diff options
Diffstat (limited to '')
| -rw-r--r-- | third_party/aom/aom_dsp/arm/mem_neon.h | 56 |
1 files changed, 49 insertions, 7 deletions
diff --git a/third_party/aom/aom_dsp/arm/mem_neon.h b/third_party/aom/aom_dsp/arm/mem_neon.h index 52c7a34e3e..32a462a186 100644 --- a/third_party/aom/aom_dsp/arm/mem_neon.h +++ b/third_party/aom/aom_dsp/arm/mem_neon.h @@ -56,17 +56,10 @@ static INLINE uint16x8x4_t vld1q_u16_x4(const uint16_t *ptr) { #elif defined(__GNUC__) && !defined(__clang__) // GCC 64-bit. #if __GNUC__ < 8 - static INLINE uint8x16x2_t vld1q_u8_x2(const uint8_t *ptr) { uint8x16x2_t res = { { vld1q_u8(ptr + 0 * 16), vld1q_u8(ptr + 1 * 16) } }; return res; } - -static INLINE uint16x8x4_t vld1q_u16_x4(const uint16_t *ptr) { - uint16x8x4_t res = { { vld1q_u16(ptr + 0 * 8), vld1q_u16(ptr + 1 * 8), - vld1q_u16(ptr + 2 * 8), vld1q_u16(ptr + 3 * 8) } }; - return res; -} #endif // __GNUC__ < 8 #if __GNUC__ < 9 @@ -76,6 +69,15 @@ static INLINE uint8x16x3_t vld1q_u8_x3(const uint8_t *ptr) { return res; } #endif // __GNUC__ < 9 + +// vld1q_u16_x4 is defined from GCC 8.5.0 and onwards. +#if ((__GNUC__ << 8) | __GNUC_MINOR__) < 0x805 +static INLINE uint16x8x4_t vld1q_u16_x4(const uint16_t *ptr) { + uint16x8x4_t res = { { vld1q_u16(ptr + 0 * 8), vld1q_u16(ptr + 1 * 8), + vld1q_u16(ptr + 2 * 8), vld1q_u16(ptr + 3 * 8) } }; + return res; +} +#endif // ((__GNUC__ << 8) | __GNUC_MINOR__) < 0x805 #endif // defined(__GNUC__) && !defined(__clang__) static INLINE void store_u8_8x2(uint8_t *s, ptrdiff_t p, const uint8x8_t s0, @@ -457,6 +459,16 @@ static INLINE void load_s16_4x4(const int16_t *s, ptrdiff_t p, *s3 = vld1_s16(s); } +static INLINE void load_s16_4x3(const int16_t *s, ptrdiff_t p, + int16x4_t *const s0, int16x4_t *const s1, + int16x4_t *const s2) { + *s0 = vld1_s16(s); + s += p; + *s1 = vld1_s16(s); + s += p; + *s2 = vld1_s16(s); +} + static INLINE void store_u8_8x8(uint8_t *s, ptrdiff_t p, const uint8x8_t s0, const uint8x8_t s1, const uint8x8_t s2, const uint8x8_t s3, const uint8x8_t s4, @@ -525,6 +537,16 @@ static INLINE void store_u16_8x8(uint16_t *s, ptrdiff_t dst_stride, vst1q_u16(s, s7); } +static INLINE void store_u16_4x3(uint16_t *s, ptrdiff_t dst_stride, + const uint16x4_t s0, const uint16x4_t s1, + const uint16x4_t s2) { + vst1_u16(s, s0); + s += dst_stride; + vst1_u16(s, s1); + s += dst_stride; + vst1_u16(s, s2); +} + static INLINE void store_u16_4x4(uint16_t *s, ptrdiff_t dst_stride, const uint16x4_t s0, const uint16x4_t s1, const uint16x4_t s2, const uint16x4_t s3) { @@ -544,6 +566,16 @@ static INLINE void store_u16_8x2(uint16_t *s, ptrdiff_t dst_stride, vst1q_u16(s, s1); } +static INLINE void store_u16_8x3(uint16_t *s, ptrdiff_t dst_stride, + const uint16x8_t s0, const uint16x8_t s1, + const uint16x8_t s2) { + vst1q_u16(s, s0); + s += dst_stride; + vst1q_u16(s, s1); + s += dst_stride; + vst1q_u16(s, s2); +} + static INLINE void store_u16_8x4(uint16_t *s, ptrdiff_t dst_stride, const uint16x8_t s0, const uint16x8_t s1, const uint16x8_t s2, const uint16x8_t s3) { @@ -857,6 +889,16 @@ static INLINE void load_s16_8x4(const int16_t *s, ptrdiff_t p, *s3 = vld1q_s16(s); } +static INLINE void load_s16_8x3(const int16_t *s, ptrdiff_t p, + int16x8_t *const s0, int16x8_t *const s1, + int16x8_t *const s2) { + *s0 = vld1q_s16(s); + s += p; + *s1 = vld1q_s16(s); + s += p; + *s2 = vld1q_s16(s); +} + // Load 2 sets of 4 bytes when alignment is not guaranteed. static INLINE uint8x8_t load_unaligned_u8(const uint8_t *buf, int stride) { uint32_t a; |