VideoFrame VideoDecoderFFmpegHW::copyToFrame(const VideoFormat& fmt, int surface_h, quint8 *src[], int pitch[], bool swapUV)
{
DPTR_D(VideoDecoderFFmpegHW);
Q_ASSERT_X(src[0] && pitch[0] > 0, "VideoDecoderFFmpegHW::copyToFrame", "src[0] and pitch[0] must be set");
const int nb_planes = fmt.planeCount();
const int chroma_pitch = nb_planes > 1 ? fmt.bytesPerLine(pitch[0], 1) : 0;
const int chroma_h = fmt.chromaHeight(surface_h);
int h[] = { surface_h, 0, 0};
for (int i = 1; i < nb_planes; ++i) {
h[i] = chroma_h;
// set chroma address and pitch if not set
if (pitch[i] <= 0)
pitch[i] = chroma_pitch;
if (!src[i])
src[i] = src[i-1] + pitch[i-1]*h[i-1];
}
if (swapUV) {
std::swap(src[1], src[2]);
std::swap(pitch[1], pitch[2]);
}
VideoFrame frame;
if (copyMode() == VideoDecoderFFmpegHW::OptimizedCopy && d.gpu_mem.isReady()) {
int yuv_size = 0;
for (int i = 0; i < nb_planes; ++i) {
yuv_size += pitch[i]*h[i];
}
// additional 15 bytes to ensure 16 bytes aligned
QByteArray buf(15 + yuv_size, 0);
const int offset_16 = (16 - ((uintptr_t)buf.data() & 0x0f)) & 0x0f;
// plane 1, 2... is aligned?
uchar* plane_ptr = (uchar*)buf.data() + offset_16;
QVector<uchar*> dst(nb_planes, 0);
for (int i = 0; i < nb_planes; ++i) {
dst[i] = plane_ptr;
// TODO: add VideoFormat::planeWidth/Height() ?
// pitch instead of surface_width
plane_ptr += pitch[i] * h[i];
d.gpu_mem.copyFrame(src[i], dst[i], pitch[i], h[i], pitch[i]);
}
frame = VideoFrame(buf, width(), height(), fmt);
frame.setBits(dst);
frame.setBytesPerLine(pitch);
} else {
frame = VideoFrame(width(), height(), fmt);
frame.setBits(src);
frame.setBytesPerLine(pitch);
// TODO: why clone is faster()?
// TODO: buffer pool and create VideoFrame when needed to avoid copy? also for other va
frame = frame.clone();
}
frame.setTimestamp(double(d.frame->pkt_pts)/1000.0);
frame.setDisplayAspectRatio(d.getDAR(d.frame));
d.updateColorDetails(&frame);
return frame;
}
VideoFrame VideoDecoderVAAPI::frame()
{
DPTR_D(VideoDecoderVAAPI);
if (!d.frame->opaque || !d.frame->data[0])
return VideoFrame();
VASurfaceID surface_id = (VASurfaceID)(uintptr_t)d.frame->data[3];
VAStatus status = VA_STATUS_SUCCESS;
if (display() == GLX) {
d.surface_interop->setSurface((va_surface_t*)d.frame->opaque, d.surface_width, d.surface_height);
VideoFrame f(d.surface_width, d.surface_height, VideoFormat::Format_RGB32);
f.setBytesPerLine(d.surface_width*4); //used by gl to compute texture size
f.setSurfaceInterop(d.surface_interop);
return f;
}
#if VA_CHECK_VERSION(0,31,0)
if ((status = vaSyncSurface(d.display, surface_id)) != VA_STATUS_SUCCESS) {
qWarning("vaSyncSurface(VADisplay:%p, VASurfaceID:%#x) == %#x", d.display, surface_id, status);
#else
if (vaSyncSurface(d.display, d.context_id, surface_id)) {
qWarning("vaSyncSurface(VADisplay:%#x, VAContextID:%#x, VASurfaceID:%#x) == %#x", d.display, d.context_id, surface_id, status);
#endif
return VideoFrame();
}
if (!d.disable_derive && d.supports_derive) {
/*
* http://web.archiveorange.com/archive/v/OAywENyq88L319OcRnHI
* vaDeriveImage is faster than vaGetImage. But VAImage is uncached memory and copying from it would be terribly slow
* TODO: copy from USWC, see vlc and https://github.com/OpenELEC/OpenELEC.tv/pull/2937.diff
* https://software.intel.com/en-us/articles/increasing-memory-throughput-with-intel-streaming-simd-extensions-4-intel-sse4-streaming-load
*/
status = vaDeriveImage(d.display, surface_id, &d.image);
if (status != VA_STATUS_SUCCESS) {
qWarning("vaDeriveImage(VADisplay:%p, VASurfaceID:%#x, VAImage*:%p) == %#x", d.display, surface_id, &d.image, status);
return VideoFrame();
}
} else {
status = vaGetImage(d.display, surface_id, 0, 0, d.surface_width, d.surface_height, d.image.image_id);
if (status != VA_STATUS_SUCCESS) {
qWarning("vaGetImage(VADisplay:%p, VASurfaceID:%#x, 0,0, %d, %d, VAImageID:%#x) == %#x", d.display, surface_id, d.surface_width, d.surface_height, d.image.image_id, status);
return VideoFrame();
}
}
void *p_base;
if ((status = vaMapBuffer(d.display, d.image.buf, &p_base)) != VA_STATUS_SUCCESS) {
qWarning("vaMapBuffer(VADisplay:%p, VABufferID:%#x, pBuf:%p) == %#x", d.display, d.image.buf, &p_base, status);
return VideoFrame();
}
VideoFormat::PixelFormat pixfmt = VideoFormat::Format_Invalid;
bool swap_uv = false;
switch (d.image.format.fourcc) {
case VA_FOURCC_YV12:
swap_uv |= d.disable_derive || !d.supports_derive;
pixfmt = VideoFormat::Format_YUV420P;
break;
case VA_FOURCC_IYUV:
swap_uv = true;
pixfmt = VideoFormat::Format_YUV420P;
break;
case VA_FOURCC_NV12:
pixfmt = VideoFormat::Format_NV12;
break;
default:
break;
}
if (pixfmt == VideoFormat::Format_Invalid) {
qWarning("unsupported vaapi pixel format: %#x", d.image.format.fourcc);
return VideoFrame();
}
const VideoFormat fmt(pixfmt);
uint8_t *src[3];
int pitch[3];
for (int i = 0; i < fmt.planeCount(); ++i) {
src[i] = (uint8_t*)p_base + d.image.offsets[i];
pitch[i] = d.image.pitches[i];
}
if (swap_uv) {
std::swap(src[1], src[2]);
std::swap(pitch[1], pitch[2]);
}
VideoFrame frame;
if (d.copy_uswc && d.gpu_mem.isReady()) {
int yuv_size = 0;
if (pixfmt == VideoFormat::Format_NV12)
yuv_size = pitch[0]*d.surface_height*3/2;
else
yuv_size = pitch[0]*d.surface_height + pitch[1]*d.surface_height/2 + pitch[2]*d.surface_height/2;
// additional 15 bytes to ensure 16 bytes aligned
QByteArray buf(15 + yuv_size, 0);
const int offset_16 = (16 - ((uintptr_t)buf.data() & 0x0f)) & 0x0f;
// plane 1, 2... is aligned?
uchar* plane_ptr = (uchar*)buf.data() + offset_16;
QVector<uchar*> dst(fmt.planeCount(), 0);
for (int i = 0; i < dst.size(); ++i) {
dst[i] = plane_ptr;
// TODO: add VideoFormat::planeWidth/Height() ?
const int plane_w = pitch[i];//(i == 0 || pixfmt == VideoFormat::Format_NV12) ? d.surface_width : fmt.chromaWidth(d.surface_width);
const int plane_h = i == 0 ? d.surface_height : fmt.chromaHeight(d.surface_height);
plane_ptr += pitch[i] * plane_h;
d.gpu_mem.copyFrame(src[i], dst[i], plane_w, plane_h, pitch[i]);
}
frame = VideoFrame(buf, d.width, d.height, fmt);
frame.setBits(dst);
frame.setBytesPerLine(pitch);
} else {
frame = VideoFrame(d.width, d.height, fmt);
frame.setBits(src);
frame.setBytesPerLine(pitch);
// TODO: why clone is faster()?
frame = frame.clone();
}
if ((status = vaUnmapBuffer(d.display, d.image.buf)) != VA_STATUS_SUCCESS) {
qWarning("vaUnmapBuffer(VADisplay:%p, VABufferID:%#x) == %#x", d.display, d.image.buf, status);
return VideoFrame();
}
if (!d.disable_derive && d.supports_derive) {
vaDestroyImage(d.display, d.image.image_id);
d.image.image_id = VA_INVALID_ID;
}
return frame;
}
struct display_names_t {
VideoDecoderVAAPI::DisplayType display;
QString name;
};
static const display_names_t display_names[] = {
{ VideoDecoderVAAPI::GLX, "GLX" },
{ VideoDecoderVAAPI::X11, "X11" },
{ VideoDecoderVAAPI::DRM, "DRM" }
};
static VideoDecoderVAAPI::DisplayType displayFromName(QString name) {
for (unsigned int i = 0; i < sizeof(display_names)/sizeof(display_names[0]); ++i) {
if (name.toUpper().contains(display_names[i].name.toUpper())) {
return display_names[i].display;
}
}
return VideoDecoderVAAPI::X11;
}
static QString displayToName(VideoDecoderVAAPI::DisplayType t) {
for (unsigned int i = 0; i < sizeof(display_names)/sizeof(display_names[0]); ++i) {
if (t == display_names[i].display) {
return display_names[i].name;
}
}
return QString();
}
void VideoDecoderVAAPI::setDisplayPriority(const QStringList &priority)
{
DPTR_D(VideoDecoderVAAPI);
d.display_priority.clear();
foreach (QString disp, priority) {
d.display_priority.push_back(displayFromName(disp));
}