Enable NEON for arm / aarch64

[thomasdangl]

This PR enables the NEON (SIMD) extension for arm / aarch64, which should give us a nice performance boost.

It does so by:

1. removing the arm / aarch64 handling from the nasm path (nasm is x86 only and cannot compile arm assembly)
2. dynamically adding the appropriate defines, includes and compilation units (compare to https://github.com/cisco/openh264/blob/cf568c83f71a18778f9a16e344effaf40c11b752/meson.build#L79-L88)

Best,
Thomas

[ralfbiedert]

Sorry, didn't see this in a flood of CI notifications ...

[ralfbiedert]

The build failure log is humongous. An agent summary of what went wrong is attached below. I'm not sure if this is caused by your changes, or if this is a CI / platform bug (we've had the latter). I'm tentatively assuming its the former though.

---

The CI build failed when compiling openh264-sys2 v0.9.1 for the ARM64 Windows target (aarch64-pc-windows-msvc).

Root Cause

The build system tried to compile ARM64 NEON assembly files (.S files) using MSVC's C++ compiler (cl.exe), which doesn't support assembly source files. This caused the linker to fail with:

LINK : fatal error LNK1181: cannot open input file
'...\out\2a171e1abf1f6202-copy_mb_aarch64_neon.o'

Key Issues

1. Missing ARM64 assembler: MSVC's cl.exe cannot compile .S assembly files. The warnings show:
   cl : Command line warning D9024 : unrecognized source file type
   'upstream/codec/common.\arm64\copy_mb_aarch64_neon.S', object file assumed
   cl : Command line warning D9027 : source file '...' ignored
2. No NASM configured: The build output states:
   No NASM configuration found for target, not using any assembly.
3. Cross-compilation issue: The build is using -Ctarget-cpu=native in RUSTFLAGS while cross-compiling from x86_64 host to aarch64 target, which is problematic.

Affected Files

• copy_mb_aarch64_neon.S
• deblocking_aarch64_neon.S
• expand_picture_aarch64_neon.S
• intra_pred_common_aarch64_neon.S
• mc_aarch64_neon.S

The build needs either a proper ARM64 assembler (like armasm64.exe) or the build script should be configured to skip assembly optimizations for this target.

[thomasdangl]

Oops, my bad! Did not consider that cl.exe does not have an assembler.
As I don't have access to a Windows aarch64 machine for testing, I disabled the proposed optimization for that target.

I added the following comment that explains how to proceed from here if someone else wants to add it in the future.

openh264-rs/openh264-sys2/build.rs

Lines 270 to 272 in 802abea

	// cl.exe cannot assemble .S files.
	// TODO: generalize try_compile_nasm and invoke armasm64.exe.
	if !matches!(target.os, TargetOs::Windows) {

If you rerun the CI, it should work fine now and yield a speed-up of around 200-250% on the respective platforms (tested for 1080p24 on Cortex A76/Linux and Apple M4/MacOS).

Best,
Thomas

[ralfbiedert]

Thanks!

[ralfbiedert]

Should be in 0.9.3