Blob summary.go
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 |
package tdigest import ( "fmt" "math" "sort" ) type summary struct { means []float64 counts []uint64 } func newSummary(initialCapacity int) *summary { s := &summary{ means: make([]float64, 0, initialCapacity), counts: make([]uint64, 0, initialCapacity), } return s } func (s *summary) Len() int { return len(s.means) } func (s *summary) Add(key float64, value uint64) error { if math.IsNaN(key) { return fmt.Errorf("Key must not be NaN") } if value == 0 { return fmt.Errorf("Count must be >0") } idx := s.findInsertionIndex(key) s.means = append(s.means, math.NaN()) s.counts = append(s.counts, 0) copy(s.means[idx+1:], s.means[idx:]) copy(s.counts[idx+1:], s.counts[idx:]) s.means[idx] = key s.counts[idx] = value return nil } // Always insert to the right func (s *summary) findInsertionIndex(x float64) int { // Binary search is only worthwhile if we have a lot of keys. if len(s.means) < 250 { for i, mean := range s.means { if mean > x { return i } } return len(s.means) } return sort.Search(len(s.means), func(i int) bool { return s.means[i] > x }) } // This method is the hotspot when calling Add(), which in turn is called by // Compress() and Merge(). func (s *summary) HeadSum(idx int) (sum float64) { return float64(sumUntilIndex(s.counts, idx)) } func (s *summary) Floor(x float64) int { return s.findIndex(x) - 1 } func (s *summary) findIndex(x float64) int { // Binary search is only worthwhile if we have a lot of keys. if len(s.means) < 250 { for i, mean := range s.means { if mean >= x { return i } } return len(s.means) } return sort.Search(len(s.means), func(i int) bool { return s.means[i] >= x }) } func (s *summary) Mean(uncheckedIndex int) float64 { return s.means[uncheckedIndex] } func (s *summary) Count(uncheckedIndex int) uint64 { return s.counts[uncheckedIndex] } // return the index of the last item which the sum of counts // of items before it is less than or equal to `sum`. -1 in // case no centroid satisfies the requirement. // Since it's cheap, this also returns the `HeadSum` until // the found index (i.e. cumSum = HeadSum(FloorSum(x))) func (s *summary) FloorSum(sum float64) (index int, cumSum float64) { index = -1 for i, count := range s.counts { if cumSum <= sum { index = i } else { break } cumSum += float64(count) } if index != -1 { cumSum -= float64(s.counts[index]) } return index, cumSum } func (s *summary) setAt(index int, mean float64, count uint64) { s.means[index] = mean s.counts[index] = count s.adjustRight(index) s.adjustLeft(index) } func (s *summary) adjustRight(index int) { for i := index + 1; i < len(s.means) && s.means[i-1] > s.means[i]; i++ { s.means[i-1], s.means[i] = s.means[i], s.means[i-1] s.counts[i-1], s.counts[i] = s.counts[i], s.counts[i-1] } } func (s *summary) adjustLeft(index int) { for i := index - 1; i >= 0 && s.means[i] > s.means[i+1]; i-- { s.means[i], s.means[i+1] = s.means[i+1], s.means[i] s.counts[i], s.counts[i+1] = s.counts[i+1], s.counts[i] } } func (s *summary) ForEach(f func(float64, uint64) bool) { for i, mean := range s.means { if !f(mean, s.counts[i]) { break } } } func (s *summary) Perm(rng RNG, f func(float64, uint64) bool) { for _, i := range perm(rng, s.Len()) { if !f(s.means[i], s.counts[i]) { break } } } func (s *summary) Clone() *summary { return &summary{ means: append([]float64{}, s.means...), counts: append([]uint64{}, s.counts...), } } // Randomly shuffles summary contents, so they can be added to another summary // with being pathological. Renders summary invalid. func (s *summary) shuffle(rng RNG) { for i := len(s.means) - 1; i > 1; i-- { s.Swap(i, rng.Intn(i+1)) } } // for sort.Interface func (s *summary) Swap(i, j int) { s.means[i], s.means[j] = s.means[j], s.means[i] s.counts[i], s.counts[j] = s.counts[j], s.counts[i] } func (s *summary) Less(i, j int) bool { return s.means[i] < s.means[j] } // A simple loop unroll saves a surprising amount of time. func sumUntilIndex(s []uint64, idx int) uint64 { var cumSum uint64 var i int for i = idx - 1; i >= 3; i -= 4 { cumSum += uint64(s[i]) cumSum += uint64(s[i-1]) cumSum += uint64(s[i-2]) cumSum += uint64(s[i-3]) } for ; i >= 0; i-- { cumSum += uint64(s[i]) } return cumSum } func perm(rng RNG, n int) []int { m := make([]int, n) for i := 1; i < n; i++ { j := rng.Intn(i + 1) m[i] = m[j] m[j] = i } return m } |