oneDNN VS asmjit

If you are talking about non-small matrix multiplication in MKL, is now in opensource as a part of oneDNN. It literally has exactly the same code, as in MKL (you can see this by inspecting constants or doing high-precision benchmarks).

For small matmul there is libxsmm. It may take tremendous efforts make something faster than oneDNN and libxsmm, as jit-based approach of https://github.com/oneapi-src/oneDNN/blob/main/src/gpu/jit/g... is too flexible: if someone finds a better sequence, oneDNN can reuse it without major change of design.

But MKL is not limited to matmul, I understand it...

For completeness, it looks like this question was posted to the oneDNN GitHub repo and the response was to stay tune for updates.

From the release note of oneDNN v2.5:

Intel are trying to support their oneAPI compute framework on Arm and IBM POWER and z/Architecture (s390x) but since they ever released only a single discrete GPU with the Xe architecture it's unclear whether they'll support Xe GPU compute on e.g. ARM https://github.com/oneapi-src/oneDNN

Some other instruction sets in some JSON: https://github.com/asmjit/asmjit/tree/master/db

The attentive reader may notice that this code is from a third-party library. So, we didn't want to include it in the article at first. However, we found something interesting. In 2017, somebody opened an issue in the asmjit project: the GCC 7.2 compiler issued a warning to the code above. The project authors fixed it:

You can use something like https://asmjit.com/ to generate and call x64 code at runtime.

But an easy to create a JIT would be to use https://github.com/asmjit/asmjit, which is used in RPCS3.

You could go the way of https://asmjit.com (or forth) and make it your assembler DSL on top of the low-level call.

But if you're willing to get closer to the hardware is https://github.com/bitfunnel/nativejit/ and https://asmjit.com/

https://github.com/asmjit/asmjit is nice. But using a JIT seems like a sledgehammer working around a lacking design.