google
external/libaom: optimization
The x86/ARM architectures make use of SSSE3 and NEON for the codec support in this library. We probably will need some optimization work done to meet parity.
FYI: optimized riscv64 vector implementations of NEON intrinsics are coming to SIMDe: https://doi.org/10.48550/arXiv.2309.16509 https://github.com/simd-everywhere/simde/issues/1087
(SIMDe already provides portable implementations of SSSE3 (and almost all of NEON) which compile and work on riscv64 with GCC or clang)
Obviously hand-optimized RVV 1.0 intrinsics would be better, but this can be a stopgap.
We should probably reach out to the libaom team and kick-start the long-term plan. At the moment, I don't know of anyone who needs this urgently enough that we need to implement a stopgap.
At a glance, it looks like we'll have to do a few operations:
- We'll have to update libyuv in the aom repository, it's version is old and doesn't contain the RVV intrinsics.
- We'll have to create equivalent files for aom_dsp/arm by implementing intrinsics for RVV. This is about 60 files to rewrite.