01:08 karolherbst: do we have a microbenchmark somehwere testing depth+stencil performance?
01:08 karolherbst: for vulkan
01:12 HdkR: karolherbst: Doesn't gpu-ratemeter have a clear test for depth+stencil?
01:13 karolherbst: good question
01:13 HdkR: Might have more as well
01:18 karolherbst: though not sure if it's tests what I want to test: stencil throughput
01:19 HdkR: Might be a good framework to have that tested added as well at least? :)
01:19 HdkR: test added*
01:21 karolherbst: maybe?
01:21 karolherbst: I just found out we have a stencil compression flag we never turned on (but we did turn on depth compression)
01:21 karolherbst: so I am curious if it changes anything
01:21 HdkR: Oh weird
01:21 HdkR: Is that new?
01:22 karolherbst: maxwell_b+
01:22 HdkR: So not really
01:22 HdkR: I guess I just completely forgot about it
01:25 karolherbst: well.. might be one of those "same as depth compression really" things :D
01:25 karolherbst: no idea why you'd only enable one of those two..
01:35 HdkR: karolherbst: Looks like the imgbw tests that include stencil would be a throughput test?
01:35 karolherbst: HdkR: it tests filling/copying the buffer, no? I meant like actual stencil throughput
01:36 karolherbst: like actual stencil testing
03:48 mareko: only depth
03:48 mareko: gpu-ratemeter doesn't have stencil perf testing yet
03:50 mareko: I guess stencil clear and copy that probably aren't relevant since the reason to test stencil test performance would be to test HiS
09:27 karolherbst: yeah... I was wondering whether compression on stencil might impact stencil testing performance directly, so I kinda was looking for a very targeted microbenchmark, because those things are pretty difficult to notice in real world workloads
18:32 mareko: adding stencil perf tests would be easy, the non-easy part is designing the tests such that the tests target HiS and stencil compression and not just some random stencil ops
18:35 mareko: the depth tests in pix that are already there cover HiZ quite well since they were designed with HiZ in mind; gpu-ratemeter doesn't explain how HiZ is tested, but it's can be inferred from the test names: depth func equal/not_equal -> no HiZ, depth test less -> HiZ
18:37 mareko: the reported results match amdgpu expectations: when HiZ is used, the depth test fail results are 4x higher than non-HiZ depth test fail results
18:41 mareko: similarly, amdgpu with HiS should get 4x faster stencil testing than without HiS
18:46 mareko: imgbw can only show the effects of stencil compression on clears and copies, which are very noticable on amdgpu
18:48 karolherbst: yeah.. I don't have any numbers on stencil testing, so that's why I'm curious
22:39 dos1: etnaviv
22:41 dos1: focus mishap, move on :D