Open davidcl opened 3 months ago
Curious, the current DGEMM kernel for Haswell has been in use for years. Have you only started testing/using OpenBLAS recently, or did something else (like the default compiler version or options) change in your setup lately ?
We were compiling without DYNAMIC_TARGET thus selecting NEHALEM arch as the minimal supported target. All the opened bugs are due to a more active cross-checking of Windows vs Linux results on our test base after the GCC 8 to GCC 11 Linux migration.
Note: all these corner-case bugs might or might not be valid at some points; eg. in that case, EPSILON might be acceptable in most codebases.
Symbolic result is e-7 off, so both code paths do very well at rounding artifacts
After upgrading Scilab OpenBLAS build, we detected a poor precision result impacting
dgemm
when using AVX2 kernels.With OpenBLAS 0.3.27:
With refBLAS 3.12.0 :
test code
```c #include