Intel NIR I/O Vectorization Ported From The AMD ACO Back-End - ~10% Performance Boost

[andrei_me]

Michael it is not clear on the news if the improvement is made on:

General NIR codebase, which is what it implies, so everyone that uses NIR would receive this improvement (radeonSI, panfrost, i965, anv, etnaviv?, etc)

ANV's NIR usage(??)

The part that says that the Intel paid Mesa developer, which is the lead ANV developer, confuses everything.

it's
hard to know who will benefit from this improvement with all this qualifications, be like Jon Snow, "I am Jon Snow"

[smitty3268]

Michael it is not clear on the news if the improvement is made on:

General NIR codebase, which is what it implies, so everyone that uses NIR would receive this improvement (radeonSI, panfrost, i965, anv, etnaviv?, etc)

ANV's NIR usage(??)

The part that says that the Intel paid Mesa developer, which is the lead ANV developer, confuses everything.

it's
hard to know who will benefit from this improvement with all this qualifications, be like Jon Snow, "I am Jon Snow"

These particular patches only affect the Intel drivers (anv and iris).

The NIR I/O vectorization itself is already common code, written originally by the Valve developers for ACO. However, while the NIR pass itself is out in common code, it's not enabled by default unless drivers choose to do so. That's because it only makes sense on some hardware, for certain situations. That's pretty standard with a lot of the different NIR passes - the code is generic enough to run on any driver, but whether you actually want to do it is hardware specific.

These changes were Intel deciding to enable it on their driver, and then making a bunch of fixes in their driver (and a couple other common/optional passes they use) to make sure it all worked as intended rather than introducing regressions.