m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

refine mkldnn logic, move reset buffers into MKLDNNLayer

Browse Source
revert-4814-Add_sequence_project_op
tensor-tang 7 years ago
parent 60b84856e7
commit c1914543b0
9 changed files with 358 additions and 506 deletions
Show Stats Download Patch File Download Diff File

233
paddle/gserver/layers/MKLDNNConvLayer.cpp
View File

File diff suppressed because it is too large Load Diff

66
paddle/gserver/layers/MKLDNNConvLayer.h
Unescape View File

@ -48,17 +48,6 @@ protected:
// save forward primitive_desc, which can be used backward
std::shared_ptr<conv_fwd::primitive_desc> fwdPD_;
// MKLDNNMatrixPtr which should be created from CPU Device
MKLDNNMatrixPtr cpuInVal_;
MKLDNNMatrixPtr cpuInGrad_;
MKLDNNMatrixPtr cpuOutVal_;
MKLDNNMatrixPtr cpuOutGrad_;
// convert handle between CPU device and MKLDNN device
std::shared_ptr<mkldnn::reorder> cvtInVal_;
std::shared_ptr<mkldnn::reorder> cvtInGrad_;
std::shared_ptr<mkldnn::reorder> cvtOutVal_;
std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
// whether the weight has been init
bool hasInitedWgt_;
@ -94,8 +83,6 @@ public:
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) override;
void updateInputData() override;
void updateWeights(const UpdateCallback& callback) override;
void convertWeightsFromPaddle() override;
@ -109,26 +96,6 @@ public:
<< ", sw: " << sw_ << ", dh: " << dh_ << ", dw: " << dw_;
}
void printValueFormatFlow() override {
if (cpuInVal_) {
VLOG(MKLDNN_FMTS) << cpuInVal_->getFormat() << " >>>";
}
MKLDNNLayer::printValueFormatFlow();
if (cpuOutVal_) {
VLOG(MKLDNN_FMTS) << " >>> " << cpuOutVal_->getFormat();
}
}
void printGradFormatFlow() override {
if (cpuInGrad_) {
VLOG(MKLDNN_FMTS) << cpuInGrad_->getFormat() << " <<<";
}
MKLDNNLayer::printGradFormatFlow();
if (cpuOutGrad_) {
VLOG(MKLDNN_FMTS) << " <<< " << cpuOutGrad_->getFormat();
}
}
protected:
/**
* load the dims settings of this conv
@ -162,23 +129,6 @@ protected:
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out);
/**
* reset MKLDNNMatrix of input value
*/
void resetInValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in);
/**
* reset MKLDNNMatrix of weight and bias value
*/
void resetWgtBiasValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias);
/**
* reset MKLDNNMatrix of output value
*/
void resetOutValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& out);
/**
* reset the backward weight primitive descriptor.
*/
@ -207,22 +157,6 @@ protected:
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out);
/**
* reset MKLDNNMatrix of output grad
*/
void resetOutGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
MKLDNNMatrixPtr& out);
/**
* reset MKLDNNMatrix of weight and bias grad
*/
void resetWgtBiasGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias);
/**
* reset MKLDNNMatrix of input grad
*/
void resetInGrad(std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
MKLDNNMatrixPtr& in);
/**
* reset MKLDNNMatrix of weight value for backward data
* since the primitive_desc would be different with wgtVal_

101
paddle/gserver/layers/MKLDNNFcLayer.cpp
Unescape View File

@ -62,7 +62,7 @@ void MKLDNNFcLayer::convertWeightsFromPaddle() {
CHECK(wgtVal_) << "should have been initialized";
bool hasNoSpatial_ = ih_ == 1 && iw_ == 1;
auto targetDim = wgtVal_->getDims();
auto srcFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo;
auto srcFmt = hasNoSpatial_ ? format::io : format::ihwo;
wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim);
hasInitedWgt_ = true;
}
@ -71,7 +71,7 @@ void MKLDNNFcLayer::convertWeightsToPaddle() {
CHECK(wgtVal_) << "should have been initialized";
bool hasNoSpatial_ = ih_ == 1 && iw_ == 1;
auto targetDim = wgtVal_->getDims();
auto dstFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo;
auto dstFmt = hasNoSpatial_ ? format::io : format::ihwo;
wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
}
@ -100,8 +100,6 @@ void MKLDNNFcLayer::resetFwd(std::vector<primitive>& pipeline,
resetFwdPD(fwdPD_, in, wgt, bias, out);
resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
printValueFormatFlow();
}
void MKLDNNFcLayer::resetBwd(std::vector<primitive>& pipeline,
@ -119,12 +117,6 @@ void MKLDNNFcLayer::resetBwd(std::vector<primitive>& pipeline,
resetBwdDataPD(bwdDataPD, in, out);
resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out);
printGradFormatFlow();
}
void MKLDNNFcLayer::updateInputData() {
inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
}
void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
@ -139,51 +131,33 @@ void MKLDNNFcLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
resetInValue(in);
CHECK(in);
in->downSpatial();
resetWgtBiasValue(wgt, bias);
resetOutValue(out);
}
// if (extInVal_) {
// extInVal_->downSpatial();
// }
void MKLDNNFcLayer::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_);
}
in->downSpatial();
}
auto outPD =
MKLDNNMatrix::createPrimitiveDesc({bs_, oc_}, format::nc, engine_);
resetOutValue(out, outPD);
void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias) {
format wgtFmt = format::oihw;
if (inVal_->getFormat() == format::nChw8c) {
if (in->getFormat() == format::nChw8c) {
wgtFmt = format::oIhw8i;
} else if (inVal_->getFormat() == format::nChw16c) {
} else if (in->getFormat() == format::nChw16c) {
wgtFmt = format::oIhw16i;
}
wgt = MKLDNNMatrix::create(
weight_->getW(), {oc_, ic_, ih_, iw_}, wgtFmt, engine_);
auto wgtPD =
MKLDNNMatrix::createPrimitiveDesc({oc_, ic_, ih_, iw_}, wgtFmt, engine_);
resetWithMatrix(wgt, weight_->getW(), wgtPD);
wgt->downSpatial();
VLOG(MKLDNN_FMTS) << "Weight value format: " << wgt->getFormat();
bias = (biases_ && biases_->getW())
? MKLDNNMatrix::create(biases_->getW(), {oc_}, format::x, engine_)
: nullptr;
}
void MKLDNNFcLayer::resetOutValue(MKLDNNMatrixPtr& out) {
out = MKLDNNMatrix::create(output_.value, {bs_, oc_}, format::nc, engine_);
if (!outputIsOnlyMKLDNN()) {
// fc cpu output value do not need create convert, just share data
getOutput(CPU_DEVICE).value->setData(out->getData());
if (biases_ == nullptr || biases_->getW() == nullptr) {
return;
}
output_.value = std::dynamic_pointer_cast<Matrix>(out);
auto biasPD = MKLDNNMatrix::createPrimitiveDesc({oc_}, format::x, engine_);
resetWithMatrix(bias, biases_->getW(), biasPD);
}
void MKLDNNFcLayer::resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
@ -219,7 +193,6 @@ void MKLDNNFcLayer::resetFwdPipeline(
} else {
fwd_.reset(new fc_fwd(*pd, *in, *wgt, *out));
}
pipeline.push_back(*fwd_);
}
@ -227,44 +200,18 @@ void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
resetOutGrad(out);
resetWgtBiasGrad(wgt, bias);
resetInGrad(in);
}
void MKLDNNFcLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
CHECK(outVal_);
if (outputIsOnlyMKLDNN()) {
MKLDNNLayer::resetOutGrad(out, outVal_->getPrimitiveDesc());
} else {
const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
output_.grad->setData(cpuOut->getData());
out = MKLDNNMatrix::create(cpuOut, outVal_->getPrimitiveDesc());
}
}
CHECK(inVal_ && outVal_);
resetOutGrad(out, outVal_->getPrimitiveDesc());
resetInGrad(in, inVal_->getPrimitiveDesc());
void MKLDNNFcLayer::resetWgtBiasGrad(MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias) {
CHECK(wgtVal_);
wgt = MKLDNNMatrix::create(weight_->getWGrad(), wgtVal_->getPrimitiveDesc());
resetWithMatrix(wgt, weight_->getWGrad(), wgtVal_->getPrimitiveDesc());
bias = nullptr;
if (biasVal_ == nullptr) {
return;
}
bias =
MKLDNNMatrix::create(biases_->getWGrad(), biasVal_->getPrimitiveDesc());
}
void MKLDNNFcLayer::resetInGrad(MKLDNNMatrixPtr& in) {
in = nullptr;
if (inputLayers_[0]->getOutput().grad == nullptr) {
return;
}
CHECK(inVal_);
MKLDNNLayer::resetInGrad(in, inVal_->getPrimitiveDesc());
resetWithMatrix(bias, biases_->getWGrad(), biasVal_->getPrimitiveDesc());
}
void MKLDNNFcLayer::resetBwdWgtPD(

8
paddle/gserver/layers/MKLDNNFcLayer.h
Unescape View File

@ -66,8 +66,6 @@ public:
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) override;
void updateInputData() override;
void updateWeights(const UpdateCallback& callback) override;
void convertWeightsFromPaddle() override;
@ -84,9 +82,6 @@ protected:
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out);
void resetInValue(MKLDNNMatrixPtr& in);
void resetWgtBiasValue(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
void resetOutValue(MKLDNNMatrixPtr& out);
void resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr in,
MKLDNNMatrixPtr wgt,
@ -109,9 +104,6 @@ protected:
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out);
void resetOutGrad(MKLDNNMatrixPtr& out);
void resetWgtBiasGrad(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
void resetInGrad(MKLDNNMatrixPtr& in);
void resetBwdWgtPD(std::shared_ptr<fc_bwdWgt::primitive_desc>& pd,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,

324
paddle/gserver/layers/MKLDNNLayer.h
View File

File diff suppressed because it is too large Load Diff

103
paddle/gserver/layers/MKLDNNPoolLayer.cpp
Unescape View File

@ -85,8 +85,6 @@ void MKLDNNPoolLayer::resetFwd(std::vector<primitive>& pipeline,
resetFwdPD(fwdPD_, in, out);
resetFwdPipeline(pipeline, fwdPD_, in, out);
printValueFormatFlow();
}
void MKLDNNPoolLayer::resetBwd(std::vector<primitive>& pipeline,
@ -101,65 +99,22 @@ void MKLDNNPoolLayer::resetBwd(std::vector<primitive>& pipeline,
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_);
// 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()) {
out = MKLDNNMatrix::create(nullptr, out->getPrimitiveDesc());
cvtOutVal_ = MKLDNNMatrix::createReorder(out, cpuOutVal_);
CHECK(cvtOutVal_) << "should not be emptry";
} else {
cpuOut->setData(output_.value->getData());
cpuOutVal_ = out;
}
output_.value = std::dynamic_pointer_cast<Matrix>(cpuOutVal_);
return;
}
output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
CHECK(in);
auto outPD =
MKLDNNMatrix::createPrimitiveDesc(outDims, in->getFormat(), engine_);
resetOutValue(out, outPD);
}
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_};
@ -194,58 +149,26 @@ void MKLDNNPoolLayer::resetFwdPipeline(
? 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) {
cpuOutGrad_ = nullptr;
cvtOutGrad_ = nullptr;
CHECK(outVal_);
if (outputIsOnlyMKLDNN()) {
MKLDNNLayer::resetOutGrad(out, outVal_->getPrimitiveDesc());
} else {
const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
// always share the same grad data of CPU output
// then the activation can get the right grad from output_.grad
output_.grad->setData(cpuOut->getData());
cpuOutGrad_ = MKLDNNMatrix::create(
cpuOut, memory::dims{bs_, oc_, oh_, ow_}, format::nchw, engine_);
if (cpuOutGrad_->getPrimitiveDesc() != outVal_->getPrimitiveDesc()) {
out = MKLDNNMatrix::create(nullptr, outVal_->getPrimitiveDesc());
cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
CHECK(cvtOutGrad_) << "should not be emptry";
} else {
out = cpuOutGrad_;
}
}
}
void MKLDNNPoolLayer::resetInGrad(MKLDNNMatrixPtr& in) {
in = nullptr;
if (inputLayers_[0]->getOutput().grad == nullptr) {
return;
}
CHECK(inVal_);
MKLDNNLayer::resetInGrad(in, inVal_->getPrimitiveDesc());
CHECK(inVal_ && outVal_);
resetOutGrad(out, outVal_->getPrimitiveDesc());
resetInGrad(in, inVal_->getPrimitiveDesc());
}
void MKLDNNPoolLayer::resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
pd = nullptr;
if (in == nullptr) {
return;
}
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(),
@ -263,8 +186,8 @@ void MKLDNNPoolLayer::resetBwdPipeline(
std::shared_ptr<pool_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
if (cvtOutGrad_) {
pipeline.push_back(*cvtOutGrad_);
if (pd == nullptr) {
return;
}
bwdData_ =

13
paddle/gserver/layers/MKLDNNPoolLayer.h
Unescape View File

@ -38,13 +38,6 @@ protected:
// 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/
@ -74,8 +67,6 @@ public:
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) override;
void updateInputData() override;
void printSizeInfo() override {
MKLDNNLayer::printSizeInfo();
VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
@ -90,8 +81,6 @@ protected:
* 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);
@ -106,8 +95,6 @@ protected:
* 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);

2
paddle/math/MKLDNNMatrix.cpp
Unescape View File

@ -46,7 +46,7 @@ MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m,
memory::format fmt,
engine& eg,
mkldnn::memory::data_type dtype) {
return create(m, memory::primitive_desc(memory::desc(dims, dtype, fmt), eg));
return create(m, createPrimitiveDesc(dims, fmt, eg, dtype));
}
std::shared_ptr<reorder> MKLDNNMatrix::createReorder(const MKLDNNMatrixPtr& src,

14
paddle/math/MKLDNNMatrix.h
Unescape View File

@ -52,12 +52,24 @@ public:
mkldnn::engine& eg,
mkldnn::memory::data_type dtype = mkldnn::memory::data_type::f32);
/**
* Create primitive descriptor.
* default with f32 dtype
*/
static mkldnn::memory::primitive_desc createPrimitiveDesc(
const mkldnn::memory::dims dims,
const mkldnn::memory::format& fmt,
const mkldnn::engine& eg,
const mkldnn::memory::data_type& dtype = mkldnn::memory::data_type::f32) {
return mkldnn::memory::primitive_desc(memory::desc(dims, dtype, fmt), eg);
}
/**
* Create Memory descriptor.
* default with any format and f32 dtype
*/
static mkldnn::memory::desc createMemoryDesc(
const mkldnn::memory::dims& dims,
const mkldnn::memory::dims dims,
const mkldnn::memory::format& fmt = mkldnn::memory::format::any,
const mkldnn::memory::data_type& dtype = mkldnn::memory::data_type::f32) {
return mkldnn::memory::desc(dims, dtype, fmt);

Write Preview
Loading…
Cancel
Save