02:27 <cphealy> dv_: I don't know the answer to that question for sure but we do know that the Vivante GPUs are IMR (immediate mode renderer) and not TBR (tile based renderer). I think discard is generally useful with TBRs to avoid having to populate the GRAM at the start of a render operation when it's not needed so it may not be useful with the Vivante GPU.

05:17 <dv_> cphealy: vivante GPUs aren't tile based? huh, I thought they were

05:31 <cphealy> dv_: The older i.MX51/53 with Adreno GPU are tile based. They have either 128KB or 256KB of GMEM for tile rendering.

05:32 <cphealy> It is a little confusing though as the Vivante GPU (unlike the Adreno GPU in the i.MX51/53) uses a tiled render buffer format. The render buffer format is completely independent of if a GPU is IMR or TBR though.

08:12 <daniels> right, there are two 'tiling's: whether or not your render progresses through a frame in raster-scanline order or in tile order (literally everyone now does tile order), and also IMR/TBR though for the latter it's more useful to think of it in memory access terms

08:13 <daniels> a2xx, Mali, PVR, and other GPUs do not have direct access to the framebuffer; in order to sample pixels/texels they need to load portions of the framebuffer into local tile memory, and in order to write rendered pixels they need to write the tile buffer out to system memory

08:13 <daniels> so Vivante/Intel/AMD/NV/etc can still access framebuffer memory directly, even though they progress through framebuffer with a tile-based access pattern

08:15 <daniels> GL_EXT_discard_framebuffer and EGL_KHR_partial_update are of most use in the indirect category, since they allow you to skip the first load into tile memory when it's not necessary