HRESULT CopyLAVFrame(LAVFrame *pSrc, LAVFrame **ppDst)
{
ASSERT(pSrc->format != LAVPixFmt_DXVA2);
*ppDst = (LAVFrame *)CoTaskMemAlloc(sizeof(LAVFrame));
if (!*ppDst) return E_OUTOFMEMORY;
**ppDst = *pSrc;
(*ppDst)->destruct = nullptr;
(*ppDst)->priv_data = nullptr;
HRESULT hr = AllocLAVFrameBuffers(*ppDst);
if (FAILED(hr))
return hr;
LAVPixFmtDesc desc = getPixelFormatDesc(pSrc->format);
for (int plane = 0; plane < desc.planes; plane++) {
size_t linesize = (pSrc->width / desc.planeWidth[plane]) * desc.codedbytes;
BYTE *dst = (*ppDst)->data[plane];
BYTE *src = pSrc->data[plane];
if (!dst || !src)
return E_FAIL;
for (int i = 0; i < (pSrc->height / desc.planeHeight[plane]); i++) {
memcpy(dst, src, linesize);
dst += (*ppDst)->stride[plane];
src += pSrc->stride[plane];
}
if (pSrc->flags & LAV_FRAME_FLAG_MVC) {
dst = (*ppDst)->stereo[plane];
src = pSrc->stereo[plane];
if (!dst || !src)
return E_FAIL;
for (int i = 0; i < (pSrc->height / desc.planeHeight[plane]); i++) {
memcpy(dst, src, linesize);
dst += (*ppDst)->stride[plane];
src += pSrc->stride[plane];
}
}
}
(*ppDst)->side_data = nullptr;
(*ppDst)->side_data_count = 0;
for (int i = 0; i < pSrc->side_data_count; i++)
{
BYTE * p = AddLAVFrameSideData(*ppDst, pSrc->side_data[i].guidType, pSrc->side_data[i].size);
if (p)
memcpy(p, pSrc->side_data[i].data, pSrc->side_data[i].size);
}
return S_OK;
}
void CDecMSDKMVC::GetOffsetSideData(LAVFrame *pFrame, mfxU64 timestamp)
{
MediaSideData3DOffset offset = { 255 };
// Go over all stored GOPs and find an entry for our timestamp
// In general it should be found in the first GOP, unless we lost frames in between or something else went wrong.
for (auto it = m_GOPs.begin(); it != m_GOPs.end(); it++) {
if (RemoveFrameFromGOP(&(*it), timestamp)) {
offset = it->offsets.front();
it->offsets.pop_front();
// Erase previous GOPs when we start accessing a new one
if (it != m_GOPs.begin()) {
#ifdef DEBUG
// Check that all to-be-erased GOPs are empty
for (auto itd = m_GOPs.begin(); itd < it; itd++) {
if (!itd->offsets.empty()) {
DbgLog((LOG_TRACE, 10, L"CDecMSDKMVC::GetOffsetSideData(): Switched to next GOP at %I64u with %Iu entries remaining", timestamp, itd->offsets.size()));
}
}
#endif
m_GOPs.erase(m_GOPs.begin(), it);
}
break;
}
}
if (offset.offset_count == 255) {
DbgLog((LOG_TRACE, 10, L"CDecMSDKMVC::GetOffsetSideData():No offset for frame at %I64u", timestamp));
offset = m_PrevOffset;
}
m_PrevOffset = offset;
// Only set the offset when data is present
if (offset.offset_count > 0) {
MediaSideData3DOffset *FrameOffset = (MediaSideData3DOffset *)AddLAVFrameSideData(pFrame, IID_MediaSideData3DOffset, sizeof(MediaSideData3DOffset));
if (FrameOffset)
*FrameOffset = offset;
}
}
STDMETHODIMP CDecAvcodec::Decode(const BYTE *buffer, int buflen, REFERENCE_TIME rtStartIn, REFERENCE_TIME rtStopIn, BOOL bSyncPoint, BOOL bDiscontinuity)
{
CheckPointer(m_pAVCtx, E_UNEXPECTED);
int got_picture = 0;
int used_bytes = 0;
BOOL bFlush = (buffer == nullptr);
BOOL bEndOfSequence = FALSE;
AVPacket avpkt;
av_init_packet(&avpkt);
if (m_pAVCtx->active_thread_type & FF_THREAD_FRAME) {
if (!m_bFFReordering) {
m_tcThreadBuffer[m_CurrentThread].rtStart = rtStartIn;
m_tcThreadBuffer[m_CurrentThread].rtStop = rtStopIn;
}
m_CurrentThread = (m_CurrentThread + 1) % m_pAVCtx->thread_count;
} else if (m_bBFrameDelay) {
m_tcBFrameDelay[m_nBFramePos].rtStart = rtStartIn;
m_tcBFrameDelay[m_nBFramePos].rtStop = rtStopIn;
m_nBFramePos = !m_nBFramePos;
}
uint8_t *pDataBuffer = nullptr;
if (!bFlush && buflen > 0) {
if (!m_bInputPadded && (!(m_pAVCtx->active_thread_type & FF_THREAD_FRAME) || m_pParser)) {
// Copy bitstream into temporary buffer to ensure overread protection
// Verify buffer size
if (buflen > m_nFFBufferSize) {
m_nFFBufferSize = buflen;
m_pFFBuffer = (BYTE *)av_realloc_f(m_pFFBuffer, m_nFFBufferSize + FF_INPUT_BUFFER_PADDING_SIZE, 1);
if (!m_pFFBuffer) {
m_nFFBufferSize = 0;
return E_OUTOFMEMORY;
}
}
memcpy(m_pFFBuffer, buffer, buflen);
memset(m_pFFBuffer+buflen, 0, FF_INPUT_BUFFER_PADDING_SIZE);
pDataBuffer = m_pFFBuffer;
} else {
pDataBuffer = (uint8_t *)buffer;
}
if (m_nCodecId == AV_CODEC_ID_VP8 && m_bWaitingForKeyFrame) {
if (!(pDataBuffer[0] & 1)) {
DbgLog((LOG_TRACE, 10, L"::Decode(): Found VP8 key-frame, resuming decoding"));
m_bWaitingForKeyFrame = FALSE;
} else {
return S_OK;
}
}
}
while (buflen > 0 || bFlush) {
REFERENCE_TIME rtStart = rtStartIn, rtStop = rtStopIn;
if (!bFlush) {
avpkt.data = pDataBuffer;
avpkt.size = buflen;
avpkt.pts = rtStartIn;
if (rtStartIn != AV_NOPTS_VALUE && rtStopIn != AV_NOPTS_VALUE)
avpkt.duration = (int)(rtStopIn - rtStartIn);
else
avpkt.duration = 0;
avpkt.flags = AV_PKT_FLAG_KEY;
if (m_bHasPalette) {
m_bHasPalette = FALSE;
uint32_t *pal = (uint32_t *)av_packet_new_side_data(&avpkt, AV_PKT_DATA_PALETTE, AVPALETTE_SIZE);
int pal_size = FFMIN((1 << m_pAVCtx->bits_per_coded_sample) << 2, m_pAVCtx->extradata_size);
uint8_t *pal_src = m_pAVCtx->extradata + m_pAVCtx->extradata_size - pal_size;
for (int i = 0; i < pal_size/4; i++)
pal[i] = 0xFF<<24 | AV_RL32(pal_src+4*i);
}
} else {
avpkt.data = nullptr;
avpkt.size = 0;
}
// Parse the data if a parser is present
// This is mandatory for MPEG-1/2
if (m_pParser) {
BYTE *pOut = nullptr;
int pOut_size = 0;
used_bytes = av_parser_parse2(m_pParser, m_pAVCtx, &pOut, &pOut_size, avpkt.data, avpkt.size, AV_NOPTS_VALUE, AV_NOPTS_VALUE, 0);
if (used_bytes == 0 && pOut_size == 0 && !bFlush) {
DbgLog((LOG_TRACE, 50, L"::Decode() - could not process buffer, starving?"));
break;
} else if (used_bytes > 0) {
buflen -= used_bytes;
pDataBuffer += used_bytes;
}
// Update start time cache
// If more data was read then output, update the cache (incomplete frame)
// If output is bigger, a frame was completed, update the actual rtStart with the cached value, and then overwrite the cache
if (used_bytes > pOut_size) {
if (rtStartIn != AV_NOPTS_VALUE)
m_rtStartCache = rtStartIn;
} else if (used_bytes == pOut_size || ((used_bytes + 9) == pOut_size)) {
// Why +9 above?
// Well, apparently there are some broken MKV muxers that like to mux the MPEG-2 PICTURE_START_CODE block (which is 9 bytes) in the package with the previous frame
// This would cause the frame timestamps to be delayed by one frame exactly, and cause timestamp reordering to go wrong.
// So instead of failing on those samples, lets just assume that 9 bytes are that case exactly.
m_rtStartCache = rtStartIn = AV_NOPTS_VALUE;
} else if (pOut_size > used_bytes) {
rtStart = m_rtStartCache;
m_rtStartCache = rtStartIn;
// The value was used once, don't use it for multiple frames, that ends up in weird timings
rtStartIn = AV_NOPTS_VALUE;
}
if (pOut_size > 0 || bFlush) {
if (pOut && pOut_size > 0) {
if (pOut_size > m_nFFBufferSize2) {
m_nFFBufferSize2 = pOut_size;
m_pFFBuffer2 = (BYTE *)av_realloc_f(m_pFFBuffer2, m_nFFBufferSize2 + FF_INPUT_BUFFER_PADDING_SIZE, 1);
if (!m_pFFBuffer2) {
m_nFFBufferSize2 = 0;
return E_OUTOFMEMORY;
}
}
memcpy(m_pFFBuffer2, pOut, pOut_size);
memset(m_pFFBuffer2+pOut_size, 0, FF_INPUT_BUFFER_PADDING_SIZE);
avpkt.data = m_pFFBuffer2;
avpkt.size = pOut_size;
avpkt.pts = rtStart;
avpkt.duration = 0;
const uint8_t *eosmarker = CheckForEndOfSequence(m_nCodecId, avpkt.data, avpkt.size, &m_MpegParserState);
if (eosmarker) {
bEndOfSequence = TRUE;
}
} else {
avpkt.data = nullptr;
avpkt.size = 0;
}
int ret2 = avcodec_decode_video2 (m_pAVCtx, m_pFrame, &got_picture, &avpkt);
if (ret2 < 0) {
DbgLog((LOG_TRACE, 50, L"::Decode() - decoding failed despite successfull parsing"));
got_picture = 0;
}
} else {
got_picture = 0;
}
} else {
used_bytes = avcodec_decode_video2 (m_pAVCtx, m_pFrame, &got_picture, &avpkt);
buflen = 0;
}
if (FAILED(PostDecode())) {
av_frame_unref(m_pFrame);
return E_FAIL;
}
// Decoding of this frame failed ... oh well!
if (used_bytes < 0) {
av_frame_unref(m_pFrame);
return S_OK;
}
// Judge frame usability
// This determines if a frame is artifact free and can be delivered.
if (m_bResumeAtKeyFrame) {
if (m_bWaitingForKeyFrame && got_picture) {
if (m_pFrame->key_frame) {
DbgLog((LOG_TRACE, 50, L"::Decode() - Found Key-Frame, resuming decoding at %I64d", m_pFrame->pkt_pts));
m_bWaitingForKeyFrame = FALSE;
} else {
got_picture = 0;
}
}
}
// Handle B-frame delay for frame threading codecs
if ((m_pAVCtx->active_thread_type & FF_THREAD_FRAME) && m_bBFrameDelay) {
m_tcBFrameDelay[m_nBFramePos] = m_tcThreadBuffer[m_CurrentThread];
m_nBFramePos = !m_nBFramePos;
}
if (!got_picture || !m_pFrame->data[0]) {
if (!avpkt.size)
bFlush = FALSE; // End flushing, no more frames
av_frame_unref(m_pFrame);
continue;
}
///////////////////////////////////////////////////////////////////////////////////////////////
// Determine the proper timestamps for the frame, based on different possible flags.
///////////////////////////////////////////////////////////////////////////////////////////////
if (m_bFFReordering) {
rtStart = m_pFrame->pkt_pts;
if (m_pFrame->pkt_duration)
rtStop = m_pFrame->pkt_pts + m_pFrame->pkt_duration;
else
rtStop = AV_NOPTS_VALUE;
} else if (m_bBFrameDelay && m_pAVCtx->has_b_frames) {
rtStart = m_tcBFrameDelay[m_nBFramePos].rtStart;
rtStop = m_tcBFrameDelay[m_nBFramePos].rtStop;
} else if (m_pAVCtx->active_thread_type & FF_THREAD_FRAME) {
unsigned index = m_CurrentThread;
rtStart = m_tcThreadBuffer[index].rtStart;
rtStop = m_tcThreadBuffer[index].rtStop;
}
if (m_bRVDropBFrameTimings && m_pFrame->pict_type == AV_PICTURE_TYPE_B) {
rtStart = AV_NOPTS_VALUE;
}
if (m_bCalculateStopTime)
rtStop = AV_NOPTS_VALUE;
///////////////////////////////////////////////////////////////////////////////////////////////
// All required values collected, deliver the frame
///////////////////////////////////////////////////////////////////////////////////////////////
LAVFrame *pOutFrame = nullptr;
AllocateFrame(&pOutFrame);
AVRational display_aspect_ratio;
int64_t num = (int64_t)m_pFrame->sample_aspect_ratio.num * m_pFrame->width;
int64_t den = (int64_t)m_pFrame->sample_aspect_ratio.den * m_pFrame->height;
av_reduce(&display_aspect_ratio.num, &display_aspect_ratio.den, num, den, INT_MAX);
pOutFrame->width = m_pFrame->width;
pOutFrame->height = m_pFrame->height;
pOutFrame->aspect_ratio = display_aspect_ratio;
pOutFrame->repeat = m_pFrame->repeat_pict;
pOutFrame->key_frame = m_pFrame->key_frame;
pOutFrame->frame_type = av_get_picture_type_char(m_pFrame->pict_type);
pOutFrame->ext_format = GetDXVA2ExtendedFlags(m_pAVCtx, m_pFrame);
if (m_pFrame->interlaced_frame || (!m_pAVCtx->progressive_sequence && (m_nCodecId == AV_CODEC_ID_H264 || m_nCodecId == AV_CODEC_ID_MPEG2VIDEO)))
m_iInterlaced = 1;
else if (m_pAVCtx->progressive_sequence)
m_iInterlaced = 0;
if ((m_nCodecId == AV_CODEC_ID_H264 || m_nCodecId == AV_CODEC_ID_MPEG2VIDEO) && m_pFrame->repeat_pict)
m_nSoftTelecine = 2;
else if (m_nSoftTelecine > 0)
m_nSoftTelecine--;
// Don't apply aggressive deinterlacing to content that looks soft-telecined, as it would destroy the content
bool bAggressiveFlag = (m_iInterlaced == 1 && m_pSettings->GetDeinterlacingMode() == DeintMode_Aggressive) && !m_nSoftTelecine;
pOutFrame->interlaced = (m_pFrame->interlaced_frame || bAggressiveFlag || m_pSettings->GetDeinterlacingMode() == DeintMode_Force) && !(m_pSettings->GetDeinterlacingMode() == DeintMode_Disable);
LAVDeintFieldOrder fo = m_pSettings->GetDeintFieldOrder();
pOutFrame->tff = (fo == DeintFieldOrder_Auto) ? m_pFrame->top_field_first : (fo == DeintFieldOrder_TopFieldFirst);
pOutFrame->rtStart = rtStart;
pOutFrame->rtStop = rtStop;
PixelFormatMapping map = getPixFmtMapping((AVPixelFormat)m_pFrame->format);
pOutFrame->format = map.lavpixfmt;
pOutFrame->bpp = map.bpp;
if (m_nCodecId == AV_CODEC_ID_MPEG2VIDEO || m_nCodecId == AV_CODEC_ID_MPEG1VIDEO)
pOutFrame->avgFrameDuration = GetFrameDuration();
AVFrameSideData * sdHDR = av_frame_get_side_data(m_pFrame, AV_FRAME_DATA_HDR_MASTERING_INFO);
if (sdHDR) {
if (sdHDR->size == 24) {
MediaSideDataHDR * hdr = (MediaSideDataHDR *)AddLAVFrameSideData(pOutFrame, IID_MediaSideDataHDR, sizeof(MediaSideDataHDR));
if (hdr) {
CByteParser hdrParser(sdHDR->data, sdHDR->size);
for (int i = 0; i < 3; i++)
{
hdr->display_primaries_x[i] = hdrParser.BitRead(16) * 0.00002;
hdr->display_primaries_y[i] = hdrParser.BitRead(16) * 0.00002;
}
hdr->white_point_x = hdrParser.BitRead(16) * 0.00002;
hdr->white_point_y = hdrParser.BitRead(16) * 0.00002;
hdr->max_display_mastering_luminance = hdrParser.BitRead(32) * 0.0001;
hdr->min_display_mastering_luminance = hdrParser.BitRead(32) * 0.0001;
}
}
else {
DbgLog((LOG_TRACE, 10, L"::Decode(): Found HDR data of an unexpected size (%d)", sdHDR->size));
}
}
if (map.conversion) {
ConvertPixFmt(m_pFrame, pOutFrame);
} else {
AVFrame *pFrameRef = av_frame_alloc();
av_frame_ref(pFrameRef, m_pFrame);
for (int i = 0; i < 4; i++) {
pOutFrame->data[i] = pFrameRef->data[i];
pOutFrame->stride[i] = pFrameRef->linesize[i];
}
pOutFrame->priv_data = pFrameRef;
pOutFrame->destruct = lav_avframe_free;
// Check alignment on rawvideo, which can be off depending on the source file
if (m_nCodecId == AV_CODEC_ID_RAWVIDEO) {
for (int i = 0; i < 4; i++) {
if ((intptr_t)pOutFrame->data[i] % 16u || pOutFrame->stride[i] % 16u) {
// copy the frame, its not aligned properly and would crash later
CopyLAVFrameInPlace(pOutFrame);
break;
}
}
}
}
if (bEndOfSequence)
pOutFrame->flags |= LAV_FRAME_FLAG_END_OF_SEQUENCE;
if (pOutFrame->format == LAVPixFmt_DXVA2) {
pOutFrame->data[0] = m_pFrame->data[4];
HandleDXVA2Frame(pOutFrame);
} else {
Deliver(pOutFrame);
}
if (bEndOfSequence) {
bEndOfSequence = FALSE;
if (pOutFrame->format == LAVPixFmt_DXVA2) {
HandleDXVA2Frame(m_pCallback->GetFlushFrame());
} else {
Deliver(m_pCallback->GetFlushFrame());
}
}
if (bFlush) {
m_CurrentThread = (m_CurrentThread + 1) % m_pAVCtx->thread_count;
}
av_frame_unref(m_pFrame);
}
return S_OK;
}
