/src/libwebp/src/dsp/cost_sse2.c
Line | Count | Source |
1 | | // Copyright 2015 Google Inc. All Rights Reserved. |
2 | | // |
3 | | // Use of this source code is governed by a BSD-style license |
4 | | // that can be found in the COPYING file in the root of the source |
5 | | // tree. An additional intellectual property rights grant can be found |
6 | | // in the file PATENTS. All contributing project authors may |
7 | | // be found in the AUTHORS file in the root of the source tree. |
8 | | // ----------------------------------------------------------------------------- |
9 | | // |
10 | | // SSE2 version of cost functions |
11 | | // |
12 | | // Author: Skal (pascal.massimino@gmail.com) |
13 | | |
14 | | #include "src/dsp/dsp.h" |
15 | | |
16 | | #if defined(WEBP_USE_SSE2) |
17 | | #include <assert.h> |
18 | | #include <emmintrin.h> |
19 | | |
20 | | #include "src/dsp/cpu.h" |
21 | | #include "src/enc/cost_enc.h" |
22 | | #include "src/enc/vp8i_enc.h" |
23 | | #include "src/utils/utils.h" |
24 | | #include "src/webp/types.h" |
25 | | |
26 | | //------------------------------------------------------------------------------ |
27 | | |
28 | | static void SetResidualCoeffs_SSE2(const int16_t* WEBP_RESTRICT const coeffs, |
29 | 346M | VP8Residual* WEBP_RESTRICT const res) { |
30 | 346M | const __m128i c0 = _mm_loadu_si128((const __m128i*)(coeffs + 0)); |
31 | 346M | const __m128i c1 = _mm_loadu_si128((const __m128i*)(coeffs + 8)); |
32 | | // Use SSE2 to compare 16 values with a single instruction. |
33 | 346M | const __m128i zero = _mm_setzero_si128(); |
34 | 346M | const __m128i m0 = _mm_packs_epi16(c0, c1); |
35 | 346M | const __m128i m1 = _mm_cmpeq_epi8(m0, zero); |
36 | | // Get the comparison results as a bitmask into 16bits. Negate the mask to get |
37 | | // the position of entries that are not equal to zero. We don't need to mask |
38 | | // out least significant bits according to res->first, since coeffs[0] is 0 |
39 | | // if res->first > 0. |
40 | 346M | const uint32_t mask = 0x0000ffffu ^ (uint32_t)_mm_movemask_epi8(m1); |
41 | | // The position of the most significant non-zero bit indicates the position of |
42 | | // the last non-zero value. |
43 | 346M | assert(res->first == 0 || coeffs[0] == 0); |
44 | 346M | res->last = mask ? BitsLog2Floor(mask) : -1; |
45 | 346M | res->coeffs = coeffs; |
46 | 346M | } |
47 | | |
48 | 283M | static int GetResidualCost_SSE2(int ctx0, const VP8Residual* const res) { |
49 | 283M | uint8_t levels[16], ctxs[16]; |
50 | 283M | uint16_t abs_levels[16]; |
51 | 283M | int n = res->first; |
52 | | // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1 |
53 | 283M | const int p0 = res->prob[n][ctx0][0]; |
54 | 283M | CostArrayPtr const costs = res->costs; |
55 | 283M | const uint16_t* t = costs[n][ctx0]; |
56 | | // bit_cost(1, p0) is already incorporated in t[] tables, but only if ctx != 0 |
57 | | // (as required by the syntax). For ctx0 == 0, we need to add it here or it'll |
58 | | // be missing during the loop. |
59 | 283M | int cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0; |
60 | | |
61 | 283M | if (res->last < 0) { |
62 | 96.1M | return VP8BitCost(0, p0); |
63 | 96.1M | } |
64 | | |
65 | 187M | { // precompute clamped levels and contexts, packed to 8b. |
66 | 187M | const __m128i zero = _mm_setzero_si128(); |
67 | 187M | const __m128i kCst2 = _mm_set1_epi8(2); |
68 | 187M | const __m128i kCst67 = _mm_set1_epi8(MAX_VARIABLE_LEVEL); |
69 | 187M | const __m128i c0 = _mm_loadu_si128((const __m128i*)&res->coeffs[0]); |
70 | 187M | const __m128i c1 = _mm_loadu_si128((const __m128i*)&res->coeffs[8]); |
71 | 187M | const __m128i D0 = _mm_sub_epi16(zero, c0); |
72 | 187M | const __m128i D1 = _mm_sub_epi16(zero, c1); |
73 | 187M | const __m128i E0 = _mm_max_epi16(c0, D0); // abs(v), 16b |
74 | 187M | const __m128i E1 = _mm_max_epi16(c1, D1); |
75 | 187M | const __m128i F = _mm_packs_epi16(E0, E1); |
76 | 187M | const __m128i G = _mm_min_epu8(F, kCst2); // context = 0,1,2 |
77 | 187M | const __m128i H = _mm_min_epu8(F, kCst67); // clamp_level in [0..67] |
78 | | |
79 | 187M | _mm_storeu_si128((__m128i*)&ctxs[0], G); |
80 | 187M | _mm_storeu_si128((__m128i*)&levels[0], H); |
81 | | |
82 | 187M | _mm_storeu_si128((__m128i*)&abs_levels[0], E0); |
83 | 187M | _mm_storeu_si128((__m128i*)&abs_levels[8], E1); |
84 | 187M | } |
85 | 1.67G | for (; n < res->last; ++n) { |
86 | 1.49G | const int ctx = ctxs[n]; |
87 | 1.49G | const int level = levels[n]; |
88 | 1.49G | const int flevel = abs_levels[n]; // full level |
89 | 1.49G | cost += VP8LevelFixedCosts[flevel] + t[level]; // simplified VP8LevelCost() |
90 | 1.49G | t = costs[n + 1][ctx]; |
91 | 1.49G | } |
92 | | // Last coefficient is always non-zero |
93 | 187M | { |
94 | 187M | const int level = levels[n]; |
95 | 187M | const int flevel = abs_levels[n]; |
96 | 187M | assert(flevel != 0); |
97 | 187M | cost += VP8LevelFixedCosts[flevel] + t[level]; |
98 | 187M | if (n < 15) { |
99 | 139M | const int b = VP8EncBands[n + 1]; |
100 | 139M | const int ctx = ctxs[n]; |
101 | 139M | const int last_p0 = res->prob[b][ctx][0]; |
102 | 139M | cost += VP8BitCost(0, last_p0); |
103 | 139M | } |
104 | 187M | } |
105 | 187M | return cost; |
106 | 283M | } |
107 | | |
108 | | //------------------------------------------------------------------------------ |
109 | | // Entry point |
110 | | |
111 | | extern void VP8EncDspCostInitSSE2(void); |
112 | | |
113 | 5.41k | WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspCostInitSSE2(void) { |
114 | 5.41k | VP8SetResidualCoeffs = SetResidualCoeffs_SSE2; |
115 | 5.41k | VP8GetResidualCost = GetResidualCost_SSE2; |
116 | 5.41k | } |
117 | | |
118 | | #else // !WEBP_USE_SSE2 |
119 | | |
120 | | WEBP_DSP_INIT_STUB(VP8EncDspCostInitSSE2) |
121 | | |
122 | | #endif // WEBP_USE_SSE2 |