Implement the vfwbdota.vv instruction from the Zvfwbdota16bf extension. This instruction performs a batched dot-product operation on BF16 vectors, accumulating into FP32
Signed-off-by: Max Chou <[email protected]> --- target/riscv/helper.h | 2 + target/riscv/insn32.decode | 3 ++ target/riscv/tcg/insn_trans/trans_rvv.c.inc | 52 +++++++++++++++++++ target/riscv/tcg/vector_helper.c | 55 +++++++++++++++++++++ 4 files changed, 112 insertions(+) diff --git a/target/riscv/helper.h b/target/riscv/helper.h index 84571a999f4..1f2dfd6d557 100644 --- a/target/riscv/helper.h +++ b/target/riscv/helper.h @@ -1395,6 +1395,8 @@ DEF_HELPER_6(vqwdotas_vv_h, void, ptr, ptr, ptr, ptr, env, i32) /* Zvfwdota16bf - BF16 Dot Product */ 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) /* Zvfqwdota8f - OCP FP8 Dot Product */ DEF_HELPER_6(vfqwdota_vv, void, ptr, ptr, ptr, ptr, env, i32) diff --git a/target/riscv/insn32.decode b/target/riscv/insn32.decode index 8a06547a570..00ed64b73fd 100644 --- a/target/riscv/insn32.decode +++ b/target/riscv/insn32.decode @@ -1105,6 +1105,9 @@ vqwdotas_vv 100111 . ..... ..... 000 ..... 1110111 @r_vm # *** Zvfwdota16bf BF16 Dot-Product Extension *** vfwdota_vv 100100 . ..... ..... 001 ..... 1110111 @r_vm +# *** Zvfwbdota16bf BF16 Batched Dot-Product Extension *** +vfwbdota_vv 101100 . ..... ..... 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 650affc045b..d4678171ac5 100644 --- a/target/riscv/tcg/insn_trans/trans_rvv.c.inc +++ b/target/riscv/tcg/insn_trans/trans_rvv.c.inc @@ -4365,3 +4365,55 @@ static bool trans_##NAME(DisasContext *s, arg_rmrr *a) \ GEN_ZVBDOTA_INT_TRANS(vqwbdotau_vv, vext_vqwbdota_check) GEN_ZVBDOTA_INT_TRANS(vqwbdotas_vv, vext_vqwbdota_check) + +/* Zvfwbdota16bf BF16 Batched Dot-Product Extension */ +static bool vext_vfwbdota_check(DisasContext *s, arg_rmrr * a) +{ + int32_t sew = s->sew; + int32_t vd_eew = sew + 1; + return s->mstatus_fs != EXT_STATUS_DISABLED && + vext_zvbdota_base_check(s, a, vd_eew) && + (s->cfg_ptr->ext_zvfwbdota16bf && s->sew == MO_16 && s->altfmt == 1); +} + +#define GEN_ZVBDOTA_FP_TRANS(NAME, CHECK, EEW_VD) \ +static bool trans_##NAME(DisasContext *s, arg_rmrr * a) \ +{ \ + if (CHECK(s, a)) { \ + TCGv_ptr dest, src1, src2, mask; \ + TCGv_i32 ci, desc; \ + uint32_t data = 0; \ + uint32_t vlenb = s->cfg_ptr->vlenb; \ + int32_t vd_emul = MAX(0, EEW_VD + 3 - ctzl(vlenb)); \ + \ + gen_set_rm(s, RISCV_FRM_DYN); \ + \ + data = FIELD_DP32(data, VDATA, VM, a->vm); \ + data = FIELD_DP32(data, VDATA, LMUL, vd_emul); \ + data = FIELD_DP32(data, VDATA, VMA, s->vma); \ + data = FIELD_DP32(data, VDATA, VTA, s->vta); \ + data = FIELD_DP32(data, VDATA, VTA_ALL_1S, s->cfg_vta_all_1s); \ + \ + dest = tcg_temp_new_ptr(); \ + mask = tcg_temp_new_ptr(); \ + src1 = tcg_temp_new_ptr(); \ + src2 = tcg_temp_new_ptr(); \ + desc = tcg_constant_i32(simd_desc(s->cfg_ptr->vlenb, \ + s->cfg_ptr->vlenb, data)); \ + ci = tcg_constant_i32((a->rs2 & 0x7) << 3); \ + \ + tcg_gen_addi_ptr(dest, tcg_env, vreg_ofs(s, a->rd)); \ + tcg_gen_addi_ptr(src1, tcg_env, vreg_ofs(s, a->rs1)); \ + tcg_gen_addi_ptr(src2, tcg_env, vreg_ofs(s, a->rs2 & ~0x7)); \ + tcg_gen_addi_ptr(mask, tcg_env, vreg_ofs(s, 0)); \ + \ + gen_helper_##NAME(dest, mask, src1, src2, ci, tcg_env, desc); \ + \ + finalize_rvv_inst(s); \ + \ + return true; \ + } \ + return false; \ +} + +GEN_ZVBDOTA_FP_TRANS(vfwbdota_vv, vext_vfwbdota_check, MO_32) diff --git a/target/riscv/tcg/vector_helper.c b/target/riscv/tcg/vector_helper.c index 873adf4d7a6..51170625aad 100644 --- a/target/riscv/tcg/vector_helper.c +++ b/target/riscv/tcg/vector_helper.c @@ -6201,3 +6201,58 @@ GEN_VEXT_ZVBDOTA_INT(vqwbdotau_vv_b, 32, 8, uint8_t, H4, H1) GEN_VEXT_ZVBDOTA_INT(vqwbdotas_vv_b, 32, 8, int8_t, H4, H1) GEN_VEXT_ZVBDOTA_INT(vqwbdotau_vv_h, 64, 16, uint16_t, H8, H2) GEN_VEXT_ZVBDOTA_INT(vqwbdotas_vv_h, 64, 16, int16_t, H8, H2) + +/* + * Zvfwbdota16bf: BF16 Batched Dot Product with Accumulation + */ +void HELPER(vfwbdota_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 total_elems = (vlenb << vd_emul) >> log_esz; + uint32_t vlmax = vlenb >> MO_16; + uint32_t n, k; + bfloat16 *a_vec, *b_vec; + float32 acc, dot_result; + + VSTART_CHECK_EARLY_EXIT(env, vl); + + a_vec = g_new0(bfloat16, vlmax); + b_vec = g_new0(bfloat16, 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] = *((bfloat16 *)vs2 + n * vlmax + H2(k)); + b_vec[k] = *((bfloat16 *)vs1 + H2(k)); + } + + dot_result = rvbna_dot_bfloat16_bfloat16(a_vec, b_vec, vlmax, + &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); + } + } +} -- 2.56.0
