1 /*
2 * Copyright (C) 2013 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #ifndef ART_LIBARTBASE_BASE_HISTOGRAM_INL_H_
18 #define ART_LIBARTBASE_BASE_HISTOGRAM_INL_H_
19
20 #include <algorithm>
21 #include <cmath>
22 #include <limits>
23 #include <ostream>
24
25 #include "histogram.h"
26
27 #include <android-base/logging.h>
28
29 #include "bit_utils.h"
30 #include "time_utils.h"
31 #include "utils.h"
32
33 namespace art {
34
AddValue(Value value)35 template <class Value> inline void Histogram<Value>::AddValue(Value value) {
36 CHECK_GE(value, static_cast<Value>(0));
37 if (value >= max_) {
38 Value new_max = ((value + 1) / bucket_width_ + 1) * bucket_width_;
39 DCHECK_GT(new_max, max_);
40 GrowBuckets(new_max);
41 }
42 BucketiseValue(value);
43 }
44
AdjustAndAddValue(Value value)45 template <class Value> inline void Histogram<Value>::AdjustAndAddValue(Value value) {
46 AddValue(value / kAdjust);
47 }
48
Histogram(const char * name)49 template <class Value> inline Histogram<Value>::Histogram(const char* name)
50 : kAdjust(0),
51 kInitialBucketCount(0),
52 name_(name),
53 max_buckets_(0),
54 sample_size_(0) {
55 }
56
57 template <class Value>
Histogram(const char * name,Value initial_bucket_width,size_t max_buckets)58 inline Histogram<Value>::Histogram(const char* name, Value initial_bucket_width,
59 size_t max_buckets)
60 : kAdjust(1000),
61 kInitialBucketCount(8),
62 name_(name),
63 max_buckets_(max_buckets),
64 bucket_width_(initial_bucket_width) {
65 Reset();
66 }
67
68 template <class Value>
GrowBuckets(Value new_max)69 inline void Histogram<Value>::GrowBuckets(Value new_max) {
70 while (max_ < new_max) {
71 // If we have reached the maximum number of buckets, merge buckets together.
72 if (frequency_.size() >= max_buckets_) {
73 CHECK_ALIGNED(frequency_.size(), 2);
74 // We double the width of each bucket to reduce the number of buckets by a factor of 2.
75 bucket_width_ *= 2;
76 const size_t limit = frequency_.size() / 2;
77 // Merge the frequencies by adding each adjacent two together.
78 for (size_t i = 0; i < limit; ++i) {
79 frequency_[i] = frequency_[i * 2] + frequency_[i * 2 + 1];
80 }
81 // Remove frequencies in the second half of the array which were added to the first half.
82 while (frequency_.size() > limit) {
83 frequency_.pop_back();
84 }
85 }
86 max_ += bucket_width_;
87 frequency_.push_back(0);
88 }
89 }
90
FindBucket(Value val)91 template <class Value> inline size_t Histogram<Value>::FindBucket(Value val) const {
92 // Since this is only a linear histogram, bucket index can be found simply with
93 // dividing the value by the bucket width.
94 DCHECK_GE(val, min_);
95 DCHECK_LE(val, max_);
96 const size_t bucket_idx = static_cast<size_t>((val - min_) / bucket_width_);
97 DCHECK_GE(bucket_idx, 0ul);
98 DCHECK_LE(bucket_idx, GetBucketCount());
99 return bucket_idx;
100 }
101
102 template <class Value>
BucketiseValue(Value val)103 inline void Histogram<Value>::BucketiseValue(Value val) {
104 CHECK_LT(val, max_);
105 sum_ += val;
106 sum_of_squares_ += val * val;
107 ++sample_size_;
108 ++frequency_[FindBucket(val)];
109 max_value_added_ = std::max(val, max_value_added_);
110 min_value_added_ = std::min(val, min_value_added_);
111 }
112
Initialize()113 template <class Value> inline void Histogram<Value>::Initialize() {
114 for (size_t idx = 0; idx < kInitialBucketCount; idx++) {
115 frequency_.push_back(0);
116 }
117 // Cumulative frequency and ranges has a length of 1 over frequency.
118 max_ = bucket_width_ * GetBucketCount();
119 }
120
GetBucketCount()121 template <class Value> inline size_t Histogram<Value>::GetBucketCount() const {
122 return frequency_.size();
123 }
124
Reset()125 template <class Value> inline void Histogram<Value>::Reset() {
126 sum_of_squares_ = 0;
127 sample_size_ = 0;
128 min_ = 0;
129 sum_ = 0;
130 min_value_added_ = std::numeric_limits<Value>::max();
131 max_value_added_ = std::numeric_limits<Value>::min();
132 frequency_.clear();
133 Initialize();
134 }
135
GetRange(size_t bucket_idx)136 template <class Value> inline Value Histogram<Value>::GetRange(size_t bucket_idx) const {
137 DCHECK_LE(bucket_idx, GetBucketCount());
138 return min_ + bucket_idx * bucket_width_;
139 }
140
Mean()141 template <class Value> inline double Histogram<Value>::Mean() const {
142 DCHECK_GT(sample_size_, 0ull);
143 return static_cast<double>(sum_) / static_cast<double>(sample_size_);
144 }
145
Variance()146 template <class Value> inline double Histogram<Value>::Variance() const {
147 DCHECK_GT(sample_size_, 0ull);
148 // Using algorithms for calculating variance over a population:
149 // http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
150 Value sum_squared = sum_ * sum_;
151 double sum_squared_by_n_squared =
152 static_cast<double>(sum_squared) /
153 static_cast<double>(sample_size_ * sample_size_);
154 double sum_of_squares_by_n =
155 static_cast<double>(sum_of_squares_) / static_cast<double>(sample_size_);
156 return sum_of_squares_by_n - sum_squared_by_n_squared;
157 }
158
159 template <class Value>
PrintBins(std::ostream & os,const CumulativeData & data)160 inline void Histogram<Value>::PrintBins(std::ostream& os, const CumulativeData& data) const {
161 DCHECK_GT(sample_size_, 0ull);
162 for (size_t bin_idx = 0; bin_idx < data.freq_.size(); ++bin_idx) {
163 if (bin_idx > 0 && data.perc_[bin_idx] == data.perc_[bin_idx - 1]) {
164 bin_idx++;
165 continue;
166 }
167 os << GetRange(bin_idx) << ": " << data.freq_[bin_idx] << "\t"
168 << data.perc_[bin_idx] * 100.0 << "%\n";
169 }
170 }
171
172 template <class Value>
DumpBins(std::ostream & os)173 inline void Histogram<Value>::DumpBins(std::ostream& os) const {
174 DCHECK_GT(sample_size_, 0ull);
175 bool dumped_one = false;
176 for (size_t bin_idx = 0; bin_idx < frequency_.size(); ++bin_idx) {
177 if (frequency_[bin_idx] != 0U) {
178 if (dumped_one) {
179 // Prepend a comma if not the first bin.
180 os << ",";
181 } else {
182 dumped_one = true;
183 }
184 os << GetRange(bin_idx) << ":" << frequency_[bin_idx];
185 }
186 }
187 }
188
189 template <class Value>
PrintConfidenceIntervals(std::ostream & os,double interval,const CumulativeData & data)190 inline void Histogram<Value>::PrintConfidenceIntervals(std::ostream &os, double interval,
191 const CumulativeData& data) const {
192 static constexpr size_t kFractionalDigits = 3;
193 DCHECK_GT(interval, 0);
194 DCHECK_LT(interval, 1.0);
195 const double per_0 = (1.0 - interval) / 2.0;
196 const double per_1 = per_0 + interval;
197 const TimeUnit unit = GetAppropriateTimeUnit(Mean() * kAdjust);
198 os << Name() << ":\tSum: " << PrettyDuration(Sum() * kAdjust) << " "
199 << (interval * 100) << "% C.I. " << FormatDuration(Percentile(per_0, data) * kAdjust, unit,
200 kFractionalDigits)
201 << "-" << FormatDuration(Percentile(per_1, data) * kAdjust, unit, kFractionalDigits) << " "
202 << "Avg: " << FormatDuration(Mean() * kAdjust, unit, kFractionalDigits) << " Max: "
203 << FormatDuration(Max() * kAdjust, unit, kFractionalDigits) << std::endl;
204 }
205
206 template <class Value>
PrintMemoryUse(std::ostream & os)207 inline void Histogram<Value>::PrintMemoryUse(std::ostream &os) const {
208 os << Name();
209 if (sample_size_ != 0u) {
210 os << ": Avg: " << PrettySize(Mean()) << " Max: "
211 << PrettySize(Max()) << " Min: " << PrettySize(Min()) << "\n";
212 } else {
213 os << ": <no data>\n";
214 }
215 }
216
217 template <class Value>
CreateHistogram(CumulativeData * out_data)218 inline void Histogram<Value>::CreateHistogram(CumulativeData* out_data) const {
219 DCHECK_GT(sample_size_, 0ull);
220 out_data->freq_.clear();
221 out_data->perc_.clear();
222 uint64_t accumulated = 0;
223 out_data->freq_.push_back(accumulated);
224 out_data->perc_.push_back(0.0);
225 for (size_t idx = 0; idx < frequency_.size(); idx++) {
226 accumulated += frequency_[idx];
227 out_data->freq_.push_back(accumulated);
228 out_data->perc_.push_back(static_cast<double>(accumulated) / static_cast<double>(sample_size_));
229 }
230 DCHECK_EQ(out_data->freq_.back(), sample_size_);
231 DCHECK_LE(std::abs(out_data->perc_.back() - 1.0), 0.001);
232 }
233
234 #pragma clang diagnostic push
235 #pragma clang diagnostic ignored "-Wfloat-equal"
236
237 template <class Value>
Percentile(double per,const CumulativeData & data)238 inline double Histogram<Value>::Percentile(double per, const CumulativeData& data) const {
239 DCHECK_GT(data.perc_.size(), 0ull);
240 size_t upper_idx = 0, lower_idx = 0;
241 for (size_t idx = 0; idx < data.perc_.size(); idx++) {
242 if (per <= data.perc_[idx]) {
243 upper_idx = idx;
244 break;
245 }
246
247 if (per >= data.perc_[idx] && idx != 0 && data.perc_[idx] != data.perc_[idx - 1]) {
248 lower_idx = idx;
249 }
250 }
251
252 const double lower_perc = data.perc_[lower_idx];
253 const double lower_value = static_cast<double>(GetRange(lower_idx));
254 if (per == lower_perc) {
255 return lower_value;
256 }
257
258 const double upper_perc = data.perc_[upper_idx];
259 const double upper_value = static_cast<double>(GetRange(upper_idx));
260 if (per == upper_perc) {
261 return upper_value;
262 }
263 DCHECK_GT(upper_perc, lower_perc);
264
265 double value = lower_value + (upper_value - lower_value) *
266 (per - lower_perc) / (upper_perc - lower_perc);
267
268 if (value < min_value_added_) {
269 value = min_value_added_;
270 } else if (value > max_value_added_) {
271 value = max_value_added_;
272 }
273
274 return value;
275 }
276
277 #pragma clang diagnostic pop
278
279 } // namespace art
280 #endif // ART_LIBARTBASE_BASE_HISTOGRAM_INL_H_
281