Shrink-wrap block-level GPU register allocations - #9343

Merged

abadams merged 5 commits into

mainfrom

abadams/gpu_register_sgemm

Aug 24, 2026

Commits on Aug 20, 2026

Give each site of a GPU register allocation its own registers

abadams
and
claude
committed
6f86729
Keep the SGEMM accumulator in registers, and stage asynchronously

abadams
and
claude
committed
92d36b6
Apply pre-commit auto-fixes

halide-ci[bot]
authored and
abadams
committed
65fa1db
Build the cuda mat mul app for the capability it needs

abadams
and
claude
committed
6f7ae72
Use the lambda visitors to find and rewrite the accesses

abadams
and
claude
committed
b49398d