static WEBP_INLINE uint32_t ClampedAddSubtractFull(const uint32_t* const c0,
const uint32_t* const c1,
const uint32_t* const c2) {
const uint8x8_t p0 = vreinterpret_u8_u64(vcreate_u64(*c0));
const uint8x8_t p1 = vreinterpret_u8_u64(vcreate_u64(*c1));
const uint8x8_t p2 = vreinterpret_u8_u64(vcreate_u64(*c2));
const uint16x8_t sum0 = vaddl_u8(p0, p1); // add and widen
const uint16x8_t sum1 = vqsubq_u16(sum0, vmovl_u8(p2)); // widen and subtract
const uint8x8_t out = vqmovn_u16(sum1); // narrow and clamp
return vget_lane_u32(vreinterpret_u32_u8(out), 0);
}
static WEBP_INLINE uint32_t ClampedAddSubtractFull(const uint32_t* const c0,
const uint32_t* const c1,
const uint32_t* const c2) {
const uint64x1_t C0 = { *c0, 0 }, C1 = { *c1, 0 }, C2 = { *c2, 0 };
const uint8x8_t p0 = vreinterpret_u8_u64(C0);
const uint8x8_t p1 = vreinterpret_u8_u64(C1);
const uint8x8_t p2 = vreinterpret_u8_u64(C2);
const uint16x8_t sum0 = vaddl_u8(p0, p1); // add and widen
const uint16x8_t sum1 = vqsubq_u16(sum0, vmovl_u8(p2)); // widen and subtract
const uint8x8_t out = vqmovn_u16(sum1); // narrow and clamp
uint32_t ret;
vst1_lane_u32(&ret, vreinterpret_u32_u8(out), 0);
return ret;
}
/* u16x8 saturated sub */
void mw_neon_mm_qsub_u16x8(unsigned short * A, int Row, int Col, unsigned short * B, unsigned short * C)
{
uint16x8_t neon_a, neon_b, neon_c;
int size = Row * Col;
int i = 0;
int k = 0;
for (i = 8; i <= size ; i+=8)
{
k = i - 8;
neon_a = vld1q_u16(A + k);
neon_b = vld1q_u16(B + k);
neon_c = vqsubq_u16(neon_a, neon_b);
vst1q_u16(C + k, neon_c);
}
k = i - 8;
for (i = 0; i < size % 8; i++)
{
C[k + i] = A[k + i] + B[k + i];
}
}
inline uint16x8_t vqsubq(const uint16x8_t & v0, const uint16x8_t & v1) { return vqsubq_u16(v0, v1); }
