void test_vst4u8 (void)
{
uint8_t *arg0_uint8_t;
uint8x8x4_t arg1_uint8x8x4_t;
vst4_u8 (arg0_uint8_t, arg1_uint8x8x4_t);
}
void SkBlitLCD16Row_neon(SkPMColor dst[], const uint16_t src[],
SkColor color, int width, SkPMColor) {
int colA = SkColorGetA(color);
int colR = SkColorGetR(color);
int colG = SkColorGetG(color);
int colB = SkColorGetB(color);
colA = SkAlpha255To256(colA);
uint8x8_t vcolR, vcolG, vcolB;
uint16x8_t vcolA;
if (width >= 8) {
vcolA = vdupq_n_u16(colA);
vcolR = vdup_n_u8(colR);
vcolG = vdup_n_u8(colG);
vcolB = vdup_n_u8(colB);
}
while (width >= 8) {
uint8x8x4_t vdst;
uint16x8_t vmask;
uint16x8_t vmaskR, vmaskG, vmaskB;
vdst = vld4_u8((uint8_t*)dst);
vmask = vld1q_u16(src);
// Get all the color masks on 5 bits
vmaskR = vshrq_n_u16(vmask, SK_R16_SHIFT);
vmaskG = vshrq_n_u16(vshlq_n_u16(vmask, SK_R16_BITS),
SK_B16_BITS + SK_R16_BITS + 1);
vmaskB = vmask & vdupq_n_u16(SK_B16_MASK);
// Upscale to 0..32
vmaskR = vmaskR + vshrq_n_u16(vmaskR, 4);
vmaskG = vmaskG + vshrq_n_u16(vmaskG, 4);
vmaskB = vmaskB + vshrq_n_u16(vmaskB, 4);
vmaskR = vshrq_n_u16(vmaskR * vcolA, 8);
vmaskG = vshrq_n_u16(vmaskG * vcolA, 8);
vmaskB = vshrq_n_u16(vmaskB * vcolA, 8);
vdst.val[NEON_A] = vdup_n_u8(0xFF);
vdst.val[NEON_R] = SkBlend32_neon8(vcolR, vdst.val[NEON_R], vmaskR);
vdst.val[NEON_G] = SkBlend32_neon8(vcolG, vdst.val[NEON_G], vmaskG);
vdst.val[NEON_B] = SkBlend32_neon8(vcolB, vdst.val[NEON_B], vmaskB);
vst4_u8((uint8_t*)dst, vdst);
dst += 8;
src += 8;
width -= 8;
}
for (int i = 0; i < width; i++) {
dst[i] = SkBlendLCD16(colA, colR, colG, colB, dst[i], src[i]);
}
}
void SkBlitLCD16OpaqueRow_neon(SkPMColor dst[], const uint16_t src[],
SkColor color, int width,
SkPMColor opaqueDst) {
int colR = SkColorGetR(color);
int colG = SkColorGetG(color);
int colB = SkColorGetB(color);
uint8x8_t vcolR, vcolG, vcolB;
uint8x8_t vopqDstA, vopqDstR, vopqDstG, vopqDstB;
if (width >= 8) {
vcolR = vdup_n_u8(colR);
vcolG = vdup_n_u8(colG);
vcolB = vdup_n_u8(colB);
vopqDstA = vdup_n_u8(SkGetPackedA32(opaqueDst));
vopqDstR = vdup_n_u8(SkGetPackedR32(opaqueDst));
vopqDstG = vdup_n_u8(SkGetPackedG32(opaqueDst));
vopqDstB = vdup_n_u8(SkGetPackedB32(opaqueDst));
}
while (width >= 8) {
uint8x8x4_t vdst;
uint16x8_t vmask;
uint16x8_t vmaskR, vmaskG, vmaskB;
uint8x8_t vsel_trans, vsel_opq;
vdst = vld4_u8((uint8_t*)dst);
vmask = vld1q_u16(src);
// Prepare compare masks
vsel_trans = vmovn_u16(vceqq_u16(vmask, vdupq_n_u16(0)));
vsel_opq = vmovn_u16(vceqq_u16(vmask, vdupq_n_u16(0xFFFF)));
// Get all the color masks on 5 bits
vmaskR = vshrq_n_u16(vmask, SK_R16_SHIFT);
vmaskG = vshrq_n_u16(vshlq_n_u16(vmask, SK_R16_BITS),
SK_B16_BITS + SK_R16_BITS + 1);
vmaskB = vmask & vdupq_n_u16(SK_B16_MASK);
// Upscale to 0..32
vmaskR = vmaskR + vshrq_n_u16(vmaskR, 4);
vmaskG = vmaskG + vshrq_n_u16(vmaskG, 4);
vmaskB = vmaskB + vshrq_n_u16(vmaskB, 4);
vdst.val[NEON_A] = vbsl_u8(vsel_trans, vdst.val[NEON_A], vdup_n_u8(0xFF));
vdst.val[NEON_A] = vbsl_u8(vsel_opq, vopqDstA, vdst.val[NEON_A]);
vdst.val[NEON_R] = SkBlend32_neon8(vcolR, vdst.val[NEON_R], vmaskR);
vdst.val[NEON_G] = SkBlend32_neon8(vcolG, vdst.val[NEON_G], vmaskG);
vdst.val[NEON_B] = SkBlend32_neon8(vcolB, vdst.val[NEON_B], vmaskB);
vdst.val[NEON_R] = vbsl_u8(vsel_opq, vopqDstR, vdst.val[NEON_R]);
vdst.val[NEON_G] = vbsl_u8(vsel_opq, vopqDstG, vdst.val[NEON_G]);
vdst.val[NEON_B] = vbsl_u8(vsel_opq, vopqDstB, vdst.val[NEON_B]);
vst4_u8((uint8_t*)dst, vdst);
dst += 8;
src += 8;
width -= 8;
}
// Leftovers
for (int i = 0; i < width; i++) {
dst[i] = SkBlendLCD16Opaque(colR, colG, colB, dst[i], src[i],
opaqueDst);
}
}
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);
}
}
}
}
}
}
static void store_pixel_block(unsigned char* dst, PixelBlock& pblock, uint8x8_t const& r, uint8x8_t const& g, uint8x8_t const& b) {
pblock.val[0] = b;
pblock.val[1] = g;
pblock.val[2] = r;
vst4_u8(dst, pblock);
}
inline void vst4(u8 * ptr, const uint8x8x4_t & v) { return vst4_u8(ptr, v); }
//
// 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 combineUYVY(const Size2D &size,
const u8 * srcyBase, ptrdiff_t srcyStride,
const u8 * srcuBase, ptrdiff_t srcuStride,
const u8 * srcvBase, ptrdiff_t srcvStride,
u8 * dstBase, ptrdiff_t dstStride)
{
internal::assertSupportedConfiguration();
#ifdef CAROTENE_NEON
#ifndef __ANDROID__
size_t roiw32 = size.width >= 31 ? size.width - 31 : 0;
#endif
size_t roiw8 = size.width >= 7 ? size.width - 7 : 0;
for (size_t i = 0u; i < size.height; ++i)
{
const u8 * srcy = internal::getRowPtr(srcyBase, srcyStride, i);
const u8 * srcu = internal::getRowPtr(srcuBase, srcuStride, i);
const u8 * srcv = internal::getRowPtr(srcvBase, srcvStride, i);
u8 * dst = internal::getRowPtr(dstBase, dstStride, i);
size_t syj = 0u, sj = 0u, dj = 0u;
#ifndef __ANDROID__
for (; sj < roiw32; sj += 32, syj += 64, dj += 128)
{
internal::prefetch(srcy + syj);
internal::prefetch(srcu + sj);
internal::prefetch(srcv + sj);
uint8x16x2_t v_y = vld2q_u8(srcy + syj);
uint8x16x4_t v_dst;
v_dst.val[0] = vld1q_u8(srcu + sj);
v_dst.val[1] = v_y.val[0];
v_dst.val[2] = vld1q_u8(srcv + sj);
v_dst.val[3] = v_y.val[1];
vst4q_u8(dst + dj, v_dst);
v_y = vld2q_u8(srcy + syj + 32);
v_dst.val[0] = vld1q_u8(srcu + sj + 16);
v_dst.val[1] = v_y.val[0];
v_dst.val[2] = vld1q_u8(srcv + sj + 16);
v_dst.val[3] = v_y.val[1];
vst4q_u8(dst + dj + 64, v_dst);
}
#endif
for (; sj < roiw8; sj += 8, syj += 16, dj += 32)
{
uint8x8x2_t v_y = vld2_u8(srcy + syj);
uint8x8x4_t v_dst;
v_dst.val[0] = vld1_u8(srcu + sj);
v_dst.val[1] = v_y.val[0];
v_dst.val[2] = vld1_u8(srcv + sj);
v_dst.val[3] = v_y.val[1];
vst4_u8(dst + dj, v_dst);
}
for (; sj < size.width; ++sj, syj += 2, dj += 4)
{
dst[dj] = srcu[sj];
dst[dj + 1] = srcy[syj];
dst[dj + 2] = srcv[sj];
dst[dj + 3] = srcy[syj + 1];
}
}
#else
(void)size;
(void)srcyBase;
(void)srcyStride;
(void)srcuBase;
(void)srcuStride;
(void)srcvBase;
(void)srcvStride;
(void)dstBase;
(void)dstStride;
#endif
}