target/arm: Implement FMOPA, FMOPS (widening)

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20220708151540.18136-27-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>

1 files changed, 74 insertions, 0 deletions

diff --git a/target/arm/sme_helper.c b/target/arm/sme_helper.c
index 690a53eee2..302f89c30b 100644
--- a/target/arm/sme_helper.c
+++ b/target/arm/sme_helper.c
@@ -1008,6 +1008,80 @@ static inline uint32_t f16mop_adj_pair(uint32_t pair, uint32_t pg, uint32_t neg)
     return pair;
 }
 
+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.
+     */
+    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) * 0x80008000u;
+    uint16_t *pn = vpn, *pm = vpm;
+    float_status fpst_odd, fpst_std;
+
+    /*
+     * Make a copy of float_status because this operation does not
+     * update the cumulative fp exception status.  It also produces
+     * default nans.  Make a second copy with round-to-odd -- see above.
+     */
+    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 prow = pn[H2(row >> 4)];
+        do {
+            void *vza_row = vza + tile_vslice_offset(row);
+            uint32_t n = *(uint32_t *)(vzn + H1_4(row));
+
+            n = f16mop_adj_pair(n, prow, neg);
+
+            for (col = 0; col < oprsz; ) {
+                uint16_t pcol = pm[H2(col >> 4)];
+                do {
+                    if (prow & pcol & 0b0101) {
+                        uint32_t *a = vza_row + H1_4(col);
+                        uint32_t m = *(uint32_t *)(vzm + H1_4(col));
+
+                        m = f16mop_adj_pair(m, pcol, 0);
+                        *a = f16_dotadd(*a, n, m, &fpst_std, &fpst_odd);
+
+                        col += 4;
+                        pcol >>= 4;
+                    }
+                } while (col & 15);
+            }
+            row += 4;
+            prow >>= 4;
+        } while (row & 15);
+    }
+}
+
 void HELPER(sme_bfmopa)(void *vza, void *vzn, void *vzm, void *vpn,
                         void *vpm, uint32_t desc)
 {