void vpx_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
const uint8x8_t XABCD_u8 = vld1_u8(above - 1);
const uint64x1_t XABCD = vreinterpret_u64_u8(XABCD_u8);
const uint64x1_t ____XABC = vshl_n_u64(XABCD, 32);
const uint32x2_t zero = vdup_n_u32(0);
const uint32x2_t IJKL = vld1_lane_u32((const uint32_t *)left, zero, 0);
const uint8x8_t IJKL_u8 = vreinterpret_u8_u32(IJKL);
const uint64x1_t LKJI____ = vreinterpret_u64_u8(vrev32_u8(IJKL_u8));
const uint64x1_t LKJIXABC = vorr_u64(LKJI____, ____XABC);
const uint8x8_t KJIXABC_ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 8));
const uint8x8_t JIXABC__ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 16));
const uint8_t D = vget_lane_u8(XABCD_u8, 4);
const uint8x8_t JIXABCD_ = vset_lane_u8(D, JIXABC__, 6);
const uint8x8_t LKJIXABC_u8 = vreinterpret_u8_u64(LKJIXABC);
const uint8x8_t avg1 = vhadd_u8(JIXABCD_, LKJIXABC_u8);
const uint8x8_t avg2 = vrhadd_u8(avg1, KJIXABC_);
const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2);
const uint32x2_t r3 = vreinterpret_u32_u8(avg2);
const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8));
const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16));
const uint32x2_t r0 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24));
vst1_lane_u32((uint32_t *)(dst + 0 * stride), r0, 0);
vst1_lane_u32((uint32_t *)(dst + 1 * stride), r1, 0);
vst1_lane_u32((uint32_t *)(dst + 2 * stride), r2, 0);
vst1_lane_u32((uint32_t *)(dst + 3 * stride), r3, 0);
}
void vpx_idct4x4_1_add_neon(
int16_t *input,
uint8_t *dest,
int dest_stride) {
uint8x8_t d6u8;
uint32x2_t d2u32 = vdup_n_u32(0);
uint16x8_t q8u16;
int16x8_t q0s16;
uint8_t *d1, *d2;
int16_t i, a1, cospi_16_64 = 11585;
int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
out = dct_const_round_shift(out * cospi_16_64);
a1 = ROUND_POWER_OF_TWO(out, 4);
q0s16 = vdupq_n_s16(a1);
// dc_only_idct_add
d1 = d2 = dest;
for (i = 0; i < 2; i++) {
d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 0);
d1 += dest_stride;
d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 1);
d1 += dest_stride;
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q0s16),
vreinterpret_u8_u32(d2u32));
d6u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 0);
d2 += dest_stride;
vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 1);
d2 += dest_stride;
}
return;
}
static INLINE void idct4x4_1_add_kernel(uint8_t **dest, const int stride,
const int16x8_t res,
uint32x2_t *const d) {
uint16x8_t a;
uint8x8_t b;
*d = vld1_lane_u32((const uint32_t *)*dest, *d, 0);
*d = vld1_lane_u32((const uint32_t *)(*dest + stride), *d, 1);
a = vaddw_u8(vreinterpretq_u16_s16(res), vreinterpret_u8_u32(*d));
b = vqmovun_s16(vreinterpretq_s16_u16(a));
vst1_lane_u32((uint32_t *)*dest, vreinterpret_u32_u8(b), 0);
*dest += stride;
vst1_lane_u32((uint32_t *)*dest, vreinterpret_u32_u8(b), 1);
*dest += stride;
}
void test_vst1_laneu32 (void)
{
uint32_t *arg0_uint32_t;
uint32x2_t arg1_uint32x2_t;
vst1_lane_u32 (arg0_uint32_t, arg1_uint32x2_t, 1);
}
void
png_read_filter_row_sub3_neon(png_row_infop row_info, png_bytep row,
png_const_bytep prev_row)
{
png_bytep rp = row;
png_bytep rp_stop = row + row_info->rowbytes;
uint8x16_t vtmp = vld1q_u8(rp);
uint8x8x2_t *vrpt = png_ptr(uint8x8x2_t, &vtmp);
uint8x8x2_t vrp = *vrpt;
uint8x8x4_t vdest;
vdest.val[3] = vdup_n_u8(0);
png_debug(1, "in png_read_filter_row_sub3_neon");
for (; rp < rp_stop;)
{
uint8x8_t vtmp1, vtmp2;
uint32x2_t *temp_pointer;
vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
vdest.val[0] = vadd_u8(vdest.val[3], vrp.val[0]);
vtmp2 = vext_u8(vrp.val[0], vrp.val[1], 6);
vdest.val[1] = vadd_u8(vdest.val[0], vtmp1);
vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
vdest.val[2] = vadd_u8(vdest.val[1], vtmp2);
vdest.val[3] = vadd_u8(vdest.val[2], vtmp1);
vtmp = vld1q_u8(rp + 12);
vrpt = png_ptr(uint8x8x2_t, &vtmp);
vrp = *vrpt;
vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
rp += 3;
vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
rp += 3;
vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
rp += 3;
vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
rp += 3;
}
PNG_UNUSED(prev_row)
}
void vp9_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
int i;
uint32x2_t d0u32 = vdup_n_u32(0);
(void)left;
d0u32 = vld1_lane_u32((const uint32_t *)above, d0u32, 0);
for (i = 0; i < 4; i++, dst += stride)
vst1_lane_u32((uint32_t *)dst, d0u32, 0);
}
static WEBP_INLINE uint32_t Average2(const uint32_t* const a,
const uint32_t* const b) {
const uint64x1_t a0 = { *a }, b0 = { *b };
const uint8x8_t a1 = vreinterpret_u8_u64(a0);
const uint8x8_t b1 = vreinterpret_u8_u64(b0);
const uint8x8_t avg = vhadd_u8(a1, b1);
uint32_t ret;
vst1_lane_u32(&ret, vreinterpret_u32_u8(avg), 0);
return ret;
}
void vp9_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
uint8x8_t d0u8 = vdup_n_u8(0);
uint32x2_t d1u32 = vdup_n_u32(0);
(void)above;
d1u32 = vld1_lane_u32((const uint32_t *)left, d1u32, 0);
d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 0);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 1);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 2);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
dst += stride;
d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 3);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
}
void vpx_d45_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
const uint64x1_t A0 = vreinterpret_u64_u8(vld1_u8(above)); // top row
const uint64x1_t A1 = vshr_n_u64(A0, 8);
const uint64x1_t A2 = vshr_n_u64(A0, 16);
const uint8x8_t ABCDEFGH = vreinterpret_u8_u64(A0);
const uint8x8_t BCDEFGH0 = vreinterpret_u8_u64(A1);
const uint8x8_t CDEFGH00 = vreinterpret_u8_u64(A2);
const uint8x8_t avg1 = vhadd_u8(ABCDEFGH, CDEFGH00);
const uint8x8_t avg2 = vrhadd_u8(avg1, BCDEFGH0);
const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2);
const uint32x2_t r0 = vreinterpret_u32_u8(avg2);
const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8));
const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16));
const uint32x2_t r3 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24));
(void)left;
vst1_lane_u32((uint32_t *)(dst + 0 * stride), r0, 0);
vst1_lane_u32((uint32_t *)(dst + 1 * stride), r1, 0);
vst1_lane_u32((uint32_t *)(dst + 2 * stride), r2, 0);
vst1_lane_u32((uint32_t *)(dst + 3 * stride), r3, 0);
dst[3 * stride + 3] = above[7];
}
static WEBP_INLINE uint32_t Average3(const uint32_t* const a,
const uint32_t* const b,
const uint32_t* const c) {
const uint64x1_t a0 = { *a }, b0 = { *b }, c0 = { *c };
const uint8x8_t a1 = vreinterpret_u8_u64(a0);
const uint8x8_t b1 = vreinterpret_u8_u64(b0);
const uint8x8_t c1 = vreinterpret_u8_u64(c0);
const uint8x8_t avg1 = vhadd_u8(a1, c1);
const uint8x8_t avg2 = vhadd_u8(avg1, b1);
uint32_t ret;
vst1_lane_u32(&ret, vreinterpret_u32_u8(avg2), 0);
return ret;
}
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;
}
static WEBP_INLINE uint32_t ClampedAddSubtractHalf(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 uint8x8_t avg = vhadd_u8(p0, p1); // Average(c0,c1)
const uint8x8_t ab = vshr_n_u8(vqsub_u8(avg, p2), 1); // (a-b)>>1 saturated
const uint8x8_t ba = vshr_n_u8(vqsub_u8(p2, avg), 1); // (b-a)>>1 saturated
const uint8x8_t out = vqsub_u8(vqadd_u8(avg, ab), ba);
uint32_t ret;
vst1_lane_u32(&ret, vreinterpret_u32_u8(out), 0);
return ret;
}
void vp9_tm_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
int i;
uint16x8_t q1u16, q3u16;
int16x8_t q1s16;
uint8x8_t d0u8 = vdup_n_u8(0);
uint32x2_t d2u32 = vdup_n_u32(0);
d0u8 = vld1_dup_u8(above - 1);
d2u32 = vld1_lane_u32((const uint32_t *)above, d2u32, 0);
q3u16 = vsubl_u8(vreinterpret_u8_u32(d2u32), d0u8);
for (i = 0; i < 4; i++, dst += stride) {
q1u16 = vdupq_n_u16((uint16_t)left[i]);
q1s16 = vaddq_s16(vreinterpretq_s16_u16(q1u16),
vreinterpretq_s16_u16(q3u16));
d0u8 = vqmovun_s16(q1s16);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
}
}
static INLINE void scaledconvolve_vert_w4(
const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
const ptrdiff_t dst_stride, const InterpKernel *const y_filters,
const int y0_q4, const int y_step_q4, const int w, const int h) {
int y;
int y_q4 = y0_q4;
src -= src_stride * (SUBPEL_TAPS / 2 - 1);
y = h;
do {
const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
if (y_q4 & SUBPEL_MASK) {
const int16x8_t filters = vld1q_s16(y_filters[y_q4 & SUBPEL_MASK]);
const int16x4_t filter3 = vdup_lane_s16(vget_low_s16(filters), 3);
const int16x4_t filter4 = vdup_lane_s16(vget_high_s16(filters), 0);
uint8x8_t s[8], d;
int16x4_t t[8], tt;
load_u8_8x8(src_y, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
&s[6], &s[7]);
t[0] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[0])));
t[1] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[1])));
t[2] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[2])));
t[3] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[3])));
t[4] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[4])));
t[5] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[5])));
t[6] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[6])));
t[7] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[7])));
tt = convolve8_4(t[0], t[1], t[2], t[3], t[4], t[5], t[6], t[7], filters,
filter3, filter4);
d = vqrshrun_n_s16(vcombine_s16(tt, tt), 7);
vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d), 0);
} else {
memcpy(dst, &src_y[3 * src_stride], w);
}
dst += dst_stride;
y_q4 += y_step_q4;
} while (--y);
}
// 'do_above' and 'do_left' facilitate branch removal when inlined.
static INLINE void dc_4x4(uint8_t *dst, ptrdiff_t stride, const uint8_t *above,
const uint8_t *left, int do_above, int do_left) {
uint16x8_t sum_top;
uint16x8_t sum_left;
uint8x8_t dc0;
if (do_above) {
const uint8x8_t A = vld1_u8(above); // top row
const uint16x4_t p0 = vpaddl_u8(A); // cascading summation of the top
const uint16x4_t p1 = vpadd_u16(p0, p0);
sum_top = vcombine_u16(p1, p1);
}
if (do_left) {
const uint8x8_t L = vld1_u8(left); // left border
const uint16x4_t p0 = vpaddl_u8(L); // cascading summation of the left
const uint16x4_t p1 = vpadd_u16(p0, p0);
sum_left = vcombine_u16(p1, p1);
}
if (do_above && do_left) {
const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
dc0 = vrshrn_n_u16(sum, 3);
} else if (do_above) {
dc0 = vrshrn_n_u16(sum_top, 2);
} else if (do_left) {
dc0 = vrshrn_n_u16(sum_left, 2);
} else {
dc0 = vdup_n_u8(0x80);
}
{
const uint8x8_t dc = vdup_lane_u8(dc0, 0);
int i;
for (i = 0; i < 4; ++i) {
vst1_lane_u32((uint32_t *)(dst + i * stride), vreinterpret_u32_u8(dc), 0);
}
}
}
void
png_read_filter_row_paeth3_neon(png_row_infop row_info, png_bytep row,
png_const_bytep prev_row)
{
png_bytep rp = row;
png_const_bytep pp = prev_row;
png_bytep rp_stop = row + row_info->rowbytes;
uint8x16_t vtmp;
uint8x8x2_t *vrpt;
uint8x8x2_t vrp;
uint8x8_t vlast = vdup_n_u8(0);
uint8x8x4_t vdest;
vdest.val[3] = vdup_n_u8(0);
vtmp = vld1q_u8(rp);
vrpt = png_ptr(uint8x8x2_t,&vtmp);
vrp = *vrpt;
png_debug(1, "in png_read_filter_row_paeth3_neon");
for (; rp < rp_stop; pp += 12)
{
uint8x8x2_t *vppt;
uint8x8x2_t vpp;
uint8x8_t vtmp1, vtmp2, vtmp3;
uint32x2_t *temp_pointer;
vtmp = vld1q_u8(pp);
vppt = png_ptr(uint8x8x2_t,&vtmp);
vpp = *vppt;
vdest.val[0] = paeth(vdest.val[3], vpp.val[0], vlast);
vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 3);
vdest.val[1] = paeth(vdest.val[0], vtmp2, vpp.val[0]);
vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 6);
vtmp3 = vext_u8(vpp.val[0], vpp.val[1], 6);
vdest.val[2] = paeth(vdest.val[1], vtmp3, vtmp2);
vdest.val[2] = vadd_u8(vdest.val[2], vtmp1);
vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1);
vtmp = vld1q_u8(rp + 12);
vrpt = png_ptr(uint8x8x2_t,&vtmp);
vrp = *vrpt;
vdest.val[3] = paeth(vdest.val[2], vtmp2, vtmp3);
vdest.val[3] = vadd_u8(vdest.val[3], vtmp1);
vlast = vtmp2;
vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
rp += 3;
vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
rp += 3;
vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
rp += 3;
vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
rp += 3;
}
}
void computeNetwork0new_neon(const float *dataf, const float *weightsf, uint8_t *d) {
const int16_t *data = (const int16_t *)dataf;
const int16_t *weights = (const int16_t *)weightsf;
int32x4_t accum0 = { 0, 0, 0, 0 };
int32x4_t accum1 = accum0;
int32x4_t accum2 = accum0;
int32x4_t accum3 = accum0;
for (int i = 0; i < 128/2; i += 8) {
int16x4x2_t d0 = vld2_s16(data + i);
int16x4x2_t w0 = vld2_s16(weights + i * 4);
int16x4x2_t w1 = vld2_s16(weights + i * 4 + 8);
int16x4x2_t w2 = vld2_s16(weights + i * 4 + 16);
int16x4x2_t w3 = vld2_s16(weights + i * 4 + 24);
accum0 = vmlal_s16(accum0, d0.val[0], w0.val[0]);
accum0 = vmlal_s16(accum0, d0.val[1], w0.val[1]);
accum1 = vmlal_s16(accum1, d0.val[0], w1.val[0]);
accum1 = vmlal_s16(accum1, d0.val[1], w1.val[1]);
accum2 = vmlal_s16(accum2, d0.val[0], w2.val[0]);
accum2 = vmlal_s16(accum2, d0.val[1], w2.val[1]);
accum3 = vmlal_s16(accum3, d0.val[0], w3.val[0]);
accum3 = vmlal_s16(accum3, d0.val[1], w3.val[1]);
}
int32x2_t sum0 = vpadd_s32(vget_low_s32(accum0), vget_high_s32(accum0));
int32x2_t sum1 = vpadd_s32(vget_low_s32(accum1), vget_high_s32(accum1));
int32x2_t sum2 = vpadd_s32(vget_low_s32(accum2), vget_high_s32(accum2));
int32x2_t sum3 = vpadd_s32(vget_low_s32(accum3), vget_high_s32(accum3));
sum0 = vpadd_s32(sum0, sum1);
sum1 = vpadd_s32(sum2, sum3);
int32x4_t sum = vcombine_s32(sum0, sum1);
float32x4_t m0 = vcvtq_f32_s32(sum);
m0 = vmulq_f32(m0, vld1q_f32(weightsf + 512/4));
m0 = vaddq_f32(m0, vld1q_f32(weightsf + 528/4));
float32x4_t m1, m2, m3, m4;
m1 = m0;
m0 = vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(m0), sign_bits_f));
m0 = vaddq_f32(m0, ones_f);
m0 = vmulq_f32(reciprocal(m0), m1);
m1 = vdupq_lane_f32(vget_low_f32(m0), 0);
m2 = vdupq_lane_f32(vget_low_f32(m0), 1);
m3 = vdupq_lane_f32(vget_high_f32(m0), 0);
m4 = vdupq_lane_f32(vget_high_f32(m0), 1);
m1 = vmulq_f32(m1, vld1q_f32(weightsf + 544/4));
m2 = vmulq_f32(m2, vld1q_f32(weightsf + 560/4));
m3 = vmulq_f32(m3, vld1q_f32(weightsf + 576/4));
m4 = vmulq_f32(m4, vld1q_f32(weightsf + 592/4));
m1 = vaddq_f32(m1, m2);
m3 = vaddq_f32(m3, m4);
m1 = vaddq_f32(m1, m3);
m1 = vaddq_f32(m1, vld1q_f32(weightsf + 608/4));
uint32x4_t gte = vcgeq_f32(m1, zeroes_f);
uint16x4_t gte_u16 = vmovn_u32(gte);
uint8x8_t gte_u8 = vmovn_u16(vcombine_u16(gte_u16, vget_low_u16(vreinterpretq_u16_u32(sign_bits_f))));
gte_u8 = vshr_n_u8(gte_u8, 7);
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(gte_u8), 0);
}
void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
uint8_t *dst, ptrdiff_t dst_stride,
const int16_t *filter_x, // unused
int x_step_q4, // unused
const int16_t *filter_y, int y_step_q4, int w,
int h) {
int height;
const uint8_t *s;
uint8_t *d;
uint32x2_t d2u32, d3u32;
uint32x2_t d16u32, d18u32, d20u32, d22u32, d24u32, d26u32;
int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16;
int16x4_t d24s16, d25s16, d26s16, d27s16;
uint16x4_t d2u16, d3u16, d4u16, d5u16;
int16x8_t q0s16;
uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
int32x4_t q1s32, q2s32, q14s32, q15s32;
assert(y_step_q4 == 16);
(void)x_step_q4;
(void)y_step_q4;
(void)filter_x;
src -= src_stride * 3;
q0s16 = vld1q_s16(filter_y);
for (; w > 0; w -= 4, src += 4, dst += 4) { // loop_vert_h
s = src;
d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 0);
s += src_stride;
d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 1);
s += src_stride;
d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 0);
s += src_stride;
d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 1);
s += src_stride;
d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 0);
s += src_stride;
d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 1);
s += src_stride;
d22u32 = vld1_lane_u32((const uint32_t *)s, d22u32, 0);
s += src_stride;
q8u16 = vmovl_u8(vreinterpret_u8_u32(d16u32));
q9u16 = vmovl_u8(vreinterpret_u8_u32(d18u32));
q10u16 = vmovl_u8(vreinterpret_u8_u32(d20u32));
q11u16 = vmovl_u8(vreinterpret_u8_u32(d22u32));
d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d = dst;
for (height = h; height > 0; height -= 4) { // loop_vert
d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 0);
s += src_stride;
d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 0);
s += src_stride;
d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 1);
s += src_stride;
d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 1);
s += src_stride;
q12u16 = vmovl_u8(vreinterpret_u8_u32(d24u32));
q13u16 = vmovl_u8(vreinterpret_u8_u32(d26u32));
d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
__builtin_prefetch(d);
__builtin_prefetch(d + dst_stride);
q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d18s16, d19s16, d20s16, d21s16,
d22s16, d24s16, q0s16);
__builtin_prefetch(d + dst_stride * 2);
__builtin_prefetch(d + dst_stride * 3);
q2s32 = MULTIPLY_BY_Q0(d17s16, d18s16, d19s16, d20s16, d21s16, d22s16,
d24s16, d26s16, q0s16);
__builtin_prefetch(s);
__builtin_prefetch(s + src_stride);
q14s32 = MULTIPLY_BY_Q0(d18s16, d19s16, d20s16, d21s16, d22s16, d24s16,
d26s16, d27s16, q0s16);
__builtin_prefetch(s + src_stride * 2);
__builtin_prefetch(s + src_stride * 3);
q15s32 = MULTIPLY_BY_Q0(d19s16, d20s16, d21s16, d22s16, d24s16, d26s16,
d27s16, d25s16, q0s16);
d2u16 = vqrshrun_n_s32(q1s32, 7);
d3u16 = vqrshrun_n_s32(q2s32, 7);
d4u16 = vqrshrun_n_s32(q14s32, 7);
d5u16 = vqrshrun_n_s32(q15s32, 7);
q1u16 = vcombine_u16(d2u16, d3u16);
q2u16 = vcombine_u16(d4u16, d5u16);
d2u32 = vreinterpret_u32_u8(vqmovn_u16(q1u16));
d3u32 = vreinterpret_u32_u8(vqmovn_u16(q2u16));
vst1_lane_u32((uint32_t *)d, d2u32, 0);
d += dst_stride;
vst1_lane_u32((uint32_t *)d, d2u32, 1);
d += dst_stride;
vst1_lane_u32((uint32_t *)d, d3u32, 0);
d += dst_stride;
vst1_lane_u32((uint32_t *)d, d3u32, 1);
d += dst_stride;
q8u16 = q10u16;
d18s16 = d22s16;
d19s16 = d24s16;
q10u16 = q13u16;
d22s16 = d25s16;
}
}
return;
}
static INLINE void scaledconvolve_horiz_w4(
const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
const ptrdiff_t dst_stride, const InterpKernel *const x_filters,
const int x0_q4, const int x_step_q4, const int w, const int h) {
DECLARE_ALIGNED(16, uint8_t, temp[4 * 4]);
int x, y, z;
src -= SUBPEL_TAPS / 2 - 1;
y = h;
do {
int x_q4 = x0_q4;
x = 0;
do {
// process 4 src_x steps
for (z = 0; z < 4; ++z) {
const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
if (x_q4 & SUBPEL_MASK) {
const int16x8_t filters = vld1q_s16(x_filters[x_q4 & SUBPEL_MASK]);
const int16x4_t filter3 = vdup_lane_s16(vget_low_s16(filters), 3);
const int16x4_t filter4 = vdup_lane_s16(vget_high_s16(filters), 0);
uint8x8_t s[8], d;
int16x8_t ss[4];
int16x4_t t[8], tt;
load_u8_8x4(src_x, src_stride, &s[0], &s[1], &s[2], &s[3]);
transpose_u8_8x4(&s[0], &s[1], &s[2], &s[3]);
ss[0] = vreinterpretq_s16_u16(vmovl_u8(s[0]));
ss[1] = vreinterpretq_s16_u16(vmovl_u8(s[1]));
ss[2] = vreinterpretq_s16_u16(vmovl_u8(s[2]));
ss[3] = vreinterpretq_s16_u16(vmovl_u8(s[3]));
t[0] = vget_low_s16(ss[0]);
t[1] = vget_low_s16(ss[1]);
t[2] = vget_low_s16(ss[2]);
t[3] = vget_low_s16(ss[3]);
t[4] = vget_high_s16(ss[0]);
t[5] = vget_high_s16(ss[1]);
t[6] = vget_high_s16(ss[2]);
t[7] = vget_high_s16(ss[3]);
tt = convolve8_4(t[0], t[1], t[2], t[3], t[4], t[5], t[6], t[7],
filters, filter3, filter4);
d = vqrshrun_n_s16(vcombine_s16(tt, tt), 7);
vst1_lane_u32((uint32_t *)&temp[4 * z], vreinterpret_u32_u8(d), 0);
} else {
int i;
for (i = 0; i < 4; ++i) {
temp[z * 4 + i] = src_x[i * src_stride + 3];
}
}
x_q4 += x_step_q4;
}
// transpose the 4x4 filters values back to dst
{
const uint8x8x4_t d4 = vld4_u8(temp);
vst1_lane_u32((uint32_t *)&dst[x + 0 * dst_stride],
vreinterpret_u32_u8(d4.val[0]), 0);
vst1_lane_u32((uint32_t *)&dst[x + 1 * dst_stride],
vreinterpret_u32_u8(d4.val[1]), 0);
vst1_lane_u32((uint32_t *)&dst[x + 2 * dst_stride],
vreinterpret_u32_u8(d4.val[2]), 0);
vst1_lane_u32((uint32_t *)&dst[x + 3 * dst_stride],
vreinterpret_u32_u8(d4.val[3]), 0);
}
x += 4;
} while (x < w);
src += src_stride * 4;
dst += dst_stride * 4;
y -= 4;
} while (y > 0);
}
void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
uint8_t *dst, ptrdiff_t dst_stride,
const int16_t *filter_x, int x_step_q4,
const int16_t *filter_y, // unused
int y_step_q4, // unused
int w, int h) {
int width;
const uint8_t *s, *psrc;
uint8_t *d, *pdst;
uint8x8_t d2u8, d3u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8;
uint32x2_t d2u32, d3u32, d28u32, d29u32, d30u32, d31u32;
uint8x16_t q12u8, q13u8, q14u8, q15u8;
int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d22s16, d23s16;
int16x4_t d24s16, d25s16, d26s16, d27s16;
uint16x4_t d2u16, d3u16, d4u16, d5u16, d16u16, d17u16, d18u16, d19u16;
int16x8_t q0s16;
uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
int32x4_t q1s32, q2s32, q14s32, q15s32;
uint16x8x2_t q0x2u16;
uint8x8x2_t d0x2u8, d1x2u8;
uint32x2x2_t d0x2u32;
uint16x4x2_t d0x2u16, d1x2u16;
uint32x4x2_t q0x2u32;
assert(x_step_q4 == 16);
(void)x_step_q4;
(void)y_step_q4;
(void)filter_y;
q0s16 = vld1q_s16(filter_x);
src -= 3; // adjust for taps
for (; h > 0; h -= 4, src += src_stride * 4,
dst += dst_stride * 4) { // loop_horiz_v
s = src;
d24u8 = vld1_u8(s);
s += src_stride;
d25u8 = vld1_u8(s);
s += src_stride;
d26u8 = vld1_u8(s);
s += src_stride;
d27u8 = vld1_u8(s);
q12u8 = vcombine_u8(d24u8, d25u8);
q13u8 = vcombine_u8(d26u8, d27u8);
q0x2u16 =
vtrnq_u16(vreinterpretq_u16_u8(q12u8), vreinterpretq_u16_u8(q13u8));
d24u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[0]));
d25u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[0]));
d26u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[1]));
d27u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[1]));
d0x2u8 = vtrn_u8(d24u8, d25u8);
d1x2u8 = vtrn_u8(d26u8, d27u8);
__builtin_prefetch(src + src_stride * 4);
__builtin_prefetch(src + src_stride * 5);
__builtin_prefetch(src + src_stride * 6);
q8u16 = vmovl_u8(d0x2u8.val[0]);
q9u16 = vmovl_u8(d0x2u8.val[1]);
q10u16 = vmovl_u8(d1x2u8.val[0]);
q11u16 = vmovl_u8(d1x2u8.val[1]);
d16u16 = vget_low_u16(q8u16);
d17u16 = vget_high_u16(q8u16);
d18u16 = vget_low_u16(q9u16);
d19u16 = vget_high_u16(q9u16);
q8u16 = vcombine_u16(d16u16, d18u16); // vswp 17 18
q9u16 = vcombine_u16(d17u16, d19u16);
d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
d23s16 = vreinterpret_s16_u16(vget_high_u16(q10u16)); // vmov 23 21
for (width = w, psrc = src + 7, pdst = dst; width > 0;
width -= 4, psrc += 4, pdst += 4) { // loop_horiz
s = psrc;
d28u32 = vld1_dup_u32((const uint32_t *)s);
s += src_stride;
d29u32 = vld1_dup_u32((const uint32_t *)s);
s += src_stride;
d31u32 = vld1_dup_u32((const uint32_t *)s);
s += src_stride;
d30u32 = vld1_dup_u32((const uint32_t *)s);
__builtin_prefetch(psrc + 64);
d0x2u16 =
vtrn_u16(vreinterpret_u16_u32(d28u32), vreinterpret_u16_u32(d31u32));
d1x2u16 =
vtrn_u16(vreinterpret_u16_u32(d29u32), vreinterpret_u16_u32(d30u32));
d0x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[0]), // d28
vreinterpret_u8_u16(d1x2u16.val[0])); // d29
d1x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[1]), // d31
vreinterpret_u8_u16(d1x2u16.val[1])); // d30
__builtin_prefetch(psrc + 64 + src_stride);
q14u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
q15u8 = vcombine_u8(d1x2u8.val[1], d1x2u8.val[0]);
q0x2u32 =
vtrnq_u32(vreinterpretq_u32_u8(q14u8), vreinterpretq_u32_u8(q15u8));
d28u8 = vreinterpret_u8_u32(vget_low_u32(q0x2u32.val[0]));
d29u8 = vreinterpret_u8_u32(vget_high_u32(q0x2u32.val[0]));
q12u16 = vmovl_u8(d28u8);
q13u16 = vmovl_u8(d29u8);
__builtin_prefetch(psrc + 64 + src_stride * 2);
d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d20s16, d22s16, d18s16, d19s16,
d23s16, d24s16, q0s16);
q2s32 = MULTIPLY_BY_Q0(d17s16, d20s16, d22s16, d18s16, d19s16, d23s16,
d24s16, d26s16, q0s16);
q14s32 = MULTIPLY_BY_Q0(d20s16, d22s16, d18s16, d19s16, d23s16, d24s16,
d26s16, d27s16, q0s16);
q15s32 = MULTIPLY_BY_Q0(d22s16, d18s16, d19s16, d23s16, d24s16, d26s16,
d27s16, d25s16, q0s16);
__builtin_prefetch(psrc + 60 + src_stride * 3);
d2u16 = vqrshrun_n_s32(q1s32, 7);
d3u16 = vqrshrun_n_s32(q2s32, 7);
d4u16 = vqrshrun_n_s32(q14s32, 7);
d5u16 = vqrshrun_n_s32(q15s32, 7);
q1u16 = vcombine_u16(d2u16, d3u16);
q2u16 = vcombine_u16(d4u16, d5u16);
d2u8 = vqmovn_u16(q1u16);
d3u8 = vqmovn_u16(q2u16);
d0x2u16 = vtrn_u16(vreinterpret_u16_u8(d2u8), vreinterpret_u16_u8(d3u8));
d0x2u32 = vtrn_u32(vreinterpret_u32_u16(d0x2u16.val[0]),
vreinterpret_u32_u16(d0x2u16.val[1]));
d0x2u8 = vtrn_u8(vreinterpret_u8_u32(d0x2u32.val[0]),
vreinterpret_u8_u32(d0x2u32.val[1]));
d2u32 = vreinterpret_u32_u8(d0x2u8.val[0]);
d3u32 = vreinterpret_u32_u8(d0x2u8.val[1]);
d = pdst;
vst1_lane_u32((uint32_t *)d, d2u32, 0);
d += dst_stride;
vst1_lane_u32((uint32_t *)d, d3u32, 0);
d += dst_stride;
vst1_lane_u32((uint32_t *)d, d2u32, 1);
d += dst_stride;
vst1_lane_u32((uint32_t *)d, d3u32, 1);
q8u16 = q9u16;
d20s16 = d23s16;
q11u16 = q12u16;
q9u16 = q13u16;
d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
}
}
return;
}
unsigned int vp8_variance16x8_neon(
const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *sse) {
int i;
int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
uint32x2_t d0u32, d10u32;
int64x1_t d0s64, d1s64;
uint8x16_t q0u8, q1u8, q2u8, q3u8;
uint16x8_t q11u16, q12u16, q13u16, q14u16;
int32x4_t q8s32, q9s32, q10s32;
int64x2_t q0s64, q1s64, q5s64;
q8s32 = vdupq_n_s32(0);
q9s32 = vdupq_n_s32(0);
q10s32 = vdupq_n_s32(0);
for (i = 0; i < 4; i++) { // variance16x8_neon_loop
q0u8 = vld1q_u8(src_ptr);
src_ptr += source_stride;
q1u8 = vld1q_u8(src_ptr);
src_ptr += source_stride;
__builtin_prefetch(src_ptr);
q2u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q3u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
__builtin_prefetch(ref_ptr);
q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
q10s32 = vmlal_s16(q10s32, d27s16, d27s16);
d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
}
q10s32 = vaddq_s32(q10s32, q9s32);
q0s64 = vpaddlq_s32(q8s32);
q1s64 = vpaddlq_s32(q10s32);
d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
vreinterpret_s32_s64(d0s64));
vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
return vget_lane_u32(d0u32, 0);
}
unsigned int vpx_mse16x16_neon(
const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *sse) {
int i;
int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
int64x1_t d0s64;
uint8x16_t q0u8, q1u8, q2u8, q3u8;
int32x4_t q7s32, q8s32, q9s32, q10s32;
uint16x8_t q11u16, q12u16, q13u16, q14u16;
int64x2_t q1s64;
q7s32 = vdupq_n_s32(0);
q8s32 = vdupq_n_s32(0);
q9s32 = vdupq_n_s32(0);
q10s32 = vdupq_n_s32(0);
for (i = 0; i < 8; i++) { // mse16x16_neon_loop
q0u8 = vld1q_u8(src_ptr);
src_ptr += source_stride;
q1u8 = vld1q_u8(src_ptr);
src_ptr += source_stride;
q2u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q3u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
q7s32 = vmlal_s16(q7s32, d22s16, d22s16);
q8s32 = vmlal_s16(q8s32, d23s16, d23s16);
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
q7s32 = vmlal_s16(q7s32, d26s16, d26s16);
q8s32 = vmlal_s16(q8s32, d27s16, d27s16);
d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
}
q7s32 = vaddq_s32(q7s32, q8s32);
q9s32 = vaddq_s32(q9s32, q10s32);
q10s32 = vaddq_s32(q7s32, q9s32);
q1s64 = vpaddlq_s32(q10s32);
d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d0s64), 0);
return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0);
}
unsigned int vp8_sub_pixel_variance16x16_neon_func(
const unsigned char *src_ptr,
int src_pixels_per_line,
int xoffset,
int yoffset,
const unsigned char *dst_ptr,
int dst_pixels_per_line,
unsigned int *sse) {
int i;
DECLARE_ALIGNED_ARRAY(16, unsigned char, tmp, 528);
unsigned char *tmpp;
unsigned char *tmpp2;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8;
uint8x8_t d10u8, d11u8, d12u8, d13u8, d14u8, d15u8, d16u8, d17u8, d18u8;
uint8x8_t d19u8, d20u8, d21u8;
int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
uint32x2_t d0u32, d10u32;
int64x1_t d0s64, d1s64, d2s64, d3s64;
uint8x16_t q0u8, q1u8, q2u8, q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8;
uint8x16_t q10u8, q11u8, q12u8, q13u8, q14u8, q15u8;
uint16x8_t q1u16, q2u16, q3u16, q4u16, q5u16, q6u16, q7u16, q8u16;
uint16x8_t q9u16, q10u16, q11u16, q12u16, q13u16, q14u16;
int32x4_t q8s32, q9s32, q10s32;
int64x2_t q0s64, q1s64, q5s64;
tmpp2 = tmp + 272;
tmpp = tmp;
if (xoffset == 0) { // secondpass_bfilter16x16_only
d0u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][0]);
d1u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][1]);
q11u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
for (i = 4; i > 0; i--) {
q12u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
q13u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
q14u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
q15u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
__builtin_prefetch(src_ptr);
__builtin_prefetch(src_ptr + src_pixels_per_line);
__builtin_prefetch(src_ptr + src_pixels_per_line * 2);
q1u16 = vmull_u8(vget_low_u8(q11u8), d0u8);
q2u16 = vmull_u8(vget_high_u8(q11u8), d0u8);
q3u16 = vmull_u8(vget_low_u8(q12u8), d0u8);
q4u16 = vmull_u8(vget_high_u8(q12u8), d0u8);
q5u16 = vmull_u8(vget_low_u8(q13u8), d0u8);
q6u16 = vmull_u8(vget_high_u8(q13u8), d0u8);
q7u16 = vmull_u8(vget_low_u8(q14u8), d0u8);
q8u16 = vmull_u8(vget_high_u8(q14u8), d0u8);
q1u16 = vmlal_u8(q1u16, vget_low_u8(q12u8), d1u8);
q2u16 = vmlal_u8(q2u16, vget_high_u8(q12u8), d1u8);
q3u16 = vmlal_u8(q3u16, vget_low_u8(q13u8), d1u8);
q4u16 = vmlal_u8(q4u16, vget_high_u8(q13u8), d1u8);
q5u16 = vmlal_u8(q5u16, vget_low_u8(q14u8), d1u8);
q6u16 = vmlal_u8(q6u16, vget_high_u8(q14u8), d1u8);
q7u16 = vmlal_u8(q7u16, vget_low_u8(q15u8), d1u8);
q8u16 = vmlal_u8(q8u16, vget_high_u8(q15u8), d1u8);
d2u8 = vqrshrn_n_u16(q1u16, 7);
d3u8 = vqrshrn_n_u16(q2u16, 7);
d4u8 = vqrshrn_n_u16(q3u16, 7);
d5u8 = vqrshrn_n_u16(q4u16, 7);
d6u8 = vqrshrn_n_u16(q5u16, 7);
d7u8 = vqrshrn_n_u16(q6u16, 7);
d8u8 = vqrshrn_n_u16(q7u16, 7);
d9u8 = vqrshrn_n_u16(q8u16, 7);
q1u8 = vcombine_u8(d2u8, d3u8);
q2u8 = vcombine_u8(d4u8, d5u8);
q3u8 = vcombine_u8(d6u8, d7u8);
q4u8 = vcombine_u8(d8u8, d9u8);
q11u8 = q15u8;
vst1q_u8((uint8_t *)tmpp2, q1u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q2u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q3u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q4u8);
tmpp2 += 16;
}
} else if (yoffset == 0) { // firstpass_bfilter16x16_only
d0u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][0]);
d1u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][1]);
for (i = 4; i > 0 ; i--) {
d2u8 = vld1_u8(src_ptr);
d3u8 = vld1_u8(src_ptr + 8);
d4u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d5u8 = vld1_u8(src_ptr);
d6u8 = vld1_u8(src_ptr + 8);
d7u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d8u8 = vld1_u8(src_ptr);
d9u8 = vld1_u8(src_ptr + 8);
d10u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d11u8 = vld1_u8(src_ptr);
d12u8 = vld1_u8(src_ptr + 8);
d13u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
__builtin_prefetch(src_ptr);
__builtin_prefetch(src_ptr + src_pixels_per_line);
__builtin_prefetch(src_ptr + src_pixels_per_line * 2);
q7u16 = vmull_u8(d2u8, d0u8);
q8u16 = vmull_u8(d3u8, d0u8);
q9u16 = vmull_u8(d5u8, d0u8);
q10u16 = vmull_u8(d6u8, d0u8);
q11u16 = vmull_u8(d8u8, d0u8);
q12u16 = vmull_u8(d9u8, d0u8);
q13u16 = vmull_u8(d11u8, d0u8);
q14u16 = vmull_u8(d12u8, d0u8);
d2u8 = vext_u8(d2u8, d3u8, 1);
d5u8 = vext_u8(d5u8, d6u8, 1);
d8u8 = vext_u8(d8u8, d9u8, 1);
d11u8 = vext_u8(d11u8, d12u8, 1);
q7u16 = vmlal_u8(q7u16, d2u8, d1u8);
q9u16 = vmlal_u8(q9u16, d5u8, d1u8);
q11u16 = vmlal_u8(q11u16, d8u8, d1u8);
q13u16 = vmlal_u8(q13u16, d11u8, d1u8);
d3u8 = vext_u8(d3u8, d4u8, 1);
d6u8 = vext_u8(d6u8, d7u8, 1);
d9u8 = vext_u8(d9u8, d10u8, 1);
d12u8 = vext_u8(d12u8, d13u8, 1);
q8u16 = vmlal_u8(q8u16, d3u8, d1u8);
q10u16 = vmlal_u8(q10u16, d6u8, d1u8);
q12u16 = vmlal_u8(q12u16, d9u8, d1u8);
q14u16 = vmlal_u8(q14u16, d12u8, d1u8);
d14u8 = vqrshrn_n_u16(q7u16, 7);
d15u8 = vqrshrn_n_u16(q8u16, 7);
d16u8 = vqrshrn_n_u16(q9u16, 7);
d17u8 = vqrshrn_n_u16(q10u16, 7);
d18u8 = vqrshrn_n_u16(q11u16, 7);
d19u8 = vqrshrn_n_u16(q12u16, 7);
d20u8 = vqrshrn_n_u16(q13u16, 7);
d21u8 = vqrshrn_n_u16(q14u16, 7);
q7u8 = vcombine_u8(d14u8, d15u8);
q8u8 = vcombine_u8(d16u8, d17u8);
q9u8 = vcombine_u8(d18u8, d19u8);
q10u8 = vcombine_u8(d20u8, d21u8);
vst1q_u8((uint8_t *)tmpp2, q7u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q8u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q9u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q10u8);
tmpp2 += 16;
}
} else {
d0u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][0]);
d1u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][1]);
d2u8 = vld1_u8(src_ptr);
d3u8 = vld1_u8(src_ptr + 8);
d4u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d5u8 = vld1_u8(src_ptr);
d6u8 = vld1_u8(src_ptr + 8);
d7u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d8u8 = vld1_u8(src_ptr);
d9u8 = vld1_u8(src_ptr + 8);
d10u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d11u8 = vld1_u8(src_ptr);
d12u8 = vld1_u8(src_ptr + 8);
d13u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
// First Pass: output_height lines x output_width columns (17x16)
for (i = 3; i > 0; i--) {
q7u16 = vmull_u8(d2u8, d0u8);
q8u16 = vmull_u8(d3u8, d0u8);
q9u16 = vmull_u8(d5u8, d0u8);
q10u16 = vmull_u8(d6u8, d0u8);
q11u16 = vmull_u8(d8u8, d0u8);
q12u16 = vmull_u8(d9u8, d0u8);
q13u16 = vmull_u8(d11u8, d0u8);
q14u16 = vmull_u8(d12u8, d0u8);
d2u8 = vext_u8(d2u8, d3u8, 1);
d5u8 = vext_u8(d5u8, d6u8, 1);
d8u8 = vext_u8(d8u8, d9u8, 1);
d11u8 = vext_u8(d11u8, d12u8, 1);
q7u16 = vmlal_u8(q7u16, d2u8, d1u8);
q9u16 = vmlal_u8(q9u16, d5u8, d1u8);
q11u16 = vmlal_u8(q11u16, d8u8, d1u8);
q13u16 = vmlal_u8(q13u16, d11u8, d1u8);
d3u8 = vext_u8(d3u8, d4u8, 1);
d6u8 = vext_u8(d6u8, d7u8, 1);
d9u8 = vext_u8(d9u8, d10u8, 1);
d12u8 = vext_u8(d12u8, d13u8, 1);
q8u16 = vmlal_u8(q8u16, d3u8, d1u8);
q10u16 = vmlal_u8(q10u16, d6u8, d1u8);
q12u16 = vmlal_u8(q12u16, d9u8, d1u8);
q14u16 = vmlal_u8(q14u16, d12u8, d1u8);
d14u8 = vqrshrn_n_u16(q7u16, 7);
d15u8 = vqrshrn_n_u16(q8u16, 7);
d16u8 = vqrshrn_n_u16(q9u16, 7);
d17u8 = vqrshrn_n_u16(q10u16, 7);
d18u8 = vqrshrn_n_u16(q11u16, 7);
d19u8 = vqrshrn_n_u16(q12u16, 7);
d20u8 = vqrshrn_n_u16(q13u16, 7);
d21u8 = vqrshrn_n_u16(q14u16, 7);
d2u8 = vld1_u8(src_ptr);
d3u8 = vld1_u8(src_ptr + 8);
d4u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d5u8 = vld1_u8(src_ptr);
d6u8 = vld1_u8(src_ptr + 8);
d7u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d8u8 = vld1_u8(src_ptr);
d9u8 = vld1_u8(src_ptr + 8);
d10u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d11u8 = vld1_u8(src_ptr);
d12u8 = vld1_u8(src_ptr + 8);
d13u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
q7u8 = vcombine_u8(d14u8, d15u8);
q8u8 = vcombine_u8(d16u8, d17u8);
q9u8 = vcombine_u8(d18u8, d19u8);
q10u8 = vcombine_u8(d20u8, d21u8);
vst1q_u8((uint8_t *)tmpp, q7u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q8u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q9u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q10u8);
tmpp += 16;
}
// First-pass filtering for rest 5 lines
d14u8 = vld1_u8(src_ptr);
d15u8 = vld1_u8(src_ptr + 8);
d16u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
q9u16 = vmull_u8(d2u8, d0u8);
q10u16 = vmull_u8(d3u8, d0u8);
q11u16 = vmull_u8(d5u8, d0u8);
q12u16 = vmull_u8(d6u8, d0u8);
q13u16 = vmull_u8(d8u8, d0u8);
q14u16 = vmull_u8(d9u8, d0u8);
d2u8 = vext_u8(d2u8, d3u8, 1);
d5u8 = vext_u8(d5u8, d6u8, 1);
d8u8 = vext_u8(d8u8, d9u8, 1);
q9u16 = vmlal_u8(q9u16, d2u8, d1u8);
q11u16 = vmlal_u8(q11u16, d5u8, d1u8);
q13u16 = vmlal_u8(q13u16, d8u8, d1u8);
d3u8 = vext_u8(d3u8, d4u8, 1);
d6u8 = vext_u8(d6u8, d7u8, 1);
d9u8 = vext_u8(d9u8, d10u8, 1);
q10u16 = vmlal_u8(q10u16, d3u8, d1u8);
q12u16 = vmlal_u8(q12u16, d6u8, d1u8);
q14u16 = vmlal_u8(q14u16, d9u8, d1u8);
q1u16 = vmull_u8(d11u8, d0u8);
q2u16 = vmull_u8(d12u8, d0u8);
q3u16 = vmull_u8(d14u8, d0u8);
q4u16 = vmull_u8(d15u8, d0u8);
d11u8 = vext_u8(d11u8, d12u8, 1);
d14u8 = vext_u8(d14u8, d15u8, 1);
q1u16 = vmlal_u8(q1u16, d11u8, d1u8);
q3u16 = vmlal_u8(q3u16, d14u8, d1u8);
d12u8 = vext_u8(d12u8, d13u8, 1);
d15u8 = vext_u8(d15u8, d16u8, 1);
q2u16 = vmlal_u8(q2u16, d12u8, d1u8);
q4u16 = vmlal_u8(q4u16, d15u8, d1u8);
d10u8 = vqrshrn_n_u16(q9u16, 7);
d11u8 = vqrshrn_n_u16(q10u16, 7);
d12u8 = vqrshrn_n_u16(q11u16, 7);
d13u8 = vqrshrn_n_u16(q12u16, 7);
d14u8 = vqrshrn_n_u16(q13u16, 7);
d15u8 = vqrshrn_n_u16(q14u16, 7);
d16u8 = vqrshrn_n_u16(q1u16, 7);
d17u8 = vqrshrn_n_u16(q2u16, 7);
d18u8 = vqrshrn_n_u16(q3u16, 7);
d19u8 = vqrshrn_n_u16(q4u16, 7);
q5u8 = vcombine_u8(d10u8, d11u8);
q6u8 = vcombine_u8(d12u8, d13u8);
q7u8 = vcombine_u8(d14u8, d15u8);
q8u8 = vcombine_u8(d16u8, d17u8);
q9u8 = vcombine_u8(d18u8, d19u8);
vst1q_u8((uint8_t *)tmpp, q5u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q6u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q7u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q8u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q9u8);
// secondpass_filter
d0u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][0]);
d1u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][1]);
tmpp = tmp;
tmpp2 = tmpp + 272;
q11u8 = vld1q_u8(tmpp);
tmpp += 16;
for (i = 4; i > 0; i--) {
q12u8 = vld1q_u8(tmpp);
tmpp += 16;
q13u8 = vld1q_u8(tmpp);
tmpp += 16;
q14u8 = vld1q_u8(tmpp);
tmpp += 16;
q15u8 = vld1q_u8(tmpp);
tmpp += 16;
q1u16 = vmull_u8(vget_low_u8(q11u8), d0u8);
q2u16 = vmull_u8(vget_high_u8(q11u8), d0u8);
q3u16 = vmull_u8(vget_low_u8(q12u8), d0u8);
q4u16 = vmull_u8(vget_high_u8(q12u8), d0u8);
q5u16 = vmull_u8(vget_low_u8(q13u8), d0u8);
q6u16 = vmull_u8(vget_high_u8(q13u8), d0u8);
q7u16 = vmull_u8(vget_low_u8(q14u8), d0u8);
q8u16 = vmull_u8(vget_high_u8(q14u8), d0u8);
q1u16 = vmlal_u8(q1u16, vget_low_u8(q12u8), d1u8);
q2u16 = vmlal_u8(q2u16, vget_high_u8(q12u8), d1u8);
q3u16 = vmlal_u8(q3u16, vget_low_u8(q13u8), d1u8);
q4u16 = vmlal_u8(q4u16, vget_high_u8(q13u8), d1u8);
q5u16 = vmlal_u8(q5u16, vget_low_u8(q14u8), d1u8);
q6u16 = vmlal_u8(q6u16, vget_high_u8(q14u8), d1u8);
q7u16 = vmlal_u8(q7u16, vget_low_u8(q15u8), d1u8);
q8u16 = vmlal_u8(q8u16, vget_high_u8(q15u8), d1u8);
d2u8 = vqrshrn_n_u16(q1u16, 7);
d3u8 = vqrshrn_n_u16(q2u16, 7);
d4u8 = vqrshrn_n_u16(q3u16, 7);
d5u8 = vqrshrn_n_u16(q4u16, 7);
d6u8 = vqrshrn_n_u16(q5u16, 7);
d7u8 = vqrshrn_n_u16(q6u16, 7);
d8u8 = vqrshrn_n_u16(q7u16, 7);
d9u8 = vqrshrn_n_u16(q8u16, 7);
q1u8 = vcombine_u8(d2u8, d3u8);
q2u8 = vcombine_u8(d4u8, d5u8);
q3u8 = vcombine_u8(d6u8, d7u8);
q4u8 = vcombine_u8(d8u8, d9u8);
q11u8 = q15u8;
vst1q_u8((uint8_t *)tmpp2, q1u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q2u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q3u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q4u8);
tmpp2 += 16;
}
}
// sub_pixel_variance16x16_neon
q8s32 = vdupq_n_s32(0);
q9s32 = vdupq_n_s32(0);
q10s32 = vdupq_n_s32(0);
tmpp = tmp + 272;
for (i = 0; i < 8; i++) { // sub_pixel_variance16x16_neon_loop
q0u8 = vld1q_u8(tmpp);
tmpp += 16;
q1u8 = vld1q_u8(tmpp);
tmpp += 16;
q2u8 = vld1q_u8(dst_ptr);
dst_ptr += dst_pixels_per_line;
q3u8 = vld1q_u8(dst_ptr);
dst_ptr += dst_pixels_per_line;
d0u8 = vget_low_u8(q0u8);
d1u8 = vget_high_u8(q0u8);
d2u8 = vget_low_u8(q1u8);
d3u8 = vget_high_u8(q1u8);
q11u16 = vsubl_u8(d0u8, vget_low_u8(q2u8));
q12u16 = vsubl_u8(d1u8, vget_high_u8(q2u8));
q13u16 = vsubl_u8(d2u8, vget_low_u8(q3u8));
q14u16 = vsubl_u8(d3u8, vget_high_u8(q3u8));
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
q10s32 = vmlal_s16(q10s32, d27s16, d27s16);
d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
}
q10s32 = vaddq_s32(q10s32, q9s32);
q0s64 = vpaddlq_s32(q8s32);
q1s64 = vpaddlq_s32(q10s32);
d0s64 = vget_low_s64(q0s64);
d1s64 = vget_high_s64(q0s64);
d2s64 = vget_low_s64(q1s64);
d3s64 = vget_high_s64(q1s64);
d0s64 = vadd_s64(d0s64, d1s64);
d1s64 = vadd_s64(d2s64, d3s64);
q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
vreinterpret_s32_s64(d0s64));
vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 8);
d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
return vget_lane_u32(d0u32, 0);
}
unsigned int vp8_variance_halfpixvar16x16_hv_neon(
const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *sse) {
int i;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
int16x4_t d0s16, d1s16, d2s16, d3s16, d10s16, d11s16, d12s16, d13s16;
int16x4_t d18s16, d19s16, d20s16, d21s16, d22s16, d23s16, d24s16, d25s16;
uint32x2_t d0u32, d10u32;
int64x1_t d0s64, d1s64, d2s64, d3s64;
uint8x16_t q0u8, q1u8, q2u8, q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8;
uint16x8_t q0u16, q1u16, q5u16, q6u16, q9u16, q10u16, q11u16, q12u16;
int32x4_t q13s32, q14s32, q15s32;
int64x2_t q0s64, q1s64, q5s64;
q13s32 = vdupq_n_s32(0);
q14s32 = vdupq_n_s32(0);
q15s32 = vdupq_n_s32(0);
q0u8 = vld1q_u8(src_ptr);
q1u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q1u8 = vextq_u8(q0u8, q1u8, 1);
q0u8 = vrhaddq_u8(q0u8, q1u8);
for (i = 0; i < 4; i++) { // vp8_filt_fpo16x16s_4_0_loop_neon
q2u8 = vld1q_u8(src_ptr);
q3u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q4u8 = vld1q_u8(src_ptr);
q5u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q6u8 = vld1q_u8(src_ptr);
q7u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q8u8 = vld1q_u8(src_ptr);
q9u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q3u8 = vextq_u8(q2u8, q3u8, 1);
q5u8 = vextq_u8(q4u8, q5u8, 1);
q7u8 = vextq_u8(q6u8, q7u8, 1);
q9u8 = vextq_u8(q8u8, q9u8, 1);
q1u8 = vrhaddq_u8(q2u8, q3u8);
q2u8 = vrhaddq_u8(q4u8, q5u8);
q3u8 = vrhaddq_u8(q6u8, q7u8);
q4u8 = vrhaddq_u8(q8u8, q9u8);
q0u8 = vrhaddq_u8(q0u8, q1u8);
q1u8 = vrhaddq_u8(q1u8, q2u8);
q2u8 = vrhaddq_u8(q2u8, q3u8);
q3u8 = vrhaddq_u8(q3u8, q4u8);
q5u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q6u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q7u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q8u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
d0u8 = vget_low_u8(q0u8);
d1u8 = vget_high_u8(q0u8);
d2u8 = vget_low_u8(q1u8);
d3u8 = vget_high_u8(q1u8);
d4u8 = vget_low_u8(q2u8);
d5u8 = vget_high_u8(q2u8);
d6u8 = vget_low_u8(q3u8);
d7u8 = vget_high_u8(q3u8);
q9u16 = vsubl_u8(d0u8, vget_low_u8(q5u8));
q10u16 = vsubl_u8(d1u8, vget_high_u8(q5u8));
q11u16 = vsubl_u8(d2u8, vget_low_u8(q6u8));
q12u16 = vsubl_u8(d3u8, vget_high_u8(q6u8));
q0u16 = vsubl_u8(d4u8, vget_low_u8(q7u8));
q1u16 = vsubl_u8(d5u8, vget_high_u8(q7u8));
q5u16 = vsubl_u8(d6u8, vget_low_u8(q8u8));
q6u16 = vsubl_u8(d7u8, vget_high_u8(q8u8));
d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q9u16));
q14s32 = vmlal_s16(q14s32, d18s16, d18s16);
q15s32 = vmlal_s16(q15s32, d19s16, d19s16);
d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q10u16));
q14s32 = vmlal_s16(q14s32, d20s16, d20s16);
q15s32 = vmlal_s16(q15s32, d21s16, d21s16);
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q11u16));
q14s32 = vmlal_s16(q14s32, d22s16, d22s16);
q15s32 = vmlal_s16(q15s32, d23s16, d23s16);
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q12u16));
q14s32 = vmlal_s16(q14s32, d24s16, d24s16);
q15s32 = vmlal_s16(q15s32, d25s16, d25s16);
d0s16 = vreinterpret_s16_u16(vget_low_u16(q0u16));
d1s16 = vreinterpret_s16_u16(vget_high_u16(q0u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q0u16));
q14s32 = vmlal_s16(q14s32, d0s16, d0s16);
q15s32 = vmlal_s16(q15s32, d1s16, d1s16);
d2s16 = vreinterpret_s16_u16(vget_low_u16(q1u16));
d3s16 = vreinterpret_s16_u16(vget_high_u16(q1u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q1u16));
q14s32 = vmlal_s16(q14s32, d2s16, d2s16);
q15s32 = vmlal_s16(q15s32, d3s16, d3s16);
d10s16 = vreinterpret_s16_u16(vget_low_u16(q5u16));
d11s16 = vreinterpret_s16_u16(vget_high_u16(q5u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q5u16));
q14s32 = vmlal_s16(q14s32, d10s16, d10s16);
q15s32 = vmlal_s16(q15s32, d11s16, d11s16);
d12s16 = vreinterpret_s16_u16(vget_low_u16(q6u16));
d13s16 = vreinterpret_s16_u16(vget_high_u16(q6u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q6u16));
q14s32 = vmlal_s16(q14s32, d12s16, d12s16);
q15s32 = vmlal_s16(q15s32, d13s16, d13s16);
q0u8 = q4u8;
}
q15s32 = vaddq_s32(q14s32, q15s32);
q0s64 = vpaddlq_s32(q13s32);
q1s64 = vpaddlq_s32(q15s32);
d0s64 = vget_low_s64(q0s64);
d1s64 = vget_high_s64(q0s64);
d2s64 = vget_low_s64(q1s64);
d3s64 = vget_high_s64(q1s64);
d0s64 = vadd_s64(d0s64, d1s64);
d1s64 = vadd_s64(d2s64, d3s64);
q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
vreinterpret_s32_s64(d0s64));
vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 8);
d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
return vget_lane_u32(d0u32, 0);
}
void aom_idct4x4_16_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
uint8x8_t d26u8, d27u8;
uint32x2_t d26u32, d27u32;
uint16x8_t q8u16, q9u16;
int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16;
int16x4_t d22s16, d23s16, d24s16, d26s16, d27s16, d28s16, d29s16;
int16x8_t q8s16, q9s16, q13s16, q14s16;
int32x4_t q1s32, q13s32, q14s32, q15s32;
int16x4x2_t d0x2s16, d1x2s16;
int32x4x2_t q0x2s32;
uint8_t *d;
d26u32 = d27u32 = vdup_n_u32(0);
q8s16 = vld1q_s16(input);
q9s16 = vld1q_s16(input + 8);
d16s16 = vget_low_s16(q8s16);
d17s16 = vget_high_s16(q8s16);
d18s16 = vget_low_s16(q9s16);
d19s16 = vget_high_s16(q9s16);
d0x2s16 = vtrn_s16(d16s16, d17s16);
d1x2s16 = vtrn_s16(d18s16, d19s16);
q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
d20s16 = vdup_n_s16((int16_t)cospi_8_64);
d21s16 = vdup_n_s16((int16_t)cospi_16_64);
q0x2s32 =
vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
d22s16 = vdup_n_s16((int16_t)cospi_24_64);
// stage 1
d23s16 = vadd_s16(d16s16, d18s16);
d24s16 = vsub_s16(d16s16, d18s16);
q15s32 = vmull_s16(d17s16, d22s16);
q1s32 = vmull_s16(d17s16, d20s16);
q13s32 = vmull_s16(d23s16, d21s16);
q14s32 = vmull_s16(d24s16, d21s16);
q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
d26s16 = vqrshrn_n_s32(q13s32, 14);
d27s16 = vqrshrn_n_s32(q14s32, 14);
d29s16 = vqrshrn_n_s32(q15s32, 14);
d28s16 = vqrshrn_n_s32(q1s32, 14);
q13s16 = vcombine_s16(d26s16, d27s16);
q14s16 = vcombine_s16(d28s16, d29s16);
// stage 2
q8s16 = vaddq_s16(q13s16, q14s16);
q9s16 = vsubq_s16(q13s16, q14s16);
d16s16 = vget_low_s16(q8s16);
d17s16 = vget_high_s16(q8s16);
d18s16 = vget_high_s16(q9s16); // vswp d18 d19
d19s16 = vget_low_s16(q9s16);
d0x2s16 = vtrn_s16(d16s16, d17s16);
d1x2s16 = vtrn_s16(d18s16, d19s16);
q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
q0x2s32 =
vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
// do the transform on columns
// stage 1
d23s16 = vadd_s16(d16s16, d18s16);
d24s16 = vsub_s16(d16s16, d18s16);
q15s32 = vmull_s16(d17s16, d22s16);
q1s32 = vmull_s16(d17s16, d20s16);
q13s32 = vmull_s16(d23s16, d21s16);
q14s32 = vmull_s16(d24s16, d21s16);
q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
d26s16 = vqrshrn_n_s32(q13s32, 14);
d27s16 = vqrshrn_n_s32(q14s32, 14);
d29s16 = vqrshrn_n_s32(q15s32, 14);
d28s16 = vqrshrn_n_s32(q1s32, 14);
q13s16 = vcombine_s16(d26s16, d27s16);
q14s16 = vcombine_s16(d28s16, d29s16);
// stage 2
q8s16 = vaddq_s16(q13s16, q14s16);
q9s16 = vsubq_s16(q13s16, q14s16);
q8s16 = vrshrq_n_s16(q8s16, 4);
q9s16 = vrshrq_n_s16(q9s16, 4);
d = dest;
d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 0);
d += dest_stride;
d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 1);
d += dest_stride;
d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 1);
d += dest_stride;
d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 0);
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32));
d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d = dest;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 0);
d += dest_stride;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 1);
d += dest_stride;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 1);
d += dest_stride;
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 0);
return;
}
void vp10_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest,
int dest_stride, int tx_type) {
uint8x8_t d26u8, d27u8;
int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16;
uint32x2_t d26u32, d27u32;
int16x8_t q3s16, q8s16, q9s16;
uint16x8_t q8u16, q9u16;
d26u32 = d27u32 = vdup_n_u32(0);
q8s16 = vld1q_s16(input);
q9s16 = vld1q_s16(input + 8);
TRANSPOSE4X4(&q8s16, &q9s16);
switch (tx_type) {
case 0: // idct_idct is not supported. Fall back to C
vp10_iht4x4_16_add_c(input, dest, dest_stride, tx_type);
return;
break;
case 1: // iadst_idct
// generate constants
GENERATE_COSINE_CONSTANTS(&d0s16, &d1s16, &d2s16);
GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
// first transform rows
IDCT4x4_1D(&d0s16, &d1s16, &d2s16, &q8s16, &q9s16);
// transpose the matrix
TRANSPOSE4X4(&q8s16, &q9s16);
// then transform columns
IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
break;
case 2: // idct_iadst
// generate constantsyy
GENERATE_COSINE_CONSTANTS(&d0s16, &d1s16, &d2s16);
GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
// first transform rows
IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
// transpose the matrix
TRANSPOSE4X4(&q8s16, &q9s16);
// then transform columns
IDCT4x4_1D(&d0s16, &d1s16, &d2s16, &q8s16, &q9s16);
break;
case 3: // iadst_iadst
// generate constants
GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
// first transform rows
IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
// transpose the matrix
TRANSPOSE4X4(&q8s16, &q9s16);
// then transform columns
IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
break;
default: // iadst_idct
assert(0);
break;
}
q8s16 = vrshrq_n_s16(q8s16, 4);
q9s16 = vrshrq_n_s16(q9s16, 4);
d26u32 = vld1_lane_u32((const uint32_t *)dest, d26u32, 0);
dest += dest_stride;
d26u32 = vld1_lane_u32((const uint32_t *)dest, d26u32, 1);
dest += dest_stride;
d27u32 = vld1_lane_u32((const uint32_t *)dest, d27u32, 0);
dest += dest_stride;
d27u32 = vld1_lane_u32((const uint32_t *)dest, d27u32, 1);
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32));
d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d27u8), 1);
dest -= dest_stride;
vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d27u8), 0);
dest -= dest_stride;
vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d26u8), 1);
dest -= dest_stride;
vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d26u8), 0);
return;
}
void vp8_short_idct4x4llm_neon(int16_t *input, unsigned char *pred_ptr,
int pred_stride, unsigned char *dst_ptr,
int dst_stride) {
int i;
uint32x2_t d6u32 = vdup_n_u32(0);
uint8x8_t d1u8;
int16x4_t d2, d3, d4, d5, d10, d11, d12, d13;
uint16x8_t q1u16;
int16x8_t q1s16, q2s16, q3s16, q4s16;
int32x2x2_t v2tmp0, v2tmp1;
int16x4x2_t v2tmp2, v2tmp3;
d2 = vld1_s16(input);
d3 = vld1_s16(input + 4);
d4 = vld1_s16(input + 8);
d5 = vld1_s16(input + 12);
// 1st for loop
q1s16 = vcombine_s16(d2, d4); // Swap d3 d4 here
q2s16 = vcombine_s16(d3, d5);
q3s16 = vqdmulhq_n_s16(q2s16, sinpi8sqrt2);
q4s16 = vqdmulhq_n_s16(q2s16, cospi8sqrt2minus1);
d12 = vqadd_s16(vget_low_s16(q1s16), vget_high_s16(q1s16)); // a1
d13 = vqsub_s16(vget_low_s16(q1s16), vget_high_s16(q1s16)); // b1
q3s16 = vshrq_n_s16(q3s16, 1);
q4s16 = vshrq_n_s16(q4s16, 1);
q3s16 = vqaddq_s16(q3s16, q2s16);
q4s16 = vqaddq_s16(q4s16, q2s16);
d10 = vqsub_s16(vget_low_s16(q3s16), vget_high_s16(q4s16)); // c1
d11 = vqadd_s16(vget_high_s16(q3s16), vget_low_s16(q4s16)); // d1
d2 = vqadd_s16(d12, d11);
d3 = vqadd_s16(d13, d10);
d4 = vqsub_s16(d13, d10);
d5 = vqsub_s16(d12, d11);
v2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
v2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
v2tmp2 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[0]),
vreinterpret_s16_s32(v2tmp1.val[0]));
v2tmp3 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[1]),
vreinterpret_s16_s32(v2tmp1.val[1]));
// 2nd for loop
q1s16 = vcombine_s16(v2tmp2.val[0], v2tmp3.val[0]);
q2s16 = vcombine_s16(v2tmp2.val[1], v2tmp3.val[1]);
q3s16 = vqdmulhq_n_s16(q2s16, sinpi8sqrt2);
q4s16 = vqdmulhq_n_s16(q2s16, cospi8sqrt2minus1);
d12 = vqadd_s16(vget_low_s16(q1s16), vget_high_s16(q1s16)); // a1
d13 = vqsub_s16(vget_low_s16(q1s16), vget_high_s16(q1s16)); // b1
q3s16 = vshrq_n_s16(q3s16, 1);
q4s16 = vshrq_n_s16(q4s16, 1);
q3s16 = vqaddq_s16(q3s16, q2s16);
q4s16 = vqaddq_s16(q4s16, q2s16);
d10 = vqsub_s16(vget_low_s16(q3s16), vget_high_s16(q4s16)); // c1
d11 = vqadd_s16(vget_high_s16(q3s16), vget_low_s16(q4s16)); // d1
d2 = vqadd_s16(d12, d11);
d3 = vqadd_s16(d13, d10);
d4 = vqsub_s16(d13, d10);
d5 = vqsub_s16(d12, d11);
d2 = vrshr_n_s16(d2, 3);
d3 = vrshr_n_s16(d3, 3);
d4 = vrshr_n_s16(d4, 3);
d5 = vrshr_n_s16(d5, 3);
v2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
v2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
v2tmp2 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[0]),
vreinterpret_s16_s32(v2tmp1.val[0]));
v2tmp3 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[1]),
vreinterpret_s16_s32(v2tmp1.val[1]));
q1s16 = vcombine_s16(v2tmp2.val[0], v2tmp2.val[1]);
q2s16 = vcombine_s16(v2tmp3.val[0], v2tmp3.val[1]);
// dc_only_idct_add
for (i = 0; i < 2; i++, q1s16 = q2s16) {
d6u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d6u32, 0);
pred_ptr += pred_stride;
d6u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d6u32, 1);
pred_ptr += pred_stride;
q1u16 = vaddw_u8(vreinterpretq_u16_s16(q1s16), vreinterpret_u8_u32(d6u32));
d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q1u16));
vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d1u8), 0);
dst_ptr += dst_stride;
vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d1u8), 1);
dst_ptr += dst_stride;
}
return;
}
int neon_new(DATA32* _p0, DATA32* _p1, DATA32* _p2, DATA32* _p3, DATA32* _ax, DATA32 _ay, DATA32* result, int len) {
int ay = _ay;
int i;
DATA32* pbuf = result;
uint16x4_t ay_16x4;
uint16x4_t p0_16x4;
uint16x4_t p2_16x4;
uint16x8_t ax_16x8;
uint16x8_t p0_p2_16x8;
uint16x8_t p1_p3_16x8;
uint16x8_t x255_16x8;
uint32x2_t p0_p2_32x2;
uint32x2_t p1_p3_32x2;
uint32x2_t res_32x2;
uint8x8_t p0_p2_8x8;
uint8x8_t p1_p3_8x8;
uint8x8_t p2_8x8;
uint16x4_t temp_16x4;
ay_16x4 = vdup_n_u16(ay);
x255_16x8 = vdupq_n_u16(0xff);
for(i = 0; i < len; i++) {
DATA32 p0 = *_p0++;
DATA32 p1 = *_p1++;
DATA32 p2 = *_p2++;
DATA32 p3 = *_p3++;
int ax = *_ax++;
if (p0 | p1 | p2 | p3)
{
ax_16x8 = vdupq_n_u16(ax);
p0_p2_32x2 = vset_lane_u32(p0, p0_p2_32x2, 0);
p0_p2_32x2 = vset_lane_u32(p2, p0_p2_32x2, 1);
p1_p3_32x2 = vset_lane_u32(p1, p1_p3_32x2, 0);
p1_p3_32x2 = vset_lane_u32(p3, p1_p3_32x2, 1);
p0_p2_8x8 = vreinterpret_u8_u32(p0_p2_32x2);
p1_p3_8x8 = vreinterpret_u8_u32(p1_p3_32x2);
p1_p3_16x8 = vmovl_u8(p1_p3_8x8);
p0_p2_16x8 = vmovl_u8(p0_p2_8x8);
p1_p3_16x8 = vsubq_u16(p1_p3_16x8, p0_p2_16x8);
p1_p3_16x8 = vmulq_u16(p1_p3_16x8, ax_16x8);
p1_p3_16x8 = vshrq_n_u16(p1_p3_16x8, 8);
p1_p3_16x8 = vaddq_u16(p1_p3_16x8, p0_p2_16x8);
p1_p3_16x8 = vandq_u16(p1_p3_16x8, x255_16x8);
p0_16x4 = vget_low_u16(p1_p3_16x8);
p2_16x4 = vget_high_u16(p1_p3_16x8);
p2_16x4 = vsub_u16(p2_16x4, p0_16x4);
p2_16x4 = vmul_u16(p2_16x4, ay_16x4);
p2_16x4 = vshr_n_u16(p2_16x4, 8);
p2_16x4 = vadd_u16(p2_16x4, p0_16x4);
p1_p3_16x8 = vcombine_u16(temp_16x4, p2_16x4);
p2_8x8 = vmovn_u16(p1_p3_16x8);
res_32x2 = vreinterpret_u32_u8(p2_8x8);
vst1_lane_u32(pbuf++, res_32x2, 1);
}
else
*pbuf++ = p0;
}
return 0;
}