cuda::GpuMat mapFrame(int picIdx, CUVIDPROCPARAMS& videoProcParams)
{
CUdeviceptr ptr;
unsigned int pitch;
cuSafeCall( cuvidMapVideoFrame(decoder_, picIdx, &ptr, &pitch, &videoProcParams) );
return cuda::GpuMat(targetHeight() * 3 / 2, targetWidth(), CV_8UC1, (void*) ptr, pitch);
}
void
VideoDecoder::mapFrame(int iPictureIndex, CUdeviceptr *ppDevice, unsigned int *pPitch, CUVIDPROCPARAMS *pVideoProcessingParameters)
{
CUresult oResult = cuvidMapVideoFrame(oDecoder_,
iPictureIndex,
ppDevice,
pPitch, pVideoProcessingParameters);
assert(CUDA_SUCCESS == oResult);
assert(0 != *ppDevice);
assert(0 != *pPitch);
}
void* InteropResource::mapToHost(const VideoFormat &format, void *handle, int picIndex, const CUVIDPROCPARAMS ¶m, int width, int height, int coded_height)
{
AutoCtxLock locker((cuda_api*)this, lock);
Q_UNUSED(locker);
CUdeviceptr devptr;
unsigned int pitch;
CUDA_ENSURE(cuvidMapVideoFrame(dec, picIndex, &devptr, &pitch, const_cast<CUVIDPROCPARAMS*>(¶m)), NULL);
CUVIDAutoUnmapper unmapper(this, dec, devptr);
Q_UNUSED(unmapper);
uchar* host_data = NULL;
const size_t host_size = pitch*coded_height*3/2;
CUDA_ENSURE(cuMemAllocHost((void**)&host_data, host_size), NULL);
// copy to the memory not allocated by cuda is possible but much slower
CUDA_ENSURE(cuMemcpyDtoH(host_data, devptr, host_size), NULL);
VideoFrame frame(width, height, VideoFormat::Format_NV12);
uchar *planes[] = {
host_data,
host_data + pitch * coded_height
};
frame.setBits(planes);
int pitches[] = { (int)pitch, (int)pitch };
frame.setBytesPerLine(pitches);
VideoFrame *f = reinterpret_cast<VideoFrame*>(handle);
frame.setTimestamp(f->timestamp());
frame.setDisplayAspectRatio(f->displayAspectRatio());
if (format == frame.format())
*f = frame.clone();
else
*f = frame.to(format);
cuMemFreeHost(host_data);
return f;
}
bool VideoDecoderCUDAPrivate::processDecodedData(CUVIDPARSERDISPINFO *cuviddisp, VideoFrame* outFrame) {
int num_fields = cuviddisp->progressive_frame ? 1 : 2+cuviddisp->repeat_first_field;
for (int active_field = 0; active_field < num_fields; ++active_field) {
CUVIDPROCPARAMS proc_params;
memset(&proc_params, 0, sizeof(CUVIDPROCPARAMS));
proc_params.progressive_frame = cuviddisp->progressive_frame; //check user config
proc_params.second_field = active_field == 1; //check user config
proc_params.top_field_first = cuviddisp->top_field_first;
proc_params.unpaired_field = cuviddisp->progressive_frame == 1;
CUdeviceptr devptr;
unsigned int pitch;
cuvidCtxLock(vid_ctx_lock, 0);
CUresult cuStatus = cuvidMapVideoFrame(dec, cuviddisp->picture_index, &devptr, &pitch, &proc_params);
if (cuStatus != CUDA_SUCCESS) {
qWarning("cuvidMapVideoFrame failed on index %d (%#x, %s)", cuviddisp->picture_index, cuStatus, _cudaGetErrorEnum(cuStatus));
cuvidUnmapVideoFrame(dec, devptr);
cuvidCtxUnlock(vid_ctx_lock, 0);
return false;
}
#define PAD_ALIGN(x,mask) ( (x + mask) & ~mask )
//uint w = dec_create_info.ulWidth;//PAD_ALIGN(dec_create_info.ulWidth, 0x3F);
uint h = dec_create_info.ulHeight;//PAD_ALIGN(dec_create_info.ulHeight, 0x0F); //?
#undef PAD_ALIGN
int size = pitch*h*3/2;
if (size > host_data_size && host_data) {
cuMemFreeHost(host_data);
host_data = 0;
host_data_size = 0;
}
if (!host_data) {
cuStatus = cuMemAllocHost((void**)&host_data, size);
if (cuStatus != CUDA_SUCCESS) {
qWarning("cuMemAllocHost failed (%#x, %s)", cuStatus, _cudaGetErrorEnum(cuStatus));
cuvidUnmapVideoFrame(dec, devptr);
cuvidCtxUnlock(vid_ctx_lock, 0);
return false;
}
host_data_size = size;
}
if (!host_data) {
qWarning("No valid staging memory!");
cuvidUnmapVideoFrame(dec, devptr);
cuvidCtxUnlock(vid_ctx_lock, 0);
return false;
}
cuStatus = cuMemcpyDtoHAsync(host_data, devptr, size, stream);
if (cuStatus != CUDA_SUCCESS) {
qWarning("cuMemcpyDtoHAsync failed (%#x, %s)", cuStatus, _cudaGetErrorEnum(cuStatus));
cuvidUnmapVideoFrame(dec, devptr);
cuvidCtxUnlock(vid_ctx_lock, 0);
return false;
}
cuStatus = cuCtxSynchronize();
if (cuStatus != CUDA_SUCCESS) {
qWarning("cuCtxSynchronize failed (%#x, %s)", cuStatus, _cudaGetErrorEnum(cuStatus));
}
cuvidUnmapVideoFrame(dec, devptr);
cuvidCtxUnlock(vid_ctx_lock, 0);
//qDebug("mark not in use pic_index: %d", cuviddisp->picture_index);
surface_in_use[cuviddisp->picture_index] = false;
uchar *planes[] = {
host_data,
host_data + pitch * h
};
int pitches[] = { (int)pitch, (int)pitch };
VideoFrame frame(codec_ctx->width, codec_ctx->height, VideoFormat::Format_NV12);
frame.setBits(planes);
frame.setBytesPerLine(pitches);
//TODO: is clone required? may crash on clone, I should review clone()
//frame = frame.clone();
if (outFrame) {
*outFrame = frame.clone();
}
#if COPY_ON_DECODE
frame_queue.put(frame.clone());
#endif
//qDebug("frame queue size: %d", frame_queue.size());
}
return true;
}
static int cuvid_output_frame(AVCodecContext *avctx, AVFrame *frame)
{
CuvidContext *ctx = avctx->priv_data;
AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
CUdeviceptr mapped_frame = 0;
int ret = 0, eret = 0;
av_log(avctx, AV_LOG_TRACE, "cuvid_output_frame\n");
if (ctx->decoder_flushing) {
ret = cuvid_decode_packet(avctx, NULL);
if (ret < 0 && ret != AVERROR_EOF)
return ret;
}
ret = CHECK_CU(cuCtxPushCurrent(cuda_ctx));
if (ret < 0)
return ret;
if (av_fifo_size(ctx->frame_queue)) {
CuvidParsedFrame parsed_frame;
CUVIDPROCPARAMS params;
unsigned int pitch = 0;
int offset = 0;
int i;
av_fifo_generic_read(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
memset(¶ms, 0, sizeof(params));
params.progressive_frame = parsed_frame.dispinfo.progressive_frame;
params.second_field = parsed_frame.second_field;
params.top_field_first = parsed_frame.dispinfo.top_field_first;
ret = CHECK_CU(cuvidMapVideoFrame(ctx->cudecoder, parsed_frame.dispinfo.picture_index, &mapped_frame, &pitch, ¶ms));
if (ret < 0)
goto error;
if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
ret = av_hwframe_get_buffer(ctx->hwframe, frame, 0);
if (ret < 0) {
av_log(avctx, AV_LOG_ERROR, "av_hwframe_get_buffer failed\n");
goto error;
}
ret = ff_decode_frame_props(avctx, frame);
if (ret < 0) {
av_log(avctx, AV_LOG_ERROR, "ff_decode_frame_props failed\n");
goto error;
}
for (i = 0; i < 2; i++) {
CUDA_MEMCPY2D cpy = {
.srcMemoryType = CU_MEMORYTYPE_DEVICE,
.dstMemoryType = CU_MEMORYTYPE_DEVICE,
.srcDevice = mapped_frame,
.dstDevice = (CUdeviceptr)frame->data[i],
.srcPitch = pitch,
.dstPitch = frame->linesize[i],
.srcY = offset,
.WidthInBytes = FFMIN(pitch, frame->linesize[i]),
.Height = avctx->height >> (i ? 1 : 0),
};
ret = CHECK_CU(cuMemcpy2D(&cpy));
if (ret < 0)
goto error;
offset += avctx->coded_height;
}
} else if (avctx->pix_fmt == AV_PIX_FMT_NV12) {
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;
}
/*
* Main transcoding thread
* Initializes CUDA device, decodes frames with NVCUVID API and adds them to frame queue, which passes them to NVENC for encoding, then output
*/
void NVENCGUI::Transcode()
{
CUresult result;
// initialize CUDA
result = cuInit(0);
if (result != CUDA_SUCCESS)
{
emit Error(ERR_CUDA_INIT);
return;
}
NVENCSTATUS nvStatus = NV_ENC_SUCCESS;
// no input file
if (encodeConfig.inputFileName == NULL)
{
emit Error(ERR_INPUT);
return;
}
// no output file
if (encodeConfig.outputFileName == NULL)
{
emit Error(ERR_OUTPUT);
return;
}
// unable to open input file
if (!fopen(encodeConfig.inputFileName, "r"))
{
emit Error(ERR_INPUT);
return;
}
encodeConfig.fOutput = fopen(encodeConfig.outputFileName, "wb");
// unable to open output file
if (encodeConfig.fOutput == NULL)
{
emit Error(ERR_OUTPUT);
return;
}
// initialize CUDA on device and set CUDA context
CUcontext cudaCtx;
CUdevice device;
result = cuDeviceGet(&device, encodeConfig.deviceID);
if (result != CUDA_SUCCESS)
{
emit Error(ERR_CUDA_DEVICE);
return;
}
result = cuCtxCreate(&cudaCtx, CU_CTX_SCHED_AUTO, device);
if (result != CUDA_SUCCESS)
{
emit Error(ERR_CUDA_CTX);
return;
}
// initialize NVCUVID context
CUcontext curCtx;
CUvideoctxlock ctxLock;
result = cuCtxPopCurrent(&curCtx);
if (result != CUDA_SUCCESS)
{
emit Error(ERR_CUDA_CTX);
return;
}
result = cuvidCtxLockCreate(&ctxLock, curCtx);
if (result != CUDA_SUCCESS)
{
emit Error(ERR_CUDA_CTX);
return;
}
CudaDecoder* pDecoder = new CudaDecoder;
FrameQueue* pFrameQueue = new CUVIDFrameQueue(ctxLock);
pDecoder->InitVideoDecoder(encodeConfig.inputFileName, ctxLock, pFrameQueue, encodeConfig.width, encodeConfig.height);
int decodedW, decodedH, decodedFRN, decodedFRD;
pDecoder->GetCodecParam(&decodedW, &decodedH, &decodedFRN, &decodedFRD);
// If the width/height is not set, set to same as source
if (encodeConfig.width <= 0 || encodeConfig.height <= 0) {
encodeConfig.width = decodedW;
encodeConfig.height = decodedH;
}
// same, except for fps
if (encodeConfig.fps <= 0) {
if (decodedFRN <= 0 || decodedFRD <= 0)
encodeConfig.fps = 30;
else
encodeConfig.fps = decodedFRN / decodedFRD;
}
// initialize frame queue with width/height
pFrameQueue->init(encodeConfig.width, encodeConfig.height);
VideoEncoder* pEncoder = new VideoEncoder(ctxLock);
assert(pEncoder->GetHWEncoder());
// initialize NVENC HW Encoder
nvStatus = pEncoder->GetHWEncoder()->Initialize(cudaCtx, NV_ENC_DEVICE_TYPE_CUDA);
if (nvStatus != NV_ENC_SUCCESS)
{
emit Error(ERR_NVENC_ENC_INIT);
return;
}
// get preset GUID
encodeConfig.presetGUID = pEncoder->GetHWEncoder()->GetPresetGUID(encodeConfig.encoderPreset, encodeConfig.codec);
// create encoder
nvStatus = pEncoder->GetHWEncoder()->CreateEncoder(&encodeConfig);
if (nvStatus != NV_ENC_SUCCESS)
{
emit Error(ERR_NVENC_ENC_CREATE);
return;
}
// create buffer
nvStatus = pEncoder->AllocateIOBuffers(&encodeConfig);
if (nvStatus != NV_ENC_SUCCESS)
{
emit Error(ERR_NVENC_ENC_BUFFER);
return;
}
// print details to text window, start counter
emit PrintDetails();
NvQueryPerformanceCounter(&results.lStart);
//start decoding thread
#ifdef _WIN32
HANDLE decodeThread = CreateThread(NULL, 0, DecodeProc, (LPVOID)pDecoder, 0, NULL);
#else
pthread_t pid;
pthread_create(&pid, NULL, DecodeProc, (void*)pDecoder);
#endif
int encodedFrames = 0;
//start encoding thread
while (!(pFrameQueue->isEndOfDecode() && pFrameQueue->isEmpty()))
{
CUVIDPARSERDISPINFO pInfo;
if (pFrameQueue->dequeue(&pInfo))
{
CUdeviceptr dMappedFrame = 0;
unsigned int pitch;
CUVIDPROCPARAMS oVPP = { 0 };
oVPP.unpaired_field = 1;
oVPP.progressive_frame = 1;
cuvidMapVideoFrame(pDecoder->GetDecoder(), pInfo.picture_index, &dMappedFrame, &pitch, &oVPP);
EncodeFrameConfig stEncodeConfig = { 0 };
stEncodeConfig.dptr = dMappedFrame;
stEncodeConfig.pitch = pitch;
stEncodeConfig.width = encodeConfig.width;
stEncodeConfig.height = encodeConfig.height;
pEncoder->EncodeFrame(&stEncodeConfig);
cuvidUnmapVideoFrame(pDecoder->GetDecoder(), dMappedFrame);
pFrameQueue->releaseFrame(&pInfo);
//emit IncrementEncodedFrames();
}
}
// flush
pEncoder->EncodeFrame(NULL, true);
// end decoding thread
#ifdef _WIN32
WaitForSingleObject(decodeThread, INFINITE);
#else
pthread_join(pid, NULL);
#endif
// print transcoding details
if (pEncoder->GetEncodedFrames() > 0)
{
results.decodedFrames = pDecoder->m_decodedFrames;
results.encodedFrames = pEncoder->GetEncodedFrames();
NvQueryPerformanceCounter(&results.lEnd);
NvQueryPerformanceFrequency(&results.lFreq);
results.elapsedTime = (double)(results.lEnd - results.lStart) / (double)results.lFreq;
}
emit TranscodingEnd();
// clean up
cuvidCtxLockDestroy(ctxLock);
pEncoder->Deinitialize();
delete pDecoder;
delete pEncoder;
delete pFrameQueue;
result = cuCtxDestroy(cudaCtx);
if (result != CUDA_SUCCESS)
{
emit Error(ERR_CUDA_CTX_DESTROY);
return;
}
return;
}
static int cuvid_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *avpkt)
{
CuvidContext *ctx = avctx->priv_data;
AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
AVFrame *frame = data;
CUVIDSOURCEDATAPACKET cupkt;
AVPacket filter_packet = { 0 };
AVPacket filtered_packet = { 0 };
CUdeviceptr mapped_frame = 0;
int ret = 0, eret = 0;
if (ctx->bsf && avpkt->size) {
if ((ret = av_packet_ref(&filter_packet, avpkt)) < 0) {
av_log(avctx, AV_LOG_ERROR, "av_packet_ref failed\n");
return ret;
}
if ((ret = av_bsf_send_packet(ctx->bsf, &filter_packet)) < 0) {
av_log(avctx, AV_LOG_ERROR, "av_bsf_send_packet failed\n");
av_packet_unref(&filter_packet);
return ret;
}
if ((ret = av_bsf_receive_packet(ctx->bsf, &filtered_packet)) < 0) {
av_log(avctx, AV_LOG_ERROR, "av_bsf_receive_packet failed\n");
return ret;
}
avpkt = &filtered_packet;
}
ret = CHECK_CU(cuCtxPushCurrent(cuda_ctx));
if (ret < 0) {
av_packet_unref(&filtered_packet);
return ret;
}
memset(&cupkt, 0, sizeof(cupkt));
if (avpkt->size) {
cupkt.payload_size = avpkt->size;
cupkt.payload = avpkt->data;
if (avpkt->pts != AV_NOPTS_VALUE) {
cupkt.flags = CUVID_PKT_TIMESTAMP;
if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
cupkt.timestamp = av_rescale_q(avpkt->pts, avctx->pkt_timebase, (AVRational){1, 10000000});
else
cupkt.timestamp = avpkt->pts;
}
} else {
cupkt.flags = CUVID_PKT_ENDOFSTREAM;
}
ret = CHECK_CU(cuvidParseVideoData(ctx->cuparser, &cupkt));
av_packet_unref(&filtered_packet);
if (ret < 0) {
goto error;
}
// cuvidParseVideoData doesn't return an error just because stuff failed...
if (ctx->internal_error) {
av_log(avctx, AV_LOG_ERROR, "cuvid decode callback error\n");
ret = ctx->internal_error;
goto error;
}
if (av_fifo_size(ctx->frame_queue)) {
CUVIDPARSERDISPINFO dispinfo;
CUVIDPROCPARAMS params;
unsigned int pitch = 0;
int offset = 0;
int i;
av_fifo_generic_read(ctx->frame_queue, &dispinfo, sizeof(CUVIDPARSERDISPINFO), NULL);
memset(¶ms, 0, sizeof(params));
params.progressive_frame = dispinfo.progressive_frame;
params.second_field = 0;
params.top_field_first = dispinfo.top_field_first;
ret = CHECK_CU(cuvidMapVideoFrame(ctx->cudecoder, dispinfo.picture_index, &mapped_frame, &pitch, ¶ms));
if (ret < 0)
goto error;
if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
ret = av_hwframe_get_buffer(ctx->hwframe, frame, 0);
if (ret < 0) {
av_log(avctx, AV_LOG_ERROR, "av_hwframe_get_buffer failed\n");
goto error;
}
ret = ff_decode_frame_props(avctx, frame);
if (ret < 0) {
av_log(avctx, AV_LOG_ERROR, "ff_decode_frame_props failed\n");
goto error;
}
for (i = 0; i < 2; i++) {
CUDA_MEMCPY2D cpy = {
.srcMemoryType = CU_MEMORYTYPE_DEVICE,
.dstMemoryType = CU_MEMORYTYPE_DEVICE,
.srcDevice = mapped_frame,
.dstDevice = (CUdeviceptr)frame->data[i],
.srcPitch = pitch,
.dstPitch = frame->linesize[i],
.srcY = offset,
.WidthInBytes = FFMIN(pitch, frame->linesize[i]),
.Height = avctx->coded_height >> (i ? 1 : 0),
};
ret = CHECK_CU(cuMemcpy2D(&cpy));
if (ret < 0)
goto error;
offset += avctx->coded_height;
}
} else if (avctx->pix_fmt == AV_PIX_FMT_NV12) {
