jyapayne/moonlight-qt
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Optimize CUDA support to avoid roundtrip to CPU memory

Browse source
This commit is contained in:
Cameron Gutman 2021-12-06 18:22:39 -06:00
commit 3e9aea1f7a
5 changed files with 141 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

101
app/streaming/video/ffmpeg-renderers/cuda.cpp
View file

@ -1,5 +1,13 @@
#include "cuda.h"
#include <ffnvcodec/dynlink_loader.h>
#include <SDL_opengl.h>
extern "C" {
#include <libavutil/hwcontext_cuda.h>
}
CUDARenderer::CUDARenderer()
: m_HwContext(nullptr)
{
@ -55,3 +63,96 @@ bool CUDARenderer::isDirectRenderingSupported()
return false;
}
bool CUDARenderer::copyCudaFrameToBoundTexture(AVFrame* frame)
{
static CudaFunctions* funcs;
CUresult err;
AVCUDADeviceContext* devCtx = (AVCUDADeviceContext*)(((AVHWFramesContext*)frame->hw_frames_ctx->data)->device_ctx->hwctx);
bool ret = false;
if (!funcs) {
// One-time init of CUDA library
cuda_load_functions(&funcs, nullptr);
if (!funcs) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "Failed to initialize CUDA library");
return false;
}
}
SDL_assert(frame->format == AV_PIX_FMT_CUDA);
// Push FFmpeg's CUDA context to use for our CUDA operations
err = funcs->cuCtxPushCurrent(devCtx->cuda_ctx);
if (err != CUDA_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "cuCtxPushCurrent() failed: %d", err);
return false;
}
// NV12 has 2 planes
for (int i = 0; i < 2; i++) {
CUgraphicsResource cudaResource;
CUarray cudaArray;
GLint tex;
// Get the ID of this plane's texture
glActiveTexture(GL_TEXTURE0 + i);
glGetIntegerv(GL_TEXTURE_BINDING_2D, &tex);
// Register it with CUDA
err = funcs->cuGraphicsGLRegisterImage(&cudaResource, tex, GL_TEXTURE_2D, CU_GRAPHICS_REGISTER_FLAGS_WRITE_DISCARD);
if (err != CUDA_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "cuGraphicsGLRegisterImage() failed: %d", err);
goto Exit;
}
// Map it to allow us to use it as a copy destination
err = funcs->cuGraphicsMapResources(1, &cudaResource, devCtx->stream);
if (err != CUDA_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "cuGraphicsMapResources() failed: %d", err);
funcs->cuGraphicsUnregisterResource(cudaResource);
goto Exit;
}
// Get a pointer to the mapped array
err = funcs->cuGraphicsSubResourceGetMappedArray(&cudaArray, cudaResource, 0, 0);
if (err != CUDA_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "cuGraphicsSubResourceGetMappedArray() failed: %d", err);
funcs->cuGraphicsUnmapResources(1, &cudaResource, devCtx->stream);
funcs->cuGraphicsUnregisterResource(cudaResource);
goto Exit;
}
CUDA_MEMCPY2D cu2d = {
.srcMemoryType = CU_MEMORYTYPE_DEVICE,
.srcDevice = (CUdeviceptr)frame->data[i],
.srcPitch = (size_t)frame->linesize[i],
.dstMemoryType = CU_MEMORYTYPE_ARRAY,
.dstArray = cudaArray,
.dstPitch = (size_t)frame->width >> i,
.WidthInBytes = (size_t)frame->width,
.Height = (size_t)frame->height >> i
};
// Do the copy
err = funcs->cuMemcpy2D(&cu2d);
if (err != CUDA_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "cuMemcpy2D() failed: %d", err);
funcs->cuGraphicsUnmapResources(1, &cudaResource, devCtx->stream);
funcs->cuGraphicsUnregisterResource(cudaResource);
goto Exit;
}
funcs->cuGraphicsUnmapResources(1, &cudaResource, devCtx->stream);
funcs->cuGraphicsUnregisterResource(cudaResource);
}
ret = true;
Exit:
{
CUcontext dummy;
funcs->cuCtxPopCurrent(&dummy);
}
return ret;
}