00:40 <HdkR> https://twitter.com/Sonicadvance1/status/1227753702552817664 So many cute Bifrost devices :P

03:21 <icecream95> There are two index minmax functions, u_vbuf_get_minmax_index_mapped and vbo_get_minmax_index, and (for ushort) one of them is twice as fast as the other...

03:28 <icecream95> tomeu: When info->has_user_indices is false the prefetch won't do anything, so maybe STK always goes through the pipe_buffer_map_range path in u_vbuf_get_minmax_index

03:33 <HdkR> Is the buffer linearly accessed? Or is it accessed by index?

03:33 <HdkR> Since prefetching wouldn't really effect linearly fetched memory

03:34 <icecream95> The buffers are linearly accessed for minmax, but accessed by index in the GPU

03:42 <icecream95> Prefetching from halfway through panfrost_draw_vbo should allow the indices to get into L1 by the time u_vbuf_get_minmax_index is executed

21:24 <alyssa> tomeu: https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3818

21:24 <alyssa> We should probably be doing this for 'frost

21:29 <alyssa> But yeah, we should probably be doing that to avoid a staging memcpy

21:37 <alyssa> Anyhow, back to ld/st witchcraft

21:37 <alyssa> Trying to crack the rest of the encoding of arg_1/arg_2

21:37 <alyssa> Guess I need to get o r g a n i z e d

21:44 <alyssa> https://people.collabora.com/~alyssa/ldst.md

21:44 <alyssa> ^ All the combiantions of arg_1/arg_2 used in panfrost

21:51 <alyssa> There's quite a bit of symmetry/structure here, it's just not "nice" int he usual sense.