Multiplatform Matrix Multiplication Kernels

Very interesting, willing to try burn

One of the author here, don't hesitate if you have any question or comment!

I'm sorry this is a low brow comment but this is the dumbest thing you can do in this space:

> Unit (thread in CUDA, invocation in Vulkan/Wgpu): the smallest execution entity performing computations.

> Plane (warp in CUDA, subgroup in Vulkan/Wgpu): a group of (typically 32) units executing in lockstep and able to share data efficiently through registers.

> Cube (thread block in CUDA, workgroup in Vulkan/Wgpu): a group of units that execute on the same SM, sharing memory and able to synchronize

It's already bad enough that the vendors themselves insisted on different names but why in the bejesus would you rename these concepts and diverge from literally all existing naming conventions when you're providing middleware. Ie when using your tool I'm still going to reference NVIDIA's or AMD's docs to understand how the hardware actually works. Like do you really think otherwise - that your thing is gonna be end of the line???

FYI the word warp isn't random techno babble but is actually a very clever pun that actually fits very well conceptually:

https://en.m.wikipedia.org/wiki/Warp_and_weft

Using the naming from one of the existing API would put too much bias towards that API. It started as a WebGPU project early on, but some features are not present so mixing terms wasn't ideal. We're also working on extending CubeCL to CPU, so we want terms not only tied to the GPU word.

Thread, group, workgroup.

There you go you've hit basically two of 3 completely (AMD and Vulkan) and are close enough to CUDA that people would get it.

I have no idea what a plane connotes and a cube literally gives a distinct enough picture from block that I will be continuously reminding myself of the mapping.

What you did was pointless - you assigned new words to objects that you don't own and now your conceptual framework is askew from the actual underlying (true) conceptual framework.

> CubeCL to CPU

There is zero affinity between GPU programing models and multicore CPU programing models. If you don't believe me go ask the OpenMP people how they're doing supporting GPUs.

burn is awesome

Why unit instead of point?

Unit, plane (as vs train), and cube?

Or point, plane, cube (1d, 2d, 3d)?

Well we can agree to disagree, CubeCL also has the concept of instruction parallelism, which would be used to target simd instructions on CPU. Our algorithms are normally flexible on both the plane size and the line size, adapting to the hardware with comptime logique. You are free to dislike the naming, but imo a mix of multiple APIs is worse than something new.

> Our algorithms are normally flexible on both the plane size and the line size

Congrats - I have no idea what this means lol.

I don't recall the reason why, point is a valid name.

Has there been much research into slightly flawed matrix multiplications?

If you have a measure of correctness, and a measure of performance. Is there a maximum value of correctness per some unit of processing that exists below a full matrix multiply

Obviously it can be done with precision, since that is what floating point is. But is there anything where you can save x% of computation and have fewer than x% incorrect values in a matrix multiplications?

Gradient descent wouldn't really care about a few (Reliably) dud values.

Could something like this be done in WebGPU?

For people who are interested Kokkos (a C++ library for writing portable kernels) also has a naming scheme for hierarchical parallelism. They use ThreadTeam, Thread (for individual threads within a group), and ThreadVector (for per thread SIMD).

Just commenting to share, personally I have no naming preference but the hierarchal abstractions in general are incredibly useful.

Randomized matrix sketching is one way to get at this (see https://arxiv.org/abs/2302.11474), the problem is hardware is heavily optimized for dense multiplies so what you save in flops doesn't translate to real runtime speeds ups.

It will make more sense once you start using CubeCL. There's now a CubeCL book available: https://burn.dev/books/cubecl/.

It does come with some mental overhead, but let’s be honest, there’s no objectively “good” choice here without introducing bias toward a specific vendor API.

Learning the core concepts takes effort, but if CubeCL is useful for your work, it’s definitely worth it.

Actually, points are zero dimensional, lines are one dimensional.

I had bet that matmult would be in transformer-optimized hardware costing a fraction of GPUs first class in torch 2 years ago with no reason to use GPUs any more. Wrong.

> matmult would be in transformer-optimized hardware

It is... it's in GPUs lol

> first class in torch

It is

> costing a fraction of GPUs

Why would anyone give you this for cheaper than GPUs lol?

CubeCL supports WebGPU and can be used with wasm!

I think they’re referring to hardware like TPUs and other ASICs. Which also exist, of course :)

Sure but GPUs literally have MMA engines now

GPUs came about because of the need for faster float 4x4 and 3x3 matrix, and 3 and 4 vector math ops like multiply, multiply-accumulate, and such, and faster pushing of pixels with things like texture mapping. All hail OpenGL and dual Voodoo2 SLI. ;)

Reminds me of ye olden days when kernel transforms were merely weighted multiplicative and/or additive matrixes applied to every point in the source arriving at pixel data in the target. Blur, sharpen, color channel filter, color swap, invert, etc. An extremely diagonalizable problem suitable for massive parallelism and concurrent calculation because there is little/no dependency on prior calculations.