LMU/DKS
5
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

OpenCL FFT using clfft and tests

Browse Source
This commit is contained in:
Uldis Locans
2016-11-21 16:14:11 +01:00
parent 4432d32480
commit b3a12e02a8
14 changed files with 482 additions and 390 deletions
Download Patch File Download Diff File Expand all files Collapse all files

149
test/testFFT3DRC.cpp
View File

@ -1,6 +1,8 @@
#include <iostream>
#include <cstdlib>
#include <complex>
#include <fstream>
#include <iomanip>
#include "Utility/TimeStamp.h"
#include "DKSBase.h"
@ -8,14 +10,20 @@
using namespace std;
void compareData(double* data1, double* data2, int NI, int NJ, int NK, int dim);
void initData(double *data, int dimsize[3]);
bool readParams(int argc, char *argv[], int &N1, int &N2, int &N3, int &loop,
char *api_name, char *device_name);
void initData(double *data, int dimsize[3], int dim);
bool readParams(int argc, char *argv[], int &N1, int &N2, int &N3, int &loop, int &dim,
char *api_name, char *device_name, char *file_name);
void printHelp();
void printData3DN4(complex<double>* &data, int N, int dim);
void printData3DN4(double* &data, int N, int dim);
double precision(double a) {
//if (a < 1e-10)
// return 0.0;
//else
return a;
}
int main(int argc, char *argv[]) {
@ -26,32 +34,29 @@ int main(int argc, char *argv[]) {
int loop = 0;
char *api_name = new char[10];
char *device_name = new char[10];
char *file_name = new char[50];
if ( readParams(argc, argv, N1, N2, N3, loop, api_name, device_name) )
if ( readParams(argc, argv, N1, N2, N3, loop, dim, api_name, device_name, file_name) )
return 0;
cout << "Use api: " << api_name << ", " << device_name << endl;
int dimsize[3] = {N3, N2, N1};
int dimsize[3] = {N1, N2, N3};
int sizereal = dimsize[0] * dimsize[1] * dimsize[2];
int sizecomp = (dimsize[0]/2+1) * dimsize[1] *dimsize[2];
double *rdata = new double[sizereal];
double *outdata = new double[sizereal];
complex<double> *cfft = new complex<double>[sizecomp];
initData(rdata, dimsize);
initData(rdata, dimsize, dim);
/* init DKSBase */
cout << "Init device and set function" << endl;
DKSBase base;
DKSBase base;
base.setAPI(api_name, strlen(api_name));
base.setDevice(device_name, strlen(device_name));
base.initDevice();
base.setupFFT(3, dimsize);
base.setupFFTRC(dim, dimsize);
/* setup backward fft (COMPLEX->REAL) */
base.setupFFTCR(dim, dimsize,1./(N1*N2*N3));
base.setupFFT(dim, dimsize);
// allocate memory on device
int ierr;
@ -63,68 +68,59 @@ DKSBase base;
// execute one run before starting the timers
base.writeData<double>(real_ptr, rdata, sizereal);
base.callR2CFFT(real_ptr, comp_ptr, dim, dimsize);
base.readData< complex<double> >(comp_ptr, cfft, sizecomp);
base.callC2RFFT(real_res_ptr, comp_ptr, dim, dimsize);
base.callNormalizeC2RFFT(real_res_ptr, dim, dimsize);
base.readData<double>(real_res_ptr, outdata, sizereal);
//timer for total loop time, FFT and IFFT calls
struct timeval timeStart, timeEnd;
struct timeval timeFFTStart[loop], timeFFTEnd[loop];
struct timeval timeIFFTStart[loop], timeIFFTEnd[loop];
gettimeofday(&timeStart, NULL);
for (int i=0; i<loop; ++i){
// write data to device
base.writeData<double>(real_ptr, rdata, sizereal);
// execute rcfft
gettimeofday(&timeFFTStart[i], NULL);
base.callR2CFFT(real_ptr, comp_ptr, dim, dimsize);
gettimeofday(&timeFFTEnd[i], NULL);
// execute crfft
gettimeofday(&timeIFFTStart[i], NULL);
base.callC2RFFT(real_res_ptr, comp_ptr, dim, dimsize);
gettimeofday(&timeIFFTEnd[i], NULL);
//normalize
base.callNormalizeC2RFFT(real_res_ptr, dim, dimsize);
// read IFFT data from device
base.readData<double>(real_res_ptr, outdata, sizereal);
ofstream myfile;
myfile.open(file_name);
myfile<< "in\tout\treal\timag\n";
for (int i = 0; i < sizereal; i++) {
//myfile << precision(rdata[i]) << "\t";
//myfile << precision(outdata[i]) << "\t";
if (i < sizecomp) {
myfile << precision(cfft[i].real()) << "\t";
myfile << precision(cfft[i].imag());
}
myfile << "\n";
}
gettimeofday(&timeEnd, NULL);
myfile.close();
/*
if (dim == 2) {
for (int i = 0; i < N2; i++) {
for (int j = 0; j < N1; j++) {
cout << rdata[i*N1 + j] << " ";
}
cout << endl;
}
cout << endl;
}
if (dim == 2) {
for (int i = 0; i < N2; i++) {
for (int j = 0; j < N1 / 2 + 1; j++) {
cout << cfft[i*(N1 / 2 + 1) + j] << " ";
}
cout << endl;
}
cout << endl;
}
*/
// free device memory
base.freeMemory< std::complex<double> >(comp_ptr, sizecomp);
base.freeMemory<double>(real_ptr, sizereal);
base.freeMemory<double>(real_res_ptr, sizereal);
// compare in and out data to see if we get back the same results
cout << "comp" << endl;
compareData(rdata, outdata, N1, N2, N3, dim);
//calculate seconds for total time and fft times
double tfft = 0;
double tifft = 0;
double ttot = ( (timeEnd.tv_sec - timeStart.tv_sec) * 1e6 +
(timeEnd.tv_usec - timeStart.tv_usec) ) * 1e-6;
for (int i = 0; i < loop; i++) {
tfft += ( (timeFFTEnd[i].tv_sec - timeFFTStart[i].tv_sec) * 1e6 +
(timeFFTEnd[i].tv_usec - timeFFTStart[i].tv_usec) ) * 1e-6;
tifft += ( (timeIFFTEnd[i].tv_sec - timeIFFTStart[i].tv_sec) * 1e6 +
(timeIFFTEnd[i].tv_usec - timeIFFTStart[i].tv_usec) ) * 1e-6;
}
//print timing results
std::cout << std::fixed << std::setprecision(5) << "\nTiming results"
<< "\nTotal time\t" << ttot << "s\tavg time\t" << ttot / loop << "s"
<< "\nFFT total\t" << tfft << "s\tFFT avg \t" << tfft / loop << "s"
<< "\nIFFT total\t" << tifft << "s\tIFFT avg\t" << tifft / loop << "s"
<< "\n\n";
cout << "done" << endl;
return 0;
}
@ -132,10 +128,10 @@ DKSBase base;
void compareData(double* data1, double* data2, int NI, int NJ, int NK, int dim) {
int id;
double sum = 0;
for (int i = 0; i < NI; i++) {
for (int i = 0; i < NK; i++) {
for (int j = 0; j < NJ; j++) {
for (int k = 0; k < NK; k++) {
id = k*NI*NJ + j*NI + i;
for (int k = 0; k < NI; k++) {
id = i*NI*NJ + j*NI + k;
sum += fabs(data1[id] - data2[id]);
}
}
@ -143,13 +139,21 @@ void compareData(double* data1, double* data2, int NI, int NJ, int NK, int dim)
std::cout << "RC <--> CR diff: " << sum << std::endl;
}
void initData(double *data, int dimsize[3]) {
for (int i = 0; i < dimsize[2]; i++) {
void initData(double *data, int dimsize[3], int dim) {
if (dim == 3) {
for (int i = 0; i < dimsize[2]; i++)
for (int j = 0; j < dimsize[1]; j++)
for (int k = 0; k < dimsize[0]; k++)
data[i*dimsize[1]*dimsize[0] + j*dimsize[0] + k] = sin(k);
} else if (dim == 2) {
for (int j = 0; j < dimsize[1]; j++) {
for (int k = 0; k < dimsize[0]; k++) {
data[i*dimsize[1]*dimsize[0] + j*dimsize[0] + k] = k;
data[j*dimsize[0] + k] = sin(k);
}
}
} else {
for (int k = 0; k < dimsize[0]; k++)
data[k] = sin(k);
}
}
@ -168,12 +172,17 @@ void printHelp() {
std::cout << std::endl;
}
bool readParams(int argc, char *argv[], int &N1, int &N2, int &N3, int &loop,
char *api_name, char *device_name)
bool readParams(int argc, char *argv[], int &N1, int &N2, int &N3, int &loop, int &dim,
char *api_name, char *device_name, char *file_name)
{
for (int i = 1; i < argc; i++) {
if ( argv[i] == std::string("-dim")) {
dim = atoi(argv[i + 1]);
i++;
}
if ( argv[i] == std::string("-grid") ) {
N1 = atoi(argv[i + 1]);
N2 = atoi(argv[i + 2]);
@ -194,21 +203,25 @@ bool readParams(int argc, char *argv[], int &N1, int &N2, int &N3, int &loop,
if (argv[i] == string("-cuda")) {
strcpy(api_name, "Cuda");
strcpy(device_name, "-gpu");
strcpy(file_name, "cuda_fft.dat");
}
if (argv[i] == string("-opencl")) {
strcpy(api_name, "OpenCL");
strcpy(device_name, "-gpu");
strcpy(file_name, "opencl_fft.dat");
}
if (argv[i] == string("-mic")) {
strcpy(api_name, "OpenMP");
strcpy(device_name, "-mic");
strcpy(file_name, "openmp_fft.dat");
}
if (argv[i] == string("-cpu")) {
strcpy(api_name, "OpenCL");
strcpy(device_name, "-cpu");
strcpy(file_name, "opencl_cpu_fft.dat");
}
}