potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-29 15:23:05 +04:00
Code Packages Releases Wiki Activity

OpenCL GPU target for Inference Engine deep learning backend

Browse Source 
Enable FP16 GPU target for DL Inference Engine backend.
This commit is contained in:
Dmitry Kurtaev
2018-03-12 17:35:28 +03:00
parent 72cb06abf0
commit 709cf5d038
12 changed files with 196 additions and 169 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/src/layers/convolution_layer.cpp
+29 -9
View File
@@ -173,21 +173,21 @@ public:
std::vector<float> biasvec;
std::vector<float> reluslope;
Ptr<ActivationLayer> activ;
bool newWeightAndBias;
bool fusedBias;
#ifdef HAVE_OPENCL
Ptr<OCL4DNNConvSpatial<float> > convolutionOp;
std::vector<UMat> umat_blobs;
bool fusedBias;
bool newWeightAndBias;
bool newActiv;
ocl4dnnFusedActiv_t activType;
float power;
#endif
ConvolutionLayerImpl(const LayerParams &params) : BaseConvolutionLayerImpl(params)
{
#ifdef HAVE_OPENCL
fusedBias = false;
newWeightAndBias = false;
fusedBias = false;
#ifdef HAVE_OPENCL
newActiv = false;
activType = OCL4DNN_CONV_FUSED_ACTIV_NONE;
power = 0.f;
@@ -350,10 +350,8 @@ public:
biasvec[i] += b.at<float>(i);
}
#ifdef HAVE_OPENCL
newWeightAndBias = !w.empty() || !b.empty();
fusedBias = hasBias() || !b.empty();
#endif
biasvec[outCn] = biasvec[outCn+1] = biasvec[outCn-1];
}
@@ -433,9 +431,31 @@ public:
ieLayer->_dilation_y = dilation.height;
ieLayer->_group = group;
ieLayer->_weights = wrapToInfEngineBlob(blobs[0]);
if (hasBias())
ieLayer->_biases = wrapToInfEngineBlob(blobs[1]);
ieLayer->_weights = wrapToInfEngineBlob(blobs[0], InferenceEngine::Layout::OIHW);
if (newWeightAndBias)
{
if (weightsMat.isContinuous())
{
Mat fusedWeights = weightsMat.reshape(1, blobs[0].dims, blobs[0].size);
ieLayer->_weights = wrapToInfEngineBlob(fusedWeights, InferenceEngine::Layout::OIHW);
}
else
{
ieLayer->_weights = InferenceEngine::make_shared_blob<float>(
InferenceEngine::Precision::FP32, InferenceEngine::Layout::OIHW,
ieLayer->_weights->dims());
ieLayer->_weights->allocate();
Mat newWeights = infEngineBlobToMat(ieLayer->_weights).reshape(1, outCn);
Mat fusedWeights = weightsMat.colRange(0, newWeights.cols);
fusedWeights.copyTo(newWeights);
}
}
if (hasBias() || fusedBias)
{
Mat biasesMat({outCn}, CV_32F, &biasvec[0]);
ieLayer->_biases = wrapToInfEngineBlob(biasesMat, {outCn}, InferenceEngine::Layout::C);
}
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();