diff options
Diffstat (limited to 'src/3rdparty/libwebp/src/dsp/dec_sse2.c')
-rw-r--r-- | src/3rdparty/libwebp/src/dsp/dec_sse2.c | 526 |
1 files changed, 274 insertions, 252 deletions
diff --git a/src/3rdparty/libwebp/src/dsp/dec_sse2.c b/src/3rdparty/libwebp/src/dsp/dec_sse2.c index 150c559..c37a637 100644 --- a/src/3rdparty/libwebp/src/dsp/dec_sse2.c +++ b/src/3rdparty/libwebp/src/dsp/dec_sse2.c @@ -26,7 +26,7 @@ //------------------------------------------------------------------------------ // Transforms (Paragraph 14.4) -static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) { +static void Transform(const int16_t* in, uint8_t* dst, int do_two) { // This implementation makes use of 16-bit fixed point versions of two // multiply constants: // K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16 @@ -246,7 +246,7 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) { #if defined(USE_TRANSFORM_AC3) #define MUL(a, b) (((a) * (b)) >> 16) -static void TransformAC3SSE2(const int16_t* in, uint8_t* dst) { +static void TransformAC3(const int16_t* in, uint8_t* dst) { static const int kC1 = 20091 + (1 << 16); static const int kC2 = 35468; const __m128i A = _mm_set1_epi16(in[0] + 4); @@ -298,20 +298,15 @@ static void TransformAC3SSE2(const int16_t* in, uint8_t* dst) { _mm_subs_epu8((q), (p)), \ _mm_subs_epu8((p), (q))) -// Shift each byte of "a" by N bits while preserving by the sign bit. -// -// It first shifts the lower bytes of the words and then the upper bytes and -// then merges the results together. -#define SIGNED_SHIFT_N(a, N) { \ - __m128i t = a; \ - t = _mm_slli_epi16(t, 8); \ - t = _mm_srai_epi16(t, N); \ - t = _mm_srli_epi16(t, 8); \ - \ - a = _mm_srai_epi16(a, N + 8); \ - a = _mm_slli_epi16(a, 8); \ - \ - a = _mm_or_si128(t, a); \ +// 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) { \ @@ -324,103 +319,123 @@ static void TransformAC3SSE2(const int16_t* in, uint8_t* dst) { FLIP_SIGN_BIT2(c, d); \ } -#define GET_NOTHEV(p1, p0, q0, q1, hev_thresh, 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 */\ -} - -#define GET_BASE_DELTA(p1, p0, q0, q1, o) { \ - const __m128i qp0 = _mm_subs_epi8(q0, p0); /* q0 - p0 */ \ - o = _mm_subs_epi8(p1, q1); /* p1 - q1 */ \ - o = _mm_adds_epi8(o, qp0); /* p1 - q1 + 1 * (q0 - p0) */ \ - o = _mm_adds_epi8(o, qp0); /* p1 - q1 + 2 * (q0 - p0) */ \ - o = _mm_adds_epi8(o, qp0); /* p1 - q1 + 3 * (q0 - p0) */ \ -} - -#define DO_SIMPLE_FILTER(p0, q0, fl) { \ - const __m128i three = _mm_set1_epi8(3); \ - const __m128i four = _mm_set1_epi8(4); \ - __m128i v3 = _mm_adds_epi8(fl, three); \ - __m128i v4 = _mm_adds_epi8(fl, four); \ - \ - /* Do +4 side */ \ - SIGNED_SHIFT_N(v4, 3); /* v4 >> 3 */ \ - q0 = _mm_subs_epi8(q0, v4); /* q0 -= v4 */ \ - \ - /* Now do +3 side */ \ - SIGNED_SHIFT_N(v3, 3); /* v3 >> 3 */ \ - p0 = _mm_adds_epi8(p0, v3); /* p0 += v3 */ \ +// 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)] -#define UPDATE_2PIXELS(pi, qi, a_lo, a_hi) { \ - const __m128i a_lo7 = _mm_srai_epi16(a_lo, 7); \ - const __m128i a_hi7 = _mm_srai_epi16(a_hi, 7); \ - const __m128i delta = _mm_packs_epi16(a_lo7, a_hi7); \ - pi = _mm_adds_epi8(pi, delta); \ - qi = _mm_subs_epi8(qi, delta); \ +// 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); } -static void NeedsFilter(const __m128i* p1, const __m128i* p0, const __m128i* q0, - const __m128i* q1, int thresh, __m128i *mask) { - __m128i t1 = MM_ABS(*p1, *q1); // abs(p1 - q1) - *mask = _mm_set1_epi8(0xFE); - t1 = _mm_and_si128(t1, *mask); // set lsb of each byte to zero - t1 = _mm_srli_epi16(t1, 1); // abs(p1 - q1) / 2 +// 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 - *mask = MM_ABS(*p0, *q0); // abs(p0 - q0) - *mask = _mm_adds_epu8(*mask, *mask); // abs(p0 - q0) * 2 - *mask = _mm_adds_epu8(*mask, t1); // abs(p0 - q0) * 2 + 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 - t1 = _mm_set1_epi8(thresh); - *mask = _mm_subs_epu8(*mask, t1); // mask <= thresh - *mask = _mm_cmpeq_epi8(*mask, _mm_setzero_si128()); + 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(const __m128i* p1, __m128i* p0, __m128i* q0, - const __m128i* q1, int thresh) { +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); - // convert to signed values FLIP_SIGN_BIT2(*p0, *q0); - - GET_BASE_DELTA(p1s, *p0, *q0, q1s, a); + GetBaseDelta(&p1s, p0, q0, &q1s, &a); a = _mm_and_si128(a, mask); // mask filter values we don't care about - DO_SIMPLE_FILTER(*p0, *q0, a); - - // unoffset + 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* p1, __m128i *p0, - __m128i* q0, __m128i* q1, - const __m128i* mask, int hev_thresh) { +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; - const __m128i sign_bit = _mm_set1_epi8(0x80); // compute hev mask - GET_NOTHEV(*p1, *p0, *q0, *q1, hev_thresh, not_hev); + GetNotHEV(p1, p0, q0, q1, hev_thresh, ¬_hev); // convert to signed values FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1); @@ -433,92 +448,83 @@ static WEBP_INLINE void DoFilter4(__m128i* p1, __m128i *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 + 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); - t2 = _mm_set1_epi8(1); - t3 = _mm_adds_epi8(t3, t2); - SIGNED_SHIFT_N(t3, 1); // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 4 + // 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 - - // unoffset - FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1); + FLIP_SIGN_BIT2(*p1, *q1); } // 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; +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 - GET_NOTHEV(*p1, *p0, *q0, *q1, hev_thresh, not_hev); + GetNotHEV(p1, p0, q0, q1, hev_thresh, ¬_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); + 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); - DO_SIMPLE_FILTER(*p0, *q0, f); + DoSimpleFilter(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 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, 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 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, sixty_three); // Filter * 9 + 63 - const __m128i a2_hi = _mm_add_epi16(f9_hi, sixty_three); // Filter * 9 + 63 + 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(f18_lo, sixty_three); // F... * 18 + 63 - const __m128i a1_hi = _mm_add_epi16(f18_hi, sixty_three); // F... * 18 + 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(f18_lo, a2_lo); // Filter * 27 + 63 - const __m128i a0_hi = _mm_add_epi16(f18_hi, a2_hi); // Filter * 27 + 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 - UPDATE_2PIXELS(*p2, *q2, a2_lo, a2_hi); - UPDATE_2PIXELS(*p1, *q1, a1_lo, a1_hi); - UPDATE_2PIXELS(*p0, *q0, a0_lo, a0_hi); + Update2Pixels(p2, q2, &a2_lo, &a2_hi); + Update2Pixels(p1, q1, &a1_lo, &a1_hi); + Update2Pixels(p0, q0, &a0_lo, &a0_hi); } - - // unoffset - FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1); - FLIP_SIGN_BIT2(*p2, *q2); } // 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* b, int stride, - __m128i* p, __m128i* q) { +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 @@ -557,10 +563,11 @@ static WEBP_INLINE void Load8x4(const uint8_t* b, int stride, *q = _mm_unpackhi_epi32(t1, t2); } -static WEBP_INLINE void Load16x4(const uint8_t* r0, const uint8_t* r8, +static WEBP_INLINE void Load16x4(const uint8_t* const r0, + const uint8_t* const r8, int stride, - __m128i* p1, __m128i* p0, - __m128i* q0, __m128i* q1) { + __m128i* const p1, __m128i* const p0, + __m128i* const q0, __m128i* const q1) { __m128i t1, t2; // Assume the pixels around the edge (|) are numbered as follows // 00 01 | 02 03 @@ -592,7 +599,7 @@ static WEBP_INLINE void Load16x4(const uint8_t* r0, const uint8_t* r8, *q1 = _mm_unpackhi_epi64(t2, *q1); } -static WEBP_INLINE void Store4x4(__m128i* x, uint8_t* dst, int stride) { +static WEBP_INLINE void Store4x4(__m128i* const x, uint8_t* dst, int stride) { int i; for (i = 0; i < 4; ++i, dst += stride) { *((int32_t*)dst) = _mm_cvtsi128_si32(*x); @@ -601,48 +608,51 @@ static WEBP_INLINE void Store4x4(__m128i* x, uint8_t* dst, int stride) { } // Transpose back and store -static WEBP_INLINE void Store16x4(uint8_t* r0, uint8_t* r8, int stride, - __m128i* p1, __m128i* p0, - __m128i* q0, __m128i* q1) { - __m128i t1; +static WEBP_INLINE void Store16x4(const __m128i* const p1, + const __m128i* const p0, + const __m128i* const q0, + const __m128i* const q1, + uint8_t* r0, uint8_t* r8, + int stride) { + __m128i t1, p1_s, p0_s, q0_s, q1_s; // p0 = 71 70 61 60 51 50 41 40 31 30 21 20 11 10 01 00 // p1 = f1 f0 e1 e0 d1 d0 c1 c0 b1 b0 a1 a0 91 90 81 80 t1 = *p0; - *p0 = _mm_unpacklo_epi8(*p1, t1); - *p1 = _mm_unpackhi_epi8(*p1, t1); + p0_s = _mm_unpacklo_epi8(*p1, t1); + p1_s = _mm_unpackhi_epi8(*p1, t1); // q0 = 73 72 63 62 53 52 43 42 33 32 23 22 13 12 03 02 // q1 = f3 f2 e3 e2 d3 d2 c3 c2 b3 b2 a3 a2 93 92 83 82 t1 = *q0; - *q0 = _mm_unpacklo_epi8(t1, *q1); - *q1 = _mm_unpackhi_epi8(t1, *q1); + q0_s = _mm_unpacklo_epi8(t1, *q1); + q1_s = _mm_unpackhi_epi8(t1, *q1); // p0 = 33 32 31 30 23 22 21 20 13 12 11 10 03 02 01 00 // q0 = 73 72 71 70 63 62 61 60 53 52 51 50 43 42 41 40 - t1 = *p0; - *p0 = _mm_unpacklo_epi16(t1, *q0); - *q0 = _mm_unpackhi_epi16(t1, *q0); + t1 = p0_s; + p0_s = _mm_unpacklo_epi16(t1, q0_s); + q0_s = _mm_unpackhi_epi16(t1, q0_s); // p1 = b3 b2 b1 b0 a3 a2 a1 a0 93 92 91 90 83 82 81 80 // q1 = f3 f2 f1 f0 e3 e2 e1 e0 d3 d2 d1 d0 c3 c2 c1 c0 - t1 = *p1; - *p1 = _mm_unpacklo_epi16(t1, *q1); - *q1 = _mm_unpackhi_epi16(t1, *q1); + t1 = p1_s; + p1_s = _mm_unpacklo_epi16(t1, q1_s); + q1_s = _mm_unpackhi_epi16(t1, q1_s); - Store4x4(p0, r0, stride); + Store4x4(&p0_s, r0, stride); r0 += 4 * stride; - Store4x4(q0, r0, stride); + Store4x4(&q0_s, r0, stride); - Store4x4(p1, r8, stride); + Store4x4(&p1_s, r8, stride); r8 += 4 * stride; - Store4x4(q1, r8, stride); + Store4x4(&q1_s, r8, stride); } //------------------------------------------------------------------------------ // Simple In-loop filtering (Paragraph 15.2) -static void SimpleVFilter16SSE2(uint8_t* p, int stride, int thresh) { +static void SimpleVFilter16(uint8_t* p, int stride, int thresh) { // Load __m128i p1 = _mm_loadu_si128((__m128i*)&p[-2 * stride]); __m128i p0 = _mm_loadu_si128((__m128i*)&p[-stride]); @@ -653,49 +663,49 @@ static void SimpleVFilter16SSE2(uint8_t* p, int stride, int thresh) { // Store _mm_storeu_si128((__m128i*)&p[-stride], p0); - _mm_storeu_si128((__m128i*)p, q0); + _mm_storeu_si128((__m128i*)&p[0], q0); } -static void SimpleHFilter16SSE2(uint8_t* p, int stride, int thresh) { +static void SimpleHFilter16(uint8_t* p, int stride, int thresh) { __m128i p1, p0, q0, q1; p -= 2; // beginning of p1 - Load16x4(p, p + 8 * stride, stride, &p1, &p0, &q0, &q1); + Load16x4(p, p + 8 * stride, stride, &p1, &p0, &q0, &q1); DoFilter2(&p1, &p0, &q0, &q1, thresh); - Store16x4(p, p + 8 * stride, stride, &p1, &p0, &q0, &q1); + Store16x4(&p1, &p0, &q0, &q1, p, p + 8 * stride, stride); } -static void SimpleVFilter16iSSE2(uint8_t* p, int stride, int thresh) { +static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) { int k; for (k = 3; k > 0; --k) { p += 4 * stride; - SimpleVFilter16SSE2(p, stride, thresh); + SimpleVFilter16(p, stride, thresh); } } -static void SimpleHFilter16iSSE2(uint8_t* p, int stride, int thresh) { +static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) { int k; for (k = 3; k > 0; --k) { p += 4; - SimpleHFilter16SSE2(p, stride, thresh); + SimpleHFilter16(p, stride, thresh); } } //------------------------------------------------------------------------------ // Complex In-loop filtering (Paragraph 15.3) -#define MAX_DIFF1(p3, p2, p1, p0, m) { \ - m = MM_ABS(p3, p2); \ +#define MAX_DIFF1(p3, p2, p1, p0, m) do { \ + m = MM_ABS(p1, p0); \ + m = _mm_max_epu8(m, MM_ABS(p3, p2)); \ m = _mm_max_epu8(m, MM_ABS(p2, p1)); \ - m = _mm_max_epu8(m, MM_ABS(p1, p0)); \ -} +} while (0) -#define MAX_DIFF2(p3, p2, p1, p0, m) { \ +#define MAX_DIFF2(p3, p2, p1, p0, m) do { \ + m = _mm_max_epu8(m, MM_ABS(p1, p0)); \ m = _mm_max_epu8(m, MM_ABS(p3, p2)); \ m = _mm_max_epu8(m, MM_ABS(p2, p1)); \ - m = _mm_max_epu8(m, MM_ABS(p1, p0)); \ -} +} while (0) #define LOAD_H_EDGES4(p, stride, e1, e2, e3, e4) { \ e1 = _mm_loadu_si128((__m128i*)&(p)[0 * stride]); \ @@ -704,10 +714,11 @@ static void SimpleHFilter16iSSE2(uint8_t* p, int stride, int thresh) { e4 = _mm_loadu_si128((__m128i*)&(p)[3 * stride]); \ } -#define LOADUV_H_EDGE(p, u, v, stride) { \ - p = _mm_loadl_epi64((__m128i*)&(u)[(stride)]); \ - p = _mm_unpacklo_epi64(p, _mm_loadl_epi64((__m128i*)&(v)[(stride)])); \ -} +#define LOADUV_H_EDGE(p, u, v, stride) do { \ + const __m128i U = _mm_loadl_epi64((__m128i*)&(u)[(stride)]); \ + const __m128i V = _mm_loadl_epi64((__m128i*)&(v)[(stride)]); \ + p = _mm_unpacklo_epi64(U, V); \ +} while (0) #define LOADUV_H_EDGES4(u, v, stride, e1, e2, e3, e4) { \ LOADUV_H_EDGE(e1, u, v, 0 * stride); \ @@ -722,18 +733,23 @@ static void SimpleHFilter16iSSE2(uint8_t* p, int stride, int thresh) { _mm_storel_epi64((__m128i*)&v[(stride)], p); \ } -#define COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask) { \ - __m128i fl_yes; \ - const __m128i it = _mm_set1_epi8(ithresh); \ - mask = _mm_subs_epu8(mask, it); \ - mask = _mm_cmpeq_epi8(mask, _mm_setzero_si128()); \ - NeedsFilter(&p1, &p0, &q0, &q1, thresh, &fl_yes); \ - mask = _mm_and_si128(mask, fl_yes); \ +static WEBP_INLINE void ComplexMask(const __m128i* const p1, + const __m128i* const p0, + const __m128i* const q0, + const __m128i* const q1, + int thresh, int ithresh, + __m128i* const mask) { + const __m128i it = _mm_set1_epi8(ithresh); + const __m128i diff = _mm_subs_epu8(*mask, it); + const __m128i thresh_mask = _mm_cmpeq_epi8(diff, _mm_setzero_si128()); + __m128i filter_mask; + NeedsFilter(p1, p0, q0, q1, thresh, &filter_mask); + *mask = _mm_and_si128(thresh_mask, filter_mask); } // on macroblock edges -static void VFilter16SSE2(uint8_t* p, int stride, - int thresh, int ithresh, int hev_thresh) { +static void VFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { __m128i t1; __m128i mask; __m128i p2, p1, p0, q0, q1, q2; @@ -746,20 +762,20 @@ static void VFilter16SSE2(uint8_t* p, int stride, LOAD_H_EDGES4(p, stride, q0, q1, q2, t1); MAX_DIFF2(t1, q2, q1, q0, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); + ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask); DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh); // Store _mm_storeu_si128((__m128i*)&p[-3 * stride], p2); _mm_storeu_si128((__m128i*)&p[-2 * stride], p1); _mm_storeu_si128((__m128i*)&p[-1 * stride], p0); - _mm_storeu_si128((__m128i*)&p[0 * stride], q0); - _mm_storeu_si128((__m128i*)&p[1 * stride], q1); - _mm_storeu_si128((__m128i*)&p[2 * stride], q2); + _mm_storeu_si128((__m128i*)&p[+0 * stride], q0); + _mm_storeu_si128((__m128i*)&p[+1 * stride], q1); + _mm_storeu_si128((__m128i*)&p[+2 * stride], q2); } -static void HFilter16SSE2(uint8_t* p, int stride, - int thresh, int ithresh, int hev_thresh) { +static void HFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { __m128i mask; __m128i p3, p2, p1, p0, q0, q1, q2, q3; @@ -770,71 +786,78 @@ static void HFilter16SSE2(uint8_t* p, int stride, Load16x4(p, p + 8 * stride, stride, &q0, &q1, &q2, &q3); // q0, q1, q2, q3 MAX_DIFF2(q3, q2, q1, q0, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); + ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask); DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh); - Store16x4(b, b + 8 * stride, stride, &p3, &p2, &p1, &p0); - Store16x4(p, p + 8 * stride, stride, &q0, &q1, &q2, &q3); + Store16x4(&p3, &p2, &p1, &p0, b, b + 8 * stride, stride); + Store16x4(&q0, &q1, &q2, &q3, p, p + 8 * stride, stride); } // on three inner edges -static void VFilter16iSSE2(uint8_t* p, int stride, - int thresh, int ithresh, int hev_thresh) { +static void VFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { int k; - __m128i mask; - __m128i t1, t2, p1, p0, q0, q1; + __m128i p3, p2, p1, p0; // loop invariants - for (k = 3; k > 0; --k) { - // Load p3, p2, p1, p0 - LOAD_H_EDGES4(p, stride, t2, t1, p1, p0); - MAX_DIFF1(t2, t1, p1, p0, mask); + LOAD_H_EDGES4(p, stride, p3, p2, p1, p0); // prologue + for (k = 3; k > 0; --k) { + __m128i mask, tmp1, tmp2; + uint8_t* const b = p + 2 * stride; // beginning of p1 p += 4 * stride; - // Load q0, q1, q2, q3 - LOAD_H_EDGES4(p, stride, q0, q1, t1, t2); - MAX_DIFF2(t2, t1, q1, q0, mask); + MAX_DIFF1(p3, p2, p1, p0, mask); // compute partial mask + LOAD_H_EDGES4(p, stride, p3, p2, tmp1, tmp2); + MAX_DIFF2(p3, p2, tmp1, tmp2, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); - DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh); + // p3 and p2 are not just temporary variables here: they will be + // re-used for next span. And q2/q3 will become p1/p0 accordingly. + ComplexMask(&p1, &p0, &p3, &p2, thresh, ithresh, &mask); + DoFilter4(&p1, &p0, &p3, &p2, &mask, hev_thresh); // Store - _mm_storeu_si128((__m128i*)&p[-2 * stride], p1); - _mm_storeu_si128((__m128i*)&p[-1 * stride], p0); - _mm_storeu_si128((__m128i*)&p[0 * stride], q0); - _mm_storeu_si128((__m128i*)&p[1 * stride], q1); + _mm_storeu_si128((__m128i*)&b[0 * stride], p1); + _mm_storeu_si128((__m128i*)&b[1 * stride], p0); + _mm_storeu_si128((__m128i*)&b[2 * stride], p3); + _mm_storeu_si128((__m128i*)&b[3 * stride], p2); + + // rotate samples + p1 = tmp1; + p0 = tmp2; } } -static void HFilter16iSSE2(uint8_t* p, int stride, - int thresh, int ithresh, int hev_thresh) { +static void HFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { int k; - uint8_t* b; - __m128i mask; - __m128i t1, t2, p1, p0, q0, q1; + __m128i p3, p2, p1, p0; // loop invariants + + Load16x4(p, p + 8 * stride, stride, &p3, &p2, &p1, &p0); // prologue for (k = 3; k > 0; --k) { - b = p; - Load16x4(b, b + 8 * stride, stride, &t2, &t1, &p1, &p0); // p3, p2, p1, p0 - MAX_DIFF1(t2, t1, p1, p0, mask); + __m128i mask, tmp1, tmp2; + uint8_t* const b = p + 2; // beginning of p1 - b += 4; // beginning of q0 - Load16x4(b, b + 8 * stride, stride, &q0, &q1, &t1, &t2); // q0, q1, q2, q3 - MAX_DIFF2(t2, t1, q1, q0, mask); + p += 4; // beginning of q0 (and next span) - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); - DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh); + MAX_DIFF1(p3, p2, p1, p0, mask); // compute partial mask + Load16x4(p, p + 8 * stride, stride, &p3, &p2, &tmp1, &tmp2); + MAX_DIFF2(p3, p2, tmp1, tmp2, mask); - b -= 2; // beginning of p1 - Store16x4(b, b + 8 * stride, stride, &p1, &p0, &q0, &q1); + ComplexMask(&p1, &p0, &p3, &p2, thresh, ithresh, &mask); + DoFilter4(&p1, &p0, &p3, &p2, &mask, hev_thresh); - p += 4; + Store16x4(&p1, &p0, &p3, &p2, b, b + 8 * stride, stride); + + // rotate samples + p1 = tmp1; + p0 = tmp2; } } // 8-pixels wide variant, for chroma filtering -static void VFilter8SSE2(uint8_t* u, uint8_t* v, int stride, - int thresh, int ithresh, int hev_thresh) { +static void VFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { __m128i mask; __m128i t1, p2, p1, p0, q0, q1, q2; @@ -846,7 +869,7 @@ static void VFilter8SSE2(uint8_t* u, uint8_t* v, int stride, LOADUV_H_EDGES4(u, v, stride, q0, q1, q2, t1); MAX_DIFF2(t1, q2, q1, q0, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); + ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask); DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh); // Store @@ -858,8 +881,8 @@ static void VFilter8SSE2(uint8_t* u, uint8_t* v, int stride, STOREUV(q2, u, v, 2 * stride); } -static void HFilter8SSE2(uint8_t* u, uint8_t* v, int stride, - int thresh, int ithresh, int hev_thresh) { +static void HFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { __m128i mask; __m128i p3, p2, p1, p0, q0, q1, q2, q3; @@ -871,15 +894,15 @@ static void HFilter8SSE2(uint8_t* u, uint8_t* v, int stride, Load16x4(u, v, stride, &q0, &q1, &q2, &q3); // q0, q1, q2, q3 MAX_DIFF2(q3, q2, q1, q0, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); + ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask); DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh); - Store16x4(tu, tv, stride, &p3, &p2, &p1, &p0); - Store16x4(u, v, stride, &q0, &q1, &q2, &q3); + Store16x4(&p3, &p2, &p1, &p0, tu, tv, stride); + Store16x4(&q0, &q1, &q2, &q3, u, v, stride); } -static void VFilter8iSSE2(uint8_t* u, uint8_t* v, int stride, - int thresh, int ithresh, int hev_thresh) { +static void VFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { __m128i mask; __m128i t1, t2, p1, p0, q0, q1; @@ -894,7 +917,7 @@ static void VFilter8iSSE2(uint8_t* u, uint8_t* v, int stride, LOADUV_H_EDGES4(u, v, stride, q0, q1, t1, t2); MAX_DIFF2(t2, t1, q1, q0, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); + ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask); DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh); // Store @@ -904,8 +927,8 @@ static void VFilter8iSSE2(uint8_t* u, uint8_t* v, int stride, STOREUV(q1, u, v, 1 * stride); } -static void HFilter8iSSE2(uint8_t* u, uint8_t* v, int stride, - int thresh, int ithresh, int hev_thresh) { +static void HFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { __m128i mask; __m128i t1, t2, p1, p0, q0, q1; Load16x4(u, v, stride, &t2, &t1, &p1, &p0); // p3, p2, p1, p0 @@ -916,12 +939,12 @@ static void HFilter8iSSE2(uint8_t* u, uint8_t* v, int stride, Load16x4(u, v, stride, &q0, &q1, &t1, &t2); // q0, q1, q2, q3 MAX_DIFF2(t2, t1, q1, q0, mask); - COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask); + ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask); DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh); u -= 2; // beginning of p1 v -= 2; - Store16x4(u, v, stride, &p1, &p0, &q0, &q1); + Store16x4(&p1, &p0, &q0, &q1, u, v, stride); } #endif // WEBP_USE_SSE2 @@ -933,24 +956,23 @@ extern void VP8DspInitSSE2(void); void VP8DspInitSSE2(void) { #if defined(WEBP_USE_SSE2) - VP8Transform = TransformSSE2; + VP8Transform = Transform; #if defined(USE_TRANSFORM_AC3) - VP8TransformAC3 = TransformAC3SSE2; + VP8TransformAC3 = TransformAC3; #endif - VP8VFilter16 = VFilter16SSE2; - VP8HFilter16 = HFilter16SSE2; - VP8VFilter8 = VFilter8SSE2; - VP8HFilter8 = HFilter8SSE2; - VP8VFilter16i = VFilter16iSSE2; - VP8HFilter16i = HFilter16iSSE2; - VP8VFilter8i = VFilter8iSSE2; - VP8HFilter8i = HFilter8iSSE2; - - VP8SimpleVFilter16 = SimpleVFilter16SSE2; - VP8SimpleHFilter16 = SimpleHFilter16SSE2; - VP8SimpleVFilter16i = SimpleVFilter16iSSE2; - VP8SimpleHFilter16i = SimpleHFilter16iSSE2; + VP8VFilter16 = VFilter16; + VP8HFilter16 = HFilter16; + VP8VFilter8 = VFilter8; + VP8HFilter8 = HFilter8; + VP8VFilter16i = VFilter16i; + VP8HFilter16i = HFilter16i; + VP8VFilter8i = VFilter8i; + VP8HFilter8i = HFilter8i; + + VP8SimpleVFilter16 = SimpleVFilter16; + VP8SimpleHFilter16 = SimpleHFilter16; + VP8SimpleVFilter16i = SimpleVFilter16i; + VP8SimpleHFilter16i = SimpleHFilter16i; #endif // WEBP_USE_SSE2 } - |