Implements the Zvfqwdota8f OCP FP8 dot-product instructions: - vfqwdota.vv: vs2 is E4M3, vs1 format selected by altfmt (0=E4M3, 1=E5M2) - vfqwdota.alt.vv: vs2 is E5M2, vs1 format selected by altfmt (0=E4M3, 1=E5M2)
Signed-off-by: Max Chou <[email protected]> --- target/riscv/helper.h | 5 ++ target/riscv/insn32.decode | 4 ++ target/riscv/tcg/insn_trans/trans_rvv.c.inc | 47 ++++++++++++++++ target/riscv/tcg/vector_helper.c | 62 +++++++++++++++++++++ 4 files changed, 118 insertions(+) diff --git a/target/riscv/helper.h b/target/riscv/helper.h index 918ff21b3e2..48c64f45748 100644 --- a/target/riscv/helper.h +++ b/target/riscv/helper.h @@ -1395,6 +1395,11 @@ 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) + +/* 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) + /* SiFive Custom int8 Matrix-Multiply */ DEF_HELPER_5(sf_vqmaccu_4x8x4, void, ptr, ptr, ptr, env, i32) DEF_HELPER_5(sf_vqmacc_4x8x4, void, ptr, ptr, ptr, env, i32) diff --git a/target/riscv/insn32.decode b/target/riscv/insn32.decode index d800ad99451..def5785f8e2 100644 --- a/target/riscv/insn32.decode +++ b/target/riscv/insn32.decode @@ -1104,3 +1104,7 @@ vqwdotas_vv 100111 . ..... ..... 000 ..... 1110111 @r_vm # *** Zvfwdota16bf BF16 Dot-Product Extension *** vfwdota_vv 100100 . ..... ..... 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 c50f863ea0b..6eb22494838 100644 --- a/target/riscv/tcg/insn_trans/trans_rvv.c.inc +++ b/target/riscv/tcg/insn_trans/trans_rvv.c.inc @@ -4242,3 +4242,50 @@ static bool trans_vfwdota_vv(DisasContext *s, arg_rmrr * a) } return false; } + +/* Zvfqwdota8f OCP FP8 Dot-Product Extension */ +static bool vext_vfqwdota_check(DisasContext *s, arg_rmrr *a) +{ + return s->mstatus_fs != EXT_STATUS_DISABLED && + vext_zvdota_base_check(s, a) && + (s->cfg_ptr->ext_zvfqwdota8f && s->sew == MO_8); +} + +#define GEN_ZVDOTA_FP8_TRANS(NAME, CHECK) \ +static bool trans_##NAME(DisasContext *s, arg_rmrr *a) \ +{ \ + if (CHECK(s, a)) { \ + TCGv_ptr dest, src1, src2, mask; \ + TCGv_i32 desc; \ + uint32_t data = 0; \ + \ + gen_set_rm(s, RISCV_FRM_DYN); \ + \ + data = FIELD_DP32(data, VDATA, VM, a->vm); \ + data = FIELD_DP32(data, VDATA, LMUL, s->lmul); \ + 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)); \ + \ + 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)); \ + tcg_gen_addi_ptr(mask, tcg_env, vreg_ofs(s, 0)); \ + \ + gen_helper_##NAME(dest, mask, src1, src2, tcg_env, desc); \ + \ + finalize_rvv_inst(s); \ + \ + return true; \ + } \ + return false; \ +} + +GEN_ZVDOTA_FP8_TRANS(vfqwdota_vv, vext_vfqwdota_check) +GEN_ZVDOTA_FP8_TRANS(vfqwdota_alt_vv, vext_vfqwdota_check) diff --git a/target/riscv/tcg/vector_helper.c b/target/riscv/tcg/vector_helper.c index a4c52ed296a..5cb44326200 100644 --- a/target/riscv/tcg/vector_helper.c +++ b/target/riscv/tcg/vector_helper.c @@ -6075,3 +6075,65 @@ void HELPER(vfwdota_vv)(void *vd, void *v0, void *vs1, void *vs2, vext_set_elems_1s(vd, vta, 1 * 4, total_elems * 4); } } + +/* + * Zvfqwdota8f: OCP FP8 Dot Product with Accumulation + */ +#define GEN_VEXT_ZVDOTA_FP8(NAME, TYPE_VS2) \ +void HELPER(NAME)(void *vd, void *v0, void *vs1, void *vs2, \ + CPURISCVState *env, uint32_t desc) \ +{ \ + uint32_t vlmax = vext_max_elems(desc, ctzl(sizeof(uint8_t))); \ + 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 total_elems = riscv_cpu_cfg(env)->vlenb >> 2; \ + uint32_t altfmt = FIELD_EX64(env->vtype, VTYPE, ALTFMT); \ + uint32_t i; \ + uint8_t *a_vec, *b_vec; \ + uint32_t dot_result, acc; \ + \ + VSTART_CHECK_EARLY_EXIT(env, vl); \ + \ + a_vec = g_new0(uint8_t, vlmax); \ + b_vec = g_new0(uint8_t, vlmax); \ + \ + for (i = 0; i < vl; i++) { \ + if (!vm && !vext_elem_mask(v0, i)) { \ + continue; \ + } \ + a_vec[i] = *((uint8_t *)vs2 + H1(i)); \ + b_vec[i] = *((uint8_t *)vs1 + H1(i)); \ + } \ + \ + FloatExceptionFlags pre_fflag = \ + get_float_exception_flags(&env->fp_status); \ + 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 = *((uint32_t *)vd + H4(0)); \ + *((uint32_t *)vd + H4(0)) = 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 (vta_all_1s) { \ + vext_set_elems_1s(vd, vta, 1 * 4, total_elems * 4); \ + } \ +} + +GEN_VEXT_ZVDOTA_FP8(vfqwdota_vv, e4m3) +GEN_VEXT_ZVDOTA_FP8(vfqwdota_alt_vv, e5m2) -- 2.56.0
