potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 07:43:03 +04:00
Code Packages Releases Wiki Activity

mvn, batch_norm and relu layer fusion

Browse Source 
Signed-off-by: Li Peng <peng.li@intel.com>
This commit is contained in:
Li Peng
2018-01-23 23:52:41 +08:00
parent e15928b49e
commit 2493083935
4 changed files with 75 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/src/layers/mvn_layer.cpp
+49 -9
View File
@@ -60,6 +60,36 @@ public:
normVariance = params.get<bool>("normalize_variance", true);
acrossChannels = params.get<bool>("across_channels", false);
eps = params.get<double>("eps", 1e-9);
fuse_batch_norm = false;
fuse_relu = false;
relu_slope = 0.f;
}
Ptr<BatchNormLayer> bnorm;
Mat scale, shift;
UMat bnorm_weight, bnorm_bias;
bool fuse_batch_norm;
bool setBatchNorm(const Ptr<BatchNormLayer>& layer )
{
bnorm = layer;
fuse_batch_norm = !bnorm.empty() && (preferableTarget == DNN_TARGET_OPENCL);
return fuse_batch_norm;
}
Ptr<ReLULayer> activ_relu;
float relu_slope;
bool fuse_relu;
bool setActivation(const Ptr<ActivationLayer>& layer)
{
if (!layer.empty() && preferableTarget == DNN_TARGET_OPENCL)
{
activ_relu = layer.dynamicCast<ReLULayer>();
if( !activ_relu.empty() )
relu_slope = activ_relu->negativeSlope;
}
fuse_relu = !activ_relu.empty();
return fuse_relu;
}
#ifdef HAVE_OPENCL
@@ -71,19 +101,24 @@ public:
inputs_.getUMatVector(inputs);
outputs_.getUMatVector(outputs);
if( fuse_batch_norm && scale.empty())
{
bnorm->getScaleShift(scale, shift);
bnorm_weight = scale.getUMat(ACCESS_READ);
bnorm_bias = shift.getUMat(ACCESS_READ);
}
for (size_t inpIdx = 0; inpIdx < inputs.size(); inpIdx++)
{
UMat &inpBlob = inputs[inpIdx];
UMat &outBlob = outputs[inpIdx];
UMat &inpMat = inputs[inpIdx];
UMat &outMat = outputs[inpIdx];
int splitDim = (acrossChannels) ? 1 : 2;
int i, newRows = 1;
for( i = 0; i < splitDim; i++ )
newRows *= inpBlob.size[i];
newRows *= inpMat.size[i];
MatShape s = shape(newRows, inpBlob.total() / newRows);
UMat& inpMat = inpBlob;
UMat& outMat = outBlob;
MatShape s = shape(newRows, inpMat.total() / newRows);
UMat oneMat = UMat::ones(s[1], 1, CV_32F);
UMat meanMat = UMat(s[0], 1, CV_32F);
UMat devMat = UMat(s[0], 1, CV_32F);
@@ -121,8 +156,9 @@ public:
}
String kname = format("mvn%d", number);
if (normVariance)
buildopt += "-DNORM_VARIANCE";
buildopt += format("%s %s %s ", (normVariance) ? "-DNORM_VARIANCE" : "",
(fuse_batch_norm) ? "-DFUSE_BATCH_NORM" : "",
(fuse_relu) ? "-DFUSE_RELU" : "");
ocl::Kernel kernel1(kname.c_str(), ocl::dnn::mvn_oclsrc, buildopt);
if (kernel1.empty())
return false;
@@ -132,7 +168,11 @@ public:
kernel1.set(3, (float)eps);
kernel1.set(4, ocl::KernelArg::PtrReadOnly(meanMat));
kernel1.set(5, ocl::KernelArg::PtrReadOnly(devMat));
kernel1.set(6, ocl::KernelArg::PtrWriteOnly(outMat));
kernel1.set(6, ocl::KernelArg::PtrReadOnly(bnorm_weight));
kernel1.set(7, ocl::KernelArg::PtrReadOnly(bnorm_bias));
kernel1.set(8, (int)inpMat.size[1]);
kernel1.set(9, (float)relu_slope);
kernel1.set(10, ocl::KernelArg::PtrWriteOnly(outMat));
ret = kernel1.run(2, global, NULL, false);
if (!ret)
return false;