generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

added implement MatPL with serocopy and writecombited

Browse Source
This commit is contained in:
Andrey Morozov
2010-08-13 14:52:50 +00:00
parent 024283ceae
commit 35ebeb21bd
3 changed files with 97 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
modules/gpu/src/matrix_operations.cpp
View File

@@ -67,7 +67,8 @@ namespace cv
void GpuMat::create(int /*_rows*/, int /*_cols*/, int /*_type*/) { throw_nogpu(); }
void GpuMat::release() { throw_nogpu(); }
void MatPL::create(int /*_rows*/, int /*_cols*/, int /*_type*/) { throw_nogpu(); }
void MatPL::create(int /*_rows*/, int /*_cols*/, int /*_type*/, int /*type_alloc*/) { throw_nogpu(); }
void MatPL::get_property_device() { throw_nogpu(); }
void MatPL::release() { throw_nogpu(); }
}
@@ -164,7 +165,7 @@ GpuMat& GpuMat::setTo(const Scalar& s, const GpuMat& mask)
else
impl::set_to_with_mask( *this, depth(), s.val, mask, channels());
return *this;
return *this;
}
@@ -209,6 +210,15 @@ GpuMat cv::gpu::GpuMat::reshape(int new_cn, int new_rows) const
return hdr;
}
bool cv::gpu::MatPL::can_device_map_to_host()
{
cudaDeviceProp prop;
cudaGetDeviceProperties(&prop, 0);
return (prop.canMapHostMemory != 0) ? true : false;
}
void cv::gpu::GpuMat::create(int _rows, int _cols, int _type)
{
_type &= TYPE_MASK;
@@ -259,8 +269,9 @@ void cv::gpu::GpuMat::release()
//////////////////////////////// MatPL ////////////////////////////////
///////////////////////////////////////////////////////////////////////
void cv::gpu::MatPL::create(int _rows, int _cols, int _type)
void cv::gpu::MatPL::create(int _rows, int _cols, int _type, int type_alloc)
{
alloc_type = type_alloc;
_type &= TYPE_MASK;
if( rows == _rows && cols == _cols && type() == _type && data )
return;
@@ -281,7 +292,24 @@ void cv::gpu::MatPL::create(int _rows, int _cols, int _type)
//datastart = data = (uchar*)fastMalloc(datasize + sizeof(*refcount));
void *ptr;
cudaSafeCall( cudaHostAlloc( &ptr, datasize, cudaHostAllocDefault) );
switch (type_alloc)
{
case ALLOC_PAGE_LOCKED: cudaSafeCall( cudaHostAlloc( &ptr, datasize, cudaHostAllocDefault) ); break;
case ALLOC_ZEROCOPY:
if (can_device_map_to_host() == true)
{
cudaSafeCall( cudaHostAlloc( &ptr, datasize, cudaHostAllocMapped) );
}
else
cv::gpu::error("ZeroCopy is not supported by current device", __FILE__, __LINE__);
break;
case ALLOC_WRITE_COMBINED: cudaSafeCall( cudaHostAlloc( &ptr, datasize, cudaHostAllocWriteCombined) ); break;
default:
cv::gpu::error("Invalid alloc type", __FILE__, __LINE__);
}
datastart = data = (uchar*)ptr;
dataend = data + nettosize;
@@ -291,6 +319,19 @@ void cv::gpu::MatPL::create(int _rows, int _cols, int _type)
}
}
inline MatPL::operator GpuMat() const
{
if (alloc_type == ALLOC_ZEROCOPY)
{
void ** pdev;
cudaHostGetDevicePointer( pdev, this->data, 0 );
GpuMat m(this->rows, this->cols, this->type(), *pdev, this->step);
return m;
}
else
cv::gpu::error("", __FILE__, __LINE__);
}
void cv::gpu::MatPL::release()
{
if( refcount && CV_XADD(refcount, -1) == 1 )