added init() function into MultiGpuMgr, added samples

modules/gpu/include/opencv2/gpu/gpu.hpp

@@ -131,15 +131,16 @@ namespace cv
 
         /////////////////////////// Multi GPU Manager //////////////////////////////
 
-        // Provides functionality for working with many GPUs. Object of this
-        // class must be created before any OpenCV GPU call and no call must
-        // be done after its destruction.
+        // Provides functionality for working with many GPUs
         class CV_EXPORTS MultiGpuMgr
         {
         public:
             MultiGpuMgr();
             ~MultiGpuMgr();
 
+            // Must be called before any other GPU calls
+            void init();
+
             // Makes the given GPU active
             void gpuOn(int gpu_id);
 

modules/gpu/src/multi_gpu_mgr.cpp

@@ -48,6 +48,7 @@ namespace cv { namespace gpu {
 
 class MultiGpuMgr::Impl {};
 MultiGpuMgr::MultiGpuMgr() { throw_nogpu(); }
+void MultiGpuMgr::init() { throw_nogpu(); }
 void MultiGpuMgr::gpuOn(int) { throw_nogpu(); }
 void MultiGpuMgr::gpuOff() { throw_nogpu(); }
 
@@ -79,6 +80,8 @@ public:
 
     void gpuOn(int gpu_id)
     {
+        if (gpu_id < 0 || gpu_id >= num_devices_)
+            CV_Error(CV_StsBadArg, "MultiGpuMgr::gpuOn: GPU ID is out of range");
         cuSafeCall(cuCtxPushCurrent(contexts_[gpu_id]));
     }
 
@@ -105,6 +108,8 @@ MultiGpuMgr::Impl::Impl(): num_devices_(0)
     num_devices_ = getCudaEnabledDeviceCount();
     contexts_.resize(num_devices_);
 
+    cuSafeCall(cuInit(0));
+
     CUdevice device;
     CUcontext prev_context;
     for (int i = 0; i < num_devices_; ++i)
@@ -116,18 +121,28 @@ MultiGpuMgr::Impl::Impl(): num_devices_(0)
 }
 
 
-MultiGpuMgr::MultiGpuMgr(): impl_(new Impl()) {}
+MultiGpuMgr::MultiGpuMgr() {}
 MultiGpuMgr::~MultiGpuMgr() {}
 
 
+void MultiGpuMgr::init()
+{
+    impl_ = Ptr<Impl>(new Impl());
+}
+
+
 void MultiGpuMgr::gpuOn(int gpu_id)
 {
+    if (impl_.empty())
+        CV_Error(CV_StsNullPtr, "MultiGpuMgr::gpuOn: must be initialized before any calls");
     impl_->gpuOn(gpu_id);
 }
 
 
 void MultiGpuMgr::gpuOff()
 {
+    if (impl_.empty())
+        CV_Error(CV_StsNullPtr, "MultiGpuMgr::gpuOff: must be initialized before any calls");
     impl_->gpuOff();
 }
 

samples/gpu/driver_api_stereo_multi.cpp

@@ -120,25 +120,25 @@ int main(int argc, char** argv)
     // Init CUDA Driver API
     safeCall(cuInit(0));
 
-    // Create context for the first GPU
+    // Create context for GPU #0
     CUdevice device;
     safeCall(cuDeviceGet(&device, 0));
     safeCall(cuCtxCreate(&contexts[0], 0, device));
     contextOff();
 
-    // Create context for the second GPU
+    // Create context for GPU #1
     safeCall(cuDeviceGet(&device, 1));
     safeCall(cuCtxCreate(&contexts[1], 0, device));
     contextOff();
 
-    // Split source images for processing on the first GPU
+    // 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));
     bm[0] = new StereoBM_GPU();
     contextOff();
 
-    // Split source images for processing on the second GPU
+    // 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));

samples/gpu/multi.cpp

@@ -0,0 +1,102 @@
+/* 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 <cvconfig.h>
+#include <opencv2/core/core.hpp>
+#include <opencv2/gpu/gpu.hpp>
+
+#if !defined(HAVE_CUDA) || !defined(HAVE_TBB)
+
+int main()
+{
+#if !defined(HAVE_CUDA)
+    cout << "CUDA support is required (CMake key 'WITH_CUDA' must be true).\n";
+#endif
+
+#if !defined(HAVE_TBB)
+    cout << "TBB support is required (CMake key 'WITH_TBB' must be true).\n";
+#endif
+
+    return 0;
+}
+
+#else
+
+#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; };
+
+MultiGpuMgr multi_gpu_mgr;
+
+int main()
+{
+    int num_devices = getCudaEnabledDeviceCount();
+    if (num_devices < 2)
+    {
+        cout << "Two or more GPUs are required\n";
+        return -1;
+    }
+    for (int i = 0; i < num_devices; ++i)
+    {
+        DeviceInfo dev_info(i);
+        if (!dev_info.isCompatible())
+        {
+            cout << "GPU module isn't built for GPU #" << i << " ("
+                 << dev_info.name() << ", CC " << dev_info.majorVersion()
+                 << dev_info.minorVersion() << "\n";
+            return -1;
+        }
+    }
+
+    multi_gpu_mgr.init();
+
+    // Execute calculation in two threads using two GPUs
+    int devices[] = {0, 2};
+    parallel_do(devices, devices + 2, Worker());
+
+    return 0;
+}
+
+
+void Worker::operator()(int device_id) const
+{
+    multi_gpu_mgr.gpuOn(device_id);
+
+    Mat src(1000, 1000, CV_32F);
+    Mat dst;
+
+    RNG rng(0);
+    rng.fill(src, RNG::UNIFORM, 0, 1);
+
+    // CPU works
+    transpose(src, dst);
+
+    // GPU works
+    GpuMat d_src(src);
+    GpuMat d_dst;
+    transpose(d_src, d_dst);
+
+    // Check results
+    bool passed = norm(dst - Mat(d_dst), NORM_INF) < 1e-3;
+    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();
+
+    multi_gpu_mgr.gpuOff();
+}
+
+#endif

samples/gpu/stereo_multi.cpp

@@ -0,0 +1,148 @@
+/* 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>
+#include <cvconfig.h>
+#include <opencv2/core/core.hpp>
+#include <opencv2/highgui/highgui.hpp>
+#include <opencv2/gpu/gpu.hpp>
+
+#if !defined(HAVE_CUDA) || !defined(HAVE_TBB)
+
+int main()
+{
+#if !defined(HAVE_CUDA)
+    cout << "CUDA support is required (CMake key 'WITH_CUDA' must be true).\n";
+#endif
+
+#if !defined(HAVE_TBB)
+    cout << "TBB support is required (CMake key 'WITH_TBB' must be true).\n";
+#endif
+
+    return 0;
+}
+
+#else
+
+#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; };
+
+MultiGpuMgr multi_gpu_mgr;
+
+// GPUs data
+GpuMat d_left[2];
+GpuMat d_right[2];
+StereoBM_GPU* bm[2];
+GpuMat d_result[2];
+
+// CPU result
+Mat result;
+
+int main(int argc, char** argv)
+{
+    if (argc < 3)
+    {
+        cout << "Usage: stereo_multi_gpu <left_image> <right_image>\n";
+        return -1;
+    }
+
+    int num_devices = getCudaEnabledDeviceCount();
+    if (num_devices < 2)
+    {
+        cout << "Two or more GPUs are required\n";
+        return -1;
+    }
+    for (int i = 0; i < num_devices; ++i)
+    {
+        DeviceInfo dev_info(i);
+        if (!dev_info.isCompatible())
+        {
+            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 = imread(argv[1], CV_LOAD_IMAGE_GRAYSCALE);
+    Mat right = imread(argv[2], CV_LOAD_IMAGE_GRAYSCALE);
+    if (left.empty())
+    {
+        cout << "Cannot open '" << argv[1] << "'\n";
+        return -1;
+    }
+    if (right.empty())
+    {
+        cout << "Cannot open '" << argv[2] << "'\n";
+        return -1;
+    }
+
+    multi_gpu_mgr.init();
+
+    // Split source images for processing on the GPU #0
+    multi_gpu_mgr.gpuOn(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();
+    multi_gpu_mgr.gpuOff();
+
+    // Split source images for processing on the GPU #1
+    multi_gpu_mgr.gpuOn(1);
+    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();
+    multi_gpu_mgr.gpuOff();
+
+    // Execute calculation in two threads using two GPUs
+    int devices[] = {0, 1};
+    parallel_do(devices, devices + 2, Worker());
+
+    // Release the first GPU resources
+    multi_gpu_mgr.gpuOn(0);
+    imshow("GPU #0 result", Mat(d_result[0]));
+    d_left[0].release();
+    d_right[0].release();
+    d_result[0].release();
+    delete bm[0];
+    multi_gpu_mgr.gpuOff();
+
+    // Release the second GPU resources
+    multi_gpu_mgr.gpuOn(1);
+    imshow("GPU #1 result", Mat(d_result[1]));
+    d_left[1].release();
+    d_right[1].release();
+    d_result[1].release();
+    delete bm[1];
+    multi_gpu_mgr.gpuOff();
+
+    waitKey();
+    return 0;
+}
+
+
+void Worker::operator()(int device_id) const
+{
+    multi_gpu_mgr.gpuOn(device_id);
+
+    bm[device_id]->operator()(d_left[device_id], d_right[device_id],
+                              d_result[device_id]);
+
+    cout << "GPU #" << device_id << " (" << DeviceInfo().name()
+        << "): finished\n";
+
+    multi_gpu_mgr.gpuOff();
+}
+
+#endif