InteropResource::~InteropResource()
{
//CUDA_WARN(cuCtxPushCurrent(ctx)); //error invalid value
if (res[0].cuRes)
CUDA_WARN(cuGraphicsUnregisterResource(res[0].cuRes));
if (res[1].cuRes)
CUDA_WARN(cuGraphicsUnregisterResource(res[1].cuRes));
if (res[0].stream)
CUDA_WARN(cuStreamDestroy(res[0].stream));
if (res[1].stream)
CUDA_WARN(cuStreamDestroy(res[1].stream));
// FIXME: we own the context. But why crash to destroy ctx? CUDA_ERROR_INVALID_VALUE
//CUDA_ENSURE(cuCtxDestroy(ctx));
}
bool GLInteropResource::ensureResource(int w, int h, int H, GLuint tex, int plane)
{
Q_ASSERT(plane < 2 && "plane number must be 0 or 1 for NV12");
TexRes &r = res[plane];
if (r.texture == tex && r.w == w && r.h == h && r.H == H && r.cuRes)
return true;
if (!ctx) {
// TODO: how to use pop/push decoder's context without the context in opengl context
CUDA_ENSURE(cuCtxCreate(&ctx, CU_CTX_SCHED_BLOCKING_SYNC, dev), false);
if (USE_STREAM) {
CUDA_WARN(cuStreamCreate(&res[0].stream, CU_STREAM_DEFAULT));
CUDA_WARN(cuStreamCreate(&res[1].stream, CU_STREAM_DEFAULT));
}
qDebug("cuda contex on gl thread: %p", ctx);
CUDA_ENSURE(cuCtxPopCurrent(&ctx), false); // TODO: why cuMemcpy2D need this
}
if (r.cuRes) {
CUDA_ENSURE(cuGraphicsUnregisterResource(r.cuRes), false);
r.cuRes = NULL;
}
// CU_GRAPHICS_REGISTER_FLAGS_WRITE_DISCARD works too for opengl, but not d3d
CUDA_ENSURE(cuGraphicsGLRegisterImage(&r.cuRes, tex, GL_TEXTURE_2D, CU_GRAPHICS_REGISTER_FLAGS_NONE), false);
r.texture = tex;
r.w = w;
r.h = h;
r.H = H;
return true;
}
//----------------------------------------------------------------
/// setup for interop : prt refers to a texture in which we write
/// results. This must be seen as a linear buffer. So we need to
/// allocate a temporary linear buffer that will be then copied
/// back to the texture
//
bool ResourceCUDA::setupAsCUDATarget()
{
int fmtSz = ResourceFormatByteSize(m_creationData.fmt);
CUresult res;
m_xByteSz = m_creationData.sz[0] * fmtSz;
m_size = m_xByteSz * m_creationData.sz[1];
if(m_dptr)
{
res = cuMemFree(m_dptr);
if(res)
return false;
}
res = cuMemAllocPitch( &m_dptr, &m_pitch, m_xByteSz, m_creationData.sz[1], 4);
if(res)
return false;
float pitchToSendToKernel = (float)m_pitch / (float)fmtSz;
#pragma MESSAGE("TODO TODO TODO TODO TODO TODO : send pitch to the kernel !")
LOGI("Event>>cuMemAllocPitch : Pitch of Target buffer (%d, %d) allocation = %d, %f\n", m_creationData.sz[0], m_xByteSz, m_pitch, pitchToSendToKernel);
//
// Register the texture to CUDA to be able to copy data back in it
//
GLenum target;
if(m_cudaResource)
{
res = cuGraphicsUnregisterResource(m_cudaResource);
if(res)
return false;
}
switch(m_type)
{
case RESTEX_1D:
target = GL_TEXTURE_1D;
break;
case RESTEX_2D:
target = GL_TEXTURE_2D;
break;
case RESTEX_2DRECT:
case RESRBUF_2D:
case RESOURCE_UNKNOWN:
//case RESTEX_3D:
//case RESTEX_CUBE_MAP:
default:
LOGE("Failed to register the resource %s for CUDA : may be a render buffer\n", m_name.c_str());
return false;
};
res = cuGraphicsGLRegisterImage(
&m_cudaResource, m_OGLId, target,
CU_GRAPHICS_MAP_RESOURCE_FLAGS_WRITE_DISCARD );
if(res)
{
LOGE("Failed to register the texture %s for CUDA (as write discard)\n", m_name.c_str());
return 0;
}
return true;
}
bool EGLInteropResource::ensureD3D9CUDA(int w, int h, int W, int H)
{
TexRes &r = res[0];// 1 NV12 texture
if (r.w == w && r.h == h && r.W == W && r.H == H && r.cuRes)
return true;
if (!ctx) {
// TODO: how to use pop/push decoder's context without the context in opengl context
if (!ensureD3DDevice())
return false;
// CUdevice is different from decoder's
CUDA_ENSURE(cuD3D9CtxCreate(&ctx, &dev, CU_CTX_SCHED_BLOCKING_SYNC, device9), false);
#if USE_STREAM
CUDA_WARN(cuStreamCreate(&res[0].stream, CU_STREAM_DEFAULT));
CUDA_WARN(cuStreamCreate(&res[1].stream, CU_STREAM_DEFAULT));
#endif //USE_STREAM
qDebug("cuda contex on gl thread: %p", ctx);
CUDA_ENSURE(cuCtxPopCurrent(&ctx), false); // TODO: why cuMemcpy2D need this
}
if (r.cuRes) {
CUDA_ENSURE(cuGraphicsUnregisterResource(r.cuRes), false);
r.cuRes = NULL;
}
// create d3d resource for interop
if (!surface9_nv12) {
// TODO: need pitch from cuvid to ensure cuMemcpy2D can copy the whole pitch
DX_ENSURE(device9->CreateTexture(W
//, H
, H*3/2
, 1
, D3DUSAGE_DYNAMIC //D3DUSAGE_DYNAMIC is lockable // 0 is from NV example. cudaD3D9.h says The primary rendertarget may not be registered with CUDA. So can not be D3DUSAGE_RENDERTARGET?
//, D3DUSAGE_RENDERTARGET
, D3DFMT_L8
//, (D3DFORMAT)MAKEFOURCC('N','V','1','2') // can not create nv12. use 2 textures L8+A8L8?
, D3DPOOL_DEFAULT // must be D3DPOOL_DEFAULT for cuda?
, &texture9_nv12
, NULL) // - Resources allocated as shared may not be registered with CUDA.
, false);
DX_ENSURE(device9->CreateOffscreenPlainSurface(W, H, (D3DFORMAT)MAKEFOURCC('N','V','1','2'), D3DPOOL_DEFAULT, &surface9_nv12, NULL), false);
}
// TODO: cudaD3D9.h says NV12 is not supported
// CUDA_ERROR_INVALID_HANDLE if register D3D9 surface
// TODO: why flag CU_GRAPHICS_REGISTER_FLAGS_WRITE_DISCARD is invalid while it's fine for opengl
CUDA_ENSURE(cuGraphicsD3D9RegisterResource(&r.cuRes, texture9_nv12, CU_GRAPHICS_REGISTER_FLAGS_NONE), false);
return true;
}
CudaSurface::~CudaSurface(void)
{
getGLTexture(); // unmap
cuGraphicsUnregisterResource(m_cudaResource);
glDeleteTextures(1, &m_glTexture);
}
