When eigenvalues are nearly degenerate, numerical accumulation in
the divide-and-conquer merge can push w[i] slightly above zero.
The subsequent sqrt(-w[i]) then produces NaN, which propagates
through the entire eigenvector output of DSYEVD/DSTEDC.
Use fmax(-w[i], 0.0) to safely clamp the sqrt argument, yielding
zero instead of NaN for spuriously positive w[i]. The fmax call
maps to a single vmaxsd instruction on x86_64 with zero overhead.
Add test/test_laed3_nan.c which exercises DSYEVD on a rank-1
perturbed identity matrix designed to create near-degenerate
eigenvalue clusters.
common_interface.h declares slamc3 as returning FLOATRET when
NEED_F2CCONV is enabled, but laed3_single.c and laed3_parallel.c
redeclared LAMC3 as returning FLOAT. This causes conflicting-type
errors in MinGW builds.
Use FLOATRET for the local LAMC3 prototype so it matches the shared
declaration. Also undefine the Windows max macro before the local
max definition in laed3_parallel.c to avoid macro redefinition
warnings.
Added HFLOAT16 support for RISCV64
Added shgemm_kernel_8x8 for RISCV64_ZVL128B and shgemm_kernel_16x8 for RISCV64_ZVL256B based on HFLOAT16
The instruction sets used are ZVFH and ZFH, which need to be supported by RVV1.0
Related to issue #5279
Co-authored-by Linjin Li <linjin_li@163.com>
Using `OpenMP::OpenMP_LANG` targets for CMake is less error-prone than
passing the compiler and linker flags manually. Furthermore, it allows
the user to customize those flags by setting `OpenMP_LANG_FLAGS`,
`OpenMP_LANG_LIB_NAMES`, and `OpenMP_omp_LIBRARY`.
Implement DYNAMIC_ARCH support for riscv64. Three cpu types are
supported, riscv64_generic, riscv64_zvl256b, riscv64_zvl128b.
The two non-generic kernels require CPU support for RVV 1.0 to
function correctly. Detecting that a riscv64 device supports
RVV 1.0 is a little complicated as there are some boards on the
market that advertise support for V via hwcap but only support
RVV 0.7.1, which is not binary compatible with RVV 1.0. The
approach taken is to first try hwprobe. If hwprobe is not
available, we fall back to hwcap + an additional check to distinguish
between RVV 1.0 and RVV 0.7.1.
Tested on a VM with VLEN=256, a CanMV K230 with VLEN=128 (with only
the big core enabled), a Lichee Pi with RVV 0.7.1 and a VF2 with no
vector.
A compiler with RVV 1.0 support must be used to build OpenBLAS for
riscv64 when DYNAMIC_ARCH=1.
Signed-off-by: Mark Ryan <markdryan@rivosinc.com>
Previously dynamic builds were either using the default SWITCH_RATIO
or one from the higher level architecture; this patch ensures the
dynamic builds can use this parameter as well.
Martin Kroeker