|
17 | 17 |
|
18 | 18 | #include <arm_neon.h> |
19 | 19 |
|
| 20 | +#if defined (__cplusplus) |
| 21 | +extern "C" { |
| 22 | +#endif |
| 23 | + |
| 24 | +static inline uint8x16_t __iso_volatile_neon_load128(const volatile uint8x16_t *ptr) |
| 25 | +{ |
| 26 | + return *ptr; |
| 27 | +} |
| 28 | + |
| 29 | +static inline uint8x16x2_t __iso_volatile_neon_load128_p(const volatile uint8x16x2_t *ptr) |
| 30 | +{ |
| 31 | + uint8x16x2_t result; |
| 32 | + __asm__ __volatile__( |
| 33 | + "ldp %q0, %q1, %2" |
| 34 | + : "=w" (result.val[0]), "=w" (result.val[1]) |
| 35 | + : "m" (*ptr) |
| 36 | + ); |
| 37 | + return result; |
| 38 | +} |
| 39 | + |
| 40 | +static inline uint8x8x2_t __iso_volatile_neon_load64_np(const volatile uint8x8x2_t *ptr) { |
| 41 | + uint8x8x2_t result; |
| 42 | + __asm__ __volatile__( |
| 43 | + "ldnp %d0, %d1, %2" |
| 44 | + : "=w" (result.val[0]), "=w" (result.val[1]) |
| 45 | + : "m" (*ptr) |
| 46 | + ); |
| 47 | + return result; |
| 48 | +} |
| 49 | + |
| 50 | +static inline uint8x16x2_t __iso_volatile_neon_load128_np(const volatile uint8x16x2_t *ptr) { |
| 51 | + uint8x16x2_t result; |
| 52 | + __asm__ __volatile__( |
| 53 | + "ldnp %q0, %q1, %2" |
| 54 | + : "=w" (result.val[0]), "=w" (result.val[1]) |
| 55 | + : "m" (*ptr) |
| 56 | + ); |
| 57 | + return result; |
| 58 | +} |
| 59 | + |
| 60 | +static inline void __iso_volatile_neon_store128(volatile uint8x16_t *p, uint8x16_t val) |
| 61 | +{ |
| 62 | + *p = val; |
| 63 | +} |
| 64 | + |
| 65 | +static inline void __iso_volatile_neon_store128_p(volatile uint8x16x2_t *ptr, uint8x16x2_t data) { |
| 66 | + __asm__ __volatile__( |
| 67 | + "stp %q[i1], %q[i2], %[out]" |
| 68 | + : [out] "=m" (*ptr) |
| 69 | + : [i1]"w" (data.val[0]), [i2]"w" (data.val[1]) |
| 70 | + ); |
| 71 | +} |
| 72 | + |
| 73 | +static inline void __iso_volatile_neon_store64_np(volatile uint8x8x2_t *ptr, uint8x8x2_t data) { |
| 74 | + __asm__ __volatile__( |
| 75 | + "stnp %d[i1], %d[i2], %[out]" |
| 76 | + : [out] "=m" (*ptr) |
| 77 | + : [i1]"w" (data.val[0]), [i2]"w" (data.val[1]) |
| 78 | + ); |
| 79 | +} |
| 80 | + |
| 81 | +static inline void __iso_volatile_neon_store128_np(volatile uint8x16x2_t *ptr, uint8x16x2_t data) { |
| 82 | + __asm__ __volatile__( |
| 83 | + "stnp %q[i1], %q[i2], %[out]" |
| 84 | + : [out] "=m" (*ptr) |
| 85 | + : [i1]"w" (data.val[0]), [i2]"w" (data.val[1]) |
| 86 | + ); |
| 87 | +} |
| 88 | + |
| 89 | +#if defined (__cplusplus) |
| 90 | +} |
| 91 | +#endif |
| 92 | + |
20 | 93 | #endif /* __ARM64_NEON_H */ |
21 | 94 | #endif /* _MSC_VER */ |
0 commit comments