target/arm: Implement FMMLA (FP8 to FP16) for AdvSIMD
Reviewed-by: Peter Maydell <peter.maydell@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> Message-id: 20260609192110.752384-44-richard.henderson@linaro.org Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Richard Henderson committed
Jun 9, 2026 at 12:21 UTC
c44e030bdbd4256509aa4e2b14fd11fb5c9ba45f
5 files changed
+33
target/arm/cpu-features.h
+5
@@ -1650,6 +1650,11 @@ static inline bool isar_feature_aa64_f8mm8(const ARMISARegisters *id)
1650
return FIELD_EX64_IDREG(id, ID_AA64FPFR0, F8MM8);
1651
}
1652
1653
+static inline bool isar_feature_aa64_f8mm4(const ARMISARegisters *id)
1654
+{
1655
+ return FIELD_EX64_IDREG(id, ID_AA64FPFR0, F8MM4);
1656
+}
1657
+
1658
/*
1659
* Combinations of feature tests, for ease of use with TRANS_FEAT.
1660
*/
target/arm/tcg/a64.decode
+1
@@ -1228,6 +1228,7 @@ FDOT_sb_v 0.00 1110 000 ..... 11111 1 ..... ..... @qrrr_s
1228
FDOT_hb_v 0.00 1110 010 ..... 11111 1 ..... ..... @qrrr_h
1229
1230
FMMLA_sb 0110 1110 100 ..... 11101 1 ..... ..... @rrr_q1e0
1231
+FMMLA_hb 0110 1110 000 ..... 11101 1 ..... ..... @rrr_q1e0
1232
1233
### Advanced SIMD scalar x indexed element
1234
target/arm/tcg/fp8_helper.c
+25
@@ -832,3 +832,28 @@ void HELPER(gvec_fmmla_sb)(void *vd, void *vn, void *vm,
832
833
clear_tail(vd, oprsz, simd_maxsz(desc));
834
}
835
+
836
+void HELPER(gvec_fmmla_hb)(void *vd, void *vn, void *vm,
837
+ CPUARMState *env, uint32_t desc)
838
+{
839
+ FP8MulContext ctx = fp8_mul_start(env, 0xf);
840
+ size_t oprsz = simd_oprsz(desc);
841
+ size_t nseg = oprsz / 8;
842
+ uint32_t *n = vn;
843
+ uint32_t *m = vm;
844
+ float16 *d = vd;
845
+
846
+ for (size_t seg = 0; seg < nseg; seg++, d += 4, n += 2, m += 2) {
847
+ float16 d0 = f8dotadd_h(n[H4(0)], m[H4(0)], 4, d[H2(0)], &ctx);
848
+ float16 d1 = f8dotadd_h(n[H4(0)], m[H4(1)], 4, d[H2(1)], &ctx);
849
+ float16 d2 = f8dotadd_h(n[H4(1)], m[H4(0)], 4, d[H2(2)], &ctx);
850
+ float16 d3 = f8dotadd_h(n[H4(1)], m[H4(1)], 4, d[H2(3)], &ctx);
851
+
852
+ d[H2(0)] = d0;
853
+ d[H2(1)] = d1;
854
+ d[H2(2)] = d2;
855
+ d[H2(3)] = d3;
856
+ }
857
+
858
+ clear_tail(vd, oprsz, simd_maxsz(desc));
859
+}
target/arm/tcg/helper-fp8-defs.h
+1
@@ -37,3 +37,4 @@ DEF_HELPER_FLAGS_5(gvec_fdot_hb, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, env, i32)
37
DEF_HELPER_FLAGS_5(gvec_fdot_idx_hb, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, env, i32)
38
39
DEF_HELPER_FLAGS_5(gvec_fmmla_sb, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, env, i32)
40
+DEF_HELPER_FLAGS_5(gvec_fmmla_hb, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, env, i32)
target/arm/tcg/translate-a64.c
+1
@@ -7516,6 +7516,7 @@ static bool do_f8dot(DisasContext *s, arg_qrrr_e *a,
7516
TRANS_FEAT(FDOT_sb_v, aa64_f8dp4, do_f8dot, a, gen_helper_gvec_fdot_sb)
7517
TRANS_FEAT(FDOT_hb_v, aa64_f8dp2, do_f8dot, a, gen_helper_gvec_fdot_hb)
7518
TRANS_FEAT(FMMLA_sb, aa64_f8mm8, do_f8dot, a, gen_helper_gvec_fmmla_sb)
7519
+TRANS_FEAT(FMMLA_hb, aa64_f8mm4, do_f8dot, a, gen_helper_gvec_fmmla_hb)
7520
7521
static bool do_f8dot_idx(DisasContext *s, arg_qrrx_e *a,
7522
gen_helper_gvec_3_ptr *fn)