Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/arm/tcg/helper.h        |  8 ++++
 target/arm/tcg/translate-sme.c | 85 ++++++++++++++++++++++++++++++++++
 target/arm/tcg/vec_helper.c    | 51 ++++++++++++++++++++
 target/arm/tcg/sme.decode      | 63 ++++++++++++++++++++++++-
 4 files changed, 206 insertions(+), 1 deletion(-)

diff --git a/target/arm/tcg/helper.h b/target/arm/tcg/helper.h
index a19955b872..c4a208e3ba 100644
--- a/target/arm/tcg/helper.h
+++ b/target/arm/tcg/helper.h
@@ -622,6 +622,9 @@ DEF_HELPER_FLAGS_5(gvec_sdot_4h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32)
 DEF_HELPER_FLAGS_5(gvec_udot_4h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32)
 DEF_HELPER_FLAGS_5(gvec_usdot_4b, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32)
 
+DEF_HELPER_FLAGS_5(gvec_sdot_2h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_5(gvec_udot_2h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32)
+
 DEF_HELPER_FLAGS_5(gvec_sdot_idx_4b, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, ptr, i32)
 DEF_HELPER_FLAGS_5(gvec_udot_idx_4b, TCG_CALL_NO_RWG,
@@ -635,6 +638,11 @@ DEF_HELPER_FLAGS_5(gvec_sudot_idx_4b, TCG_CALL_NO_RWG,
 DEF_HELPER_FLAGS_5(gvec_usdot_idx_4b, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, ptr, i32)
 
+DEF_HELPER_FLAGS_5(gvec_sdot_idx_2h, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_5(gvec_udot_idx_2h, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
+
 DEF_HELPER_FLAGS_5(gvec_fcaddh, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, fpst, i32)
 DEF_HELPER_FLAGS_5(gvec_fcadds, TCG_CALL_NO_RWG,
diff --git a/target/arm/tcg/translate-sme.c b/target/arm/tcg/translate-sme.c
index 410a8d037c..341f4495e9 100644
--- a/target/arm/tcg/translate-sme.c
+++ b/target/arm/tcg/translate-sme.c
@@ -962,3 +962,88 @@ static bool do_vdot(DisasContext *s, arg_azx_n *a, gen_helper_gvec_4_ptr *fn)
 
 TRANS_FEAT(FVDOT, aa64_sme, do_vdot, a, gen_helper_sme2_fvdot_idx_h)
 TRANS_FEAT(BFVDOT, aa64_sme, do_vdot, a, gen_helper_sme2_bfvdot_idx)
+
+/*
+ * Expand array multi-vector single (n1), array multi-vector (nn),
+ * and array multi-vector indexed (nx), for integer accumulate.
+ *   multi: true for nn, false for n1.
+ *   data: stuff for simd_data, including any index.
+ */
+static bool do_azz_acc(DisasContext *s, int nreg, int nsel,
+                       int rv, int off, int zn, int zm,
+                       int data, int shsel, bool multi,
+                       gen_helper_gvec_4 *fn)
+{
+    if (sme_smza_enabled_check(s)) {
+        int svl = streaming_vec_reg_size(s);
+        int vstride = svl / nreg;
+        TCGv_ptr t_za = get_zarray(s, rv, off, nreg, nsel);
+        TCGv_ptr t = tcg_temp_new_ptr();
+
+        for (int r = 0; r < nreg; ++r) {
+            TCGv_ptr t_zn = vec_full_reg_ptr(s, zn);
+            TCGv_ptr t_zm = vec_full_reg_ptr(s, zm);
+
+            for (int i = 0; i < nsel; ++i) {
+                int o_za = (r * vstride + i) * sizeof(ARMVectorReg);
+                int desc = simd_desc(svl, svl, data | (i << shsel));
+
+                tcg_gen_addi_ptr(t, t_za, o_za);
+                fn(t, t_zn, t_zm, t, tcg_constant_i32(desc));
+            }
+
+            /*
+             * For multiple-and-single vectors, Zn may wrap.
+             * For multiple vectors, both Zn and Zm are aligned.
+             */
+            zn = (zn + 1) % 32;
+            zm += multi;
+        }
+    }
+    return true;
+}
+
+static bool do_dot(DisasContext *s, arg_azz_n *a, bool multi,
+                   gen_helper_gvec_4 *fn)
+{
+    return do_azz_acc(s, a->n, 1, a->rv, a->off, a->zn, a->zm,
+                      0, 0, multi, fn);
+}
+
+static void gen_helper_gvec_sudot_4b(TCGv_ptr d, TCGv_ptr n, TCGv_ptr m,
+                                     TCGv_ptr a, TCGv_i32 desc)
+{
+    gen_helper_gvec_usdot_4b(d, m, n, a, desc);
+}
+
+TRANS_FEAT(USDOT_n1, aa64_sme2, do_dot, a, false, gen_helper_gvec_usdot_4b)
+TRANS_FEAT(SUDOT_n1, aa64_sme2, do_dot, a, false, gen_helper_gvec_sudot_4b)
+TRANS_FEAT(SDOT_n1_2h, aa64_sme2, do_dot, a, false, gen_helper_gvec_sdot_2h)
+TRANS_FEAT(UDOT_n1_2h, aa64_sme2, do_dot, a, false, gen_helper_gvec_udot_2h)
+TRANS_FEAT(SDOT_n1_4b, aa64_sme2, do_dot, a, false, gen_helper_gvec_sdot_4b)
+TRANS_FEAT(UDOT_n1_4b, aa64_sme2, do_dot, a, false, gen_helper_gvec_udot_4b)
+TRANS_FEAT(SDOT_n1_4h, aa64_sme2_i16i64, do_dot, a, false, gen_helper_gvec_sdot_4h)
+TRANS_FEAT(UDOT_n1_4h, aa64_sme2_i16i64, do_dot, a, false, gen_helper_gvec_udot_4h)
+
+TRANS_FEAT(USDOT_nn, aa64_sme2, do_dot, a, true, gen_helper_gvec_usdot_4b)
+TRANS_FEAT(SDOT_nn_2h, aa64_sme2, do_dot, a, true, gen_helper_gvec_sdot_2h)
+TRANS_FEAT(UDOT_nn_2h, aa64_sme2, do_dot, a, true, gen_helper_gvec_udot_2h)
+TRANS_FEAT(SDOT_nn_4b, aa64_sme2, do_dot, a, true, gen_helper_gvec_sdot_4b)
+TRANS_FEAT(UDOT_nn_4b, aa64_sme2, do_dot, a, true, gen_helper_gvec_udot_4b)
+TRANS_FEAT(SDOT_nn_4h, aa64_sme2_i16i64, do_dot, a, true, gen_helper_gvec_sdot_4h)
+TRANS_FEAT(UDOT_nn_4h, aa64_sme2_i16i64, do_dot, a, true, gen_helper_gvec_udot_4h)
+
+static bool do_dot_nx(DisasContext *s, arg_azx_n *a, gen_helper_gvec_4 *fn)
+{
+    return do_azz_acc(s, a->n, 1, a->rv, a->off, a->zn, a->zm,
+                      a->idx, 0, false, fn);
+}
+
+TRANS_FEAT(USDOT_nx, aa64_sme2, do_dot_nx, a, gen_helper_gvec_usdot_idx_4b)
+TRANS_FEAT(SUDOT_nx, aa64_sme2, do_dot_nx, a, gen_helper_gvec_sudot_idx_4b)
+TRANS_FEAT(SDOT_nx_2h, aa64_sme2, do_dot_nx, a, gen_helper_gvec_sdot_idx_2h)
+TRANS_FEAT(UDOT_nx_2h, aa64_sme2, do_dot_nx, a, gen_helper_gvec_udot_idx_2h)
+TRANS_FEAT(SDOT_nx_4b, aa64_sme2, do_dot_nx, a, gen_helper_gvec_sdot_idx_4b)
+TRANS_FEAT(UDOT_nx_4b, aa64_sme2, do_dot_nx, a, gen_helper_gvec_udot_idx_4b)
+TRANS_FEAT(SDOT_nx_4h, aa64_sme2_i16i64, do_dot_nx, a, gen_helper_gvec_sdot_idx_4h)
+TRANS_FEAT(UDOT_nx_4h, aa64_sme2_i16i64, do_dot_nx, a, gen_helper_gvec_udot_idx_4h)
diff --git a/target/arm/tcg/vec_helper.c b/target/arm/tcg/vec_helper.c
index 4fedaa7293..c9d1b09268 100644
--- a/target/arm/tcg/vec_helper.c
+++ b/target/arm/tcg/vec_helper.c
@@ -872,6 +872,57 @@ DO_DOT_IDX(gvec_usdot_idx_4b, int32_t, uint8_t, int8_t, H4)
 DO_DOT_IDX(gvec_sdot_idx_4h, int64_t, int16_t, int16_t, H8)
 DO_DOT_IDX(gvec_udot_idx_4h, uint64_t, uint16_t, uint16_t, H8)
 
+#undef DO_DOT
+#undef DO_DOT_IDX
+
+/* Similar for 2-way dot product */
+#define DO_DOT(NAME, TYPED, TYPEN, TYPEM) \
+void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc)  \
+{                                                                         \
+    intptr_t i, opr_sz = simd_oprsz(desc);                                \
+    TYPED *d = vd, *a = va;                                               \
+    TYPEN *n = vn;                                                        \
+    TYPEM *m = vm;                                                        \
+    for (i = 0; i < opr_sz / sizeof(TYPED); ++i) {                        \
+        d[i] = (a[i] +                                                    \
+                (TYPED)n[i * 2 + 0] * m[i * 2 + 0] +                      \
+                (TYPED)n[i * 2 + 1] * m[i * 2 + 1]);                      \
+    }                                                                     \
+    clear_tail(d, opr_sz, simd_maxsz(desc));                              \
+}
+
+#define DO_DOT_IDX(NAME, TYPED, TYPEN, TYPEM, HD) \
+void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc)  \
+{                                                                         \
+    intptr_t i = 0, opr_sz = simd_oprsz(desc);                            \
+    intptr_t opr_sz_n = opr_sz / sizeof(TYPED);                           \
+    intptr_t segend = MIN(16 / sizeof(TYPED), opr_sz_n);                  \
+    intptr_t index = simd_data(desc);                                     \
+    TYPED *d = vd, *a = va;                                               \
+    TYPEN *n = vn;                                                        \
+    TYPEM *m_indexed = (TYPEM *)vm + HD(index) * 2;                       \
+    do {                                                                  \
+        TYPED m0 = m_indexed[i * 2 + 0];                                  \
+        TYPED m1 = m_indexed[i * 2 + 1];                                  \
+        do {                                                              \
+            d[i] = (a[i] +                                                \
+                    n[i * 2 + 0] * m0 +                                   \
+                    n[i * 2 + 1] * m1);                                   \
+        } while (++i < segend);                                           \
+        segend = i + (16 / sizeof(TYPED));                                \
+    } while (i < opr_sz_n);                                               \
+    clear_tail(d, opr_sz, simd_maxsz(desc));                              \
+}
+
+DO_DOT(gvec_sdot_2h, int32_t, int16_t, int16_t)
+DO_DOT(gvec_udot_2h, uint32_t, uint16_t, uint16_t)
+
+DO_DOT_IDX(gvec_sdot_idx_2h, int32_t, int16_t, int16_t, H4)
+DO_DOT_IDX(gvec_udot_idx_2h, uint32_t, uint16_t, uint16_t, H4)
+
+#undef DO_DOT
+#undef DO_DOT_IDX
+
 void HELPER(gvec_fcaddh)(void *vd, void *vn, void *vm,
                          float_status *fpst, uint32_t desc)
 {
diff --git a/target/arm/tcg/sme.decode b/target/arm/tcg/sme.decode
index 7c057bcad2..338637decd 100644
--- a/target/arm/tcg/sme.decode
+++ b/target/arm/tcg/sme.decode
@@ -291,6 +291,26 @@ FDOT_n1         11000001 001 1 .... 0 .. 100 ..... 00 ...   @azz_nx1_o3 n=4
 BFDOT_n1        11000001 001 0 .... 0 .. 100 ..... 10 ...   @azz_nx1_o3 n=2
 BFDOT_n1        11000001 001 1 .... 0 .. 100 ..... 10 ...   @azz_nx1_o3 n=4
 
+USDOT_n1        11000001 001 0 .... 0 .. 101 ..... 01 ...   @azz_nx1_o3 n=2
+USDOT_n1        11000001 001 1 .... 0 .. 101 ..... 01 ...   @azz_nx1_o3 n=4
+
+SUDOT_n1        11000001 001 0 .... 0 .. 101 ..... 11 ...   @azz_nx1_o3 n=2
+SUDOT_n1        11000001 001 1 .... 0 .. 101 ..... 11 ...   @azz_nx1_o3 n=4
+
+SDOT_n1_4b      11000001 001 0 .... 0 .. 101 ..... 00 ...   @azz_nx1_o3 n=2
+SDOT_n1_4b      11000001 001 1 .... 0 .. 101 ..... 00 ...   @azz_nx1_o3 n=4
+SDOT_n1_4h      11000001 011 0 .... 0 .. 101 ..... 00 ...   @azz_nx1_o3 n=2
+SDOT_n1_4h      11000001 011 1 .... 0 .. 101 ..... 00 ...   @azz_nx1_o3 n=4
+SDOT_n1_2h      11000001 011 0 .... 0 .. 101 ..... 01 ...   @azz_nx1_o3 n=2
+SDOT_n1_2h      11000001 011 1 .... 0 .. 101 ..... 01 ...   @azz_nx1_o3 n=4
+
+UDOT_n1_4b      11000001 001 0 .... 0 .. 101 ..... 10 ...   @azz_nx1_o3 n=2
+UDOT_n1_4b      11000001 001 1 .... 0 .. 101 ..... 10 ...   @azz_nx1_o3 n=4
+UDOT_n1_4h      11000001 011 0 .... 0 .. 101 ..... 10 ...   @azz_nx1_o3 n=2
+UDOT_n1_4h      11000001 011 1 .... 0 .. 101 ..... 10 ...   @azz_nx1_o3 n=4
+UDOT_n1_2h      11000001 011 0 .... 0 .. 101 ..... 11 ...   @azz_nx1_o3 n=2
+UDOT_n1_2h      11000001 011 1 .... 0 .. 101 ..... 11 ...   @azz_nx1_o3 n=4
+
 ### SME2 Multi-vector Multiple Array Vectors
 
 %zn_ax2         6:4 !function=times_2
@@ -334,6 +354,23 @@ FDOT_nn         11000001 101 ...01 0 .. 100 ...00 00 ...    @azz_4x4_o3
 BFDOT_nn        11000001 101 ....0 0 .. 100 ....0 10 ...    @azz_2x2_o3
 BFDOT_nn        11000001 101 ...01 0 .. 100 ...00 10 ...    @azz_4x4_o3
 
+USDOT_nn        11000001 101 ....0 0 .. 101 ....0 01 ...    @azz_2x2_o3
+USDOT_nn        11000001 101 ...01 0 .. 101 ...00 01 ...    @azz_4x4_o3
+
+SDOT_nn_4b      11000001 101 ....0 0 .. 101 ....0 00 ...    @azz_2x2_o3
+SDOT_nn_4b      11000001 101 ...01 0 .. 101 ...00 00 ...    @azz_4x4_o3
+SDOT_nn_4h      11000001 111 ....0 0 .. 101 ....0 00 ...    @azz_2x2_o3
+SDOT_nn_4h      11000001 111 ...01 0 .. 101 ...00 00 ...    @azz_4x4_o3
+SDOT_nn_2h      11000001 111 ....0 0 .. 101 ....0 01 ...    @azz_2x2_o3
+SDOT_nn_2h      11000001 111 ...01 0 .. 101 ...00 01 ...    @azz_4x4_o3
+
+UDOT_nn_4b      11000001 101 ....0 0 .. 101 ....0 10 ...    @azz_2x2_o3
+UDOT_nn_4b      11000001 101 ...01 0 .. 101 ...00 10 ...    @azz_4x4_o3
+UDOT_nn_4h      11000001 111 ....0 0 .. 101 ....0 10 ...    @azz_2x2_o3
+UDOT_nn_4h      11000001 111 ...01 0 .. 101 ...00 10 ...    @azz_4x4_o3
+UDOT_nn_2h      11000001 111 ....0 0 .. 101 ....0 11 ...    @azz_2x2_o3
+UDOT_nn_2h      11000001 111 ...01 0 .. 101 ...00 11 ...    @azz_4x4_o3
+
 ### SME2 Multi-vector Indexed
 
 &azx_n          n off rv zn zm idx
@@ -367,7 +404,11 @@ BFMLSL_nx       11000001 1001 .... 1 .. 1 .. ...00 11 ...   @azx_4x1_o2x2
 @azx_2x1_i2_o3  ........ .... zm:4 . .. . idx:2 .... ... off:3 \
                 &azx_n n=2 rv=%mova_rv zn=%zn_ax2
 @azx_4x1_i2_o3  ........ .... zm:4 . .. . idx:2 .... ... off:3 \
-                &azx_n n=2 rv=%mova_rv zn=%zn_ax4
+                &azx_n n=4 rv=%mova_rv zn=%zn_ax4
+@azx_2x1_i1_o3  ........ .... zm:4 . .. .. idx:1 .... ... off:3 \
+                &azx_n n=2 rv=%mova_rv zn=%zn_ax2
+@azx_4x1_i1_o3  ........ .... zm:4 . .. .. idx:1 .... ... off:3 \
+                &azx_n n=4 rv=%mova_rv zn=%zn_ax4
 
 FDOT_nx         11000001 0101 .... 0 .. 1 .. ....0 01 ...   @azx_2x1_i2_o3
 FDOT_nx         11000001 0101 .... 1 .. 1 .. ...00 01 ...   @azx_4x1_i2_o3
@@ -377,3 +418,23 @@ BFDOT_nx        11000001 0101 .... 1 .. 1 .. ...00 11 ...   @azx_4x1_i2_o3
 
 FVDOT           11000001 0101 .... 0 .. 0 .. ....0 01 ...   @azx_2x1_i2_o3
 BFVDOT          11000001 0101 .... 0 .. 0 .. ....0 11 ...   @azx_2x1_i2_o3
+
+SDOT_nx_2h      11000001 0101 .... 0 .. 1 .. ....0 00 ...   @azx_2x1_i2_o3
+SDOT_nx_2h      11000001 0101 .... 1 .. 1 .. ...00 00 ...   @azx_4x1_i2_o3
+SDOT_nx_4b      11000001 0101 .... 0 .. 1 .. ....1 00 ...   @azx_2x1_i2_o3
+SDOT_nx_4b      11000001 0101 .... 1 .. 1 .. ...01 00 ...   @azx_4x1_i2_o3
+SDOT_nx_4h      11000001 1101 .... 0 .. 00 . ....0 01 ...   @azx_2x1_i1_o3
+SDOT_nx_4h      11000001 1101 .... 1 .. 00 . ...00 01 ...   @azx_4x1_i1_o3
+
+UDOT_nx_2h      11000001 0101 .... 0 .. 1 .. ....0 10 ...   @azx_2x1_i2_o3
+UDOT_nx_2h      11000001 0101 .... 1 .. 1 .. ...00 10 ...   @azx_4x1_i2_o3
+UDOT_nx_4b      11000001 0101 .... 0 .. 1 .. ....1 10 ...   @azx_2x1_i2_o3
+UDOT_nx_4b      11000001 0101 .... 1 .. 1 .. ...01 10 ...   @azx_4x1_i2_o3
+UDOT_nx_4h      11000001 1101 .... 0 .. 00 . ....0 11 ...   @azx_2x1_i1_o3
+UDOT_nx_4h      11000001 1101 .... 1 .. 00 . ...00 11 ...   @azx_4x1_i1_o3
+
+USDOT_nx        11000001 0101 .... 0 .. 1 .. ....1 01 ...   @azx_2x1_i2_o3
+USDOT_nx        11000001 0101 .... 1 .. 1 .. ...01 01 ...   @azx_4x1_i2_o3
+
+SUDOT_nx        11000001 0101 .... 0 .. 1 .. ....1 11 ...   @azx_2x1_i2_o3
+SUDOT_nx        11000001 0101 .... 1 .. 1 .. ...01 11 ...   @azx_4x1_i2_o3
-- 
2.43.0

On Wed, 2 Jul 2025 at 13:34, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>



> +/* Similar for 2-way dot product */
> +#define DO_DOT(NAME, TYPED, TYPEN, TYPEM) \
> +void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc)  \
> +{                                                                         \
> +    intptr_t i, opr_sz = simd_oprsz(desc);                                \
> +    TYPED *d = vd, *a = va;                                               \
> +    TYPEN *n = vn;                                                        \
> +    TYPEM *m = vm;                                                        \
> +    for (i = 0; i < opr_sz / sizeof(TYPED); ++i) {                        \
> +        d[i] = (a[i] +                                                    \
> +                (TYPED)n[i * 2 + 0] * m[i * 2 + 0] +                      \
> +                (TYPED)n[i * 2 + 1] * m[i * 2 + 1]);                      \

Don't we need some H macros here for the big-endian host case?
(For that matter, the existing 4-way dot product helpers also
look like they won't work on big-endian...)

> +    }                                                                     \
> +    clear_tail(d, opr_sz, simd_maxsz(desc));                              \
> +}
> +
> +#define DO_DOT_IDX(NAME, TYPED, TYPEN, TYPEM, HD) \
> +void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc)  \
> +{                                                                         \
> +    intptr_t i = 0, opr_sz = simd_oprsz(desc);                            \
> +    intptr_t opr_sz_n = opr_sz / sizeof(TYPED);                           \
> +    intptr_t segend = MIN(16 / sizeof(TYPED), opr_sz_n);                  \
> +    intptr_t index = simd_data(desc);                                     \
> +    TYPED *d = vd, *a = va;                                               \
> +    TYPEN *n = vn;                                                        \
> +    TYPEM *m_indexed = (TYPEM *)vm + HD(index) * 2;                       \
> +    do {                                                                  \
> +        TYPED m0 = m_indexed[i * 2 + 0];                                  \
> +        TYPED m1 = m_indexed[i * 2 + 1];                                  \
> +        do {                                                              \
> +            d[i] = (a[i] +                                                \
> +                    n[i * 2 + 0] * m0 +                                   \
> +                    n[i * 2 + 1] * m1);                                   \

Similarly here.

> +        } while (++i < segend);                                           \
> +        segend = i + (16 / sizeof(TYPED));                                \
> +    } while (i < opr_sz_n);                                               \
> +    clear_tail(d, opr_sz, simd_maxsz(desc));                              \
> +}

Otherwise
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>

thanks
-- PMM

On 7/3/25 03:45, Peter Maydell wrote:
> On Wed, 2 Jul 2025 at 13:34, Richard Henderson
> <richard.henderson@linaro.org> wrote:
>>
>> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> 
> 
> 
>> +/* Similar for 2-way dot product */
>> +#define DO_DOT(NAME, TYPED, TYPEN, TYPEM) \
>> +void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc)  \
>> +{                                                                         \
>> +    intptr_t i, opr_sz = simd_oprsz(desc);                                \
>> +    TYPED *d = vd, *a = va;                                               \
>> +    TYPEN *n = vn;                                                        \
>> +    TYPEM *m = vm;                                                        \
>> +    for (i = 0; i < opr_sz / sizeof(TYPED); ++i) {                        \
>> +        d[i] = (a[i] +                                                    \
>> +                (TYPED)n[i * 2 + 0] * m[i * 2 + 0] +                      \
>> +                (TYPED)n[i * 2 + 1] * m[i * 2 + 1]);                      \
> 
> Don't we need some H macros here for the big-endian host case?
> (For that matter, the existing 4-way dot product helpers also
> look like they won't work on big-endian...)

The logic here is that all columns are treated identically.

...a0... ...a1...
.n0..n1. .n2..n3.
.m0..m1. .m2..m3.

vs

...a1... ...a0...
.n3..n2. .n1..n0.
.m3..m2. .m1..m0.

d0 = a0 + n0 * m0 + n1 * m1 -- it doesn't matter if n0 or n1 is at the lowest or highest 
address, because it still gets multiplied by the corresponding element in m, and then the 
two products are added to the sum that is addressed the same way.

The existing 4-way dot product uses the same endian independent logic, fwiw.


r~