Update/add to qr_ks_vs_whole_k_prefetch pipeline#3485

Open

qianfengz wants to merge 25 commits intodevelopfrom

whole_k_prefetch_n0loop

Commits on Dec 4, 2025

Initial re-implementation of pipeline qr_ks_vs_whole_k_prefetch in looping Gemm0 along n0 dimension
qianfengz
committed
Add prefetching whole next iteration K path in the pipeline
qianfengz
committed

Commits on Dec 5, 2025

Change in GetKVBlockGemm to let gemm1 to use WarpTile-16x16x16/32x32x8 on mi350
qianfengz
committed
Switch the codes based on the iteration index (first/intermediate/last)
qianfengz
committed

Commits on Dec 6, 2025

Simplify the block_gemm codes
qianfengz
committed

Commits on Dec 7, 2025

Commits on Dec 8, 2025

Using explicit vgpr-saved partition_index with store_tile(lds_window, ...)
qianfengz
committed

Commits on Dec 11, 2025

Separate kN0Sub from kK0 to be used for flexible tile tuning for whole_k_prefetch pipeline
qianfengz
committed

Commits on Dec 14, 2025

Load Q through Lds
qianfengz
committed

Commits on Dec 15, 2025

Commits on Dec 17, 2025

Add support of loading QK tiles of hdim96 without padding to hdim128
qianfengz
committed

Commits on Dec 19, 2025

Add qr_ks_vs_whole_k_prefetch_trload pipeline
qianfengz
committed

Commits on Dec 20, 2025

Commits on Dec 21, 2025

Commits on Dec 22, 2025

Commits on Dec 23, 2025