IDirect3DDevice9* CreateDevice9(HINSTANCE dll, IDirect3D9** d3d9, D3DADAPTER_IDENTIFIER9 *d3dai)
{
qDebug("creating d3d9 device...");
typedef IDirect3D9* (WINAPI *Create9Func)(UINT SDKVersion);
Create9Func Create9 = (Create9Func)GetProcAddress(dll, "Direct3DCreate9");
if (!Create9) {
qWarning("Symbol not found: Direct3DCreate9");
return NULL;
}
*d3d9 = Create9(D3D_SDK_VERSION);
if (!(*d3d9)) {
qWarning("Direct3DCreate9 failed");
return NULL;
}
if (d3dai)
DX_WARN((*d3d9)->GetAdapterIdentifier(D3DADAPTER_DEFAULT, 0, d3dai));
D3DPRESENT_PARAMETERS d3dpp;
InitParameters(&d3dpp);
DWORD flags = D3DCREATE_FPU_PRESERVE | D3DCREATE_MULTITHREADED | D3DCREATE_MIXED_VERTEXPROCESSING;
IDirect3DDevice9 *d3d9dev = NULL;
DX_ENSURE(((*d3d9)->CreateDevice(D3DADAPTER_DEFAULT,
D3DDEVTYPE_HAL, GetShellWindow(),// GetDesktopWindow(), //GetShellWindow()?
flags,
&d3dpp, &d3d9dev))
, NULL);
qDebug("IDirect3DDevice9 created");
return d3d9dev;
}
bool EGLInteropResource::ensureD3DDevice()
{
if (device9)
return true;
if (!dll9)
dll9 = LoadLibrary(TEXT("D3D9.DLL"));
if (!dll9) {
qWarning("cuda::EGLInteropResource cannot load d3d9.dll");
return false;
}
D3DADAPTER_IDENTIFIER9 ai9;
ZeroMemory(&ai9, sizeof(ai9));
device9 = DXHelper::CreateDevice9Ex(dll9, (IDirect3D9Ex**)(&d3d9), &ai9);
if (!device9) {
qWarning("Failed to create d3d9 device ex, fallback to d3d9 device");
device9 = DXHelper::CreateDevice9(dll9, &d3d9, &ai9);
}
if (!device9)
return false;
qDebug() << QString().sprintf("CUDA.D3D9 (%.*s, vendor %lu, device %lu, revision %lu)",
sizeof(ai9.Description), ai9.Description,
ai9.VendorId, ai9.DeviceId, ai9.Revision);
// move to ensureResouce
DX_ENSURE(device9->CreateQuery(D3DQUERYTYPE_EVENT, &query9), false);
query9->Issue(D3DISSUE_END);
return !!device9;
}
bool EGLInteropResource::map(IDirect3DSurface9* surface, GLuint tex, int w, int h, int)
{
if (!ensureSurface(w, h)) {
releaseEGL();
releaseDX();
return false;
}
const RECT src = { 0, 0, (~0-1)&w, (~0-1)&h};
DX_ENSURE(d3ddev->StretchRect(surface, &src, dx_surface, NULL, D3DTEXF_NONE), false);
if (dx_query) {
// Flush the draw command now. Ideally, this should be done immediately before the draw call that uses the texture. Flush it once here though.
dx_query->Issue(D3DISSUE_END); //StretchRect does not supports odd values
// ensure data is copied to egl surface. Solution and comment is from chromium
// The DXVA decoder has its own device which it uses for decoding. ANGLE has its own device which we don't have access to.
// The above code attempts to copy the decoded picture into a surface which is owned by ANGLE.
// As there are multiple devices involved in this, the StretchRect call above is not synchronous.
// We attempt to flush the batched operations to ensure that the picture is copied to the surface owned by ANGLE.
// We need to do this in a loop and call flush multiple times.
// We have seen the GetData call for flushing the command buffer fail to return success occassionally on multi core machines, leading to an infinite loop.
// Workaround is to have an upper limit of 10 on the number of iterations to wait for the Flush to finish.
int k = 0;
while ((dx_query->GetData(NULL, 0, D3DGETDATA_FLUSH) == FALSE) && ++k < 10) {
Sleep(1);
}
}
DYGL(glBindTexture(GL_TEXTURE_2D, tex));
eglBindTexImage(egl->dpy, egl->surface, EGL_BACK_BUFFER);
DYGL(glBindTexture(GL_TEXTURE_2D, 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;
}
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;
}
bool EGLInteropResource::ensureD3D9EGL(int w, int h) {
if (surface9 && res[0].w == w && res[0].h == h)
return true;
#if QTAV_HAVE(GUI_PRIVATE)
QPlatformNativeInterface *nativeInterface = QGuiApplication::platformNativeInterface();
egl->dpy = static_cast<EGLDisplay>(nativeInterface->nativeResourceForContext("eglDisplay", QOpenGLContext::currentContext()));
EGLConfig egl_cfg = static_cast<EGLConfig>(nativeInterface->nativeResourceForContext("eglConfig", QOpenGLContext::currentContext()));
#else
#ifdef Q_OS_WIN
#if QT_VERSION < QT_VERSION_CHECK(5, 5, 0)
#ifdef _MSC_VER
#pragma message("ANGLE version in Qt<5.5 does not support eglQueryContext. You must upgrade your runtime ANGLE libraries")
#else
#warning "ANGLE version in Qt<5.5 does not support eglQueryContext. You must upgrade your runtime ANGLE libraries"
#endif //_MSC_VER
#endif
#endif //Q_OS_WIN
// eglQueryContext() added (Feb 2015): https://github.com/google/angle/commit/8310797003c44005da4143774293ea69671b0e2a
egl->dpy = eglGetCurrentDisplay();
qDebug("EGL version: %s, client api: %s", eglQueryString(egl->dpy, EGL_VERSION), eglQueryString(egl->dpy, EGL_CLIENT_APIS));
// TODO: check runtime egl>=1.4 for eglGetCurrentContext()
EGLint cfg_id = 0;
EGL_ENSURE(eglQueryContext(egl->dpy, eglGetCurrentContext(), EGL_CONFIG_ID , &cfg_id) == EGL_TRUE, false);
qDebug("egl config id: %d", cfg_id);
EGLint nb_cfg = 0;
EGL_ENSURE(eglGetConfigs(egl->dpy, NULL, 0, &nb_cfg) == EGL_TRUE, false);
qDebug("eglGetConfigs number: %d", nb_cfg);
QVector<EGLConfig> cfgs(nb_cfg); //check > 0
EGL_ENSURE(eglGetConfigs(egl->dpy, cfgs.data(), cfgs.size(), &nb_cfg) == EGL_TRUE, false);
EGLConfig egl_cfg = NULL;
for (int i = 0; i < nb_cfg; ++i) {
EGLint id = 0;
eglGetConfigAttrib(egl->dpy, cfgs[i], EGL_CONFIG_ID, &id);
if (id == cfg_id) {
egl_cfg = cfgs[i];
break;
}
}
#endif
qDebug("egl display:%p config: %p", egl->dpy, egl_cfg);
// check extensions
QList<QByteArray> extensions = QByteArray(eglQueryString(egl->dpy, EGL_EXTENSIONS)).split(' ');
// ANGLE_d3d_share_handle_client_buffer will be used if possible
const bool kEGL_ANGLE_d3d_share_handle_client_buffer = extensions.contains("EGL_ANGLE_d3d_share_handle_client_buffer");
const bool kEGL_ANGLE_query_surface_pointer = extensions.contains("EGL_ANGLE_query_surface_pointer");
if (!kEGL_ANGLE_d3d_share_handle_client_buffer && !kEGL_ANGLE_query_surface_pointer) {
qWarning("EGL extension 'kEGL_ANGLE_query_surface_pointer' or 'ANGLE_d3d_share_handle_client_buffer' is required!");
return false;
}
GLint has_alpha = 1; //QOpenGLContext::currentContext()->format().hasAlpha()
eglGetConfigAttrib(egl->dpy, egl_cfg, EGL_BIND_TO_TEXTURE_RGBA, &has_alpha); //EGL_ALPHA_SIZE
EGLint attribs[] = {
EGL_WIDTH, w,
EGL_HEIGHT, h,
EGL_TEXTURE_FORMAT, has_alpha ? EGL_TEXTURE_RGBA : EGL_TEXTURE_RGB,
EGL_TEXTURE_TARGET, EGL_TEXTURE_2D,
EGL_NONE
};
HANDLE share_handle = NULL;
if (!kEGL_ANGLE_d3d_share_handle_client_buffer && kEGL_ANGLE_query_surface_pointer) {
EGL_ENSURE((egl->surface = eglCreatePbufferSurface(egl->dpy, egl_cfg, attribs)) != EGL_NO_SURFACE, false);
qDebug("pbuffer surface: %p", egl->surface);
PFNEGLQUERYSURFACEPOINTERANGLEPROC eglQuerySurfacePointerANGLE = reinterpret_cast<PFNEGLQUERYSURFACEPOINTERANGLEPROC>(eglGetProcAddress("eglQuerySurfacePointerANGLE"));
if (!eglQuerySurfacePointerANGLE) {
qWarning("EGL_ANGLE_query_surface_pointer is not supported");
return false;
}
EGL_ENSURE(eglQuerySurfacePointerANGLE(egl->dpy, egl->surface, EGL_D3D_TEXTURE_2D_SHARE_HANDLE_ANGLE, &share_handle), false);
}
SafeRelease(&surface9);
SafeRelease(&texture9);
// _A8 for a yuv plane
/*
* d3d resource share requires windows >= vista: https://msdn.microsoft.com/en-us/library/windows/desktop/bb219800(v=vs.85).aspx
* from extension files:
* d3d9: level must be 1, dimensions must match EGL surface's
* d3d9ex or d3d10:
*/
DX_ENSURE(device9->CreateTexture(w, h, 1,
D3DUSAGE_RENDERTARGET,
has_alpha ? D3DFMT_A8R8G8B8 : D3DFMT_X8R8G8B8,
D3DPOOL_DEFAULT,
&texture9,
&share_handle) , false);
DX_ENSURE(texture9->GetSurfaceLevel(0, &surface9), false);
if (kEGL_ANGLE_d3d_share_handle_client_buffer) {
// requires extension EGL_ANGLE_d3d_share_handle_client_buffer
// egl surface size must match d3d texture's
// d3d9ex or d3d10 is required
EGL_ENSURE((egl->surface = eglCreatePbufferFromClientBuffer(egl->dpy, EGL_D3D_TEXTURE_2D_SHARE_HANDLE_ANGLE, share_handle, egl_cfg, attribs)), false);
qDebug("pbuffer surface from client buffer: %p", egl->surface);
}
return true;
}
