potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-29 07:13:02 +04:00
Code Packages Releases Wiki Activity

dnn(ocl4dnn): add fusion support for Power activation and eltwise add

Browse Source 
Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>
This commit is contained in:
Wu Zhiwen
2017-11-20 11:29:18 +08:00
parent c7f1843584
commit 45d11dde57
6 changed files with 613 additions and 449 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/src/layers/convolution_layer.cpp
+39 -4
View File
@@ -142,6 +142,9 @@ public:
}
};
#define IS_POWER_LAYER(layer) \
(!layer.empty() && !layer->type.compare("Power"))
//TODO: simultaneously convolution and bias addition for cache optimization
class ConvolutionLayerImpl : public BaseConvolutionLayerImpl
{
@@ -161,6 +164,7 @@ public:
bool newWeightAndBias;
bool newActiv;
ocl4dnnFusedActiv_t activType;
float power;
#endif
ConvolutionLayerImpl()
{
@@ -169,6 +173,7 @@ public:
newWeightAndBias = false;
newActiv = false;
activType = OCL4DNN_CONV_FUSED_ACTIV_NONE;
power = 0.f;
#endif
}
@@ -225,6 +230,22 @@ public:
#ifdef HAVE_OPENCL
newActiv = true;
activType = OCL4DNN_CONV_FUSED_ACTIV_NONE;
if (preferableTarget == DNN_TARGET_OPENCL)
{
Ptr<PowerLayer> activ_power = activ.dynamicCast<PowerLayer>();
if (!activ_power.empty())
{
if (activ_power->scale != 1.f || activ_power->shift != 0.f)
newWeightAndBias = true;
if (activ_power->scale != 1.f)
weightsMat.release();
power = activ_power->power;
activType = OCL4DNN_CONV_FUSED_ACTIV_POWER;
}
}
#endif
return !activ.empty();
}
@@ -727,11 +748,12 @@ public:
biasvec[k] = biasMat.at<float>(k);
}
if( !bnorm.empty() || !scaleLayer.empty() )
if( !bnorm.empty() || !scaleLayer.empty() || IS_POWER_LAYER(activ))
{
Mat scale, shift, scale2, shift2;
const float *scaleptr = 0, *shiftptr = 0;
const float *scaleptr2 = 0, *shiftptr2 = 0;
float a = 1.f, b = 0.f;
if( !bnorm.empty() )
{
@@ -758,7 +780,14 @@ public:
}
}
if (shiftptr || shiftptr2)
if( IS_POWER_LAYER(activ) )
{
Ptr<PowerLayer> activ_power = activ.dynamicCast<PowerLayer>();
a = activ_power->scale;
b = activ_power->shift;
}
if (shiftptr || shiftptr2 || b != 0.f)
fusedBias = true;
for( int i = 0; i < outCn; i++ )
@@ -771,9 +800,9 @@ public:
int j, wcols = weightsMat.cols;
for( j = 0; j < wcols; j++ )
w_i[j] *= (s1*s2);
w_i[j] *= (s1*s2*a);
biasvec[i] = biasvec[i]*(s1*s2) + (delta1*s2 + delta2);
biasvec[i] = biasvec[i]*(s1*s2*a) + (delta1*s2*a + delta2*a + b);
}
}
biasvec[outCn] = biasvec[outCn+1] = biasvec[outCn-1];
@@ -827,10 +856,15 @@ public:
CV_Assert(!reluslope.empty());
convolutionOp->setActivPReLU(true, reluslope);
}
else if ( activType == OCL4DNN_CONV_FUSED_ACTIV_POWER)
{
convolutionOp->setActivPower(true, power);
}
else
{
convolutionOp->setActivReLU(false, 0);
convolutionOp->setActivPReLU(false, reluslope);
convolutionOp->setActivPower(false, 1.f);
}
newActiv = false;
}
@@ -840,6 +874,7 @@ public:
int batch_size = inpMat.size[0];
return convolutionOp->Forward(inpMat,
inputs.size() == 2 ? inputs[1] : UMat(),
umat_blobs[0],
(hasBias() || fusedBias) ? umat_blobs[1] : UMat(),
outMat,