Alien-FreeImage
view release on metacpan or search on metacpan
src/Source/LibWebP/src/dsp/dsp.dec_sse2.c view on Meta::CPAN
__m128i dst2 = _mm_cvtsi32_si128(*(int*)(dst + 2 * BPS));
__m128i dst3 = _mm_cvtsi32_si128(*(int*)(dst + 3 * BPS));
// Convert to 16b.
dst0 = _mm_unpacklo_epi8(dst0, zero);
dst1 = _mm_unpacklo_epi8(dst1, zero);
dst2 = _mm_unpacklo_epi8(dst2, zero);
dst3 = _mm_unpacklo_epi8(dst3, zero);
// Add the inverse transform.
dst0 = _mm_adds_epi16(dst0, _mm_srai_epi16(m0, 3));
dst1 = _mm_adds_epi16(dst1, _mm_srai_epi16(m1, 3));
dst2 = _mm_adds_epi16(dst2, _mm_srai_epi16(m2, 3));
dst3 = _mm_adds_epi16(dst3, _mm_srai_epi16(m3, 3));
// Unsigned saturate to 8b.
dst0 = _mm_packus_epi16(dst0, dst0);
dst1 = _mm_packus_epi16(dst1, dst1);
dst2 = _mm_packus_epi16(dst2, dst2);
dst3 = _mm_packus_epi16(dst3, dst3);
// Store the results.
*(int*)(dst + 0 * BPS) = _mm_cvtsi128_si32(dst0);
*(int*)(dst + 1 * BPS) = _mm_cvtsi128_si32(dst1);
*(int*)(dst + 2 * BPS) = _mm_cvtsi128_si32(dst2);
*(int*)(dst + 3 * BPS) = _mm_cvtsi128_si32(dst3);
}
#undef MUL
#endif // USE_TRANSFORM_AC3
//------------------------------------------------------------------------------
// Loop Filter (Paragraph 15)
// Compute abs(p - q) = subs(p - q) OR subs(q - p)
#define MM_ABS(p, q) _mm_or_si128( \
_mm_subs_epu8((q), (p)), \
_mm_subs_epu8((p), (q)))
// Shift each byte of "x" by 3 bits while preserving by the sign bit.
static WEBP_INLINE void SignedShift8b(__m128i* const x) {
const __m128i zero = _mm_setzero_si128();
const __m128i signs = _mm_cmpgt_epi8(zero, *x);
const __m128i lo_0 = _mm_unpacklo_epi8(*x, signs); // s8 -> s16 sign extend
const __m128i hi_0 = _mm_unpackhi_epi8(*x, signs);
const __m128i lo_1 = _mm_srai_epi16(lo_0, 3);
const __m128i hi_1 = _mm_srai_epi16(hi_0, 3);
*x = _mm_packs_epi16(lo_1, hi_1);
}
#define FLIP_SIGN_BIT2(a, b) { \
a = _mm_xor_si128(a, sign_bit); \
b = _mm_xor_si128(b, sign_bit); \
}
#define FLIP_SIGN_BIT4(a, b, c, d) { \
FLIP_SIGN_BIT2(a, b); \
FLIP_SIGN_BIT2(c, d); \
}
// input/output is uint8_t
static WEBP_INLINE void GetNotHEV(const __m128i* const p1,
const __m128i* const p0,
const __m128i* const q0,
const __m128i* const q1,
int hev_thresh, __m128i* const not_hev) {
const __m128i zero = _mm_setzero_si128();
const __m128i t_1 = MM_ABS(*p1, *p0);
const __m128i t_2 = MM_ABS(*q1, *q0);
const __m128i h = _mm_set1_epi8(hev_thresh);
const __m128i t_3 = _mm_subs_epu8(t_1, h); // abs(p1 - p0) - hev_tresh
const __m128i t_4 = _mm_subs_epu8(t_2, h); // abs(q1 - q0) - hev_tresh
*not_hev = _mm_or_si128(t_3, t_4);
*not_hev = _mm_cmpeq_epi8(*not_hev, zero); // not_hev <= t1 && not_hev <= t2
}
// input pixels are int8_t
static WEBP_INLINE void GetBaseDelta(const __m128i* const p1,
const __m128i* const p0,
const __m128i* const q0,
const __m128i* const q1,
__m128i* const delta) {
// beware of addition order, for saturation!
const __m128i p1_q1 = _mm_subs_epi8(*p1, *q1); // p1 - q1
const __m128i q0_p0 = _mm_subs_epi8(*q0, *p0); // q0 - p0
const __m128i s1 = _mm_adds_epi8(p1_q1, q0_p0); // p1 - q1 + 1 * (q0 - p0)
const __m128i s2 = _mm_adds_epi8(q0_p0, s1); // p1 - q1 + 2 * (q0 - p0)
const __m128i s3 = _mm_adds_epi8(q0_p0, s2); // p1 - q1 + 3 * (q0 - p0)
*delta = s3;
}
// input and output are int8_t
static WEBP_INLINE void DoSimpleFilter(__m128i* const p0, __m128i* const q0,
const __m128i* const fl) {
const __m128i k3 = _mm_set1_epi8(3);
const __m128i k4 = _mm_set1_epi8(4);
__m128i v3 = _mm_adds_epi8(*fl, k3);
__m128i v4 = _mm_adds_epi8(*fl, k4);
SignedShift8b(&v4); // v4 >> 3
SignedShift8b(&v3); // v3 >> 3
*q0 = _mm_subs_epi8(*q0, v4); // q0 -= v4
*p0 = _mm_adds_epi8(*p0, v3); // p0 += v3
}
// Updates values of 2 pixels at MB edge during complex filtering.
// Update operations:
// q = q - delta and p = p + delta; where delta = [(a_hi >> 7), (a_lo >> 7)]
// Pixels 'pi' and 'qi' are int8_t on input, uint8_t on output (sign flip).
static WEBP_INLINE void Update2Pixels(__m128i* const pi, __m128i* const qi,
const __m128i* const a0_lo,
const __m128i* const a0_hi) {
const __m128i a1_lo = _mm_srai_epi16(*a0_lo, 7);
const __m128i a1_hi = _mm_srai_epi16(*a0_hi, 7);
const __m128i delta = _mm_packs_epi16(a1_lo, a1_hi);
const __m128i sign_bit = _mm_set1_epi8(0x80);
*pi = _mm_adds_epi8(*pi, delta);
*qi = _mm_subs_epi8(*qi, delta);
FLIP_SIGN_BIT2(*pi, *qi);
}
// input pixels are uint8_t
static WEBP_INLINE void NeedsFilter(const __m128i* const p1,
const __m128i* const p0,
const __m128i* const q0,
const __m128i* const q1,
int thresh, __m128i* const mask) {
const __m128i m_thresh = _mm_set1_epi8(thresh);
const __m128i t1 = MM_ABS(*p1, *q1); // abs(p1 - q1)
const __m128i kFE = _mm_set1_epi8(0xFE);
const __m128i t2 = _mm_and_si128(t1, kFE); // set lsb of each byte to zero
const __m128i t3 = _mm_srli_epi16(t2, 1); // abs(p1 - q1) / 2
const __m128i t4 = MM_ABS(*p0, *q0); // abs(p0 - q0)
const __m128i t5 = _mm_adds_epu8(t4, t4); // abs(p0 - q0) * 2
const __m128i t6 = _mm_adds_epu8(t5, t3); // abs(p0-q0)*2 + abs(p1-q1)/2
const __m128i t7 = _mm_subs_epu8(t6, m_thresh); // mask <= m_thresh
*mask = _mm_cmpeq_epi8(t7, _mm_setzero_si128());
}
//------------------------------------------------------------------------------
// Edge filtering functions
// Applies filter on 2 pixels (p0 and q0)
static WEBP_INLINE void DoFilter2(__m128i* const p1, __m128i* const p0,
__m128i* const q0, __m128i* const q1,
int thresh) {
__m128i a, mask;
const __m128i sign_bit = _mm_set1_epi8(0x80);
// convert p1/q1 to int8_t (for GetBaseDelta)
const __m128i p1s = _mm_xor_si128(*p1, sign_bit);
const __m128i q1s = _mm_xor_si128(*q1, sign_bit);
NeedsFilter(p1, p0, q0, q1, thresh, &mask);
FLIP_SIGN_BIT2(*p0, *q0);
GetBaseDelta(&p1s, p0, q0, &q1s, &a);
a = _mm_and_si128(a, mask); // mask filter values we don't care about
DoSimpleFilter(p0, q0, &a);
FLIP_SIGN_BIT2(*p0, *q0);
}
// 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);
}
// 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);
const __m128i f = _mm_and_si128(a, m);
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
const __m128i a2_lo = _mm_add_epi16(f9_lo, k63); // Filter * 9 + 63
const __m128i a2_hi = _mm_add_epi16(f9_hi, k63); // Filter * 9 + 63
const __m128i a1_lo = _mm_add_epi16(a2_lo, f9_lo); // Filter * 18 + 63
const __m128i a1_hi = _mm_add_epi16(a2_hi, f9_hi); // Filter * 18 + 63
const __m128i a0_lo = _mm_add_epi16(a1_lo, f9_lo); // Filter * 27 + 63
const __m128i a0_hi = _mm_add_epi16(a1_hi, f9_hi); // Filter * 27 + 63
Update2Pixels(p2, q2, &a2_lo, &a2_hi);
Update2Pixels(p1, q1, &a1_lo, &a1_hi);
Update2Pixels(p0, q0, &a0_lo, &a0_hi);
}
}
// reads 8 rows across a vertical edge.
//
// TODO(somnath): Investigate _mm_shuffle* also see if it can be broken into
// two Load4x4() to avoid code duplication.
static WEBP_INLINE void Load8x4(const uint8_t* const b, int stride,
__m128i* const p, __m128i* const q) {
__m128i t1, t2;
// Load 0th, 1st, 4th and 5th rows
__m128i r0 = _mm_cvtsi32_si128(*(const int*)&b[0 * stride]); // 03 02 01 00
__m128i r1 = _mm_cvtsi32_si128(*(const int*)&b[1 * stride]); // 13 12 11 10
__m128i r4 = _mm_cvtsi32_si128(*(const int*)&b[4 * stride]); // 43 42 41 40
__m128i r5 = _mm_cvtsi32_si128(*(const int*)&b[5 * stride]); // 53 52 51 50
r0 = _mm_unpacklo_epi32(r0, r4); // 43 42 41 40 03 02 01 00
r1 = _mm_unpacklo_epi32(r1, r5); // 53 52 51 50 13 12 11 10
// t1 = 53 43 52 42 51 41 50 40 13 03 12 02 11 01 10 00
t1 = _mm_unpacklo_epi8(r0, r1);
// Load 2nd, 3rd, 6th and 7th rows
r0 = _mm_cvtsi32_si128(*(const int*)&b[2 * stride]); // 23 22 21 22
r1 = _mm_cvtsi32_si128(*(const int*)&b[3 * stride]); // 33 32 31 30
r4 = _mm_cvtsi32_si128(*(const int*)&b[6 * stride]); // 63 62 61 60
r5 = _mm_cvtsi32_si128(*(const int*)&b[7 * stride]); // 73 72 71 70
r0 = _mm_unpacklo_epi32(r0, r4); // 63 62 61 60 23 22 21 20
r1 = _mm_unpacklo_epi32(r1, r5); // 73 72 71 70 33 32 31 30
// t2 = 73 63 72 62 71 61 70 60 33 23 32 22 31 21 30 20
t2 = _mm_unpacklo_epi8(r0, r1);
// t1 = 33 23 13 03 32 22 12 02 31 21 11 01 30 20 10 00
// t2 = 73 63 53 43 72 62 52 42 71 61 51 41 70 60 50 40
r0 = t1;
t1 = _mm_unpacklo_epi16(t1, t2);
t2 = _mm_unpackhi_epi16(r0, t2);
( run in 1.233 second using v1.01-cache-2.11-cpan-941387dca55 )