Skip to content

Commit

Permalink
Refactor MUL functions, pass object reference instead of shared_ptr.
Browse files Browse the repository at this point in the history
  • Loading branch information
xutianbing committed Dec 20, 2016
1 parent 706c572 commit 4fbf949
Show file tree
Hide file tree
Showing 21 changed files with 144 additions and 154 deletions.
6 changes: 3 additions & 3 deletions paddle/gserver/layers/ConvexCombinationLayer.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -113,7 +113,7 @@ void ConvexCombinationLayer::forward(PassType passType) {
tmpRow0->setData(inV0->getData() + i * weightDim);
tmpRow1->setData(outV->getData() + i * dataDim);

tmpRow1->mul(tmpRow0, tmpMtx0, 1, 0);
tmpRow1->mul(*tmpRow0, *tmpMtx0, 1, 0);
}
}

Expand All @@ -136,7 +136,7 @@ void ConvexCombinationLayer::backward(const UpdateCallback& callback) {
tmpRow1->setData(outG->getData() + i * dataDim);
tmpMtx0->setData(inV1->getData() + i * weightDim * dataDim);

tmpRow0->mul(tmpRow1, tmpMtx0->getTranspose(), 1, 1);
tmpRow0->mul(*tmpRow1, *(tmpMtx0->getTranspose()), 1, 1);
}
}

Expand All @@ -146,7 +146,7 @@ void ConvexCombinationLayer::backward(const UpdateCallback& callback) {
tmpRow1->setData(outG->getData() + i * dataDim);
tmpMtx0->setData(inG1->getData() + i * weightDim * dataDim);

tmpMtx0->mul(tmpRow0->getTranspose(), tmpRow1, 1, 1);
tmpMtx0->mul(*(tmpRow0->getTranspose()), *tmpRow1, 1, 1);
}
}
}
Expand Down
6 changes: 3 additions & 3 deletions paddle/gserver/layers/ExpandConvBaseLayer.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -150,7 +150,7 @@ void ExpandConvBaseLayer::expandFwdOnce(MatrixPtr image,
Matrix::create(wgtData, subM, subK, false, useGpu_); // mark transpose
MatrixPtr B = Matrix::create(expInData, subK, subN, false, useGpu_);
MatrixPtr C = Matrix::create(outData, subM, subN, false, useGpu_);
C->mul(A, B, 1, 1);
C->mul(*A, *B, 1, 1);

A->clear();
B->clear();
Expand Down Expand Up @@ -185,7 +185,7 @@ void ExpandConvBaseLayer::bpropActs(MatrixPtr out,
MatrixPtr C = Matrix::create(expandInData, subK, subN, false, useGpu_);
MatrixPtr B = Matrix::create(localGradData, subM, subN, false, useGpu_);
MatrixPtr A = Matrix::create(wgtData, subM, subK, true, useGpu_);
C->mul(A, B); // mul
C->mul(*A, *B); // mul

// clear the temporary matrix
A->clear();
Expand Down Expand Up @@ -252,7 +252,7 @@ void ExpandConvBaseLayer::bpropWeights(MatrixPtr image,
MatrixPtr A = Matrix::create(expandInData, subK, subN, true, useGpu_);
MatrixPtr B = Matrix::create(gradData, subM, subN, false, useGpu_);
MatrixPtr C = Matrix::create(wGradData, subM, subK, false, useGpu_);
C->mul(B, A, 1, 1);
C->mul(*B, *A, 1, 1);

A->clear();
B->clear();
Expand Down
7 changes: 4 additions & 3 deletions paddle/gserver/layers/FullMatrixProjection.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,7 @@ FullMatrixProjection::FullMatrixProjection(const ProjectionConfig& config,

void FullMatrixProjection::forward() {
REGISTER_TIMER_INFO("FwMulTimer", getName().c_str());
out_->value->mul(in_->value, weight_->getW(), 1, 1);
out_->value->mul(*(in_->value), *(weight_->getW()), 1, 1);
}

void FullMatrixProjection::backward(const UpdateCallback& callback) {
Expand All @@ -37,7 +37,8 @@ void FullMatrixProjection::backward(const UpdateCallback& callback) {
/* Calculate the W-gradient for the current layer */
if (weight_->getWGrad()) {
REGISTER_TIMER_INFO("GradMulTimer", getName().c_str());
weight_->getWGrad()->mul(in_->value->getTranspose(), out_->grad, 1, 1);
weight_->getWGrad()->mul(
*(in_->value->getTranspose()), *(out_->grad), 1, 1);
}

// If callback does not change value, backward propagation error
Expand All @@ -47,7 +48,7 @@ void FullMatrixProjection::backward(const UpdateCallback& callback) {
/* Calculate the input layers error */
if (in_->grad) {
REGISTER_TIMER_INFO("BpMulTimer", getName().c_str());
in_->grad->mul(out_->grad, weight_->getW()->getTranspose(), 1, 1);
in_->grad->mul(*(out_->grad), *(weight_->getW()->getTranspose()), 1, 1);
}

hl_set_sync_flag(syncFlag);
Expand Down
8 changes: 4 additions & 4 deletions paddle/gserver/layers/FullyConnectedLayer.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -84,8 +84,8 @@ void FullyConnectedLayer::forward(PassType passType) {
auto input = getInput(i);
CHECK(input.value) << "The input of 'fc' layer must be matrix";
REGISTER_TIMER_INFO("FwMulTimer", getName().c_str());
i == 0 ? outV->mul(input.value, weights_[i]->getW(), 1, 0)
: outV->mul(input.value, weights_[i]->getW(), 1, 1);
i == 0 ? outV->mul(*input.value, *weights_[i]->getW(), 1, 0)
: outV->mul(*input.value, *weights_[i]->getW(), 1, 1);
}

/* add the bias-vector */
Expand Down Expand Up @@ -123,7 +123,7 @@ void FullyConnectedLayer::backward(const UpdateCallback& callback) {
MatrixPtr oGrad = getOutputGrad();
{
REGISTER_TIMER_INFO("GradMulTimer", getName().c_str());
weights_[i]->getWGrad()->mul(input_T, oGrad, 1, 1);
weights_[i]->getWGrad()->mul(*input_T, *oGrad, 1, 1);
}
}

Expand All @@ -136,7 +136,7 @@ void FullyConnectedLayer::backward(const UpdateCallback& callback) {
if (NULL != preGrad) {
MatrixPtr weights_T = weights_[i]->getW()->getTranspose();
REGISTER_TIMER_INFO("BpMulTimer", getName().c_str());
preGrad->mul(getOutputGrad(), weights_T, 1, 1);
preGrad->mul(*getOutputGrad(), *weights_T, 1, 1);
}

hl_set_sync_flag(syncFlag);
Expand Down
2 changes: 1 addition & 1 deletion paddle/gserver/layers/LinearChainCRF.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -59,7 +59,7 @@ real LinearChainCRF::forward(real* x, int* s, int length) {
matX->rowMax(*maxX_);
expX_->assign(*matX);
// subtract max to avoid overflow or underflow
expX_->mul(maxX_, ones_, (real)-1, (real)1);
expX_->mul(*maxX_, *ones_, (real)-1, (real)1);
expX_->exp2();

real* a = a_->getData();
Expand Down
26 changes: 13 additions & 13 deletions paddle/gserver/layers/LstmLayer.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -316,7 +316,7 @@ void LstmLayer::forwardSequence(int batchSize,
}
if (prevOutput_) {
frameGate->setData(lstmValue.gateValue);
frameGate->mul(prevOutput_, weight_->getW(), 1, 1);
frameGate->mul(*prevOutput_, *weight_->getW(), 1, 1);
}
}
AsyncGpuBlock asyncGpuBlock;
Expand All @@ -338,7 +338,7 @@ void LstmLayer::forwardSequence(int batchSize,
frameOutput->setData(lstmValue.outputValue);
nextFrame(reversed_, getSize());
frameGate->setData(lstmValue.gateValue);
frameGate->mul(frameOutput, weight_->getW(), 1, 1);
frameGate->mul(*frameOutput, *weight_->getW(), 1, 1);
}
}
if (n != numSequences - 1) {
Expand All @@ -348,7 +348,7 @@ void LstmLayer::forwardSequence(int batchSize,
if (!reversed_) {
if (!prevState_) lstmValue.prevStateValue = nullptr;
if (prevOutput_) {
frameGate->mul(frameOutput, weight_->getW(), 1, 1);
frameGate->mul(*frameOutput, *weight_->getW(), 1, 1);
}
} else {
lstmValue.prevStateValue = nullptr;
Expand Down Expand Up @@ -470,7 +470,7 @@ void LstmLayer::backwardSequence(int batchSize,
frameGate->setData(lstmGrad.gateGrad);
nextFrame(reversed_, getSize());
frameOutput->setData(lstmGrad.outputGrad);
frameOutput->mul(frameGate, weightT, 1, 1);
frameOutput->mul(*frameGate, *weightT, 1, 1);
} else {
nextFrame(reversed_, getSize());
}
Expand All @@ -479,14 +479,14 @@ void LstmLayer::backwardSequence(int batchSize,
if (weight_->getWGrad()) {
if (!reversed_) {
weight_->getWGrad()->mul(
output_.value->subMatrix(start, length - 1)->getTranspose(),
gate_.grad->subMatrix(start + 1, length - 1),
*output_.value->subMatrix(start, length - 1)->getTranspose(),
*gate_.grad->subMatrix(start + 1, length - 1),
1,
1);
} else {
weight_->getWGrad()->mul(
output_.value->subMatrix(start + 1, length - 1)->getTranspose(),
gate_.grad->subMatrix(start, length - 1),
*output_.value->subMatrix(start + 1, length - 1)->getTranspose(),
*gate_.grad->subMatrix(start, length - 1),
1,
1);
}
Expand Down Expand Up @@ -541,15 +541,15 @@ void LstmLayer::forwardBatch(int batchSize,

if (n != 0) {
MatrixPtr batch1 = batchValue_->getBatchValue(n - 1, batchSize);
gateValue->mul(batch1, weight_->getW(), 1, 1);
gateValue->mul(*batch1, *weight_->getW(), 1, 1);
} else if (prevOutput_) {
Matrix::resizeOrCreate(prevBatchOutput2_,
gateValue->getHeight(),
getSize(),
false,
useGpu_);
batchValue_->prevOutput2Batch(*prevOutput_, *prevBatchOutput2_);
gateValue->mul(prevBatchOutput2_, weight_->getW(), 1, 1);
gateValue->mul(*prevBatchOutput2_, *weight_->getW(), 1, 1);

batchValue_->prevOutput2Batch(*prevState_,
*totalState_->subMatrix(0, numSequences));
Expand Down Expand Up @@ -672,16 +672,16 @@ void LstmLayer::backwardBatch(int batchSize,

if (n != 0) {
MatrixPtr tmp = batchGrad_->getBatchValue(n - 1, batchSize);
tmp->mul(gateGrad, weightT, 1, 1);
tmp->mul(*gateGrad, *weightT, 1, 1);
}

if (n != 0 && weight_->getWGrad()) {
/* backward weight */
MatrixPtr outputValue = batchValue_->getBatchValue(n - 1, batchSize);
weight_->getWGrad()->mul(outputValue->getTranspose(), gateGrad, 1, 1);
weight_->getWGrad()->mul(*outputValue->getTranspose(), *gateGrad, 1, 1);
} else if (prevOutput_ && weight_->getWGrad()) {
weight_->getWGrad()->mul(
prevBatchOutput2_->getTranspose(), gateGrad, 1, 1);
*prevBatchOutput2_->getTranspose(), *gateGrad, 1, 1);
}
}
}
Expand Down
8 changes: 4 additions & 4 deletions paddle/gserver/layers/MDLstmLayer.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -547,7 +547,7 @@ void MDLstmLayer::forwardOneSequence(int start, CoordIterator& coordIter) {
if (coordIter.getPrePos(delays_, i, prePos)) {
int preOffset = coordIter.offset(prePos);
frameGate_[start + offset].value->mul(
frameOutput_[start + preOffset].value, weight_->getW(), 1.0, 1.0);
*frameOutput_[start + preOffset].value, *weight_->getW(), 1.0, 1.0);
}
}
forwardGate2OutputSequence(start, coordIter);
Expand Down Expand Up @@ -747,11 +747,11 @@ void MDLstmLayer::backwardOneSequence(int start, CoordIterator& coordIter) {
if (coordIter.getPrePos(delays_, i, prePos)) {
int preOffset = coordIter.offset(prePos);
frameOutput_[start + preOffset].grad->mul(
frameGate_[start + offset].grad, weightT, 1.0, 1.0);
*frameGate_[start + offset].grad, *weightT, 1.0, 1.0);
if (weight_->getWGrad()) {
weight_->getWGrad()->mul(
frameOutput_[start + preOffset].value->getTranspose(),
frameGate_[start + offset].grad,
*frameOutput_[start + preOffset].value->getTranspose(),
*frameGate_[start + offset].grad,
1.0,
1.0);
}
Expand Down
6 changes: 3 additions & 3 deletions paddle/gserver/layers/OuterProdLayer.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -96,7 +96,7 @@ void OuterProdLayer::forward(PassType passType) {
tmpRow0->setData(inV0->getData() + i * dim0);
tmpRow1->setData(inV1->getData() + i * dim1);

tmpMtx0->mul(tmpRow0->getTranspose(), tmpRow1);
tmpMtx0->mul(*tmpRow0->getTranspose(), *tmpRow1);
}
}
}
Expand All @@ -121,7 +121,7 @@ void OuterProdLayer::backward(const UpdateCallback& callback) {
tmpRow0->setData(inG0->getData() + i * dim0);
tmpRow1->setData(inV1->getData() + i * dim1);

tmpRow0->mul(tmpRow1, tmpMtx0->getTranspose(), 1, 1);
tmpRow0->mul(*tmpRow1, *tmpMtx0->getTranspose(), 1, 1);
}
}

Expand All @@ -131,7 +131,7 @@ void OuterProdLayer::backward(const UpdateCallback& callback) {
tmpRow0->setData(inV0->getData() + i * dim0);
tmpRow1->setData(inG1->getData() + i * dim1);

tmpRow1->mul(tmpRow0, tmpMtx0, 1, 1);
tmpRow1->mul(*tmpRow0, *tmpMtx0, 1, 1);
}
}
}
Expand Down
32 changes: 16 additions & 16 deletions paddle/gserver/layers/RecurrentLayer.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -215,12 +215,12 @@ void RecurrentLayer::forwardSequence(int batchSize,
void RecurrentLayer::forwardOneSequence(int start, int length) {
if (!reversed_) {
if (prevOutput_) {
frameOutput_[start].value->mul(prevOutput_, weight_->getW(), 1, 1);
frameOutput_[start].value->mul(*prevOutput_, *weight_->getW(), 1, 1);
}
activation_->forward(frameOutput_[start]);
for (int i = 1; i < length; ++i) {
frameOutput_[start + i].value->mul(
frameOutput_[start + i - 1].value, weight_->getW(), 1, 1);
*frameOutput_[start + i - 1].value, *weight_->getW(), 1, 1);
activation_->forward(frameOutput_[start + i]);
}
if (prevOutput_) {
Expand All @@ -230,7 +230,7 @@ void RecurrentLayer::forwardOneSequence(int start, int length) {
activation_->forward(frameOutput_[start + length - 1]);
for (int i = length - 2; i >= 0; --i) {
frameOutput_[start + i].value->mul(
frameOutput_[start + i + 1].value, weight_->getW(), 1, 1);
*frameOutput_[start + i + 1].value, *weight_->getW(), 1, 1);
activation_->forward(frameOutput_[start + i]);
}
}
Expand Down Expand Up @@ -282,27 +282,27 @@ void RecurrentLayer::backwardOneSequence(int start, int length) {
for (int i = length - 1; i > 0; --i) {
activation_->backward(frameOutput_[start + i]);
frameOutput_[start + i - 1].grad->mul(
frameOutput_[start + i].grad, weightT, 1, 1);
*frameOutput_[start + i].grad, *weightT, 1, 1);
}
activation_->backward(frameOutput_[start]);
if (weight_->getWGrad()) {
weight_->getWGrad()->mul(
output_.value->subMatrix(start, length - 1)->getTranspose(),
output_.grad->subMatrix(start + 1, length - 1),
*output_.value->subMatrix(start, length - 1)->getTranspose(),
*output_.grad->subMatrix(start + 1, length - 1),
1,
1);
}
} else {
for (int i = 0; i < length - 1; ++i) {
activation_->backward(frameOutput_[start + i]);
frameOutput_[start + i + 1].grad->mul(
frameOutput_[start + i].grad, weightT, 1, 1);
*frameOutput_[start + i].grad, *weightT, 1, 1);
}
activation_->backward(frameOutput_[start + length - 1]);
if (weight_->getWGrad()) {
weight_->getWGrad()->mul(
output_.value->subMatrix(start + 1, length - 1)->getTranspose(),
output_.grad->subMatrix(start, length - 1),
*output_.value->subMatrix(start + 1, length - 1)->getTranspose(),
*output_.grad->subMatrix(start, length - 1),
1,
1);
}
Expand All @@ -329,7 +329,7 @@ void RecurrentLayer::forwardBatch(int batchSize,
if (n != 0) {
MatrixPtr batch1 =
batchValue_->getBatchValue(n - 1, batch2->getHeight());
batch2->mul(batch1, weight_->getW(), 1, 1);
batch2->mul(*batch1, *weight_->getW(), 1, 1);
}
Argument arg;
arg.value = batch2;
Expand Down Expand Up @@ -367,14 +367,14 @@ void RecurrentLayer::backwardBatch(int batchSize,

if (n != 0) {
batch1 = batchGrad_->getBatchValue(n - 1, batch2->getHeight());
batch1->mul(batch2, weightT, 1, 1);
batch1->mul(*batch2, *weightT, 1, 1);
}

if (backwardByBatch && weight_->getWGrad()) {
if (n != 0) {
/* backward weight */
batch1 = batchValue_->getBatchValue(n - 1, batch2->getHeight());
weight_->getWGrad()->mul(batch1->getTranspose(), batch2, 1, 1);
weight_->getWGrad()->mul(*batch1->getTranspose(), *batch2, 1, 1);
}
}
}
Expand All @@ -389,14 +389,14 @@ void RecurrentLayer::backwardBatch(int batchSize,
int len = starts[seq + 1] - starts[seq];
if (!reversed_) {
weight_->getWGrad()->mul(
output_.value->subMatrix(starts[seq], len - 1)->getTranspose(),
output_.grad->subMatrix(starts[seq] + 1, len - 1),
*output_.value->subMatrix(starts[seq], len - 1)->getTranspose(),
*output_.grad->subMatrix(starts[seq] + 1, len - 1),
1,
1);
} else {
weight_->getWGrad()->mul(
output_.value->subMatrix(starts[seq] + 1, len - 1)->getTranspose(),
output_.grad->subMatrix(starts[seq], len - 1),
*output_.value->subMatrix(starts[seq] + 1, len - 1)->getTranspose(),
*output_.grad->subMatrix(starts[seq], len - 1),
1,
1);
}
Expand Down
Loading

0 comments on commit 4fbf949

Please sign in to comment.