Re: [PATCH v5 26/45] target/arm: Implement FMOPA, FMOPS (widening)

[Peter Maydell]

On Wed, 6 Jul 2022 at 10:26, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>


> +static float32 f16_dotadd(float32 sum, uint32_t e1, uint32_t e2,
> +                          float_status *s_std, float_status *s_odd)
> +{
> +    float64 e1r = float16_to_float64(e1 & 0xffff, true, s_std);
> +    float64 e1c = float16_to_float64(e1 >> 16, true, s_std);
> +    float64 e2r = float16_to_float64(e2 & 0xffff, true, s_std);
> +    float64 e2c = float16_to_float64(e2 >> 16, true, s_std);
> +    float64 t64;
> +    float32 t32;
> +
> +    /*
> +     * The ARM pseudocode function FPDot performs both multiplies
> +     * and the add with a single rounding operation.  Emulate this
> +     * by performing the first multiply in round-to-odd, then doing
> +     * the second multiply as fused multiply-add, and rounding to
> +     * float32 all in one step.
> +     */

I guess if we find we're not producing quite bit-accurate results
we can come back and revisit this :-)

> +    t64 = float64_mul(e1r, e2r, s_odd);
> +    t64 = float64r32_muladd(e1c, e2c, t64, 0, s_std);
> +
> +    /* This conversion is exact, because we've already rounded. */
> +    t32 = float64_to_float32(t64, s_std);
> +
> +    /* The final accumulation step is not fused. */
> +    return float32_add(sum, t32, s_std);
> +}
> +
> +void HELPER(sme_fmopa_h)(void *vza, void *vzn, void *vzm, void *vpn,
> +                         void *vpm, void *vst, uint32_t desc)
> +{
> +    intptr_t row, col, oprsz = simd_maxsz(desc);
> +    uint32_t neg = simd_data(desc) << 15;
> +    uint16_t *pn = vpn, *pm = vpm;
> +    float_status fpst_odd, fpst_std = *(float_status *)vst;
> +
> +    set_default_nan_mode(true, &fpst_std);
> +    fpst_odd = fpst_std;
> +    set_float_rounding_mode(float_round_to_odd, &fpst_odd);
> +
> +    for (row = 0; row < oprsz; ) {
> +        uint16_t pa = pn[H2(row >> 4)];
> +        do {
> +            void *vza_row = vza + tile_vslice_offset(row);
> +            uint32_t n = *(uint32_t *)(vzn + row);

More missing H macros.

> +
> +            n = f16mop_adj_pair(n, pa, neg);
> +
> +            for (col = 0; col < oprsz; ) {
> +                uint16_t pb = pm[H2(col >> 4)];
> +                do {
> +                    if ((pa & 0b0101) == 0b0101 || (pb & 0b0101) == 0b0101) {

Wrong condition again?

> +                        uint32_t *a = vza_row + col;
> +                        uint32_t m = *(uint32_t *)(vzm + col);
> +
> +                        m = f16mop_adj_pair(m, pb, neg);
> +                        *a = f16_dotadd(*a, n, m, &fpst_std, &fpst_odd);
> +
> +                        col += 4;
> +                        pb >>= 4;
> +                    }
> +                } while (col & 15);
> +            }
> +            row += 4;
> +            pa >>= 4;
> +        } while (row & 15);
> +    }
> +}

thanks
-- PMM