08:58colinmarc: how do people here develop on multiple gpus? do you use eGPU enclosures?
09:00tnt: colinmarc: I have 2 GPUs in my workstation (nvidia/amd) and then a laptop with intel graphics.
09:01colinmarc: my workstation has a second PCI slot, but I feel like I hit so many issues trying to switch at boot or runtime (I currently have a 6650xt, I wanted to run a 4070 alongside). Also I learned too late it doesn't really have enough lanes...
09:04landofgeeks: boolean combinations can not be more than 1024 for 32bit alu, since 64rows of partial products is 32*32 boolean entries, so 1024 boolean combinations. But this is not as to how i solve the compute engine by default, instead i have invariant 561 format. so max combinations of single number powers is 561 that's for one operand , and another 561 combinations for another operand (they have
09:04landofgeeks: indexes shared, and asymmetric selection from 883 440+443 cause -115-768 of it is 0 and incoming verts are operands then 883-72-512=184 and second is 186 as from another bank then 184+186-512=-142), you select 2 of these numbers you have two negative values -72 and -70 and you have X number of paired combinations cause the weighted powers are invariant any pairs sum is invariant, and
09:04landofgeeks: finally a LOGICS intrinsic to do the selection in packed format during runtime. It's an engine, data banks work differently, they sample the distances and measure the coefs and input them to formula from table/matrice. I develooped a filesystem guidelines or addressing shematic.
09:12mareko: karolherbst: from my notes: gallium changes: pipe_scissor_state, pipe_surface, pipe_framebuffer_state, pipe_sampler_view, pipe_image_view can’t represent width/height/maxx/maxy/row_stride=64K
09:13mareko: some of our hw can do 64k textures
09:16landofgeeks: so forgive me, but my code so much powerful than yours, i do not contribute this code, for my snb two laptops and before releases i lack vulkan API and driver for. I have two vulkan capable laptops and one fitpc though, intel and two amd's, so to decode a video with your stack one needs a vulkan driver, and 2012 macbooks which i use currently do not have vulkan. I have much hardware
09:16landofgeeks: available though.
09:26landofgeeks: so what i do is write my own code for those hardware's and have enough happy with the outcome, cause i myself wrote it, yeah to write code systems like linux os are enough good , the last which i use to input my own paradigm, i am not sure who to thank for this but any other than that, i fork all your efforts for my own purpose.
09:26landofgeeks: or own use.
10:19karolherbst: mareko: I see. not really caring about anything besides sampler_view and image_view though. But if you think it makes sense to bump it for GL as well, then I guess we could do it more globally
10:42alyssa: colinmarc: for me, ssh
16:44mareko: karolherbst: also pipe_resource
16:45karolherbst: right...
16:45karolherbst: though I think that's the easiest to argue about. Not sure how perf critical the other things are, but maybe one should figure out how high the costs are, so we can decide if it's worth it or not?
16:46mareko: should be fine
16:46mareko: also I'd be fine with just 32k
16:46karolherbst: might need to reshuffle things in pipe_resource, because I think there are still gaps anyway
16:47karolherbst: 24 bits after usage
16:47karolherbst: in pipe_resource at least
16:47mareko: amd hw limits: array_size: 8k, depth: 16k, width/height: 64k
16:47karolherbst: but that's also the one I'm the least worried about
16:47karolherbst: right
16:48karolherbst: the thing with adreno hw apparently is, that their sampled texture limit is 16k, but unsampled it's 64k
16:48karolherbst: or something along those lines
16:50mareko: there is also the issue of binding more than 4G buffer range
16:51mareko: our hw doesn't support 64-bit load offsets in shaders though
16:52mareko: need to use pointers in shaders instead of bindings
16:52karolherbst: rusticl doesn't use bindings anyway
16:53karolherbst: it's all raw pointers