target/arm: Implement FDOT (FP8 to FP32) for AdvSIMD
Reviewed-by: Peter Maydell <peter.maydell@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> Message-id: 20260609192110.752384-35-richard.henderson@linaro.org Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Richard Henderson committed
Jun 9, 2026 at 12:20 UTC
9e599b5cd74b765dc18ed8094c47f36441c0d14c
5 files changed
+81
target/arm/cpu-features.h
+5
@@ -1625,6 +1625,11 @@ static inline bool isar_feature_aa64_f8fma(const ARMISARegisters *id)
1625
return FIELD_EX64_IDREG(id, ID_AA64FPFR0, F8FMA);
1626
}
1627
1628
+static inline bool isar_feature_aa64_f8dp4(const ARMISARegisters *id)
1629
+{
1630
+ return FIELD_EX64_IDREG(id, ID_AA64FPFR0, F8DP4);
1631
+}
1632
+
1633
/*
1634
* Combinations of feature tests, for ease of use with TRANS_FEAT.
1635
*/
target/arm/tcg/a64.decode
+4
@@ -1224,6 +1224,8 @@ FMLAL_hb_v 0 idxn:1 00 1110 110 rm:5 11111 1 rn:5 rd:5 \
1224
FMLALL_sb_v 0.00 1110 0.0 rm:5 110001 rn:5 rd:5 \
1225
&rxx idxm=0 idxn=%fmlall_idxn
1226
1227
+FDOT_sb_v 0.00 1110 000 ..... 11111 1 ..... ..... @qrrr_s
1228
+
1229
### Advanced SIMD scalar x indexed element
1230
1231
FMUL_si 0101 1111 00 .. .... 1001 . 0 ..... ..... @rrx_h
@@ -1348,6 +1350,8 @@ FMLAL_hb_vi 0 idxn:1 00 1111 11 ... rm:3 0000 . 0 rn:5 rd:5 \
1350
FMLALL_sb_vi 0 . 10 1111 0 . ... rm:3 1000 . 0 rn:5 rd:5 \
1351
&rxx idxm=%hlm4 idxn=%fmlall_idxn
1352
1353
+FDOT_sb_vi 0.00 1111 00 . ..... 0000 . 0 ..... ..... @qrrx_s
1354
+
1355
# Floating-point conditional select
1356
1357
FCSEL 0001 1110 .. 1 rm:5 cond:4 11 rn:5 rd:5 esz=%esz_hsd
target/arm/tcg/fp8_helper.c
+39
@@ -729,3 +729,42 @@ void HELPER(gvec_fmla_idx_sb)(void *vd, void *vn, void *vm,
729
730
clear_tail(vd, oprsz, simd_maxsz(desc));
731
}
732
+
733
+void HELPER(gvec_fdot_sb)(void *vd, void *vn, void *vm,
734
+ CPUARMState *env, uint32_t desc)
735
+{
736
+ FP8MulContext ctx = fp8_mul_start(env, -1);
737
+ size_t oprsz = simd_oprsz(desc);
738
+ size_t nelem = oprsz / 4;
739
+ uint32_t *n = vn;
740
+ uint32_t *m = vm;
741
+ float32 *d = vd;
742
+
743
+ for (size_t i = 0; i < nelem; i++) {
744
+ d[i] = f8dotadd_s(n[i], m[i], 4, d[i], &ctx);
745
+ }
746
+
747
+ clear_tail(vd, oprsz, simd_maxsz(desc));
748
+}
749
+
750
+void HELPER(gvec_fdot_idx_sb)(void *vd, void *vn, void *vm,
751
+ CPUARMState *env, uint32_t desc)
752
+{
753
+ FP8MulContext ctx = fp8_mul_start(env, -1);
754
+ size_t idx = simd_data(desc);
755
+ size_t oprsz = simd_oprsz(desc);
756
+ size_t nelem = oprsz / 4;
757
+ uint32_t *n = vn;
758
+ uint32_t *m = vm;
759
+ float32 *d = vd;
760
+ size_t i = 0;
761
+
762
+ do {
763
+ uint32_t e1 = m[i + H4(idx)];
764
+ do {
765
+ d[i] = f8dotadd_s(n[i], e1, 4, d[i], &ctx);
766
+ } while (++i % 4 != 0);
767
+ } while (i < nelem);
768
+
769
+ clear_tail(vd, oprsz, simd_maxsz(desc));
770
+}
target/arm/tcg/helper-fp8-defs.h
+3
@@ -29,3 +29,6 @@ DEF_HELPER_FLAGS_5(gvec_fmla_idx_hb, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, env,
29
30
DEF_HELPER_FLAGS_5(gvec_fmla_sb, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, env, i32)
31
DEF_HELPER_FLAGS_5(gvec_fmla_idx_sb, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, env, i32)
32
+
33
+DEF_HELPER_FLAGS_5(gvec_fdot_sb, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, env, i32)
34
+DEF_HELPER_FLAGS_5(gvec_fdot_idx_sb, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, env, i32)
target/arm/tcg/translate-a64.c
+30
@@ -7500,6 +7500,36 @@ TRANS_FEAT(FMLAL_hb_vi, aa64_f8fma, do_fmla_fp8, a, gen_helper_gvec_fmla_idx_hb)
7500
TRANS_FEAT(FMLALL_sb_v, aa64_f8fma, do_fmla_fp8, a, gen_helper_gvec_fmla_sb)
7501
TRANS_FEAT(FMLALL_sb_vi, aa64_f8fma, do_fmla_fp8, a, gen_helper_gvec_fmla_idx_sb)
7502
7503
+static bool do_f8dot(DisasContext *s, arg_qrrr_e *a,
7504
+ gen_helper_gvec_3_ptr *fn)
7505
+{
7506
+ if (fpmr_access_check(s) && fp_access_check(s)) {
7507
+ tcg_gen_gvec_3_ptr(vec_full_reg_offset(s, a->rd),
7508
+ vec_full_reg_offset(s, a->rn),
7509
+ vec_full_reg_offset(s, a->rm),
7510
+ tcg_env, a->q ? 16 : 8, vec_full_reg_size(s),
7511
+ 0, fn);
7512
+ }
7513
+ return true;
7514
+}
7515
+
7516
+TRANS_FEAT(FDOT_sb_v, aa64_f8dp4, do_f8dot, a, gen_helper_gvec_fdot_sb)
7517
+
7518
+static bool do_f8dot_idx(DisasContext *s, arg_qrrx_e *a,
7519
+ gen_helper_gvec_3_ptr *fn)
7520
+{
7521
+ if (fpmr_access_check(s) && fp_access_check(s)) {
7522
+ tcg_gen_gvec_3_ptr(vec_full_reg_offset(s, a->rd),
7523
+ vec_full_reg_offset(s, a->rn),
7524
+ vec_full_reg_offset(s, a->rm),
7525
+ tcg_env, a->q ? 16 : 8, vec_full_reg_size(s),
7526
+ a->idx, fn);
7527
+ }
7528
+ return true;
7529
+}
7530
+
7531
+TRANS_FEAT(FDOT_sb_vi, aa64_f8dp4, do_f8dot_idx, a, gen_helper_gvec_fdot_idx_sb)
7532
+
7533
static bool do_int3_vector_idx(DisasContext *s, arg_qrrx_e *a,
7534
gen_helper_gvec_3 * const fns[2])
7535
{