Add support for the Zvfbdota32f FP32 batched dot-product instruction.
The vfbdota.vv instruction performs 8 parallel dot products between
FP32 vectors, accumulating into FP32 elements.

Signed-off-by: Max Chou <[email protected]>
---
 target/riscv/helper.h                       |  3 ++
 target/riscv/insn32.decode                  |  3 ++
 target/riscv/tcg/insn_trans/trans_rvv.c.inc | 11 ++++
 target/riscv/tcg/vector_helper.c            | 58 +++++++++++++++++++++
 4 files changed, 75 insertions(+)

diff --git a/target/riscv/helper.h b/target/riscv/helper.h
index a257436fb31..94f4d9c40c6 100644
--- a/target/riscv/helper.h
+++ b/target/riscv/helper.h
@@ -1398,6 +1398,9 @@ DEF_HELPER_6(vfwdota_vv, void, ptr, ptr, ptr, ptr, env, 
i32)
 /* Zvfwbdota16bf - BF16 Batched Dot Product */
 DEF_HELPER_7(vfwbdota_vv, void, ptr, ptr, ptr, ptr, i32, env, i32)
 
+/* Zvfbdota32f - FP32 Batched Dot Product */
+DEF_HELPER_7(vfbdota_vv, void, ptr, ptr, ptr, ptr, i32, env, i32)
+
 /* Zvfqwbdota8f - OCP FP8 Batched Dot Product */
 DEF_HELPER_7(vfqwbdota_vv, void, ptr, ptr, ptr, ptr, i32, env, i32)
 DEF_HELPER_7(vfqwbdota_alt_vv, void, ptr, ptr, ptr, ptr, i32, env, i32)
diff --git a/target/riscv/insn32.decode b/target/riscv/insn32.decode
index 48597915f7b..7fdb7eb54da 100644
--- a/target/riscv/insn32.decode
+++ b/target/riscv/insn32.decode
@@ -1112,6 +1112,9 @@ vfwbdota_vv 101100 . ..... ..... 001 ..... 1110111 @r_vm
 vfqwbdota_vv     101110 . ..... ..... 001 ..... 1110111 @r_vm
 vfqwbdota_alt_vv 101111 . ..... ..... 001 ..... 1110111 @r_vm
 
+# *** Zvfbdota32f FP32 Batched Dot-Product Extension ***
+vfbdota_vv  101011 . ..... ..... 001 ..... 1110111 @r_vm
+
 # *** Zvfqwdota8f OCP FP8 Dot-Product Extension ***
 vfqwdota_vv     100110 . ..... ..... 001 ..... 1110111 @r_vm
 vfqwdota_alt_vv 100111 . ..... ..... 001 ..... 1110111 @r_vm
diff --git a/target/riscv/tcg/insn_trans/trans_rvv.c.inc 
b/target/riscv/tcg/insn_trans/trans_rvv.c.inc
index 50f0f67ea60..9b06c6a782c 100644
--- a/target/riscv/tcg/insn_trans/trans_rvv.c.inc
+++ b/target/riscv/tcg/insn_trans/trans_rvv.c.inc
@@ -4430,3 +4430,14 @@ static bool vext_vfqwbdota_check(DisasContext *s, 
arg_rmrr * a)
 
 GEN_ZVBDOTA_FP_TRANS(vfqwbdota_vv, vext_vfqwbdota_check, MO_32)
 GEN_ZVBDOTA_FP_TRANS(vfqwbdota_alt_vv, vext_vfqwbdota_check, MO_32)
+
+/* Zvfbdota32f FP32 Batched Dot-Product Extension */
+static bool vext_vfbdota_check(DisasContext *s, arg_rmrr * a)
+{
+    int32_t sew = s->sew;
+    return s->mstatus_fs != EXT_STATUS_DISABLED &&
+           vext_zvbdota_base_check(s, a, sew) &&
+           (s->cfg_ptr->ext_zvfbdota32f && sew == MO_32);
+}
+
+GEN_ZVBDOTA_FP_TRANS(vfbdota_vv, vext_vfbdota_check, MO_32)
diff --git a/target/riscv/tcg/vector_helper.c b/target/riscv/tcg/vector_helper.c
index e641e21d088..7d680747ddd 100644
--- a/target/riscv/tcg/vector_helper.c
+++ b/target/riscv/tcg/vector_helper.c
@@ -6327,3 +6327,61 @@ void HELPER(NAME)(void *vd, void *v0, void *vs1, void 
*vs2,              \
 
 GEN_VEXT_ZVBDOTA_FP8(vfqwbdota_vv, e4m3)
 GEN_VEXT_ZVBDOTA_FP8(vfqwbdota_alt_vv, e5m2)
+
+/*
+ * Zvfbdota32f: FP32 Batched Dot Product with Accumulation
+ */
+void HELPER(vfbdota_vv)(void *vd, void *v0, void *vs1, void *vs2, uint32_t ci,
+                        CPURISCVState *env, uint32_t desc)
+{
+    uint32_t vm = vext_vm(desc);
+    uint32_t vl = env->vl;
+    uint32_t vta = vext_vta(desc);
+    uint32_t vta_all_1s = vext_vta_all_1s(desc);
+    uint32_t vlenb = riscv_cpu_cfg(env)->vlenb;
+    uint32_t vd_emul = vext_lmul(desc);
+    uint32_t log_esz = 2;
+    uint32_t vlmax = vlenb >> MO_32;
+    uint32_t total_elems = (vlenb << vd_emul) >> log_esz;
+    uint32_t n, k;
+    float32 *a_vec, *b_vec;
+    float32 acc, prod, a, b;
+
+    VSTART_CHECK_EARLY_EXIT(env, vl);
+
+    a_vec = g_new0(float32, vlmax);
+    b_vec = g_new0(float32, vlmax);
+
+    FloatExceptionFlags pre_fflag = get_float_exception_flags(&env->fp_status);
+    for (n = 0; n < 8; n++) {
+        if (!vm && !vext_elem_mask(v0, ci + n)) {
+            continue;
+        }
+
+        for (k = 0; k < vl; k++) {
+            a_vec[k] = *((float32 *)vs2 + n * vlmax + H4(k));
+            b_vec[k] = *((float32 *)vs1 + H4(k));
+        }
+
+        acc = *((float32 *)vd + H4(ci + n));
+        for (k = 0; k < vlmax; k++) {
+            a = a_vec[k];
+            b = b_vec[k];
+            prod = float32_mul(a, b, &env->fp_status);
+            acc = float32_add(acc, prod, &env->fp_status);
+        }
+        *((float32 *)vd + H4(ci + n)) = acc;
+    }
+    riscv_cpu_check_fflags(env, pre_fflag);
+
+    g_free(a_vec);
+    g_free(b_vec);
+
+    env->vstart = 0;
+
+    if (ci + 8 < total_elems) {
+        if (vta_all_1s) {
+            vext_set_elems_1s(vd, vta, (ci + 8) * 4, total_elems * 4);
+        }
+    }
+}
-- 
2.56.0


Reply via email to