OpenMathLib/OpenBLAS - CodSpeed

OpenBLAS

Blog Docs Changelog

Performance History

Latest Results

Improve f_check when assembly has no vendor identity Vendor detection in f_check greps the probe assembly (ftest.s) for a compiler identity string. Some compilers, notably Intel ifx, emit no such string, so the primary `case "$data"` scan fails to set a vendor. The name-based fallback only recognizes bare compiler names, so an ifx-backed wrapper (e.g. mpif90, mpifort or the Cray ftn wrapper) is not matched either, leaving the vendor empty and resulting in a false conclusion that no Fortran compilers are available. Append the output of `$compiler --version` to $data so such compilers are identified by their version banner (ifx's contains "ifx", "Intel").

jczhang07:fix-fortran-vendor-detection

46 minutes ago

CORTEXA72: fix DTRSM LN rem indexing for non-PoT UNROLL_M Rem panels of size 4/5 used (m & ~(i-1))-i, which assumes power-of-two M alignment. With MR=6 that misplaces the rem-4 block (cblas N=5/35). Walk the rem start position from the end instead. Co-authored-by: Cursor <cursoragent@cursor.com>

hugomeiland:cortexa72-l3-fix

4 hours ago

CORTEXA72: set DGEMM_UNROLL_MN=LCM(6,8)=24 for SYRK panels SYRK/SYR2K diagonal kernels advance packed A/B by UNROLL_MN and require panel-aligned starts for both MR and NR. MN=8 fixed NR packing but broke MR=6 A panels; MN=24 is the common multiple. Keeps non-PoT division stepping. Co-authored-by: Cursor <cursoragent@cursor.com>

hugomeiland:cortexa72-l3-fix

13 hours ago

syr2k: use division for UNROLL_MN stepping when MN is non-PoT CORTEXA72 sets DGEMM_UNROLL_MN=6 so SYRK diagonals match MR=6 packs. The ~(MN-1) form only works for power-of-two MN; use the same division as syrk_kernel.c. Co-authored-by: Cursor <cursoragent@cursor.com>

hugomeiland:cortexa72-l3-fix

14 hours ago

wasm: use relaxed-simd madd for v_muladd when available When emcc is invoked with -mrelaxed-simd, __wasm_relaxed_simd__ is defined and wasm_f{32x4,64x2}_relaxed_madd intrinsics are available. These lower to hardware FMA on engines that support relaxed-simd (Chrome >=114, Firefox >=120, Safari >=16.4, Node >=22) and to separate mul+add otherwise, so callers get FMA where possible with no ABI change. Off by default: builds not passing -mrelaxed-simd fall through to the existing separate mul+add definitions.

MMesch:wasm-relaxed-simd-madd

16 hours ago

Merge pull request #5996 from MMesch/wasm-kernel-ifndef-guards wasm: guard remaining KERNEL assignments so KERNEL.$(TARGET_CORE) overrides survive

develop

3 days ago

Merge pull request #5998 from OpenMathLib/revert-5885-lapack1317 Revert "Increase deflation tolerance in ?LAED to match recent changes in ?SLASD" (Reference-LAPACK PR 1372)

develop

3 days ago

Merge pull request #5999 from martin-frbg/ci-graviton-ubuntuports CI: Work around Graviton jobs failing due to an unreliable package server

develop

3 days ago

Latest Branches

0%

Improve f_check when assembly has no vendor identity#6003

10 hours ago

db09161

jczhang07:fix-fortran-vendor-detection

0%

Fix CORTEXA72 Level-3 correctness (SYMM/TRMM/SYRK)#6002

5 hours ago

c8a6e79

hugomeiland:cortexa72-l3-fix

0%

wasm: use relaxed-simd madd for v_muladd when available#6001

4 days ago

f85a214

MMesch:wasm-relaxed-simd-madd

© 2026 CodSpeed Technology

Home Terms Privacy Docs