#asahi-gpu · 2021-04-10 16:42
bloom(372, 6),#asahi-gpu · 2021-04-10 16:42
glibcactually what bothers me is the 372/388 boundary :) 20/36 -> computation could simply be in unit of 32 bytes#asahi-gpu · 2021-04-10 16:42
bloomWorse:#asahi-gpu · 2021-04-10 16:42
bloomIf it's log2, then why is the there are threshold betweeen 20 and 36?#asahi-gpu · 2021-04-10 16:42
bloom^ This should be bugging you#asahi-gpu · 2021-04-10 16:41
bloom (20, 1),#asahi-gpu · 2021-04-10 16:41
bloom (36, 2),#asahi-gpu · 2021-04-10 16:40
bloombut the coefficients get all weird for small n#asahi-gpu · 2021-04-10 16:40
bloomtable.py shows a clear logarithmic pattern (as jn__ also identified) so this doesn't seem far off from AGX#asahi-gpu · 2021-04-10 16:40
bloomso we need to round up the number of 16-byte slots, and then round up the power-of-two#asahi-gpu · 2021-04-10 16:40
bloomOn Mali, the idea is that the hardware allocates a power-of-two number of 16-byte slots#asahi-gpu · 2021-04-10 16:39
bloomYeah#asahi-gpu · 2021-04-10 16:39
glibcah, the input number is TLS byte size in shader?#asahi-gpu · 2021-04-10 16:39
bloomceil(log2(ceil(s / 16))) I guess it was#asahi-gpu · 2021-04-10 16:39
bloomer#asahi-gpu · 2021-04-10 16:37
bloomuh. apparently that comment is wrnog, should be ceil(log2(s / 16)) but I digress#asahi-gpu · 2021-04-10 16:37
bloomfwiw, the equivalent field on Mali is https://cgit.freedesktop.org/mesa/mesa/tree/src/panfrost/lib/pan_scratch.c#n70 explanation in the comments#asahi-gpu · 2021-04-10 16:36
jn__else 0x30 would be in it#asahi-gpu · 2021-04-10 16:36
jn__no, less#asahi-gpu · 2021-04-10 16:36
bloomheh#asahi-gpu · 2021-04-10 16:35
jn__actually two bits of mantissa?#asahi-gpu · 2021-04-10 16:35
jn__or, wait#asahi-gpu · 2021-04-10 16:35
jn__my approximation is more or less a floating point with a mantissa of between one and two bits#asahi-gpu · 2021-04-10 16:33
bloomclearly you are right, that's in the data. but why?#asahi-gpu · 2021-04-10 16:33
bloomIt's the division-by-3 that's baffling me#asahi-gpu · 2021-04-10 16:33
bloomit's certainly a possibility :)#asahi-gpu · 2021-04-10 16:33
bloomgot it#asahi-gpu · 2021-04-10 16:33
jn__it generates the boundaries (0x20, 0x40, etc.) along with the output numbers (2,3,4,..)#asahi-gpu · 2021-04-10 16:32
bloomjn__: what's f2gen?#asahi-gpu · 2021-04-10 16:30
DarkShadow44hm interesting#asahi-gpu · 2021-04-10 16:26
jn__https://susepaste.org/view/raw/14872569 f2()#asahi-gpu · 2021-04-10 16:26
DarkShadow44jn__: mind showing?#asahi-gpu · 2021-04-10 16:23
jn__i have a fitted function and it seems reasonably compact, but it could be overfitted still#asahi-gpu · 2021-04-10 16:13
jn__bloom: how much data do you have for the validation of functions?#asahi-gpu · 2021-04-10 16:05
glibc@bloom: thank you for your suggestion. I will have a look.#asahi-gpu · 2021-04-10 15:56
bloomSpecifically, can you write a function that takes in the left number (340) and outputs the right one (6), fitting so well that it can predict the results for future data?#asahi-gpu · 2021-04-10 15:55
bloomCan you find the pattern in this?#asahi-gpu · 2021-04-10 15:55
bloomglibc: https://github.com/AsahiLinux/gpu/blob/0f4f417ef85f2dcb5840427fc07a1162062f7594/docs/table.py#asahi-gpu · 2021-04-10 15:50
DarkShadow44bloom: Would you mind sharing a shader that uses that? Would be helpful to understand that.#asahi-gpu · 2021-04-10 14:11
bloomeven have spilling disassembled correctly, nice!#asahi-gpu · 2021-04-10 13:58
bloomdougall: I have to hand it to you, no clue how you figured out the control flow stuff <3#asahi-gpu · 2021-04-10 12:07
jn__i'll wait for bloom or someone else to come around with good ideas :)#asahi-gpu · 2021-04-10 12:07
glibceither GPU reverse engineering, or driver work#asahi-gpu · 2021-04-10 12:06
jn__glibc: it somewhat depends on your field of interest, too — what kind of work are you particularly interested in?#asahi-gpu · 2021-04-10 12:03
glibcthanks#asahi-gpu · 2021-04-10 11:58
robinpglibc: your probably better off asking in the main channel #asahi. I think at the moment they are just getting through the low level initial support - I'm hoping that there will be more JJ once they get that sorted#asahi-gpu · 2021-04-10 11:10
glibcDo you have suggestions of "relatively" easy tasks to help the project?#asahi-gpu · 2021-04-09 00:02
bloomhaving one of those nights#asahi-gpu · 2021-04-09 00:02
bloommixed up her endians