In 0.7.1 RVV implementation multiplication had performed unnecessary operations unpacking and then packing values back again. Removed this part. Performance on LicheePi 4A with Xuantie 2.8.0 toolchain have increased 5.82 ms -> 1.06 ms (1920x1080 / CV_8UC1 - BinaryOpTest.multiply/20). Accuracy tests for core and imgproc pass with the same failures as before the fix.

vuint16m2_t res = vwmulu_vv_u16m2(a, b, 16);

// following calls are not needed - they unpack values and pack them back again
vuint16m1_t c = vget_v_u16m2_u16m1(res, 0);
vuint16m1_t d = vget_v_u16m2_u16m1(res, 1);
vuint16m2_t im = vundefined_u16m2();
im = vset_v_u16m1_u16m2(im, 0, c);
im = vset_v_u16m1_u16m2(im, 1, d);
// emd - we can pass 'res' directly to 'vnclipu'

vuint8m1_t fin = vnclipu_wx_u8m1(im, 0, 16);

Note: RVV-scalable implementation already uses shortened code