Etnaviv NPU Optimizations Make It Into Mesa 24.1

In addition to many RadeonSI driver optimizations that were merged just prior to yesterday's code branching and Mesa 24.1-rc1 release, a number of Etnaviv driver improvements were also merged for benefiting that recent Vivante NPU IP open-source driver work.

Merged earlier this cycle for Mesa 24.1 is Teflon as a Gallium3D front-end for TensorFlow Lite and initially used by Etnaviv, the reverse-engineered Vivante graphics driver. The Etnaviv NPU driver support has come together nicely and for some inferencing workloads can be nearly as fast as the proprietary driver.

Tomeu Vizoso who has been leading this NPU effort landed "bunch of fixes and optimizations" that he's been accumulating for the Etnaviv driver.

Among the optimizations are enabling an image cache using the on-chip SRAM that is helping reduce times such as MobileNetV1 dropping from 9.9 ms to 6.2 ms or SSDLite MobileDet from 27 ms to 24.3 ms. Plus a few other Etnaviv improvements as a nice end to the Mesa 24.1 feature work.