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 layer ocl implementation

Browse Source 
Signed-off-by: Li Peng <peng.li@intel.com>
This commit is contained in:
Li Peng
2018-01-10 21:56:42 +08:00
parent 7bc017601f
commit e77af4ae33
3 changed files with 201 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/src/layers/mvn_layer.cpp
+84
View File
@@ -43,6 +43,8 @@
#include "../precomp.hpp"
#include "layers_common.hpp"
#include <opencv2/dnn/shape_utils.hpp>
#include "math_functions.hpp"
#include "opencl_kernels_dnn.hpp"
namespace cv
{
@@ -60,11 +62,93 @@ public:
eps = params.get<double>("eps", 1e-9);
}
#ifdef HAVE_OPENCL
bool forward_ocl(InputArrayOfArrays inputs_, OutputArrayOfArrays outputs_, OutputArrayOfArrays internals_)
{
std::vector<UMat> inputs;
std::vector<UMat> outputs;
inputs_.getUMatVector(inputs);
outputs_.getUMatVector(outputs);
for (size_t inpIdx = 0; inpIdx < inputs.size(); inpIdx++)
{
UMat &inpBlob = inputs[inpIdx];
UMat &outBlob = outputs[inpIdx];
int splitDim = (acrossChannels) ? 1 : 2;
int i, newRows = 1;
for( i = 0; i < splitDim; i++ )
newRows *= inpBlob.size[i];
MatShape s = shape(newRows, inpBlob.total() / newRows);
UMat& inpMat = inpBlob;
UMat& outMat = outBlob;
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);
UMat tmpMat = UMat(s[0], s[1], CV_32F);
float alpha = 1.0f / s[1];
bool ret = ocl4dnn::ocl4dnnGEMV<float>(ocl4dnn::CblasNoTrans, s[0], s[1], alpha,
inpMat, 0, oneMat, 0, 0.0f, meanMat, 0);
if (!ret)
return false;
int number = (s[1] % 8 == 0) ? 8 : ((s[1] % 4 == 0) ? 4 : 1);
String buildopt = format("-DNUM=%d ", number);
String kname = format("calc_mean%d", number);
ocl::Kernel kernel(kname.c_str(), ocl::dnn::mvn_oclsrc, buildopt);
if (kernel.empty())
return false;
size_t global[] = { (size_t)s[0], (size_t)(s[1] / number) };
kernel.set(0, ocl::KernelArg::PtrReadOnly(inpMat));
kernel.set(1, (int)s[0]);
kernel.set(2, (int)s[1]);
kernel.set(3, ocl::KernelArg::PtrReadOnly(meanMat));
kernel.set(4, ocl::KernelArg::PtrWriteOnly(tmpMat));
ret = kernel.run(2, global, NULL, false);
if (!ret)
return false;
if (normVariance)
{
ret = ocl4dnn::ocl4dnnGEMV<float>(ocl4dnn::CblasNoTrans, s[0], s[1], alpha,
tmpMat, 0, oneMat, 0, 0.0f, devMat, 0);
if (!ret)
return false;
}
kname = format("mvn%d", number);
if (normVariance)
buildopt += "-DNORM_VARIANCE";
ocl::Kernel kernel1(kname.c_str(), ocl::dnn::mvn_oclsrc, buildopt);
if (kernel1.empty())
return false;
kernel1.set(0, ocl::KernelArg::PtrReadOnly(inpMat));
kernel1.set(1, (int)s[0]);
kernel1.set(2, (int)s[1]);
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));
ret = kernel1.run(2, global, NULL, false);
if (!ret)
return false;
}
return true;
}
#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);
}