00:05 <cphealy> mmind00: Any chance you know the bus architecture with the RK3266 in Odroid-GO? I'm curious if there are any bus HW counters that can exposed to see DRAM read/write bytes or busy cycles.

02:50 <macc24> cphealy: look who made that commit :D

03:06 <alyssa> mmind00: is all knowing rockchip guru

03:28 <macc24> oh yeah i should release dtg-ng, time to build 90% of the os in a weekend

04:18 <alyssa> bbrezillon: spent much of today cleaning up varying handling in gallium

04:18 <alyssa> separate_shader.random.* s all passing now, dunno what the issue was, don't especially care to bisect and find out :-p

06:31 <icecream95> huh? perf is showing panfrost_emit_texture_descriptors at 22% of CPU time for a drawoverhead test, but if I comment out the call to panfrost_emit_const_buf (which doesn't call emit_texture_descriptors) it drops to only 0.74%

06:33 <amonakov> icecream95: this can happen if the library is stripped (hot function does not have a dynamic symbol, panfrost_emit_texture_descriptors is the immediately preceding dynamic symbol)

06:45 <amonakov> (btw, instead of building full debuginfo with -g, for use with perf you can build with -g1 (line info), it is much smaller)

07:12 <icecream95> amonakov: That's not happening here, the debug info is all there and correct

07:18 <icecream95> panfrost_emit_texture_descriptors gets called immediately after panfrost_emit_const_buf, so maybe emit_const_buf is thrashing the cache and causing the next function to spend a lot of time waiting for RAM

07:24 <icecream95> This effect only happens on the big cores, for LITTLE cores panfrost_emit_const_buf gets blamed for as many CPU cycles as it should be

07:30 <amonakov> perhaps you're seeing sampling skid then? (PMU interrupts are slightly asynchronous, so on deeply OoO CPUs you see events recorded "later" than the instructions causing them)

07:31 <amonakov> "precise" events usually need extra hw support ("PEBS" on Intel, "IBS" on AMD)

07:32 <icecream95> panfrost_emit_const_buf does a couple of reads from uncached memory, so those might take a while to complete

15:02 <alyssa> as usual Mali < AGX architecturally

15:03 <alyssa> AGX uniform registers are specified indirectly as separate ranges, so the driver can do arbitrary UBO pushing with no CPU side copying / readback of GPU memory / stalls

22:28 <kinkinkijkin> just noticed this channel was removed from my autojoin for some reason