21:25glehmann: alyssa: karolherbst: cmarcelo: I think I figured it out: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39180/diffs?commit_id=bc9a303ac83d99bb25008f4b362f6d2d04170305
21:25karolherbst: nice
21:26karolherbst: but yeah.. I think you want to call late somewhere after the last optimization loop
21:27glehmann: well brw has a helper for that
21:27glehmann: brw_postprocess_nir_opts
21:28glehmann: but then, it also calls their mid process opt helper again after simd size selection
21:28glehmann: so that messes everything up
21:29glehmann: ofc it only does this for compute/task/mesh, and all the vk tests for denorm flushing are fs/vs...
21:30glehmann: I was really confused why vk cts passes but cl doesn't