tinygrad + rusticl + aco: why not?

I recently came across tinygrad as a small powerful nn framework that had an OpenCL backend target and could run LLaMA model.

I've been looking out for rusticl workloads, and this seemed like a good one, and I could jump on the AI train, and run an LLM in my house!

I started it going on my Radeon 6700XT with the latest rusticl using radeonsi with the LLVM backend, and I could slowly interrogate a model with a question, and it would respond. I've no idea how performant it is vs ROCm yet which seems to be where tinygrad is more directed, but I may get to that next week.

While I was there though I decided to give the Mesa ACO compiler backend a go, it's been tied into radeonsi recently, and I done some hacks before to get compute kernels to run. I reproduced said hacks on the modern code and gave it a run.

tinygrad comes with a benchmark script called benchmark_train_efficientnet so I started playing with it to see what low hanging fruit I could find in an LLVM vs ACO shootout.

The bench does 10 runs, the first is where lots of compilation happens, the last is well primed cache wise. There are the figures from the first and last runs with a release build of llvm and mesa. (and the ACO hacks).

LLVM:

215.78 ms cpy,  12245.04 ms run,  120.33 ms build, 12019.45 ms realize,  105.26 ms CL,   -0.12 loss,  421 tensors, 0.04 GB used,      0.94 GFLOPS

10.25 ms cpy,   221.02 ms run,   83.50 ms build,   36.25 ms realize,  101.27 ms CL,   -0.01 loss,  421 tensors, 0.04 GB used,     52.11 GFLOPS

ACO:

71.10 ms cpy,  3443.04 ms run,  112.58 ms build, 3214.13 ms realize,  116.34 ms CL,   -0.04 loss,  421 tensors, 0.04 GB used,      3.35 GFLOPS
10.36 ms cpy,   234.90 ms run,   84.84 ms build,   36.51 ms realize,  113.54 ms CL,    0.05 loss,  421 tensors, 0.04 GB used,     49.03 GFLOPS

So ACO is about 4 times faster to compile but produces binaries that are less optimised.

The benchmark produces 148 shaders:

LLVM:

126 Max Waves: 16 
  6 Max Waves: 10
  5 Max Waves: 9
  6 Max Waves: 8
  5 Max Waves: 4


ACO:

 96 Max Waves: 16
 36 Max Waves: 12
  2 Max Waves: 10
 10 Max Waves: 8
  4 Max Waves: 4

So ACO doesn't quite get the optimal shaders for a bunch of paths, even with some local hackery I've done to make it do better.[1]

I'll investigate ROCm next week maybe, got a bit of a cold/flu, and large GPU stacks usually make me want to wipe the machine after I test them :-P

[1] https://gitlab.freedesktop.org/airlied/mesa/-/commits/radeonsi-rusticl-aco-wip


Comments

Post a Comment

Popular posts from this blog

On Rust, Linux, developers, maintainers

There's been a couple of mentions of Rust4Linux in the past week or two, one from Linus on the speed of engagement and one about Wedson departing the project due to non-technical concerns. This got me thinking about project phases and developer types. Archetypes: I will regret making an analogy, in an area I have no experience in, but let's give it a go with a road building analogy.   Let's sort developers into 3 rough categories. Let's preface by saying not all developers fit in a single category throughout their careers, and some developers can do different roles on different projects, or on the same project simultaneously. 1. Wayfinders/Mapmakers I want to go build a hotel somewhere but there exists no map or path. I need to travel through a bunch of mountains, valleys, rivers, weather, animals, friendly humans, antagonistic humans and some unknowns. I don't care deeply about them, I want to make a path to where I want to go. I hit a roadblock, I don't focus...
Read more

radv: vulkan av1 video decode status

The Khronos Group announced VK_KHR_video_decode_av1 [1], this extension adds AV1 decoding to the Vulkan specification. There is a radv branch [2] and merge request [3]. I did some AV1 work on this in the past, but I need to take some time to see if it has made any progress since. I'll post an ANV update once I figure that out. This extension is one of the ones I've been wanting for a long time, since having royalty-free codec is something I can actually care about and ship, as opposed to the painful ones. I started working on a MESA extension for this a year or so ago with Lynne from the ffmpeg project and we made great progress with it. We submitted that to Khronos and it has gone through the committee process and been refined and validated amongst the hardware vendors. I'd like to say thanks to Charlie Turner and Igalia for taking over a lot of the porting to the Khronos extension and fixing up bugs that their CTS development brought up. This is a great feature of having ...
Read more

anv: vulkan av1 decode status

 Vulkan Video AV1 decode has been released, and I had some partly working support on Intel ANV driver previously, but I let it lapse. The branch is currently [1]. It builds, but is totally untested, I'll get some time next week to plug in my DG2 and see if I can persuade it to decode some frames. Update: the current branch decodes one frame properly, reference frames need more work unfortunately. [1] https://gitlab.freedesktop.org/airlied/mesa/-/commits/anv-vulkan-video-decode-av1
Read more