added enqueueHostCallback method to gpu::Stream

modules/gpu/doc/data_structures.rst

@@ -271,41 +271,37 @@ This class encapsulates a queue of asynchronous calls. Some functions have overl
     class CV_EXPORTS Stream
     {
     public:
-            Stream();
-            ~Stream();
+        Stream();
+        ~Stream();
 
-            Stream(const Stream&);
-            Stream& operator=(const Stream&);
+        Stream(const Stream&);
+        Stream& operator=(const Stream&);
 
-            bool queryIfComplete();
-            void waitForCompletion();
+        bool queryIfComplete();
+        void waitForCompletion();
 
-            //! downloads asynchronously.
-            // Warning! cv::Mat must point to page locked memory
-                     (i.e. to CudaMem data or to its subMat)
-            void enqueueDownload(const GpuMat& src, CudaMem& dst);
-            void enqueueDownload(const GpuMat& src, Mat& dst);
+        void enqueueDownload(const GpuMat& src, CudaMem& dst);
+        void enqueueDownload(const GpuMat& src, Mat& dst);
 
-            //! uploads asynchronously.
-            // Warning! cv::Mat must point to page locked memory
-                     (i.e. to CudaMem data or to its ROI)
-            void enqueueUpload(const CudaMem& src, GpuMat& dst);
-            void enqueueUpload(const Mat& src, GpuMat& dst);
+        void enqueueUpload(const CudaMem& src, GpuMat& dst);
+        void enqueueUpload(const Mat& src, GpuMat& dst);
 
-            void enqueueCopy(const GpuMat& src, GpuMat& dst);
+        void enqueueCopy(const GpuMat& src, GpuMat& dst);
 
-            void enqueueMemSet(const GpuMat& src, Scalar val);
-            void enqueueMemSet(const GpuMat& src, Scalar val, const GpuMat& mask);
+        void enqueueMemSet(const GpuMat& src, Scalar val);
+        void enqueueMemSet(const GpuMat& src, Scalar val, const GpuMat& mask);
 
-            // converts matrix type, ex from float to uchar depending on type
-            void enqueueConvert(const GpuMat& src, GpuMat& dst, int type,
-                    double a = 1, double b = 0);
+        void enqueueConvert(const GpuMat& src, GpuMat& dst, int type,
+                            double a = 1, double b = 0);
+
+        typedef void (*StreamCallback)(Stream& stream, int status, void* userData);
+        void enqueueHostCallback(StreamCallback callback, void* userData);
     };
 
 
 
 gpu::Stream::queryIfComplete
---------------------------------
+----------------------------
 Returns ``true`` if the current stream queue is finished. Otherwise, it returns false.
 
 .. ocv:function:: bool gpu::Stream::queryIfComplete()
@@ -313,13 +309,73 @@ Returns ``true`` if the current stream queue is finished. Otherwise, it returns
 
 
 gpu::Stream::waitForCompletion
-----------------------------------
+------------------------------
 Blocks the current CPU thread until all operations in the stream are complete.
 
 .. ocv:function:: void gpu::Stream::waitForCompletion()
 
 
 
+gpu::Stream::enqueueDownload
+----------------------------
+Copies data from device to host.
+
+.. ocv:function:: void gpu::Stream::enqueueDownload(const GpuMat& src, CudaMem& dst)
+
+.. ocv:function:: void gpu::Stream::enqueueDownload(const GpuMat& src, Mat& dst)
+
+.. note:: ``cv::Mat`` must point to page locked memory (i.e. to ``CudaMem`` data or to its subMat) or must be registered with :ocv:func:`gpu::registerPageLocked` .
+
+
+
+gpu::Stream::enqueueUpload
+--------------------------
+Copies data from host to device.
+
+.. ocv:function:: void gpu::Stream::enqueueUpload(const CudaMem& src, GpuMat& dst)
+
+.. ocv:function:: void gpu::Stream::enqueueUpload(const Mat& src, GpuMat& dst)
+
+.. note:: ``cv::Mat`` must point to page locked memory (i.e. to ``CudaMem`` data or to its subMat) or must be registered with :ocv:func:`gpu::registerPageLocked` .
+
+
+
+gpu::Stream::enqueueCopy
+------------------------
+Copies data from device to device.
+
+.. ocv:function:: void gpu::Stream::enqueueCopy(const GpuMat& src, GpuMat& dst)
+
+
+
+gpu::Stream::enqueueMemSet
+--------------------------
+Initializes or sets device memory to a value.
+
+.. ocv:function:: void gpu::Stream::enqueueMemSet(const GpuMat& src, Scalar val)
+
+.. ocv:function:: void gpu::Stream::enqueueMemSet(const GpuMat& src, Scalar val, const GpuMat& mask)
+
+
+
+gpu::Stream::enqueueConvert
+---------------------------
+Converts matrix type, ex from float to uchar depending on type.
+
+.. ocv:function:: void gpu::Stream::enqueueConvert(const GpuMat& src, GpuMat& dst, int type, double a = 1, double b = 0)
+
+
+
+gpu::Stream::enqueueHostCallback
+--------------------------------
+Adds a callback to be called on the host after all currently enqueued items in the stream have completed.
+
+.. ocv:function:: void gpu::Stream::enqueueHostCallback(StreamCallback callback, void* userData)
+
+.. note:: Callbacks must not make any CUDA API calls. Callbacks must not perform any synchronization that may depend on outstanding device work or other callbacks that are not mandated to run earlier.  Callbacks without a mandated order (in independent streams) execute in undefined order and may be serialized.
+
+
+
 gpu::StreamAccessor
 -------------------
 .. ocv:struct:: gpu::StreamAccessor

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

@@ -145,43 +145,49 @@ public:
     ~Stream();
 
     Stream(const Stream&);
-    Stream& operator=(const Stream&);
+    Stream& operator =(const Stream&);
 
     bool queryIfComplete();
     void waitForCompletion();
 
-    //! downloads asynchronously.
+    //! downloads asynchronously
     // Warning! cv::Mat must point to page locked memory (i.e. to CudaMem data or to its subMat)
     void enqueueDownload(const GpuMat& src, CudaMem& dst);
     void enqueueDownload(const GpuMat& src, Mat& dst);
 
-    //! uploads asynchronously.
+    //! uploads asynchronously
     // Warning! cv::Mat must point to page locked memory (i.e. to CudaMem data or to its ROI)
     void enqueueUpload(const CudaMem& src, GpuMat& dst);
     void enqueueUpload(const Mat& src, GpuMat& dst);
 
+    //! copy asynchronously
     void enqueueCopy(const GpuMat& src, GpuMat& dst);
 
+    //! memory set asynchronously
     void enqueueMemSet(GpuMat& src, Scalar val);
     void enqueueMemSet(GpuMat& src, Scalar val, const GpuMat& mask);
 
-    // converts matrix type, ex from float to uchar depending on type
-    void enqueueConvert(const GpuMat& src, GpuMat& dst, int type, double a = 1, double b = 0);
+    //! converts matrix type, ex from float to uchar depending on type
+    void enqueueConvert(const GpuMat& src, GpuMat& dst, int dtype, double a = 1, double b = 0);
+
+    //! adds a callback to be called on the host after all currently enqueued items in the stream have completed
+    typedef void (*StreamCallback)(Stream& stream, int status, void* userData);
+    void enqueueHostCallback(StreamCallback callback, void* userData);
 
     static Stream& Null();
 
     operator bool() const;
 
 private:
+    struct Impl;
+
+    explicit Stream(Impl* impl);
     void create();
     void release();
 
-    struct Impl;
     Impl *impl;
 
     friend struct StreamAccessor;
-
-    explicit Stream(Impl* impl);
 };
 
 

modules/gpu/src/cudastream.cpp

@@ -42,51 +42,37 @@
 
 #include "precomp.hpp"
 
+using namespace std;
 using namespace cv;
 using namespace cv::gpu;
 
-#if defined HAVE_CUDA
-
-struct Stream::Impl
-{
-    static cudaStream_t getStream(const Impl* impl) { return impl ? impl->stream : 0; }
-    cudaStream_t stream;
-    int ref_counter;
-};
-
-#include "opencv2/gpu/stream_accessor.hpp"
-
-CV_EXPORTS cudaStream_t cv::gpu::StreamAccessor::getStream(const Stream& stream)
-{
-    return Stream::Impl::getStream(stream.impl);
-};
-
-#endif /* !defined (HAVE_CUDA) */
-
-
 #if !defined (HAVE_CUDA)
 
-void cv::gpu::Stream::create() { throw_nogpu(); }
-void cv::gpu::Stream::release() { throw_nogpu(); }
-cv::gpu::Stream::Stream() : impl(0) { throw_nogpu(); }
-cv::gpu::Stream::~Stream() { throw_nogpu(); }
-cv::gpu::Stream::Stream(const Stream& /*stream*/) { throw_nogpu(); }
-Stream& cv::gpu::Stream::operator=(const Stream& /*stream*/) { throw_nogpu(); return *this; }
-bool cv::gpu::Stream::queryIfComplete() { throw_nogpu(); return true; }
+cv::gpu::Stream::Stream() { throw_nogpu(); }
+cv::gpu::Stream::~Stream() {}
+cv::gpu::Stream::Stream(const Stream&) { throw_nogpu(); }
+Stream& cv::gpu::Stream::operator=(const Stream&) { throw_nogpu(); return *this; }
+bool cv::gpu::Stream::queryIfComplete() { throw_nogpu(); return false; }
 void cv::gpu::Stream::waitForCompletion() { throw_nogpu(); }
-void cv::gpu::Stream::enqueueDownload(const GpuMat& /*src*/, Mat& /*dst*/) { throw_nogpu(); }
-void cv::gpu::Stream::enqueueDownload(const GpuMat& /*src*/, CudaMem& /*dst*/) { throw_nogpu(); }
-void cv::gpu::Stream::enqueueUpload(const CudaMem& /*src*/, GpuMat& /*dst*/) { throw_nogpu(); }
-void cv::gpu::Stream::enqueueUpload(const Mat& /*src*/, GpuMat& /*dst*/) { throw_nogpu(); }
-void cv::gpu::Stream::enqueueCopy(const GpuMat& /*src*/, GpuMat& /*dst*/) { throw_nogpu(); }
-void cv::gpu::Stream::enqueueMemSet(GpuMat& /*src*/, Scalar /*val*/) { throw_nogpu(); }
-void cv::gpu::Stream::enqueueMemSet(GpuMat& /*src*/, Scalar /*val*/, const GpuMat& /*mask*/) { throw_nogpu(); }
-void cv::gpu::Stream::enqueueConvert(const GpuMat& /*src*/, GpuMat& /*dst*/, int /*type*/, double /*a*/, double /*b*/) { throw_nogpu(); }
+void cv::gpu::Stream::enqueueDownload(const GpuMat&, Mat&) { throw_nogpu(); }
+void cv::gpu::Stream::enqueueDownload(const GpuMat&, CudaMem&) { throw_nogpu(); }
+void cv::gpu::Stream::enqueueUpload(const CudaMem&, GpuMat&) { throw_nogpu(); }
+void cv::gpu::Stream::enqueueUpload(const Mat&, GpuMat&) { throw_nogpu(); }
+void cv::gpu::Stream::enqueueCopy(const GpuMat&, GpuMat&) { throw_nogpu(); }
+void cv::gpu::Stream::enqueueMemSet(GpuMat&, Scalar) { throw_nogpu(); }
+void cv::gpu::Stream::enqueueMemSet(GpuMat&, Scalar, const GpuMat&) { throw_nogpu(); }
+void cv::gpu::Stream::enqueueConvert(const GpuMat&, GpuMat&, int, double, double) { throw_nogpu(); }
+void cv::gpu::Stream::enqueueHostCallback(StreamCallback, void*) { throw_nogpu(); }
 Stream& cv::gpu::Stream::Null() { throw_nogpu(); static Stream s; return s; }
 cv::gpu::Stream::operator bool() const { throw_nogpu(); return false; }
+cv::gpu::Stream::Stream(Impl*) { throw_nogpu(); }
+void cv::gpu::Stream::create() { throw_nogpu(); }
+void cv::gpu::Stream::release() { throw_nogpu(); }
 
 #else /* !defined (HAVE_CUDA) */
 
+#include "opencv2/gpu/stream_accessor.hpp"
+
 namespace cv { namespace gpu
 {
     void copyWithMask(const GpuMat& src, GpuMat& dst, const GpuMat& mask, cudaStream_t stream);
@@ -95,63 +81,55 @@ namespace cv { namespace gpu
     void setTo(GpuMat& src, Scalar s, const GpuMat& mask, cudaStream_t stream);
 }}
 
-namespace
+struct Stream::Impl
 {
-    template<class S, class D> void devcopy(const S& src, D& dst, cudaStream_t s, cudaMemcpyKind k)
+    static cudaStream_t getStream(const Impl* impl)
     {
-        dst.create(src.size(), src.type());
-        size_t bwidth = src.cols * src.elemSize();
-        cudaSafeCall( cudaMemcpy2DAsync(dst.data, dst.step, src.data, src.step, bwidth, src.rows, k, s) );
-    };
-}
-
-void cv::gpu::Stream::create()
-{
-    if (impl)
-        release();
+        return impl ? impl->stream : 0;
+    }
 
     cudaStream_t stream;
-    cudaSafeCall( cudaStreamCreate( &stream ) );
-
-    impl = (Stream::Impl*)fastMalloc(sizeof(Stream::Impl));
+    int ref_counter;
+};
 
-    impl->stream = stream;
-    impl->ref_counter = 1;
+cudaStream_t cv::gpu::StreamAccessor::getStream(const Stream& stream)
+{
+    return Stream::Impl::getStream(stream.impl);
 }
 
-void cv::gpu::Stream::release()
+cv::gpu::Stream::Stream() : impl(0)
 {
-    if( impl && CV_XADD(&impl->ref_counter, -1) == 1 )
-    {
-        cudaSafeCall( cudaStreamDestroy( impl->stream ) );
-        cv::fastFree( impl );
-    }
+    create();
 }
 
-cv::gpu::Stream::Stream() : impl(0) { create(); }
-cv::gpu::Stream::~Stream() { release(); }
+cv::gpu::Stream::~Stream()
+{
+    release();
+}
 
 cv::gpu::Stream::Stream(const Stream& stream) : impl(stream.impl)
 {
-    if( impl )
+    if (impl)
         CV_XADD(&impl->ref_counter, 1);
 }
-Stream& cv::gpu::Stream::operator=(const Stream& stream)
+
+Stream& cv::gpu::Stream::operator =(const Stream& stream)
 {
-    if( this != &stream )
+    if (this != &stream)
     {
-        if( stream.impl )
-            CV_XADD(&stream.impl->ref_counter, 1);
-
         release();
         impl = stream.impl;
+        if (impl)
+            CV_XADD(&impl->ref_counter, 1);
     }
+
     return *this;
 }
 
 bool cv::gpu::Stream::queryIfComplete()
 {
-    cudaError_t err = cudaStreamQuery( Impl::getStream(impl) );
+    cudaStream_t stream = Impl::getStream(impl);
+    cudaError_t err = cudaStreamQuery(stream);
 
     if (err == cudaErrorNotReady || err == cudaSuccess)
         return err == cudaSuccess;
@@ -160,94 +138,213 @@ bool cv::gpu::Stream::queryIfComplete()
     return false;
 }
 
-void cv::gpu::Stream::waitForCompletion() { cudaSafeCall( cudaStreamSynchronize( Impl::getStream(impl) ) ); }
+void cv::gpu::Stream::waitForCompletion()
+{
+    cudaStream_t stream = Impl::getStream(impl);
+    cudaSafeCall( cudaStreamSynchronize(stream) );
+}
 
 void cv::gpu::Stream::enqueueDownload(const GpuMat& src, Mat& dst)
 {
     // if not -> allocation will be done, but after that dst will not point to page locked memory
-    CV_Assert(src.cols == dst.cols && src.rows == dst.rows && src.type() == dst.type() );
-    devcopy(src, dst, Impl::getStream(impl), cudaMemcpyDeviceToHost);
+    CV_Assert( src.size() == dst.size() && src.type() == dst.type() );
+
+    cudaStream_t stream = Impl::getStream(impl);
+    size_t bwidth = src.cols * src.elemSize();
+    cudaSafeCall( cudaMemcpy2DAsync(dst.data, dst.step, src.data, src.step, bwidth, src.rows, cudaMemcpyDeviceToHost, stream) );
+}
+
+void cv::gpu::Stream::enqueueDownload(const GpuMat& src, CudaMem& dst)
+{
+    dst.create(src.size(), src.type(), CudaMem::ALLOC_PAGE_LOCKED);
+
+    cudaStream_t stream = Impl::getStream(impl);
+    size_t bwidth = src.cols * src.elemSize();
+    cudaSafeCall( cudaMemcpy2DAsync(dst.data, dst.step, src.data, src.step, bwidth, src.rows, cudaMemcpyDeviceToHost, stream) );
+}
+
+void cv::gpu::Stream::enqueueUpload(const CudaMem& src, GpuMat& dst)
+{
+    dst.create(src.size(), src.type());
+
+    cudaStream_t stream = Impl::getStream(impl);
+    size_t bwidth = src.cols * src.elemSize();
+    cudaSafeCall( cudaMemcpy2DAsync(dst.data, dst.step, src.data, src.step, bwidth, src.rows, cudaMemcpyHostToDevice, stream) );
 }
-void cv::gpu::Stream::enqueueDownload(const GpuMat& src, CudaMem& dst) { devcopy(src, dst, Impl::getStream(impl), cudaMemcpyDeviceToHost); }
 
-void cv::gpu::Stream::enqueueUpload(const CudaMem& src, GpuMat& dst){ devcopy(src, dst, Impl::getStream(impl),   cudaMemcpyHostToDevice); }
-void cv::gpu::Stream::enqueueUpload(const Mat& src, GpuMat& dst)  { devcopy(src, dst, Impl::getStream(impl),   cudaMemcpyHostToDevice); }
-void cv::gpu::Stream::enqueueCopy(const GpuMat& src, GpuMat& dst) { devcopy(src, dst, Impl::getStream(impl), cudaMemcpyDeviceToDevice); }
+void cv::gpu::Stream::enqueueUpload(const Mat& src, GpuMat& dst)
+{
+    dst.create(src.size(), src.type());
+
+    cudaStream_t stream = Impl::getStream(impl);
+    size_t bwidth = src.cols * src.elemSize();
+    cudaSafeCall( cudaMemcpy2DAsync(dst.data, dst.step, src.data, src.step, bwidth, src.rows, cudaMemcpyHostToDevice, stream) );
+}
+
+void cv::gpu::Stream::enqueueCopy(const GpuMat& src, GpuMat& dst)
+{
+    dst.create(src.size(), src.type());
+
+    cudaStream_t stream = Impl::getStream(impl);
+    size_t bwidth = src.cols * src.elemSize();
+    cudaSafeCall( cudaMemcpy2DAsync(dst.data, dst.step, src.data, src.step, bwidth, src.rows, cudaMemcpyDeviceToDevice, stream) );
+}
 
-void cv::gpu::Stream::enqueueMemSet(GpuMat& src, Scalar s)
+void cv::gpu::Stream::enqueueMemSet(GpuMat& src, Scalar val)
 {
-    CV_Assert((src.depth() != CV_64F) ||
-        (TargetArchs::builtWith(NATIVE_DOUBLE) && DeviceInfo().supports(NATIVE_DOUBLE)));
+    const int sdepth = src.depth();
 
-    if (s[0] == 0.0 && s[1] == 0.0 && s[2] == 0.0 && s[3] == 0.0)
+    if (sdepth == CV_64F)
     {
-        cudaSafeCall( cudaMemset2DAsync(src.data, src.step, 0, src.cols * src.elemSize(), src.rows, Impl::getStream(impl)) );
+        if (!deviceSupports(NATIVE_DOUBLE))
+            CV_Error(CV_StsUnsupportedFormat, "The device doesn't support double");
+    }
+
+    cudaStream_t stream = Impl::getStream(impl);
+
+    if (val[0] == 0.0 && val[1] == 0.0 && val[2] == 0.0 && val[3] == 0.0)
+    {
+        cudaSafeCall( cudaMemset2DAsync(src.data, src.step, 0, src.cols * src.elemSize(), src.rows, stream) );
         return;
     }
-    if (src.depth() == CV_8U)
+
+    if (sdepth == CV_8U)
     {
         int cn = src.channels();
 
-        if (cn == 1 || (cn == 2 && s[0] == s[1]) || (cn == 3 && s[0] == s[1] && s[0] == s[2]) || (cn == 4 && s[0] == s[1] && s[0] == s[2] && s[0] == s[3]))
+        if (cn == 1 || (cn == 2 && val[0] == val[1]) || (cn == 3 && val[0] == val[1] && val[0] == val[2]) || (cn == 4 && val[0] == val[1] && val[0] == val[2] && val[0] == val[3]))
         {
-            int val = saturate_cast<uchar>(s[0]);
-            cudaSafeCall( cudaMemset2DAsync(src.data, src.step, val, src.cols * src.elemSize(), src.rows, Impl::getStream(impl)) );
+            int ival = saturate_cast<uchar>(val[0]);
+            cudaSafeCall( cudaMemset2DAsync(src.data, src.step, ival, src.cols * src.elemSize(), src.rows, stream) );
             return;
         }
     }
 
-    setTo(src, s, Impl::getStream(impl));
+    setTo(src, val, stream);
 }
 
 void cv::gpu::Stream::enqueueMemSet(GpuMat& src, Scalar val, const GpuMat& mask)
 {
-    CV_Assert((src.depth() != CV_64F) ||
-        (TargetArchs::builtWith(NATIVE_DOUBLE) && DeviceInfo().supports(NATIVE_DOUBLE)));
+    const int sdepth = src.depth();
+
+    if (sdepth == CV_64F)
+    {
+        if (!deviceSupports(NATIVE_DOUBLE))
+            CV_Error(CV_StsUnsupportedFormat, "The device doesn't support double");
+    }
 
     CV_Assert(mask.type() == CV_8UC1);
 
-    setTo(src, val, mask, Impl::getStream(impl));
+    cudaStream_t stream = Impl::getStream(impl);
+
+    setTo(src, val, mask, stream);
 }
 
-void cv::gpu::Stream::enqueueConvert(const GpuMat& src, GpuMat& dst, int rtype, double alpha, double beta)
+void cv::gpu::Stream::enqueueConvert(const GpuMat& src, GpuMat& dst, int dtype, double alpha, double beta)
 {
-    CV_Assert((src.depth() != CV_64F && CV_MAT_DEPTH(rtype) != CV_64F) ||
-        (TargetArchs::builtWith(NATIVE_DOUBLE) && DeviceInfo().supports(NATIVE_DOUBLE)));
+    if (dtype < 0)
+        dtype = src.type();
+    else
+        dtype = CV_MAKE_TYPE(CV_MAT_DEPTH(dtype), src.channels());
 
-    bool noScale = fabs(alpha-1) < std::numeric_limits<double>::epsilon() && fabs(beta) < std::numeric_limits<double>::epsilon();
+    const int sdepth = src.depth();
+    const int ddepth = CV_MAT_DEPTH(dtype);
 
-    if( rtype < 0 )
-        rtype = src.type();
-    else
-        rtype = CV_MAKETYPE(CV_MAT_DEPTH(rtype), src.channels());
+    if (sdepth == CV_64F || ddepth == CV_64F)
+    {
+        if (!deviceSupports(NATIVE_DOUBLE))
+            CV_Error(CV_StsUnsupportedFormat, "The device doesn't support double");
+    }
+
+    bool noScale = fabs(alpha - 1) < numeric_limits<double>::epsilon() && fabs(beta) < numeric_limits<double>::epsilon();
 
-    int sdepth = src.depth(), ddepth = CV_MAT_DEPTH(rtype);
-    if( sdepth == ddepth && noScale )
+    if (sdepth == ddepth && noScale)
     {
-        src.copyTo(dst);
+        enqueueCopy(src, dst);
         return;
     }
 
-    GpuMat temp;
-    const GpuMat* psrc = &src;
-    if( sdepth != ddepth && psrc == &dst )
-        psrc = &(temp = src);
+    dst.create(src.size(), dtype);
 
-    dst.create( src.size(), rtype );
-    convertTo(src, dst, alpha, beta, Impl::getStream(impl));
+    cudaStream_t stream = Impl::getStream(impl);
+    convertTo(src, dst, alpha, beta, stream);
 }
 
-cv::gpu::Stream::operator bool() const
+#if CUDA_VERSION >= 5000
+
+namespace
 {
-    return impl && impl->stream;
+    struct CallbackData
+    {
+        cv::gpu::Stream::StreamCallback callback;
+        void* userData;
+        Stream stream;
+    };
+
+    void CUDART_CB cudaStreamCallback(cudaStream_t, cudaError_t status, void* userData)
+    {
+        CallbackData* data = reinterpret_cast<CallbackData*>(userData);
+        data->callback(data->stream, static_cast<int>(status), data->userData);
+        delete data;
+    }
 }
 
-cv::gpu::Stream::Stream(Impl* impl_) : impl(impl_) {}
+#endif
+
+void cv::gpu::Stream::enqueueHostCallback(StreamCallback callback, void* userData)
+{
+#if CUDA_VERSION >= 5000
+    CallbackData* data = new CallbackData;
+    data->callback = callback;
+    data->userData = userData;
+    data->stream = *this;
+
+    cudaStream_t stream = Impl::getStream(impl);
+
+    cudaSafeCall( cudaStreamAddCallback(stream, cudaStreamCallback, data, 0) );
+#else
+    (void) callback;
+    (void) userData;
+    CV_Error(CV_StsNotImplemented, "This function requires CUDA 5.0");
+#endif
+}
 
 cv::gpu::Stream& cv::gpu::Stream::Null()
 {
-    static Stream s((Impl*)0);
+    static Stream s((Impl*) 0);
     return s;
 }
 
+cv::gpu::Stream::operator bool() const
+{
+    return impl && impl->stream;
+}
+
+cv::gpu::Stream::Stream(Impl* impl_) : impl(impl_)
+{
+}
+
+void cv::gpu::Stream::create()
+{
+    if (impl)
+        release();
+
+    cudaStream_t stream;
+    cudaSafeCall( cudaStreamCreate( &stream ) );
+
+    impl = (Stream::Impl*) fastMalloc(sizeof(Stream::Impl));
+
+    impl->stream = stream;
+    impl->ref_counter = 1;
+}
+
+void cv::gpu::Stream::release()
+{
+    if (impl && CV_XADD(&impl->ref_counter, -1) == 1)
+    {
+        cudaSafeCall( cudaStreamDestroy(impl->stream) );
+        cv::fastFree(impl);
+    }
+}
+
 #endif /* !defined (HAVE_CUDA) */

modules/gpu/test/test_stream.cpp

@@ -0,0 +1,130 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other GpuMaterials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or bpied warranties, including, but not limited to, the bpied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#include "test_precomp.hpp"
+
+#ifdef HAVE_CUDA
+
+#if CUDA_VERSION >= 5000
+
+struct Async : testing::TestWithParam<cv::gpu::DeviceInfo>
+{
+    cv::gpu::CudaMem src;
+    cv::gpu::GpuMat d_src;
+
+    cv::gpu::CudaMem dst;
+    cv::gpu::GpuMat d_dst;
+
+    virtual void SetUp()
+    {
+        cv::gpu::DeviceInfo devInfo = GetParam();
+        cv::gpu::setDevice(devInfo.deviceID());
+
+        cv::Mat m = randomMat(cv::Size(128, 128), CV_8UC1);
+        src.create(m.size(), m.type(), cv::gpu::CudaMem::ALLOC_PAGE_LOCKED);
+        m.copyTo(src.createMatHeader());
+    }
+};
+
+void checkMemSet(cv::gpu::Stream&, int status, void* userData)
+{
+    ASSERT_EQ(cudaSuccess, status);
+
+    Async* test = reinterpret_cast<Async*>(userData);
+
+    cv::Mat src = test->src;
+    cv::Mat dst = test->dst;
+
+    cv::Mat dst_gold = cv::Mat::zeros(src.size(), src.type());
+
+    ASSERT_MAT_NEAR(dst_gold, dst, 0);
+}
+
+GPU_TEST_P(Async, MemSet)
+{
+    cv::gpu::Stream stream;
+
+    d_dst.upload(src);
+
+    stream.enqueueMemSet(d_dst, cv::Scalar::all(0));
+    stream.enqueueDownload(d_dst, dst);
+
+    Async* test = this;
+    stream.enqueueHostCallback(checkMemSet, test);
+
+    stream.waitForCompletion();
+}
+
+void checkConvert(cv::gpu::Stream&, int status, void* userData)
+{
+    ASSERT_EQ(cudaSuccess, status);
+
+    Async* test = reinterpret_cast<Async*>(userData);
+
+    cv::Mat src = test->src;
+    cv::Mat dst = test->dst;
+
+    cv::Mat dst_gold;
+    src.convertTo(dst_gold, CV_32S);
+
+    ASSERT_MAT_NEAR(dst_gold, dst, 0);
+}
+
+GPU_TEST_P(Async, Convert)
+{
+    cv::gpu::Stream stream;
+
+    stream.enqueueUpload(src, d_src);
+    stream.enqueueConvert(d_src, d_dst, CV_32S);
+    stream.enqueueDownload(d_dst, dst);
+
+    Async* test = this;
+    stream.enqueueHostCallback(checkConvert, test);
+
+    stream.waitForCompletion();
+}
+
+INSTANTIATE_TEST_CASE_P(GPU_Stream, Async, ALL_DEVICES);
+
+#endif
+
+#endif // HAVE_CUDA