George Hotz | Programming | tinygrad: monday afternoon at the tiny corp making things fast

Date of stream 8 Nov 2022. Live-stream chat added as Subtitles/CC - English (Twitch Chat) - three-dot menu icon - Show transcript Source files: - https://github.com/geohot/tinygrad/tree/chonker - https://github.com/geohot/tinygrad Follow for notifications: - https://twitch.tv/georgehotz Support George: - https://twitch.tv/subs/georgehotz Programming playlist: - https://www.youtube.com/playlist?list=PLzFUMGbVxlQs5s-LNAyKgcq5SL28ZLLKC Chapters: 00:00:00 intro 00:00:25 not voting 00:03:00 overview of work (tests) 00:03:39 whiteboard 00:03:00 running tests 00:05:35 what makes things fast, slow 00:08:10 review of pull requests 00:09:26 chonker 00:11:00 bug 1, 50$, 15 min phone call 00:12:20 bug 2, test_max 00:13:10 price only if solved live 00:14:25 full time at tiny corp 00:14:59 call llvm.maximum 00:19:30 current results 00:20:30 amx coprocessor 00:24:40 stable diffusion 3x faster on M1 00:25:20 don't want to talk about chess 00:26:45 checkers 00:27:10 stability founding 00:27:40 working with tiny corp 00:28:20 dumb questions, andrew tate 00:29:00 cache coherency, order of computation 00:30:15 no more edge tpu 00:30:40 messy first clean later 00:35:00 cache aware, order 00:35:30 theoretical maximum 00:37:15 Auto-Vectorization in LLVM, doing by hand 00:42:35 50$ for max or fma 00:42:53 llvmlite 00:46:00 llvm unsafe-fp-math 00:52:10 enable-unsafe-fp-math llvmlite 00:59:40 not voting 01:01:18 llvmlite create_module_pass_manager 01:02:13 add_instruction_combining_pass 01:03:40 llvm 11 download 01:07:00 InstructionMix llvm 01:07:25 somewhere in there 01:09:00 llvm 14 01:11:30 ffp-contract clang argument 01:15:25 auto vectorising 01:17:05 raining outside in San Diego, crazy emails from people 01:18:25 tbaa llvm 01:23:16 not ir problem 01:34:10 brew install llvm 11 01:34:57 aarch-64-use-aa 01:35:59 llvm 11 configuration not correct 01:36:50 don't like llvmlite 01:37:30 polyhedral loop optimizer 01:41:05 debug only 01:44:02 not a debug build 01:52:25 numba output llvm bytec