* Enable HEVC/H.265 software video decoding on Linux
Adds FFmpeg-based HEVC/H.265 software decoding support for Linux,
providing a fallback when hardware decode is unavailable (e.g., NVIDIA
GPUs where VA-API is not functional in Chromium).
- chromium_src override for supported_types.cc (HEVC as built-in codec)
- brave/third_party/ffmpeg/sources.gni with HEVC source lists,
imported via ffmpeg_generated.gni patch
- Minimal patches for ffmpeg_common.cc and ffmpeg_video_decoder.cc
(anonymous namespace / file-local functions, not overridable)
- All HEVC code paths gated on BUILDFLAG(ENABLE_PLATFORM_HEVC) to
allow disabling via build args
- FFmpeg config patches for Linux x64 (config.h, config_components.h,
codec_list.c, parser_list.c, autorename wrappers)
- Enable enable_platform_hevc and enable_hevc_parser_and_hw_decoder
- Add ffmpeg to patch apply/update infrastructure with s.path guard
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Address review: advertise HEVC support to web content on Linux
On Linux, upstream's IsDecoderHevcProfileSupported takes the
PLATFORM_HAS_OPTIONAL_HEVC_DECODE_SUPPORT path and consults the
supplemental profile cache, which is populated only by hardware
decoders registering via the GPU mojo interface. Since we ship a
bundled FFmpeg software decoder that never touches that cache,
MediaSource.isTypeSupported('video/mp4;codecs="hev1.1.6.L93.B0"')
returned false even though the decoder could handle the stream.
Override IsDefaultDecoderSupportedVideoType so that when upstream
rejects HEVC for that reason alone (cache empty), we fall back to
returning true. HDR, color space, and proprietary-codec gating from
upstream are preserved.
* Address review: round-trip new-file patches through apply/update_patches
Two related issues with how brave's patch tooling handles patches that
create new files (new file mode diffs):
apply_patches would fail on a fresh checkout with "Target file does
not exist" because isPatchStale treats a missing target as SRC_REMOVED
and bails out before apply. Teach it to recognize "new file mode"
patches and re-route through SRC_CHANGED so apply proceeds. Also
unlink any leftover target before apply so re-runs on an existing
checkout don't hit "already exists in working directory".
update_patches would silently delete new-file patches because
getModifiedPaths only inspected tracked/modified files. git apply
creates new files without staging them, so they show up as untracked
(??) and never reach the diff. Extend it to pick up untracked paths
that correspond to an existing patch, and generate their diff via
git diff --no-index /dev/null <path> to match the new-file-mode
format produced by git apply.
The four autorename_*.c.patch files in third_party/ffmpeg now round-
trip cleanly across apply/update cycles, which addresses the
"update_patches currently removes some patches you added here" review
comment.
* Address review: rebase HEVC patches onto Chromium 147 ffmpeg
Refresh patch index hashes to the current third_party/ffmpeg tree,
tighten the ffmpeg_generated.gni patch to avoid an empty line at the
hunk boundary (presubmit warning), and pick up libavcodec/x86/hevc/
dequant.asm, which upstream added since the original patch set and
whose absence caused an ff_hevc_dequant_8_ssse3 link error.
All four autorename_*.c patches, the config/config_components toggles,
the generated-list import, and the BUILD.gn include_dirs tweak now
apply cleanly on Chromium 147 and the resulting build links and plays
HEVC streams end-to-end.
* Address review: host autorename stubs in brave/third_party/ffmpeg
Moves the four HEVC autorename .c stubs from patches/third_party/ffmpeg/
into brave/third_party/ffmpeg/libavcodec/ (and libavcodec/hevc/) so
they're regular brave-owned sources rather than new-file patches against
upstream ffmpeg. This is what goodov suggested in review.
The stubs use fully-qualified includes like "libavcodec/hevc/cabac.c",
which resolve through the existing -I "." on ffmpeg_internal — so the
upstream source is what actually gets compiled, just under a different
object name to keep the linker happy. sources.gni references the new
files with absolute //brave/... paths.
One thing I missed on the first build: libavcodec/cabac.c and
libavcodec/hevc/cabac.c are two different files, and the hevc stub
needs to point at the latter or the linker complains about duplicate
ff_init_cabac_decoder. Fixed before pushing.
Since no new-file patches are left in the PR, the apply/update_patches
tooling changes from c1b2b937cf9 don't have a reason to be here anymore.
Reverted them — gitPatcher.js and updatePatches.js are back to where
they were before this PR.
Tested locally on Linux x64 component build: apply_patches and
update_patches round-trip cleanly, gn_check passes, ffmpeg and media
targets both build, and HEVC playback still works end to end.
* Address review: centralize HEVC enabled check via FFmpegSupportsHEVC helper
Introduces FFmpegSupportsHEVC() and HEVCAsH264IfFFmpegSupportsHEVC() in
chromium_src/media/base/supported_types.{h,cc}. The helpers wrap the
BUILDFLAG(IS_LINUX) && BUILDFLAG(ENABLE_PLATFORM_HEVC) check in one
place so callers don't have to repeat it.
Most of this is shrinking the existing patches:
- patches/media-filters-ffmpeg_video_decoder.cc.patch: drops the nested
#if BUILDFLAG blocks around the kHEVC switch case and IsCodecSupported.
The switch now reads `HEVCAsH264IfFFmpegSupportsHEVC(config.codec())`
so kHEVC reuses the kH264 threading branch, and IsCodecSupported maps
the incoming codec the same way before its kH264 comparison.
- patches/media-ffmpeg-ffmpeg_common.cc.patch: one-line
`if (FFmpegSupportsHEVC()) return "h264,hevc";` instead of a
C-preprocessor branch.
- chromium_src/media/base/supported_types.cc: IsDecoderBuiltInVideoCodec,
IsDefaultDecoderSupportedVideoType and IsDecoderSupportedVideoType all
go through FFmpegSupportsHEVC() now. The third one matters: blink
(webcodecs, media_capabilities) calls the public
IsDecoderSupportedVideoType, and the upstream definition calls
IsDefaultDecoderSupportedVideoType internally — which under our
#define rebrand routes to _ChromiumImpl and bypasses our override.
Overriding the public entry point too keeps the helper authoritative
whichever path the caller takes.
The override file now includes media/base/supported_types.h *before*
the rebrand #defines and forward-declares the *_ChromiumImpl variants.
That way the MEDIA_EXPORT declarations of the three Is* functions are
visible with their original names when our override definitions are
compiled, so the symbols get the right visibility without having to
repeat MEDIA_EXPORT on every override body. Verified by checking
`nm -D libmedia.so` — IsDecoderBuiltInVideoCodec,
IsDefaultDecoderSupportedVideoType and IsDecoderSupportedVideoType are
all exported, and media:media_unittests links cleanly.
sources.gni: collapses the two duplicated `is_linux && ffmpeg_branding`
conditions into a single `enable_ffmpeg_hevc_support` flag derived from
the upstream `enable_hevc_parser_and_hw_decoder` build arg. Drops
chromeos / fuchsia from the condition (not relevant for this PR).
Test adjustments for the new HEVC-supported reality on Linux:
- media_unittests: SupportedTypesTest.IsDecoderBuiltInVideoCodec and
IsDecoderSupportedVideoTypeBasics asserted HEVC=false unconditionally;
switched to `EXPECT_EQ(..., FFmpegSupportsHEVC())` so they pass on
Linux with the FFmpeg HEVC decoder built in and stay correct
everywhere else.
- browser_tests: EncryptedMediaSupportedTypesTest::CheckPlatformHevcSupport
asserted UNSUPPORTED for HEVC on Linux. With the software decoder
built in HEVC is supported via Clear Key too, so the assertion now
flips to EXPECT_ECK_PROPRIETARY when FFmpegSupportsHEVC() is true.
This fixes EncryptedMediaSupportedTypesClearKeyTest.Video_MP4 and the
two ExternalClearKey variants (Basic, Video_MP4).
Also renames the four autorename stubs to drop the "autorename_" prefix
and removes the "Auto-generated" wording from their comments — they
aren't produced by a script, they're hand-placed brave sources. The
sources.gni references are updated to match.
Verified on Linux x64 component build with chromium 149: gn_check,
format, presubmit, full brave build, SupportedTypesTest.* (32 tests)
and the four EME browser_tests above all green. HEVC playback + MSE
isTypeSupported confirmed end to end on a Main / Main10 / Level 5 test
clip.
* Address review: enable HEVC software decoder on linux-arm64
Mirrors the linux-x64 enablement on linux-arm64 so HEVC isn't a
brave-on-x64-only feature.
ffmpeg config patches: four new patches under chromium/config/Chrome/
linux/arm64/, symmetric to the existing x64 ones — flip CONFIG_BSWAPDSP,
CONFIG_HEVC_DECODER and CONFIG_HEVC_PARSER to 1, and add ff_hevc_decoder
/ ff_hevc_parser to codec_list.c / parser_list.c.
sources.gni: new `if (enable_ffmpeg_hevc_support && current_cpu ==
"arm64")` block pulling in the libavcodec/aarch64 HEVC NEON files that
upstream's Makefile lists for CONFIG_HEVC_DECODER (hevcdsp_init,
hevcdsp_{deblock,dequant,idct}_neon and h26x/{epel,qpel,sao}_neon).
Also introduce a `brave_third_party_ffmpeg_gas_sources` list — arm64
hand-written assembly goes through ffmpeg_gas_sources upstream rather
than ffmpeg_asm_sources (which is nasm-only). The ffmpeg_generated.gni
patch is extended to append it.
Verified: gn_check, format, presubmit all green. autoninja
third_party/ffmpeg:ffmpeg in a cross-compiled Component_arm64 build dir
links cleanly. Couldn't get a full media:media_unittests build past
unrelated clang segfaults on V8 torque-generated sources locally, but
that's host-side flake — the brave CI arm64 builders should hit a clean
environment.
Don't have arm64 hardware to runtime-verify; relying on CI for that.
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
