05-20-2023, 03:27 AM
I feel your pain. I wish there was an "--autotune-classic" or similar that would revert to the old method.
Today, the only way to disable autotuning is to manually add your own tuning. On the command line, this is done with -n/--kernel-accel and -u/--kernel-loops. If you do a run, and watch for the accel/loop values that are calculated by autotuning, you can just steal those values and use them and they won't suck.
Side note: if each mask only takes a few seconds to run and you're doing 229k of them, you can probably combine quite a few of them into larger masks.
Today, the only way to disable autotuning is to manually add your own tuning. On the command line, this is done with -n/--kernel-accel and -u/--kernel-loops. If you do a run, and watch for the accel/loop values that are calculated by autotuning, you can just steal those values and use them and they won't suck.
Side note: if each mask only takes a few seconds to run and you're doing 229k of them, you can probably combine quite a few of them into larger masks.
~