00:18airlied: bnieuwen1uizen, mareko : I noticed radv sets 8 simd waves in one place for iceland/tonga as well as polaris, but it seems inconsistent where we work out for shader stats
00:19airlied: I've sent a patch to refactor a bit in the area, but it might need iceland/tonga added
00:45mareko: airlied: 8 is implied by LLVM on tonga/iceland due to fixed SGPR usage, while not on Polaris
00:46airlied: mareko: so the stats code in mesa should use 8 there as well then?
00:47mareko: airlied: Polaris needs to clamp the wave count in Mesa, while tonga/iceland don't
00:48airlied: mareko: ah
00:51mareko: because Polaris can use 80 SGPRs, which is 10 waves on Fiji, but Polaris has a limit of 8
00:52mareko: Tonga/Iceland don't have the limit of 8, but the compiler sets SGPRs = 96, which is 8 waves, so the max waves computation can
00:52mareko: can't return more than 8
00:54airlied: mareko: cool thanks for that, make sense how it is then,
00:54airlied: I was just reducing the number of places we store per-family info between radv/radeonsi
02:46alyssa: Is there a GLES extension that provides glLogicOp?
02:49airlied: alyssa: not that i know off
02:51alyssa: alright, thanks :)
02:53imirkin: airlied: fyi, it's part of ES 1.1
02:53imirkin: (and maybe 1.0)
02:53airlied: imirkin: yup but nobody ever brought it back to gles2/3
02:57imirkin: [chances are that whatever core driver the RE is being done from would support gles1...]
05:02Kayden: huh, are people not using NIR for ARB vp / fixed function vp in gallium drivers?
05:02Kayden: looks like it still goes through TGSI
05:02Kayden: yeah...looks like i965 is the only prog_to_nir user :(
05:05Kayden: I suppose people are probably using TTN in that case
05:05Kayden: so it still gets the NIR optimizations...
05:12imirkin: Kayden: you probably know this, but for desktop (where one is likely to encounter ARB / ff shaders), that's radeonsi and nouveau. so just plain TGSI.
05:18Kayden: right, good point
05:24imirkin: that said, freedreon/vc4 support desktop GL, so for piglits and whatnot, it goes through there
06:53pmoreau: karolherbst: I did not include OpTypeBool because it is not a valid kernel argument (see Section 2.8 of the OpenCL SPIR-V Environment Specification).
06:55pmoreau: karolherbst: Also, from the OpenCL C Specification (2.0, Section 6.9): “Arguments to kernel functions in a program cannot be declared with the built-in scalar types bool, size_t [...] or a struct and/or union that contain fields declared to be one of these built-in scalar types.”.
06:57HdkR: Curious, I wonder what that justification was
06:58pmoreau: It’s the line right below that “The size in bytes of these types are implementation-defined and in addition can also be different for the OpenCL device and the host processor making it difficult to allocate buffer objects [...]”
09:36karolherbst: pmoreau: it fails parsing the spirv though
09:37karolherbst: with "unknown type"
09:37pmoreau: Could you share a link to that SPIR-V please?
09:39karolherbst: it isn't getting used as a param
09:39karolherbst: https://gist.githubusercontent.com/karolherbst/faecb461ad86b34cb87a815156687340/raw/853e40c686ff013ad35f4bcd57ce08ba1117b2bd/gistfile1.txt 09:39pmoreau: Ah yes, I see how it can fail, if you have a composite object of bools.
09:40karolherbst: pmoreau: can you pass bools inside structs?
09:40karolherbst: into the kernel?
09:40pmoreau: Not as kernel arguments
09:41pmoreau: I’ll rework the code identifying the types, because we will have the same issue with all the types that are not valid as kernel arguments.
09:41karolherbst: pmoreau: fails: 179 passes: 393 by the way (subtests of the CTS)
09:42pmoreau: Cool! Where do those fails come from? And which tests are you currently ignoring?
09:43karolherbst: pmoreau: look at my runner: https://gist.github.com/karolherbst/1373fa96baa1e1045573dbb91d125290 09:43karolherbst: this includes the tests I am running
09:44karolherbst: fails are due to random reasons
09:44pmoreau: Nice; you are even testing printf? :-D
09:44karolherbst: well, it fails
09:44karolherbst: but yeah
09:44karolherbst: it is required, no?
09:45pmoreau: I think so, yes, but it wasn’t even being tested IIRC.