Re: [PATCH v1 08/11] target/arm: Implement bfloat16 matrix multiply accumulate

[Richard Henderson]

On 5/18/21 7:37 AM, Peter Maydell wrote:
> On Sat, 17 Apr 2021 at 01:00, Richard Henderson
> <richard.henderson@linaro.org> wrote:
> > This is BFMMLA for both AArch64 AdvSIMD and SVE,
> > and VMMLA.BF16 for AArch32 NEON.
> >
> > Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>
> > +void HELPER(gvec_bfmmla)(void *vd, void *vn, void *vm, void *va, uint32_t desc)
> > +{
> > +    intptr_t s, opr_sz = simd_oprsz(desc);
> > +    float32 *d = vd, *a = va;
> > +    uint32_t *n = vn, *m = vm;
> > +
> > +    for (s = 0; s < opr_sz / 4; s += 4) {
> > +        float32 sum00, sum01, sum10, sum11;
> > +
> > +        /*
> > +         * Process the entire segment at once, writing back the
> > +         * results only after we've consumed all of the inputs.
> > +         *
> > +         * Key to indicies by column:
>
> "indices"
>
> > +         *               i   j           i   k             j   k
> > +         */
> > +        sum00 = a[s + H4(0 + 0)];
> > +        sum00 = bfdotadd(sum00, n[s + H4(0 + 0)], m[s + H4(0 + 0)]);
> > +        sum00 = bfdotadd(sum00, n[s + H4(0 + 1)], m[s + H4(0 + 1)]);
>
> I can't make these indices match up with the arm arm pseudocode ones,
> which index by "4*i + 2*k + 0" and "4*i + 2*k + 1", not "2*i + k";
> are we hiding a division by 2 somewhere?

Yes.  We're passing BFloat16 pairs via uint32_t[] to bfdotadd().


r~