target/arm: Implement bfloat16 dot product (vector)

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: 20210525225817.400336-7-richard.henderson@linaro.org
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>

target/arm/helper.h

@@ -1002,6 +1002,9 @@ DEF_HELPER_FLAGS_5(gvec_ummla_b, TCG_CALL_NO_RWG,
 DEF_HELPER_FLAGS_5(gvec_usmmla_b, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, ptr, i32)
 
+DEF_HELPER_FLAGS_5(gvec_bfdot, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
+
 #ifdef TARGET_AARCH64
 #include "helper-a64.h"
 #include "helper-sve.h"

target/arm/neon-shared.decode

@@ -52,6 +52,8 @@ VUDOT          1111 110 00 . 10 .... .... 1101 . q:1 . 1 .... \
                vm=%vm_dp vn=%vn_dp vd=%vd_dp
 VUSDOT         1111 110 01 . 10 .... .... 1101 . q:1 . 0 .... \
                vm=%vm_dp vn=%vn_dp vd=%vd_dp
+VDOT_b16       1111 110 00 . 00 .... .... 1101 . q:1 . 0 .... \
+               vm=%vm_dp vn=%vn_dp vd=%vd_dp
 
 # VFM[AS]L
 VFML           1111 110 0 s:1 . 10 .... .... 1000 . 0 . 1 .... \

target/arm/sve.decode

@@ -1625,6 +1625,9 @@ FMLALT_zzzw     01100100 10 1 ..... 10 0 00 1 ..... .....  @rda_rn_rm_e0
 FMLSLB_zzzw     01100100 10 1 ..... 10 1 00 0 ..... .....  @rda_rn_rm_e0
 FMLSLT_zzzw     01100100 10 1 ..... 10 1 00 1 ..... .....  @rda_rn_rm_e0
 
+### SVE2 floating-point bfloat16 dot-product
+BFDOT_zzzz      01100100 01 1 ..... 10 0 00 0 ..... .....  @rda_rn_rm_e0
+
 ### SVE2 floating-point multiply-add long (indexed)
 FMLALB_zzxw     01100100 10 1 ..... 0100.0 ..... .....     @rrxr_3a esz=2
 FMLALT_zzxw     01100100 10 1 ..... 0100.1 ..... .....     @rrxr_3a esz=2

target/arm/translate-a64.c

@@ -12235,6 +12235,16 @@ static void disas_simd_three_reg_same_extra(DisasContext *s, uint32_t insn)
         }
         feature = dc_isar_feature(aa64_fcma, s);
         break;
+    case 0x1f: /* BFDOT */
+        switch (size) {
+        case 1:
+            feature = dc_isar_feature(aa64_bf16, s);
+            break;
+        default:
+            unallocated_encoding(s);
+            return;
+        }
+        break;
     default:
         unallocated_encoding(s);
         return;
@@ -12318,6 +12328,16 @@ static void disas_simd_three_reg_same_extra(DisasContext *s, uint32_t insn)
         }
         return;
 
+    case 0xf: /* BFDOT */
+        switch (size) {
+        case 1:
+            gen_gvec_op4_ool(s, is_q, rd, rn, rm, rd, 0, gen_helper_gvec_bfdot);
+            break;
+        default:
+            g_assert_not_reached();
+        }
+        return;
+
     default:
         g_assert_not_reached();
     }

target/arm/translate-neon.c

@@ -296,6 +296,15 @@ static bool trans_VUSDOT(DisasContext *s, arg_VUSDOT *a)
                         gen_helper_gvec_usdot_b);
 }
 
+static bool trans_VDOT_b16(DisasContext *s, arg_VDOT_b16 *a)
+{
+    if (!dc_isar_feature(aa32_bf16, s)) {
+        return false;
+    }
+    return do_neon_ddda(s, a->q * 7, a->vd, a->vn, a->vm, 0,
+                        gen_helper_gvec_bfdot);
+}
+
 static bool trans_VFML(DisasContext *s, arg_VFML *a)
 {
     int opr_sz;

target/arm/translate-sve.c

@@ -8653,3 +8653,15 @@ static bool trans_UMMLA(DisasContext *s, arg_rrrr_esz *a)
 {
     return do_i8mm_zzzz_ool(s, a, gen_helper_gvec_ummla_b, 0);
 }
+
+static bool trans_BFDOT_zzzz(DisasContext *s, arg_rrrr_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,
+                          a->rd, a->rn, a->rm, a->ra, 0);
+    }
+    return true;
+}

target/arm/vec_helper.c

@@ -2412,3 +2412,43 @@ static void do_mmla_b(void *vd, void *vn, void *vm, void *va, uint32_t desc,
 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)
+
+/*
+ * BFloat16 Dot Product
+ */
+
+static float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2)
+{
+    /* FPCR is ignored for BFDOT and BFMMLA. */
+    float_status bf_status = {
+        .tininess_before_rounding = float_tininess_before_rounding,
+        .float_rounding_mode = float_round_to_odd_inf,
+        .flush_to_zero = true,
+        .flush_inputs_to_zero = true,
+        .default_nan_mode = true,
+    };
+    float32 t1, t2;
+
+    /*
+     * Extract each BFloat16 from the element pair, and shift
+     * them such that they become float32.
+     */
+    t1 = float32_mul(e1 << 16, e2 << 16, &bf_status);
+    t2 = float32_mul(e1 & 0xffff0000u, e2 & 0xffff0000u, &bf_status);
+    t1 = float32_add(t1, t2, &bf_status);
+    t1 = float32_add(sum, t1, &bf_status);
+
+    return t1;
+}
+
+void HELPER(gvec_bfdot)(void *vd, void *vn, void *vm, void *va, uint32_t desc)
+{
+    intptr_t i, opr_sz = simd_oprsz(desc);
+    float32 *d = vd, *a = va;
+    uint32_t *n = vn, *m = vm;
+
+    for (i = 0; i < opr_sz / 4; ++i) {
+        d[i] = bfdotadd(a[i], n[i], m[i]);
+    }
+    clear_tail(d, opr_sz, simd_maxsz(desc));
+}