wonnx VS cutlass

The two I know of are IREE and Kompute[1]. I'm not sure how much momentum the latter has, I don't see it referenced much. There's also a growing body of work that uses Vulkan indirectly through WebGPU. This is currently lagging in performance due to lack of subgroups and cooperative matrix mult, but I see that gap closing. There I think wonnx[2] has the most momentum, but I am aware of other efforts.

[1]: https://kompute.cc/

[2]: https://github.com/webonnx/wonnx

To a first approximation, Kompute[1] is that. It doesn't seem to be catching on, I'm seeing more buzz around WebGPU solutions, including wonnx[2] and more hand-rolled approaches, and IREE[3], the latter of which has a Vulkan back-end.

[1]: https://kompute.cc/

[2]: https://github.com/webonnx/wonnx

[3]: https://github.com/openxla/iree

There's also a third-party WebGPU implementation: https://github.com/webonnx/wonnx

By experimental I meant e.g. using WGPU to run compute shaders like wonnx, which is working fine but only on a very restricted set of devices and browsers.

https://news.ycombinator.com/item?id=35696031 ... TIL about wonnx: https://github.com/webonnx/wonnx#in-the-browser-using-webgpu...

microsoft/onnxruntime: https://github.com/microsoft/onnxruntime

Apache/arrow has language-portable Tensors for cpp: https://arrow.apache.org/docs/cpp/api/tensor.html and rust: https://docs.rs/arrow/latest/arrow/tensor/struct.Tensor.html and Python: https://arrow.apache.org/docs/python/api/tables.html#tensors https://arrow.apache.org/docs/python/generated/pyarrow.Tenso...

Fwiw it looks like the llama.cpp Tensor is from ggml, for which there are CUDA and OpenCL implementations (but not yet ROCm, or a WebGPU shim for use with emscripten transpilation to WASM): https://github.com/ggerganov/llama.cpp/blob/master/ggml.h

Are the recommendable ways to cast e.g. arrow Tensors to pytorch/tensorflow?

FWIU, Rust has a better compilation to WASM; and that's probably faster than already-compiled-to-JS/ES TensorFlow + WebGPU.

What's a fair benchmark?

wonnx has done some fantastic work in this regard, so that's where we plan to start once we get there. In terms of general discussion of alternate backends, see this issue.

> GPU in other ways, such as training ML models and then using them via an inference engine all powered by your local GPU?

Have a look at wonnix https://github.com/webonnx/wonnx

A WebGPU-accelerated ONNX inference run-time written 100% in Rust, ready for native and the web

Looking forward to your WebGPU ML runtime! Also, why not contribute back to WONNX? (https://github.com/webonnx/wonnx)

You can indeed perform inference using WebGPU (see e.g. [1] for GPU-accelerated inference of ONNX models on WebGPU; I am one of the authors).

The point made above is that WebGPU can only be used for GPU's and not really for other types of 'neural accelerators' (like e.g. the ANE on Apple devices).

[1] https://github.com/webonnx/wonnx

I would recommend the course from Oxford (https://people.maths.ox.ac.uk/gilesm/cuda/). Also explore the tutorial section of cutlass (https://github.com/NVIDIA/cutlass/blob/main/media/docs/cute/...) if you want to learn more about high performance gemm.

The latter (and I guess you were asking about this one) is designed to accelerate NN inference in reduced precision. It is possible to use Tensor Cores for you own purposes, mainly through CUTLASS. But because Tensor Cores are designed to execute matrix multiplications, it can be hard to adapt your problem to them. The performance with them is insane (IIRC 32x the performance of the INT32 pipeline), but only for matrix multiplication…

If you want to learn about controlling Tensor Cores, the main way is through the CUTLASS library, that wraps the complexity of Tensor Cores into higher level abstractions. You can also look for mma/wmma instructions in the PTX specification, or for the WMMA API in CUDA.

> we used tensor cores and managed to get back fp32 accuracy with 3 rounds of the things

Hey are you referring to 3xTF32 (https://github.com/NVIDIA/cutlass/tree/master/examples/28_am...)? IMO this is a perfect example where proper abstraction could save engineers non-trivial amount of time - imagine a compiler stack which allows 3xTF32 as a normal dtype and subsequent analysis compatible with this special dtype :-)

This is a great post for people who are new to optimizing GPU code.

It is interesting to see that the author got this far without interchanging the innermost loop over k to the outermost loop, as is done in CUTLASS (https://github.com/NVIDIA/cutlass).

As you can see in this blog post the code ends up with a lot of compile-time constants (e.g. BLOCKSIZE, BM, BN, BK, TM, TN) one way to optimize this code further is to use an auto-tuner to find the optimal value for all of these parameters for your GPU and problem size, for example Kernel Tuner (https://github.com/KernelTuner/kernel_tuner)

Some examples here have a benchmark: https://github.com/NVIDIA/cutlass/blob/master/examples/24_gemm_grouped/gemm_grouped.cu

I would like to make a minimum CMakeLists to use the CUDA CUTLASS library in another project. The build system is CMake, however I have little experience with CMake.

If you already know some C++, the Nvidia devblog is a great resource. Going further, Cub and Cutlass provide examples of efficient implementations for key operations at all hardware levels. Finally, this is more anecdotal but I always start my lectures on Cuda programming with the pictures in this doc page, to provide some intuition on the different memory layers that you can leverage to speed up a program. In any case, good luck :-)