Implements the Zvqwbdota8i INT8 dot-product instructions: - vqwbdotau.vv - vqwbdotas.vv
The vqwbdotau.vv and vqwbdotas.vv instructions perform batched dot-product operations on 8-bit integer vectors, computing 8 parallel dot products and accumulating into 32-bit integers. Signed-off-by: Max Chou <[email protected]> --- target/riscv/helper.h | 4 ++ target/riscv/insn32.decode | 4 ++ target/riscv/tcg/insn_trans/trans_rvv.c.inc | 75 +++++++++++++++++++++ target/riscv/tcg/vector_helper.c | 56 +++++++++++++++ 4 files changed, 139 insertions(+) diff --git a/target/riscv/helper.h b/target/riscv/helper.h index 48c64f45748..e48b9b4edec 100644 --- a/target/riscv/helper.h +++ b/target/riscv/helper.h @@ -1400,6 +1400,10 @@ DEF_HELPER_6(vfwdota_vv, void, ptr, ptr, ptr, ptr, env, i32) 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) +/* Zvqwbdota8i - 8-bit Integer Batched Dot Product */ +DEF_HELPER_7(vqwbdotau_vv_b, void, ptr, ptr, ptr, ptr, i32, env, i32) +DEF_HELPER_7(vqwbdotas_vv_b, void, ptr, ptr, ptr, ptr, i32, 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 def5785f8e2..bf325229100 100644 --- a/target/riscv/insn32.decode +++ b/target/riscv/insn32.decode @@ -1108,3 +1108,7 @@ 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 + +# *** Zvqwbdota8i 8-bit Integer Batched Dot-Product Extension *** +vqwbdotau_vv 101110 . ..... ..... 000 ..... 1110111 @r_vm +vqwbdotas_vv 101111 . ..... ..... 000 ..... 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 6eb22494838..bc12bcb9b72 100644 --- a/target/riscv/tcg/insn_trans/trans_rvv.c.inc +++ b/target/riscv/tcg/insn_trans/trans_rvv.c.inc @@ -4289,3 +4289,78 @@ static bool trans_##NAME(DisasContext *s, arg_rmrr *a) \ GEN_ZVDOTA_FP8_TRANS(vfqwdota_vv, vext_vfqwdota_check) GEN_ZVDOTA_FP8_TRANS(vfqwdota_alt_vv, vext_vfqwdota_check) + +static bool vext_zvbdota_base_check(DisasContext *s, arg_rmrr * a, + int32_t vd_eew) +{ + int32_t vlenb = s->cfg_ptr->vlenb; + int32_t vd_emul = MAX(0, vd_eew + 3 - ctzl(vlenb)); + uint8_t ci = a->rs2 & 0b111; + int rs2 = a->rs2 & ~0x7; + + return require_rvv(s) && + vext_check_isa_ill(s) && + s->vstart_eq_zero && + (s->lmul == 0) && + ((ci << (3 + vd_eew)) < vlenb) && + require_vm(a->vm, a->rd) && + require_align(a->rd, vd_emul) && + !is_overlapped(a->rd, 1 << vd_emul, a->rs1, 1) && + !is_overlapped(a->rd, 1 << vd_emul, rs2, 8) && + vext_check_input_eew(s, a->rs1, s->sew, rs2, s->sew, a->vm); +} + +static bool vext_vqwbdota_check(DisasContext *s, arg_rmrr *a) +{ + int32_t sew = s->sew; + int32_t vd_eew = sew + 2; + return vext_zvbdota_base_check(s, a, vd_eew) && + (s->cfg_ptr->ext_zvqwbdota8i && sew == MO_8); +} + +typedef void gen_helper_zvbdota_int(TCGv_ptr, TCGv_ptr, TCGv_ptr, TCGv_ptr, + TCGv_i32, TCGv_env, TCGv_i32); + +#define GEN_ZVBDOTA_INT_TRANS(NAME, CHECK) \ +static bool trans_##NAME(DisasContext *s, arg_rmrr *a) \ +{ \ + static gen_helper_zvbdota_int * const fns[2] = { \ + gen_helper_##NAME##_b, NULL, \ + }; \ + if (CHECK(s, a)) { \ + TCGv_ptr dest, src1, src2, mask; \ + TCGv_i32 ci, desc; \ + uint32_t data = 0; \ + int32_t vd_eew = s->sew + 2; \ + int32_t vd_emul = MAX(0, vd_eew + 3 - ctzl(s->cfg_ptr->vlenb)); \ + \ + data = FIELD_DP32(data, VDATA, VM, a->vm); \ + data = FIELD_DP32(data, VDATA, VTA, s->vta); \ + data = FIELD_DP32(data, VDATA, LMUL, vd_emul); \ + data = FIELD_DP32(data, VDATA, VMA, s->vma); \ + 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)); \ + \ + fns[s->sew](dest, mask, src1, src2, ci, tcg_env, desc); \ + \ + finalize_rvv_inst(s); \ + \ + return true; \ + } \ + return false; \ +} + +GEN_ZVBDOTA_INT_TRANS(vqwbdotau_vv, vext_vqwbdota_check) +GEN_ZVBDOTA_INT_TRANS(vqwbdotas_vv, vext_vqwbdota_check) diff --git a/target/riscv/tcg/vector_helper.c b/target/riscv/tcg/vector_helper.c index b1ca584f9ef..9e872eccaeb 100644 --- a/target/riscv/tcg/vector_helper.c +++ b/target/riscv/tcg/vector_helper.c @@ -6139,3 +6139,59 @@ void HELPER(NAME)(void *vd, void *v0, void *vs1, void *vs2, \ GEN_VEXT_ZVDOTA_FP8(vfqwdota_vv, e4m3) GEN_VEXT_ZVDOTA_FP8(vfqwdota_alt_vv, e5m2) + +/* + * Zvqwbdota8i: 8-bit Integer Batched Dot Product with Accumulation + */ +#define GEN_VEXT_ZVBDOTA_INT(NAME, WD, W1, T2, HD, HS) \ +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) >> ctzl(sizeof(uint##WD##_t)); \ + uint8_t altfmt = FIELD_EX64(env->vtype, VTYPE, ALTFMT); \ + uint32_t n, k; \ + uint32_t stride = vlenb >> ctzl(sizeof(uint##W1##_t)); \ + uint##WD##_t acc, dot_result, a; \ + \ + VSTART_CHECK_EARLY_EXIT(env, vl); \ + \ + for (n = 0; n < 8; n++) { \ + if (!vm && !vext_elem_mask(v0, ci + n)) { \ + continue; \ + } \ + \ + acc = *((uint##WD##_t *)vd + HD(ci + n)); \ + dot_result = 0; \ + \ + for (k = 0; k < vl; k++) { \ + uint##WD##_t b = *((T2 *)vs2 + n * stride + HS(k)); \ + if (altfmt) { \ + a = *((int##W1##_t *)vs1 + HS(k)); \ + } else { \ + a = *((uint##W1##_t *)vs1 + HS(k)); \ + } \ + dot_result += (a * b); \ + } \ + \ + *((uint##WD##_t *)vd + HD(ci + n)) = acc + dot_result; \ + } \ + \ + env->vstart = 0; \ + \ + if (ci + 8 < total_elems) { \ + if (vta_all_1s) { \ + vext_set_elems_1s(vd, vta, \ + (ci + 8) * sizeof(int##WD##_t), \ + total_elems * sizeof(int##WD##_t)); \ + } \ + } \ +} + +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) -- 2.56.0
