Build extremely slow for large GEMM mapping (M=N=K=2048, m=n=k=32)

[ec-935]

Hi,

When running tests/dataflow/aie/test_mapping_large_gemm.py with the following parameters:

M, N, K = 2048, 2048, 2048
m, n, k = 32, 32, 32

the build process takes an extremely long time, though it does eventually complete with no errors.

For example:

• At dtype=int8, the build took over 2 hours.

• At dtype=int16 and bf16, it runs even longer.

Is there any way to reduce the build time or optimize the generation for such large problem sizes?
I’m wondering whether this is expected due to code generation complexity or if there are recommended flags or configuration changes that could help.

Thank you!

[Fangtangtang]

Larger sizes take longer to compile, that’s expected. There’s currently no flag to reduce the build time.
From the matrix parameter settings, it’s better to avoid very large M//m * N//n * K//n. If you don’t need extremely large MNK, I think it’s not necessary to test cases with such large settings. Using larger m, n, k to better utilize data memory usually gives better performance.

At dtype=int8, the build took over 2 hours.

the efficiency shouldn’t be affected this much. I’ll check.