09:26snowycoder[d]: Test run totals:
09:26snowycoder[d]: Passed: 78/3444 (2.3%)
09:26snowycoder[d]: Failed: 507/3444 (14.7%)
09:26snowycoder[d]: format lowering works!
11:54karolherbst[d]: nice
11:55karolherbst[d]: kinda lucky it's not such a disaster as on v3d, where format lowering is not really viable...
12:16snowycoder[d]: karolherbst[d]: Why not?
12:26karolherbst[d]: snowycoder[d]: you need to pack the destination for stores e.g.
12:26karolherbst[d]: so if it's a 1 byte format, you need to pack them into a single 32 bit register
12:27karolherbst[d]: if it's 4 byte, then it's 4 32 bit register
12:27karolherbst[d]: it's not that it's impossible, it's just a giant pita
12:42snowycoder[d]: That's odd 0_o
12:44karolherbst[d]: well.. just how the hardware works
12:45karolherbst[d]: I think there is some implicit format conversion going on on most recent hardware versions, but I think the packing thing still applies
16:36snowycoder[d]: Ok, "simple" imageLoad and imageStore work fairly well across formats on Kepler, we miss the strange stuff: linear layout, msaa and atomics, looking into linear.
16:36snowycoder[d]: Codegen didn't handle it (does opengl even support it?), but it shouldn't be too hard
16:52gfxstrand[d]: We can disable linear on Kepler for now.
17:05gfxstrand[d]: MSAA is a pain but it's not too bad
17:06gfxstrand[d]: Atomics should be easy(ish) assuming the address calculations yield an actual address.
17:16blisto[d]: MSAAAAARG is what people working with it scream
17:31snowycoder[d]: gfxstrand[d]: sucalc ops are made so that you can switch between block_linear/pitch_linear by only changing descriptors, I just need to check what parameters to put in there (thanks to the instruction RE I roughly know what goes where)
17:32gfxstrand[d]: Okay