162 lines
3.7 KiB
C++
162 lines
3.7 KiB
C++
/* This sample demonstrates the way you can perform independed tasks
|
|
on the different GPUs */
|
|
|
|
// Disable some warnings which are caused with CUDA headers
|
|
#if defined(_MSC_VER)
|
|
#pragma warning(disable: 4201 4408 4100)
|
|
#endif
|
|
|
|
#include <iostream>
|
|
#include "opencv2/cvconfig.h"
|
|
#include "opencv2/core.hpp"
|
|
#include "opencv2/cudaarithm.hpp"
|
|
|
|
#ifdef HAVE_TBB
|
|
# include "tbb/tbb_stddef.h"
|
|
# if TBB_VERSION_MAJOR*100 + TBB_VERSION_MINOR >= 202
|
|
# include "tbb/tbb.h"
|
|
# include "tbb/task.h"
|
|
# undef min
|
|
# undef max
|
|
# else
|
|
# undef HAVE_TBB
|
|
# endif
|
|
#endif
|
|
|
|
#if !defined(HAVE_CUDA) || !defined(HAVE_TBB) || defined(__arm__)
|
|
|
|
int main()
|
|
{
|
|
#if !defined(HAVE_CUDA)
|
|
std::cout << "CUDA support is required (CMake key 'WITH_CUDA' must be true).\n";
|
|
#endif
|
|
|
|
#if !defined(HAVE_TBB)
|
|
std::cout << "TBB support is required (CMake key 'WITH_TBB' must be true).\n";
|
|
#endif
|
|
|
|
#if defined(__arm__)
|
|
std::cout << "Unsupported for ARM CUDA library." << std::endl;
|
|
#endif
|
|
|
|
return 0;
|
|
}
|
|
|
|
#else
|
|
|
|
#include <cuda.h>
|
|
#include <cuda_runtime.h>
|
|
|
|
using namespace std;
|
|
using namespace cv;
|
|
using namespace cv::cuda;
|
|
|
|
struct Worker { void operator()(int device_id) const; };
|
|
void destroyContexts();
|
|
|
|
#define safeCall(expr) safeCall_(expr, #expr, __FILE__, __LINE__)
|
|
inline void safeCall_(int code, const char* expr, const char* file, int line)
|
|
{
|
|
if (code != CUDA_SUCCESS)
|
|
{
|
|
std::cout << "CUDA driver API error: code " << code << ", expr " << expr
|
|
<< ", file " << file << ", line " << line << endl;
|
|
destroyContexts();
|
|
exit(-1);
|
|
}
|
|
}
|
|
|
|
// Each GPU is associated with its own context
|
|
CUcontext contexts[2];
|
|
|
|
int main()
|
|
{
|
|
int num_devices = getCudaEnabledDeviceCount();
|
|
if (num_devices < 2)
|
|
{
|
|
std::cout << "Two or more GPUs are required\n";
|
|
return -1;
|
|
}
|
|
|
|
for (int i = 0; i < num_devices; ++i)
|
|
{
|
|
cv::cuda::printShortCudaDeviceInfo(i);
|
|
|
|
DeviceInfo dev_info(i);
|
|
if (!dev_info.isCompatible())
|
|
{
|
|
std::cout << "CUDA module isn't built for GPU #" << i << " ("
|
|
<< dev_info.name() << ", CC " << dev_info.majorVersion()
|
|
<< dev_info.minorVersion() << "\n";
|
|
return -1;
|
|
}
|
|
}
|
|
|
|
// Init CUDA Driver API
|
|
safeCall(cuInit(0));
|
|
|
|
// Create context for GPU #0
|
|
CUdevice device;
|
|
safeCall(cuDeviceGet(&device, 0));
|
|
safeCall(cuCtxCreate(&contexts[0], 0, device));
|
|
|
|
CUcontext prev_context;
|
|
safeCall(cuCtxPopCurrent(&prev_context));
|
|
|
|
// Create context for GPU #1
|
|
safeCall(cuDeviceGet(&device, 1));
|
|
safeCall(cuCtxCreate(&contexts[1], 0, device));
|
|
|
|
safeCall(cuCtxPopCurrent(&prev_context));
|
|
|
|
// Execute calculation in two threads using two GPUs
|
|
int devices[] = {0, 1};
|
|
tbb::parallel_do(devices, devices + 2, Worker());
|
|
|
|
destroyContexts();
|
|
return 0;
|
|
}
|
|
|
|
|
|
void Worker::operator()(int device_id) const
|
|
{
|
|
// Set the proper context
|
|
safeCall(cuCtxPushCurrent(contexts[device_id]));
|
|
|
|
Mat src(1000, 1000, CV_32F);
|
|
Mat dst;
|
|
|
|
RNG rng(0);
|
|
rng.fill(src, RNG::UNIFORM, 0, 1);
|
|
|
|
// CPU works
|
|
cv::transpose(src, dst);
|
|
|
|
// GPU works
|
|
GpuMat d_src(src);
|
|
GpuMat d_dst;
|
|
cuda::transpose(d_src, d_dst);
|
|
|
|
// Check results
|
|
bool passed = cv::norm(dst - Mat(d_dst), NORM_INF) < 1e-3;
|
|
std::cout << "GPU #" << device_id << " (" << DeviceInfo().name() << "): "
|
|
<< (passed ? "passed" : "FAILED") << endl;
|
|
|
|
// Deallocate data here, otherwise deallocation will be performed
|
|
// after context is extracted from the stack
|
|
d_src.release();
|
|
d_dst.release();
|
|
|
|
CUcontext prev_context;
|
|
safeCall(cuCtxPopCurrent(&prev_context));
|
|
}
|
|
|
|
|
|
void destroyContexts()
|
|
{
|
|
safeCall(cuCtxDestroy(contexts[0]));
|
|
safeCall(cuCtxDestroy(contexts[1]));
|
|
}
|
|
|
|
#endif
|