The misdetection affects double precision on Jammy, not only sbgemm. On an Intel Xeon w9-3575X (Sapphire Rapids workstation; avx512f, avx512_bf16, amx_tile), Ubuntu 22.04.5, libopenblas0-pthread 0.3.20+ds-1, R 4.6.1 linked to `/usr/lib/x86_64-linux-gnu/openblas-pthread/`:
- `OPENBLAS_VERBOSE=2` reports `Core: Cooperlake`. - With that kernel, LAPACK results are wrong: `solve()` already for matrices of order 12 and 20 (max |A^-1 A - I| of 3.1 and 1.6 on random SPD matrices; up to 19 at order 40), `chol()` from order 40 (fails on well-conditioned SPD matrices), and `eigen()` (reconstruction error about 3e4 at order 200). - `OPENBLAS_CORETYPE=SkylakeX` or `Haswell` gives errors of about 1e-16 on the same matrices; `OPENBLAS_CORETYPE=Cooperlake` reproduces the error. - `OPENBLAS_NUM_THREADS=1` does not change the result. Reproduction (R; the matrix is the AR(1) correlation 0.5^|i-j|, condition number about 9): ```r A <- 0.5^abs(outer(1:64, 1:64, "-")) max(abs(crossprod(chol(A)) - A)) # 0.25 (Cooperlake), 1e-16 (SkylakeX) B <- 0.5^abs(outer(1:20, 1:20, "-")) max(abs(solve(B) %*% B - diag(20))) # 0.5 (Cooperlake), 1e-16 (SkylakeX) ``` ```sh OPENBLAS_VERBOSE=2 Rscript -e 'A <- 0.5^abs(outer(1:64,1:64,"-")); max(abs(crossprod(chol(A))-A))' OPENBLAS_VERBOSE=2 OPENBLAS_CORETYPE=SkylakeX Rscript -e 'A <- 0.5^abs(outer(1:64,1:64,"-")); max(abs(crossprod(chol(A))-A))' ``` The results are wrong without any warning, so any numerical software on Jammy with this CPU (R, NumPy linked to the system OpenBLAS, ...) can return incorrect results. Upstream reworked the Cooperlake / Sapphire Rapids runtime selection in 0.3.22 and 0.3.24 (Changelog: "fixed runtime selection of COOPERLAKE in DYNAMIC_ARCH builds"; "fixed runtime detection of Cooperlake and Sapphire Rapids in DYNAMIC_ARCH"). -- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/2040027 Title: [SPR][ICX] OpenBLAS: Issues while testing sbgemm for Sapphire Rapids (SPR) To manage notifications about this bug go to: https://bugs.launchpad.net/intel/+bug/2040027/+subscriptions -- ubuntu-bugs mailing list [email protected] https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
