3, 5, 5, 1, 1, 1, 1});
testPoolLayer({2, 8, 56, 56, 29, 29, 3, 3, 1, 1, 2, 2});
}
// TODO(TJ): add branch test // TODO(TJ): add branch test
int main(int argc, char** argv) { int main(int argc, char** argv) {
......
...@@ -2286,8 +2286,15 @@ class NormLayer(LayerBase): ...@@ -2286,8 +2286,15 @@ class NormLayer(LayerBase):
@config_layer('pool') @config_layer('pool')
class PoolLayer(LayerBase): class PoolLayer(LayerBase):
layer_type = 'pool'
def __init__(self, name, inputs, ceil_mode=True, **xargs): def __init__(self, name, inputs, ceil_mode=True, **xargs):
super(PoolLayer, self).__init__(name, 'pool', 0, inputs=inputs, **xargs) use_mkldnn = int(g_command_config_args.get("use_mkldnn", 0))
if self.layer_type == "mkldnn_pool":
config_assert(use_mkldnn, "mkldnn_pool only support MKLDNN")
self.layer_type = 'mkldnn_pool' if use_mkldnn else 'pool'
super(PoolLayer, self).__init__(
name, self.layer_type, 0, inputs=inputs, **xargs)
for input_index in xrange(len(self.inputs)): for input_index in xrange(len(self.inputs)):
input_layer = self.get_input_layer(input_index) input_layer = self.get_input_layer(input_index)
pool_conf = self.config.inputs[input_index].pool_conf pool_conf = self.config.inputs[input_index].pool_conf
...@@ -2297,6 +2304,11 @@ class PoolLayer(LayerBase): ...@@ -2297,6 +2304,11 @@ class PoolLayer(LayerBase):
pool_conf.channels) pool_conf.channels)
@config_layer('mkldnn_pool')
class MKLDNNPoolLayer(PoolLayer):
layer_type = 'mkldnn_pool'
@config_layer('pool3d') @config_layer('pool3d')
class Pool3DLayer(LayerBase): class Pool3DLayer(LayerBase):
def __init__(self, name, inputs, ceil_mode=True, **xargs): def __init__(self, name, inputs, ceil_mode=True, **xargs):
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册