1/* NOLINT(build/header_guard) */
2/* Copyright 2013 Google Inc. All Rights Reserved.
3
4 Distributed under MIT license.
5 See file LICENSE for detail or copy at https://opensource.org/licenses/MIT
6*/
7
8/* template parameters: FN */
9
10#define HistogramType FN(Histogram)
11
12double FN(BrotliPopulationCost)(const HistogramType* histogram) {
13 static const double kOneSymbolHistogramCost = 12;
14 static const double kTwoSymbolHistogramCost = 20;
15 static const double kThreeSymbolHistogramCost = 28;
16 static const double kFourSymbolHistogramCost = 37;
17 const size_t data_size = FN(HistogramDataSize)();
18 int count = 0;
19 size_t s[5];
20 double bits = 0.0;
21 size_t i;
22 if (histogram->total_count_ == 0) {
23 return kOneSymbolHistogramCost;
24 }
25 for (i = 0; i < data_size; ++i) {
26 if (histogram->data_[i] > 0) {
27 s[count] = i;
28 ++count;
29 if (count > 4) break;
30 }
31 }
32 if (count == 1) {
33 return kOneSymbolHistogramCost;
34 }
35 if (count == 2) {
36 return (kTwoSymbolHistogramCost + (double)histogram->total_count_);
37 }
38 if (count == 3) {
39 const uint32_t histo0 = histogram->data_[s[0]];
40 const uint32_t histo1 = histogram->data_[s[1]];
41 const uint32_t histo2 = histogram->data_[s[2]];
42 const uint32_t histomax =
43 BROTLI_MAX(uint32_t, histo0, BROTLI_MAX(uint32_t, histo1, histo2));
44 return (kThreeSymbolHistogramCost +
45 2 * (histo0 + histo1 + histo2) - histomax);
46 }
47 if (count == 4) {
48 uint32_t histo[4];
49 uint32_t h23;
50 uint32_t histomax;
51 for (i = 0; i < 4; ++i) {
52 histo[i] = histogram->data_[s[i]];
53 }
54 /* Sort */
55 for (i = 0; i < 4; ++i) {
56 size_t j;
57 for (j = i + 1; j < 4; ++j) {
58 if (histo[j] > histo[i]) {
59 BROTLI_SWAP(uint32_t, histo, j, i);
60 }
61 }
62 }
63 h23 = histo[2] + histo[3];
64 histomax = BROTLI_MAX(uint32_t, h23, histo[0]);
65 return (kFourSymbolHistogramCost +
66 3 * h23 + 2 * (histo[0] + histo[1]) - histomax);
67 }
68
69 {
70 /* In this loop we compute the entropy of the histogram and simultaneously
71 build a simplified histogram of the code length codes where we use the
72 zero repeat code 17, but we don't use the non-zero repeat code 16. */
73 size_t max_depth = 1;
74 uint32_t depth_histo[BROTLI_CODE_LENGTH_CODES] = { 0 };
75 const double log2total = FastLog2(histogram->total_count_);
76 for (i = 0; i < data_size;) {
77 if (histogram->data_[i] > 0) {
78 /* Compute -log2(P(symbol)) = -log2(count(symbol)/total_count) =
79 = log2(total_count) - log2(count(symbol)) */
80 double log2p = log2total - FastLog2(histogram->data_[i]);
81 /* Approximate the bit depth by round(-log2(P(symbol))) */
82 size_t depth = (size_t)(log2p + 0.5);
83 bits += histogram->data_[i] * log2p;
84 if (depth > 15) {
85 depth = 15;
86 }
87 if (depth > max_depth) {
88 max_depth = depth;
89 }
90 ++depth_histo[depth];
91 ++i;
92 } else {
93 /* Compute the run length of zeros and add the appropriate number of 0
94 and 17 code length codes to the code length code histogram. */
95 uint32_t reps = 1;
96 size_t k;
97 for (k = i + 1; k < data_size && histogram->data_[k] == 0; ++k) {
98 ++reps;
99 }
100 i += reps;
101 if (i == data_size) {
102 /* Don't add any cost for the last zero run, since these are encoded
103 only implicitly. */
104 break;
105 }
106 if (reps < 3) {
107 depth_histo[0] += reps;
108 } else {
109 reps -= 2;
110 while (reps > 0) {
111 ++depth_histo[BROTLI_REPEAT_ZERO_CODE_LENGTH];
112 /* Add the 3 extra bits for the 17 code length code. */
113 bits += 3;
114 reps >>= 3;
115 }
116 }
117 }
118 }
119 /* Add the estimated encoding cost of the code length code histogram. */
120 bits += (double)(18 + 2 * max_depth);
121 /* Add the entropy of the code length code histogram. */
122 bits += BitsEntropy(depth_histo, BROTLI_CODE_LENGTH_CODES);
123 }
124 return bits;
125}
126
127#undef HistogramType
128