Source code
Revision control
Copy as Markdown
Other Tools
/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
/* vim:set ts=2 sw=2 sts=2 et cindent: */
/* This Source Code Form is subject to the terms of the Mozilla Public
* License, v. 2.0. If a copy of the MPL was not distributed with this
#include "DAV1DDecoder.h"
#include "gfxUtils.h"
#include "ImageContainer.h"
#include "mozilla/StaticPrefs_media.h"
#include "mozilla/TaskQueue.h"
#include "nsThreadUtils.h"
#include "PerformanceRecorder.h"
#include "VideoUtils.h"
#undef LOG
#define LOG(arg, ...) \
DDMOZ_LOG(sPDMLog, mozilla::LogLevel::Debug, "::%s: " arg, __func__, \
##__VA_ARGS__)
namespace mozilla {
static int GetDecodingThreadCount(uint32_t aCodedHeight) {
/**
* Based on the result we print out from the dav1decoder [1], the
* following information shows the number of tiles for AV1 videos served on
* Youtube. Each Tile can be decoded in parallel, so we would like to make
* sure we at least use enough threads to match the number of tiles.
*
* ----------------------------
* | resolution row col total |
* | 480p 2 1 2 |
* | 720p 2 2 4 |
* | 1080p 4 2 8 |
* | 1440p 4 2 8 |
* | 2160p 8 4 32 |
* ----------------------------
*
* Besides the tile thread count, the frame thread count also needs to be
* considered. As we didn't find anything about what the best number is for
* the count of frame thread, just simply use 2 for parallel jobs, which
* is similar with Chromium's implementation. They uses 3 frame threads for
* 720p+ but less tile threads, so we will still use more total threads. In
* addition, their data is measured on 2019, our data should be closer to the
* current real world situation.
* [1]
*/
int tileThreads = 2, frameThreads = 2;
if (aCodedHeight >= 2160) {
tileThreads = 32;
} else if (aCodedHeight >= 1080) {
tileThreads = 8;
} else if (aCodedHeight >= 720) {
tileThreads = 4;
}
return tileThreads * frameThreads;
}
DAV1DDecoder::DAV1DDecoder(const CreateDecoderParams& aParams)
: mInfo(aParams.VideoConfig()),
mTaskQueue(TaskQueue::Create(
GetMediaThreadPool(MediaThreadType::PLATFORM_DECODER),
"Dav1dDecoder")),
mImageContainer(aParams.mImageContainer),
mImageAllocator(aParams.mKnowsCompositor),
mTrackingId(aParams.mTrackingId),
mLowLatency(
aParams.mOptions.contains(CreateDecoderParams::Option::LowLatency)) {}
DAV1DDecoder::~DAV1DDecoder() = default;
RefPtr<MediaDataDecoder::InitPromise> DAV1DDecoder::Init() {
Dav1dSettings settings;
dav1d_default_settings(&settings);
if (mLowLatency) {
settings.max_frame_delay = 1;
}
size_t decoder_threads = 2;
if (mInfo.mDisplay.width >= 2048) {
decoder_threads = 8;
} else if (mInfo.mDisplay.width >= 1024) {
decoder_threads = 4;
}
if (StaticPrefs::media_av1_new_thread_count_strategy()) {
decoder_threads = GetDecodingThreadCount(mInfo.mImage.Height());
}
// Still need to consider the amount of physical cores in order to achieve
// best performance.
settings.n_threads =
static_cast<int>(std::min(decoder_threads, GetNumberOfProcessors()));
if (int32_t count = StaticPrefs::media_av1_force_thread_count(); count > 0) {
settings.n_threads = count;
}
int res = dav1d_open(&mContext, &settings);
if (res < 0) {
return DAV1DDecoder::InitPromise::CreateAndReject(
MediaResult(NS_ERROR_DOM_MEDIA_FATAL_ERR,
RESULT_DETAIL("Couldn't get dAV1d decoder interface.")),
__func__);
}
return DAV1DDecoder::InitPromise::CreateAndResolve(TrackInfo::kVideoTrack,
__func__);
}
RefPtr<MediaDataDecoder::DecodePromise> DAV1DDecoder::Decode(
MediaRawData* aSample) {
return InvokeAsync<MediaRawData*>(mTaskQueue, this, __func__,
&DAV1DDecoder::InvokeDecode, aSample);
}
void ReleaseDataBuffer_s(const uint8_t* buf, void* user_data) {
MOZ_ASSERT(user_data);
MOZ_ASSERT(buf);
DAV1DDecoder* d = static_cast<DAV1DDecoder*>(user_data);
d->ReleaseDataBuffer(buf);
}
void DAV1DDecoder::ReleaseDataBuffer(const uint8_t* buf) {
// The release callback may be called on a different thread defined by the
// third party dav1d execution. In that case post a task into TaskQueue to
// ensure that mDecodingBuffers is only ever accessed on the TaskQueue.
RefPtr<DAV1DDecoder> self = this;
auto releaseBuffer = [self, buf] {
MOZ_ASSERT(self->mTaskQueue->IsCurrentThreadIn());
DebugOnly<bool> found = self->mDecodingBuffers.Remove(buf);
MOZ_ASSERT(found);
};
if (mTaskQueue->IsCurrentThreadIn()) {
releaseBuffer();
} else {
nsresult rv = mTaskQueue->Dispatch(NS_NewRunnableFunction(
"DAV1DDecoder::ReleaseDataBuffer", std::move(releaseBuffer)));
MOZ_DIAGNOSTIC_ASSERT(NS_SUCCEEDED(rv));
Unused << rv;
}
}
RefPtr<MediaDataDecoder::DecodePromise> DAV1DDecoder::InvokeDecode(
MediaRawData* aSample) {
MOZ_ASSERT(mTaskQueue->IsCurrentThreadIn());
MOZ_ASSERT(aSample);
MediaInfoFlag flag = MediaInfoFlag::None;
flag |= (aSample->mKeyframe ? MediaInfoFlag::KeyFrame
: MediaInfoFlag::NonKeyFrame);
flag |= MediaInfoFlag::SoftwareDecoding;
flag |= MediaInfoFlag::VIDEO_AV1;
mTrackingId.apply([&](const auto& aId) {
mPerformanceRecorder.Start(aSample->mTimecode.ToMicroseconds(),
"DAV1DDecoder"_ns, aId, flag);
});
// Add the buffer to the hashtable in order to increase
// the ref counter and keep it alive. When dav1d does not
// need it any more will call it's release callback. Remove
// the buffer, in there, to reduce the ref counter and eventually
// free it. We need a hashtable and not an array because the
// release callback are not coming in the same order that the
// buffers have been added in the decoder (threading ordering
// inside decoder)
mDecodingBuffers.InsertOrUpdate(aSample->Data(), RefPtr{aSample});
Dav1dData data;
int res = dav1d_data_wrap(&data, aSample->Data(), aSample->Size(),
ReleaseDataBuffer_s, this);
data.m.timestamp = aSample->mTimecode.ToMicroseconds();
data.m.duration = aSample->mDuration.ToMicroseconds();
data.m.offset = aSample->mOffset;
if (res < 0) {
LOG("Create decoder data error.");
return DecodePromise::CreateAndReject(
MediaResult(NS_ERROR_OUT_OF_MEMORY, __func__), __func__);
}
DecodedData results;
do {
res = dav1d_send_data(mContext, &data);
if (res < 0 && res != DAV1D_ERR(EAGAIN)) {
LOG("Decode error: %d", res);
return DecodePromise::CreateAndReject(
MediaResult(NS_ERROR_DOM_MEDIA_DECODE_ERR, __func__), __func__);
}
// Alway consume the whole buffer on success.
// At this point only DAV1D_ERR(EAGAIN) is expected.
MOZ_ASSERT((res == 0 && !data.sz) ||
(res == DAV1D_ERR(EAGAIN) && data.sz == aSample->Size()));
Result<already_AddRefed<VideoData>, MediaResult> r = GetPicture();
if (r.isOk()) {
results.AppendElement(r.unwrap());
} else {
MediaResult rs = r.unwrapErr();
if (rs.Code() == NS_ERROR_DOM_MEDIA_WAITING_FOR_DATA) {
// No frames ready to return. This is not an error, in some
// circumstances, we need to feed it with a certain amount of frames
// before we get a picture.
continue;
}
// Skip if rs is NS_OK, which can happen if picture layout is I400.
if (NS_FAILED(rs.Code())) {
return DecodePromise::CreateAndReject(rs, __func__);
}
}
} while (data.sz > 0);
return DecodePromise::CreateAndResolve(std::move(results), __func__);
}
Result<already_AddRefed<VideoData>, MediaResult> DAV1DDecoder::GetPicture() {
class Dav1dPictureWrapper {
public:
Dav1dPicture* operator&() { return &p; }
const Dav1dPicture& operator*() const { return p; }
~Dav1dPictureWrapper() { dav1d_picture_unref(&p); }
private:
Dav1dPicture p = Dav1dPicture();
};
Dav1dPictureWrapper picture;
int res = dav1d_get_picture(mContext, &picture);
if (res < 0) {
auto r = MediaResult(res == DAV1D_ERR(EAGAIN)
? NS_ERROR_DOM_MEDIA_WAITING_FOR_DATA
: NS_ERROR_DOM_MEDIA_DECODE_ERR,
RESULT_DETAIL("dav1d_get_picture: %d", res));
LOG("%s", r.Message().get());
return Err(r);
}
if ((*picture).p.layout == DAV1D_PIXEL_LAYOUT_I400) {
// Use NS_OK to indicate that this picture should be skipped.
auto r = MediaResult(
NS_OK,
RESULT_DETAIL("I400 picture: No chroma data to construct an image"));
LOG("%s", r.Message().get());
return Err(r);
}
Result<already_AddRefed<VideoData>, MediaResult> r = ConstructImage(*picture);
return r.mapErr([&](const MediaResult& aResult) {
LOG("ConstructImage (%ux%u display %ux%u picture %ux%u ) error - %s: %s",
(*picture).p.w, (*picture).p.h, mInfo.mDisplay.width,
mInfo.mDisplay.height, mInfo.mImage.width, mInfo.mImage.height,
aResult.ErrorName().get(), aResult.Message().get());
return aResult;
});
}
/* static */
Maybe<gfx::YUVColorSpace> DAV1DDecoder::GetColorSpace(
const Dav1dPicture& aPicture, LazyLogModule& aLogger) {
// When returning Nothing(), the caller chooses the appropriate default.
if (!aPicture.seq_hdr || !aPicture.seq_hdr->color_description_present) {
return Nothing();
}
return gfxUtils::CicpToColorSpace(
static_cast<gfx::CICP::MatrixCoefficients>(aPicture.seq_hdr->mtrx),
static_cast<gfx::CICP::ColourPrimaries>(aPicture.seq_hdr->pri), aLogger);
}
/* static */
Maybe<gfx::ColorSpace2> DAV1DDecoder::GetColorPrimaries(
const Dav1dPicture& aPicture, LazyLogModule& aLogger) {
// When returning Nothing(), the caller chooses the appropriate default.
if (!aPicture.seq_hdr || !aPicture.seq_hdr->color_description_present) {
return Nothing();
}
return gfxUtils::CicpToColorPrimaries(
static_cast<gfx::CICP::ColourPrimaries>(aPicture.seq_hdr->pri), aLogger);
}
Result<already_AddRefed<VideoData>, MediaResult> DAV1DDecoder::ConstructImage(
const Dav1dPicture& aPicture) {
VideoData::YCbCrBuffer b;
if (aPicture.p.bpc == 10) {
b.mColorDepth = gfx::ColorDepth::COLOR_10;
} else if (aPicture.p.bpc == 12) {
b.mColorDepth = gfx::ColorDepth::COLOR_12;
} else {
b.mColorDepth = gfx::ColorDepth::COLOR_8;
}
b.mYUVColorSpace =
DAV1DDecoder::GetColorSpace(aPicture, sPDMLog)
.valueOr(DefaultColorSpace({aPicture.p.w, aPicture.p.h}));
b.mColorPrimaries = DAV1DDecoder::GetColorPrimaries(aPicture, sPDMLog)
.valueOr(gfx::ColorSpace2::BT709);
b.mColorRange = aPicture.seq_hdr->color_range ? gfx::ColorRange::FULL
: gfx::ColorRange::LIMITED;
b.mPlanes[0].mData = static_cast<uint8_t*>(aPicture.data[0]);
b.mPlanes[0].mStride = aPicture.stride[0];
b.mPlanes[0].mHeight = aPicture.p.h;
b.mPlanes[0].mWidth = aPicture.p.w;
b.mPlanes[0].mSkip = 0;
b.mPlanes[1].mData = static_cast<uint8_t*>(aPicture.data[1]);
b.mPlanes[1].mStride = aPicture.stride[1];
b.mPlanes[1].mSkip = 0;
b.mPlanes[2].mData = static_cast<uint8_t*>(aPicture.data[2]);
b.mPlanes[2].mStride = aPicture.stride[1];
b.mPlanes[2].mSkip = 0;
const int ss_ver = aPicture.p.layout == DAV1D_PIXEL_LAYOUT_I420;
const int ss_hor = aPicture.p.layout != DAV1D_PIXEL_LAYOUT_I444;
b.mPlanes[1].mHeight = (aPicture.p.h + ss_ver) >> ss_ver;
b.mPlanes[1].mWidth = (aPicture.p.w + ss_hor) >> ss_hor;
b.mPlanes[2].mHeight = (aPicture.p.h + ss_ver) >> ss_ver;
b.mPlanes[2].mWidth = (aPicture.p.w + ss_hor) >> ss_hor;
if (ss_ver) {
b.mChromaSubsampling = gfx::ChromaSubsampling::HALF_WIDTH_AND_HEIGHT;
} else if (ss_hor) {
b.mChromaSubsampling = gfx::ChromaSubsampling::HALF_WIDTH;
}
// Timestamp, duration and offset used here are wrong.
// We need to take those values from the decoder. Latest
// dav1d version allows for that.
media::TimeUnit timecode =
media::TimeUnit::FromMicroseconds(aPicture.m.timestamp);
media::TimeUnit duration =
media::TimeUnit::FromMicroseconds(aPicture.m.duration);
int64_t offset = aPicture.m.offset;
bool keyframe = aPicture.frame_hdr->frame_type == DAV1D_FRAME_TYPE_KEY;
mPerformanceRecorder.Record(aPicture.m.timestamp, [&](DecodeStage& aStage) {
aStage.SetResolution(aPicture.p.w, aPicture.p.h);
auto format = [&]() -> Maybe<DecodeStage::ImageFormat> {
switch (aPicture.p.layout) {
case DAV1D_PIXEL_LAYOUT_I420:
return Some(DecodeStage::YUV420P);
case DAV1D_PIXEL_LAYOUT_I422:
return Some(DecodeStage::YUV422P);
case DAV1D_PIXEL_LAYOUT_I444:
return Some(DecodeStage::YUV444P);
default:
return Nothing();
}
}();
format.apply([&](auto& aFmt) { aStage.SetImageFormat(aFmt); });
aStage.SetYUVColorSpace(b.mYUVColorSpace);
aStage.SetColorRange(b.mColorRange);
aStage.SetColorDepth(b.mColorDepth);
});
return VideoData::CreateAndCopyData(
mInfo, mImageContainer, offset, timecode, duration, b, keyframe, timecode,
mInfo.ScaledImageRect(aPicture.p.w, aPicture.p.h), mImageAllocator);
}
RefPtr<MediaDataDecoder::DecodePromise> DAV1DDecoder::Drain() {
RefPtr<DAV1DDecoder> self = this;
return InvokeAsync(mTaskQueue, __func__, [self, this] {
DecodedData results;
while (true) {
Result<already_AddRefed<VideoData>, MediaResult> r = GetPicture();
if (r.isOk()) {
results.AppendElement(r.unwrap());
} else {
MediaResult rs = r.unwrapErr();
if (rs.Code() == NS_ERROR_DOM_MEDIA_WAITING_FOR_DATA) {
break;
}
// Skip if rs is NS_OK, which can happen if picture layout is I400.
if (NS_FAILED(rs.Code())) {
return DecodePromise::CreateAndReject(rs, __func__);
}
}
}
return DecodePromise::CreateAndResolve(std::move(results), __func__);
});
}
RefPtr<MediaDataDecoder::FlushPromise> DAV1DDecoder::Flush() {
RefPtr<DAV1DDecoder> self = this;
return InvokeAsync(mTaskQueue, __func__, [this, self]() {
dav1d_flush(self->mContext);
mPerformanceRecorder.Record(std::numeric_limits<int64_t>::max());
return FlushPromise::CreateAndResolve(true, __func__);
});
}
RefPtr<ShutdownPromise> DAV1DDecoder::Shutdown() {
RefPtr<DAV1DDecoder> self = this;
return InvokeAsync(mTaskQueue, __func__, [self]() {
dav1d_close(&self->mContext);
return self->mTaskQueue->BeginShutdown();
});
}
} // namespace mozilla
#undef LOG