diff --git a/paddle/gserver/layers/MKLDNNFcLayer.cpp b/paddle/gserver/layers/MKLDNNFcLayer.cpp index 8318c8c519a4cec1610eadd28320ee5ce0b4147d..53433cef35a377a73f87b041fdcfadd848dd2ec9 100644 --- a/paddle/gserver/layers/MKLDNNFcLayer.cpp +++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp @@ -77,24 +77,6 @@ void MKLDNNFcLayer::convertWeightsToPaddle() { wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim); } -void MKLDNNFcLayer::convertOutputToOtherDevice() { - copyOutputInfoToOtherDevice(); - // find other cpu device and reorder output to cpu device - int cnt = 0; - for (size_t i = 0; i < outputOtherDevice_.size(); i++) { - if (outputOtherDevice_[i].deviceId == CPU_DEVICE) { - // fc cpu output value do not need convert - // just share point - outputOtherDevice_[i].value = output_.value; - ++cnt; - } - } - - if (cnt > 1) { - LOG(WARNING) << "should not have more than one CPU devie"; - } -} - void MKLDNNFcLayer::reshape() { const Argument& input = getInput(0, getPrev(0)->getDeviceId()); int batchSize = input.getBatchSize(); @@ -155,7 +137,10 @@ void MKLDNNFcLayer::resetFwd() { // change original output value to mkldnn output value output_.value = std::dynamic_pointer_cast(outVal_); if (!outputIsOnlyMKLDNN()) { - convertOutputToOtherDevice(); + copyOutputInfoToOtherDevice(); + // fc cpu output value do not need create convert + // just share point + getOutput(CPU_DEVICE).value->setData(output_.value->getData()); } // create forward handle @@ -235,13 +220,12 @@ void MKLDNNFcLayer::resetBwd() { pipelineBwd_.push_back(*bwdWgt_); /// backward data - device = inputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE; - const MatrixPtr& in = getInputGrad(0, device); + const MatrixPtr& in = inputLayers_[0]->getOutput().grad; if (in == nullptr) { return; } - if (getInput(0, device).getAllCount() > 1) { - // TODO(TJ): use outputMaps_ ways when merge outgrad done + if (getInput(0, MKLDNN_DEVICE).getAllCount() > 1) { + // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done } else { inGrad_ = MKLDNNMatrix::create(in, inVal_->getPrimitiveDesc()); } @@ -258,13 +242,21 @@ void MKLDNNFcLayer::resetBwd() { pipelineBwd_.push_back(*bwdData_); } +void MKLDNNFcLayer::updateInputData() { + if (inputLayers_[0]->getType() != "data") { + return; + } + real* iData = getInputValue(0, CPU_DEVICE)->getData(); + inVal_->setData(iData); +} + void MKLDNNFcLayer::forward(PassType passType) { Layer::forward(passType); reshape(); { REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str()); - syncInputValue(); + updateInputData(); // just submit forward pipeline stream_->submit(pipelineFwd_); @@ -286,7 +278,6 @@ void MKLDNNFcLayer::backward(const UpdateCallback& callback) { REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str()); resetBwd(); - syncOutputGrad(); // just sumbmit backward pipeline stream_->submit(pipelineBwd_); } diff --git a/paddle/gserver/layers/MKLDNNFcLayer.h b/paddle/gserver/layers/MKLDNNFcLayer.h index e138a6faf181c412949218458e7ecf800a0d6a07..4ad67a16e056a718c45a28babcf22a7cd571b15c 100644 --- a/paddle/gserver/layers/MKLDNNFcLayer.h +++ b/paddle/gserver/layers/MKLDNNFcLayer.h @@ -53,6 +53,8 @@ public: void backward(const UpdateCallback& callback) override; + void updateInputData() override; + protected: /** * reshape the input image sizes @@ -72,8 +74,6 @@ protected: * only would be called when needed */ void resetBwd(); - - void convertOutputToOtherDevice() override; }; } // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h index b983b833d510b823c5d4cff0b9390173e4cefc89..543364edceff684bdcd002a8f4f10e7ce5e6953b 100644 --- a/paddle/gserver/layers/MKLDNNLayer.h +++ b/paddle/gserver/layers/MKLDNNLayer.h @@ -114,10 +114,10 @@ public: virtual void convertWeightsToPaddle() {} /** - * convert MKLDNN output to other device. - * only support CPU device yet + * Update input value data when input layer is "data" type. + * Since the input value data address might be changed. */ - virtual void convertOutputToOtherDevice() {} + virtual void updateInputData() {} /** * print info about sizes @@ -155,6 +155,7 @@ protected: * 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()); @@ -163,6 +164,12 @@ protected: 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"; } } @@ -193,32 +200,6 @@ protected: return outputOtherDevice_.size() == 0; } - /** - * Sync input value data - */ - void syncInputValue() { - if (inputIsOnlyMKLDNN()) { - return; - } - real* iData = getInputValue(0, CPU_DEVICE)->getData(); - // update input data - // since it might be changed if this is after data layer - inVal_->updateData(iData); - } - - /** - * Sync output grad data - */ - void syncOutputGrad() { - if (outputIsOnlyMKLDNN()) { - return; - } - - // update diff - real* oDiff = getOutput(CPU_DEVICE).grad->getData(); - outGrad_->updateData(oDiff); - } - /** * Set deviceId of this layer. */ diff --git a/paddle/math/MKLDNNMatrix.cpp b/paddle/math/MKLDNNMatrix.cpp index 0a355e2644cce572ce90ecf5c9d2a5b7b395bc61..c4063e5069854242d9f93886b66580385557ca73 100644 --- a/paddle/math/MKLDNNMatrix.cpp +++ b/paddle/math/MKLDNNMatrix.cpp @@ -33,14 +33,12 @@ MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, memory::primitive_desc pd) { size_t width = cnts / dims[0]; m = Matrix::create(height, width, false, false); } - CHECK(m) << " Matrix should not be empty"; + CpuMatrixPtr cpuMatrix = std::dynamic_pointer_cast(m); CHECK(cpuMatrix) << "Only support create from CPU matrix yet"; - - CHECK_EQ(cnts, m->getElementCnt()) << "Count size does not match"; - return std::make_shared( - m->getData(), m->getHeight(), m->getWidth(), pd); + CHECK_EQ(cpuMatrix->getElementCnt(), cnts) << "Count size does not match"; + return std::make_shared(cpuMatrix, pd); } MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, @@ -138,7 +136,7 @@ void MKLDNNMatrix::downSpatial() { mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr), "could not create a memory primitive"); reset(result); - set_data_handle(getData()); + set_data_handle(data_); } } // namespace paddle diff --git a/paddle/math/MKLDNNMatrix.h b/paddle/math/MKLDNNMatrix.h index e50f698b495713e6f15ab7a12a7ee7487662040f..eef3b429e6fa0087aeac3f5aed9dff983b06e826 100644 --- a/paddle/math/MKLDNNMatrix.h +++ b/paddle/math/MKLDNNMatrix.h @@ -30,11 +30,10 @@ typedef std::shared_ptr MKLDNNMatrixPtr; */ class MKLDNNMatrix : public CpuMatrix, public mkldnn::memory { public: - MKLDNNMatrix(real* data, - size_t height, - size_t width, - mkldnn::memory::primitive_desc pd) - : CpuMatrix(data, height, width, false), mkldnn::memory(pd, data) {} + MKLDNNMatrix(CpuMatrixPtr m, mkldnn::memory::primitive_desc pd) + : CpuMatrix(m->getData(), m->getHeight(), m->getWidth(), false), + mkldnn::memory(pd, m->getData()), + m_(m) {} ~MKLDNNMatrix() {} @@ -81,11 +80,29 @@ public: void downSpatial(); /** - * Update the memory data handle. + * set the memory data handle. * Caution: This will not check the buffer size of the data, * it should be coverd by user. */ - void updateData(void* data) { set_data_handle(data); } + void setData(real* data) { + set_data_handle(data); + CpuMatrix::setData(data); + m_.reset(); + } + + /** + * override Matrix::getData + * check data before return + */ + real* getData() override { + CHECK_EQ((void*)data_, get_data_handle()); + return data_; + } + + const real* getData() const override { + CHECK_EQ((void*)data_, get_data_handle()); + return data_; + } /** * Get primitive descriptor. @@ -143,6 +160,10 @@ protected: memory::format srcFmt, memory::format dstFmt, memory::dims dm); + +private: + // save the CpuMatrixPtr in case the buffer released outside + CpuMatrixPtr m_; }; } // namespace paddle