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

Add new layer forward interface

Browse Source 
Add layer forward interface with InputArrayOfArrays and
OutputArrayOfArrays parameters, it allows UMat buffer to be
processed and transferred in the layers.

Signed-off-by: Li Peng <peng.li@intel.com>
This commit is contained in:
Li Peng
2017-11-09 12:57:37 +08:00
parent 6e4f9433d0
commit 8f99083726
31 changed files with 802 additions and 133 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/src/layers/convolution_layer.cpp
+137 -21
View File
@@ -671,14 +671,20 @@ public:
};
#ifdef HAVE_OPENCL
bool forward_ocl(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
{
int group = inputs[0]->size[1] / umat_blobs[0].size[1];
std::vector<UMat> inputs;
std::vector<UMat> outputs;
inps.getUMatVector(inputs);
outs.getUMatVector(outputs);
int group = inputs[0].size[1] / umat_blobs[0].size[1];
if (convolutionOp.empty())
{
OCL4DNNConvConfig config;
config.in_shape = shape(*inputs[0]);
config.in_shape = shape(inputs[0]);
config.out_shape = shape(outputs[0]);
config.kernel = kernel;
config.pad = pad;
@@ -690,6 +696,112 @@ public:
convolutionOp = Ptr<OCL4DNNConvSpatial<float> >(new OCL4DNNConvSpatial<float>(config));
}
int k, outCn = umat_blobs[0].size[0];
if( weightsMat.empty() )
{
// prepare weightsMat where each row is aligned and has enough zero padding on the right to
// use vectorized (i.e. with intrinsics) loops without tail processing
Mat wm = blobs[0].reshape(1, outCn).clone();
if( wm.step1() % VEC_ALIGN != 0 )
{
int newcols = (int)alignSize(wm.step1(), VEC_ALIGN);
Mat wm_buffer = Mat(outCn, newcols, wm.type());
Mat wm_padding = wm_buffer.colRange(wm.cols, newcols);
wm_padding.setTo(Scalar::all(0.));
Mat wm_aligned = wm_buffer.colRange(0, wm.cols);
wm.copyTo(wm_aligned);
wm = wm_aligned;
}
weightsMat = wm;
Mat biasMat = hasBias() ? blobs[1].reshape(1, outCn) : Mat();
biasvec.resize(outCn+2);
if( biasMat.empty() )
{
for( k = 0; k < outCn; k++ )
biasvec[k] = 0.f;
}
else
{
for( k = 0; k < outCn; k++ )
biasvec[k] = biasMat.at<float>(k);
}
if( !bnorm.empty() || !scaleLayer.empty() )
{
Mat scale, shift, scale2, shift2;
const float *scaleptr = 0, *shiftptr = 0;
const float *scaleptr2 = 0, *shiftptr2 = 0;
if( !bnorm.empty() )
{
bnorm->getScaleShift(scale, shift);
CV_Assert( scale.isContinuous() && shift.isContinuous() &&
scale.type() == CV_32F && shift.type() == CV_32F &&
scale.total() == (size_t)outCn &&
shift.total() == (size_t)outCn );
scaleptr = scale.ptr<float>();
shiftptr = shift.ptr<float>();
}
if( !scaleLayer.empty() )
{
scale2 = scaleLayer->blobs[0];
CV_Assert( scale2.isContinuous() && scale2.type() == CV_32F &&
scale2.total() == (size_t)outCn );
scaleptr2 = scale2.ptr<float>();
if( scaleLayer->hasBias )
{
shift2 = scaleLayer->blobs[1];
CV_Assert( shift2.isContinuous() && shift2.type() == CV_32F &&
shift2.total() == (size_t)outCn );
shiftptr2 = shift2.ptr<float>();
}
}
if (shiftptr || shiftptr2)
fusedBias = true;
for( int i = 0; i < outCn; i++ )
{
float s1 = scaleptr ? scaleptr[i] : 1.f;
float delta1 = shiftptr ? shiftptr[i] : 0.f;
float s2 = scaleptr2 ? scaleptr2[i] : 1.f;
float delta2 = shiftptr2 ? shiftptr2[i] : 0.f;
float* w_i = weightsMat.ptr<float>(i);
int j, wcols = weightsMat.cols;
for( j = 0; j < wcols; j++ )
w_i[j] *= (s1*s2);
biasvec[i] = biasvec[i]*(s1*s2) + (delta1*s2 + delta2);
}
}
biasvec[outCn] = biasvec[outCn+1] = biasvec[outCn-1];
}
reluslope.clear();
if( activ )
{
Ptr<ReLULayer> activ_relu = activ.dynamicCast<ReLULayer>();
if( !activ_relu.empty() )
{
reluslope.assign(outCn+2, activ_relu->negativeSlope);
activType = OCL4DNN_CONV_FUSED_ACTIV_RELU;
}
Ptr<ChannelsPReLULayer> activ_chprelu = activ.dynamicCast<ChannelsPReLULayer>();
if( !activ_chprelu.empty() )
{
const Mat& m = activ_chprelu->blobs[0];
CV_Assert(m.isContinuous() && m.type() == CV_32F && (int)m.total() == outCn);
const float* mdata = m.ptr<float>();
reluslope.resize(outCn+2);
std::copy(mdata, mdata + outCn, reluslope.begin());
reluslope[outCn] = reluslope[outCn+1] = reluslope[outCn-1];
activType = OCL4DNN_CONV_FUSED_ACTIV_PRELU;
}
}
if ( newWeightAndBias )
{
weightsMat.copyTo(umat_blobs[0]);
@@ -723,9 +835,8 @@ public:
newActiv = false;
}
UMat inpMat, outMat;
inpMat = inputs[0]->getUMat(ACCESS_READ);
outMat = outputs[0].getUMat(ACCESS_WRITE);
UMat& inpMat = inputs[0];
UMat& outMat = outputs[0];
int batch_size = inpMat.size[0];
return convolutionOp->Forward(inpMat,
@@ -736,6 +847,18 @@ public:
}
#endif
void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
{
CV_TRACE_FUNCTION();
CV_TRACE_ARG_VALUE(name, "name", name.c_str());
CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
forward_ocl(inputs_arr, outputs_arr, internals_arr))
Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
}
void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
{
CV_TRACE_FUNCTION();
@@ -811,11 +934,6 @@ public:
}
}
#ifdef HAVE_OPENCL
if (shiftptr || shiftptr2)
fusedBias = true;
#endif
for( int i = 0; i < outCn; i++ )
{
float s1 = scaleptr ? scaleptr[i] : 1.f;
@@ -841,9 +959,6 @@ public:
if( !activ_relu.empty() )
{
reluslope.assign(outCn+2, activ_relu->negativeSlope);
#ifdef HAVE_OPENCL
activType = OCL4DNN_CONV_FUSED_ACTIV_RELU;
#endif
}
Ptr<ChannelsPReLULayer> activ_chprelu = activ.dynamicCast<ChannelsPReLULayer>();
@@ -855,16 +970,9 @@ public:
reluslope.resize(outCn+2);
std::copy(mdata, mdata + outCn, reluslope.begin());
reluslope[outCn] = reluslope[outCn+1] = reluslope[outCn-1];
#ifdef HAVE_OPENCL
activType = OCL4DNN_CONV_FUSED_ACTIV_PRELU;
#endif
}
}
CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
forward_ocl(inputs, outputs, internals))
int nstripes = std::max(getNumThreads(), 1);
ParallelConv::run(*inputs[0], outputs[0], weightsMat, biasvec, reluslope,
@@ -1173,6 +1281,14 @@ public:
}
};
void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
{
CV_TRACE_FUNCTION();
CV_TRACE_ARG_VALUE(name, "name", name.c_str());
Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
}
void forward(std::vector<Mat *> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
{
CV_TRACE_FUNCTION();