On Tue, Sep 2, 2025 at 22:38 Richard Henderson <richard.hender...@linaro.org>
wrote:
> On 9/1/25 23:38, Max Chou wrote:
> > +#define OPMVV_VQDOTQ(NAME, TD, T1, T2, TX1, TX2, HD, HS1, HS2)
> \
> > +static void do_##NAME(void *vd, void *vs1, void *vs2, int i)
> \
> > +{
> \
> > + int idx;
> \
> > + T1 r1;
> \
> > + T2 r2;
> \
> > + TX1 *r1_buf = (TX1 *)vs1 + HD(i);
> \
> > + TX2 *r2_buf = (TX2 *)vs2 + HD(i);
> \
> > + TD acc = *((TD *)vd + HD(i));
> \
> > + int64_t partial_sum = 0;
> \
>
> I think it's clear partial_sum should be the 32-bit type TD.
> Indeed, I'm not sure why you don't just have
>
> TD acc = ((TD *)vd)[HD(i)];
Thanks for the suggestion. I’ll update version 2 for this part.
>
> > +
> \
> > + for (idx = 0; idx < 4; ++idx) {
> \
> > + r1 = *((T1 *)r1_buf + HS1(idx));
> \
> > + r2 = *((T2 *)r2_buf + HS2(idx));
> \
> > + partial_sum += (r1 * r2);
> \
>
> acc += r1 * r2;
>
> > + }
> \
> > + *((TD *)vd + HD(i)) = (acc + partial_sum) & MAKE_64BIT_MASK(0, 32);
> \
>
> ((TD *)vd)[HD(i)] = acc;
>
> because that final mask is bogus.
>
The partial_sum and the final mask are created to ensure the behavior
described in the Zvqdotq isa spec section 3 as follows:
“Finally, the four products are accumulated into the corresponding element
of vd, wrapping around signed overflow.”
Thanks,
Max.
