[HiderX]

Summary

The RVV index kernels iamax_rvv.c, iamin_rvv.c, imax_rvv.c,
imin_rvv.c, izamax_rvv.c and izamin_rvv.c do not always return the first
occurrence of the extreme value when that value appears more than once, which
BLAS requires. These kernels are used by RISCV64_ZVL128B and x280.

How the kernels work: each vector lane keeps the index of the first extreme
value it has seen. After the loop, they pick the lowest lane that holds the
overall extreme, using vfirst, and return that lane's index after a
vslidedown. When the extreme value is duplicated, a lower lane can hold a
later index. For example, with n = 65 and equal maxima at positions 21 and 64,
cblas_isamax returns 64.

The patch instead returns the smallest index recorded by the lanes that hold
the extreme value: it sets every other lane to all-ones with vmerge and then
reduces with vredminu. The search loops are unchanged.

Testing

Platform: SpacemiT K3 (X100, VLEN 256), TARGET=RISCV64_ZVL128B, GCC 15.2.

• Targeted check of i?amax, i?amin, i?max, i?min and of
  icamax/izamax/icamin/izamin, 15312 cases in total:

  • S, D, C and Z;
  • strides 1, 2 and 3;
  • 47 lengths from 1 to 32775;
  • the extreme is planted at strip boundaries, in the tail, at a random
    position, or twice (a tie).

  Results:

  • develop (5bab605): 552 mismatches, all ties;
  • patched: 0 mismatches.

• OpenBLAS tests. The default make goal's tests give identical results
  before and after the patch: utest 141/141, extension tests 1479/1479, and all
  BLAS/CBLAS tests PASSED.

• Performance is unchanged. ISAMAX/IDAMAX at n = 1024, 32768 and 4M all
  stay within ±0.35 %.

RISCV64_ZVL256B uses the *_vector.c index kernels, which already return the
first occurrence. Its ISMIN/IDMIN bug is fixed separately in #6169.