AI Performance @ AMD
Ex Varjo, Remedy, UX3D, Grimlore Games
GitHub: https://github.com/rAzoR8
Twitter: https://twitter.com/singul4rity
We're coming back to Breda for this year's edition of the Graphics Programming Conference, November 18th-20th. Submissions are open: https://graphicsprogrammingconference.com/participate/
If you want to present, now is the time to get your proposal done! Looking forward to seeing you all there again!
After years of continuous development, we just tagged a new release of D3D12 Memory Allocator library v3.0.0 🎉
🏷️ So if you want the latest and greatest D3D12MA, or just more info - head over to GPUOpen!
This was a great read. https://cookieplmonster.github.io/2025/04/23/gta-san-andreas-win11-24h2-bug/
We're doing the thing!
🔼 Triangles may have won the hardware-acceleration rendering battle ... 🔼
⏹️ ... but the method in our latest blog enables emulation of quadrilateral rasterization, through bilinear interpolation of quadrilateral vertices! ⏹️
See how it squares up: https://gpuopen.com/learn/bilinear-interpolation-quadrilateral-barycentric-coordinates/?utm_source=mastodon&utm_medium=social&utm_campaign=quad
I also tried to use them to control the kernel group size. Different devices prefer different sizes, so it would have been useful to be able to tweak this without shipping multiple spir-v variants.
That however was a bit more problematic. This is supposed to be a spir-v 1.2 feature, but in practice only works when targeting spirv 1.6.
I filed a bug report that as usual didn't get much attention:
@aras new consensus method: proof-of-boilerplate
New updates to the Decompilation Wiki by harpend (on GitHub). We have a new in-depth Switch structuring section and a new Loop Reduction section.
https://decompilation.wiki/fundamentals/structuring/schema-based/switch-structuring/
Ported https://salykova.github.io/sgemm-gpu to Vulkan (nice article!) 🐢
it's 2x slower than Cuda. That one was tricky to port, (e.g. need to alias shared buffer to allow LDS/STS.128), half of it with AI, 2nd half going over lines 1-by-1
Ported the Kernel 5 from https://seb-v.github.io/optimization/update/2025/01/20/Fast-GPU-Matrix-multiplication.html to Vulkan 🚀
It's only 15% slower than Cuda (and it works on AMD)
In both cases, it's quite difficult to reason about SPIR-V ➡️ ISA, apart the AMD GPU Analyzer that is helping!
Digging deeper 🏗️
Another Tool for Checking Library Level API and ABI Compatibility
http://github.com/isuckatcs/abicorn-on-graduation-ceremony
Discussions: https://discu.eu/q/http://github.com/isuckatcs/abicorn-on-graduation-ceremony
Amd tldr:
$549 9070 16Gb
$599 9070 XT 16Gb
Claimed 2x RT uplift, OBBs for bhv? Some boring ai stuff. Better compute units.
LLVM 20 Promotes SPIR-V To Official Backend, Enabled By Default
https://www.phoronix.com/news/LLVM-20-SPIR-V-Official-Target
#cpp
#cplusplus
Awesome and very deep exploration of fp32 GEMMs on RDNA3: https://seb-v.github.io/optimization/update/2025/01/20/Fast-GPU-Matrix-multiplication.html
New release of #romusage 1.3.1
- Experimental GBDK NES
- Allow leading area underscores for .noi/.map
- Fix: Brief mode header area multi-counting
Win/Linux/Mac console app that measures free ROM & RAM space for #GameBoy games. Works with #gbdk, #rgbds, #gbstudio & ROM files
Hacker News, but for game dev:
https://gamedev.city/
The new DirectX 12 Agility SDK 1.615.0 and 1.716.0-preview adds some exciting features! I plan to write more about it on my blog soon. D3D12 Memory Allocator library and D3d12info tool will need an update.
https://devblogs.microsoft.com/directx/agilitysdk-1-716-0-preview-and-1-615-retail/
#directx12 #dx12
More seriously, though, if anyone follows me and works for or has contacts at NVIDIA and can help me actually get one this week so I can test NVK on it, that'd be pretty cool.
A #WebAssembly compiler that fits in a tweet: https://wasmgroundup.com/blog/wasm-compiler-in-a-tweet/. This was a fun read on both #Wasm and JavaScript code golfing ⛳.
Graphics Programming weekly - Issue 375 - January 19th, 2025 https://www.jendrikillner.com/post/graphics-programming-weekly-issue-375/
"NESFab is a new programming language for creating NES games. Designed with 8-bit limitations in mind, the language is more ergonomic to use than C, while also producing faster assembly code."
