OpenBLAS commit 632ef87 (0.3.34.dev)
produces incorrect double-precision results on a real Raspberry Pi 400 when
built for TARGET=CORTEXA72.
Hardware: Raspberry Pi 400, Cortex-A72 r0p3, 4 CPUs
OS: Debian arm64, Linux 6.18.39+rpt-rpi-v8
Compiler: GCC 14.2.0-19
Build settings:
TARGET=CORTEXA72 DYNAMIC_ARCH=0 NO_SHARED=1 NO_LAPACK=1 ONLY_CBLAS=1
BUILD_DOUBLE=1, with OPENBLAS_NUM_THREADS=1
The Level-3 checks fail consistently:
DSYMM max_abs_error=57.47919143876338
DTRMM max_abs_error=1.9381688466111771
DSYRK max_abs_error=39.17479191438764
DSYR2K max_abs_error=75.115338882283027
failures=4
The generic ARMV8 and Cortex-A57 targets pass. The previous Cortex-A72
version also passed.
The first bad commit is:
560caa5
Add Cortex-A72 DGEMM 6x8 microkernel with MR=6 packers and TRSM
Its parent is good:
05518e0
Merge pull request #5969 from martin-frbg/lapack134
OpenBLAS commit 632ef87 (0.3.34.dev)
produces incorrect double-precision results on a real Raspberry Pi 400 when
built for TARGET=CORTEXA72.
Hardware: Raspberry Pi 400, Cortex-A72 r0p3, 4 CPUs
OS: Debian arm64, Linux 6.18.39+rpt-rpi-v8
Compiler: GCC 14.2.0-19
Build settings:
TARGET=CORTEXA72 DYNAMIC_ARCH=0 NO_SHARED=1 NO_LAPACK=1 ONLY_CBLAS=1
BUILD_DOUBLE=1, with OPENBLAS_NUM_THREADS=1
The Level-3 checks fail consistently:
DSYMM max_abs_error=57.47919143876338
DTRMM max_abs_error=1.9381688466111771
DSYRK max_abs_error=39.17479191438764
DSYR2K max_abs_error=75.115338882283027
failures=4
The generic ARMV8 and Cortex-A57 targets pass. The previous Cortex-A72
version also passed.
The first bad commit is:
560caa5
Add Cortex-A72 DGEMM 6x8 microkernel with MR=6 packers and TRSM
Its parent is good:
05518e0
Merge pull request #5969 from martin-frbg/lapack134