2011-02-03 13:02:39 +01:00
|
|
|
/* This sample demonstrates working on one piece of data using two GPUs.
|
|
|
|
It splits input into two parts and processes them separately on different
|
|
|
|
GPUs. */
|
|
|
|
|
|
|
|
// Disable some warnings which are caused with CUDA headers
|
2011-02-14 14:13:05 +01:00
|
|
|
#if defined(_MSC_VER)
|
2011-02-03 13:02:39 +01:00
|
|
|
#pragma warning(disable: 4201 4408 4100)
|
2011-02-14 14:13:05 +01:00
|
|
|
#endif
|
2011-02-03 13:02:39 +01:00
|
|
|
|
|
|
|
#include <iostream>
|
2011-05-21 20:32:34 +02:00
|
|
|
#include "cvconfig.h"
|
|
|
|
#include "opencv2/core/core.hpp"
|
|
|
|
#include "opencv2/highgui/highgui.hpp"
|
|
|
|
#include "opencv2/gpu/gpu.hpp"
|
2011-02-03 13:02:39 +01:00
|
|
|
|
|
|
|
#if !defined(HAVE_CUDA) || !defined(HAVE_TBB)
|
|
|
|
|
|
|
|
int main()
|
|
|
|
{
|
|
|
|
#if !defined(HAVE_CUDA)
|
2011-02-16 07:36:15 +01:00
|
|
|
std::cout << "CUDA support is required (CMake key 'WITH_CUDA' must be true).\n";
|
2011-02-03 13:02:39 +01:00
|
|
|
#endif
|
|
|
|
|
|
|
|
#if !defined(HAVE_TBB)
|
2011-02-16 07:36:15 +01:00
|
|
|
std::cout << "TBB support is required (CMake key 'WITH_TBB' must be true).\n";
|
2011-02-03 13:02:39 +01:00
|
|
|
#endif
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
#else
|
|
|
|
|
|
|
|
#include <cuda.h>
|
|
|
|
#include <cuda_runtime.h>
|
|
|
|
#include "opencv2/core/internal.hpp" // For TBB wrappers
|
|
|
|
|
|
|
|
using namespace std;
|
|
|
|
using namespace cv;
|
|
|
|
using namespace cv::gpu;
|
|
|
|
|
|
|
|
struct Worker { void operator()(int device_id) const; };
|
|
|
|
void destroyContexts();
|
|
|
|
|
2011-02-14 16:09:45 +01:00
|
|
|
#define safeCall(expr) safeCall_(expr, #expr, __FILE__, __LINE__)
|
|
|
|
inline void safeCall_(int code, const char* expr, const char* file, int line)
|
|
|
|
{
|
2011-02-14 16:36:07 +01:00
|
|
|
if (code != CUDA_SUCCESS)
|
|
|
|
{
|
2011-02-16 07:36:15 +01:00
|
|
|
std::cout << "CUDA driver API error: code " << code << ", expr " << expr
|
2011-02-14 16:36:07 +01:00
|
|
|
<< ", file " << file << ", line " << line << endl;
|
|
|
|
destroyContexts();
|
|
|
|
exit(-1);
|
|
|
|
}
|
2011-02-03 13:02:39 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
// Each GPU is associated with its own context
|
2011-06-09 12:21:02 +02:00
|
|
|
CUcontext contexts[2];
|
2011-02-03 13:02:39 +01:00
|
|
|
|
2011-06-09 12:21:02 +02:00
|
|
|
void inline contextOn(int id)
|
2011-02-03 13:02:39 +01:00
|
|
|
{
|
|
|
|
safeCall(cuCtxPushCurrent(contexts[id]));
|
|
|
|
}
|
|
|
|
|
2011-06-09 12:21:02 +02:00
|
|
|
void inline contextOff()
|
2011-02-03 13:02:39 +01:00
|
|
|
{
|
|
|
|
CUcontext prev_context;
|
|
|
|
safeCall(cuCtxPopCurrent(&prev_context));
|
|
|
|
}
|
|
|
|
|
2011-02-04 09:16:09 +01:00
|
|
|
// GPUs data
|
2011-02-03 13:02:39 +01:00
|
|
|
GpuMat d_left[2];
|
|
|
|
GpuMat d_right[2];
|
|
|
|
StereoBM_GPU* bm[2];
|
|
|
|
GpuMat d_result[2];
|
2011-02-04 09:16:09 +01:00
|
|
|
|
|
|
|
// CPU result
|
2011-02-03 13:02:39 +01:00
|
|
|
Mat result;
|
|
|
|
|
2011-10-24 10:34:15 +02:00
|
|
|
void printHelp()
|
|
|
|
{
|
|
|
|
std::cout << "Usage: driver_api_stereo_multi_gpu --left <left_image> --right <right_image>\n";
|
|
|
|
}
|
|
|
|
|
2011-02-03 13:02:39 +01:00
|
|
|
int main(int argc, char** argv)
|
|
|
|
{
|
2011-10-24 10:34:15 +02:00
|
|
|
if (argc < 5)
|
2011-02-03 13:02:39 +01:00
|
|
|
{
|
2011-10-24 10:34:15 +02:00
|
|
|
printHelp();
|
2011-02-03 13:02:39 +01:00
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
int num_devices = getCudaEnabledDeviceCount();
|
2011-06-09 12:21:02 +02:00
|
|
|
if (num_devices < 2)
|
|
|
|
{
|
|
|
|
std::cout << "Two or more GPUs are required\n";
|
|
|
|
return -1;
|
|
|
|
}
|
2011-02-03 13:02:39 +01:00
|
|
|
|
|
|
|
for (int i = 0; i < num_devices; ++i)
|
|
|
|
{
|
2011-02-03 15:51:58 +01:00
|
|
|
DeviceInfo dev_info(i);
|
|
|
|
if (!dev_info.isCompatible())
|
2011-02-03 13:02:39 +01:00
|
|
|
{
|
2011-02-16 07:36:15 +01:00
|
|
|
std::cout << "GPU module isn't built for GPU #" << i << " ("
|
2011-02-03 15:51:58 +01:00
|
|
|
<< dev_info.name() << ", CC " << dev_info.majorVersion()
|
|
|
|
<< dev_info.minorVersion() << "\n";
|
2011-02-03 13:02:39 +01:00
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Load input data
|
2011-10-24 10:34:15 +02:00
|
|
|
Mat left, right;
|
|
|
|
for (int i = 1; i < argc; ++i)
|
2011-02-03 13:02:39 +01:00
|
|
|
{
|
2011-10-24 10:34:15 +02:00
|
|
|
if (string(argv[i]) == "--left")
|
|
|
|
{
|
|
|
|
left = imread(argv[++i], CV_LOAD_IMAGE_GRAYSCALE);
|
|
|
|
CV_Assert(!left.empty());
|
|
|
|
}
|
|
|
|
else if (string(argv[i]) == "--right")
|
|
|
|
{
|
|
|
|
right = imread(argv[++i], CV_LOAD_IMAGE_GRAYSCALE);
|
|
|
|
CV_Assert(!right.empty());
|
|
|
|
}
|
|
|
|
else if (string(argv[i]) == "--help")
|
|
|
|
{
|
|
|
|
printHelp();
|
|
|
|
return -1;
|
|
|
|
}
|
2011-02-03 13:02:39 +01:00
|
|
|
}
|
|
|
|
|
2011-10-24 10:34:15 +02:00
|
|
|
|
2011-02-10 15:33:02 +01:00
|
|
|
// Init CUDA Driver API
|
2011-02-03 13:02:39 +01:00
|
|
|
safeCall(cuInit(0));
|
|
|
|
|
2011-02-15 11:03:26 +01:00
|
|
|
// Create context for GPU #0
|
2011-02-03 13:02:39 +01:00
|
|
|
CUdevice device;
|
|
|
|
safeCall(cuDeviceGet(&device, 0));
|
2011-06-09 12:21:02 +02:00
|
|
|
safeCall(cuCtxCreate(&contexts[0], 0, device));
|
2011-02-03 13:02:39 +01:00
|
|
|
contextOff();
|
|
|
|
|
2011-06-09 12:21:02 +02:00
|
|
|
// Create context for GPU #1
|
|
|
|
safeCall(cuDeviceGet(&device, 1));
|
|
|
|
safeCall(cuCtxCreate(&contexts[1], 0, device));
|
|
|
|
contextOff();
|
2011-02-03 13:02:39 +01:00
|
|
|
|
2011-02-15 11:03:26 +01:00
|
|
|
// Split source images for processing on GPU #0
|
2011-02-03 13:02:39 +01:00
|
|
|
contextOn(0);
|
|
|
|
d_left[0].upload(left.rowRange(0, left.rows / 2));
|
|
|
|
d_right[0].upload(right.rowRange(0, right.rows / 2));
|
|
|
|
bm[0] = new StereoBM_GPU();
|
|
|
|
contextOff();
|
|
|
|
|
2011-02-15 11:03:26 +01:00
|
|
|
// Split source images for processing on the GPU #1
|
2011-06-09 12:21:02 +02:00
|
|
|
contextOn(1);
|
2011-02-03 13:02:39 +01:00
|
|
|
d_left[1].upload(left.rowRange(left.rows / 2, left.rows));
|
|
|
|
d_right[1].upload(right.rowRange(right.rows / 2, right.rows));
|
|
|
|
bm[1] = new StereoBM_GPU();
|
|
|
|
contextOff();
|
|
|
|
|
|
|
|
// Execute calculation in two threads using two GPUs
|
2011-06-09 12:21:02 +02:00
|
|
|
int devices[] = {0, 1};
|
|
|
|
parallel_do(devices, devices + 2, Worker());
|
2011-02-03 13:02:39 +01:00
|
|
|
|
|
|
|
// Release the first GPU resources
|
|
|
|
contextOn(0);
|
|
|
|
imshow("GPU #0 result", Mat(d_result[0]));
|
|
|
|
d_left[0].release();
|
|
|
|
d_right[0].release();
|
|
|
|
d_result[0].release();
|
|
|
|
delete bm[0];
|
|
|
|
contextOff();
|
|
|
|
|
|
|
|
// Release the second GPU resources
|
2011-06-09 12:21:02 +02:00
|
|
|
contextOn(1);
|
2011-02-03 13:02:39 +01:00
|
|
|
imshow("GPU #1 result", Mat(d_result[1]));
|
|
|
|
d_left[1].release();
|
|
|
|
d_right[1].release();
|
|
|
|
d_result[1].release();
|
|
|
|
delete bm[1];
|
|
|
|
contextOff();
|
|
|
|
|
|
|
|
waitKey();
|
|
|
|
destroyContexts();
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
void Worker::operator()(int device_id) const
|
|
|
|
{
|
2011-06-09 12:21:02 +02:00
|
|
|
contextOn(device_id);
|
2011-02-03 13:02:39 +01:00
|
|
|
|
|
|
|
bm[device_id]->operator()(d_left[device_id], d_right[device_id],
|
|
|
|
d_result[device_id]);
|
|
|
|
|
2011-02-16 07:36:15 +01:00
|
|
|
std::cout << "GPU #" << device_id << " (" << DeviceInfo().name()
|
2011-02-03 13:02:39 +01:00
|
|
|
<< "): finished\n";
|
|
|
|
|
|
|
|
contextOff();
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
void destroyContexts()
|
|
|
|
{
|
|
|
|
safeCall(cuCtxDestroy(contexts[0]));
|
2011-06-09 12:21:02 +02:00
|
|
|
safeCall(cuCtxDestroy(contexts[1]));
|
2011-02-03 13:02:39 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
#endif
|