Re: [PATCH v2 035/101] target/arm: Implement SME2 FDOT

[Richard Henderson]

On 6/23/25 09:38, Peter Maydell wrote:
On Sun, 22 Jun 2025 at 00:54, Richard Henderson
<richard.hender...@linaro.org> wrote:

Signed-off-by: Richard Henderson <richard.hender...@linaro.org>
---
  target/arm/tcg/helper-sme.h    |  5 ++++
  target/arm/tcg/sme_helper.c    | 44 ++++++++++++++++++++++++++++++++++
  target/arm/tcg/translate-sme.c | 18 ++++++++++++++
  target/arm/tcg/translate-sve.c |  5 ++++
  target/arm/tcg/sme.decode      | 14 +++++++++++
  target/arm/tcg/sve.decode      |  8 +++++--
  6 files changed, 92 insertions(+), 2 deletions(-)

diff --git a/target/arm/tcg/helper-sme.h b/target/arm/tcg/helper-sme.h
index cdd7058aed..ec93ff57ff 100644
--- a/target/arm/tcg/helper-sme.h
+++ b/target/arm/tcg/helper-sme.h
@@ -173,3 +173,8 @@ DEF_HELPER_FLAGS_5(gvec_fmaxnum_b16, TCG_CALL_NO_RWG,
                     void, ptr, ptr, ptr, fpst, i32)
  DEF_HELPER_FLAGS_5(gvec_fminnum_b16, TCG_CALL_NO_RWG,
                     void, ptr, ptr, ptr, fpst, i32)
+
+DEF_HELPER_FLAGS_6(sme2_fdot_h, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, env, i32)
+DEF_HELPER_FLAGS_6(sme2_fdot_idx_h, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, env, i32)
diff --git a/target/arm/tcg/sme_helper.c b/target/arm/tcg/sme_helper.c
index 194560eafa..bd9b81d5aa 100644
--- a/target/arm/tcg/sme_helper.c
+++ b/target/arm/tcg/sme_helper.c
@@ -1122,6 +1122,50 @@ void HELPER(sme_fmopa_h)(void *vza, void *vzn, void 
*vzm, void *vpn,
      }
  }

+void HELPER(sme2_fdot_h)(void *vd, void *vn, void *vm, void *va,
+                         CPUARMState *env, uint32_t desc)
+{
+    intptr_t i, oprsz = simd_maxsz(desc);
+    bool za = extract32(desc, SIMD_DATA_SHIFT, 1);
+    float_status *fpst_std = &env->vfp.fp_status[za ? FPST_ZA : FPST_A64];
+    float_status *fpst_f16 = &env->vfp.fp_status[za ? FPST_ZA_F16 : 
FPST_A64_F16];
+    float_status fpst_odd = *fpst_std;
+    float32 *d = vd, *a = va;
+    uint32_t *n = vn, *m = vm;
+
+    set_float_rounding_mode(float_round_to_odd, &fpst_odd);
+
+    for (i = 0; i < oprsz / sizeof(float32); ++i) {
+        d[H4(i)] = f16_dotadd(a[H4(i)], n[H4(i)], m[H4(i)],
+                              fpst_f16, fpst_std, &fpst_odd);

I can never figure out with these helpers when we need the H
macros and when we don't...
If the access is smaller than uint64_t, and the set of operations isn't completely uniform 
across all columns, then H is required.

As it happens, here the set of operations *is* uniform, so I probably didn't 
need the H.

r~