On Wed, 6 Jul 2022 at 10:17, Richard Henderson
<[email protected]> wrote:
>
> Signed-off-by: Richard Henderson <[email protected]>
> +void HELPER(sme_addha_s)(void *vzda, void *vzn, void *vpn,
> + void *vpm, uint32_t desc)
> +{
> + intptr_t row, col, oprsz = simd_oprsz(desc) / 4;
> + uint64_t *pn = vpn, *pm = vpm;
> + uint32_t *zda = vzda, *zn = vzn;
> +
> + for (row = 0; row < oprsz; ) {
> + uint64_t pa = pn[row >> 4];
> + do {
> + if (pa & 1) {
> + for (col = 0; col < oprsz; ) {
> + uint64_t pb = pm[col >> 4];
> + do {
> + if (pb & 1) {
> + zda[tile_vslice_index(row) + col] += zn[col];
Doesn't this need some H macros to handle the bigendian case?
We process the predicate from architecturally least to most
significant, but because zda is a uint32_t* we don't handle
that in the same order I think. Similarly sme_addva_s().
> + }
> + pb >>= 4;
> + } while (++col & 15);
> + }
> + }
> + pa >>= 4;
> + } while (++row & 15);
> + }
> +}
Otherwise
Reviewed-by: Peter Maydell <[email protected]>
thanks
-- PMM
