target/arm: Implement bfloat16 dot product (indexed)

This is BFDOT for both AArch64 AdvSIMD and SVE,
and VDOT.BF16 for AArch32 NEON.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210525225817.400336-8-richard.henderson@linaro.org
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>

7 files changed, 80 insertions, 9 deletions

diff --git a/target/arm/helper.h b/target/arm/helper.h
index de2f5331dc..376c1cef0f 100644
--- a/target/arm/helper.h
+++ b/target/arm/helper.h
@@ -1004,6 +1004,8 @@ DEF_HELPER_FLAGS_5(gvec_usmmla_b, TCG_CALL_NO_RWG,
 
 DEF_HELPER_FLAGS_5(gvec_bfdot, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_5(gvec_bfdot_idx, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
 
 #ifdef TARGET_AARCH64
 #include "helper-a64.h"
diff --git a/target/arm/neon-shared.decode b/target/arm/neon-shared.decode
index 31a0839bbb..fa3cf14e3a 100644
--- a/target/arm/neon-shared.decode
+++ b/target/arm/neon-shared.decode
@@ -81,6 +81,8 @@ VUSDOT_scalar  1111 1110 1 . 00 .... .... 1101 . q:1 index:1 0 vm:4 \
                vn=%vn_dp vd=%vd_dp
 VSUDOT_scalar  1111 1110 1 . 00 .... .... 1101 . q:1 index:1 1 vm:4 \
                vn=%vn_dp vd=%vd_dp
+VDOT_b16_scal  1111 1110 0 . 00 .... .... 1101 . q:1 index:1 0 vm:4 \
+               vn=%vn_dp vd=%vd_dp
 
 %vfml_scalar_q0_rm 0:3 5:1
 %vfml_scalar_q1_index 5:1 3:1
diff --git a/target/arm/sve.decode b/target/arm/sve.decode
index a7429b293f..51f87e8937 100644
--- a/target/arm/sve.decode
+++ b/target/arm/sve.decode
@@ -1633,3 +1633,6 @@ FMLALB_zzxw     01100100 10 1 ..... 0100.0 ..... .....     @rrxr_3a esz=2
 FMLALT_zzxw     01100100 10 1 ..... 0100.1 ..... .....     @rrxr_3a esz=2
 FMLSLB_zzxw     01100100 10 1 ..... 0110.0 ..... .....     @rrxr_3a esz=2
 FMLSLT_zzxw     01100100 10 1 ..... 0110.1 ..... .....     @rrxr_3a esz=2
+
+### SVE2 floating-point bfloat16 dot-product (indexed)
+BFDOT_zzxz      01100100 01 1 ..... 010000 ..... .....     @rrxr_2 esz=2
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 3c36de3df2..71de75e568 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -13449,8 +13449,22 @@ static void disas_simd_indexed(DisasContext *s, uint32_t insn)
             return;
         }
         break;
-    case 0x0f: /* SUDOT, USDOT */
-        if (is_scalar || (size & 1) || !dc_isar_feature(aa64_i8mm, s)) {
+    case 0x0f:
+        switch (size) {
+        case 0: /* SUDOT */
+        case 2: /* USDOT */
+            if (is_scalar || !dc_isar_feature(aa64_i8mm, s)) {
+                unallocated_encoding(s);
+                return;
+            }
+            break;
+        case 1: /* BFDOT */
+            if (is_scalar || !dc_isar_feature(aa64_bf16, s)) {
+                unallocated_encoding(s);
+                return;
+            }
+            break;
+        default:
             unallocated_encoding(s);
             return;
         }
@@ -13570,13 +13584,22 @@ static void disas_simd_indexed(DisasContext *s, uint32_t insn)
                          u ? gen_helper_gvec_udot_idx_b
                          : gen_helper_gvec_sdot_idx_b);
         return;
-    case 0x0f: /* SUDOT, USDOT */
-        gen_gvec_op4_ool(s, is_q, rd, rn, rm, rd, index,
-                         extract32(insn, 23, 1)
-                         ? gen_helper_gvec_usdot_idx_b
-                         : gen_helper_gvec_sudot_idx_b);
-        return;
-
+    case 0x0f:
+        switch (extract32(insn, 22, 2)) {
+        case 0: /* SUDOT */
+            gen_gvec_op4_ool(s, is_q, rd, rn, rm, rd, index,
+                             gen_helper_gvec_sudot_idx_b);
+            return;
+        case 1: /* BFDOT */
+            gen_gvec_op4_ool(s, is_q, rd, rn, rm, rd, index,
+                             gen_helper_gvec_bfdot_idx);
+            return;
+        case 2: /* USDOT */
+            gen_gvec_op4_ool(s, is_q, rd, rn, rm, rd, index,
+                             gen_helper_gvec_usdot_idx_b);
+            return;
+        }
+        g_assert_not_reached();
     case 0x11: /* FCMLA #0 */
     case 0x13: /* FCMLA #90 */
     case 0x15: /* FCMLA #180 */
diff --git a/target/arm/translate-neon.c b/target/arm/translate-neon.c
index 9460857b2a..8099767792 100644
--- a/target/arm/translate-neon.c
+++ b/target/arm/translate-neon.c
@@ -390,6 +390,15 @@ static bool trans_VSUDOT_scalar(DisasContext *s, arg_VSUDOT_scalar *a)
                         gen_helper_gvec_sudot_idx_b);
 }
 
+static bool trans_VDOT_b16_scal(DisasContext *s, arg_VDOT_b16_scal *a)
+{
+    if (!dc_isar_feature(aa32_bf16, s)) {
+        return false;
+    }
+    return do_neon_ddda(s, a->q * 6, a->vd, a->vn, a->vm, a->index,
+                        gen_helper_gvec_bfdot_idx);
+}
+
 static bool trans_VFML_scalar(DisasContext *s, arg_VFML_scalar *a)
 {
     int opr_sz;
diff --git a/target/arm/translate-sve.c b/target/arm/translate-sve.c
index ed290827ad..6f02030635 100644
--- a/target/arm/translate-sve.c
+++ b/target/arm/translate-sve.c
@@ -8665,3 +8665,15 @@ static bool trans_BFDOT_zzzz(DisasContext *s, arg_rrrr_esz *a)
     }
     return true;
 }
+
+static bool trans_BFDOT_zzxz(DisasContext *s, arg_rrxr_esz *a)
+{
+    if (!dc_isar_feature(aa64_sve_bf16, s)) {
+        return false;
+    }
+    if (sve_access_check(s)) {
+        gen_gvec_ool_zzzz(s, gen_helper_gvec_bfdot_idx,
+                          a->rd, a->rn, a->rm, a->ra, a->index);
+    }
+    return true;
+}
diff --git a/target/arm/vec_helper.c b/target/arm/vec_helper.c
index 7eefcd06ea..74a497f38c 100644
--- a/target/arm/vec_helper.c
+++ b/target/arm/vec_helper.c
@@ -2452,3 +2452,23 @@ void HELPER(gvec_bfdot)(void *vd, void *vn, void *vm, void *va, uint32_t desc)
     }
     clear_tail(d, opr_sz, simd_maxsz(desc));
 }
+
+void HELPER(gvec_bfdot_idx)(void *vd, void *vn, void *vm,
+                            void *va, uint32_t desc)
+{
+    intptr_t i, j, opr_sz = simd_oprsz(desc);
+    intptr_t index = simd_data(desc);
+    intptr_t elements = opr_sz / 4;
+    intptr_t eltspersegment = MIN(16 / 4, elements);
+    float32 *d = vd, *a = va;
+    uint32_t *n = vn, *m = vm;
+
+    for (i = 0; i < elements; i += eltspersegment) {
+        uint32_t m_idx = m[i + H4(index)];
+
+        for (j = i; j < i + eltspersegment; j++) {
+            d[j] = bfdotadd(a[j], n[j], m_idx);
+        }
+    }
+    clear_tail(d, opr_sz, simd_maxsz(desc));
+}