#asahi-gpu · 2021-01-10 02:31
bloomTrying to render into a buffer, I created it as SHARED and used newTextureWithDescriptor:offset:bytesPerRow to get a texture object which I assigned to the color attachment#asahi-gpu · 2021-01-10 02:31
bloomDoes anyone here know Metal well? Having an issue which I assume is a dumb user error but wouldn't really be possible in OpenGL so...#asahi-gpu · 2021-01-10 01:47
bloomWhere is this framebuffer?!#asahi-gpu · 2021-01-07 22:24
roxfano/#asahi-gpu · 2021-01-07 22:23
bloomroxfan: That's orthogonal, but certainly fun to see :)#asahi-gpu · 2021-01-07 22:11
roxfanhttps://twitter.com/astarasikov/status/1347287775389884417 not sure if useful#asahi-gpu · 2021-01-07 20:24
Mary_bloom: just got my metallib thing working if you are interested ^^ https://github.com/Thog/gpu/tree/metallib#asahi-gpu · 2021-01-07 20:15
ConeOfAttack[m]With regards to 32 v 16 bit ALU performance, is there a way to see if 32-bit ALUs are repurposed as two 16-bit units? Iirc the last couple of generations from nvidia pushed this for ints and fps probably for neural network workloads.#asahi-gpu · 2021-01-07 20:06
bloomthanks :)#asahi-gpu · 2021-01-07 20:04
q3k|mbloom: impressive work#asahi-gpu · 2021-01-07 20:02
bloomContrast eg Adreno#asahi-gpu · 2021-01-07 20:02
bloomYuzu: No(t yet), which is why I'm concluding there's quite a bit of hw scheduling.#asahi-gpu · 2021-01-07 20:01
Yuzubloom: Fair - was just wondering if you had seen any patterns in generated code that would be obviously inefficient unless there's delays#asahi-gpu · 2021-01-07 20:01
bloomdhewg: I am embarassed by the code, yes#asahi-gpu · 2021-01-07 20:01
bloomYuzu: Too early to tell and it would be speculation anyway, compiler can only see so much.#asahi-gpu · 2021-01-07 20:00
bloomsven: thank you :)#asahi-gpu · 2021-01-07 19:09
dhewgnvm, that's like exactly line 264 further down :)#asahi-gpu · 2021-01-07 19:07
dhewgyou mean "(op[0] & 0x7f) | (op[1] & 0x80)", right?#asahi-gpu · 2021-01-07 19:05
dhewgalthough i found "with the 8th bit from the 8th bit of the *second* byte" weird to parse#asahi-gpu · 2021-01-07 19:03
Yuzuso AGX is a fully dynamically scheduled machine? have you seen any sign of mandatory delays before dependent ops or anything like that?#asahi-gpu · 2021-01-07 19:01
dhewgyeah, seconded, just read your post ;)#asahi-gpu · 2021-01-07 18:28
svenbloom: nice work :)#asahi-gpu · 2021-01-07 18:11
bloomthank you :)#asahi-gpu · 2021-01-07 18:10
Shizbloom: very nice going#asahi-gpu · 2021-01-07 17:39
marcan(sorry, was doing dinner and things)#asahi-gpu · 2021-01-07 17:39
marcanbloom: want a tweet from @AsahiLinux?#asahi-gpu · 2021-01-07 17:14
bloomwrap code pushed but probably fairly broken rn, whoops#asahi-gpu · 2021-01-07 16:20
Mary_there is the two variants#asahi-gpu · 2021-01-07 16:19
bloomOh, right. I didn't know a priori if that was actual GPU binaries instead of LLVM IR serialization#asahi-gpu · 2021-01-07 16:18
Mary_(All of that is standard metal stuffs)#asahi-gpu · 2021-01-07 16:18
Mary_I was meaning using Metal APIs to generate a GPU compiled file and dissassembling the result of it#asahi-gpu · 2021-01-07 16:18
bloomNor was I, sorry#asahi-gpu · 2021-01-07 16:18
Mary_ah okay I wasn't clear#asahi-gpu · 2021-01-07 16:18
bloomUsing an arm64 disassembler to disassemble Apples's proprietary code -- noto ok#asahi-gpu · 2021-01-07 16:17
bloomWriting out own disassembler to disassemble shaders we write ourselves, for ISA r/e -- ok#asahi-gpu · 2021-01-07 16:17
bloomDisassembly != disassembly#asahi-gpu · 2021-01-07 16:17
bloom---Correct.#asahi-gpu · 2021-01-07 16:17
Mary_I mean you aren't disassembling any proprietary code here tho bloom#asahi-gpu · 2021-01-07 16:16
bloomMary_: Disassembly is absolutely a no-no.#asahi-gpu · 2021-01-07 16:16
Mary_bloom, is disassembling Metal dynamic library a no no for you? I noticed that you can ask Metal api to prebuild shaders for the current target you are running on and export it.#asahi-gpu · 2021-01-07 16:12
bloomthe wrap library will be released shortly, need to disentangle it from unrelated experiment code#asahi-gpu · 2021-01-07 16:08
Mary_I think I will try to take a look at it a bit later today see if I can help a bit#asahi-gpu · 2021-01-07 16:04
bloomMary_: Thanks :)#asahi-gpu · 2021-01-07 16:04
Mary_bloom, nice work!#asahi-gpu · 2021-01-07 15:57
bloom_More code to be uploaded to that repo soon, need to clean some stuff up, maybe will push to a branch sooner#asahi-gpu · 2021-01-07 15:54
bloom_...or if you'd rather cut to the chase, code is https://github.com/AsahiLinux/gpu/blob/main/disasm/disasm.c#asahi-gpu · 2021-01-07 15:53
bloom_Hot off the press!#asahi-gpu · 2021-01-07 15:53
bloom_https://rosenzweig.io/blog/asahi-gpu-part-1.html#asahi-gpu · 2021-01-07 15:44
bloom_marcan: ping#asahi-gpu · 2021-01-07 15:26
bloom_blogs