triton

I was expecting DirectX in triton-lang https://github.com/triton-lang/triton

https://github.com/triton-lang/triton https://github.com/tenstorrent/tt-mlir https://github.com/onnx/onnx-mlir https://github.com/openxla/stablehlo plenty more - just google

Perhaps any new language targetting GPU acceleration would consider TILE based concept and primitive recently supported by major GPU vendors including Nvidia [1],[2],[3],[4]. For more generic GPU targets there's TRITON [5],[6]. [1] NVIDIA CUDA 13.1 Powers Next-Gen GPU Programming with NVIDIA CUDA Tile and Performance Gains: https://developer.nvidia.com/blog/nvidia-cuda-13-1-powers-ne... [2] Nvidia Tilus: A Tile-Level GPU Kernel Programming Language: https://github.com/NVIDIA/tilus [3] Simplify GPU Programming with NVIDIA CUDA Tile in Python: https://developer.nvidia.com/blog/simplify-gpu-programming-w... [4] Tile Language: https://github.com/tile-ai/tilelang [5] Triton: An Intermediate Language and Compiler for Tiled Neural Network Computations: https://dl.acm.org/doi/10.1145/3315508.3329973 [6] Triton: https://github.com/triton-lang/trit

Here is a recent example that lack of archived discussion causes problem: https://github.com/triton-lang/triton/issues/9539

https://github.com/triton-lang/triton/pull/7298#discussion_r... > By disassembly of ptxas, it is indeed hard-coded that they have logic like: strstr(kernel_name, "cutlass"). > it is likely that, this is an unstable, experimental, aggressive optimization by NVIDIA, and blindly always enabling it may produce some elusive bugs.

It sounds like they share that goal. Gluon is a thing because the Triton team realized over the last few months that Blackwell is a significant departure from the Hopper, and achieving >80% SoL kernels is becoming intractable as the triton middle-end simply can't keep up. Some more info in this issue: https://github.com/triton-lang/triton/issues/7392

My experience searching for "nvidia triton" coughed up oppressive number of results for the similarly named inference server but I think the Triton discussed here is https://github.com/triton-lang/triton (although the commits that I spot checked were all from @openai.com emails, not nvidia)

AMD already has Composable Kernels[1], and supports for example Triton[2]. Then there is also HIP[3], and there are tools to automatically convert from from CUDA to HIP. But since CUDA is the de-facto standard, there is always friction to use something else (unless you need to support also AMD stack). Making something just CUDA-compatible is non-trivial, and since Nvidia decides its direction and new features then the alternatives would always be lagging behind. Currently there are also major hardware differences between Nvidia and AMD, which may make highly optimized CUDA code inefficient or even buggy. [1] https://github.com/ROCm/composable_kernel?tab=readme-ov-file#composable-kernel [2] https://github.com/triton-lang/triton?tab=readme-ov-file#triton [3] https://github.com/ROCm/HIP?tab=readme-ov-file#what-is-this-repository-for

https://github.com/triton-lang/triton/pull/7298/commits/a5e2... It's literally in the code.

actual link: https://github.com/triton-lang/triton/pull/7298