10:53 <rellla> anarsoul, enunes, cwabbott: i pushed a branch with extensive (and ugly) logging of gpir to see, what is happening, when we run into the 512 instr limit

10:54 <rellla> it's here: https://gitlab.freedesktop.org/rellla/mesa/-/commits/lima-gpir-branch-opt-rebased_WIP

10:55 <cwabbott> rellla: sorry, i haven't had much time recently to look at it... i did have time a while ago to finish my rework of the scheduler to fix some infinite loops in the scheduler w/ some piglit tests on that branch, and i was hoping to finish that first, but you can go ahead if that's ok with you

10:57 <rellla> cwabbott: np, we definitly run into an infinite loop, but i guess i'm not able to solve that issue. probably if someone gives me a hint, whats wrong :)

10:57 <rellla> wait a second, i push a logfile and i guess you will see very fast, whats going wrong.

10:58 <rellla> the problem is, that we end up with scheduling just moves

11:00 <rellla> http://imkreisrum.de/deqp/512_c/log is the log from dEQP-GLES2.functional.shaders.indexing.tmp_array.vec3_dynamic_loop_write_dynamic_read_vertex , which is one of the failing tests

11:00 <rellla> be aware, it's 32M size

11:02 <rellla> scheduler is limited to 256 instructions, but the problem is beginning in block 5 with instr nr. 142 (in the last scheduling iteration)

11:06 <cwabbott> ah, that might be the same infinite-loop problem

11:07 <cwabbott> i don't have the details paged in now, but it's pretty involved to fix it in a way where it's guaranteed to work

11:08 <cwabbott> just so, so many corner cases

11:16 <rellla> :)

18:38 <anarsoul> cwabbott: I wonder if it would help to always lower select to if-else, in this case we are getting rid of instruction that limits already limited input bandwidth

18:39 <anarsoul> rellla: ^^

18:46 <cwabbott> i think that for small things select is still gonna be less instructions

18:46 <cwabbott> a branch is a whole instruction wasted

18:57 <anarsoul> cwabbott: more instructions but easier to schedule, no need to have 3 inputs ready for a single op