[HiderX]

Summary

On RISCV64_ZVL256B, cblas_ismin/cblas_idmin (and ISMIN/IDMIN) return a wrong
index for many inputs longer than one vector strip. imin_vector.c has two bugs
in its RVV 1.0 intrinsics branch:

1. The merge operand is missing. The masked vid/vadd instructions that
   record each lane's running-minimum index use the mask-agnostic
   __riscv_vid_v_*_m / __riscv_vadd_vx_*_m forms, which take no merge
   operand. A lane whose minimum was found in an earlier strip therefore loses
   its index.

   • This was introduced by the v1.0 intrinsics port (2406958).
   • In that same commit, imax_vector.c got the mask-undisturbed _mu forms.
   • The RISCV_0p10_INTRINSICS branch of imin_vector.c still passes
     v_min_index as the merge operand.

   The patch uses the _mu forms, with v_min_index as the merge operand.

2. The tail comparison is inverted. The tail checks cur_minf > minf. That
   can never be true, because cur_minf is reduced from minf, so a minimum in
   the last partial strip is ignored. The check is now cur_minf < minf, as in
   iamin_vector.c.

Both the unit-stride and the strided paths are fixed. Not affected:

• C910V, which uses the 0.10 intrinsics branch;
• the targets that use imin_rvv.c.

Testing

Platform: SpacemiT K3, one X100 core, TARGET=RISCV64_ZVL256B, GCC 15.2.

• Targeted check of i?amax, i?amin, i?max, i?min (S and D):

  • strides 1, 2 and 3;
  • 47 lengths from 1 to 32775, covering strip boundaries and tails;
  • the extreme is planted at strip boundaries, in the tail, at a random
    position, and as a tie;
  • 9624 cases in total.

  Results:

  • develop (5bab605): 1321 mismatches, all in ISMIN/IDMIN. For example,
    ISMIN with n = 65 returns 6 instead of 64.
  • 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. The patched kernel runs exactly as fast as its sibling
  imax_vector.c. At n = 1024, ISMIN/IDMIN take 472/821 ns per call, and
  ISMAX/IDMAX take 473/821 ns. The old kernel was faster on in-cache sizes
  (257/387 ns) only because dropping the merge operand also removed the
  loop-carried dependency on the index vector, at the cost of losing the
  indices. At n = 4M the patched kernel is about 3 % faster.

This bug was found while comparing the _vector and _rvv level-1 kernels on
the X100.

Disclosure: an AI coding agent (Anthropic Claude) prepared this patch, its
validation runs and this description on behalf of the account owner. All
numbers above come from archived runs on the board.