Add scheduling directive to cap the max GPU registers - #9389

Open

abadams wants to merge 13 commits into

mainfrom

abadams/gpu_max_registers

Commits on Aug 25, 2026

Add a gpu_max_registers scheduling directive

abadams
and
claude
committed
8922f9d
Use gpu_max_registers in the depthwise separable conv app

abadams
and
claude
committed
a3bd9bd
Call gpu_max_registers on the output buffer directly

abadams
and
claude
committed
7c3d420
Say that gpu_max_registers works in both directions

abadams
and
claude
committed
bfcac2e
Describe gpu_max_registers without claiming a mechanism

abadams
and
claude
committed
ff76877
Drop the remaining unfounded claims about gpu_max_registers

abadams
and
claude
committed
59258c2
Let gpu_max_registers(0) ask for the automatic choice

abadams
and
claude
committed
b023522
Report the depthwise conv GPU runtime on the card it was tuned on

abadams
and
claude
committed
53e8943
Compare the depthwise conv app against a baseline we can rerun

abadams
and
claude
committed
05c6bad
Retune the depthwise conv GPU schedule for the current card

abadams
and
claude
committed
684e484
Say what gpu_max_registers actually buys the depthwise conv app

abadams
and
claude
committed
fa8b28e
Note the occupancy gpu_max_registers actually achieves here

abadams
and
claude
committed
b824f3d
Apply pre-commit auto-fixes
halide-ci[bot]
committed
6d0e76e