SIMD_INLINE uint32x4_t SquaredDifferenceSumMasked(const uint8x16_t & a, const uint8x16_t & b, const uint8x16_t & mask)
{
uint8x16_t ad = vandq_u8(vabdq_u8(a, b), mask);
uint16x8_t lo = Square(vget_low_u8(ad));
uint16x8_t hi = Square(vget_high_u8(ad));
return vaddq_u32(vpaddlq_u16(lo), vpaddlq_u16(hi));
}
static INLINE unsigned int horizontal_add_u16x8(const uint16x8_t v_16x8) {
const uint32x4_t a = vpaddlq_u16(v_16x8);
const uint64x2_t b = vpaddlq_u32(a);
const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
vreinterpret_u32_u64(vget_high_u64(b)));
return vget_lane_u32(c, 0);
}
static uint64_t SharpYUVUpdateY_NEON(const uint16_t* ref, const uint16_t* src,
uint16_t* dst, int len) {
int i;
const int16x8_t zero = vdupq_n_s16(0);
const int16x8_t max = vdupq_n_s16(MAX_Y);
uint64x2_t sum = vdupq_n_u64(0);
uint64_t diff;
for (i = 0; i + 8 <= len; i += 8) {
const int16x8_t A = vreinterpretq_s16_u16(vld1q_u16(ref + i));
const int16x8_t B = vreinterpretq_s16_u16(vld1q_u16(src + i));
const int16x8_t C = vreinterpretq_s16_u16(vld1q_u16(dst + i));
const int16x8_t D = vsubq_s16(A, B); // diff_y
const int16x8_t F = vaddq_s16(C, D); // new_y
const uint16x8_t H =
vreinterpretq_u16_s16(vmaxq_s16(vminq_s16(F, max), zero));
const int16x8_t I = vabsq_s16(D); // abs(diff_y)
vst1q_u16(dst + i, H);
sum = vpadalq_u32(sum, vpaddlq_u16(vreinterpretq_u16_s16(I)));
}
diff = vgetq_lane_u64(sum, 0) + vgetq_lane_u64(sum, 1);
for (; i < len; ++i) {
const int diff_y = ref[i] - src[i];
const int new_y = (int)(dst[i]) + diff_y;
dst[i] = clip_y(new_y);
diff += (uint64_t)(abs(diff_y));
}
return diff;
}
inline int v_signmask(const v_uint16x8& a)
{
int16x4_t m0 = vcreate_s16(CV_BIG_UINT(0x0003000200010000));
uint16x8_t v0 = vshlq_u16(vshrq_n_u16(a.val, 15), vcombine_s16(m0, m0));
uint64x2_t v1 = vpaddlq_u32(vpaddlq_u16(v0));
return (int)vgetq_lane_u64(v1, 0) + ((int)vgetq_lane_u64(v1, 1) << 4);
}
inline int v_signmask(const v_uint8x16& a)
{
int8x8_t m0 = vcreate_s8(CV_BIG_UINT(0x0706050403020100));
uint8x16_t v0 = vshlq_u8(vshrq_n_u8(a.val, 7), vcombine_s8(m0, m0));
uint64x2_t v1 = vpaddlq_u32(vpaddlq_u16(vpaddlq_u8(v0)));
return (int)vgetq_lane_u64(v1, 0) + ((int)vgetq_lane_u64(v1, 1) << 8);
}
uint32x2_t FORCE_INLINE popcnt_neon_qreg(const uint8x16_t reg) {
const uint8x16_t pcnt = vcntq_u8(reg);
const uint16x8_t t0 = vpaddlq_u8(pcnt);
const uint32x4_t t1 = vpaddlq_u16(t0);
const uint32x2_t t2 = vadd_u32(vget_low_u32(t1), vget_high_u32(t1));
return t2;
}
inline ResultType operator()(Iterator1 a, Iterator2 b, size_t size) const
{
ResultType result = 0;
#if (defined __GNUC__ || defined __clang__) && defined USE_SSE
#ifdef __ARM_NEON__
{
uint32x4_t bits = vmovq_n_u32(0);
for (size_t i = 0; i < size; i += 16) {
uint8x16_t A_vec = vld1q_u8 (a + i);
uint8x16_t B_vec = vld1q_u8 (b + i);
uint8x16_t AxorB = veorq_u8 (A_vec, B_vec);
uint8x16_t bitsSet = vcntq_u8 (AxorB);
uint16x8_t bitSet8 = vpaddlq_u8 (bitsSet);
uint32x4_t bitSet4 = vpaddlq_u16 (bitSet8);
bits = vaddq_u32(bits, bitSet4);
}
uint64x2_t bitSet2 = vpaddlq_u32 (bits);
result = vgetq_lane_s32 (vreinterpretq_s32_u64(bitSet2),0);
result += vgetq_lane_s32 (vreinterpretq_s32_u64(bitSet2),2);
}
#else
{
//for portability just use unsigned long -- and use the __builtin_popcountll (see docs for __builtin_popcountll)
typedef unsigned long long pop_t;
const size_t modulo = size % sizeof(pop_t);
const pop_t* a2 = reinterpret_cast<const pop_t*> (a);
const pop_t* b2 = reinterpret_cast<const pop_t*> (b);
const pop_t* a2_end = a2 + (size / sizeof(pop_t));
for (; a2 != a2_end; ++a2, ++b2) result += __builtin_popcountll((*a2) ^ (*b2));
if (modulo) {
//in the case where size is not dividable by sizeof(pop_t)
//need to mask off the bits at the end
pop_t a_final = 0, b_final = 0;
memcpy(&a_final, a2, modulo);
memcpy(&b_final, b2, modulo);
result += __builtin_popcountll(a_final ^ b_final);
}
}
#endif //NEON
return result;
#endif
#ifdef PLATFORM_64_BIT
if(size%64 == 0)
{
const uint64_t* pa = reinterpret_cast<const uint64_t*>(a);
const uint64_t* pb = reinterpret_cast<const uint64_t*>(b);
size /= (sizeof(uint64_t)/sizeof(unsigned char));
for(size_t i = 0; i < size; ++i, ++pa, ++pb ) {
result += popcnt64(*pa ^ *pb);
}
}
else
{
const uint32_t* pa = reinterpret_cast<const uint32_t*>(a);
const uint32_t* pb = reinterpret_cast<const uint32_t*>(b);
size /= (sizeof(uint32_t)/sizeof(unsigned char));
for(size_t i = 0; i < size; ++i, ++pa, ++pb ) {
result += popcnt32(*pa ^ *pb);
}
}
#else
const uint32_t* pa = reinterpret_cast<const uint32_t*>(a);
const uint32_t* pb = reinterpret_cast<const uint32_t*>(b);
size /= (sizeof(uint32_t)/sizeof(unsigned char));
for(size_t i = 0; i < size; ++i, ++pa, ++pb ) {
result += popcnt32(*pa ^ *pb);
}
#endif
return result;
}
f64 dotProduct(const Size2D &_size,
const u8 * src0Base, ptrdiff_t src0Stride,
const u8 * src1Base, ptrdiff_t src1Stride)
{
internal::assertSupportedConfiguration();
#ifdef CAROTENE_NEON
Size2D size(_size);
if (src0Stride == src1Stride &&
src0Stride == (ptrdiff_t)(size.width))
{
size.width *= size.height;
size.height = 1;
}
// It is possible to accumulate up to 66051 uchar multiplication results in uint32 without overflow
// We process 16 elements and accumulate two new elements per step. So we could handle 66051/2*16 elements
#define DOT_UINT_BLOCKSIZE 66050*8
f64 result = 0.0;
for (size_t row = 0; row < size.height; ++row)
{
const u8 * src0 = internal::getRowPtr(src0Base, src0Stride, row);
const u8 * src1 = internal::getRowPtr(src1Base, src1Stride, row);
size_t i = 0;
uint64x2_t ws = vmovq_n_u64(0);
while(i + 16 <= size.width)
{
size_t lim = std::min(i + DOT_UINT_BLOCKSIZE, size.width) - 16;
uint32x4_t s1 = vmovq_n_u32(0);
uint32x4_t s2 = vmovq_n_u32(0);
for (; i <= lim; i += 16)
{
internal::prefetch(src0 + i);
internal::prefetch(src1 + i);
uint8x16_t vs1 = vld1q_u8(src0 + i);
uint8x16_t vs2 = vld1q_u8(src1 + i);
uint16x8_t vdot1 = vmull_u8(vget_low_u8(vs1), vget_low_u8(vs2));
uint16x8_t vdot2 = vmull_u8(vget_high_u8(vs1), vget_high_u8(vs2));
s1 = vpadalq_u16(s1, vdot1);
s2 = vpadalq_u16(s2, vdot2);
}
ws = vpadalq_u32(ws, s1);
ws = vpadalq_u32(ws, s2);
}
if(i + 8 <= size.width)
{
uint8x8_t vs1 = vld1_u8(src0 + i);
uint8x8_t vs2 = vld1_u8(src1 + i);
ws = vpadalq_u32(ws, vpaddlq_u16(vmull_u8(vs1, vs2)));
i += 8;
}
result += (double)vget_lane_u64(vadd_u64(vget_low_u64(ws), vget_high_u64(ws)), 0);
for (; i < size.width; ++i)
result += s32(src0[i]) * s32(src1[i]);
}
return result;
#else
(void)_size;
(void)src0Base;
(void)src0Stride;
(void)src1Base;
(void)src1Stride;
return 0;
#endif
}
