Implement algorithms for GPU and CPU without to duplicate the code
- OpenCV compiled with CUDA
- Eigen
example in main.cu and UserCode.cu :
template<class Target>
CUDA_BOTH void UserCode<Target>::addToAllElements(int iRow, int iCol, Matrix<float>& matrix, int i)
{
matrix(iRow, iCol) += i;
}
template<class Target>
CUDA_BOTH void UserCode<Target>::multiplyBy2AllElements(int iRow, int iCol, Matrix<float>& matrix)
{
matrix(iRow, iCol) *= 2.0f;
}
template<typename Target>
void profile(cv::Mat_<float>& toModify, const char* title)
{
using namespace std::chrono;
const auto start = high_resolution_clock::now();
{
using Pipeline = Pipeline<Target>;
using UserCode = UserCode<Target>;
typename Target::host::template Matrix<float> d_toModify;
Pipeline::upload(d_toModify, toModify);
for(int i = 0; i < 10; i++)
{
Pipeline::template parallel_for_each_cell<UserCode::addToAllElements>(d_toModify, i);
Pipeline::template parallel_for_each_cell<UserCode::multiplyBy2AllElements>(d_toModify);
}
Pipeline::download(toModify, d_toModify);
}
const auto end = high_resolution_clock::now();
const float elapsed = duration_cast<std::chrono::duration<float>>(end - start).count();
std::cout << "Elapsed for " << title << ": " << std::fixed << std::setprecision(2) << elapsed << std::endl;
};
int main()
{
const int rows = 1000;
const int cols = 10000;
cv::Mat_<float> cpuMat(rows, cols, 0.0f);
cv::Mat_<float> gpuMat(rows, cols, 0.0f);
profile<CPU>(cpuMat, "CPU");
profile<GPU>(gpuMat, "GPU");
std::cout << "diff = " << cv::norm(cpuMat - gpuMat) << std::endl;
}output :
Elapsed for CPU: 2.31
Elapsed for GPU: 0.20
diff = 0.00
Add your own methods to UserCode class.