void UpsampleRgbaLinePairNEON(const uint8_t *top_y, const uint8_t *bottom_y, const uint8_t *top_u, const uint8_t *top_v, const uint8_t *cur_u, const uint8_t *cur_v, uint8_t *top_dst, uint8_t *bottom_dst, int len)
{
int block;
uint8_t uv_buf[2 * 32 + 15];
uint8_t *const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15);
const int uv_len = (len + 1) >> 1;
const int num_blocks = (uv_len - 1) >> 3;
const int leftover = uv_len - num_blocks * 8;
const int last_pos = 1 + 16 * num_blocks;
const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1;
const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1;
const int16x4_t cf16 = vld1_s16(coef);
const int32x2_t cf32 = vmov_n_s32(76283);
const uint8x8_t u16 = vmov_n_u8(16);
const uint8x8_t u128 = vmov_n_u8(128);
for (block = 0; block < num_blocks; ++block) {
{
uint8x8_t a = vld1_u8(top_u);
uint8x8_t b = vld1_u8(top_u + 1);
uint8x8_t c = vld1_u8(cur_u);
uint8x8_t d = vld1_u8(cur_u + 1);
uint16x8_t al = vshll_n_u8(a, 1);
uint16x8_t bl = vshll_n_u8(b, 1);
uint16x8_t cl = vshll_n_u8(c, 1);
uint16x8_t dl = vshll_n_u8(d, 1);
uint8x8_t diag1, diag2;
uint16x8_t sl;
sl = vaddl_u8(a, b);
sl = vaddw_u8(sl, c);
sl = vaddw_u8(sl, d);
al = vaddq_u16(sl, al);
bl = vaddq_u16(sl, bl);
al = vaddq_u16(al, dl);
bl = vaddq_u16(bl, cl);
diag2 = vshrn_n_u16(al, 3);
diag1 = vshrn_n_u16(bl, 3);
a = vrhadd_u8(a, diag1);
b = vrhadd_u8(b, diag2);
c = vrhadd_u8(c, diag2);
d = vrhadd_u8(d, diag1);
{
const uint8x8x2_t a_b = {{ a, b }};
const uint8x8x2_t c_d = {{ c, d }};
vst2_u8(r_uv, a_b);
vst2_u8(r_uv + 32, c_d);
}
}
{
uint8x8_t a = vld1_u8(top_v);
uint8x8_t b = vld1_u8(top_v + 1);
uint8x8_t c = vld1_u8(cur_v);
uint8x8_t d = vld1_u8(cur_v + 1);
uint16x8_t al = vshll_n_u8(a, 1);
uint16x8_t bl = vshll_n_u8(b, 1);
uint16x8_t cl = vshll_n_u8(c, 1);
uint16x8_t dl = vshll_n_u8(d, 1);
uint8x8_t diag1, diag2;
uint16x8_t sl;
sl = vaddl_u8(a, b);
sl = vaddw_u8(sl, c);
sl = vaddw_u8(sl, d);
al = vaddq_u16(sl, al);
bl = vaddq_u16(sl, bl);
al = vaddq_u16(al, dl);
bl = vaddq_u16(bl, cl);
diag2 = vshrn_n_u16(al, 3);
diag1 = vshrn_n_u16(bl, 3);
a = vrhadd_u8(a, diag1);
b = vrhadd_u8(b, diag2);
c = vrhadd_u8(c, diag2);
d = vrhadd_u8(d, diag1);
{
const uint8x8x2_t a_b = {{ a, b }};
const uint8x8x2_t c_d = {{ c, d }};
vst2_u8(r_uv + 16, a_b);
vst2_u8(r_uv + 16 + 32, c_d);
}
}
{
if (top_y) {
{
int i;
for (i = 0; i < 16; i += 8) {
int off = ((16 * block + 1) + i) * 4;
uint8x8_t y = vld1_u8(top_y + (16 * block + 1) + i);
uint8x8_t u = vld1_u8((r_uv) + i);
uint8x8_t v = vld1_u8((r_uv) + i + 16);
int16x8_t yy = vreinterpretq_s16_u16(vsubl_u8(y, u16));
int16x8_t uu = vreinterpretq_s16_u16(vsubl_u8(u, u128));
int16x8_t vv = vreinterpretq_s16_u16(vsubl_u8(v, u128));
int16x8_t ud = vshlq_n_s16(uu, 1);
int16x8_t vd = vshlq_n_s16(vv, 1);
int32x4_t vrl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(vv), 1), vget_low_s16(vd), cf16, 0);
int32x4_t vrh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(vv), 1), vget_high_s16(vd), cf16, 0);
int16x8_t vr = vcombine_s16(vrshrn_n_s32(vrl, 16), vrshrn_n_s32(vrh, 16));
int32x4_t vl = vmovl_s16(vget_low_s16(vv));
int32x4_t vh = vmovl_s16(vget_high_s16(vv));
int32x4_t ugl = vmlal_lane_s16(vl, vget_low_s16(uu), cf16, 1);
int32x4_t ugh = vmlal_lane_s16(vh, vget_high_s16(uu), cf16, 1);
int32x4_t gcl = vqdmlal_lane_s16(ugl, vget_low_s16(vv), cf16, 2);
int32x4_t gch = vqdmlal_lane_s16(ugh, vget_high_s16(vv), cf16, 2);
int16x8_t gc = vcombine_s16(vrshrn_n_s32(gcl, 16), vrshrn_n_s32(gch, 16));
int32x4_t ubl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(uu), 1), vget_low_s16(ud), cf16, 3);
int32x4_t ubh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(uu), 1), vget_high_s16(ud), cf16, 3);
int16x8_t ub = vcombine_s16(vrshrn_n_s32(ubl, 16), vrshrn_n_s32(ubh, 16));
int32x4_t rl = vaddl_s16(vget_low_s16(yy), vget_low_s16(vr));
int32x4_t rh = vaddl_s16(vget_high_s16(yy), vget_high_s16(vr));
int32x4_t gl = vsubl_s16(vget_low_s16(yy), vget_low_s16(gc));
int32x4_t gh = vsubl_s16(vget_high_s16(yy), vget_high_s16(gc));
int32x4_t bl = vaddl_s16(vget_low_s16(yy), vget_low_s16(ub));
int32x4_t bh = vaddl_s16(vget_high_s16(yy), vget_high_s16(ub));
rl = vmulq_lane_s32(rl, cf32, 0);
rh = vmulq_lane_s32(rh, cf32, 0);
gl = vmulq_lane_s32(gl, cf32, 0);
gh = vmulq_lane_s32(gh, cf32, 0);
bl = vmulq_lane_s32(bl, cf32, 0);
bh = vmulq_lane_s32(bh, cf32, 0);
y = vqmovun_s16(vcombine_s16(vrshrn_n_s32(rl, 16), vrshrn_n_s32(rh, 16)));
u = vqmovun_s16(vcombine_s16(vrshrn_n_s32(gl, 16), vrshrn_n_s32(gh, 16)));
v = vqmovun_s16(vcombine_s16(vrshrn_n_s32(bl, 16), vrshrn_n_s32(bh, 16)));
do {
const uint8x8x4_t r_g_b_v255 = {{ y, u, v, vmov_n_u8(255) }};
vst4_u8(top_dst + off, r_g_b_v255);
} while (0);
}
}
}
if (bottom_y) {
{
int i;
for (i = 0; i < 16; i += 8) {
int off = ((16 * block + 1) + i) * 4;
uint8x8_t y = vld1_u8(bottom_y + (16 * block + 1) + i);
uint8x8_t u = vld1_u8(((r_uv) + 32) + i);
uint8x8_t v = vld1_u8(((r_uv) + 32) + i + 16);
int16x8_t yy = vreinterpretq_s16_u16(vsubl_u8(y, u16));
int16x8_t uu = vreinterpretq_s16_u16(vsubl_u8(u, u128));
int16x8_t vv = vreinterpretq_s16_u16(vsubl_u8(v, u128));
int16x8_t ud = vshlq_n_s16(uu, 1);
int16x8_t vd = vshlq_n_s16(vv, 1);
int32x4_t vrl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(vv), 1), vget_low_s16(vd), cf16, 0);
int32x4_t vrh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(vv), 1), vget_high_s16(vd), cf16, 0);
int16x8_t vr = vcombine_s16(vrshrn_n_s32(vrl, 16), vrshrn_n_s32(vrh, 16));
int32x4_t vl = vmovl_s16(vget_low_s16(vv));
int32x4_t vh = vmovl_s16(vget_high_s16(vv));
int32x4_t ugl = vmlal_lane_s16(vl, vget_low_s16(uu), cf16, 1);
int32x4_t ugh = vmlal_lane_s16(vh, vget_high_s16(uu), cf16, 1);
int32x4_t gcl = vqdmlal_lane_s16(ugl, vget_low_s16(vv), cf16, 2);
int32x4_t gch = vqdmlal_lane_s16(ugh, vget_high_s16(vv), cf16, 2);
int16x8_t gc = vcombine_s16(vrshrn_n_s32(gcl, 16), vrshrn_n_s32(gch, 16));
int32x4_t ubl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(uu), 1), vget_low_s16(ud), cf16, 3);
int32x4_t ubh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(uu), 1), vget_high_s16(ud), cf16, 3);
int16x8_t ub = vcombine_s16(vrshrn_n_s32(ubl, 16), vrshrn_n_s32(ubh, 16));
int32x4_t rl = vaddl_s16(vget_low_s16(yy), vget_low_s16(vr));
int32x4_t rh = vaddl_s16(vget_high_s16(yy), vget_high_s16(vr));
int32x4_t gl = vsubl_s16(vget_low_s16(yy), vget_low_s16(gc));
int32x4_t gh = vsubl_s16(vget_high_s16(yy), vget_high_s16(gc));
int32x4_t bl = vaddl_s16(vget_low_s16(yy), vget_low_s16(ub));
int32x4_t bh = vaddl_s16(vget_high_s16(yy), vget_high_s16(ub));
rl = vmulq_lane_s32(rl, cf32, 0);
rh = vmulq_lane_s32(rh, cf32, 0);
gl = vmulq_lane_s32(gl, cf32, 0);
gh = vmulq_lane_s32(gh, cf32, 0);
bl = vmulq_lane_s32(bl, cf32, 0);
bh = vmulq_lane_s32(bh, cf32, 0);
y = vqmovun_s16(vcombine_s16(vrshrn_n_s32(rl, 16), vrshrn_n_s32(rh, 16)));
u = vqmovun_s16(vcombine_s16(vrshrn_n_s32(gl, 16), vrshrn_n_s32(gh, 16)));
v = vqmovun_s16(vcombine_s16(vrshrn_n_s32(bl, 16), vrshrn_n_s32(bh, 16)));
do {
const uint8x8x4_t r_g_b_v255 = {{ y, u, v, vmov_n_u8(255) }};
vst4_u8(bottom_dst + off, r_g_b_v255);
} while (0);
}
}
}
}
}
}
void yuv422rgb_neon_int(const unsigned char * sourcep, int source_byte_count,
unsigned char * destp)
{
const unsigned char *source_endp;
const unsigned char *vector_endp;
int remainder;
const int16x8_t u_coeff = {0, -22, 113, 0, 0, -22, 113, 0};
const int16x8_t v_coeff = {90, -46, 0, 0, 90, -46, 0, 0};
const uint8x8_t zeroalpha = {0x0, 0x0, 0x0, 0xFF, 0x0, 0x0, 0x0, 0xFF};
const int16x8_t uvbias = {0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80};
int16x8_t mp0_rgba; /* macropixel 0's resulting RGBA RGBA pixels */
int16x8_t mp1_rgba; /* macropixel 1's resulting RGBA RGBA pixels */
uint8x8_t rawpixels; /* source pixels as {[YUYV]0 [YUYV]1} */
uint8x8_t rgba0, rgba1; /* rgba values as bytes */
uint8x16_t bothrgba;
uint8_t * destinationp; /* pointer into output buffer destp */
int16x8_t widerpixels; /* rawpixels promoted to shorts per component */
const uint8x8_t yselect = {0xff, 0xff, 0xff, 0xff,
0x00, 0x00, 0x00, 0x00};
/* we're working with things in 4-byte macropixels */
remainder = source_byte_count % 4;
source_endp = sourcep + source_byte_count;
vector_endp = source_endp - remainder;
destinationp = (uint8_t *)destp;
while (sourcep < vector_endp)
{
/* pull YUYV from 2 four byte macropixels starting at sourcep. */
/* we'll increment sourcep as we go to save the array dereference */
/* and separate increment instruction at the end of the loop */
/* load rawpixels with {[YUYV]0 [YUYV]1 } with byte components */
rawpixels = vld1_u8(sourcep);
sourcep += sizeof(rawpixels);
widerpixels = vreinterpretq_s16_u16(vmovl_u8(rawpixels));
/* ---------- process macropixel 0 --------------- */
/* take macropixel zero ([YUYV]0) from rawpixels and */
/* compute the two RGBA pixels that come from it. store */
/* those two pixels in mp0_rgba */
{
int16x8_t wider_yalpha;
int16x8_t u_vec, v_vec, uv_vec;
uint8x8_t narrow_yalpha;
uint8x8_t y0_vec, y1_vec;
int16x4_t yuyv;
/* narrow_yalpha is drawn from [YUYV]0 and formed into */
/* {Y0, Y0, Y0, alpha, Y1, Y1, Y1, alpha} */
/* this would have been a nice place for vtbx1_u8, but i */
/* can't get it to work. so i'll have to use vbsl_u8 instead. */
y0_vec = vdup_lane_u8(rawpixels, MP0_Y0);
y1_vec = vdup_lane_u8(rawpixels, MP0_Y1);
narrow_yalpha = vbsl_u8(yselect, y0_vec, y1_vec);
/* store ALPHA in elements 3 and 7 (after the RGB components) */
narrow_yalpha = vset_lane_u8(ALPHA, narrow_yalpha, 3);
narrow_yalpha = vset_lane_u8(ALPHA, narrow_yalpha, 7);
/* use vmovl_u8 to go from being unsigned 8-bit to */
/* unsigned 16-bit, the use vreinterpretq_s16_u16 to */
/* change interpretation from unsigned 16-bit to signed */
/* 16-bit. */
wider_yalpha = vreinterpretq_s16_u16(vmovl_u8(narrow_yalpha));
yuyv = vget_low_s16(widerpixels);
/* form a vector of the U component from MP0 */
u_vec = vdupq_lane_s16(yuyv, MP0_U);
/* subtract uvbias from u_vec */
u_vec = vsubq_s16(u_vec, uvbias);
/* form a vector of the V component from MP0 */
v_vec = vdupq_lane_s16(yuyv, MP0_V);
/* subtract uvbias from v_vec */
v_vec = vsubq_s16(v_vec, uvbias);
/* Multiply eight 16-bit values in u_vec by eight 16-bit */
/* values in u_coeff and store the results in u_vec. */
u_vec = vmulq_s16(u_vec, u_coeff);
/* likewise multiply eight 16-bit values in v_vec by */
/* v_coeff and store the results in v_vec */
v_vec = vmulq_s16(v_vec, v_coeff);
/* form uv_vec as the sum of u_vec & v_vec, then shift 6 places */
/* (dividing by 64) */
uv_vec = vaddq_s16(u_vec, v_vec);
uv_vec = vshrq_n_s16(uv_vec, 6);
/* now mp0_rgba = y_vec + u_vec + v_vec */
mp0_rgba = vaddq_s16(wider_yalpha, uv_vec);
}
/* ---------- process macropixel 1 --------------- */
/* take macropixel one ([YUYV]1) from rawpixels and */
/* compute the two RGBA pixels that come from it. store */
/* those two pixels in mp1_rgba */
{
int16x8_t wider_yalpha;
int16x8_t u_vec, v_vec, uv_vec;
uint8x8_t narrow_yalpha;
uint8x8_t y0_vec, y1_vec;
int16x4_t yuyv;
/* narrow_yalpha is drawn from [YUYV]1 and formed into */
/* {Y0, Y0, Y0, alpha, Y1, Y1, Y1, alpha} */
/* this would have been a nice place for vtbx1_u8, but i */
/* can't get it to work. so i'll have to use vbsl_u8 instead. */
y0_vec = vdup_lane_u8(rawpixels, MP1_Y0);
y1_vec = vdup_lane_u8(rawpixels, MP1_Y1);
narrow_yalpha = vbsl_u8(yselect, y0_vec, y1_vec);
narrow_yalpha = vset_lane_u8(ALPHA, narrow_yalpha, 3);
narrow_yalpha = vset_lane_u8(ALPHA, narrow_yalpha, 7);
/* use vmovl_u8 to go from being unsigned 8-bit to */
/* unsigned 16-bit, the use vreinterpretq_s16_u16 to */
wider_yalpha = vreinterpretq_s16_u16(vmovl_u8(narrow_yalpha));
yuyv = vget_high_s16(widerpixels);
u_vec = vdupq_lane_s16(yuyv, 1);
u_vec = vsubq_s16(u_vec, uvbias);
v_vec = vdupq_lane_s16(yuyv, 3);
v_vec = vsubq_s16(v_vec, uvbias);
/* Multiply eight 16-bit values in u_vec by eight 16-bit */
/* values in u_coeff and store the results in u_vec. */
u_vec = vmulq_s16(u_vec, u_coeff);
/* likewise multiply eight 16-bit values in v_vec by */
/* v_coeff and store the results in v_vec */
v_vec = vmulq_s16(v_vec, v_coeff);
/* form uv_vec as the sum of u_vec & v_vec, then shift 6 places */
/* (dividing by 64) */
uv_vec = vaddq_s16(u_vec, v_vec);
uv_vec = vshrq_n_s16(uv_vec, 6);
/* now mp1_rgba = y_vec + u_vec + v_vec */
mp1_rgba = vaddq_s16(wider_yalpha, uv_vec);
}
/* turn mp0_rgba from a vector of shorts to a vector of */
/* unsigned unsigned chars. this will saturate: clipping */
/* the values between 0 and 255. */
rgba0 = vqmovun_s16(mp0_rgba);
rgba1 = vqmovun_s16(mp1_rgba);
/* make it faster to copy these back out of vector registers into */
/* memory by combining rgba0 and rgba1 into the larger bothrgba. */
/* then store that back into memory at destinationp. */
bothrgba = vcombine_u8(rgba0, rgba1);
vst1q_u8(destinationp, bothrgba);
destinationp += 16;
}
}
void av1_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest,
int dest_stride, const TxfmParam *txfm_param) {
int i;
uint8_t *d1, *d2;
uint8x8_t d0u8, d1u8, d2u8, d3u8;
uint64x1_t d0u64, d1u64, d2u64, d3u64;
int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
uint16x8_t q8u16, q9u16, q10u16, q11u16;
q8s16 = vld1q_s16(input);
q9s16 = vld1q_s16(input + 8);
q10s16 = vld1q_s16(input + 8 * 2);
q11s16 = vld1q_s16(input + 8 * 3);
q12s16 = vld1q_s16(input + 8 * 4);
q13s16 = vld1q_s16(input + 8 * 5);
q14s16 = vld1q_s16(input + 8 * 6);
q15s16 = vld1q_s16(input + 8 * 7);
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
const TX_TYPE tx_type = txfm_param->tx_type;
switch (tx_type) {
case DCT_DCT: // idct_idct is not supported. Fall back to C
av1_iht8x8_64_add_c(input, dest, dest_stride, txfm_param);
return;
break;
case ADST_DCT: // iadst_idct
// generate IDCT constants
// GENERATE_IDCT_CONSTANTS
// first transform rows
IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
// transpose the matrix
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
// generate IADST constants
// GENERATE_IADST_CONSTANTS
// then transform columns
IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
break;
case DCT_ADST: // idct_iadst
// generate IADST constants
// GENERATE_IADST_CONSTANTS
// first transform rows
IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
// transpose the matrix
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
// generate IDCT constants
// GENERATE_IDCT_CONSTANTS
// then transform columns
IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
break;
case ADST_ADST: // iadst_iadst
// generate IADST constants
// GENERATE_IADST_CONSTANTS
// first transform rows
IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
// transpose the matrix
TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
// then transform columns
IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
&q15s16);
break;
default: // iadst_idct
assert(0);
break;
}
q8s16 = vrshrq_n_s16(q8s16, 5);
q9s16 = vrshrq_n_s16(q9s16, 5);
q10s16 = vrshrq_n_s16(q10s16, 5);
q11s16 = vrshrq_n_s16(q11s16, 5);
q12s16 = vrshrq_n_s16(q12s16, 5);
q13s16 = vrshrq_n_s16(q13s16, 5);
q14s16 = vrshrq_n_s16(q14s16, 5);
q15s16 = vrshrq_n_s16(q15s16, 5);
for (d1 = d2 = dest, i = 0; i < 2; i++) {
if (i != 0) {
q8s16 = q12s16;
q9s16 = q13s16;
q10s16 = q14s16;
q11s16 = q15s16;
}
d0u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d1u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d2u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
d3u64 = vld1_u64((uint64_t *)d1);
d1 += dest_stride;
q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
q10u16 =
vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
q11u16 =
vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
d2 += dest_stride;
vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
d2 += dest_stride;
}
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 vpx_idct4x4_16_add_neon(const tran_low_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 = load_tran_low_to_s16(input);
q9s16 = load_tran_low_to_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);
}
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;
}
}
}
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 idct_dequant_full_2x_neon(
int16_t *q,
int16_t *dq,
unsigned char *dst,
int stride) {
unsigned char *dst0, *dst1;
int32x2_t d28, d29, d30, d31;
int16x8_t q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11;
int16x8_t qEmpty = vdupq_n_s16(0);
int32x4x2_t q2tmp0, q2tmp1;
int16x8x2_t q2tmp2, q2tmp3;
int16x4_t dLow0, dLow1, dHigh0, dHigh1;
d28 = d29 = d30 = d31 = vdup_n_s32(0);
// load dq
q0 = vld1q_s16(dq);
dq += 8;
q1 = vld1q_s16(dq);
// load q
q2 = vld1q_s16(q);
vst1q_s16(q, qEmpty);
q += 8;
q3 = vld1q_s16(q);
vst1q_s16(q, qEmpty);
q += 8;
q4 = vld1q_s16(q);
vst1q_s16(q, qEmpty);
q += 8;
q5 = vld1q_s16(q);
vst1q_s16(q, qEmpty);
// load src from dst
dst0 = dst;
dst1 = dst + 4;
d28 = vld1_lane_s32((const int32_t *)dst0, d28, 0);
dst0 += stride;
d28 = vld1_lane_s32((const int32_t *)dst1, d28, 1);
dst1 += stride;
d29 = vld1_lane_s32((const int32_t *)dst0, d29, 0);
dst0 += stride;
d29 = vld1_lane_s32((const int32_t *)dst1, d29, 1);
dst1 += stride;
d30 = vld1_lane_s32((const int32_t *)dst0, d30, 0);
dst0 += stride;
d30 = vld1_lane_s32((const int32_t *)dst1, d30, 1);
dst1 += stride;
d31 = vld1_lane_s32((const int32_t *)dst0, d31, 0);
d31 = vld1_lane_s32((const int32_t *)dst1, d31, 1);
q2 = vmulq_s16(q2, q0);
q3 = vmulq_s16(q3, q1);
q4 = vmulq_s16(q4, q0);
q5 = vmulq_s16(q5, q1);
// vswp
dLow0 = vget_low_s16(q2);
dHigh0 = vget_high_s16(q2);
dLow1 = vget_low_s16(q4);
dHigh1 = vget_high_s16(q4);
q2 = vcombine_s16(dLow0, dLow1);
q4 = vcombine_s16(dHigh0, dHigh1);
dLow0 = vget_low_s16(q3);
dHigh0 = vget_high_s16(q3);
dLow1 = vget_low_s16(q5);
dHigh1 = vget_high_s16(q5);
q3 = vcombine_s16(dLow0, dLow1);
q5 = vcombine_s16(dHigh0, dHigh1);
q6 = vqdmulhq_n_s16(q4, sinpi8sqrt2);
q7 = vqdmulhq_n_s16(q5, sinpi8sqrt2);
q8 = vqdmulhq_n_s16(q4, cospi8sqrt2minus1);
q9 = vqdmulhq_n_s16(q5, cospi8sqrt2minus1);
q10 = vqaddq_s16(q2, q3);
q11 = vqsubq_s16(q2, q3);
q8 = vshrq_n_s16(q8, 1);
q9 = vshrq_n_s16(q9, 1);
q4 = vqaddq_s16(q4, q8);
q5 = vqaddq_s16(q5, q9);
q2 = vqsubq_s16(q6, q5);
q3 = vqaddq_s16(q7, q4);
q4 = vqaddq_s16(q10, q3);
q5 = vqaddq_s16(q11, q2);
q6 = vqsubq_s16(q11, q2);
q7 = vqsubq_s16(q10, q3);
q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6));
q2tmp1 = vtrnq_s32(vreinterpretq_s32_s16(q5), vreinterpretq_s32_s16(q7));
q2tmp2 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[0]),
vreinterpretq_s16_s32(q2tmp1.val[0]));
q2tmp3 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[1]),
vreinterpretq_s16_s32(q2tmp1.val[1]));
// loop 2
q8 = vqdmulhq_n_s16(q2tmp2.val[1], sinpi8sqrt2);
q9 = vqdmulhq_n_s16(q2tmp3.val[1], sinpi8sqrt2);
q10 = vqdmulhq_n_s16(q2tmp2.val[1], cospi8sqrt2minus1);
q11 = vqdmulhq_n_s16(q2tmp3.val[1], cospi8sqrt2minus1);
q2 = vqaddq_s16(q2tmp2.val[0], q2tmp3.val[0]);
q3 = vqsubq_s16(q2tmp2.val[0], q2tmp3.val[0]);
q10 = vshrq_n_s16(q10, 1);
q11 = vshrq_n_s16(q11, 1);
q10 = vqaddq_s16(q2tmp2.val[1], q10);
q11 = vqaddq_s16(q2tmp3.val[1], q11);
q8 = vqsubq_s16(q8, q11);
q9 = vqaddq_s16(q9, q10);
q4 = vqaddq_s16(q2, q9);
q5 = vqaddq_s16(q3, q8);
q6 = vqsubq_s16(q3, q8);
q7 = vqsubq_s16(q2, q9);
q4 = vrshrq_n_s16(q4, 3);
q5 = vrshrq_n_s16(q5, 3);
q6 = vrshrq_n_s16(q6, 3);
q7 = vrshrq_n_s16(q7, 3);
q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6));
q2tmp1 = vtrnq_s32(vreinterpretq_s32_s16(q5), vreinterpretq_s32_s16(q7));
q2tmp2 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[0]),
vreinterpretq_s16_s32(q2tmp1.val[0]));
q2tmp3 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[1]),
vreinterpretq_s16_s32(q2tmp1.val[1]));
q4 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp2.val[0]),
vreinterpret_u8_s32(d28)));
q5 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp2.val[1]),
vreinterpret_u8_s32(d29)));
q6 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp3.val[0]),
vreinterpret_u8_s32(d30)));
q7 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp3.val[1]),
vreinterpret_u8_s32(d31)));
d28 = vreinterpret_s32_u8(vqmovun_s16(q4));
d29 = vreinterpret_s32_u8(vqmovun_s16(q5));
d30 = vreinterpret_s32_u8(vqmovun_s16(q6));
d31 = vreinterpret_s32_u8(vqmovun_s16(q7));
dst0 = dst;
dst1 = dst + 4;
vst1_lane_s32((int32_t *)dst0, d28, 0);
dst0 += stride;
vst1_lane_s32((int32_t *)dst1, d28, 1);
dst1 += stride;
vst1_lane_s32((int32_t *)dst0, d29, 0);
dst0 += stride;
vst1_lane_s32((int32_t *)dst1, d29, 1);
dst1 += stride;
vst1_lane_s32((int32_t *)dst0, d30, 0);
dst0 += stride;
vst1_lane_s32((int32_t *)dst1, d30, 1);
dst1 += stride;
vst1_lane_s32((int32_t *)dst0, d31, 0);
vst1_lane_s32((int32_t *)dst1, d31, 1);
return;
}
