[kevinmoran1988]

A small one I noticed while reading the DP matrix code. It doesn't change results, and I couldn't measure a speed difference, so there's no urgency here.

p7_omx_GrowTo_avx() and p7_omx_GrowTo_dpf_avx() decide whether a matrix is already wide enough by comparing the model length to allocQ4_avx * 4:

if (ox->allocQ4_avx * 4 >= allocM && ox->validR > allocL && ox->allocXR >= allocXL + 1)
  return eslOK;
...
if (allocM > ox->allocQ4_avx * 4) reset_row_pointers = TRUE;

That 4 is the SSE vector width; an AVX stripe holds 8 floats. For M = 300, p7O_NQF_AVX(300) is 38 stripes, which hold 304 positions, but the check computes 38 × 4 = 152 and concludes the matrix is too narrow. So for any model longer than 8 the early return never fires, and every call lays out all the row pointers again.

Nothing goes wrong because of it: the reset path is correct, and the DP routines' own guards already use 8 (om->M > ox->allocQ4_avx*8 in fwdback_avx.c and fwdback_fs_avx.c). This PR just changes the four * 4 to * 8 so the grow functions agree with them. The ncells bookkeeping is left alone; it uses the same unit on both sides of its comparison.

What I checked, with and without the change, on an AVX2 machine:

• Output is identical in default, --fs and --splice modes (23 models against 14 Mb), and in default and --fs on a larger search (200 models against a 352 Mb genome).
• Timing differences were within run-to-run noise (about ±3%) in every case, including 523 short single-thread searches.

🤖 Generated with Claude Code