static INLINE void mbloop_filter_neon(uint8x8_t dblimit, // mblimit
uint8x8_t dlimit, // limit
uint8x8_t dthresh, // thresh
uint8x8_t d3u8, // p2
uint8x8_t d4u8, // p2
uint8x8_t d5u8, // p1
uint8x8_t d6u8, // p0
uint8x8_t d7u8, // q0
uint8x8_t d16u8, // q1
uint8x8_t d17u8, // q2
uint8x8_t d18u8, // q3
uint8x8_t *d0ru8, // p1
uint8x8_t *d1ru8, // p1
uint8x8_t *d2ru8, // p0
uint8x8_t *d3ru8, // q0
uint8x8_t *d4ru8, // q1
uint8x8_t *d5ru8) { // q1
uint32_t flat;
uint8x8_t d0u8, d1u8, d2u8, d19u8, d20u8, d21u8, d22u8, d23u8, d24u8;
uint8x8_t d25u8, d26u8, d27u8, d28u8, d29u8, d30u8, d31u8;
int16x8_t q15s16;
uint16x8_t q10u16, q14u16;
int8x8_t d21s8, d24s8, d25s8, d26s8, d28s8, d29s8, d30s8;
d19u8 = vabd_u8(d3u8, d4u8);
d20u8 = vabd_u8(d4u8, d5u8);
d21u8 = vabd_u8(d5u8, d6u8);
d22u8 = vabd_u8(d16u8, d7u8);
d23u8 = vabd_u8(d17u8, d16u8);
d24u8 = vabd_u8(d18u8, d17u8);
d19u8 = vmax_u8(d19u8, d20u8);
d20u8 = vmax_u8(d21u8, d22u8);
d25u8 = vabd_u8(d6u8, d4u8);
d23u8 = vmax_u8(d23u8, d24u8);
d26u8 = vabd_u8(d7u8, d17u8);
d19u8 = vmax_u8(d19u8, d20u8);
d24u8 = vabd_u8(d6u8, d7u8);
d27u8 = vabd_u8(d3u8, d6u8);
d28u8 = vabd_u8(d18u8, d7u8);
d19u8 = vmax_u8(d19u8, d23u8);
d23u8 = vabd_u8(d5u8, d16u8);
d24u8 = vqadd_u8(d24u8, d24u8);
d19u8 = vcge_u8(dlimit, d19u8);
d25u8 = vmax_u8(d25u8, d26u8);
d26u8 = vmax_u8(d27u8, d28u8);
d23u8 = vshr_n_u8(d23u8, 1);
d25u8 = vmax_u8(d25u8, d26u8);
d24u8 = vqadd_u8(d24u8, d23u8);
d20u8 = vmax_u8(d20u8, d25u8);
d23u8 = vdup_n_u8(1);
d24u8 = vcge_u8(dblimit, d24u8);
d21u8 = vcgt_u8(d21u8, dthresh);
d20u8 = vcge_u8(d23u8, d20u8);
d19u8 = vand_u8(d19u8, d24u8);
d23u8 = vcgt_u8(d22u8, dthresh);
d20u8 = vand_u8(d20u8, d19u8);
d22u8 = vdup_n_u8(0x80);
d23u8 = vorr_u8(d21u8, d23u8);
q10u16 = vcombine_u16(vreinterpret_u16_u8(d20u8), vreinterpret_u16_u8(d21u8));
d30u8 = vshrn_n_u16(q10u16, 4);
flat = vget_lane_u32(vreinterpret_u32_u8(d30u8), 0);
if (flat == 0xffffffff) { // Check for all 1's, power_branch_only
d27u8 = vdup_n_u8(3);
d21u8 = vdup_n_u8(2);
q14u16 = vaddl_u8(d6u8, d7u8);
q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
q14u16 = vmlal_u8(q14u16, d4u8, d21u8);
q14u16 = vaddw_u8(q14u16, d5u8);
*d0ru8 = vqrshrn_n_u16(q14u16, 3);
q14u16 = vsubw_u8(q14u16, d3u8);
q14u16 = vsubw_u8(q14u16, d4u8);
q14u16 = vaddw_u8(q14u16, d5u8);
q14u16 = vaddw_u8(q14u16, d16u8);
*d1ru8 = vqrshrn_n_u16(q14u16, 3);
q14u16 = vsubw_u8(q14u16, d3u8);
q14u16 = vsubw_u8(q14u16, d5u8);
q14u16 = vaddw_u8(q14u16, d6u8);
q14u16 = vaddw_u8(q14u16, d17u8);
*d2ru8 = vqrshrn_n_u16(q14u16, 3);
q14u16 = vsubw_u8(q14u16, d3u8);
q14u16 = vsubw_u8(q14u16, d6u8);
q14u16 = vaddw_u8(q14u16, d7u8);
q14u16 = vaddw_u8(q14u16, d18u8);
*d3ru8 = vqrshrn_n_u16(q14u16, 3);
q14u16 = vsubw_u8(q14u16, d4u8);
q14u16 = vsubw_u8(q14u16, d7u8);
q14u16 = vaddw_u8(q14u16, d16u8);
q14u16 = vaddw_u8(q14u16, d18u8);
*d4ru8 = vqrshrn_n_u16(q14u16, 3);
q14u16 = vsubw_u8(q14u16, d5u8);
q14u16 = vsubw_u8(q14u16, d16u8);
q14u16 = vaddw_u8(q14u16, d17u8);
q14u16 = vaddw_u8(q14u16, d18u8);
*d5ru8 = vqrshrn_n_u16(q14u16, 3);
} else {
d21u8 = veor_u8(d7u8, d22u8);
d24u8 = veor_u8(d6u8, d22u8);
d25u8 = veor_u8(d5u8, d22u8);
d26u8 = veor_u8(d16u8, d22u8);
d27u8 = vdup_n_u8(3);
d28s8 = vsub_s8(vreinterpret_s8_u8(d21u8), vreinterpret_s8_u8(d24u8));
d29s8 = vqsub_s8(vreinterpret_s8_u8(d25u8), vreinterpret_s8_u8(d26u8));
q15s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d27u8));
d29s8 = vand_s8(d29s8, vreinterpret_s8_u8(d23u8));
q15s16 = vaddw_s8(q15s16, d29s8);
d29u8 = vdup_n_u8(4);
d28s8 = vqmovn_s16(q15s16);
d28s8 = vand_s8(d28s8, vreinterpret_s8_u8(d19u8));
d30s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d27u8));
d29s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d29u8));
d30s8 = vshr_n_s8(d30s8, 3);
d29s8 = vshr_n_s8(d29s8, 3);
d24s8 = vqadd_s8(vreinterpret_s8_u8(d24u8), d30s8);
d21s8 = vqsub_s8(vreinterpret_s8_u8(d21u8), d29s8);
d29s8 = vrshr_n_s8(d29s8, 1);
d29s8 = vbic_s8(d29s8, vreinterpret_s8_u8(d23u8));
d25s8 = vqadd_s8(vreinterpret_s8_u8(d25u8), d29s8);
d26s8 = vqsub_s8(vreinterpret_s8_u8(d26u8), d29s8);
if (flat == 0) { // filter_branch_only
*d0ru8 = d4u8;
*d1ru8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
*d2ru8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
*d3ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
*d4ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
*d5ru8 = d17u8;
return;
}
d21u8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
d24u8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
d25u8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
d26u8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
d23u8 = vdup_n_u8(2);
q14u16 = vaddl_u8(d6u8, d7u8);
q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
q14u16 = vmlal_u8(q14u16, d4u8, d23u8);
d0u8 = vbsl_u8(d20u8, dblimit, d4u8);
q14u16 = vaddw_u8(q14u16, d5u8);
d1u8 = vbsl_u8(d20u8, dlimit, d25u8);
d30u8 = vqrshrn_n_u16(q14u16, 3);
q14u16 = vsubw_u8(q14u16, d3u8);
q14u16 = vsubw_u8(q14u16, d4u8);
q14u16 = vaddw_u8(q14u16, d5u8);
q14u16 = vaddw_u8(q14u16, d16u8);
d2u8 = vbsl_u8(d20u8, dthresh, d24u8);
d31u8 = vqrshrn_n_u16(q14u16, 3);
q14u16 = vsubw_u8(q14u16, d3u8);
q14u16 = vsubw_u8(q14u16, d5u8);
q14u16 = vaddw_u8(q14u16, d6u8);
q14u16 = vaddw_u8(q14u16, d17u8);
*d0ru8 = vbsl_u8(d20u8, d30u8, d0u8);
d23u8 = vqrshrn_n_u16(q14u16, 3);
q14u16 = vsubw_u8(q14u16, d3u8);
q14u16 = vsubw_u8(q14u16, d6u8);
q14u16 = vaddw_u8(q14u16, d7u8);
*d1ru8 = vbsl_u8(d20u8, d31u8, d1u8);
q14u16 = vaddw_u8(q14u16, d18u8);
*d2ru8 = vbsl_u8(d20u8, d23u8, d2u8);
d22u8 = vqrshrn_n_u16(q14u16, 3);
q14u16 = vsubw_u8(q14u16, d4u8);
q14u16 = vsubw_u8(q14u16, d7u8);
q14u16 = vaddw_u8(q14u16, d16u8);
d3u8 = vbsl_u8(d20u8, d3u8, d21u8);
q14u16 = vaddw_u8(q14u16, d18u8);
d4u8 = vbsl_u8(d20u8, d4u8, d26u8);
d6u8 = vqrshrn_n_u16(q14u16, 3);
q14u16 = vsubw_u8(q14u16, d5u8);
q14u16 = vsubw_u8(q14u16, d16u8);
q14u16 = vaddw_u8(q14u16, d17u8);
q14u16 = vaddw_u8(q14u16, d18u8);
d5u8 = vbsl_u8(d20u8, d5u8, d17u8);
d7u8 = vqrshrn_n_u16(q14u16, 3);
*d3ru8 = vbsl_u8(d20u8, d22u8, d3u8);
*d4ru8 = vbsl_u8(d20u8, d6u8, d4u8);
*d5ru8 = vbsl_u8(d20u8, d7u8, d5u8);
}
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);
}
}
}
}