zstd/DecompressSequences_LdsFseCache: Prefer TG size of 32 instead of 64 and some refactoring.#73

Draft

Jonathan-Weinstein-AMD wants to merge 5 commits intomicrosoft:developmentfrom

Jonathan-Weinstein-AMD:decompress-sequences-lds-fse-cleanup

Commits on Feb 19, 2026

zstd: Prefer DecompressSequences_LdsFseCache32 (not 64) on RDNA.
Jonathan-Weinstein-AMD
committed
zstd: DecompressSequences_LdsFseCache: have all threads but thread 0 bail after LDS stores. Also remove the condition for the GroupMemoryBarrierWithGroupSync. The IHV compiler should not emit a bar…
Jonathan-Weinstein-AMD
committed
zstd: Refactor DecompressSequences_LdsFseCache to not invoke ZSTGPU_DECODE_SEQ twice. The benefit is less compiled code for easier performance analysis. Performance seems about the same. The ZSTGPU…
Jonathan-Weinstein-AMD
committed
zstd: Remove zstdgpu_Backward_BitBuffer_V0_CanRefill_Huffman. I tested/ran this at one point before and it was faster for the loop to only have 1 termination condition. Added a comment near the #if…
Jonathan-Weinstein-AMD
committed

Commits on Feb 20, 2026

zstd: Do alternate firstbithigh for known nonzero input to workaround current AMD GPU compiler from not emitting scalar instructions. This tidies up the ISA for DecompressSequences_LdsFseCache a bi…
Jonathan-Weinstein-AMD
committed