BufferPool for use with CUDA streams. More...

#include "cuda.hpp"

Public Member Functions
	BufferPool (Stream &stream)
	Gets the BufferPool for the given stream. More...

Ptr< GpuMat::Allocator >	getAllocator () const
	Returns the allocator associated with the stream. More...

GpuMat	getBuffer (int rows, int cols, int type)
	Allocates a new GpuMat of given size and type. More...

GpuMat	getBuffer (Size size, int type)
	Allocates a new GpuMat of given size and type. More...

Detailed Description

BufferPool for use with CUDA streams.

BufferPool utilizes Stream's allocator to create new buffers for GpuMat's. It is only useful when enabled with setBufferPoolUsage.

setBufferPoolUsage(true);

Note: setBufferPoolUsage must be called before any Stream declaration.

Users may specify custom allocator for Stream and may implement their own stream based functions utilizing the same underlying GPU memory management.

If custom allocator is not specified, BufferPool utilizes StackAllocator by default. StackAllocator allocates a chunk of GPU device memory beforehand, and when GpuMat is declared later on, it is given the pre-allocated memory. This kind of strategy reduces the number of calls for memory allocating APIs such as cudaMalloc or cudaMallocPitch.

Below is an example that utilizes BufferPool with StackAllocator:

#include <opencv2/opencv.hpp>
using namespace cv;
using namespace cv::cuda
int main()
{
    setBufferPoolUsage(true);                               // Tell OpenCV that we are going to utilize BufferPool
    setBufferPoolConfig(getDevice(), 1024 * 1024 * 64, 2);  // Allocate 64 MB, 2 stacks (default is 10 MB, 5 stacks)
    Stream stream1, stream2;                                // Each stream uses 1 stack
    BufferPool pool1(stream1), pool2(stream2);
    GpuMat d_src1 = pool1.getBuffer(4096, 4096, CV_8UC1);   // 16MB
    GpuMat d_dst1 = pool1.getBuffer(4096, 4096, CV_8UC3);   // 48MB, pool1 is now full
    GpuMat d_src2 = pool2.getBuffer(1024, 1024, CV_8UC1);   // 1MB
    GpuMat d_dst2 = pool2.getBuffer(1024, 1024, CV_8UC3);   // 3MB
    cvtColor(d_src1, d_dst1, CV_GRAY2BGR, 0, stream1);
    cvtColor(d_src2, d_dst2, CV_GRAY2BGR, 0, stream2);
}

If we allocate another GpuMat on pool1 in the above example, it will be carried out by the DefaultAllocator since the stack for pool1 is full.

GpuMat d_add1 = pool1.getBuffer(1024, 1024, CV_8UC1); // Stack for pool1 is full, memory is allocated with DefaultAllocator

If a third stream is declared in the above example, allocating with getBuffer within that stream will also be carried out by the DefaultAllocator because we've run out of stacks.

Stream stream3;                                         // Only 2 stacks were allocated, we've run out of stacks
BufferPool pool3(stream3);
GpuMat d_src3 = pool3.getBuffer(1024, 1024, CV_8UC1);   // Memory is allocated with DefaultAllocator

Warning: When utilizing StackAllocator, deallocation order is important.

Just like a stack, deallocation must be done in LIFO order. Below is an example of erroneous usage that violates LIFO rule. If OpenCV is compiled in Debug mode, this sample code will emit CV_Assert error.

int main()
{
    setBufferPoolUsage(true);                               // Tell OpenCV that we are going to utilize BufferPool
    Stream stream;                                          // A default size (10 MB) stack is allocated to this stream
    BufferPool pool(stream);
    GpuMat mat1 = pool.getBuffer(1024, 1024, CV_8UC1);      // Allocate mat1 (1MB)
    GpuMat mat2 = pool.getBuffer(1024, 1024, CV_8UC1);      // Allocate mat2 (1MB)
    mat1.release();                                         // erroneous usage : mat2 must be deallocated before mat1
}

Since C++ local variables are destroyed in the reverse order of construction, the code sample below satisfies the LIFO rule. Local GpuMat's are deallocated and the corresponding memory is automatically returned to the pool for later usage.

int main()
{
    setBufferPoolUsage(true);                               // Tell OpenCV that we are going to utilize BufferPool
    setBufferPoolConfig(getDevice(), 1024 * 1024 * 64, 2);  // Allocate 64 MB, 2 stacks (default is 10 MB, 5 stacks)
    Stream stream1, stream2;                                // Each stream uses 1 stack
    BufferPool pool1(stream1), pool2(stream2);
    for (int i = 0; i < 10; i++)
    {
        GpuMat d_src1 = pool1.getBuffer(4096, 4096, CV_8UC1);   // 16MB
        GpuMat d_dst1 = pool1.getBuffer(4096, 4096, CV_8UC3);   // 48MB, pool1 is now full
        GpuMat d_src2 = pool2.getBuffer(1024, 1024, CV_8UC1);   // 1MB
        GpuMat d_dst2 = pool2.getBuffer(1024, 1024, CV_8UC3);   // 3MB
        d_src1.setTo(Scalar(i), stream1);
        d_src2.setTo(Scalar(i), stream2);
        cvtColor(d_src1, d_dst1, CV_GRAY2BGR, 0, stream1);
        cvtColor(d_src2, d_dst2, CV_GRAY2BGR, 0, stream2);
                                                                // The order of destruction of the local variables is:
                                                                //   d_dst2 => d_src2 => d_dst1 => d_src1
                                                                // LIFO rule is satisfied, this code runs without error
    }
}

Constructor & Destructor Documentation

◆ BufferPool()

cv::cuda::BufferPool::BufferPool ( Stream & stream )

explicit

Gets the BufferPool for the given stream.

Member Function Documentation

◆ getAllocator()

Ptr<GpuMat::Allocator> cv::cuda::BufferPool::getAllocator ( ) const

inline

Returns the allocator associated with the stream.

References CV_EXPORTS_W, cv::cuda::setBufferPoolConfig(), and cv::cuda::setBufferPoolUsage().

Here is the call graph for this function:

◆ getBuffer() [1/2]

GpuMat cv::cuda::BufferPool::getBuffer	(	int	rows,
		int	cols,
		int	type
	)

Allocates a new GpuMat of given size and type.

◆ getBuffer() [2/2]

GpuMat cv::cuda::BufferPool::getBuffer	(	Size	size,
		int	type
	)

inline

Allocates a new GpuMat of given size and type.

References getBuffer(), cv::Size_< _Tp >::height, and cv::Size_< _Tp >::width.

Referenced by getBuffer().