Dear Intel OpenCL team,
Recently I was with a report on SGEMM broken with 1.6.2 on Intel ARC · Issue #533 · CNugteren/CLBlast · GitHub regarding the wrong results with SGEMM with CLBLAST.
CLBLAST is a fast opensource for blas (faster than clBLAS and usable like cuBLAS). I was the volunteer to help CLBLAST to tune different devices for CLBLAST. It was the first time we encountered a report from llama.cpp LLM community where their contributor witnessed an error after we tuned the SGEMM. However, when I tried to run the sgemm test cases with all the latest three versions (1.5.3, 1.6.1, 1.6.2) of CLBLAST on my own A770, it seemed that this GPU was reporting wrong results compared to the results from openBLAS or clBLAS where version 1.5.3 has no tuning for A770. It has no errors with my Xe based laptop. The tests were based on the latest Feb. 26 driver.
It would be great if you help me to have a look at it since the original CLBLAST author does not have an Intel GPU (I have one ARC 770 16GB).
Compile with test cases: CLBlast/doc/testing.md at master · CNugteren/CLBlast · GitHub
Also, you may wish to pay attention to the DLL files if you are trying to compile the test on Windows (Tests don't run on Intel Xe/ARC GPU · Issue #534 · CNugteren/CLBlast · GitHub).
I have attached two exe for your reference.
Thank you very much!
Best wishes,
Jinchuan Tang
