提交 94ea8ee0 编写于 作者: T tensor-tang

refine MKLDNNLayer logical: move forward and backward to MKLDNNLayer and...

refine MKLDNNLayer logical: move forward and backward to MKLDNNLayer and remove copyOutputInfoToOtherDevice
上级 f40d5f58
......@@ -14,7 +14,6 @@ limitations under the License. */
#include "MKLDNNFcLayer.h"
#include "paddle/utils/Logging.h"
#include "paddle/utils/Stat.h"
using namespace mkldnn; // NOLINT
typedef memory::format format;
......@@ -40,6 +39,8 @@ bool MKLDNNFcLayer::init(const LayerMap& layerMap,
oc_ = getSize();
oh_ = 1;
ow_ = 1;
ih_ = 1;
iw_ = 1;
// input size can not change in FC
iLayerSize_ = inputLayers_[0]->getSize();
......@@ -78,36 +79,17 @@ void MKLDNNFcLayer::convertWeightsToPaddle() {
}
void MKLDNNFcLayer::reshape() {
const Argument& input = getInput(0, getPrev(0)->getDeviceId());
int batchSize = input.getBatchSize();
if (bs_ == batchSize) {
return;
}
bs_ = batchSize;
ih_ = input.getFrameHeight();
iw_ = input.getFrameWidth();
if (ih_ == 0) {
ih_ = 1;
}
if (iw_ == 0) {
iw_ = 1;
}
reshapeInput();
CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize());
ic_ = iLayerSize_ / (ih_ * iw_);
CHECK_EQ(size_t(ic_ * ih_ * iw_), iLayerSize_) << "not divisible";
CHECK_EQ(size_t(oc_), getSize());
printSizeInfo();
// reset output
output_.setFrameHeight(oh_);
output_.setFrameWidth(ow_);
resetOutput(bs_, oc_);
// reset mkldnn forward
resetFwd();
needResetBwd_ = true;
reshapeOutput(oh_, ow_);
resizeOutput(bs_, oc_);
convertWeightsFromPaddle();
printSizeInfo();
}
void MKLDNNFcLayer::resetFwd() {
......@@ -137,7 +119,6 @@ void MKLDNNFcLayer::resetFwd() {
// change original output value to mkldnn output value
output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
if (!outputIsOnlyMKLDNN()) {
copyOutputInfoToOtherDevice();
// fc cpu output value do not need create convert
// just share point
getOutput(CPU_DEVICE).value->setData(output_.value->getData());
......@@ -243,51 +224,13 @@ void MKLDNNFcLayer::resetBwd() {
}
void MKLDNNFcLayer::updateInputData() {
if (inputLayers_[0]->getType() != "data") {
return;
}
real* iData = getInputValue(0, CPU_DEVICE)->getData();
inVal_->setData(iData);
inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
}
void MKLDNNFcLayer::forward(PassType passType) {
Layer::forward(passType);
reshape();
{
REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
updateInputData();
// just submit forward pipeline
stream_->submit(pipelineFwd_);
}
/* activation */ {
REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
forwardActivation();
}
}
void MKLDNNFcLayer::backward(const UpdateCallback& callback) {
/* Do derivation */ {
REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
backwardActivation();
}
{
REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
resetBwd();
// just sumbmit backward pipeline
stream_->submit(pipelineBwd_);
}
{
REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
weight_->getParameterPtr()->incUpdate(callback);
if (biases_ && biases_->getWGrad()) {
biases_->getParameterPtr()->incUpdate(callback);
}
}
}
} // namespace paddle
......@@ -45,35 +45,19 @@ public:
bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) override;
void convertWeightsFromPaddle() override;
void convertWeightsToPaddle() override;
void reshape() override;
void forward(PassType passType) override;
void resetFwd() override;
void backward(const UpdateCallback& callback) override;
void resetBwd() override;
void updateInputData() override;
protected:
/**
* reshape the input image sizes
* and reset output buffer size
* and reset mkldnn forward
*/
void reshape();
void updateWeights(const UpdateCallback& callback) override;
/**
* reset the forward primitve and memory
* only would be called when input size changes
*/
void resetFwd();
void convertWeightsFromPaddle() override;
/**
* reset the backward primitve and memory for mkldnn fc
* only would be called when needed
*/
void resetBwd();
void convertWeightsToPaddle() override;
};
} // namespace paddle
......@@ -19,6 +19,7 @@ limitations under the License. */
#include "MKLDNNBase.h"
#include "mkldnn.hpp"
#include "paddle/math/MKLDNNMatrix.h"
#include "paddle/utils/Stat.h"
DECLARE_bool(use_mkldnn);
......@@ -33,6 +34,8 @@ typedef std::shared_ptr<MKLDNNLayer> MKLDNNLayerPtr;
*/
class MKLDNNLayer : public Layer {
protected:
// input value element count
size_t inputElemenCnt_;
// batch size
int bs_;
// input image channel, height and width
......@@ -52,7 +55,7 @@ protected:
std::vector<mkldnn::primitive> pipelineFwd_;
std::vector<mkldnn::primitive> pipelineBwd_;
// MKLDNNMatrixPtr
// MKLDNNMatrixPtr with internal format
MKLDNNMatrixPtr inVal_;
MKLDNNMatrixPtr inGrad_;
MKLDNNMatrixPtr outVal_;
......@@ -65,6 +68,7 @@ protected:
public:
explicit MKLDNNLayer(const LayerConfig& config)
: Layer(config),
inputElemenCnt_(0),
bs_(0),
ic_(0),
ih_(0),
......@@ -95,12 +99,93 @@ public:
if (!Layer::init(layerMap, parameterMap)) {
return false;
}
checkCPUOutputsNumber();
stream_.reset(new MKLDNNStream());
engine_ = CPUEngine::Instance().getEngine();
return true;
}
void forward(PassType passType) override {
passType_ = passType;
{
REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
copySeqInfoToOutputs();
CHECK(!inputLayers_.empty());
size_t elemenCnt = inputLayers_[0]->getOutput().value->getElementCnt();
if (inputElemenCnt_ != elemenCnt) {
inputElemenCnt_ = elemenCnt;
reshape();
resetFwd();
convertWeightsFromPaddle();
needResetBwd_ = true;
}
if (inputLayers_[0]->getType() == "data") {
updateInputData();
}
stream_->submit(pipelineFwd_);
}
/* activation */ {
REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
forwardActivation();
}
}
void backward(const UpdateCallback& callback) override {
/* Do derivation */ {
REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
backwardActivation();
}
{
REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
if (needResetBwd_) {
resetBwd();
needResetBwd_ = false;
}
stream_->submit(pipelineBwd_);
}
{
REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
updateWeights(callback);
}
}
/**
* reshape the input image sizes
* and reset output image and buffer size
*/
virtual void reshape() = 0;
/**
* reset the mkldnn forward primitve and memory
* only would be called when input size changes
*/
virtual void resetFwd() = 0;
/**
* reset the mkldnn backward primitve and memory for mkldnn fc
* only would be called when needed
*/
virtual void resetBwd() = 0;
/**
* Update input value data when input layer is "data" type.
* Since the input value data address might be changed.
*/
virtual void updateInputData() {}
/**
* Update weights and biases if necessary.
*/
virtual void updateWeights(const UpdateCallback& callback) {}
/**
* convert weight from paddle format to mkldnn format
* weight_ will be override
......@@ -114,10 +199,38 @@ public:
virtual void convertWeightsToPaddle() {}
/**
* Update input value data when input layer is "data" type.
* Since the input value data address might be changed.
* add this interface as public for unit test
*/
virtual void updateInputData() {}
void addOutputArgument(int deviceId) { Layer::addOutputArgument(deviceId); }
protected:
/**
* reshape the input image sizes and input batchsize
*/
virtual void reshapeInput() {
const Argument& input = inputLayers_[0]->getOutput();
bs_ = input.getBatchSize();
int height = input.getFrameHeight();
int width = input.getFrameWidth();
if (height != 0) {
ih_ = height;
}
if (width != 0) {
iw_ = width;
}
}
/**
* reshape output image sizes
*/
virtual void reshapeOutput(size_t height, size_t width) {
output_.setFrameHeight(height);
output_.setFrameWidth(width);
for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
outputOtherDevice_[i].setFrameHeight(height);
outputOtherDevice_[i].setFrameWidth(width);
}
}
/**
* print info about sizes
......@@ -133,8 +246,8 @@ public:
*/
virtual void printValueFormatFlow() {
if (inVal_ && outVal_) {
VLOG(MKLDNN_FMTS) << "value format flow --- " << inVal_->getFormat()
<< " >>> " << outVal_->getFormat();
VLOG(MKLDNN_FMTS) << inVal_->getFormat() << " >>> "
<< outVal_->getFormat();
}
}
......@@ -143,36 +256,12 @@ public:
*/
virtual void printGradFormatFlow() {
if (inGrad_ && outGrad_) {
VLOG(MKLDNN_FMTS) << "grad format flow --- " << inGrad_->getFormat()
<< " <<< " << outGrad_->getFormat();
VLOG(MKLDNN_FMTS) << inGrad_->getFormat() << " <<< "
<< outGrad_->getFormat();
}
}
protected:
/**
* copy image size and sequence info to other device
* @note: can not directly use Layer::copyOutputToOtherDevice since here only
* copy base info and do not copy data value
*/
void copyOutputInfoToOtherDevice() {
int cnt = 0;
for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
outputOtherDevice_[i].setFrameHeight(output_.getFrameHeight());
outputOtherDevice_[i].setFrameWidth(output_.getFrameWidth());
outputOtherDevice_[i].sequenceStartPositions =
output_.sequenceStartPositions;
outputOtherDevice_[i].subSequenceStartPositions =
output_.subSequenceStartPositions;
outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims;
if (outputOtherDevice_[i].deviceId == CPU_DEVICE) {
++cnt;
}
}
if (cnt > 1) {
LOG(WARNING) << "should not have more than one CPU devie";
}
}
/**
* If input only has MKLDNN device.
* Otherwise, only support the previous layer using CPU device.
......@@ -205,6 +294,7 @@ protected:
*/
void setDevice(int id) { deviceId_ = id; }
private:
/**
* Set deviceId of the params used in this layer.
*/
......@@ -228,6 +318,42 @@ protected:
parameter->setDevice(id);
}
}
/**
* Check the cpu device number of outputOtherDevice_.
* should have only one at most.
*/
void checkCPUOutputsNumber(int max = 1) {
int cnt = 0;
for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
if (outputOtherDevice_[i].deviceId == CPU_DEVICE) {
++cnt;
}
}
CHECK_LE(cnt, max) << "too much CPU devies";
}
/**
* copy SeqInfo from input layer to this output and other output devices.
* @note: do not use getInput(0) since it used this deviceId_,
* use "inputLayers_[0]->getOutput()" instead.
*/
void copySeqInfoToOutputs() {
if (inputLayers_.empty() || !needSequenceInfo_) {
return;
}
const Argument& input = inputLayers_[0]->getOutput();
output_.sequenceStartPositions = input.sequenceStartPositions;
output_.subSequenceStartPositions = input.subSequenceStartPositions;
output_.cpuSequenceDims = input.cpuSequenceDims;
for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
outputOtherDevice_[i].sequenceStartPositions =
output_.sequenceStartPositions;
outputOtherDevice_[i].subSequenceStartPositions =
output_.subSequenceStartPositions;
outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims;
}
}
};
} // namespace paddle
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