Merge pull request #1575 from jet47:gpu-buffer-pool

modules/core/include/opencv2/core/cuda.hpp

@@ -61,16 +61,30 @@ namespace cv { namespace cuda {
 class CV_EXPORTS GpuMat
 {
 public:
+    class CV_EXPORTS Allocator
+    {
+    public:
+        virtual ~Allocator() {}
+
+        // allocator must fill data, step and refcount fields
+        virtual bool allocate(GpuMat* mat, int rows, int cols, size_t elemSize) = 0;
+        virtual void free(GpuMat* mat) = 0;
+    };
+
+    //! default allocator
+    static Allocator* defaultAllocator();
+    static void setDefaultAllocator(Allocator* allocator);
+
     //! default constructor
-    GpuMat();
+    explicit GpuMat(Allocator* allocator = defaultAllocator());
 
     //! constructs GpuMat of the specified size and type
-    GpuMat(int rows, int cols, int type);
-    GpuMat(Size size, int type);
+    GpuMat(int rows, int cols, int type, Allocator* allocator = defaultAllocator());
+    GpuMat(Size size, int type, Allocator* allocator = defaultAllocator());
 
     //! constucts GpuMat and fills it with the specified value _s
-    GpuMat(int rows, int cols, int type, Scalar s);
-    GpuMat(Size size, int type, Scalar s);
+    GpuMat(int rows, int cols, int type, Scalar s, Allocator* allocator = defaultAllocator());
+    GpuMat(Size size, int type, Scalar s, Allocator* allocator = defaultAllocator());
 
     //! copy constructor
     GpuMat(const GpuMat& m);
@@ -84,7 +98,7 @@ public:
     GpuMat(const GpuMat& m, Rect roi);
 
     //! builds GpuMat from host memory (Blocking call)
-    explicit GpuMat(InputArray arr);
+    explicit GpuMat(InputArray arr, Allocator* allocator = defaultAllocator());
 
     //! destructor - calls release()
     ~GpuMat();
@@ -249,6 +263,9 @@ public:
     //! helper fields used in locateROI and adjustROI
     uchar* datastart;
     uchar* dataend;
+
+    //! allocator
+    Allocator* allocator;
 };
 
 //! creates continuous matrix
@@ -260,6 +277,10 @@ CV_EXPORTS void ensureSizeIsEnough(int rows, int cols, int type, OutputArray arr
 
 CV_EXPORTS GpuMat allocMatFromBuf(int rows, int cols, int type, GpuMat& mat);
 
+//! BufferPool management (must be called before Stream creation)
+CV_EXPORTS void setBufferPoolUsage(bool on);
+CV_EXPORTS void setBufferPoolConfig(int deviceId, size_t stackSize, int stackCount);
+
 //////////////////////////////// CudaMem ////////////////////////////////
 
 // CudaMem is limited cv::Mat with page locked memory allocation.
@@ -382,6 +403,7 @@ private:
     Stream(const Ptr<Impl>& impl);
 
     friend struct StreamAccessor;
+    friend class BufferPool;
 };
 
 class CV_EXPORTS Event

modules/core/include/opencv2/core/cuda.inl.hpp

@@ -51,29 +51,29 @@ namespace cv { namespace cuda {
 //////////////////////////////// GpuMat ///////////////////////////////
 
 inline
-GpuMat::GpuMat()
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0)
+GpuMat::GpuMat(Allocator* allocator_)
+    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), allocator(allocator_)
 {}
 
 inline
-GpuMat::GpuMat(int rows_, int cols_, int type_)
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0)
+GpuMat::GpuMat(int rows_, int cols_, int type_, Allocator* allocator_)
+    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), allocator(allocator_)
 {
     if (rows_ > 0 && cols_ > 0)
         create(rows_, cols_, type_);
 }
 
 inline
-GpuMat::GpuMat(Size size_, int type_)
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0)
+GpuMat::GpuMat(Size size_, int type_, Allocator* allocator_)
+    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), allocator(allocator_)
 {
     if (size_.height > 0 && size_.width > 0)
         create(size_.height, size_.width, type_);
 }
 
 inline
-GpuMat::GpuMat(int rows_, int cols_, int type_, Scalar s_)
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0)
+GpuMat::GpuMat(int rows_, int cols_, int type_, Scalar s_, Allocator* allocator_)
+    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), allocator(allocator_)
 {
     if (rows_ > 0 && cols_ > 0)
     {
@@ -83,8 +83,8 @@ GpuMat::GpuMat(int rows_, int cols_, int type_, Scalar s_)
 }
 
 inline
-GpuMat::GpuMat(Size size_, int type_, Scalar s_)
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0)
+GpuMat::GpuMat(Size size_, int type_, Scalar s_, Allocator* allocator_)
+    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), allocator(allocator_)
 {
     if (size_.height > 0 && size_.width > 0)
     {
@@ -95,15 +95,15 @@ GpuMat::GpuMat(Size size_, int type_, Scalar s_)
 
 inline
 GpuMat::GpuMat(const GpuMat& m)
-    : flags(m.flags), rows(m.rows), cols(m.cols), step(m.step), data(m.data), refcount(m.refcount), datastart(m.datastart), dataend(m.dataend)
+    : flags(m.flags), rows(m.rows), cols(m.cols), step(m.step), data(m.data), refcount(m.refcount), datastart(m.datastart), dataend(m.dataend), allocator(m.allocator)
 {
     if (refcount)
         CV_XADD(refcount, 1);
 }
 
 inline
-GpuMat::GpuMat(InputArray arr) :
-    flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0)
+GpuMat::GpuMat(InputArray arr, Allocator* allocator_) :
+    flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), allocator(allocator_)
 {
     upload(arr);
 }

modules/core/include/opencv2/core/private.cuda.hpp

@@ -90,6 +90,38 @@ static inline void throw_no_cuda() { CV_Error(cv::Error::StsNotImplemented, "The
 
 namespace cv { namespace cuda
 {
+    class MemoryStack;
+
+    class CV_EXPORTS StackAllocator : public GpuMat::Allocator
+    {
+    public:
+        explicit StackAllocator(cudaStream_t stream);
+        ~StackAllocator();
+
+        bool allocate(GpuMat* mat, int rows, int cols, size_t elemSize);
+        void free(GpuMat* mat);
+
+    private:
+        StackAllocator(const StackAllocator&);
+        StackAllocator& operator =(const StackAllocator&);
+
+        cudaStream_t stream_;
+        MemoryStack* memStack_;
+        size_t alignment_;
+    };
+
+    class CV_EXPORTS BufferPool
+    {
+    public:
+        explicit BufferPool(Stream& stream);
+
+        GpuMat getBuffer(int rows, int cols, int type);
+        GpuMat getBuffer(Size size, int type) { return getBuffer(size.height, size.width, type); }
+
+    private:
+        GpuMat::Allocator* allocator_;
+    };
+
     static inline void checkNppError(int code, const char* file, const int line, const char* func)
     {
         if (code < 0)

modules/core/src/cuda/gpu_mat.cu

@@ -55,6 +55,54 @@ using namespace cv;
 using namespace cv::cuda;
 using namespace cv::cudev;
 
+namespace
+{
+    class DefaultAllocator : public GpuMat::Allocator
+    {
+    public:
+        bool allocate(GpuMat* mat, int rows, int cols, size_t elemSize);
+        void free(GpuMat* mat);
+    };
+
+    bool DefaultAllocator::allocate(GpuMat* mat, int rows, int cols, size_t elemSize)
+    {
+        if (rows > 1 && cols > 1)
+        {
+            CV_CUDEV_SAFE_CALL( cudaMallocPitch(&mat->data, &mat->step, elemSize * cols, rows) );
+        }
+        else
+        {
+            // Single row or single column must be continuous
+            CV_CUDEV_SAFE_CALL( cudaMalloc(&mat->data, elemSize * cols * rows) );
+            mat->step = elemSize * cols;
+        }
+
+        mat->refcount = (int*) fastMalloc(sizeof(int));
+
+        return true;
+    }
+
+    void DefaultAllocator::free(GpuMat* mat)
+    {
+        cudaFree(mat->datastart);
+        fastFree(mat->refcount);
+    }
+
+    DefaultAllocator cudaDefaultAllocator;
+    GpuMat::Allocator* g_defaultAllocator = &cudaDefaultAllocator;
+}
+
+GpuMat::Allocator* cv::cuda::GpuMat::defaultAllocator()
+{
+    return g_defaultAllocator;
+}
+
+void cv::cuda::GpuMat::setDefaultAllocator(Allocator* allocator)
+{
+    CV_Assert( allocator != 0 );
+    g_defaultAllocator = allocator;
+}
+
 /////////////////////////////////////////////////////
 /// create
 
@@ -76,19 +124,16 @@ void cv::cuda::GpuMat::create(int _rows, int _cols, int _type)
         rows = _rows;
         cols = _cols;
 
-        size_t esz = elemSize();
+        const size_t esz = elemSize();
 
-        void* devPtr;
+        bool allocSuccess = allocator->allocate(this, rows, cols, esz);
 
-        if (rows > 1 && cols > 1)
+        if (!allocSuccess)
         {
-            CV_CUDEV_SAFE_CALL( cudaMallocPitch(&devPtr, &step, esz * cols, rows) );
-        }
-        else
-        {
-            // Single row or single column must be continuous
-            CV_CUDEV_SAFE_CALL( cudaMalloc(&devPtr, esz * cols * rows) );
-            step = esz * cols;
+            // custom allocator fails, try default allocator
+            allocator = defaultAllocator();
+            allocSuccess = allocator->allocate(this, rows, cols, esz);
+            CV_Assert( allocSuccess );
         }
 
         if (esz * cols == step)
@@ -97,11 +142,11 @@ void cv::cuda::GpuMat::create(int _rows, int _cols, int _type)
         int64 _nettosize = static_cast<int64>(step) * rows;
         size_t nettosize = static_cast<size_t>(_nettosize);
 
-        datastart = data = static_cast<uchar*>(devPtr);
+        datastart = data;
         dataend = data + nettosize;
 
-        refcount = static_cast<int*>(fastMalloc(sizeof(*refcount)));
-        *refcount = 1;
+        if (refcount)
+            *refcount = 1;
     }
 }
 
@@ -110,11 +155,10 @@ void cv::cuda::GpuMat::create(int _rows, int _cols, int _type)
 
 void cv::cuda::GpuMat::release()
 {
+    CV_DbgAssert( allocator != 0 );
+
     if (refcount && CV_XADD(refcount, -1) == 1)
-    {
-        cudaFree(datastart);
-        fastFree(refcount);
-    }
+        allocator->free(this);
 
     data = datastart = dataend = 0;
     step = rows = cols = 0;

modules/core/src/cuda_buffer_pool.cpp

@@ -0,0 +1,418 @@
+/*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.
+// Copyright (C) 2013, OpenCV Foundation, 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 materials 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 implied warranties, including, but not limited to, the implied
+// 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 "precomp.hpp"
+
+using namespace cv;
+using namespace cv::cuda;
+
+#ifdef HAVE_CUDA
+
+#include "opencv2/cudev/common.hpp"
+
+/////////////////////////////////////////////////////////////
+/// MemoryStack
+
+namespace
+{
+    class MemoryPool;
+}
+
+class cv::cuda::MemoryStack
+{
+public:
+    uchar* requestMemory(size_t size);
+    void returnMemory(uchar* ptr);
+
+    uchar* datastart;
+    uchar* dataend;
+    uchar* tip;
+
+    bool isFree;
+    MemoryPool* pool;
+
+#if defined(DEBUG) || defined(_DEBUG)
+    std::vector<size_t> allocations;
+#endif
+};
+
+uchar* cv::cuda::MemoryStack::requestMemory(size_t size)
+{
+    const size_t freeMem = dataend - tip;
+
+    if (size > freeMem)
+        return 0;
+
+    uchar* ptr = tip;
+
+    tip += size;
+
+#if defined(DEBUG) || defined(_DEBUG)
+    allocations.push_back(size);
+#endif
+
+    return ptr;
+}
+
+void cv::cuda::MemoryStack::returnMemory(uchar* ptr)
+{
+    CV_DbgAssert( ptr >= datastart && ptr < dataend );
+
+#if defined(DEBUG) || defined(_DEBUG)
+    const size_t allocSize = tip - ptr;
+    CV_Assert( allocSize == allocations.back() );
+    allocations.pop_back();
+#endif
+
+    tip = ptr;
+}
+
+/////////////////////////////////////////////////////////////
+/// MemoryPool
+
+namespace
+{
+    class MemoryPool
+    {
+    public:
+        MemoryPool();
+
+        void initialize(size_t stackSize, int stackCount);
+        void release();
+
+        MemoryStack* getFreeMemStack();
+        void returnMemStack(MemoryStack* memStack);
+
+    private:
+        void initilizeImpl();
+
+        Mutex mtx_;
+
+        bool initialized_;
+        size_t stackSize_;
+        int stackCount_;
+
+        uchar* mem_;
+
+        std::vector<MemoryStack> stacks_;
+    };
+
+    MemoryPool::MemoryPool() : initialized_(false), mem_(0)
+    {
+        // default : 10 Mb, 5 stacks
+        stackSize_ = 10 * 1024 * 1024;
+        stackCount_ = 5;
+    }
+
+    void MemoryPool::initialize(size_t stackSize, int stackCount)
+    {
+        AutoLock lock(mtx_);
+
+        release();
+
+        stackSize_ = stackSize;
+        stackCount_ = stackCount;
+
+        initilizeImpl();
+    }
+
+    void MemoryPool::initilizeImpl()
+    {
+        const size_t totalSize = stackSize_ * stackCount_;
+
+        if (totalSize > 0)
+        {
+            cudaError_t err = cudaMalloc(&mem_, totalSize);
+            if (err != cudaSuccess)
+                return;
+
+            stacks_.resize(stackCount_);
+
+            uchar* ptr = mem_;
+
+            for (int i = 0; i < stackCount_; ++i)
+            {
+                stacks_[i].datastart = ptr;
+                stacks_[i].dataend = ptr + stackSize_;
+                stacks_[i].tip = ptr;
+                stacks_[i].isFree = true;
+                stacks_[i].pool = this;
+
+                ptr += stackSize_;
+            }
+
+            initialized_ = true;
+        }
+    }
+
+    void MemoryPool::release()
+    {
+        if (mem_)
+        {
+#if defined(DEBUG) || defined(_DEBUG)
+            for (int i = 0; i < stackCount_; ++i)
+            {
+                CV_DbgAssert( stacks_[i].isFree );
+                CV_DbgAssert( stacks_[i].tip == stacks_[i].datastart );
+            }
+#endif
+
+            cudaFree( mem_ );
+
+            mem_ = 0;
+            initialized_ = false;
+        }
+    }
+
+    MemoryStack* MemoryPool::getFreeMemStack()
+    {
+        AutoLock lock(mtx_);
+
+        if (!initialized_)
+            initilizeImpl();
+
+        if (!mem_)
+            return 0;
+
+        for (int i = 0; i < stackCount_; ++i)
+        {
+            if (stacks_[i].isFree)
+            {
+                stacks_[i].isFree = false;
+                return &stacks_[i];
+            }
+        }
+
+        return 0;
+    }
+
+    void MemoryPool::returnMemStack(MemoryStack* memStack)
+    {
+        AutoLock lock(mtx_);
+
+        CV_DbgAssert( !memStack->isFree );
+
+#if defined(DEBUG) || defined(_DEBUG)
+        bool found = false;
+        for (int i = 0; i < stackCount_; ++i)
+        {
+            if (memStack == &stacks_[i])
+            {
+                found = true;
+                break;
+            }
+        }
+        CV_DbgAssert( found );
+#endif
+
+        CV_DbgAssert( memStack->tip == memStack->datastart );
+
+        memStack->isFree = true;
+    }
+}
+
+/////////////////////////////////////////////////////////////
+/// MemoryPoolManager
+
+namespace
+{
+    class MemoryPoolManager
+    {
+    public:
+        MemoryPoolManager();
+        ~MemoryPoolManager();
+
+        MemoryPool* getPool(int deviceId);
+
+    private:
+        std::vector<MemoryPool> pools_;
+    };
+
+    MemoryPoolManager::MemoryPoolManager()
+    {
+        int deviceCount = getCudaEnabledDeviceCount();
+
+        if (deviceCount > 0)
+            pools_.resize(deviceCount);
+    }
+
+    MemoryPoolManager::~MemoryPoolManager()
+    {
+        for (size_t i = 0; i < pools_.size(); ++i)
+        {
+            cudaSetDevice(i);
+            pools_[i].release();
+        }
+    }
+
+    MemoryPool* MemoryPoolManager::getPool(int deviceId)
+    {
+        CV_DbgAssert( deviceId >= 0 && deviceId < static_cast<int>(pools_.size()) );
+        return &pools_[deviceId];
+    }
+
+    MemoryPool* memPool(int deviceId)
+    {
+        static MemoryPoolManager manager;
+        return manager.getPool(deviceId);
+    }
+}
+
+/////////////////////////////////////////////////////////////
+/// StackAllocator
+
+namespace
+{
+    bool enableMemoryPool = true;
+}
+
+cv::cuda::StackAllocator::StackAllocator(cudaStream_t stream) : stream_(stream), memStack_(0)
+{
+    if (enableMemoryPool)
+    {
+        const int deviceId = getDevice();
+        memStack_ = memPool(deviceId)->getFreeMemStack();
+
+        DeviceInfo devInfo(deviceId);
+        alignment_ = devInfo.textureAlignment();
+    }
+}
+
+cv::cuda::StackAllocator::~StackAllocator()
+{
+    cudaStreamSynchronize(stream_);
+
+    if (memStack_ != 0)
+        memStack_->pool->returnMemStack(memStack_);
+}
+
+namespace
+{
+    size_t alignUp(size_t what, size_t alignment)
+    {
+        size_t alignMask = alignment-1;
+        size_t inverseAlignMask = ~alignMask;
+        size_t res = (what + alignMask) & inverseAlignMask;
+        return res;
+    }
+}
+
+bool cv::cuda::StackAllocator::allocate(GpuMat* mat, int rows, int cols, size_t elemSize)
+{
+    if (memStack_ == 0)
+        return false;
+
+    size_t pitch, memSize;
+
+    if (rows > 1 && cols > 1)
+    {
+        pitch = alignUp(cols * elemSize, alignment_);
+        memSize = pitch * rows;
+    }
+    else
+    {
+        // Single row or single column must be continuous
+        pitch = elemSize * cols;
+        memSize = alignUp(elemSize * cols * rows, 64);
+    }
+
+    uchar* ptr = memStack_->requestMemory(memSize);
+
+    if (ptr == 0)
+        return false;
+
+    mat->data = ptr;
+    mat->step = pitch;
+    mat->refcount = (int*) fastMalloc(sizeof(int));
+
+    return true;
+}
+
+void cv::cuda::StackAllocator::free(GpuMat* mat)
+{
+    if (memStack_ == 0)
+        return;
+
+    memStack_->returnMemory(mat->datastart);
+    fastFree(mat->refcount);
+}
+
+void cv::cuda::setBufferPoolUsage(bool on)
+{
+    enableMemoryPool = on;
+}
+
+void cv::cuda::setBufferPoolConfig(int deviceId, size_t stackSize, int stackCount)
+{
+    const int currentDevice = getDevice();
+
+    if (deviceId >= 0)
+    {
+        setDevice(deviceId);
+        memPool(deviceId)->initialize(stackSize, stackCount);
+    }
+    else
+    {
+        const int deviceCount = getCudaEnabledDeviceCount();
+
+        for (deviceId = 0; deviceId < deviceCount; ++deviceId)
+        {
+            setDevice(deviceId);
+            memPool(deviceId)->initialize(stackSize, stackCount);
+        }
+    }
+
+    setDevice(currentDevice);
+}
+
+/////////////////////////////////////////////////////////////
+/// BufferPool
+
+GpuMat cv::cuda::BufferPool::getBuffer(int rows, int cols, int type)
+{
+    GpuMat buf(allocator_);
+    buf.create(rows, cols, type);
+    return buf;
+}
+
+#endif

modules/core/src/cuda_gpu_mat.cpp

@@ -49,7 +49,8 @@ using namespace cv::cuda;
 cv::cuda::GpuMat::GpuMat(int rows_, int cols_, int type_, void* data_, size_t step_) :
     flags(Mat::MAGIC_VAL + (type_ & Mat::TYPE_MASK)), rows(rows_), cols(cols_),
     step(step_), data((uchar*)data_), refcount(0),
-    datastart((uchar*)data_), dataend((uchar*)data_)
+    datastart((uchar*)data_), dataend((uchar*)data_),
+    allocator(defaultAllocator())
 {
     size_t minstep = cols * elemSize();
 
@@ -74,7 +75,8 @@ cv::cuda::GpuMat::GpuMat(int rows_, int cols_, int type_, void* data_, size_t st
 cv::cuda::GpuMat::GpuMat(Size size_, int type_, void* data_, size_t step_) :
     flags(Mat::MAGIC_VAL + (type_ & Mat::TYPE_MASK)), rows(size_.height), cols(size_.width),
     step(step_), data((uchar*)data_), refcount(0),
-    datastart((uchar*)data_), dataend((uchar*)data_)
+    datastart((uchar*)data_), dataend((uchar*)data_),
+    allocator(defaultAllocator())
 {
     size_t minstep = cols * elemSize();
 
@@ -92,6 +94,7 @@ cv::cuda::GpuMat::GpuMat(Size size_, int type_, void* data_, size_t step_) :
 
         flags |= step == minstep ? Mat::CONTINUOUS_FLAG : 0;
     }
+
     dataend += step * (rows - 1) + minstep;
 }
 
@@ -100,6 +103,7 @@ cv::cuda::GpuMat::GpuMat(const GpuMat& m, Range rowRange_, Range colRange_)
     flags = m.flags;
     step = m.step; refcount = m.refcount;
     data = m.data; datastart = m.datastart; dataend = m.dataend;
+    allocator = m.allocator;
 
     if (rowRange_ == Range::all())
     {
@@ -139,7 +143,8 @@ cv::cuda::GpuMat::GpuMat(const GpuMat& m, Range rowRange_, Range colRange_)
 cv::cuda::GpuMat::GpuMat(const GpuMat& m, Rect roi) :
     flags(m.flags), rows(roi.height), cols(roi.width),
     step(m.step), data(m.data + roi.y*step), refcount(m.refcount),
-    datastart(m.datastart), dataend(m.dataend)
+    datastart(m.datastart), dataend(m.dataend),
+    allocator(m.allocator)
 {
     flags &= roi.width < m.cols ? ~Mat::CONTINUOUS_FLAG : -1;
     data += roi.x * elemSize();
@@ -347,6 +352,17 @@ GpuMat cv::cuda::allocMatFromBuf(int rows, int cols, int type, GpuMat& mat)
 
 #ifndef HAVE_CUDA
 
+GpuMat::Allocator* cv::cuda::GpuMat::defaultAllocator()
+{
+    return 0;
+}
+
+void cv::cuda::GpuMat::setDefaultAllocator(Allocator* allocator)
+{
+    (void) allocator;
+    throw_no_cuda();
+}
+
 void cv::cuda::GpuMat::create(int _rows, int _cols, int _type)
 {
     (void) _rows;

modules/core/src/cuda_stream.cpp

@@ -66,6 +66,7 @@ class cv::cuda::Stream::Impl
 {
 public:
     cudaStream_t stream;
+    Ptr<StackAllocator> stackAllocator_;
 
     Impl();
     Impl(cudaStream_t stream);
@@ -73,17 +74,26 @@ public:
     ~Impl();
 };
 
+cv::cuda::BufferPool::BufferPool(Stream& stream) : allocator_(stream.impl_->stackAllocator_.get())
+{
+}
+
 cv::cuda::Stream::Impl::Impl() : stream(0)
 {
     cudaSafeCall( cudaStreamCreate(&stream) );
+
+    stackAllocator_ = makePtr<StackAllocator>(stream);
 }
 
 cv::cuda::Stream::Impl::Impl(cudaStream_t stream_) : stream(stream_)
 {
+    stackAllocator_ = makePtr<StackAllocator>(stream);
 }
 
 cv::cuda::Stream::Impl::~Impl()
 {
+    stackAllocator_.release();
+
     if (stream)
         cudaStreamDestroy(stream);
 }
@@ -197,7 +207,7 @@ cv::cuda::Stream::operator bool_type() const
 
 
 ////////////////////////////////////////////////////////////////
-// Stream
+// Event
 
 #ifndef HAVE_CUDA
 

modules/cuda/perf/perf_buffer_pool.cpp

@@ -0,0 +1,114 @@
+/*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 materials 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 implied warranties, including, but not limited to, the implied
+// 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 "perf_precomp.hpp"
+
+#ifdef HAVE_CUDA
+
+#include "opencv2/cudaarithm.hpp"
+#include "opencv2/core/private.cuda.hpp"
+
+using namespace testing;
+using namespace perf;
+using namespace cv;
+using namespace cv::cuda;
+
+namespace
+{
+    void func1(const GpuMat& src, GpuMat& dst, Stream& stream)
+    {
+        BufferPool pool(stream);
+
+        GpuMat buf = pool.getBuffer(src.size(), CV_32FC(src.channels()));
+
+        src.convertTo(buf, CV_32F, 1.0 / 255.0, stream);
+
+        cuda::exp(buf, dst, stream);
+    }
+
+    void func2(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, Stream& stream)
+    {
+        BufferPool pool(stream);
+
+        GpuMat buf1 = pool.getBuffer(src1.size(), CV_32FC(src1.channels()));
+
+        func1(src1, buf1, stream);
+
+        GpuMat buf2 = pool.getBuffer(src2.size(), CV_32FC(src2.channels()));
+
+        func1(src2, buf2, stream);
+
+        cuda::add(buf1, buf2, dst, noArray(), -1, stream);
+    }
+}
+
+PERF_TEST_P(Sz, BufferPool, CUDA_TYPICAL_MAT_SIZES)
+{
+    static bool first = true;
+
+    const Size size = GetParam();
+
+    const bool useBufferPool = PERF_RUN_CUDA();
+
+    Mat host_src(size, CV_8UC1);
+    declare.in(host_src, WARMUP_RNG);
+
+    GpuMat src1(host_src), src2(host_src);
+    GpuMat dst;
+
+    setBufferPoolUsage(useBufferPool);
+    if (useBufferPool && first)
+    {
+        setBufferPoolConfig(-1, 25 * 1024 * 1024, 2);
+        first = false;
+    }
+
+    TEST_CYCLE()
+    {
+        func2(src1, src2, dst, Stream::Null());
+    }
+
+    Mat h_dst(dst);
+    SANITY_CHECK(h_dst);
+}
+
+#endif

modules/cuda/test/test_buffer_pool.cpp

@@ -0,0 +1,120 @@
+/*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 materials 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 implied warranties, including, but not limited to, the implied
+// 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
+
+#include "opencv2/cudaarithm.hpp"
+#include "opencv2/cudawarping.hpp"
+#include "opencv2/core/private.cuda.hpp"
+
+using namespace testing;
+using namespace cv;
+using namespace cv::cuda;
+
+struct BufferPoolTest : TestWithParam<DeviceInfo>
+{
+};
+
+namespace
+{
+    void func1(const GpuMat& src, GpuMat& dst, Stream& stream)
+    {
+        BufferPool pool(stream);
+
+        GpuMat buf = pool.getBuffer(src.size(), CV_32FC(src.channels()));
+
+        src.convertTo(buf, CV_32F, 1.0 / 255.0, stream);
+
+        cuda::exp(buf, dst, stream);
+    }
+
+    void func2(const GpuMat& src, GpuMat& dst, Stream& stream)
+    {
+        BufferPool pool(stream);
+
+        GpuMat buf1 = pool.getBuffer(saturate_cast<int>(src.rows * 0.5), saturate_cast<int>(src.cols * 0.5), src.type());
+
+        cuda::resize(src, buf1, Size(), 0.5, 0.5, cv::INTER_NEAREST, stream);
+
+        GpuMat buf2 = pool.getBuffer(buf1.size(), CV_32FC(buf1.channels()));
+
+        func1(buf1, buf2, stream);
+
+        GpuMat buf3 = pool.getBuffer(src.size(), buf2.type());
+
+        cuda::resize(buf2, buf3, src.size(), 0, 0, cv::INTER_NEAREST, stream);
+
+        buf3.convertTo(dst, CV_8U, stream);
+    }
+}
+
+CUDA_TEST_P(BufferPoolTest, SimpleUsage)
+{
+    DeviceInfo devInfo = GetParam();
+    setDevice(devInfo.deviceID());
+
+    GpuMat src(200, 200, CV_8UC1);
+    GpuMat dst;
+
+    Stream stream;
+
+    func2(src, dst, stream);
+
+    stream.waitForCompletion();
+
+    GpuMat buf, buf1, buf2, buf3;
+    GpuMat dst_gold;
+
+    cuda::resize(src, buf1, Size(), 0.5, 0.5, cv::INTER_NEAREST);
+    buf1.convertTo(buf, CV_32F, 1.0 / 255.0);
+    cuda::exp(buf, buf2);
+    cuda::resize(buf2, buf3, src.size(), 0, 0, cv::INTER_NEAREST);
+    buf3.convertTo(dst_gold, CV_8U);
+
+    ASSERT_MAT_NEAR(dst_gold, dst, 0);
+}
+
+INSTANTIATE_TEST_CASE_P(CUDA_Stream, BufferPoolTest, ALL_DEVICES);
+
+#endif // HAVE_CUDA