1 /* { dg-do compile { target { aarch64*-*-* } } } */
2 /* { dg-additional-options "-O3 -march=armv8.2-a+dotprod" } */
3
4 #include <arm_neon.h>
5
6 /* Unsigned Dot Product instructions. */
7
ufoo(uint32x2_t r,uint8x8_t x,uint8x8_t y)8 uint32x2_t ufoo (uint32x2_t r, uint8x8_t x, uint8x8_t y)
9 {
10 return vdot_u32 (r, x, y);
11 }
12
ufooq(uint32x4_t r,uint8x16_t x,uint8x16_t y)13 uint32x4_t ufooq (uint32x4_t r, uint8x16_t x, uint8x16_t y)
14 {
15 return vdotq_u32 (r, x, y);
16 }
17
ufoo_lane(uint32x2_t r,uint8x8_t x,uint8x8_t y)18 uint32x2_t ufoo_lane (uint32x2_t r, uint8x8_t x, uint8x8_t y)
19 {
20 return vdot_lane_u32 (r, x, y, 0);
21 }
22
ufoo_laneq(uint32x2_t r,uint8x8_t x,uint8x16_t y)23 uint32x2_t ufoo_laneq (uint32x2_t r, uint8x8_t x, uint8x16_t y)
24 {
25 return vdot_laneq_u32 (r, x, y, 0);
26 }
27
ufooq_lane(uint32x4_t r,uint8x16_t x,uint8x8_t y)28 uint32x4_t ufooq_lane (uint32x4_t r, uint8x16_t x, uint8x8_t y)
29 {
30 return vdotq_lane_u32 (r, x, y, 0);
31 }
32
ufooq_laneq(uint32x4_t r,uint8x16_t x,uint8x16_t y)33 uint32x4_t ufooq_laneq (uint32x4_t r, uint8x16_t x, uint8x16_t y)
34 {
35 return vdotq_laneq_u32 (r, x, y, 0);
36 }
37
38 /* Signed Dot Product instructions. */
39
sfoo(int32x2_t r,int8x8_t x,int8x8_t y)40 int32x2_t sfoo (int32x2_t r, int8x8_t x, int8x8_t y)
41 {
42 return vdot_s32 (r, x, y);
43 }
44
sfooq(int32x4_t r,int8x16_t x,int8x16_t y)45 int32x4_t sfooq (int32x4_t r, int8x16_t x, int8x16_t y)
46 {
47 return vdotq_s32 (r, x, y);
48 }
49
sfoo_lane(int32x2_t r,int8x8_t x,int8x8_t y)50 int32x2_t sfoo_lane (int32x2_t r, int8x8_t x, int8x8_t y)
51 {
52 return vdot_lane_s32 (r, x, y, 0);
53 }
54
sfoo_laneq(int32x2_t r,int8x8_t x,int8x16_t y)55 int32x2_t sfoo_laneq (int32x2_t r, int8x8_t x, int8x16_t y)
56 {
57 return vdot_laneq_s32 (r, x, y, 0);
58 }
59
sfooq_lane(int32x4_t r,int8x16_t x,int8x8_t y)60 int32x4_t sfooq_lane (int32x4_t r, int8x16_t x, int8x8_t y)
61 {
62 return vdotq_lane_s32 (r, x, y, 0);
63 }
64
sfooq_laneq(int32x4_t r,int8x16_t x,int8x16_t y)65 int32x4_t sfooq_laneq (int32x4_t r, int8x16_t x, int8x16_t y)
66 {
67 return vdotq_laneq_s32 (r, x, y, 0);
68 }
69
70 /* { dg-final { scan-assembler-times {[us]dot\tv[0-9]+\.2s, v[0-9]+\.8b, v[0-9]+\.8b} 2 } } */
71 /* { dg-final { scan-assembler-times {[us]dot\tv[0-9]+\.2s, v[0-9]+\.8b, v[0-9]+\.4b\[[0-9]+\]} 4 } } */
72 /* { dg-final { scan-assembler-times {[us]dot\tv[0-9]+\.4s, v[0-9]+\.16b, v[0-9]+\.16b} 2 } } */
73 /* { dg-final { scan-assembler-times {[us]dot\tv[0-9]+\.4s, v[0-9]+\.16b, v[0-9]+\.4b\[[0-9]+\]} 4 } } */
74