[PATCH v4 77/78] target/arm: Implement SVE2 complex integer dot product

[Richard Henderson]

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/arm/helper-sve.h    | 10 ++++
 target/arm/sve.decode      |  9 ++++
 target/arm/sve_helper.c    | 99 ++++++++++++++++++++++++++++++++++++++
 target/arm/translate-sve.c | 17 +++++++
 4 files changed, 135 insertions(+)

diff --git a/target/arm/helper-sve.h b/target/arm/helper-sve.h
index 43ee4238de..955d14b7ac 100644
--- a/target/arm/helper-sve.h
+++ b/target/arm/helper-sve.h
@@ -2784,3 +2784,13 @@ DEF_HELPER_FLAGS_5(sve2_fcvtlt_sd, TCG_CALL_NO_RWG,
 DEF_HELPER_FLAGS_5(flogb_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32)
 DEF_HELPER_FLAGS_5(flogb_s, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32)
 DEF_HELPER_FLAGS_5(flogb_d, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_5(sve2_cdot_zzzz_s, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_5(sve2_cdot_zzzz_d, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_5(sve2_cdot_idx_s, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_5(sve2_cdot_idx_d, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
diff --git a/target/arm/sve.decode b/target/arm/sve.decode
index b5c53b4c40..289a3576c3 100644
--- a/target/arm/sve.decode
+++ b/target/arm/sve.decode
@@ -811,6 +811,9 @@ MUL_zzi         00100101 .. 110 000 110 ........ .....      
    @rdn_i8s
 DOT_zzzz        01000100 1 sz:1 0 rm:5 00000 u:1 rn:5 rd:5 \
                 ra=%reg_movprfx
 
+# SVE2 complex dot product (vectors)
+CDOT_zzzz       01000100 esz:2 0 rm:5 0001 rot:2 rn:5 rd:5  ra=%reg_movprfx
+
 #### SVE Multiply - Indexed
 
 # SVE integer dot product (indexed)
@@ -849,6 +852,12 @@ SQDMLSLB_zzxw_d 01000100 .. 1 ..... 0011.0 ..... .....     
     @rrxw_d
 SQDMLSLT_zzxw_s 01000100 .. 1 ..... 0011.1 ..... .....          @rrxw_s
 SQDMLSLT_zzxw_d 01000100 .. 1 ..... 0011.1 ..... .....          @rrxw_d
 
+# SVE2 complex integer dot product (indexed)
+CDOT_zzxw_s     01000100 10 1 index:2 rm:3 0100 rot:2 rn:5 rd:5 \
+                ra=%reg_movprfx
+CDOT_zzxw_d     01000100 11 1 index:1 rm:4 0100 rot:2 rn:5 rd:5 \
+                ra=%reg_movprfx
+
 # SVE2 complex integer multiply-add (indexed)
 CMLA_zzxz_h     01000100 10 1 index:2 rm:3 0110 rot:2 rn:5 rd:5 \
                 ra=%reg_movprfx
diff --git a/target/arm/sve_helper.c b/target/arm/sve_helper.c
index d08da9fdc1..0a583b9898 100644
--- a/target/arm/sve_helper.c
+++ b/target/arm/sve_helper.c
@@ -1592,6 +1592,105 @@ void HELPER(sve2_sqrdcmlah_idx_s)(void *vd, void *vn, 
void *vm,
     do_cmla_idx_s(vd, vn, vm, va, desc, do_sqrdcmlah_s);
 }
 
+/* Note N and M are 4 elements bundled into one unit. */
+static int32_t do_cdot_s(uint32_t n, uint32_t m, int32_t a,
+                         int sel_a, int sel_b, int sub_i)
+{
+    for (int i = 0; i <= 1; i++) {
+        int32_t elt1_r = (int8_t)(n >> (16 * i));
+        int32_t elt1_i = (int8_t)(n >> (16 * i + 8));
+        int32_t elt2_a = (int8_t)(m >> (16 * i + 8 * sel_a));
+        int32_t elt2_b = (int8_t)(m >> (16 * i + 8 * sel_b));
+
+        a += elt1_r * elt2_a + elt1_i * elt2_b * sub_i;
+    }
+    return a;
+}
+
+static int64_t do_cdot_d(uint64_t n, uint64_t m, int64_t a,
+                         int sel_a, int sel_b, int sub_i)
+{
+    for (int i = 0; i <= 1; i++) {
+        int64_t elt1_r = (int16_t)(n >> (32 * i + 0));
+        int64_t elt1_i = (int16_t)(n >> (32 * i + 16));
+        int64_t elt2_a = (int16_t)(m >> (32 * i + 16 * sel_a));
+        int64_t elt2_b = (int16_t)(m >> (32 * i + 16 * sel_b));
+
+        a += elt1_r * elt2_a + elt1_i * elt2_b * sub_i;
+    }
+    return a;
+}
+
+void HELPER(sve2_cdot_zzzz_s)(void *vd, void *vn, void *vm,
+                              void *va, uint32_t desc)
+{
+    int opr_sz = simd_oprsz(desc);
+    int rot = simd_data(desc);
+    int sel_a = rot & 1;
+    int sel_b = sel_a ^ 1;
+    int sub_i = (rot == 0 || rot == 3 ? -1 : 1);
+    uint32_t *d = vd, *n = vn, *m = vm, *a = va;
+
+    for (int e = 0; e < opr_sz / 4; e++) {
+        d[e] = do_cdot_s(n[e], m[e], a[e], sel_a, sel_b, sub_i);
+    }
+}
+
+void HELPER(sve2_cdot_zzzz_d)(void *vd, void *vn, void *vm,
+                              void *va, uint32_t desc)
+{
+    int opr_sz = simd_oprsz(desc);
+    int rot = simd_data(desc);
+    int sel_a = rot & 1;
+    int sel_b = sel_a ^ 1;
+    int sub_i = (rot == 0 || rot == 3 ? -1 : 1);
+    uint64_t *d = vd, *n = vn, *m = vm, *a = va;
+
+    for (int e = 0; e < opr_sz / 8; e++) {
+        d[e] = do_cdot_d(n[e], m[e], a[e], sel_a, sel_b, sub_i);
+    }
+}
+
+void HELPER(sve2_cdot_idx_s)(void *vd, void *vn, void *vm,
+                             void *va, uint32_t desc)
+{
+    int opr_sz = simd_oprsz(desc);
+    int rot = extract32(desc, SIMD_DATA_SHIFT, 2);
+    int idx = H4(extract32(desc, SIMD_DATA_SHIFT + 2, 2));
+    int sel_a = rot & 1;
+    int sel_b = sel_a ^ 1;
+    int sub_i = (rot == 0 || rot == 3 ? -1 : 1);
+    uint32_t *d = vd, *n = vn, *m = vm, *a = va;
+
+    for (int seg = 0; seg < opr_sz / 4; seg += 4) {
+        uint32_t seg_m = m[seg + idx];
+        for (int e = 0; e < 4; e++) {
+            d[seg + e] = do_cdot_s(n[seg + e], seg_m, a[seg + e],
+                                   sel_a, sel_b, sub_i);
+        }
+    }
+}
+
+void HELPER(sve2_cdot_idx_d)(void *vd, void *vn, void *vm,
+                             void *va, uint32_t desc)
+{
+    int seg, opr_sz = simd_oprsz(desc);
+    int rot = extract32(desc, SIMD_DATA_SHIFT, 2);
+    int idx = extract32(desc, SIMD_DATA_SHIFT + 2, 2);
+    int sel_a = rot & 1;
+    int sel_b = sel_a ^ 1;
+    int sub_i = (rot == 0 || rot == 3 ? -1 : 1);
+    uint64_t *d = vd, *n = vn, *m = vm, *a = va;
+
+    for (seg = 0; seg < opr_sz / 8; seg += 2) {
+        uint64_t seg_m = m[seg + idx];
+        for (int e = 0; e < 2; e++) {
+            d[seg + e] = do_cdot_d(n[seg + e], seg_m, a[seg + e],
+                                   sel_a, sel_b, sub_i);
+        }
+    }
+}
+
 #define DO_ZZXZ(NAME, TYPE, H, OP) \
 void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc) \
 {                                                                       \
diff --git a/target/arm/translate-sve.c b/target/arm/translate-sve.c
index af2c6fbb9e..3c44a0d182 100644
--- a/target/arm/translate-sve.c
+++ b/target/arm/translate-sve.c
@@ -4147,6 +4147,9 @@ DO_SVE2_RRXR_ROT(CMLA_zzxz_s, gen_helper_sve2_cmla_idx_s)
 DO_SVE2_RRXR_ROT(SQRDCMLAH_zzxz_h, gen_helper_sve2_sqrdcmlah_idx_h)
 DO_SVE2_RRXR_ROT(SQRDCMLAH_zzxz_s, gen_helper_sve2_sqrdcmlah_idx_s)
 
+DO_SVE2_RRXR_ROT(CDOT_zzxw_s, gen_helper_sve2_cdot_idx_s)
+DO_SVE2_RRXR_ROT(CDOT_zzxw_d, gen_helper_sve2_cdot_idx_d)
+
 #undef DO_SVE2_RRXR_ROT
 
 /*
@@ -8351,6 +8354,20 @@ static bool trans_CMLA_zzzz(DisasContext *s, 
arg_CMLA_zzzz *a)
     return true;
 }
 
+static bool trans_CDOT_zzzz(DisasContext *s, arg_CMLA_zzzz *a)
+{
+    if (!dc_isar_feature(aa64_sve2, s) || a->esz < MO_32) {
+        return false;
+    }
+    if (sve_access_check(s)) {
+        gen_helper_gvec_4 *fn = (a->esz == MO_32
+                                 ? gen_helper_sve2_cdot_zzzz_s
+                                 : gen_helper_sve2_cdot_zzzz_d);
+        gen_gvec_ool_zzzz(s, fn, a->rd, a->rn, a->rm, a->ra, a->rot);
+    }
+    return true;
+}
+
 static bool trans_SQRDCMLAH_zzzz(DisasContext *s, arg_SQRDCMLAH_zzzz *a)
 {
     static gen_helper_gvec_4 * const fns[] = {
-- 
2.25.1