@samitouri / QOSamiQemu / commits / 0c00be3786

target/arm: Implement FDOT (FP8 to FP16) for SVE

Reviewed-by: Peter Maydell <peter.maydell@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> Message-id: 20260609192110.752384-39-richard.henderson@linaro.org Signed-off-by: Peter Maydell <peter.maydell@linaro.org>

Richard Henderson committed Jun 9, 2026 at 12:21 UTC 0c00be3786c2b8441d20dd092276100f03deeb75
3 files changed +42
target/arm/cpu-features.h
+5
@@ -1580,6 +1580,11 @@ static inline bool isar_feature_aa64_ssve_f8dp4(const ARMISARegisters *id)
1580 return FIELD_EX64_IDREG(id, ID_AA64SMFR0, SF8DP4);
1581 }
1582
1583 +static inline bool isar_feature_aa64_ssve_f8dp2(const ARMISARegisters *id)
1584 +{
1585 + return FIELD_EX64_IDREG(id, ID_AA64SMFR0, SF8DP2);
1586 +}
1587 +
1588 static inline bool isar_feature_aa64_sme_b16b16(const ARMISARegisters *id)
1589 {
1590 return FIELD_EX64_IDREG(id, ID_AA64SMFR0, B16B16);
target/arm/tcg/sve.decode
+2
@@ -1875,6 +1875,7 @@ FDOT_zzzz 01100100 00 1 ..... 10 0 00 0 ..... ..... @rda_rn_rm_ex esz=2
1875 BFDOT_zzzz 01100100 01 1 ..... 10 0 00 0 ..... ..... @rda_rn_rm_ex esz=2
1876
1877 FDOT_sb 01100100 01 1 ..... 10 0 00 1 ..... ..... @rda_rn_rm_ex esz=2
1878 +FDOT_hb 01100100 00 1 ..... 10 0 00 1 ..... ..... @rda_rn_rm_ex esz=1
1879
1880 ### SVE2 floating-point multiply-add long (indexed)
1881
@@ -1900,6 +1901,7 @@ FDOT_zzxz 01100100 00 1 ..... 010000 ..... ..... @rrxr_2 esz=2
1901 BFDOT_zzxz 01100100 01 1 ..... 010000 ..... ..... @rrxr_2 esz=2
1902
1903 FDOT_idx_sb 01100100 01 1 ..... 010001 ..... ..... @rrxr_2 esz=2
1904 +FDOT_idx_hb 01100100 00 1 ..... 0100.1 ..... ..... @rrx_3a esz=1
1905
1906 ### SVE broadcast predicate element
1907
target/arm/tcg/translate-sve.c
+35
@@ -8426,3 +8426,38 @@ static bool do_f8dp4(DisasContext *s, gen_helper_gvec_3_ptr *fn,
8426 TRANS(FDOT_sb, do_f8dp4, gen_helper_gvec_fdot_sb, a->rd, a->rn, a->rm, 0)
8427 TRANS(FDOT_idx_sb, do_f8dp4, gen_helper_gvec_fdot_idx_sb,
8428 a->rd, a->rn, a->rm, a->index)
8429 +
8430 +static bool do_f8dp2(DisasContext *s, gen_helper_gvec_3_ptr *fn,
8431 + int rd, int rn, int rm, int index)
8432 +{
8433 + bool fp8dp2 = dc_isar_feature(aa64_f8dp2, s);
8434 + bool ssve_fp8dp2 = dc_isar_feature(aa64_ssve_f8dp2, s);
8435 + bool ok = false;
8436 +
8437 + /* Feature detection and enabling are complex here. */
8438 + if (!(ssve_fp8dp2 || (fp8dp2 && dc_isar_feature(aa64_sve2, s)))) {
8439 + return false;
8440 + }
8441 + if (fpmr_access_check(s)) {
8442 + if (fp8dp2) {
8443 + s->is_nonstreaming = !ssve_fp8dp2;
8444 + ok = sve_access_check(s);
8445 + } else {
8446 + ok = sme_sm_enabled_check(s);
8447 + }
8448 + }
8449 +
8450 + if (ok) {
8451 + unsigned vsz = vec_full_reg_size(s);
8452 + tcg_gen_gvec_3_ptr(vec_full_reg_offset(s, rd),
8453 + vec_full_reg_offset(s, rn),
8454 + vec_full_reg_offset(s, rm),
8455 + tcg_env, vsz, vsz,
8456 + index, fn);
8457 + }
8458 + return true;
8459 +}
8460 +
8461 +TRANS(FDOT_hb, do_f8dp2, gen_helper_gvec_fdot_hb, a->rd, a->rn, a->rm, 0)
8462 +TRANS(FDOT_idx_hb, do_f8dp2, gen_helper_gvec_fdot_idx_hb,
8463 + a->rd, a->rn, a->rm, a->index)