// Applies filter on 6 pixels (p2, p1, p0, q0, q1 and q2)
static WEBP_INLINE void DoFilter6(__m128i *p2, __m128i* p1, __m128i *p0,
__m128i* q0, __m128i* q1, __m128i *q2,
const __m128i* mask, int hev_thresh) {
__m128i a, not_hev;
const __m128i sign_bit = _mm_set1_epi8(0x80);
// compute hev mask
GET_NOTHEV(*p1, *p0, *q0, *q1, hev_thresh, not_hev);
// convert to signed values
FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
FLIP_SIGN_BIT2(*p2, *q2);
GET_BASE_DELTA(*p1, *p0, *q0, *q1, a);
{ // do simple filter on pixels with hev
const __m128i m = _mm_andnot_si128(not_hev, *mask);
const __m128i f = _mm_and_si128(a, m);
DO_SIMPLE_FILTER(*p0, *q0, f);
}
{ // do strong filter on pixels with not hev
const __m128i zero = _mm_setzero_si128();
const __m128i nine = _mm_set1_epi16(0x0900);
const __m128i sixty_three = _mm_set1_epi16(63);
const __m128i m = _mm_and_si128(not_hev, *mask);
const __m128i f = _mm_and_si128(a, m);
const __m128i f_lo = _mm_unpacklo_epi8(zero, f);
const __m128i f_hi = _mm_unpackhi_epi8(zero, f);
const __m128i f9_lo = _mm_mulhi_epi16(f_lo, nine); // Filter (lo) * 9
const __m128i f9_hi = _mm_mulhi_epi16(f_hi, nine); // Filter (hi) * 9
const __m128i f18_lo = _mm_add_epi16(f9_lo, f9_lo); // Filter (lo) * 18
const __m128i f18_hi = _mm_add_epi16(f9_hi, f9_hi); // Filter (hi) * 18
const __m128i a2_lo = _mm_add_epi16(f9_lo, sixty_three); // Filter * 9 + 63
const __m128i a2_hi = _mm_add_epi16(f9_hi, sixty_three); // Filter * 9 + 63
const __m128i a1_lo = _mm_add_epi16(f18_lo, sixty_three); // F... * 18 + 63
const __m128i a1_hi = _mm_add_epi16(f18_hi, sixty_three); // F... * 18 + 63
const __m128i a0_lo = _mm_add_epi16(f18_lo, a2_lo); // Filter * 27 + 63
const __m128i a0_hi = _mm_add_epi16(f18_hi, a2_hi); // Filter * 27 + 63
UPDATE_2PIXELS(*p2, *q2, a2_lo, a2_hi);
UPDATE_2PIXELS(*p1, *q1, a1_lo, a1_hi);
UPDATE_2PIXELS(*p0, *q0, a0_lo, a0_hi);
}
// unoffset
FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
FLIP_SIGN_BIT2(*p2, *q2);
}
// Applies filter on 4 pixels (p1, p0, q0 and q1)
static WEBP_INLINE void DoFilter4(__m128i* p1, __m128i *p0,
__m128i* q0, __m128i* q1,
const __m128i* mask, int hev_thresh) {
__m128i not_hev;
__m128i t1, t2, t3;
const __m128i sign_bit = _mm_set1_epi8(0x80);
// compute hev mask
GET_NOTHEV(*p1, *p0, *q0, *q1, hev_thresh, not_hev);
// convert to signed values
FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
t1 = _mm_subs_epi8(*p1, *q1); // p1 - q1
t1 = _mm_andnot_si128(not_hev, t1); // hev(p1 - q1)
t2 = _mm_subs_epi8(*q0, *p0); // q0 - p0
t1 = _mm_adds_epi8(t1, t2); // hev(p1 - q1) + 1 * (q0 - p0)
t1 = _mm_adds_epi8(t1, t2); // hev(p1 - q1) + 2 * (q0 - p0)
t1 = _mm_adds_epi8(t1, t2); // hev(p1 - q1) + 3 * (q0 - p0)
t1 = _mm_and_si128(t1, *mask); // mask filter values we don't care about
// Do +4 side
t2 = _mm_set1_epi8(4);
t2 = _mm_adds_epi8(t1, t2); // 3 * (q0 - p0) + (p1 - q1) + 4
SIGNED_SHIFT_N(t2, 3); // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 3
t3 = t2; // save t2
*q0 = _mm_subs_epi8(*q0, t2); // q0 -= t2
// Now do +3 side
t2 = _mm_set1_epi8(3);
t2 = _mm_adds_epi8(t1, t2); // +3 instead of +4
SIGNED_SHIFT_N(t2, 3); // (3 * (q0 - p0) + hev(p1 - q1) + 3) >> 3
*p0 = _mm_adds_epi8(*p0, t2); // p0 += t2
t2 = _mm_set1_epi8(1);
t3 = _mm_adds_epi8(t3, t2);
SIGNED_SHIFT_N(t3, 1); // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 4
t3 = _mm_and_si128(not_hev, t3); // if !hev
*q1 = _mm_subs_epi8(*q1, t3); // q1 -= t3
*p1 = _mm_adds_epi8(*p1, t3); // p1 += t3
// unoffset
FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
}
// Applies filter on 6 pixels (p2, p1, p0, q0, q1 and q2)
static WEBP_INLINE void DoFilter6(__m128i* const p2, __m128i* const p1,
__m128i* const p0, __m128i* const q0,
__m128i* const q1, __m128i* const q2,
const __m128i* const mask, int hev_thresh) {
const __m128i zero = _mm_setzero_si128();
const __m128i sign_bit = _mm_set1_epi8(0x80);
__m128i a, not_hev;
// compute hev mask
GetNotHEV(p1, p0, q0, q1, hev_thresh, ¬_hev);
FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
FLIP_SIGN_BIT2(*p2, *q2);
GetBaseDelta(p1, p0, q0, q1, &a);
{ // do simple filter on pixels with hev
const __m128i m = _mm_andnot_si128(not_hev, *mask);
__m128i f = _mm_and_si128(a, m); // insieme: dropped const
DoSimpleFilter(p0, q0, &f);
}
{ // do strong filter on pixels with not hev
const __m128i k9 = _mm_set1_epi16(0x0900);
const __m128i k63 = _mm_set1_epi16(63);
const __m128i m = _mm_and_si128(not_hev, *mask);
const __m128i f = _mm_and_si128(a, m);
const __m128i f_lo = _mm_unpacklo_epi8(zero, f);
const __m128i f_hi = _mm_unpackhi_epi8(zero, f);
const __m128i f9_lo = _mm_mulhi_epi16(f_lo, k9); // Filter (lo) * 9
const __m128i f9_hi = _mm_mulhi_epi16(f_hi, k9); // Filter (hi) * 9
__m128i a2_lo = _mm_add_epi16(f9_lo, k63); // Filter * 9 + 63 // insieme: dropped const
__m128i a2_hi = _mm_add_epi16(f9_hi, k63); // Filter * 9 + 63 // insieme: dropped const
__m128i a1_lo = _mm_add_epi16(a2_lo, f9_lo); // Filter * 18 + 63 // insieme: dropped const
__m128i a1_hi = _mm_add_epi16(a2_hi, f9_hi); // Filter * 18 + 63 // insieme: dropped const
__m128i a0_lo = _mm_add_epi16(a1_lo, f9_lo); // Filter * 27 + 63 // insieme: dropped const
__m128i a0_hi = _mm_add_epi16(a1_hi, f9_hi); // Filter * 27 + 63 // insieme: dropped const
Update2Pixels(p2, q2, &a2_lo, &a2_hi);
Update2Pixels(p1, q1, &a1_lo, &a1_hi);
Update2Pixels(p0, q0, &a0_lo, &a0_hi);
}
}
// Applies filter on 4 pixels (p1, p0, q0 and q1)
static WEBP_INLINE void DoFilter4(__m128i* const p1, __m128i* const p0,
__m128i* const q0, __m128i* const q1,
const __m128i* const mask, int hev_thresh) {
const __m128i sign_bit = _mm_set1_epi8(0x80);
const __m128i k64 = _mm_set1_epi8(0x40);
const __m128i zero = _mm_setzero_si128();
__m128i not_hev;
__m128i t1, t2, t3;
// compute hev mask
GetNotHEV(p1, p0, q0, q1, hev_thresh, ¬_hev);
// convert to signed values
FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
t1 = _mm_subs_epi8(*p1, *q1); // p1 - q1
t1 = _mm_andnot_si128(not_hev, t1); // hev(p1 - q1)
t2 = _mm_subs_epi8(*q0, *p0); // q0 - p0
t1 = _mm_adds_epi8(t1, t2); // hev(p1 - q1) + 1 * (q0 - p0)
t1 = _mm_adds_epi8(t1, t2); // hev(p1 - q1) + 2 * (q0 - p0)
t1 = _mm_adds_epi8(t1, t2); // hev(p1 - q1) + 3 * (q0 - p0)
t1 = _mm_and_si128(t1, *mask); // mask filter values we don't care about
t2 = _mm_set1_epi8(3);
t3 = _mm_set1_epi8(4);
t2 = _mm_adds_epi8(t1, t2); // 3 * (q0 - p0) + (p1 - q1) + 3
t3 = _mm_adds_epi8(t1, t3); // 3 * (q0 - p0) + (p1 - q1) + 4
SignedShift8b(&t2); // (3 * (q0 - p0) + hev(p1 - q1) + 3) >> 3
SignedShift8b(&t3); // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 3
*p0 = _mm_adds_epi8(*p0, t2); // p0 += t2
*q0 = _mm_subs_epi8(*q0, t3); // q0 -= t3
FLIP_SIGN_BIT2(*p0, *q0);
// this is equivalent to signed (a + 1) >> 1 calculation
t2 = _mm_add_epi8(t3, sign_bit);
t3 = _mm_avg_epu8(t2, zero);
t3 = _mm_sub_epi8(t3, k64);
t3 = _mm_and_si128(not_hev, t3); // if !hev
*q1 = _mm_subs_epi8(*q1, t3); // q1 -= t3
*p1 = _mm_adds_epi8(*p1, t3); // p1 += t3
FLIP_SIGN_BIT2(*p1, *q1);
}
