[HiderX]

Summary

kernel/riscv64/sum_vector.c is the S/DSUM kernel of RISCV64_ZVL256B. In its unrolled loop for incx > 1 it never
advanced ix, and it tripled inc_xv on every iteration:

                                j += gvl * 2;
-                               inc_xv += inc_xv * 2;
+                               ix += inc_xv * 2;

For incx > 1 and n >= 2*vl, this summed the first block again on every iteration, plus elements further and further
past the end of x:

• with incx = 2 or 3 at n = 128 and 129, cblas_dsum returned wrong results, e.g. −2.15 instead of 3.55;
• longer vectors read outside the mapping, and cblas_ssum(1000, x, 2) and cblas_dsum(1000, x, 2) crashed with
  SIGSEGV.

The fix advances ix by two blocks, as zsum_vector.c already does. No other target uses this file: C910V and GENERIC
use ../arm/sum.c, and ZVL128B uses sum_rvv.c.

Testing

All runs were on a SpacemiT K3 (X100 core), GCC 15.2, RISCV64_ZVL256B.

• OpenBLAS test suite (make TARGET=RISCV64_ZVL256B NO_FORTRAN=1 NO_LAPACK=1 NO_LAPACKE=1 USE_OPENMP=0 NUM_THREADS=8): identical to develop, with utest 141/141, extension tests 1479/1479 and every BLAS/CBLAS test PASSED.

• A level-1 sweep through CBLAS, for all four precisions. It covers copy, swap, scal, axpy, axpby, rot, dot,
  asum, sum, i?amax and i?amin:

  • n from 0 to 1031, and increment pairs including negative ones;
  • each vector sits in a buffer with guard zones and sentinel gaps, which are checked after every call;
  • results are compared bitwise where exact, otherwise against a binary128 reference.

  	crash	failures
  develop	?SUM, n = 1000, incx = 2	?SUM, incx > 1, n ≥ 2·vl
  this PR	none	0 of 6486 cases

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.