driver_api_stereo_multi.cpp 4.79 KB
Newer Older
wester committed
1 2 3 4 5 6 7 8 9 10
/* 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
#if defined(_MSC_VER)
#pragma warning(disable: 4201 4408 4100)
#endif

#include <iostream>
a  
Kai Westerkamp committed
11 12
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
wester committed
13
#include "opencv2/gpu/gpu.hpp"
wester committed
14

wester committed
15
#if defined(__arm__)
wester committed
16 17 18 19 20 21 22 23 24
int main()
{
    std::cout << "Unsupported for ARM CUDA library." << std::endl;
    return 0;
}
#else

#include <cuda.h>
#include <cuda_runtime.h>
wester committed
25
#include "opencv2/core/internal.hpp" // For TBB wrappers
wester committed
26 27 28

using namespace std;
using namespace cv;
wester committed
29
using namespace cv::gpu;
wester committed
30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61

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];

void inline contextOn(int id)
{
    safeCall(cuCtxPushCurrent(contexts[id]));
}

void inline contextOff()
{
    CUcontext prev_context;
    safeCall(cuCtxPopCurrent(&prev_context));
}

// GPUs data
GpuMat d_left[2];
GpuMat d_right[2];
wester committed
62
StereoBM_GPU* bm[2];
wester committed
63 64
GpuMat d_result[2];

wester committed
65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85

struct Worker: public ParallelLoopBody
{
    virtual void operator() (const Range& range) const
    {
        for (int device_id = range.start; device_id != range.end; ++device_id)
        {
            contextOn(device_id);

            bm[device_id]->operator()(d_left[device_id], d_right[device_id],
                                      d_result[device_id]);

            std::cout << "GPU #" << device_id << " (" << DeviceInfo().name()
            << "): finished\n";

            contextOff();

        }
    }
};

wester committed
86 87
static void printHelp()
{
a  
Kai Westerkamp committed
88
    std::cout << "Usage: driver_api_stereo_multi_gpu --left <left_image> --right <right_image>\n";
wester committed
89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107
}

int main(int argc, char** argv)
{
    if (argc < 5)
    {
        printHelp();
        return -1;
    }

    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)
    {
wester committed
108
        cv::gpu::printShortCudaDeviceInfo(i);
wester committed
109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125

        DeviceInfo dev_info(i);
        if (!dev_info.isCompatible())
        {
            std::cout << "GPU module isn't built for GPU #" << i << " ("
                 << dev_info.name() << ", CC " << dev_info.majorVersion()
                 << dev_info.minorVersion() << "\n";
            return -1;
        }
    }

    // Load input data
    Mat left, right;
    for (int i = 1; i < argc; ++i)
    {
        if (string(argv[i]) == "--left")
        {
wester committed
126
            left = imread(argv[++i], CV_LOAD_IMAGE_GRAYSCALE);
wester committed
127 128 129 130
            CV_Assert(!left.empty());
        }
        else if (string(argv[i]) == "--right")
        {
wester committed
131
            right = imread(argv[++i], CV_LOAD_IMAGE_GRAYSCALE);
wester committed
132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159
            CV_Assert(!right.empty());
        }
        else if (string(argv[i]) == "--help")
        {
            printHelp();
            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));
    contextOff();

    // Create context for GPU #1
    safeCall(cuDeviceGet(&device, 1));
    safeCall(cuCtxCreate(&contexts[1], 0, device));
    contextOff();

    // Split source images for processing on GPU #0
    contextOn(0);
    d_left[0].upload(left.rowRange(0, left.rows / 2));
    d_right[0].upload(right.rowRange(0, right.rows / 2));
wester committed
160
    bm[0] = new StereoBM_GPU();
wester committed
161 162 163 164 165 166
    contextOff();

    // Split source images for processing on the GPU #1
    contextOn(1);
    d_left[1].upload(left.rowRange(left.rows / 2, left.rows));
    d_right[1].upload(right.rowRange(right.rows / 2, right.rows));
wester committed
167
    bm[1] = new StereoBM_GPU();
wester committed
168 169 170
    contextOff();

    // Execute calculation in two threads using two GPUs
wester committed
171
    parallel_for_(cv::Range(0, 2), Worker());
wester committed
172 173 174 175 176 177 178

    // 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();
wester committed
179
    delete bm[0];
wester committed
180 181 182 183 184 185 186 187
    contextOff();

    // Release the second GPU resources
    contextOn(1);
    imshow("GPU #1 result", Mat(d_result[1]));
    d_left[1].release();
    d_right[1].release();
    d_result[1].release();
wester committed
188
    delete bm[1];
wester committed
189 190 191 192 193 194 195 196 197 198 199 200 201 202
    contextOff();

    waitKey();
    destroyContexts();
    return 0;
}

void destroyContexts()
{
    safeCall(cuCtxDestroy(contexts[0]));
    safeCall(cuCtxDestroy(contexts[1]));
}

#endif