void memory_t<CUDA>::free(){
if(!isTexture){
cuMemFree(*((CUdeviceptr*) handle));
if(!isAWrapper)
delete (CUdeviceptr*) handle;
}
else{
CUarray &array = ((CUDATextureData_t*) handle)->array;
CUsurfObject &surface = ((CUDATextureData_t*) handle)->surface;
cuArrayDestroy(array);
cuSurfObjectDestroy(surface);
if(!isAWrapper){
delete (CUDATextureData_t*) handle;
delete (CUaddress_mode*) textureInfo.arg;
}
}
size = 0;
}
bool ResourceCUDA::updateResourceFromCUDA(CUstream streamID)
{
CUresult res;
res = cuGraphicsMapResources( 1, &m_cudaResource, streamID );
if(res)
{
LOGE("Error>> CUDA failed map some target resources\n");
return false;
}
//
// Walk through output resources and perform the copies
//
CUarray cuArray;
//# define DBGDUMMYCOPY
# ifdef DBGDUMMYCOPY
// interop has issues... let's compare with a copy to a basic cuda array
int www = m_xByteSz/4;
CUDA_ARRAY_DESCRIPTOR descr = {
www,//unsigned int Width;
m_creationData.sz[1],//unsigned int Height;
CU_AD_FORMAT_UNSIGNED_INT8,//CUarray_format Format;
4//unsigned int NumChannels;
};
res = cuArrayCreate(&cuArray, &descr);
# else
res = cuGraphicsSubResourceGetMappedArray( &cuArray, m_cudaResource, 0/*arrayIndex*/, 0/*mipLevel*/);
# endif
if(res)
{
res = cuGraphicsUnmapResources( 1, &m_cudaResource, streamID );
return false;
}
CUDA_MEMCPY2D copyInfo = {
0, ///< Source X in bytes
0, ///< Source Y
CU_MEMORYTYPE_DEVICE,//< Source memory type (host, device, array)
NULL, ///< Source host pointer
m_dptr, ///< Source device pointer
NULL, ///< Source array reference
m_pitch, ///< Source pitch (ignored when src is array)
0, ///< Destination X in bytes
0, ///< Destination Y
CU_MEMORYTYPE_ARRAY,///< Destination memory type (host, device, array)
NULL, ///< Destination host pointer
NULL, ///< Destination device pointer
cuArray, ///< Destination array reference
0, ///< Destination pitch (ignored when dst is array)
m_xByteSz, ///< Width of 2D memory copy in bytes
m_creationData.sz[1] ///< Height of 2D memory copy
};
//LOGI("cuMemcpy2D(): CU_MEMORYTYPE_DEVICE source=%x pitch=%d CU_MEMORYTYPE_ARRAY=%x widthBytes=%d height=%d\n",m_dptr, m_pitch, cuArray, m_xByteSz, m_creationData.sz[1]);
res = cuMemcpy2D( ©Info );
if(res)
{
LOGE("Error>> CUDA failed to copy linear memory to texture (array memory)\n");
res = cuGraphicsUnmapResources( 1, &m_cudaResource, streamID );
return false;
}
# ifdef DBGDUMMYCOPY
res = cuArrayDestroy(cuArray);
# endif
res = cuGraphicsUnmapResources( 1, &m_cudaResource, streamID );
if(res)
{
LOGE("Error>> CUDA failed unmap the resource for output result of the kernel\n");
return false;
}
return true;
}
CuTexture::~CuTexture() {
if(_texture) cuArrayDestroy(_texture);
}
