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);
}
inline void v_mul_expand(const v_uint16x8& a, const v_uint16x8& b,
v_uint32x4& c, v_uint32x4& d)
{
c.val = vmull_u16(vget_low_u16(a.val), vget_low_u16(b.val));
d.val = vmull_u16(vget_high_u16(a.val), vget_high_u16(b.val));
}
uint16x4_t cv_16x4;
uint16x8_t c1_c2_16x8;
uint16x8_t c1_val1_16x8;
uint16x8_t c2_val3_16x8;
uint16x8_t cv_rv_16x8;
uint32x2_t c1_c2_32x2;
uint8x8_t c1_c2_8x8;
uint8x8_t val3_8x8;
uint16x8_t val3_16x8;
c1_c2_32x2 = vset_lane_u32(c1, c1_c2_32x2, 0);
c1_c2_32x2 = vset_lane_u32(c2, c1_c2_32x2, 1);
c1_c2_8x8 = vreinterpret_u8_u32(c1_c2_32x2);
c1_c2_16x8 = vmovl_u8(c1_c2_8x8);
c1_16x4 = vget_low_u16(c1_c2_16x8);
c2_16x4 = vget_high_u16(c1_c2_16x8);
# endif //COLSAME
# else
uint8x8_t val3_8x8;
uint16x8_t val3_16x8;
# endif //COLMUL
# endif //COLBLACK
# endif //SCALE_USING_NEON
while (ww > 0)
{
# ifdef COLBLACK
*d = 0xff000000; // col
# else
DATA32 ru = *_ru++;
DATA32 rv = *_rv++;
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);
}
void vp9_tm_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
int j, k;
uint16x8_t q0u16, q2u16, q3u16, q8u16, q10u16;
uint8x16_t q0u8, q1u8;
int16x8_t q0s16, q1s16, q8s16, q11s16;
uint16x4_t d20u16;
uint8x8_t d2u8, d3u8, d18u8, d22u8, d23u8;
q0u8 = vld1q_dup_u8(above - 1);
q1u8 = vld1q_u8(above);
q2u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
q3u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
for (k = 0; k < 2; k++, left += 8) {
d18u8 = vld1_u8(left);
q10u16 = vmovl_u8(d18u8);
d20u16 = vget_low_u16(q10u16);
for (j = 0; j < 2; j++, d20u16 = vget_high_u16(q10u16)) {
q0u16 = vdupq_lane_u16(d20u16, 0);
q8u16 = vdupq_lane_u16(d20u16, 1);
q1s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q2u16));
q0s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q3u16));
q11s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
vreinterpretq_s16_u16(q2u16));
q8s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
vreinterpretq_s16_u16(q3u16));
d2u8 = vqmovun_s16(q1s16);
d3u8 = vqmovun_s16(q0s16);
d22u8 = vqmovun_s16(q11s16);
d23u8 = vqmovun_s16(q8s16);
vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
dst += stride;
vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
dst += stride;
q0u16 = vdupq_lane_u16(d20u16, 2);
q8u16 = vdupq_lane_u16(d20u16, 3);
q1s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q2u16));
q0s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q3u16));
q11s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
vreinterpretq_s16_u16(q2u16));
q8s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
vreinterpretq_s16_u16(q3u16));
d2u8 = vqmovun_s16(q1s16);
d3u8 = vqmovun_s16(q0s16);
d22u8 = vqmovun_s16(q11s16);
d23u8 = vqmovun_s16(q8s16);
vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
dst += stride;
vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
dst += stride;
}
}
}
void vp8_short_fdct8x4_neon(
int16_t *input,
int16_t *output,
int pitch) {
int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
int16x4_t d16s16, d17s16, d26s16, d27s16, d28s16, d29s16;
uint16x4_t d28u16, d29u16;
uint16x8_t q14u16;
int16x8_t q0s16, q1s16, q2s16, q3s16;
int16x8_t q11s16, q12s16, q13s16, q14s16, q15s16, qEmptys16;
int32x4_t q9s32, q10s32, q11s32, q12s32;
int16x8x2_t v2tmp0, v2tmp1;
int32x4x2_t v2tmp2, v2tmp3;
d16s16 = vdup_n_s16(5352);
d17s16 = vdup_n_s16(2217);
q9s32 = vdupq_n_s32(14500);
q10s32 = vdupq_n_s32(7500);
// Part one
pitch >>= 1;
q0s16 = vld1q_s16(input);
input += pitch;
q1s16 = vld1q_s16(input);
input += pitch;
q2s16 = vld1q_s16(input);
input += pitch;
q3s16 = vld1q_s16(input);
v2tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q0s16),
vreinterpretq_s32_s16(q2s16));
v2tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q1s16),
vreinterpretq_s32_s16(q3s16));
v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]), // q0
vreinterpretq_s16_s32(v2tmp3.val[0])); // q1
v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]), // q2
vreinterpretq_s16_s32(v2tmp3.val[1])); // q3
q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]);
q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]);
q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]);
q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]);
q11s16 = vshlq_n_s16(q11s16, 3);
q12s16 = vshlq_n_s16(q12s16, 3);
q13s16 = vshlq_n_s16(q13s16, 3);
q14s16 = vshlq_n_s16(q14s16, 3);
q0s16 = vaddq_s16(q11s16, q12s16);
q2s16 = vsubq_s16(q11s16, q12s16);
q11s32 = q9s32;
q12s32 = q10s32;
d26s16 = vget_low_s16(q13s16);
d27s16 = vget_high_s16(q13s16);
d28s16 = vget_low_s16(q14s16);
d29s16 = vget_high_s16(q14s16);
q9s32 = vmlal_s16(q9s32, d28s16, d16s16);
q10s32 = vmlal_s16(q10s32, d28s16, d17s16);
q11s32 = vmlal_s16(q11s32, d29s16, d16s16);
q12s32 = vmlal_s16(q12s32, d29s16, d17s16);
q9s32 = vmlal_s16(q9s32, d26s16, d17s16);
q10s32 = vmlsl_s16(q10s32, d26s16, d16s16);
q11s32 = vmlal_s16(q11s32, d27s16, d17s16);
q12s32 = vmlsl_s16(q12s32, d27s16, d16s16);
d2s16 = vshrn_n_s32(q9s32, 12);
d6s16 = vshrn_n_s32(q10s32, 12);
d3s16 = vshrn_n_s32(q11s32, 12);
d7s16 = vshrn_n_s32(q12s32, 12);
q1s16 = vcombine_s16(d2s16, d3s16);
q3s16 = vcombine_s16(d6s16, d7s16);
// Part two
q9s32 = vdupq_n_s32(12000);
q10s32 = vdupq_n_s32(51000);
v2tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q0s16),
vreinterpretq_s32_s16(q2s16));
v2tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q1s16),
vreinterpretq_s32_s16(q3s16));
v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]), // q0
vreinterpretq_s16_s32(v2tmp3.val[0])); // q1
v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]), // q2
vreinterpretq_s16_s32(v2tmp3.val[1])); // q3
q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]);
q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]);
q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]);
q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]);
q15s16 = vdupq_n_s16(7);
q11s16 = vaddq_s16(q11s16, q15s16);
q0s16 = vaddq_s16(q11s16, q12s16);
q1s16 = vsubq_s16(q11s16, q12s16);
q11s32 = q9s32;
q12s32 = q10s32;
d0s16 = vget_low_s16(q0s16);
d1s16 = vget_high_s16(q0s16);
d2s16 = vget_low_s16(q1s16);
d3s16 = vget_high_s16(q1s16);
d0s16 = vshr_n_s16(d0s16, 4);
d4s16 = vshr_n_s16(d1s16, 4);
d2s16 = vshr_n_s16(d2s16, 4);
d6s16 = vshr_n_s16(d3s16, 4);
d26s16 = vget_low_s16(q13s16);
d27s16 = vget_high_s16(q13s16);
d28s16 = vget_low_s16(q14s16);
d29s16 = vget_high_s16(q14s16);
q9s32 = vmlal_s16(q9s32, d28s16, d16s16);
q10s32 = vmlal_s16(q10s32, d28s16, d17s16);
q11s32 = vmlal_s16(q11s32, d29s16, d16s16);
q12s32 = vmlal_s16(q12s32, d29s16, d17s16);
q9s32 = vmlal_s16(q9s32, d26s16, d17s16);
q10s32 = vmlsl_s16(q10s32, d26s16, d16s16);
q11s32 = vmlal_s16(q11s32, d27s16, d17s16);
q12s32 = vmlsl_s16(q12s32, d27s16, d16s16);
d1s16 = vshrn_n_s32(q9s32, 16);
d3s16 = vshrn_n_s32(q10s32, 16);
d5s16 = vshrn_n_s32(q11s32, 16);
d7s16 = vshrn_n_s32(q12s32, 16);
qEmptys16 = vdupq_n_s16(0);
q14u16 = vceqq_s16(q14s16, qEmptys16);
q14u16 = vmvnq_u16(q14u16);
d28u16 = vget_low_u16(q14u16);
d29u16 = vget_high_u16(q14u16);
d1s16 = vsub_s16(d1s16, vreinterpret_s16_u16(d28u16));
d5s16 = vsub_s16(d5s16, vreinterpret_s16_u16(d29u16));
q0s16 = vcombine_s16(d0s16, d1s16);
q1s16 = vcombine_s16(d2s16, d3s16);
q2s16 = vcombine_s16(d4s16, d5s16);
q3s16 = vcombine_s16(d6s16, d7s16);
vst1q_s16(output, q0s16);
vst1q_s16(output + 8, q1s16);
vst1q_s16(output + 16, q2s16);
vst1q_s16(output + 24, q3s16);
return;
}
//
// box blur a square array of pixels (power of 2, actually)
// if we insist on powers of 2, we don't need to special case some end-of-row/col conditions
// to a specific blur width
//
// also, we're using NEON to vectorize our arithmetic.
// we need to do a division along the way, but NEON doesn't support integer division.
// so rather than divide by, say "w", we multiply by magic(w).
// magic(w) is chosen so that the result of multiplying by it will be the same as
// dividing by w, except that the result will be in the high half of the result.
// yes, dorothy... this is what compilers do, too...
void NEONboxBlur(pixel *src, pixel *dest, unsigned int size, unsigned int blurRad) {
unsigned int wid = 2 * blurRad + 1;
// because NEON doesn't have integer division, we use "magic constants" that will give
// use the result of division by multiplication -- the upper half of the result will be
// (more or less) the result of the division.
// for this, we need to compute the magic numbers corresponding to a given divisor
struct magicu_info minfo = compute_unsigned_magic_info(wid, 16);
int16x8_t preshift = vdupq_n_s16(-minfo.pre_shift); // negative means shift right
int32x4_t postshift = vdupq_n_s32(-(minfo.post_shift+16)); // negative means shift right
uint16x4_t magic = vdup_n_u16(minfo.multiplier);
// fprintf(stderr,"width %5d, preshift %d, postshift %d + 16, increment %d, magic %d\n", wid,
// minfo.pre_shift, minfo.post_shift, minfo.increment, minfo.multiplier);
// if (minfo.pre_shift > 0) fprintf(stderr,"hey, not an odd number!\n");
int i, j, k, ch;
for (i = 0 ; i < size ; i+=8) {
// first, initialize the sum so that we can loop from 0 to size-1
// we'll initialize boxsum for index -1, so that we can move into 0 as part of our loop
uint16x8x4_t boxsum;
uint8x8x4_t firstpixel = vld4_u8((uint8_t *)(src + 0 * size + i));
for (ch = 0 ; ch < 4 ; ch++) {
// boxsum[ch] = blurRad * srcpixel[ch]
boxsum.val[ch] = vmulq_n_u16(vmovl_u8(firstpixel.val[ch]),(blurRad+1)+1);
}
for ( k = 1 ; k < blurRad ; k++) {
uint8x8x4_t srcpixel = vld4_u8((uint8_t *)(src + k * size + i));
for (ch = 0 ; ch < 4 ; ch++ ) {
boxsum.val[ch] = vaddw_u8(boxsum.val[ch], srcpixel.val[ch]);
}
}
int right = blurRad-1;
int left = -blurRad-1;
if (minfo.increment) {
for ( k = 0 ; k < size ; k++) {
// move to next pixel
unsigned int l = (left < 0)?0:left; // take off the old left
left++;
right++;
unsigned int r = (right < size)?right:(size-1); // but add the new right
uint8x8x4_t addpixel = vld4_u8((uint8_t *)(src + r * size + i));
uint8x8x4_t subpixel = vld4_u8((uint8_t *)(src + l * size + i));
for (ch = 0 ; ch < 4 ; ch++ ) {
// boxsum[ch] += addpixel[ch] - subpixel[ch];
boxsum.val[ch] = vsubw_u8(vaddw_u8(boxsum.val[ch], addpixel.val[ch]), subpixel.val[ch]);
}
uint8x8x4_t destpixel;
for (ch = 0 ; ch < 4 ; ch++ ) { // compute: destpixel = boxsum / wid
// since 16bit multiplication leads to 32bit results, we need to
// split our task into two chunks, for the hi and low half of our vector
// (because otherwise, it won't all fit into 128 bits)
// this is the meat of the magic division algorithm (see the include file...)
uint16x8_t bsum_preshifted = vshlq_u16(boxsum.val[ch],preshift);
// multiply by the magic number
uint32x4_t res_hi = vmull_u16(vget_high_u16(bsum_preshifted), magic);
res_hi = vaddw_u16(res_hi, magic);
// take the high half and post-shift
uint16x4_t q_hi = vmovn_u32(vshlq_u32(res_hi, postshift));
// pre-shift and multiply by the magic number
uint32x4_t res_lo = vmull_u16(vget_low_u16(bsum_preshifted), magic);
res_lo = vaddw_u16(res_lo, magic);
// take the high half and post-shift
uint16x4_t q_lo = vmovn_u32(vshlq_u32(res_lo, postshift));
destpixel.val[ch] = vqmovn_u16(vcombine_u16(q_lo, q_hi));
}
pixel block[8];
vst4_u8((uint8_t *)&block, destpixel);
for (j = 0 ; j < 8 ; j++ ) {
dest[(i + j)*size + k] = block[j];
}
// vst4_u8((uint8_t *)(dest + k * size + i), destpixel);
}
} else {
for ( k = 0 ; k < size ; k++) {
// move to next pixel
unsigned int l = (left < 0)?0:left; // take off the old left
left++;
right++;
unsigned int r = (right < size)?right:(size-1); // but add the new right
uint8x8x4_t addpixel = vld4_u8((uint8_t *)(src + r * size + i));
uint8x8x4_t subpixel = vld4_u8((uint8_t *)(src + l * size + i));
for (ch = 0 ; ch < 4 ; ch++ ) {
// boxsum[ch] += addpixel[ch] - subpixel[ch];
boxsum.val[ch] = vsubw_u8(vaddw_u8(boxsum.val[ch], addpixel.val[ch]), subpixel.val[ch]);
}
uint8x8x4_t destpixel;
for (ch = 0 ; ch < 4 ; ch++ ) { // compute: destpixel = boxsum / wid
// since 16bit multiplication leads to 32bit results, we need to
// split our task into two chunks, for the hi and low half of our vector
// (because otherwise, it won't all fit into 128 bits)
// this is the meat of the magic division algorithm (see the include file...)
uint16x8_t bsum_preshifted = vshlq_u16(boxsum.val[ch],preshift);
// multiply by the magic number
// take the high half and post-shift
uint32x4_t res_hi = vmull_u16(vget_high_u16(bsum_preshifted), magic);
uint16x4_t q_hi = vmovn_u32(vshlq_u32(res_hi, postshift));
// multiply by the magic number
// take the high half and post-shift
uint32x4_t res_lo = vmull_u16(vget_low_u16(bsum_preshifted), magic);
uint16x4_t q_lo = vmovn_u32(vshlq_u32(res_lo, postshift));
destpixel.val[ch] = vqmovn_u16(vcombine_u16(q_lo, q_hi));
}
pixel block[8];
vst4_u8((uint8_t *)&block, destpixel);
for (j = 0 ; j < 8 ; j++ ) {
dest[(i + j)*size + k] = block[j];
}
// vst4_u8((uint8_t *)(dest + k * size + i), destpixel);
}
}
}
}
void vpx_highbd_convolve_avg_neon(const uint8_t *src8, ptrdiff_t src_stride,
uint8_t *dst8, ptrdiff_t dst_stride,
const int16_t *filter_x, int filter_x_stride,
const int16_t *filter_y, int filter_y_stride,
int w, int h, int bd) {
const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
(void)filter_x;
(void)filter_x_stride;
(void)filter_y;
(void)filter_y_stride;
(void)bd;
if (w < 8) { // avg4
uint16x4_t s0, s1, d0, d1;
uint16x8_t s01, d01;
do {
s0 = vld1_u16(src);
d0 = vld1_u16(dst);
src += src_stride;
s1 = vld1_u16(src);
d1 = vld1_u16(dst + dst_stride);
src += src_stride;
s01 = vcombine_u16(s0, s1);
d01 = vcombine_u16(d0, d1);
d01 = vrhaddq_u16(s01, d01);
vst1_u16(dst, vget_low_u16(d01));
dst += dst_stride;
vst1_u16(dst, vget_high_u16(d01));
dst += dst_stride;
h -= 2;
} while (h > 0);
} else if (w == 8) { // avg8
uint16x8_t s0, s1, d0, d1;
do {
s0 = vld1q_u16(src);
d0 = vld1q_u16(dst);
src += src_stride;
s1 = vld1q_u16(src);
d1 = vld1q_u16(dst + dst_stride);
src += src_stride;
d0 = vrhaddq_u16(s0, d0);
d1 = vrhaddq_u16(s1, d1);
vst1q_u16(dst, d0);
dst += dst_stride;
vst1q_u16(dst, d1);
dst += dst_stride;
h -= 2;
} while (h > 0);
} else if (w < 32) { // avg16
uint16x8_t s0l, s0h, s1l, s1h, d0l, d0h, d1l, d1h;
do {
s0l = vld1q_u16(src);
s0h = vld1q_u16(src + 8);
d0l = vld1q_u16(dst);
d0h = vld1q_u16(dst + 8);
src += src_stride;
s1l = vld1q_u16(src);
s1h = vld1q_u16(src + 8);
d1l = vld1q_u16(dst + dst_stride);
d1h = vld1q_u16(dst + dst_stride + 8);
src += src_stride;
d0l = vrhaddq_u16(s0l, d0l);
d0h = vrhaddq_u16(s0h, d0h);
d1l = vrhaddq_u16(s1l, d1l);
d1h = vrhaddq_u16(s1h, d1h);
vst1q_u16(dst, d0l);
vst1q_u16(dst + 8, d0h);
dst += dst_stride;
vst1q_u16(dst, d1l);
vst1q_u16(dst + 8, d1h);
dst += dst_stride;
h -= 2;
} while (h > 0);
} else if (w == 32) { // avg32
uint16x8_t s0, s1, s2, s3, d0, d1, d2, d3;
do {
s0 = vld1q_u16(src);
s1 = vld1q_u16(src + 8);
s2 = vld1q_u16(src + 16);
s3 = vld1q_u16(src + 24);
d0 = vld1q_u16(dst);
d1 = vld1q_u16(dst + 8);
d2 = vld1q_u16(dst + 16);
d3 = vld1q_u16(dst + 24);
src += src_stride;
d0 = vrhaddq_u16(s0, d0);
d1 = vrhaddq_u16(s1, d1);
d2 = vrhaddq_u16(s2, d2);
d3 = vrhaddq_u16(s3, d3);
vst1q_u16(dst, d0);
vst1q_u16(dst + 8, d1);
vst1q_u16(dst + 16, d2);
vst1q_u16(dst + 24, d3);
dst += dst_stride;
s0 = vld1q_u16(src);
s1 = vld1q_u16(src + 8);
s2 = vld1q_u16(src + 16);
s3 = vld1q_u16(src + 24);
d0 = vld1q_u16(dst);
d1 = vld1q_u16(dst + 8);
d2 = vld1q_u16(dst + 16);
d3 = vld1q_u16(dst + 24);
src += src_stride;
d0 = vrhaddq_u16(s0, d0);
d1 = vrhaddq_u16(s1, d1);
d2 = vrhaddq_u16(s2, d2);
d3 = vrhaddq_u16(s3, d3);
vst1q_u16(dst, d0);
vst1q_u16(dst + 8, d1);
vst1q_u16(dst + 16, d2);
vst1q_u16(dst + 24, d3);
dst += dst_stride;
h -= 2;
} while (h > 0);
} else { // avg64
uint16x8_t s0, s1, s2, s3, d0, d1, d2, d3;
do {
s0 = vld1q_u16(src);
s1 = vld1q_u16(src + 8);
s2 = vld1q_u16(src + 16);
s3 = vld1q_u16(src + 24);
d0 = vld1q_u16(dst);
d1 = vld1q_u16(dst + 8);
d2 = vld1q_u16(dst + 16);
d3 = vld1q_u16(dst + 24);
d0 = vrhaddq_u16(s0, d0);
d1 = vrhaddq_u16(s1, d1);
d2 = vrhaddq_u16(s2, d2);
d3 = vrhaddq_u16(s3, d3);
vst1q_u16(dst, d0);
vst1q_u16(dst + 8, d1);
vst1q_u16(dst + 16, d2);
vst1q_u16(dst + 24, d3);
s0 = vld1q_u16(src + 32);
s1 = vld1q_u16(src + 40);
s2 = vld1q_u16(src + 48);
s3 = vld1q_u16(src + 56);
d0 = vld1q_u16(dst + 32);
d1 = vld1q_u16(dst + 40);
d2 = vld1q_u16(dst + 48);
d3 = vld1q_u16(dst + 56);
d0 = vrhaddq_u16(s0, d0);
d1 = vrhaddq_u16(s1, d1);
d2 = vrhaddq_u16(s2, d2);
d3 = vrhaddq_u16(s3, d3);
vst1q_u16(dst + 32, d0);
vst1q_u16(dst + 40, d1);
vst1q_u16(dst + 48, d2);
vst1q_u16(dst + 56, d3);
src += src_stride;
dst += dst_stride;
} while (--h);
}
}
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;
}
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;
}
void WebRtcAecm_CalcLinearEnergiesNeon(AecmCore* aecm,
const uint16_t* far_spectrum,
int32_t* echo_est,
uint32_t* far_energy,
uint32_t* echo_energy_adapt,
uint32_t* echo_energy_stored) {
int16_t* start_stored_p = aecm->channelStored;
int16_t* start_adapt_p = aecm->channelAdapt16;
int32_t* echo_est_p = echo_est;
const int16_t* end_stored_p = aecm->channelStored + PART_LEN;
const uint16_t* far_spectrum_p = far_spectrum;
int16x8_t store_v, adapt_v;
uint16x8_t spectrum_v;
uint32x4_t echo_est_v_low, echo_est_v_high;
uint32x4_t far_energy_v, echo_stored_v, echo_adapt_v;
far_energy_v = vdupq_n_u32(0);
echo_adapt_v = vdupq_n_u32(0);
echo_stored_v = vdupq_n_u32(0);
// Get energy for the delayed far end signal and estimated
// echo using both stored and adapted channels.
// The C code:
// for (i = 0; i < PART_LEN1; i++) {
// echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i],
// far_spectrum[i]);
// (*far_energy) += (uint32_t)(far_spectrum[i]);
// *echo_energy_adapt += aecm->channelAdapt16[i] * far_spectrum[i];
// (*echo_energy_stored) += (uint32_t)echo_est[i];
// }
while (start_stored_p < end_stored_p) {
spectrum_v = vld1q_u16(far_spectrum_p);
adapt_v = vld1q_s16(start_adapt_p);
store_v = vld1q_s16(start_stored_p);
far_energy_v = vaddw_u16(far_energy_v, vget_low_u16(spectrum_v));
far_energy_v = vaddw_u16(far_energy_v, vget_high_u16(spectrum_v));
echo_est_v_low = vmull_u16(vreinterpret_u16_s16(vget_low_s16(store_v)),
vget_low_u16(spectrum_v));
echo_est_v_high = vmull_u16(vreinterpret_u16_s16(vget_high_s16(store_v)),
vget_high_u16(spectrum_v));
vst1q_s32(echo_est_p, vreinterpretq_s32_u32(echo_est_v_low));
vst1q_s32(echo_est_p + 4, vreinterpretq_s32_u32(echo_est_v_high));
echo_stored_v = vaddq_u32(echo_est_v_low, echo_stored_v);
echo_stored_v = vaddq_u32(echo_est_v_high, echo_stored_v);
echo_adapt_v = vmlal_u16(echo_adapt_v,
vreinterpret_u16_s16(vget_low_s16(adapt_v)),
vget_low_u16(spectrum_v));
echo_adapt_v = vmlal_u16(echo_adapt_v,
vreinterpret_u16_s16(vget_high_s16(adapt_v)),
vget_high_u16(spectrum_v));
start_stored_p += 8;
start_adapt_p += 8;
far_spectrum_p += 8;
echo_est_p += 8;
}
AddLanes(far_energy, far_energy_v);
AddLanes(echo_energy_stored, echo_stored_v);
AddLanes(echo_energy_adapt, echo_adapt_v);
echo_est[PART_LEN] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[PART_LEN],
far_spectrum[PART_LEN]);
*echo_energy_stored += (uint32_t)echo_est[PART_LEN];
*far_energy += (uint32_t)far_spectrum[PART_LEN];
*echo_energy_adapt += aecm->channelAdapt16[PART_LEN] * far_spectrum[PART_LEN];
}
// CHECK-LABEL: define <4 x i16> @test_vget_high_u16(<8 x i16> %a) #0 {
// CHECK: [[SHUFFLE_I:%.*]] = shufflevector <8 x i16> %a, <8 x i16> %a, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
// CHECK: ret <4 x i16> [[SHUFFLE_I]]
uint16x4_t test_vget_high_u16(uint16x8_t a) {
return vget_high_u16(a);
}
void vp9_tm_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
const uint8_t *above, const uint8_t *left) {
int j, k;
uint16x8_t q0u16, q3u16, q8u16, q9u16, q10u16, q11u16;
uint8x16_t q0u8, q1u8, q2u8;
int16x8_t q12s16, q13s16, q14s16, q15s16;
uint16x4_t d6u16;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d26u8;
q0u8 = vld1q_dup_u8(above - 1);
q1u8 = vld1q_u8(above);
q2u8 = vld1q_u8(above + 16);
q8u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
q9u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
q10u16 = vsubl_u8(vget_low_u8(q2u8), vget_low_u8(q0u8));
q11u16 = vsubl_u8(vget_high_u8(q2u8), vget_high_u8(q0u8));
for (k = 0; k < 4; k++, left += 8) {
d26u8 = vld1_u8(left);
q3u16 = vmovl_u8(d26u8);
d6u16 = vget_low_u16(q3u16);
for (j = 0; j < 2; j++, d6u16 = vget_high_u16(q3u16)) {
q0u16 = vdupq_lane_u16(d6u16, 0);
q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q8u16));
q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q9u16));
q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q10u16));
q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q11u16));
d0u8 = vqmovun_s16(q12s16);
d1u8 = vqmovun_s16(q13s16);
d2u8 = vqmovun_s16(q14s16);
d3u8 = vqmovun_s16(q15s16);
q0u8 = vcombine_u8(d0u8, d1u8);
q1u8 = vcombine_u8(d2u8, d3u8);
vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
dst += stride;
q0u16 = vdupq_lane_u16(d6u16, 1);
q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q8u16));
q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q9u16));
q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q10u16));
q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q11u16));
d0u8 = vqmovun_s16(q12s16);
d1u8 = vqmovun_s16(q13s16);
d2u8 = vqmovun_s16(q14s16);
d3u8 = vqmovun_s16(q15s16);
q0u8 = vcombine_u8(d0u8, d1u8);
q1u8 = vcombine_u8(d2u8, d3u8);
vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
dst += stride;
q0u16 = vdupq_lane_u16(d6u16, 2);
q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q8u16));
q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q9u16));
q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q10u16));
q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q11u16));
d0u8 = vqmovun_s16(q12s16);
d1u8 = vqmovun_s16(q13s16);
d2u8 = vqmovun_s16(q14s16);
d3u8 = vqmovun_s16(q15s16);
q0u8 = vcombine_u8(d0u8, d1u8);
q1u8 = vcombine_u8(d2u8, d3u8);
vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
dst += stride;
q0u16 = vdupq_lane_u16(d6u16, 3);
q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q8u16));
q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q9u16));
q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q10u16));
q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
vreinterpretq_s16_u16(q11u16));
d0u8 = vqmovun_s16(q12s16);
d1u8 = vqmovun_s16(q13s16);
d2u8 = vqmovun_s16(q14s16);
d3u8 = vqmovun_s16(q15s16);
q0u8 = vcombine_u8(d0u8, d1u8);
q1u8 = vcombine_u8(d2u8, d3u8);
vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
dst += stride;
}
}
}
inline uint16x4_t vget_high(const uint16x8_t & v) { return vget_high_u16(v); }
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);
}
void meanStdDev(const Size2D &size,
const u16 * srcBase, ptrdiff_t srcStride,
f32 * pMean, f32 * pStdDev)
{
internal::assertSupportedConfiguration();
#ifdef CAROTENE_NEON
size_t blockSize0 = 1 << 10, roiw4 = size.width & ~3;
f64 fsum = 0.0f, fsqsum = 0.0f;
f32 arsum[8];
uint32x4_t v_zero = vdupq_n_u32(0u), v_sum;
float32x4_t v_zero_f = vdupq_n_f32(0.0f), v_sqsum;
for (size_t i = 0; i < size.height; ++i)
{
const u16 * src = internal::getRowPtr(srcBase, srcStride, i);
size_t j = 0u;
while (j < roiw4)
{
size_t blockSize = std::min(roiw4 - j, blockSize0) + j;
v_sum = v_zero;
v_sqsum = v_zero_f;
for ( ; j + 16 < blockSize ; j += 16)
{
internal::prefetch(src + j);
uint16x8_t v_src0 = vld1q_u16(src + j), v_src1 = vld1q_u16(src + j + 8);
// 0
uint32x4_t v_srclo = vmovl_u16(vget_low_u16(v_src0));
uint32x4_t v_srchi = vmovl_u16(vget_high_u16(v_src0));
v_sum = vaddq_u32(v_sum, vaddq_u32(v_srclo, v_srchi));
float32x4_t v_srclo_f = vcvtq_f32_u32(v_srclo);
float32x4_t v_srchi_f = vcvtq_f32_u32(v_srchi);
v_sqsum = vmlaq_f32(v_sqsum, v_srclo_f, v_srclo_f);
v_sqsum = vmlaq_f32(v_sqsum, v_srchi_f, v_srchi_f);
// 1
v_srclo = vmovl_u16(vget_low_u16(v_src1));
v_srchi = vmovl_u16(vget_high_u16(v_src1));
v_sum = vaddq_u32(v_sum, vaddq_u32(v_srclo, v_srchi));
v_srclo_f = vcvtq_f32_u32(v_srclo);
v_srchi_f = vcvtq_f32_u32(v_srchi);
v_sqsum = vmlaq_f32(v_sqsum, v_srclo_f, v_srclo_f);
v_sqsum = vmlaq_f32(v_sqsum, v_srchi_f, v_srchi_f);
}
for ( ; j < blockSize; j += 4)
{
uint32x4_t v_src = vmovl_u16(vld1_u16(src + j));
float32x4_t v_src_f = vcvtq_f32_u32(v_src);
v_sum = vaddq_u32(v_sum, v_src);
v_sqsum = vmlaq_f32(v_sqsum, v_src_f, v_src_f);
}
vst1q_f32(arsum, vcvtq_f32_u32(v_sum));
vst1q_f32(arsum + 4, v_sqsum);
fsum += (f64)arsum[0] + arsum[1] + arsum[2] + arsum[3];
fsqsum += (f64)arsum[4] + arsum[5] + arsum[6] + arsum[7];
}
// collect a few last elements in the current row
for ( ; j < size.width; ++j)
{
f32 srcval = src[j];
fsum += srcval;
fsqsum += srcval * srcval;
}
}
// calc mean and stddev
f64 itotal = 1.0 / size.total();
f64 mean = fsum * itotal;
f64 stddev = sqrt(std::max(fsqsum * itotal - mean * mean, 0.0));
if (pMean)
*pMean = mean;
if (pStdDev)
*pStdDev = stddev;
#else
(void)size;
(void)srcBase;
(void)srcStride;
(void)pMean;
(void)pStdDev;
#endif
}
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);
}
bool decode_yuv_neon(unsigned char* out, unsigned char const* y, unsigned char const* uv, int width, int height, unsigned char fill_alpha=0xff)
{
// pre-condition : width, height must be even
if (0!=(width&1) || width<2 || 0!=(height&1) || height<2 || !out || !y || !uv)
return false;
// in & out pointers
unsigned char* dst = out;
// constants
int const stride = width*trait::bytes_per_pixel;
int const itHeight = height>>1;
int const itWidth = width>>3;
uint8x8_t const Yshift = vdup_n_u8(16);
int16x8_t const half = vdupq_n_u16(128);
int32x4_t const rounding = vdupq_n_s32(128);
// tmp variable
uint16x8_t t;
// pixel block to temporary store 8 pixels
typename trait::PixelBlock pblock = trait::init_pixelblock(fill_alpha);
for (int j=0; j<itHeight; ++j, y+=width, dst+=stride) {
for (int i=0; i<itWidth; ++i, y+=8, uv+=8, dst+=(8*trait::bytes_per_pixel)) {
t = vmovl_u8(vqsub_u8(vld1_u8(y), Yshift));
int32x4_t const Y00 = vmulq_n_u32(vmovl_u16(vget_low_u16(t)), 298);
int32x4_t const Y01 = vmulq_n_u32(vmovl_u16(vget_high_u16(t)), 298);
t = vmovl_u8(vqsub_u8(vld1_u8(y+width), Yshift));
int32x4_t const Y10 = vmulq_n_u32(vmovl_u16(vget_low_u16(t)), 298);
int32x4_t const Y11 = vmulq_n_u32(vmovl_u16(vget_high_u16(t)), 298);
// trait::loadvu pack 4 sets of uv into a uint8x8_t, layout : { v0,u0, v1,u1, v2,u2, v3,u3 }
t = vsubq_s16((int16x8_t)vmovl_u8(trait::loadvu(uv)), half);
// UV.val[0] : v0, v1, v2, v3
// UV.val[1] : u0, u1, u2, u3
int16x4x2_t const UV = vuzp_s16(vget_low_s16(t), vget_high_s16(t));
// tR : 128+409V
// tG : 128-100U-208V
// tB : 128+516U
int32x4_t const tR = vmlal_n_s16(rounding, UV.val[0], 409);
int32x4_t const tG = vmlal_n_s16(vmlal_n_s16(rounding, UV.val[0], -208), UV.val[1], -100);
int32x4_t const tB = vmlal_n_s16(rounding, UV.val[1], 516);
int32x4x2_t const R = vzipq_s32(tR, tR); // [tR0, tR0, tR1, tR1] [ tR2, tR2, tR3, tR3]
int32x4x2_t const G = vzipq_s32(tG, tG); // [tG0, tG0, tG1, tG1] [ tG2, tG2, tG3, tG3]
int32x4x2_t const B = vzipq_s32(tB, tB); // [tB0, tB0, tB1, tB1] [ tB2, tB2, tB3, tB3]
// upper 8 pixels
trait::store_pixel_block(dst, pblock,
vshrn_n_u16(vcombine_u16(vqmovun_s32(vaddq_s32(R.val[0], Y00)), vqmovun_s32(vaddq_s32(R.val[1], Y01))), 8),
vshrn_n_u16(vcombine_u16(vqmovun_s32(vaddq_s32(G.val[0], Y00)), vqmovun_s32(vaddq_s32(G.val[1], Y01))), 8),
vshrn_n_u16(vcombine_u16(vqmovun_s32(vaddq_s32(B.val[0], Y00)), vqmovun_s32(vaddq_s32(B.val[1], Y01))), 8));
// lower 8 pixels
trait::store_pixel_block(dst+stride, pblock,
vshrn_n_u16(vcombine_u16(vqmovun_s32(vaddq_s32(R.val[0], Y10)), vqmovun_s32(vaddq_s32(R.val[1], Y11))), 8),
vshrn_n_u16(vcombine_u16(vqmovun_s32(vaddq_s32(G.val[0], Y10)), vqmovun_s32(vaddq_s32(G.val[1], Y11))), 8),
vshrn_n_u16(vcombine_u16(vqmovun_s32(vaddq_s32(B.val[0], Y10)), vqmovun_s32(vaddq_s32(B.val[1], Y11))), 8));
}
}
return true;
}
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;
}
