15:16gfxstrand[d]: Someone should probably start sm30 first
15:16gfxstrand[d]: There's zero code for it right now
15:16gfxstrand[d]: The good news is that it doesn't need dependency information so it should be a little easier to get going.
15:17karolherbst[d]: right...
15:17karolherbst[d]: and the stall count thing can even be disabled I think? At least there is hardware verification and it fixes it if you get it wrong, it's just expensive to get it wrong
15:18karolherbst[d]: the stall count is important for dual issuing, because apparently it's a real thing there
15:19gfxstrand[d]: Sure but that's not really an intern project.
15:19gfxstrand[d]: But if someone wanted to get enough working to pass NAK's sanity unit test and then let an intern start filling in ALU ops, that'd be a decent project.
15:20karolherbst[d]: yeah..., it's still entirely TBD what they'll be working on, just collecting a bunch of ideas for now
15:20gfxstrand[d]: sure
15:21gfxstrand[d]: The other big thing needed on Kepler is porting the storage image code to NIR
15:22karolherbst[d]: mhhh, right
15:29gfxstrand[d]: And I'm a bit unsure exactly how I want that to work. I'm a bit inclined to bind the image descriptor table as a cbuf and have it just read the HW descriptor.
15:30karolherbst[d]: isn't it already a cbuf anyway?
15:30gfxstrand[d]: Not today
15:31karolherbst[d]: I mean the image handle makes the tex ops fetch stuff from a cbuf, but maybe it's another layer of indirection?
15:31gfxstrand[d]: But it's that or we pass a bunch of data as part of the descriptor that's in a different format than the texture header
15:31karolherbst[d]: right
15:31gfxstrand[d]: karolherbst[d]: There are so many layers of indirection....
15:31karolherbst[d]: yeah.. I always get confused which one is the relevant one
15:32karolherbst[d]: though binding it as a cbuf might be a bit too much
15:33karolherbst[d]: could also just push the address somewhere and fetch the desc at runtime through global reads
15:33karolherbst[d]: mhh
15:33karolherbst[d]: but that might suck
15:33karolherbst[d]: or it's formatted load/stores just aren't support and cope
15:34karolherbst[d]: though not sure how terrible that would be
15:55gfxstrand[d]: Yeah, we need the start address, size, and tiling information.
15:56gfxstrand[d]: That's all in the texture header and it shouldn't be too horrible to decode.
15:57gfxstrand[d]: But format conversion is a pain. For that, we'd about need function pointers.
19:09AshiskumarNaik[m]: Hi everyone
19:09AshiskumarNaik[m]: I want ask some questions regarding the ideas listed on the Xorg EVoC site for nouveau project ideas. Is this the right place to ask ?
19:11AshiskumarNaik[m]: Hi everyone
19:11AshiskumarNaik[m]: I want ask some questions regarding the ideas listed on the Xorg EVoC site for nouveau project ideas. Is this the right place to ask ?
20:01gfxstrand[d]: Ask away. It may take a bit before the relevant person answers but if you're bridged to Matrix, you'll get the ping.
21:11AshiskumarNaik[m]: Hi... (full message at <https://matrix.org/oftc/media/v1/media/download/ATk0JNlDrH6UBKnudD1XEOXNkSGKboufs4Mz848hIMx4RIT-958eXMTOepLT64IZ8bfoE3FuSPkJBZghfZCsoFlCeSaomu3gAG1hdHJpeC5vcmcvb3ZvYVJQRVZpaW90dEhUb090Q0FuU0Jw>)