ARM64 neon fixes
SIMD-accelerated libjpeg-compatible JPEG codec library
Brought to you by:
dcommander
Menu
▾
▴
1 Attachments
#64 ARM64 neon fixes
I found issues while running make test on target aarch64 device, previously Make test never gave me such a comprehensive log on this target. I found 4 major issues and i am attaching the fixes for those. Make test runs well after applying these fixes, have attached make test log for reference.
There is another issue which surface when running make test, decoding with float fails and it looks more like gcc toolchain issue and its definelty not realted to neon assembly.
Ticket moved from /p/libjpeg-turbo/bugs/75/
Comments:
-- Please format new code consistently and using the conventions established by the existing code.
-- The floating point test can be compiler-dependent. It works on all of the platforms I've tested (including Android/ARM and iOS/ARM), but it doesn't surprise me that a different compiler or set of compiler flags would break it, particularly if you're using a non-default floating point setting. If there is a way I can validate the ARM64 code using an emulator, please post instructions.
-- 'make test' was extended in r1267 and r1274 so that it now covers all of the algorithms that are currently SIMD-accelerated. That makes it much easier to validate the conformance of new SIMD implementations. The new regression tests were added in May, so if you were using an earlier code drop than that, you wouldn't have been exercising all of the new regression tests.