On 11/4/2026 4:41 PM, Max Chou wrote:
Support the Zvfqwbdota8f OCP FP8 batched dot-product instructions.
These instructions perform 8 parallel dot products between OCP FP8 vectors,
accumulating into FP32 elements.

Signed-off-by: Max Chou <[email protected]>
---

Reviewed-by: Daniel Henrique Barboza <[email protected]>

  target/riscv/helper.h                       |  4 ++
  target/riscv/insn32.decode                  |  4 ++
  target/riscv/tcg/insn_trans/trans_rvv.c.inc | 13 ++++
  target/riscv/tcg/vector_helper.c            | 69 +++++++++++++++++++++
  4 files changed, 90 insertions(+)

diff --git a/target/riscv/helper.h b/target/riscv/helper.h
index 1f2dfd6d557..a257436fb31 100644
--- a/target/riscv/helper.h
+++ b/target/riscv/helper.h
@@ -1398,6 +1398,10 @@ 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)
+/* 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)
+
  /* Zvfqwdota8f - OCP FP8 Dot Product */
  DEF_HELPER_6(vfqwdota_vv, void, ptr, ptr, ptr, ptr, env, i32)
  DEF_HELPER_6(vfqwdota_alt_vv, void, ptr, ptr, ptr, ptr, env, i32)
diff --git a/target/riscv/insn32.decode b/target/riscv/insn32.decode
index 00ed64b73fd..48597915f7b 100644
--- a/target/riscv/insn32.decode
+++ b/target/riscv/insn32.decode
@@ -1108,6 +1108,10 @@ vfwdota_vv  100100 . ..... ..... 001 ..... 1110111 @r_vm
  # *** Zvfwbdota16bf BF16 Batched Dot-Product Extension ***
  vfwbdota_vv 101100 . ..... ..... 001 ..... 1110111 @r_vm
+# *** Zvfqwbdota8f OCP FP8 Batched Dot-Product Extension ***
+vfqwbdota_vv     101110 . ..... ..... 001 ..... 1110111 @r_vm
+vfqwbdota_alt_vv 101111 . ..... ..... 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 d4678171ac5..50f0f67ea60 100644
--- a/target/riscv/tcg/insn_trans/trans_rvv.c.inc
+++ b/target/riscv/tcg/insn_trans/trans_rvv.c.inc
@@ -4417,3 +4417,16 @@ static bool trans_##NAME(DisasContext *s, arg_rmrr * a)  
               \
  }
GEN_ZVBDOTA_FP_TRANS(vfwbdota_vv, vext_vfwbdota_check, MO_32)
+
+/* Zvfqwbdota8f OCP FP8 Batched Dot-Product Extension */
+static bool vext_vfqwbdota_check(DisasContext *s, arg_rmrr * a)
+{
+    int32_t sew = s->sew;
+    int32_t vd_eew = sew + 2;
+    return s->mstatus_fs != EXT_STATUS_DISABLED &&
+           vext_zvbdota_base_check(s, a, vd_eew) &&
+           (s->cfg_ptr->ext_zvfqwbdota8f && sew == MO_8);
+}
+
+GEN_ZVBDOTA_FP_TRANS(vfqwbdota_vv, vext_vfqwbdota_check, MO_32)
+GEN_ZVBDOTA_FP_TRANS(vfqwbdota_alt_vv, vext_vfqwbdota_check, MO_32)
diff --git a/target/riscv/tcg/vector_helper.c b/target/riscv/tcg/vector_helper.c
index 6f3b666784e..e641e21d088 100644
--- a/target/riscv/tcg/vector_helper.c
+++ b/target/riscv/tcg/vector_helper.c
@@ -6258,3 +6258,72 @@ void HELPER(vfwbdota_vv)(void *vd, void *v0, void *vs1, 
void *vs2, uint32_t ci,
          }
      }
  }
+
+/*
+ * Zvfqwdota8f: OCP FP8 Dot Product with Accumulation
+ */
+#define GEN_VEXT_ZVBDOTA_FP8(NAME, TYPE_VS2)                             \
+void HELPER(NAME)(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 total_elems = (vlenb << vd_emul) >> 2;                      \
+    uint32_t vlmax = vlenb;                                              \
+    uint32_t altfmt = FIELD_EX64(env->vtype, VTYPE, ALTFMT);             \
+    uint32_t n, k;                                                       \
+    uint8_t *a_vec, *b_vec;                                              \
+    float32 acc, dot_result;                                             \
+                                                                         \
+    VSTART_CHECK_EARLY_EXIT(env, vl);                                    \
+                                                                         \
+    a_vec = g_new0(uint8_t, vlmax);                                      \
+    b_vec = g_new0(uint8_t, 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] = *((uint8_t *)vs2 + n * vlmax + H1(k));            \
+            b_vec[k] = *((uint8_t *)vs1 + H1(k));                        \
+        }                                                                \
+                                                                         \
+        FloatSNaNRule pre_snan_rule = get_snan_rule(&env->fp_status);    \
+        set_snan_rule(float_snan_never, &env->fp_status);                \
+        if (altfmt == 0) {                                               \
+            dot_result = rvbna_dot_float8_##TYPE_VS2##_float8_e4m3(      \
+                    a_vec, b_vec, vlmax, &env->fp_status);               \
+        } else {                                                         \
+            dot_result = rvbna_dot_float8_##TYPE_VS2##_float8_e5m2(      \
+                    a_vec, b_vec, vlmax, &env->fp_status);               \
+        }                                                                \
+        set_snan_rule(pre_snan_rule, &env->fp_status);                   \
+                                                                         \
+        acc = *((float32 *)vd + H4(ci + n));                             \
+        *((float32 *)vd + H4(ci + n)) = rvbna_final_f32_add(             \
+                dot_result, acc, &env->fp_status);                       \
+    }                                                                    \
+    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);   \
+        }                                                                \
+    }                                                                    \
+}
+
+GEN_VEXT_ZVBDOTA_FP8(vfqwbdota_vv, e4m3)
+GEN_VEXT_ZVBDOTA_FP8(vfqwbdota_alt_vv, e5m2)


Reply via email to