target/arm: Implement integer matrix multiply accumulate

This is {S,U,US}MMLA for both AArch64 AdvSIMD and SVE,
and V{S,U,US}MMLA.S8 for AArch32 NEON.

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210525010358.152808-91-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>

1 files changed, 77 insertions, 0 deletions

diff --git a/target/arm/vec_helper.c b/target/arm/vec_helper.c
index 19c4ba1bdf..e84b438340 100644
--- a/target/arm/vec_helper.c
+++ b/target/arm/vec_helper.c
@@ -2335,3 +2335,80 @@ void HELPER(gvec_xar_d)(void *vd, void *vn, void *vm, uint32_t desc)
     }
     clear_tail(d, opr_sz * 8, simd_maxsz(desc));
 }
+
+/*
+ * Integer matrix-multiply accumulate
+ */
+
+static uint32_t do_smmla_b(uint32_t sum, void *vn, void *vm)
+{
+    int8_t *n = vn, *m = vm;
+
+    for (intptr_t k = 0; k < 8; ++k) {
+        sum += n[H1(k)] * m[H1(k)];
+    }
+    return sum;
+}
+
+static uint32_t do_ummla_b(uint32_t sum, void *vn, void *vm)
+{
+    uint8_t *n = vn, *m = vm;
+
+    for (intptr_t k = 0; k < 8; ++k) {
+        sum += n[H1(k)] * m[H1(k)];
+    }
+    return sum;
+}
+
+static uint32_t do_usmmla_b(uint32_t sum, void *vn, void *vm)
+{
+    uint8_t *n = vn;
+    int8_t *m = vm;
+
+    for (intptr_t k = 0; k < 8; ++k) {
+        sum += n[H1(k)] * m[H1(k)];
+    }
+    return sum;
+}
+
+static void do_mmla_b(void *vd, void *vn, void *vm, void *va, uint32_t desc,
+                      uint32_t (*inner_loop)(uint32_t, void *, void *))
+{
+    intptr_t seg, opr_sz = simd_oprsz(desc);
+
+    for (seg = 0; seg < opr_sz; seg += 16) {
+        uint32_t *d = vd + seg;
+        uint32_t *a = va + seg;
+        uint32_t sum0, sum1, sum2, sum3;
+
+        /*
+         * Process the entire segment at once, writing back the
+         * results only after we've consumed all of the inputs.
+         *
+         * Key to indicies by column:
+         *          i   j                  i             j
+         */
+        sum0 = a[H4(0 + 0)];
+        sum0 = inner_loop(sum0, vn + seg + 0, vm + seg + 0);
+        sum1 = a[H4(0 + 1)];
+        sum1 = inner_loop(sum1, vn + seg + 0, vm + seg + 8);
+        sum2 = a[H4(2 + 0)];
+        sum2 = inner_loop(sum2, vn + seg + 8, vm + seg + 0);
+        sum3 = a[H4(2 + 1)];
+        sum3 = inner_loop(sum3, vn + seg + 8, vm + seg + 8);
+
+        d[H4(0)] = sum0;
+        d[H4(1)] = sum1;
+        d[H4(2)] = sum2;
+        d[H4(3)] = sum3;
+    }
+    clear_tail(vd, opr_sz, simd_maxsz(desc));
+}
+
+#define DO_MMLA_B(NAME, INNER) \
+    void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc) \
+    { do_mmla_b(vd, vn, vm, va, desc, INNER); }
+
+DO_MMLA_B(gvec_smmla_b, do_smmla_b)
+DO_MMLA_B(gvec_ummla_b, do_ummla_b)
+DO_MMLA_B(gvec_usmmla_b, do_usmmla_b)