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;
}
void test_vpadalQu32 (void)
{
uint64x2_t out_uint64x2_t;
uint64x2_t arg0_uint64x2_t;
uint32x4_t arg1_uint32x4_t;
out_uint64x2_t = vpadalq_u32 (arg0_uint64x2_t, arg1_uint32x4_t);
}
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
}
