static inline CUresult cuMemcpyDtoH2DAsync(void * A, size_t lda, size_t ai, size_t aj,
CUdeviceptr B, size_t ldb, size_t bi, size_t bj,
size_t m, size_t n, size_t elemSize, CUstream stream) {
CUDA_MEMCPY2D copy = {
bi * elemSize, bj, CU_MEMORYTYPE_DEVICE, NULL, B, 0, ldb * elemSize,
ai * elemSize, aj, CU_MEMORYTYPE_HOST, A, 0, 0, lda * elemSize,
m * elemSize, n };
return cuMemcpy2DAsync(©, stream);
}
bool GLInteropResource::map(int picIndex, const CUVIDPROCPARAMS ¶m, GLuint tex, int w, int h, int H, int plane)
{
AutoCtxLock locker((cuda_api*)this, lock);
Q_UNUSED(locker);
if (!ensureResource(w, h, H, tex, plane)) // TODO surface size instead of frame size because we copy the device data
return false;
//CUDA_ENSURE(cuCtxPushCurrent(ctx), false);
CUdeviceptr devptr;
unsigned int pitch;
CUDA_ENSURE(cuvidMapVideoFrame(dec, picIndex, &devptr, &pitch, const_cast<CUVIDPROCPARAMS*>(¶m)), false);
CUVIDAutoUnmapper unmapper(this, dec, devptr);
Q_UNUSED(unmapper);
// TODO: why can not use res[plane].stream? CUDA_ERROR_INVALID_HANDLE
CUDA_ENSURE(cuGraphicsMapResources(1, &res[plane].cuRes, 0), false);
CUarray array;
CUDA_ENSURE(cuGraphicsSubResourceGetMappedArray(&array, res[plane].cuRes, 0, 0), false);
CUDA_MEMCPY2D cu2d;
memset(&cu2d, 0, sizeof(cu2d));
cu2d.srcDevice = devptr;
cu2d.srcMemoryType = CU_MEMORYTYPE_DEVICE;
cu2d.srcPitch = pitch;
cu2d.dstArray = array;
cu2d.dstMemoryType = CU_MEMORYTYPE_ARRAY;
cu2d.dstPitch = pitch;
// the whole size or copy size?
cu2d.WidthInBytes = pitch;
cu2d.Height = h;
if (plane == 1) {
cu2d.srcXInBytes = 0;// +srcY*srcPitch + srcXInBytes
cu2d.srcY = H; // skip the padding height
cu2d.Height /= 2;
}
if (res[plane].stream)
CUDA_ENSURE(cuMemcpy2DAsync(&cu2d, res[plane].stream), false);
else
CUDA_ENSURE(cuMemcpy2D(&cu2d), false);
//TODO: delay cuCtxSynchronize && unmap. do it in unmap(tex)?
// map to an already mapped resource will crash. sometimes I can not unmap the resource in unmap(tex) because if context switch error
// so I simply unmap the resource here
if (WORKAROUND_UNMAP_CONTEXT_SWITCH) {
if (res[plane].stream) {
//CUDA_WARN(cuCtxSynchronize(), false); //wait too long time? use cuStreamQuery?
CUDA_WARN(cuStreamSynchronize(res[plane].stream)); //slower than CtxSynchronize
}
/*
* This function provides the synchronization guarantee that any CUDA work issued
* in \p stream before ::cuGraphicsUnmapResources() will complete before any
* subsequently issued graphics work begins.
* The graphics API from which \p resources were registered
* should not access any resources while they are mapped by CUDA. If an
* application does so, the results are undefined.
*/
CUDA_ENSURE(cuGraphicsUnmapResources(1, &res[plane].cuRes, 0), false);
} else {
// call it at last. current context will be used by other cuda calls (unmap() for example)
CUDA_ENSURE(cuCtxPopCurrent(&ctx), false); // not required
}
return true;
}
bool EGLInteropResource::map(int picIndex, const CUVIDPROCPARAMS ¶m, GLuint tex, int w, int h, int H, int plane)
{
// plane is always 0 because frame is rgb
AutoCtxLock locker((cuda_api*)this, lock);
Q_UNUSED(locker);
if (!ensureResource(w, h, param.Reserved[0], H, tex)) // TODO surface size instead of frame size because we copy the device data
return false;
//CUDA_ENSURE(cuCtxPushCurrent(ctx), false);
CUdeviceptr devptr;
unsigned int pitch;
CUDA_ENSURE(cuvidMapVideoFrame(dec, picIndex, &devptr, &pitch, const_cast<CUVIDPROCPARAMS*>(¶m)), false);
CUVIDAutoUnmapper unmapper(this, dec, devptr);
Q_UNUSED(unmapper);
// TODO: why can not use res[plane].stream? CUDA_ERROR_INVALID_HANDLE
CUDA_ENSURE(cuGraphicsMapResources(1, &res[plane].cuRes, 0), false);
CUarray array;
CUDA_ENSURE(cuGraphicsSubResourceGetMappedArray(&array, res[plane].cuRes, 0, 0), false);
CUDA_ENSURE(cuGraphicsUnmapResources(1, &res[plane].cuRes, 0), false); // mapped array still accessible!
CUDA_MEMCPY2D cu2d;
memset(&cu2d, 0, sizeof(cu2d));
// Y plane
cu2d.srcDevice = devptr;
cu2d.srcMemoryType = CU_MEMORYTYPE_DEVICE;
cu2d.srcPitch = pitch;
cu2d.dstArray = array;
cu2d.dstMemoryType = CU_MEMORYTYPE_ARRAY;
cu2d.dstPitch = pitch;
// the whole size or copy size?
cu2d.WidthInBytes = res[plane].W; // the same value as texture9_nv12
cu2d.Height = H*3/2;
if (res[plane].stream)
CUDA_ENSURE(cuMemcpy2DAsync(&cu2d, res[plane].stream), false);
else
CUDA_ENSURE(cuMemcpy2D(&cu2d), false);
//TODO: delay cuCtxSynchronize && unmap. do it in unmap(tex)?
// map to an already mapped resource will crash. sometimes I can not unmap the resource in unmap(tex) because if context switch error
// so I simply unmap the resource here
if (WORKAROUND_UNMAP_CONTEXT_SWITCH) {
if (res[plane].stream) {
//CUDA_WARN(cuCtxSynchronize(), false); //wait too long time? use cuStreamQuery?
CUDA_WARN(cuStreamSynchronize(res[plane].stream)); //slower than CtxSynchronize
}
/*
* This function provides the synchronization guarantee that any CUDA work issued
* in \p stream before ::cuGraphicsUnmapResources() will complete before any
* subsequently issued graphics work begins.
* The graphics API from which \p resources were registered
* should not access any resources while they are mapped by CUDA. If an
* application does so, the results are undefined.
*/
// CUDA_ENSURE(cuGraphicsUnmapResources(1, &res[plane].cuRes, 0), false);
}
D3DLOCKED_RECT rect_src, rect_dst;
DX_ENSURE(texture9_nv12->LockRect(0, &rect_src, NULL, D3DLOCK_READONLY), false);
DX_ENSURE(surface9_nv12->LockRect(&rect_dst, NULL, D3DLOCK_DISCARD), false);
memcpy(rect_dst.pBits, rect_src.pBits, res[plane].W*H*3/2); // exactly w and h
DX_ENSURE(surface9_nv12->UnlockRect(), false);
DX_ENSURE(texture9_nv12->UnlockRect(0), false);
#if 0
//IDirect3DSurface9 *raw_surface = NULL;
//DX_ENSURE(texture9_nv12->GetSurfaceLevel(0, &raw_surface), false);
const RECT src = { 0, 0, w, h*3/2};
DX_ENSURE(device9->StretchRect(raw_surface, &src, surface9_nv12, NULL, D3DTEXF_NONE), false);
#endif
if (!map(surface9_nv12, tex, w, h, H))
return false;
return true;
}
