The ffmpeg submodule held prebuilt libraries for every platform next to the
sources (about 1.5 GB checked out), and nothing uses those any more. Only the
needed sources are copied in, from the submodule's commit, along with every
header and template they include. The path is new, so checking out commits on
either side of this works.
It's also cut down to what PPSSPP uses: the H.264, AAC, MP3, ATRAC3 and ATRAC3+
decoders, the H.264 parser, the demuxers as before, and the FFV1 encoder, AVI
muxer and file protocol for AVI dumping. config.h takes its CONFIG_ values from
FFmpeg 3.0.2's own configure run with just those enabled. The raw H.264 demuxer
stays even though nothing opens raw H.264: probing PSMF data uses it to identify
the video stream's codec, which otherwise comes out as none.
On top of that:
- The files nothing links (configure would build them, but they're never pulled
out of the static library, going by the GCC link map) are left out: mostly
libavutil's crypto and hashing and the alternative DCT/IDCT/FFT versions.
- libswresample isn't built, now that nothing uses it.
- The H.264 decoder only does 8-bit video, which is all the PSP has. The SPS
parser rejects any other bit depth, and the DSP files no longer build their 9-
to 16-bit variants, which the linker kept because the init functions pick one
by bit depth at run time.
That's 160 files to compile instead of 277, and 391 of FFmpeg's vendored, about
5.8 MB. The README in ext/ffmpeg has what was done, the PPSSPP patches in the
copy, how to add a component, and the tradeoffs.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The ndk-build builds (android/jni and libretro/jni) get a static library module
with the files in ext/ffmpeg-build/sources.mk, in place of the prebuilt
libraries. The libretro Makefile compiles them itself, with FFmpeg's include
directories first so its config.h wins over the others on the include path, and
puts the objects in obj-ffmpeg/ rather than next to the sources, where a
configure build may have left its own.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The vertex decoder's C++ steps may fuse, like the vertex JITs do; with
contraction off everywhere, RISC-V's VertexJit test found the JIT and
the steps disagreeing on morphed float UVs. The flag now applies to
Core/MIPS only, in CMake and the libretro Makefile. ndk-build has no
per-file flags, so the legacy Android.mk goes back to the default.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Clang and GCC fuse a * b + c into one rounding on CPUs that have fused
multiply-add, which on ARM64 made the interpreter's dot products, vavg,
vdet and vqmul round differently from x86-64. Fused operations now only
happen where the code asks for them. MSVC doesn't contract by default.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
GLEW resolves its entry points through GLX unless built with GLEW_EGL. On a
display stack that provides no GLX -- Wayland, or GBM -- glxewInit() returns
GLEW_ERROR_NO_GLX_DISPLAY, glewInit() fails, and the GL backend never starts.
Building the EGL backend resolves them through eglGetProcAddress instead, which
works on both. It is opt-in rather than the default so existing GLX builds are
untouched.
Note that changing this needs ext/glew/glew.o deleted to take effect, since the
Makefile is not a prerequisite of the object.
