moonlight-stream/moonlight-qt
1
0
Fork 0
mirror of https://github.com/moonlight-stream/moonlight-qt.git synced 2025-07-01 23:35:55 +00:00
Code Issues Packages Projects Releases Wiki Activity
moonlight-qt/app/streaming/video/ffmpeg-renderers/vaapi.cpp
Cameron Gutman d3219ae24a Don't use VAAPI fallback driver names on libva 2.20+ 
libva 2.20 is good enough at detecting driver names (even under XWayland)
that we don't need to use the fallback name list anymore. This saves time
probing drivers, avoids excessive log output from failed probes, and avoids
tickling bugs in VAAPI drivers that are installed but unused.
2024-06-18 00:31:59 -05:00

1204 lines
47 KiB
C++
Raw Blame History

#include <QString>
// HACK: Include before vaapi.h to prevent conflicts with Xlib.h
#include <streaming/session.h>
#include "vaapi.h"
#include "utils.h"
#include <streaming/streamutils.h>
#ifdef HAVE_LIBVA_DRM
#include <xf86drm.h>
#endif
#include <SDL_syswm.h>
#include <unistd.h>
#include <fcntl.h>
VAAPIRenderer::VAAPIRenderer(int decoderSelectionPass)
: m_DecoderSelectionPass(decoderSelectionPass),
m_HwContext(nullptr),
m_BlacklistedForDirectRendering(false),
m_OverlayMutex(nullptr)
#ifdef HAVE_EGL
, m_EglExportType(EglExportType::Unknown),
m_EglImageFactory(this)
#endif
{
#ifdef HAVE_EGL
SDL_zero(m_PrimeDescriptor);
#endif
#ifdef HAVE_LIBVA_DRM
m_DrmFd = -1;
#endif
SDL_zero(m_OverlayImage);
SDL_zero(m_OverlaySubpicture);
SDL_zero(m_OverlayFormat);
}
VAAPIRenderer::~VAAPIRenderer()
{
if (m_HwContext != nullptr) {
AVHWDeviceContext* deviceContext = (AVHWDeviceContext*)m_HwContext->data;
AVVAAPIDeviceContext* vaDeviceContext = (AVVAAPIDeviceContext*)deviceContext->hwctx;
// Hold onto this VADisplay since we'll need it to uninitialize VAAPI
VADisplay display = vaDeviceContext->display;
for (int i = 0; i < Overlay::OverlayMax; i++) {
if (m_OverlaySubpicture[i] != 0) {
vaDestroySubpicture(display, m_OverlaySubpicture[i]);
}
if (m_OverlayImage[i].image_id != 0) {
vaDestroyImage(display, m_OverlayImage[i].image_id);
}
}
av_buffer_unref(&m_HwContext);
if (display) {
vaTerminate(display);
}
}
#ifdef HAVE_LIBVA_DRM
if (m_DrmFd >= 0) {
close(m_DrmFd);
}
#endif
if (m_OverlayMutex != nullptr) {
SDL_DestroyMutex(m_OverlayMutex);
}
}
VADisplay
VAAPIRenderer::openDisplay(SDL_Window* window)
{
SDL_SysWMinfo info;
VADisplay display;
SDL_VERSION(&info.version);
if (!SDL_GetWindowWMInfo(window, &info)) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"SDL_GetWindowWMInfo() failed: %s",
SDL_GetError());
return nullptr;
}
m_WindowSystem = info.subsystem;
if (info.subsystem == SDL_SYSWM_X11) {
#ifdef HAVE_LIBVA_X11
m_XWindow = info.info.x11.window;
display = vaGetDisplay(info.info.x11.display);
if (display == nullptr) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"Unable to open X11 display for VAAPI");
return nullptr;
}
#else
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"Moonlight not compiled with VAAPI X11 support!");
return nullptr;
#endif
}
else if (info.subsystem == SDL_SYSWM_WAYLAND) {
#ifdef HAVE_LIBVA_WAYLAND
display = vaGetDisplayWl(info.info.wl.display);
if (display == nullptr) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"Unable to open Wayland display for VAAPI");
return nullptr;
}
#else
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"Moonlight not compiled with VAAPI Wayland support!");
return nullptr;
#endif
}
#if defined(SDL_VIDEO_DRIVER_KMSDRM) && defined(HAVE_LIBVA_DRM) && SDL_VERSION_ATLEAST(2, 0, 15)
else if (info.subsystem == SDL_SYSWM_KMSDRM) {
SDL_assert(info.info.kmsdrm.drm_fd >= 0);
// It's possible to enter this function several times as we're probing VA drivers.
// Make sure to only duplicate the DRM FD the first time through.
if (m_DrmFd < 0) {
// If the KMSDRM FD is not a render node FD, open the render node for libva to use.
// Since libva 2.20, using a primary node will fail in vaGetDriverNames().
if (drmGetNodeTypeFromFd(info.info.kmsdrm.drm_fd) != DRM_NODE_RENDER) {
char* renderNodePath = drmGetRenderDeviceNameFromFd(info.info.kmsdrm.drm_fd);
if (renderNodePath) {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Opening render node for VAAPI: %s",
renderNodePath);
m_DrmFd = open(renderNodePath, O_RDWR | O_CLOEXEC);
free(renderNodePath);
if (m_DrmFd < 0) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"Failed to open render node: %d",
errno);
return nullptr;
}
}
else {
SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
"Failed to get render node path. Using the SDL FD directly.");
m_DrmFd = dup(info.info.kmsdrm.drm_fd);
}
}
else {
SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
"KMSDRM FD is already a render node. Using the SDL FD directly.");
m_DrmFd = dup(info.info.kmsdrm.drm_fd);
}
}
display = vaGetDisplayDRM(m_DrmFd);
if (display == nullptr) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"Unable to open DRM display for VAAPI");
return nullptr;
}
}
#endif
else {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"Unsupported VAAPI rendering subsystem: %d",
info.subsystem);
return nullptr;
}
return display;
}
VAStatus
VAAPIRenderer::tryVaInitialize(AVVAAPIDeviceContext* vaDeviceContext, PDECODER_PARAMETERS params, int* major, int* minor)
{
VAStatus status;
SDL_assert(vaDeviceContext->display == nullptr);
vaDeviceContext->display = openDisplay(params->window);
if (vaDeviceContext->display == nullptr) {
// openDisplay() logs the error
return VA_STATUS_ERROR_INVALID_DISPLAY;
}
status = vaInitialize(vaDeviceContext->display, major, minor);
if (status != VA_STATUS_SUCCESS) {
// vaInitialize() stores state into the VADisplay even on failure, so we must still
// call vaTerminate() even if vaInitialize() failed. Similarly, calling vaInitialize()
// more than once on the same VADisplay can cause resource leaks, even if it failed
// in the prior call. https://github.com/intel/libva/issues/741
vaTerminate(vaDeviceContext->display);
vaDeviceContext->display = nullptr;
}
return status;
}
bool
VAAPIRenderer::initialize(PDECODER_PARAMETERS params)
{
int err;
m_Window = params->window;
m_VideoFormat = params->videoFormat;
m_HwContext = av_hwdevice_ctx_alloc(AV_HWDEVICE_TYPE_VAAPI);
if (!m_HwContext) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"Failed to allocate VAAPI context");
return false;
}
AVHWDeviceContext* deviceContext = (AVHWDeviceContext*)m_HwContext->data;
AVVAAPIDeviceContext* vaDeviceContext = (AVVAAPIDeviceContext*)deviceContext->hwctx;
int major, minor;
VAStatus status;
bool setPathVar = false;
for (;;) {
// vaInitialize() will return the libva library version even if the function
// fails. This has been the case since libva v2.6 from 5 years ago. This
// doesn't seem to be documented anywhere, so we will be conservative to
// protect against changes in libva behavior by reinitializing major/minor
// each time and clamping it to the valid range of versions based upon
// the version of libva that we compiled with.
major = minor = 0;
status = tryVaInitialize(vaDeviceContext, params, &major, &minor);
if (status != VA_STATUS_SUCCESS) {
major = std::max(major, VA_MAJOR_VERSION);
minor = std::max(minor, VA_MINOR_VERSION);
// If LIBVA_DRIVER_NAME has not been set manually and we're running a
// version of libva less than 2.20, we'll try our own fallback names.
// Beginning in libva 2.20, the driver name detection code is much
// more robust than earlier versions and it includes DRI3 support for
// driver name detection under Xwayland.
if (!qEnvironmentVariableIsEmpty("LIBVA_DRIVER_NAME")) {
SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
"Skipping VAAPI fallback driver names due to LIBVA_DRIVER_NAME");
}
else if (major > 1 || (major == 1 && minor >= 20)) {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Skipping VAAPI fallback driver names on libva 2.20+");
}
else {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Trying fallback VAAPI driver names");
// It would be nice to use vaSetDriverName() here, but there's no way to unset
// it and get back to the default driver selection logic once we've overridden
// the driver name using that API. As a result, we must use LIBVA_DRIVER_NAME.
if (status != VA_STATUS_SUCCESS) {
// The iHD driver supports newer hardware like Ice Lake and Comet Lake.
// It should be picked by default on those platforms, but that doesn't
// always seem to be the case for some reason.
qputenv("LIBVA_DRIVER_NAME", "iHD");
status = tryVaInitialize(vaDeviceContext, params, &major, &minor);
}
if (status != VA_STATUS_SUCCESS) {
// The Iris driver in Mesa 20.0 returns a bogus VA driver (iris_drv_video.so)
// even though the correct driver is still i965. If we hit this path, we'll
// explicitly try i965 to handle this case.
qputenv("LIBVA_DRIVER_NAME", "i965");
status = tryVaInitialize(vaDeviceContext, params, &major, &minor);
}
if (status != VA_STATUS_SUCCESS) {
// The RadeonSI driver is compatible with XWayland but can't be detected by libva
// so try it too if all else fails.
qputenv("LIBVA_DRIVER_NAME", "radeonsi");
status = tryVaInitialize(vaDeviceContext, params, &major, &minor);
}
if (status != VA_STATUS_SUCCESS && (m_WindowSystem != SDL_SYSWM_X11 || m_DecoderSelectionPass > 0)) {
// The unofficial nvidia VAAPI driver over NVDEC/CUDA works well on Wayland,
// but we'd rather use CUDA for XWayland and VDPAU for regular X11.
// NB: Remember to update the VA-API NVDEC condition below when modifying this!
qputenv("LIBVA_DRIVER_NAME", "nvidia");
status = tryVaInitialize(vaDeviceContext, params, &major, &minor);
}
if (status != VA_STATUS_SUCCESS) {
// Unset LIBVA_DRIVER_NAME if none of the drivers we tried worked. This ensures
// we will get a fresh start using the default driver selection behavior after
// setting LIBVA_DRIVERS_PATH in the code below.
qunsetenv("LIBVA_DRIVER_NAME");
}
}
}
if (status == VA_STATUS_SUCCESS) {
// Success!
break;
}
#if defined(APP_IMAGE) || defined(USE_FALLBACK_DRIVER_PATHS)
// AppImages will be running with our libva.so which means they don't know about
// distro-specific driver paths. To avoid failing in this scenario, we'll hardcode
// some such paths here for common distros. Non-AppImage packaging mechanisms won't
// need this fallback because either:
// a) They are using both distro libva.so and distro libva drivers (native packages)
// b) They are using both runtime libva.so and runtime libva drivers (Flatpak/Snap)
if (qEnvironmentVariableIsEmpty("LIBVA_DRIVERS_PATH")) {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Trying fallback VAAPI driver paths");
qputenv("LIBVA_DRIVERS_PATH",
#if Q_PROCESSOR_WORDSIZE == 8
"/usr/lib64/dri-nonfree:" // Fedora x86_64
"/usr/lib64/dri-freeworld:" // Fedora x86_64
"/usr/lib64/dri:" // Fedora x86_64
"/usr/lib64/va/drivers:" // Gentoo x86_64
#endif
"/usr/lib/dri:" // Arch i386 & x86_64, Fedora i386
"/usr/lib/va/drivers:" // Gentoo i386
#if defined(Q_PROCESSOR_X86_64)
"/usr/lib/x86_64-linux-gnu/dri:" // Ubuntu/Debian x86_64
#elif defined(Q_PROCESSOR_X86_32)
"/usr/lib/i386-linux-gnu/dri:" // Ubuntu/Debian i386
#endif
);
setPathVar = true;
}
else
#endif
{
if (setPathVar) {
// Unset LIBVA_DRIVERS_PATH if we set it ourselves
// and we didn't find any working VAAPI drivers.
qunsetenv("LIBVA_DRIVERS_PATH");
}
// Give up
break;
}
}
if (status != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"Failed to initialize VAAPI: %d",
status);
return false;
}
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Initialized VAAPI %d.%d",
major, minor);
const char* vendorString = vaQueryVendorString(vaDeviceContext->display);
QString vendorStr(vendorString);
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Driver: %s",
vendorString ? vendorString : "<unknown>");
// This is the libva-vdpau-driver which is not supported by our VAAPI renderer.
if (vendorStr.contains("Splitted-Desktop Systems VDPAU backend for VA-API")) {
// Fail and let our VDPAU renderer pick this up
SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
"Avoiding VDPAU wrapper for VAAPI decoding");
return false;
}
// The Snap (core22) and Focal/Jammy Mesa drivers have a bug that causes
// a large amount of video latency when using more than one reference frame
// and severe rendering glitches on my Ryzen 3300U system.
m_HasRfiLatencyBug = vendorStr.contains("Gallium", Qt::CaseInsensitive) && qgetenv("IGNORE_RFI_LATENCY_BUG") != "1";
if (m_HasRfiLatencyBug) {
SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
"VAAPI driver is affected by RFI latency bug");
}
if (m_DecoderSelectionPass == 0 && qgetenv("FORCE_VAAPI") != "1") {
// Older versions of the Gallium VAAPI driver have a nasty memory leak that
// causes memory to be leaked for each submitted frame. I believe this is
// resolved in the libva2 drivers (VAAPI 1.x). We will try to use VDPAU
// instead for old VAAPI versions or drivers affected by the RFI latency bug
// as long as we're not streaming HDR (which is unsupported by VDPAU).
if ((major == 0 || (m_HasRfiLatencyBug && !(m_VideoFormat & VIDEO_FORMAT_MASK_10BIT))) &&
vendorStr.contains("Gallium", Qt::CaseInsensitive)) {
// Fail and let VDPAU pick this up
SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
"Deprioritizing VAAPI on Gallium driver. Set FORCE_VAAPI=1 to override.");
return false;
}
// Prefer CUDA for XWayland and VDPAU for regular X11.
if (m_WindowSystem == SDL_SYSWM_X11 && vendorStr.contains("VA-API NVDEC", Qt::CaseInsensitive)) {
SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
"Deprioritizing VAAPI for NVIDIA driver on X11/XWayland. Set FORCE_VAAPI=1 to override.");
return false;
}
}
if (WMUtils::isRunningWayland()) {
// The iHD VAAPI driver can initialize on XWayland but it crashes in
// vaPutSurface() so we must also not directly render on XWayland.
m_BlacklistedForDirectRendering = vendorStr.contains("iHD");
}
// This will populate the driver_quirks
err = av_hwdevice_ctx_init(m_HwContext);
if (err < 0) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"Failed to initialize VAAPI context: %d",
err);
return false;
}
// Allocate mutex to synchronize overlay updates and rendering
m_OverlayMutex = SDL_CreateMutex();
if (m_OverlayMutex == nullptr) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"Failed to create overlay mutex");
return false;
}
unsigned int formatCount = vaMaxNumSubpictureFormats(vaDeviceContext->display);
if (formatCount != 0) {
auto formats = new VAImageFormat[formatCount];
auto flags = new unsigned int[formatCount];
status = vaQuerySubpictureFormats(vaDeviceContext->display, formats, flags, &formatCount);
if (status == VA_STATUS_SUCCESS) {
for (unsigned int i = 0; i < formatCount; i++) {
// Format must have 32-bit color depth
if (formats[i].depth != 32) {
continue;
}
// Select an RGB format with alpha
if (formats[i].byte_order == VA_MSB_FIRST) {
switch (formats[i].fourcc) {
case VA_FOURCC_RGBA:
m_OverlaySdlPixelFormat = SDL_PIXELFORMAT_RGBA8888;
break;
case VA_FOURCC_ARGB:
m_OverlaySdlPixelFormat = SDL_PIXELFORMAT_ARGB8888;
break;
case VA_FOURCC_BGRA:
m_OverlaySdlPixelFormat = SDL_PIXELFORMAT_BGRA8888;
break;
case VA_FOURCC_ABGR:
m_OverlaySdlPixelFormat = SDL_PIXELFORMAT_ABGR8888;
break;
default:
continue;
}
}
else {
SDL_assert(formats[i].byte_order == VA_LSB_FIRST);
switch (formats[i].fourcc) {
case VA_FOURCC_RGBA:
m_OverlaySdlPixelFormat = SDL_PIXELFORMAT_ABGR8888;
break;
case VA_FOURCC_ARGB:
m_OverlaySdlPixelFormat = SDL_PIXELFORMAT_BGRA8888;
break;
case VA_FOURCC_BGRA:
m_OverlaySdlPixelFormat = SDL_PIXELFORMAT_ARGB8888;
break;
case VA_FOURCC_ABGR:
m_OverlaySdlPixelFormat = SDL_PIXELFORMAT_RGBA8888;
break;
default:
continue;
}
}
// If we made it here, we found a format that works for us
m_OverlayFormat = formats[i];
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Selected overlay subpicture format: %c%c%c%c8888",
(m_OverlayFormat.fourcc >> 0) & 0xff,
(m_OverlayFormat.fourcc >> 8) & 0xff,
(m_OverlayFormat.fourcc >> 16) & 0xff,
(m_OverlayFormat.fourcc >> 24) & 0xff);
break;
}
}
else {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaQuerySubpictureFormats() failed: %d",
status);
}
delete[] formats;
delete[] flags;
}
return true;
}
bool
VAAPIRenderer::prepareDecoderContext(AVCodecContext* context, AVDictionary**)
{
context->hw_device_ctx = av_buffer_ref(m_HwContext);
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Using VAAPI accelerated renderer on %s",
SDL_GetCurrentVideoDriver());
return true;
}
bool
VAAPIRenderer::needsTestFrame()
{
// We need a test frame to see if this VAAPI driver
// supports the profile used for streaming
return true;
}
bool
VAAPIRenderer::isDirectRenderingSupported()
{
if (qgetenv("VAAPI_FORCE_DIRECT") == "1") {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Using direct rendering due to environment variable");
return true;
}
else if (qgetenv("VAAPI_FORCE_INDIRECT") == "1") {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Using indirect rendering due to environment variable");
return false;
}
// We only support direct rendering on X11 with VAEntrypointVideoProc support
if (m_WindowSystem != SDL_SYSWM_X11 || m_BlacklistedForDirectRendering) {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Using indirect rendering due to WM or blacklist");
return false;
}
else if (m_VideoFormat & VIDEO_FORMAT_MASK_10BIT) {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Using indirect rendering for 10-bit video");
return false;
}
AVHWDeviceContext* deviceContext = (AVHWDeviceContext*)m_HwContext->data;
AVVAAPIDeviceContext* vaDeviceContext = (AVVAAPIDeviceContext*)deviceContext->hwctx;
VAEntrypoint entrypoints[vaMaxNumEntrypoints(vaDeviceContext->display)];
int entrypointCount;
VAStatus status = vaQueryConfigEntrypoints(vaDeviceContext->display, VAProfileNone, entrypoints, &entrypointCount);
if (status == VA_STATUS_SUCCESS) {
for (int i = 0; i < entrypointCount; i++) {
// Without VAEntrypointVideoProc support, the driver will crash inside vaPutSurface()
if (entrypoints[i] == VAEntrypointVideoProc) {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Using direct rendering with VAEntrypointVideoProc");
if (m_OverlayFormat.fourcc == 0) {
SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
"Unable to find supported subpicture format. Overlays will be unavailable!");
}
return true;
}
}
}
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Using indirect rendering due to lack of VAEntrypointVideoProc");
return false;
}
int VAAPIRenderer::getDecoderColorspace()
{
// Gallium drivers don't support Rec 709 yet - https://gitlab.freedesktop.org/mesa/mesa/issues/1915
// Intel-vaapi-driver defaults to Rec 601 - https://github.com/intel/intel-vaapi-driver/blob/021bcb79d1bd873bbd9fbca55f40320344bab866/src/i965_output_dri.c#L186
return COLORSPACE_REC_601;
}
int VAAPIRenderer::getDecoderCapabilities()
{
int caps = 0;
if (!m_HasRfiLatencyBug) {
caps |= CAPABILITY_REFERENCE_FRAME_INVALIDATION_HEVC |
CAPABILITY_REFERENCE_FRAME_INVALIDATION_AV1;
}
return caps;
}
void VAAPIRenderer::notifyOverlayUpdated(Overlay::OverlayType type)
{
AVHWDeviceContext* deviceContext = (AVHWDeviceContext*)m_HwContext->data;
AVVAAPIDeviceContext* vaDeviceContext = (AVVAAPIDeviceContext*)deviceContext->hwctx;
VAStatus status;
if (m_OverlayFormat.fourcc == 0) {
// We already logged for this in isDirectRenderingSupported()
return;
}
SDL_Surface* newSurface = Session::get()->getOverlayManager().getUpdatedOverlaySurface(type);
bool overlayEnabled = Session::get()->getOverlayManager().isOverlayEnabled(type);
if (newSurface == nullptr && overlayEnabled) {
// There's no updated surface and the overlay is enabled, so just leave the old surface alone.
return;
}
// Destroy the old image and subpicture
// NB: The mutex ensures the overlay is not currently being read for rendering.
// NB 2: It is safe to unlock here because this thread is the only surface producer.
SDL_LockMutex(m_OverlayMutex);
VAImageID oldImageId = m_OverlayImage[type].image_id;
SDL_zero(m_OverlayImage[type]);
VASubpictureID oldSubpictureId = m_OverlaySubpicture[type];
m_OverlaySubpicture[type] = 0;
SDL_UnlockMutex(m_OverlayMutex);
if (oldSubpictureId != 0) {
status = vaDestroySubpicture(vaDeviceContext->display, oldSubpictureId);
if (status != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaDestroySubpicture() failed: %d",
status);
}
}
if (oldImageId != 0) {
status = vaDestroyImage(vaDeviceContext->display, oldImageId);
if (status != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaDestroyImage() failed: %d",
status);
}
}
if (!overlayEnabled) {
SDL_FreeSurface(newSurface);
return;
}
if (newSurface != nullptr) {
VAImage newImage;
SDL_assert(!SDL_MUSTLOCK(newSurface));
status = vaCreateImage(vaDeviceContext->display, &m_OverlayFormat, newSurface->w, newSurface->h, &newImage);
if (status != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaCreateImage() failed: %d",
status);
SDL_FreeSurface(newSurface);
return;
}
void* imagePixels;
status = vaMapBuffer(vaDeviceContext->display, newImage.buf, &imagePixels);
if (status != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaMapBuffer() failed: %d",
status);
SDL_FreeSurface(newSurface);
vaDestroyImage(vaDeviceContext->display, newImage.image_id);
return;
}
// Convert the surface to the proper format for the VAImage
SDL_ConvertPixels(newSurface->w, newSurface->h, newSurface->format->format,
newSurface->pixels, newSurface->pitch, m_OverlaySdlPixelFormat,
imagePixels, (int)newImage.pitches[0]);
status = vaUnmapBuffer(vaDeviceContext->display, newImage.buf);
if (status != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaUnmapBuffer() failed: %d",
status);
SDL_FreeSurface(newSurface);
vaDestroyImage(vaDeviceContext->display, newImage.image_id);
return;
}
SDL_Rect overlayRect;
if (type == Overlay::OverlayStatusUpdate) {
// Bottom Left
overlayRect.x = 0;
overlayRect.y = -newSurface->h;
}
else if (type == Overlay::OverlayDebug) {
// Top left
overlayRect.x = 0;
overlayRect.y = 0;
}
overlayRect.w = newSurface->w;
overlayRect.h = newSurface->h;
// Surface data is no longer needed
SDL_FreeSurface(newSurface);
VASubpictureID newSubpicture;
status = vaCreateSubpicture(vaDeviceContext->display, newImage.image_id, &newSubpicture);
if (status != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaCreateSubpicture() failed: %d",
status);
vaDestroyImage(vaDeviceContext->display, newImage.image_id);
return;
}
SDL_LockMutex(m_OverlayMutex);
m_OverlayImage[type] = newImage;
m_OverlaySubpicture[type] = newSubpicture;
m_OverlayRect[type] = overlayRect;
SDL_UnlockMutex(m_OverlayMutex);
}
}
bool VAAPIRenderer::notifyWindowChanged(PWINDOW_STATE_CHANGE_INFO info)
{
// We can transparently handle size and display changes
return !(info->stateChangeFlags & ~(WINDOW_STATE_CHANGE_SIZE | WINDOW_STATE_CHANGE_DISPLAY));
}
void
VAAPIRenderer::renderFrame(AVFrame* frame)
{
VASurfaceID surface = (VASurfaceID)(uintptr_t)frame->data[3];
AVHWDeviceContext* deviceContext = (AVHWDeviceContext*)m_HwContext->data;
AVVAAPIDeviceContext* vaDeviceContext = (AVVAAPIDeviceContext*)deviceContext->hwctx;
int windowWidth, windowHeight;
SDL_GetWindowSize(m_Window, &windowWidth, &windowHeight);
SDL_Rect src, dst;
src.x = src.y = 0;
src.w = frame->width;
src.h = frame->height;
dst.x = dst.y = 0;
dst.w = windowWidth;
dst.h = windowHeight;
StreamUtils::scaleSourceToDestinationSurface(&src, &dst);
if (m_WindowSystem == SDL_SYSWM_X11) {
#ifdef HAVE_LIBVA_X11
unsigned int flags = 0;
// NB: Not all VAAPI drivers respect these flags. Many drivers
// just ignore them and do the color conversion as Rec 601.
switch (getFrameColorspace(frame)) {
case COLORSPACE_REC_709:
flags |= VA_SRC_BT709;
break;
case COLORSPACE_REC_601:
flags |= VA_SRC_BT601;
break;
default:
// Unsupported colorspace
SDL_assert(false);
break;
}
SDL_LockMutex(m_OverlayMutex);
VAImage associatedOverlayImages[Overlay::OverlayMax] = {};
VASubpictureID associatedOverlaySubpictures[Overlay::OverlayMax] = {};
// Associate our overlay subpictures to the current surface
for (int type = 0; type < Overlay::OverlayMax; type++) {
VAStatus status;
if (m_OverlaySubpicture[type] == 0) {
continue;
}
SDL_Rect overlayRect = m_OverlayRect[type];
// Negative values are relative to the other side of the window
if (overlayRect.x < 0) {
overlayRect.x += windowWidth;
}
if (overlayRect.y < 0) {
overlayRect.y += windowHeight;
}
status = vaAssociateSubpicture(vaDeviceContext->display,
m_OverlaySubpicture[type],
&surface,
1,
0,
0,
m_OverlayImage[type].width,
m_OverlayImage[type].height,
overlayRect.x,
overlayRect.y,
overlayRect.w,
overlayRect.h,
0);
if (status == VA_STATUS_SUCCESS) {
// Take temporary ownership of the overlay to prevent notifyOverlayUpdated()
// from freeing them frum underneath us. We need to release the lock while
// we render for performance reasons.
associatedOverlayImages[type] = m_OverlayImage[type];
associatedOverlaySubpictures[type] = m_OverlaySubpicture[type];
SDL_zero(m_OverlayImage[type]);
m_OverlaySubpicture[type] = 0;
}
else {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaAssociateSubpicture() failed: %d",
status);
}
}
SDL_UnlockMutex(m_OverlayMutex);
// This will draw the surface and any associated subpictures
// NB: This can take a full VBlank period to complete!
vaPutSurface(vaDeviceContext->display,
surface,
m_XWindow,
0, 0,
frame->width, frame->height,
dst.x, dst.y,
dst.w, dst.h,
NULL, 0, flags);
SDL_LockMutex(m_OverlayMutex);
// Now that we've reacquired the lock, we will need to reconcile the current
// state of the overlay with our saved state from before we unlocked.
for (int type = 0; type < Overlay::OverlayMax; type++) {
VAStatus status;
if (associatedOverlaySubpictures[type] == 0) {
continue;
}
// Deassociate the subpicture so it can be safely destroyed/replaced
status = vaDeassociateSubpicture(vaDeviceContext->display, associatedOverlaySubpictures[type], &surface, 1);
if (status != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaDeassociateSubpicture() failed: %d",
status);
}
// If a new subpicture was populated while we were unlocked, free the old one we took ownership of
if (m_OverlaySubpicture[type] != 0) {
status = vaDestroySubpicture(vaDeviceContext->display, associatedOverlaySubpictures[type]);
if (status != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaDestroySubpicture() failed: %d",
status);
}
}
else {
// If no new subpicture was populated, return ownership of this one
m_OverlaySubpicture[type] = associatedOverlaySubpictures[type];
}
// If a new image was populated while we were unlocked, free the one old we took ownership of
if (m_OverlayImage[type].image_id != 0) {
status = vaDestroyImage(vaDeviceContext->display, associatedOverlayImages[type].image_id);
if (status != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaDestroyImage() failed: %d",
status);
}
}
else {
// If no new image was populated, return ownership of this one
m_OverlayImage[type] = associatedOverlayImages[type];
}
}
SDL_UnlockMutex(m_OverlayMutex);
#endif
}
else if (m_WindowSystem == SDL_SYSWM_WAYLAND) {
// We don't support direct rendering on Wayland, so we should
// never get called there. Many common Wayland compositors don't
// support YUV surfaces, so direct rendering would fail.
SDL_assert(false);
}
else {
// We don't accept anything else in initialize().
SDL_assert(false);
}
}
#if defined(HAVE_EGL) || defined(HAVE_DRM)
// Ensure that vaExportSurfaceHandle() is supported by the VA-API driver
bool
VAAPIRenderer::canExportSurfaceHandle(int layerTypeFlag, VADRMPRIMESurfaceDescriptor* descriptor) {
AVHWDeviceContext* deviceContext = (AVHWDeviceContext*)m_HwContext->data;
AVVAAPIDeviceContext* vaDeviceContext = (AVVAAPIDeviceContext*)deviceContext->hwctx;
VASurfaceID surfaceId;
VAStatus st;
VASurfaceAttrib attrs[2];
int attributeCount = 0;
if (qgetenv("VAAPI_FORCE_DIRECT") == "1") {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Using direct rendering due to environment variable");
return false;
}
// FFmpeg handles setting these quirk flags for us
if (!(vaDeviceContext->driver_quirks & AV_VAAPI_DRIVER_QUIRK_ATTRIB_MEMTYPE)) {
attrs[attributeCount].type = VASurfaceAttribMemoryType;
attrs[attributeCount].flags = VA_SURFACE_ATTRIB_SETTABLE;
attrs[attributeCount].value.type = VAGenericValueTypeInteger;
attrs[attributeCount].value.value.i = VA_SURFACE_ATTRIB_MEM_TYPE_VA;
attributeCount++;
}
// These attributes are required for i965 to create a surface that can
// be successfully exported via vaExportSurfaceHandle(). iHD doesn't
// need these, but it doesn't seem to hurt either.
attrs[attributeCount].type = VASurfaceAttribPixelFormat;
attrs[attributeCount].flags = VA_SURFACE_ATTRIB_SETTABLE;
attrs[attributeCount].value.type = VAGenericValueTypeInteger;
attrs[attributeCount].value.value.i = (m_VideoFormat & VIDEO_FORMAT_MASK_10BIT) ?
VA_FOURCC_P010 : VA_FOURCC_NV12;
attributeCount++;
st = vaCreateSurfaces(vaDeviceContext->display,
(m_VideoFormat & VIDEO_FORMAT_MASK_10BIT) ?
VA_RT_FORMAT_YUV420_10 : VA_RT_FORMAT_YUV420,
1280,
720,
&surfaceId,
1,
attrs,
attributeCount);
if (st != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaCreateSurfaces() failed: %d", st);
return false;
}
st = vaExportSurfaceHandle(vaDeviceContext->display,
surfaceId,
VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME_2,
VA_EXPORT_SURFACE_READ_ONLY | layerTypeFlag,
descriptor);
vaDestroySurfaces(vaDeviceContext->display, &surfaceId, 1);
if (st != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaExportSurfaceHandle() failed: %d", st);
return false;
}
for (size_t i = 0; i < descriptor->num_objects; ++i) {
close(descriptor->objects[i].fd);
descriptor->objects[i].fd = -1;
}
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"VAAPI driver supports exporting DRM PRIME surface handles with %s layers",
layerTypeFlag == VA_EXPORT_SURFACE_COMPOSED_LAYERS ? "composed" : "separate");
return true;
}
#endif
#ifdef HAVE_EGL
bool
VAAPIRenderer::canExportEGL() {
VADRMPRIMESurfaceDescriptor descriptor;
return (qgetenv("VAAPI_EGL_SEPARATE_LAYERS") != "1" && canExportSurfaceHandle(VA_EXPORT_SURFACE_COMPOSED_LAYERS, &descriptor)) ||
canExportSurfaceHandle(VA_EXPORT_SURFACE_SEPARATE_LAYERS, &descriptor);
}
AVPixelFormat VAAPIRenderer::getEGLImagePixelFormat() {
switch (m_EglExportType) {
case EglExportType::Separate:
return (m_VideoFormat & VIDEO_FORMAT_MASK_10BIT) ?
AV_PIX_FMT_P010 : AV_PIX_FMT_NV12;
case EglExportType::Composed:
// This tells EGLRenderer to treat the EGLImage as a single opaque texture
return AV_PIX_FMT_DRM_PRIME;
case EglExportType::Unknown:
SDL_assert(m_EglExportType != EglExportType::Unknown);
break;
}
return AV_PIX_FMT_NONE;
}
bool
VAAPIRenderer::initializeEGL(EGLDisplay dpy,
const EGLExtensions &ext) {
VADRMPRIMESurfaceDescriptor descriptor;
if (!m_EglImageFactory.initializeEGL(dpy, ext)) {
return false;
}
// Prefer exporting composed images absent a user override or lack of support for exporting or importing
if (qgetenv("VAAPI_EGL_SEPARATE_LAYERS") == "1") {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Exporting separate layers due to environment variable override");
m_EglExportType = EglExportType::Separate;
}
else if (!canExportSurfaceHandle(VA_EXPORT_SURFACE_COMPOSED_LAYERS, &descriptor)) {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Exporting separate layers due to lack of support for VA_EXPORT_SURFACE_COMPOSED_LAYERS");
m_EglExportType = EglExportType::Separate;
}
else if (!m_EglImageFactory.supportsImportingFormat(dpy, descriptor.layers[0].drm_format)) {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Exporting separate layers due to lack of support for importing format: %08x", descriptor.layers[0].drm_format);
m_EglExportType = EglExportType::Separate;
}
else if (!m_EglImageFactory.supportsImportingModifier(dpy, descriptor.layers[0].drm_format, descriptor.objects[0].drm_format_modifier)) {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Exporting separate layers due to lack of support for importing format and modifier: %08x %016" PRIx64,
descriptor.layers[0].drm_format,
descriptor.objects[0].drm_format_modifier);
m_EglExportType = EglExportType::Separate;
}
else {
SDL_LogInfo(SDL_LOG_CATEGORY_APPLICATION,
"Exporting composed layers with format and modifier: %08x %016" PRIx64,
descriptor.layers[0].drm_format,
descriptor.objects[0].drm_format_modifier);
m_EglExportType = EglExportType::Composed;
}
// Let's probe for EGL import support on separate layers too, but only warn if it's not supported
if (m_EglExportType == EglExportType::Separate) {
if (!canExportSurfaceHandle(VA_EXPORT_SURFACE_SEPARATE_LAYERS, &descriptor)) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"Exporting separate layers is not supported by the VAAPI driver");
return false;
}
for (uint32_t i = 0; i < descriptor.num_layers; i++) {
if (!m_EglImageFactory.supportsImportingFormat(dpy, descriptor.layers[i].drm_format)) {
SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
"EGL implementation lacks support for importing format: %08x", descriptor.layers[0].drm_format);
}
else if (!m_EglImageFactory.supportsImportingModifier(dpy, descriptor.layers[i].drm_format,
descriptor.objects[descriptor.layers[i].object_index[0]].drm_format_modifier)) {
SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
"EGL implementation lacks support for importing format and modifier: %08x %016" PRIx64,
descriptor.layers[i].drm_format,
descriptor.objects[descriptor.layers[i].object_index[0]].drm_format_modifier);
}
}
}
return true;
}
ssize_t
VAAPIRenderer::exportEGLImages(AVFrame *frame, EGLDisplay dpy,
EGLImage images[EGL_MAX_PLANES]) {
ssize_t count;
uint32_t exportFlags = VA_EXPORT_SURFACE_READ_ONLY;
switch (m_EglExportType) {
case EglExportType::Separate:
exportFlags |= VA_EXPORT_SURFACE_SEPARATE_LAYERS;
break;
case EglExportType::Composed:
exportFlags |= VA_EXPORT_SURFACE_COMPOSED_LAYERS;
break;
case EglExportType::Unknown:
SDL_assert(m_EglExportType != EglExportType::Unknown);
return -1;
}
auto hwFrameCtx = (AVHWFramesContext*)frame->hw_frames_ctx->data;
AVVAAPIDeviceContext* vaDeviceContext = (AVVAAPIDeviceContext*)hwFrameCtx->device_ctx->hwctx;
VASurfaceID surface_id = (VASurfaceID)(uintptr_t)frame->data[3];
VAStatus st = vaExportSurfaceHandle(vaDeviceContext->display,
surface_id,
VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME_2,
exportFlags,
&m_PrimeDescriptor);
if (st != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaExportSurfaceHandle failed: %d", st);
return -1;
}
st = vaSyncSurface(vaDeviceContext->display, surface_id);
if (st != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaSyncSurface() failed: %d", st);
goto fail;
}
count = m_EglImageFactory.exportVAImages(frame, &m_PrimeDescriptor, dpy, images);
if (count < 0) {
goto fail;
}
return count;
fail:
for (size_t i = 0; i < m_PrimeDescriptor.num_objects; ++i) {
close(m_PrimeDescriptor.objects[i].fd);
}
m_PrimeDescriptor.num_layers = 0;
m_PrimeDescriptor.num_objects = 0;
return -1;
}
void
VAAPIRenderer::freeEGLImages(EGLDisplay dpy, EGLImage images[EGL_MAX_PLANES]) {
m_EglImageFactory.freeEGLImages(dpy, images);
for (size_t i = 0; i < m_PrimeDescriptor.num_objects; ++i) {
close(m_PrimeDescriptor.objects[i].fd);
}
m_PrimeDescriptor.num_layers = 0;
m_PrimeDescriptor.num_objects = 0;
}
#endif
#ifdef HAVE_DRM
bool VAAPIRenderer::canExportDrmPrime()
{
// Our DRM renderer requires composed layers
VADRMPRIMESurfaceDescriptor descriptor;
return canExportSurfaceHandle(VA_EXPORT_SURFACE_COMPOSED_LAYERS, &descriptor);
}
bool VAAPIRenderer::mapDrmPrimeFrame(AVFrame* frame, AVDRMFrameDescriptor* drmDescriptor)
{
auto hwFrameCtx = (AVHWFramesContext*)frame->hw_frames_ctx->data;
AVVAAPIDeviceContext* vaDeviceContext = (AVVAAPIDeviceContext*)hwFrameCtx->device_ctx->hwctx;
VASurfaceID vaSurfaceId = (VASurfaceID)(uintptr_t)frame->data[3];
VADRMPRIMESurfaceDescriptor vaDrmPrimeDescriptor;
VAStatus st = vaExportSurfaceHandle(vaDeviceContext->display,
vaSurfaceId,
VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME_2,
VA_EXPORT_SURFACE_READ_ONLY | VA_EXPORT_SURFACE_COMPOSED_LAYERS,
&vaDrmPrimeDescriptor);
if (st != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaExportSurfaceHandle() failed: %d", st);
return false;
}
st = vaSyncSurface(vaDeviceContext->display, vaSurfaceId);
if (st != VA_STATUS_SUCCESS) {
SDL_LogError(SDL_LOG_CATEGORY_APPLICATION,
"vaSyncSurface() failed: %d", st);
for (uint32_t i = 0; i < vaDrmPrimeDescriptor.num_objects; i++) {
close(vaDrmPrimeDescriptor.objects[i].fd);
}
return false;
}
// Map our VADRMPRIMESurfaceDescriptor to the AVDRMFrameDescriptor our caller wants
drmDescriptor->nb_objects = vaDrmPrimeDescriptor.num_objects;
for (uint32_t i = 0; i < vaDrmPrimeDescriptor.num_objects; i++) {
drmDescriptor->objects[i].fd = vaDrmPrimeDescriptor.objects[i].fd;
drmDescriptor->objects[i].size = vaDrmPrimeDescriptor.objects[i].size;
drmDescriptor->objects[i].format_modifier = vaDrmPrimeDescriptor.objects[i].drm_format_modifier;
}
drmDescriptor->nb_layers = vaDrmPrimeDescriptor.num_layers;
for (uint32_t i = 0; i < vaDrmPrimeDescriptor.num_layers; i++) {
drmDescriptor->layers[i].format = vaDrmPrimeDescriptor.layers[i].drm_format;
drmDescriptor->layers[i].nb_planes = vaDrmPrimeDescriptor.layers[i].num_planes;
for (uint32_t j = 0; j < vaDrmPrimeDescriptor.layers[i].num_planes; j++) {
drmDescriptor->layers[i].planes[j].object_index = vaDrmPrimeDescriptor.layers[i].object_index[j];
drmDescriptor->layers[i].planes[j].offset = vaDrmPrimeDescriptor.layers[i].offset[j];
drmDescriptor->layers[i].planes[j].pitch = vaDrmPrimeDescriptor.layers[i].pitch[j];
}
}
return true;
}
void VAAPIRenderer::unmapDrmPrimeFrame(AVDRMFrameDescriptor* drmDescriptor)
{
for (int i = 0; i < drmDescriptor->nb_objects; i++) {
close(drmDescriptor->objects[i].fd);
}
}
#endif
Reference in New Issue View Git Blame Copy Permalink