conv2d.cpp

Go to the documentation of this file.

#define SHARK_USE_SIMD
#include <shark/LinAlg/BLAS/blas.h>
#include <shark/LinAlg/BLAS/kernels/conv2d.hpp>
#include <shark/Core/Timer.h>
#include <iostream>
using namespace shark;
using namespace std;

template<class E1, class E2>
void benchmark(
    blas::matrix_expression<E1, blas::cpu_tag> const& image,
    blas::matrix_expression<E2, blas::cpu_tag> const& filter,
    std::size_t num_channels,
    std::size_t num_filters
){
    std::size_t filter_size = filter().size2();
    std::size_t image_size1 = image().size1()/num_channels;
    std::size_t image_size2 = image().size2();
    std::size_t output_size1 = image_size1 - filter_size +1;
    std::size_t output_size2 = image_size2 - filter_size +1;
    typedef typename E1::value_type value_type;
    
    blas::matrix<value_type> out(output_size1 * num_filters, output_size2 ,0.0);
    double minOptTime = std::numeric_limits<double>::max();
    for(std::size_t i = 0; i != 20; ++i){
        Timer time;
        blas::kernels::conv2d(image,filter,out, num_channels, num_filters);
        minOptTime = min(minOptTime,time.stop());
    }
    
    double mults = output_size1 * output_size2 * filter_size * filter_size * num_filters * num_channels;
    double flops = mults /1024/1024/minOptTime;
    
    std::cout<<output_size1<<"\t"<<filter_size<<"\t"<<num_channels<<"\t"<< num_filters<<"\t";
    std::cout<<"\t"<<flops<< std::endl;
}


int main(int argc, char **argv) {
    std::cout<<"Flops"<<std::endl;
    std::size_t num_channels = 8;
    std::size_t num_outputs = 16;
    std::cout<<"performance float"<<std::endl;
    for(std::size_t filterSize = 4; filterSize != 32; filterSize *= 2){
        for(std::size_t iter = 0; iter != 6; ++iter){
            std::size_t sizeOut1 = (3+16 * 2<<iter);
            std::size_t sizeOut2 = (3+16 * 2<<iter);
            std::size_t sizeIm1 = sizeOut1 + filterSize-1;
            std::size_t sizeIm2 = sizeOut2 + filterSize-1;

            blas::matrix<float> image(num_channels * sizeIm1 , sizeIm2);
            blas::matrix<float> filter(num_channels * num_outputs *  filterSize, filterSize);
            
            for(std::size_t i = 0; i != num_channels * sizeIm1; ++i){
                for(std::size_t j = 0; j != sizeIm2; ++j){
                    image(i,j)  = 1.0/(num_channels * sizeOut1)*i + 0.1 - (0.1/sizeOut2)*j;
                }
            }
            for(std::size_t i = 0; i != num_channels * num_outputs * filterSize; ++i){
                for(std::size_t j = 0; j != filterSize; ++j){
                    filter(i,j)  = 1.0/(num_channels * filterSize)*i + 0.1 - (0.1/filterSize)*j;
                }
            }
            
            benchmark(image,filter,num_channels,num_outputs);           
        }
    }
    num_outputs = 8;
    std::cout<<"performance double"<<std::endl;
    for(std::size_t filterSize = 4; filterSize != 32; filterSize *= 2){
        for(std::size_t iter = 0; iter != 6; ++iter){
            std::size_t sizeOut1 = (3+16 * 2<<iter);
            std::size_t sizeOut2 = (3+16 * 2<<iter);
            std::size_t sizeIm1 = sizeOut1 + filterSize-1;
            std::size_t sizeIm2 = sizeOut2 + filterSize-1;

            blas::matrix<double> image(num_channels * sizeIm1 , sizeIm2);
            blas::matrix<double> filter(num_channels * num_outputs *  filterSize, filterSize);
            
            for(std::size_t i = 0; i != num_channels * sizeIm1; ++i){
                for(std::size_t j = 0; j != sizeIm2; ++j){
                    image(i,j)  = 1.0/(num_channels * sizeOut1)*i + 0.1 - (0.1/sizeOut2)*j;
                }
            }
            for(std::size_t i = 0; i != num_channels * num_outputs * filterSize; ++i){
                for(std::size_t j = 0; j != filterSize; ++j){
                    filter(i,j)  = 1.0/(num_channels * filterSize)*i + 0.1 - (0.1/filterSize)*j;
                }
            }
            
            benchmark(image,filter,num_channels,num_outputs);           
        }
    }
}