Source
- URL: https://x.com/real_redp/status/2091228915142472141
- Author: red plait (@real_redp)
- Posted: 2026-08-22 18:20:01
Branch
1/ @real_redp
@AgileJebrim
ptx, my ass: https://redplait.blogspot.com/2026/08/parser-of-ptx-instructions.html
2/ @AgileJebrim
@real_redp
FWIW I target SPIR-V in part because it’s a binary format (ASCII asm sucks) and in part because it’s more portable than PTX.
3/ @AgileJebrim
@real_redp
The main advantage of PTX is that it gets you direct access to a lower level API into NVIDIA hardware than anything else available unless we manage to reverse engineer PSO binary caches (something I’m still debating attempting) to write directly in SASS.
4/ @real_redp
@AgileJebrim
he-he
I made tool to patch sass
5/ @AgileJebrim
@real_redp
Which generation? Is it available?
6/ @AgileJebrim
@real_redp
I see you have a disassembler. It looks to be CUDA-specific? I’d love to see you do one for Vulkan PSOs too, especially for Ampere.
https://github.com/redplait/denvdis/tree/master
7/ @AgileJebrim
@real_redp
Actually, I think this path might work as is combined with the Cubin extension in Vulkan.
“sed-like tool for inline patching of sass instructions within cubin files.”
Lots of possibilities! 🤔
How mature is this?
8/ @real_redp
@AgileJebrim
enough to do crazy things like https://redplait.blogspot.com/2026/07/optimization-of-sass-stall-counts.html