├── .gitattributes ├── .github ├── FUNDING.yml ├── pull_request_template.md └── workflows │ └── ci.yml ├── .gitignore ├── LICENSE ├── LICENSE-APACHE ├── LICENSE-MIT ├── README.md ├── build.zig └── src ├── MeasuredAllocator.zig ├── benchmark.zig ├── binaryfusefilter.zig ├── fusefilter.zig ├── main.zig ├── unique.zig ├── util.zig └── xorfilter.zig /.gitattributes: -------------------------------------------------------------------------------- 1 | * text=auto eol=lf 2 | 3 | -------------------------------------------------------------------------------- /.github/FUNDING.yml: -------------------------------------------------------------------------------- 1 | github: emidoots 2 | -------------------------------------------------------------------------------- /.github/pull_request_template.md: -------------------------------------------------------------------------------- 1 | 2 | 3 | - [ ] By selecting this checkbox, I agree to license my contributions to this project under the license(s) described in the LICENSE file, and I have the right to do so or have received permission to do so by an employer or client I am producing work for whom has this right. -------------------------------------------------------------------------------- /.github/workflows/ci.yml: -------------------------------------------------------------------------------- 1 | name: CI 2 | on: 3 | - push 4 | - pull_request 5 | jobs: 6 | test: 7 | runs-on: ubuntu-latest 8 | steps: 9 | - name: Checkout 10 | uses: actions/checkout@v2 11 | - name: Setup Zig 12 | run: | 13 | sudo apt install xz-utils 14 | sudo sh -c 'wget -c https://pkg.machengine.org/zig/zig-linux-x86_64-0.14.0-dev.2577+271452d22.tar.xz -O - | tar -xJ --strip-components=1 -C /usr/local/bin' 15 | - name: build 16 | run: zig build 17 | - name: test 18 | run: zig build test 19 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # This file is for zig-specific build artifacts. 2 | # If you have OS-specific or editor-specific files to ignore, 3 | # such as *.swp or .DS_Store, put those in your global 4 | # ~/.gitignore and put this in your ~/.gitconfig: 5 | # 6 | # [core] 7 | # excludesfile = ~/.gitignore 8 | # 9 | # Cheers! 10 | # -andrewrk 11 | 12 | .zig-cache/ 13 | zig-out/ 14 | /release/ 15 | /debug/ 16 | /build/ 17 | /build-*/ 18 | /docgen_tmp/ 19 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Copyright 2021, Hexops Contributors (given via the Git commit history). 2 | 3 | Licensed under the Apache License, Version 2.0 (see LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0) 4 | or the MIT license (see LICENSE-MIT or http://opensource.org/licenses/MIT), at 5 | your option. All files in the project without exclusions may not be copied, 6 | modified, or distributed except according to those terms. 7 | -------------------------------------------------------------------------------- /LICENSE-APACHE: -------------------------------------------------------------------------------- 1 | 2 | Apache License 3 | Version 2.0, January 2004 4 | http://www.apache.org/licenses/ 5 | 6 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 7 | 8 | 1. Definitions. 9 | 10 | "License" shall mean the terms and conditions for use, reproduction, 11 | and distribution as defined by Sections 1 through 9 of this document. 12 | 13 | "Licensor" shall mean the copyright owner or entity authorized by 14 | the copyright owner that is granting the License. 15 | 16 | "Legal Entity" shall mean the union of the acting entity and all 17 | other entities that control, are controlled by, or are under common 18 | control with that entity. For the purposes of this definition, 19 | "control" means (i) the power, direct or indirect, to cause the 20 | direction or management of such entity, whether by contract or 21 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 22 | outstanding shares, or (iii) beneficial ownership of such entity. 23 | 24 | "You" (or "Your") shall mean an individual or Legal Entity 25 | exercising permissions granted by this License. 26 | 27 | "Source" form shall mean the preferred form for making modifications, 28 | including but not limited to software source code, documentation 29 | source, and configuration files. 30 | 31 | "Object" form shall mean any form resulting from mechanical 32 | transformation or translation of a Source form, including but 33 | not limited to compiled object code, generated documentation, 34 | and conversions to other media types. 35 | 36 | "Work" shall mean the work of authorship, whether in Source or 37 | Object form, made available under the License, as indicated by a 38 | copyright notice that is included in or attached to the work 39 | (an example is provided in the Appendix below). 40 | 41 | "Derivative Works" shall mean any work, whether in Source or Object 42 | form, that is based on (or derived from) the Work and for which the 43 | editorial revisions, annotations, elaborations, or other modifications 44 | represent, as a whole, an original work of authorship. For the purposes 45 | of this License, Derivative Works shall not include works that remain 46 | separable from, or merely link (or bind by name) to the interfaces of, 47 | the Work and Derivative Works thereof. 48 | 49 | "Contribution" shall mean any work of authorship, including 50 | the original version of the Work and any modifications or additions 51 | to that Work or Derivative Works thereof, that is intentionally 52 | submitted to Licensor for inclusion in the Work by the copyright owner 53 | or by an individual or Legal Entity authorized to submit on behalf of 54 | the copyright owner. For the purposes of this definition, "submitted" 55 | means any form of electronic, verbal, or written communication sent 56 | to the Licensor or its representatives, including but not limited to 57 | communication on electronic mailing lists, source code control systems, 58 | and issue tracking systems that are managed by, or on behalf of, the 59 | Licensor for the purpose of discussing and improving the Work, but 60 | excluding communication that is conspicuously marked or otherwise 61 | designated in writing by the copyright owner as "Not a Contribution." 62 | 63 | "Contributor" shall mean Licensor and any individual or Legal Entity 64 | on behalf of whom a Contribution has been received by Licensor and 65 | subsequently incorporated within the Work. 66 | 67 | 2. Grant of Copyright License. Subject to the terms and conditions of 68 | this License, each Contributor hereby grants to You a perpetual, 69 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 70 | copyright license to reproduce, prepare Derivative Works of, 71 | publicly display, publicly perform, sublicense, and distribute the 72 | Work and such Derivative Works in Source or Object form. 73 | 74 | 3. Grant of Patent License. Subject to the terms and conditions of 75 | this License, each Contributor hereby grants to You a perpetual, 76 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 77 | (except as stated in this section) patent license to make, have made, 78 | use, offer to sell, sell, import, and otherwise transfer the Work, 79 | where such license applies only to those patent claims licensable 80 | by such Contributor that are necessarily infringed by their 81 | Contribution(s) alone or by combination of their Contribution(s) 82 | with the Work to which such Contribution(s) was submitted. If You 83 | institute patent litigation against any entity (including a 84 | cross-claim or counterclaim in a lawsuit) alleging that the Work 85 | or a Contribution incorporated within the Work constitutes direct 86 | or contributory patent infringement, then any patent licenses 87 | granted to You under this License for that Work shall terminate 88 | as of the date such litigation is filed. 89 | 90 | 4. Redistribution. You may reproduce and distribute copies of the 91 | Work or Derivative Works thereof in any medium, with or without 92 | modifications, and in Source or Object form, provided that You 93 | meet the following conditions: 94 | 95 | (a) You must give any other recipients of the Work or 96 | Derivative Works a copy of this License; and 97 | 98 | (b) You must cause any modified files to carry prominent notices 99 | stating that You changed the files; and 100 | 101 | (c) You must retain, in the Source form of any Derivative Works 102 | that You distribute, all copyright, patent, trademark, and 103 | attribution notices from the Source form of the Work, 104 | excluding those notices that do not pertain to any part of 105 | the Derivative Works; and 106 | 107 | (d) If the Work includes a "NOTICE" text file as part of its 108 | distribution, then any Derivative Works that You distribute must 109 | include a readable copy of the attribution notices contained 110 | within such NOTICE file, excluding those notices that do not 111 | pertain to any part of the Derivative Works, in at least one 112 | of the following places: within a NOTICE text file distributed 113 | as part of the Derivative Works; within the Source form or 114 | documentation, if provided along with the Derivative Works; or, 115 | within a display generated by the Derivative Works, if and 116 | wherever such third-party notices normally appear. The contents 117 | of the NOTICE file are for informational purposes only and 118 | do not modify the License. You may add Your own attribution 119 | notices within Derivative Works that You distribute, alongside 120 | or as an addendum to the NOTICE text from the Work, provided 121 | that such additional attribution notices cannot be construed 122 | as modifying the License. 123 | 124 | You may add Your own copyright statement to Your modifications and 125 | may provide additional or different license terms and conditions 126 | for use, reproduction, or distribution of Your modifications, or 127 | for any such Derivative Works as a whole, provided Your use, 128 | reproduction, and distribution of the Work otherwise complies with 129 | the conditions stated in this License. 130 | 131 | 5. Submission of Contributions. Unless You explicitly state otherwise, 132 | any Contribution intentionally submitted for inclusion in the Work 133 | by You to the Licensor shall be under the terms and conditions of 134 | this License, without any additional terms or conditions. 135 | Notwithstanding the above, nothing herein shall supersede or modify 136 | the terms of any separate license agreement you may have executed 137 | with Licensor regarding such Contributions. 138 | 139 | 6. Trademarks. This License does not grant permission to use the trade 140 | names, trademarks, service marks, or product names of the Licensor, 141 | except as required for reasonable and customary use in describing the 142 | origin of the Work and reproducing the content of the NOTICE file. 143 | 144 | 7. Disclaimer of Warranty. Unless required by applicable law or 145 | agreed to in writing, Licensor provides the Work (and each 146 | Contributor provides its Contributions) on an "AS IS" BASIS, 147 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 148 | implied, including, without limitation, any warranties or conditions 149 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 150 | PARTICULAR PURPOSE. You are solely responsible for determining the 151 | appropriateness of using or redistributing the Work and assume any 152 | risks associated with Your exercise of permissions under this License. 153 | 154 | 8. Limitation of Liability. In no event and under no legal theory, 155 | whether in tort (including negligence), contract, or otherwise, 156 | unless required by applicable law (such as deliberate and grossly 157 | negligent acts) or agreed to in writing, shall any Contributor be 158 | liable to You for damages, including any direct, indirect, special, 159 | incidental, or consequential damages of any character arising as a 160 | result of this License or out of the use or inability to use the 161 | Work (including but not limited to damages for loss of goodwill, 162 | work stoppage, computer failure or malfunction, or any and all 163 | other commercial damages or losses), even if such Contributor 164 | has been advised of the possibility of such damages. 165 | 166 | 9. Accepting Warranty or Additional Liability. While redistributing 167 | the Work or Derivative Works thereof, You may choose to offer, 168 | and charge a fee for, acceptance of support, warranty, indemnity, 169 | or other liability obligations and/or rights consistent with this 170 | License. However, in accepting such obligations, You may act only 171 | on Your own behalf and on Your sole responsibility, not on behalf 172 | of any other Contributor, and only if You agree to indemnify, 173 | defend, and hold each Contributor harmless for any liability 174 | incurred by, or claims asserted against, such Contributor by reason 175 | of your accepting any such warranty or additional liability. 176 | 177 | END OF TERMS AND CONDITIONS 178 | 179 | APPENDIX: How to apply the Apache License to your work. 180 | 181 | To apply the Apache License to your work, attach the following 182 | boilerplate notice, with the fields enclosed by brackets "[]" 183 | replaced with your own identifying information. (Don't include 184 | the brackets!) The text should be enclosed in the appropriate 185 | comment syntax for the file format. We also recommend that a 186 | file or class name and description of purpose be included on the 187 | same "printed page" as the copyright notice for easier 188 | identification within third-party archives. 189 | 190 | Copyright [yyyy] [name of copyright owner] 191 | 192 | Licensed under the Apache License, Version 2.0 (the "License"); 193 | you may not use this file except in compliance with the License. 194 | You may obtain a copy of the License at 195 | 196 | http://www.apache.org/licenses/LICENSE-2.0 197 | 198 | Unless required by applicable law or agreed to in writing, software 199 | distributed under the License is distributed on an "AS IS" BASIS, 200 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 201 | See the License for the specific language governing permissions and 202 | limitations under the License. 203 | -------------------------------------------------------------------------------- /LICENSE-MIT: -------------------------------------------------------------------------------- 1 | Copyright (c) 2021 Hexops Contributors (given via the Git commit history). 2 | 3 | Permission is hereby granted, free of charge, to any 4 | person obtaining a copy of this software and associated 5 | documentation files (the "Software"), to deal in the 6 | Software without restriction, including without 7 | limitation the rights to use, copy, modify, merge, 8 | publish, distribute, sublicense, and/or sell copies of 9 | the Software, and to permit persons to whom the Software 10 | is furnished to do so, subject to the following 11 | conditions: 12 | 13 | The above copyright notice and this permission notice 14 | shall be included in all copies or substantial portions 15 | of the Software. 16 | 17 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF 18 | ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED 19 | TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A 20 | PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT 21 | SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY 22 | CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION 23 | OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR 24 | IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER 25 | DEALINGS IN THE SOFTWARE. 26 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # fastfilter: Binary fuse & xor filters for Zig Hexops logo 2 | 3 | [![CI](https://github.com/hexops/fastfilter/workflows/CI/badge.svg)](https://github.com/hexops/fastfilter/actions) 4 | 5 | comparison 6 | 7 | Binary fuse filters & xor filters are probabilistic data structures which allow for quickly checking whether an element is part of a set. 8 | 9 | Both are faster and more concise than Bloom filters, and smaller than Cuckoo filters. Binary fuse filters are a bleeding-edge development and are competitive with Facebook's ribbon filters: 10 | 11 | * Thomas Mueller Graf, Daniel Lemire, Binary Fuse Filters: Fast and Smaller Than Xor Filters (_not yet published_) 12 | * Thomas Mueller Graf, Daniel Lemire, [Xor Filters: Faster and Smaller Than Bloom and Cuckoo Filters](https://arxiv.org/abs/1912.08258), Journal of Experimental Algorithmics 25 (1), 2020. DOI: 10.1145/3376122 13 | 14 | ## Benefits of Zig implementation 15 | 16 | This is a [Zig](https://ziglang.org) implementation, which provides many practical benefits: 17 | 18 | 1. **Iterator-based:** you can populate xor or binary fuse filters using an iterator, without keeping your entire key set in-memory and without it being a contiguous array of keys. This can reduce memory usage when populating filters substantially. 19 | 2. **Distinct allocators:** you can provide separate Zig `std.mem.Allocator` implementations for the filter itself and population, enabling interesting opportunities like mmap-backed population of filters with low physical memory usage. 20 | 3. **Generic implementation:** use `Xor(u8)`, `Xor(u16)`, `BinaryFuse(u8)`, `BinaryFuse(u16)`, or experiment with more exotic variants like `Xor(u4)` thanks to Zig's [bit-width integers](https://ziglang.org/documentation/master/#Runtime-Integer-Values) and generic type system. 21 | 22 | Zig's safety-checking and checked overflows has also enabled us to improve the upstream C/Go implementations where overflow and undefined behavior went unnoticed.[[1]](https://github.com/FastFilter/xor_singleheader/issues/26) 23 | 24 | ## Usage 25 | 26 | Decide if xor or binary fuse filters fit your use case better: [should I use binary fuse filters or xor filters?](#should-i-use-binary-fuse-filters-or-xor-filters) 27 | 28 | Get your keys into `u64` values. If you have strings, structs, etc. then use something like Zig's [`std.hash_map.getAutoHashFn`](https://ziglang.org/documentation/master/std/#std;hash_map.getAutoHashFn) to convert your keys to `u64` first. ("It is not important to have a good hash function, but collisions should be unlikely (~1/2^64).") 29 | 30 | Create a `build.zig.zon` file in your project (replace `$LATEST_COMMIT` with the latest commit hash): 31 | 32 | ``` 33 | .{ 34 | .name = "mypkg", 35 | .version = "0.1.0", 36 | .dependencies = .{ 37 | .fastfilter = .{ 38 | .url = "https://github.com/hexops/fastfilter/archive/$LATEST_COMMIT.tar.gz", 39 | }, 40 | }, 41 | } 42 | ``` 43 | 44 | Run `zig build` in your project, and the compiler instruct you to add a `.hash = "..."` field next to `.url`. 45 | 46 | Then use the dependency in your `build.zig`: 47 | 48 | ```zig 49 | pub fn build(b: *std.Build) void { 50 | ... 51 | exe.addModule("fastfilter", b.dependency("fastfilter", .{ 52 | .target = target, 53 | .optimize = optimize, 54 | }).module("fastfilter")); 55 | } 56 | ``` 57 | 58 | In your `main.zig`, make use of the library: 59 | 60 | ```zig 61 | const std = @import("std"); 62 | const testing = std.testing; 63 | const fastfilter = @import("fastfilter"); 64 | 65 | test "mytest" { 66 | const allocator = std.heap.page_allocator; 67 | 68 | // Initialize the binary fuse filter with room for 1 million keys. 69 | const size = 1_000_000; 70 | var filter = try fastfilter.BinaryFuse8.init(allocator, size); 71 | defer filter.deinit(allocator); 72 | 73 | // Generate some consecutive keys. 74 | var keys = try allocator.alloc(u64, size); 75 | defer allocator.free(keys); 76 | for (keys, 0..) |key, i| { 77 | _ = key; 78 | keys[i] = i; 79 | } 80 | 81 | // Populate the filter with our keys. You can't update a xor / binary fuse filter after the 82 | // fact, instead you should build a new one. 83 | try filter.populate(allocator, keys[0..]); 84 | 85 | // Now we can quickly test for containment. So fast! 86 | try testing.expect(filter.contain(1) == true); 87 | } 88 | ``` 89 | 90 | (you can just add this project as a Git submodule in yours for now, as [Zig's official package manager is still under way](https://github.com/ziglang/zig/issues/943).) 91 | 92 | Binary fuse filters automatically deduplicate any keys during population. If you are using a different filter type (you probably shouldn't be!) then keys must be unique or else filter population will fail. You can use the `fastfilter.AutoUnique(u64)(keys)` helper to deduplicate (in typically O(N) time complexity), see the tests in `src/unique.zig` for usage examples. 93 | 94 | ## Serialization 95 | 96 | To serialize the filters, you only need to encode these struct fields: 97 | 98 | ```zig 99 | pub fn BinaryFuse(comptime T: type) type { 100 | return struct { 101 | ... 102 | seed: u64, 103 | segment_length: u32, 104 | segment_length_mask: u32, 105 | segment_count: u32, 106 | segment_count_length: u32, 107 | fingerprints: []T, 108 | ... 109 | ``` 110 | 111 | `T` will be the chosen fingerprint size, e.g. `u8` for `BinaryFuse8` or `Xor8`. 112 | 113 | Look at [`std.io.Writer`](https://sourcegraph.com/github.com/ziglang/zig/-/blob/lib/std/io/writer.zig) and [`std.io.BitWriter`](https://sourcegraph.com/github.com/ziglang/zig/-/blob/lib/std/io/bit_writer.zig) for ideas on actual serialization. 114 | 115 | Similarly, for xor filters you only need these struct fields: 116 | 117 | ```zig 118 | pub fn Xor(comptime T: type) type { 119 | return struct { 120 | seed: u64, 121 | blockLength: u64, 122 | fingerprints: []T, 123 | ... 124 | ``` 125 | 126 | ## Should I use binary fuse filters or xor filters? 127 | 128 | If you're not sure, start with `BinaryFuse8` filters. They're fast, and have a false-positive probability rate of 1/256 (or 0.4%). 129 | 130 | There are many tradeoffs, primarily between: 131 | 132 | * Memory usage 133 | * Containment check time 134 | * Population / creation time & memory usage 135 | 136 | See the [benchmarks](#benchmarks) section for a comparison of the tradeoffs between binary fuse filters and xor filters, as well as how larger bit sizes (e.g. `BinaryFuse(u16)`) consume more memory in exchange for a lower false-positive probability rate. 137 | 138 | Note that _fuse filters_ are not to be confused with _binary fuse filters_, the former have issues with construction, often failing unless you have a large number of unique keys. Binary fuse filters do not suffer from this and are generally better than traditional ones in several ways. For this reason, we consider traditional fuse filters deprecated. 139 | 140 | ## Note about extremely large datasets 141 | 142 | This implementation supports key iterators, so you do not need to have all of your keys in-memory, see `BinaryFuse8.populateIter` and `Xor8.populateIter`. 143 | 144 | If you intend to use a xor filter with datasets of 100m+ keys, there is a possible faster implementation _for construction_ found in the C implementation [`xor8_buffered_populate`](https://github.com/FastFilter/xor_singleheader) which is not implemented here. 145 | 146 | ## Changelog 147 | 148 | The API is generally finalized, but we may make some adjustments as Zig changes or we learn of more idiomatic ways to express things. We will release v1.0 once Zig v1.0 is released. 149 | 150 | ### **v0.11.0** 151 | 152 | - fastfilter is now available via the Zig package manager. 153 | - Updated to the latest version of Zig nightly `0.12.0-dev.706+62a0fbdae` 154 | 155 | ### **v0.10.3** 156 | 157 | - Updated to the latest version of Zig `0.12.0-dev.706+62a0fbdae` (`build.zig` `.path` -> `.source` change.) 158 | 159 | ### **v0.10.2** 160 | 161 | - Fixed a few correctness / integer overflow/underflow possibilities where we were inconsistent with the Go/C implementations of binary fuse filters. 162 | - Added debug-mode checks for iterator correctness (wraparound behavior.) 163 | 164 | ### **v0.10.1** 165 | 166 | - Updated to the latest version of Zig `0.12.0-dev.706+62a0fbdae` 167 | 168 | ### **v0.10.0** 169 | 170 | - All types are now unmanaged (allocator must be passed via parameters) 171 | - Renamed `util.sliceIterator` to `fastfilter.SliceIterator` 172 | - `SliceIterator` is now unmanaged / does not store an allocator. 173 | - `SliceIterator` now stores `[]const T` instead of `[]T` internally. 174 | - `BinaryFuseFilter.max_iterations` is now a constant. 175 | - Added `fastfilter.MeasuredAllocator` for measuring allocations. 176 | - Improved usage example. 177 | - Properly free xorfilter/fusefilter fingerprints. 178 | - Updated benchmark to latest Zig version. 179 | 180 | ### **v0.9.3** 181 | 182 | - Fixed potential integer overflow. 183 | 184 | ### **v0.9.2** 185 | 186 | - Handle duplicated keys automatically 187 | - Added a `std.build.Pkg` definition 188 | - Fixed an unlikely bug 189 | - Updated usage instructions 190 | - Updated to Zig v0.10.0-dev.1736 191 | 192 | ### **v0.9.1** 193 | 194 | - Updated to Zig v0.10.0-dev.36 195 | 196 | ### **v0.9.0** 197 | 198 | - Renamed repository github.com/hexops/xorfilter -> github.com/hexops/fastfilter to account for binary fuse filters. 199 | - Implemented bleeding-edge (paper not yet published) "Binary Fuse Filters: Fast and Smaller Than Xor Filters" algorithm by Thomas Mueller Graf, Daniel Lemire 200 | - `BinaryFuse` filters are now recommended by default, are generally better than Xor and Fuse filters. 201 | - Deprecated traditional `Fuse` filters (`BinaryFuse` are much better.) 202 | - Added much improved benchmarking suite with more details on memory consumption during filter population, etc. 203 | 204 | ### **v0.8.0** 205 | 206 | initial release with support for Xor and traditional Fuse filters of varying bit sizes, key iterators, serialization, and a slice de-duplication helper. 207 | 208 | ## Benchmarks 209 | 210 | Benchmarks were ran on both a 2019 Macbook Pro and Windows 10 desktop machine using e.g.: 211 | 212 | ``` 213 | zig run -O ReleaseFast src/benchmark.zig -- --xor 8 --num-keys 1000000 214 | ``` 215 | 216 |
217 | Benchmarks: 2019 Macbook Pro, Intel i9 (1M - 100M keys) 218 | 219 | * CPU: 2.3 GHz 8-Core Intel Core i9 220 | * Memory: 16 GB 2667 MHz DDR4 221 | * Zig version: `0.12.0-dev.706+62a0fbdae` 222 | 223 | | Algorithm | # of keys | populate | contains(k) | false+ prob. | bits per entry | peak populate | filter total | 224 | |--------------|------------|------------|-------------|--------------|----------------|---------------|--------------| 225 | | binaryfuse8 | 1000000 | 37.5ms | 24.0ns | 0.00391115 | 9.04 | 22 MiB | 1 MiB | 226 | | binaryfuse16 | 1000000 | 45.5ms | 24.0ns | 0.00001524 | 18.09 | 24 MiB | 2 MiB | 227 | | binaryfuse32 | 1000000 | 56.0ms | 24.0ns | 0 | 36.18 | 28 MiB | 4 MiB | 228 | | xor2 | 1000000 | 108.0ms | 25.0ns | 0.2500479 | 9.84 | 52 MiB | 1 MiB | 229 | | xor4 | 1000000 | 99.0ms | 25.0ns | 0.06253865 | 9.84 | 52 MiB | 1 MiB | 230 | | xor8 | 1000000 | 103.4ms | 25.0ns | 0.0039055 | 9.84 | 52 MiB | 1 MiB | 231 | | xor16 | 1000000 | 104.7ms | 26.0ns | 0.00001509 | 19.68 | 52 MiB | 2 MiB | 232 | | xor32 | 1000000 | 102.2ms | 25.0ns | 0 | 39.36 | 52 MiB | 4 MiB | 233 | | | | | | | | | | 234 | | binaryfuse8 | 10000000 | 621.2ms | 36.0ns | 0.0039169 | 9.02 | 225 MiB | 10 MiB | 235 | | binaryfuse16 | 10000000 | 666.6ms | 102.0ns | 0.0000147 | 18.04 | 245 MiB | 21 MiB | 236 | | binaryfuse32 | 10000000 | 769.0ms | 135.0ns | 0 | 36.07 | 286 MiB | 43 MiB | 237 | | xor2 | 10000000 | 1.9s | 43.0ns | 0.2500703 | 9.84 | 527 MiB | 11 MiB | 238 | | xor4 | 10000000 | 2.0s | 41.0ns | 0.0626137 | 9.84 | 527 MiB | 11 MiB | 239 | | xor8 | 10000000 | 1.9s | 42.0ns | 0.0039369 | 9.84 | 527 MiB | 11 MiB | 240 | | xor16 | 10000000 | 2.2s | 106.0ns | 0.0000173 | 19.68 | 527 MiB | 23 MiB | 241 | | xor32 | 10000000 | 2.2s | 140.0ns | 0 | 39.36 | 527 MiB | 46 MiB | 242 | | | | | | | | | | 243 | | binaryfuse8 | 100000000 | 7.4s | 145.0ns | 0.003989 | 9.01 | 2 GiB | 107 MiB | 244 | | binaryfuse16 | 100000000 | 8.4s | 169.0ns | 0.000016 | 18.01 | 2 GiB | 214 MiB | 245 | | binaryfuse32 | 100000000 | 10.2s | 173.0ns | 0 | 36.03 | 2 GiB | 429 MiB | 246 | | xor2 | 100000000 | 28.5s | 144.0ns | 0.249843 | 9.84 | 5 GiB | 117 MiB | 247 | | xor4 | 100000000 | 27.4s | 154.0ns | 0.062338 | 9.84 | 5 GiB | 117 MiB | 248 | | xor8 | 100000000 | 28.0s | 153.0ns | 0.004016 | 9.84 | 5 GiB | 117 MiB | 249 | | xor16 | 100000000 | 29.5s | 161.0ns | 0.000012 | 19.68 | 5 GiB | 234 MiB | 250 | | xor32 | 100000000 | 29.4s | 157.0ns | 0 | 39.36 | 5 GiB | 469 MiB | 251 | | | | | | | | | | 252 | 253 | Legend: 254 | 255 | * **contains(k)**: The time taken to check if a key is in the filter 256 | * **false+ prob.**: False positive probability, the probability that a containment check will erroneously return true for a key that has not actually been added to the filter. 257 | * **bits per entry**: The amount of memory in bits the filter uses to store a single entry. 258 | * **peak populate**: Amount of memory consumed during filter population, excluding keys themselves (8 bytes * num_keys.) 259 | * **filter total**: Amount of memory consumed for filter itself in total (bits per entry * entries.) 260 | 261 |
262 | 263 |
264 | Benchmarks: Windows 10, AMD Ryzen 9 3900X (1M - 100M keys) 265 | 266 | * CPU: 3.79Ghz AMD Ryzen 9 3900X 267 | * Memory: 32 GB 2133 MHz DDR4 268 | * Zig version: `0.12.0-dev.706+62a0fbdae` 269 | 270 | | Algorithm | # of keys | populate | contains(k) | false+ prob. | bits per entry | peak populate | filter total | 271 | |--------------|------------|------------|-------------|--------------|----------------|---------------|--------------| 272 | | binaryfuse8 | 1000000 | 44.6ms | 24.0ns | 0.00390796 | 9.04 | 22 MiB | 1 MiB | 273 | | binaryfuse16 | 1000000 | 48.9ms | 25.0ns | 0.00001553 | 18.09 | 24 MiB | 2 MiB | 274 | | binaryfuse32 | 1000000 | 49.9ms | 25.0ns | 0.00000001 | 36.18 | 28 MiB | 4 MiB | 275 | | xor2 | 1000000 | 77.3ms | 25.0ns | 0.25000163 | 9.84 | 52 MiB | 1 MiB | 276 | | xor4 | 1000000 | 80.0ms | 25.0ns | 0.06250427 | 9.84 | 52 MiB | 1 MiB | 277 | | xor8 | 1000000 | 76.0ms | 25.0ns | 0.00391662 | 9.84 | 52 MiB | 1 MiB | 278 | | xor16 | 1000000 | 83.7ms | 26.0ns | 0.00001536 | 19.68 | 52 MiB | 2 MiB | 279 | | xor32 | 1000000 | 79.1ms | 27.0ns | 0 | 39.36 | 52 MiB | 4 MiB | 280 | | fuse8 | 1000000 | 69.4ms | 25.0ns | 0.00390663 | 9.10 | 49 MiB | 1 MiB | 281 | | fuse16 | 1000000 | 71.5ms | 27.0ns | 0.00001516 | 18.20 | 49 MiB | 2 MiB | 282 | | fuse32 | 1000000 | 71.1ms | 27.0ns | 0 | 36.40 | 49 MiB | 4 MiB | 283 | | | | | | | | | | 284 | | binaryfuse8 | 10000000 | 572.3ms | 33.0ns | 0.0038867 | 9.02 | 225 MiB | 10 MiB | 285 | | binaryfuse16 | 10000000 | 610.6ms | 108.0ns | 0.0000127 | 18.04 | 245 MiB | 21 MiB | 286 | | binaryfuse32 | 10000000 | 658.2ms | 144.0ns | 0 | 36.07 | 286 MiB | 43 MiB | 287 | | xor2 | 10000000 | 1.2s | 39.0ns | 0.249876 | 9.84 | 527 MiB | 11 MiB | 288 | | xor4 | 10000000 | 1.2s | 39.0ns | 0.0625026 | 9.84 | 527 MiB | 11 MiB | 289 | | xor8 | 10000000 | 1.2s | 41.0ns | 0.0038881 | 9.84 | 527 MiB | 11 MiB | 290 | | xor16 | 10000000 | 1.3s | 117.0ns | 0.0000134 | 19.68 | 527 MiB | 23 MiB | 291 | | xor32 | 10000000 | 1.3s | 147.0ns | 0 | 39.36 | 527 MiB | 46 MiB | 292 | | fuse8 | 10000000 | 1.1s | 36.0ns | 0.0039089 | 9.10 | 499 MiB | 10 MiB | 293 | | fuse16 | 10000000 | 1.1s | 112.0ns | 0.0000172 | 18.20 | 499 MiB | 21 MiB | 294 | | fuse32 | 10000000 | 1.1s | 145.0ns | 0 | 36.40 | 499 MiB | 43 MiB | 295 | | | | | | | | | | 296 | | binaryfuse8 | 100000000 | 6.9s | 167.0ns | 0.00381 | 9.01 | 2 GiB | 107 MiB | 297 | | binaryfuse16 | 100000000 | 7.2s | 171.0ns | 0.000009 | 18.01 | 2 GiB | 214 MiB | 298 | | binaryfuse32 | 100000000 | 8.5s | 174.0ns | 0 | 36.03 | 2 GiB | 429 MiB | 299 | | xor2 | 100000000 | 16.8s | 166.0ns | 0.249868 | 9.84 | 5 GiB | 117 MiB | 300 | | xor4 | 100000000 | 18.9s | 183.0ns | 0.062417 | 9.84 | 5 GiB | 117 MiB | 301 | | xor8 | 100000000 | 19.1s | 168.0ns | 0.003873 | 9.84 | 5 GiB | 117 MiB | 302 | | xor16 | 100000000 | 16.9s | 171.0ns | 0.000021 | 19.68 | 5 GiB | 234 MiB | 303 | | xor32 | 100000000 | 19.4s | 189.0ns | 0 | 39.36 | 5 GiB | 469 MiB | 304 | | fuse8 | 100000000 | 19.6s | 167.0ns | 0.003797 | 9.10 | 4 GiB | 108 MiB | 305 | | fuse16 | 100000000 | 20.8s | 171.0ns | 0.000015 | 18.20 | 4 GiB | 216 MiB | 306 | | fuse32 | 100000000 | 21.5s | 176.0ns | 0 | 36.40 | 4 GiB | 433 MiB | 307 | | | | | | | | | | 308 | 309 | Legend: 310 | 311 | * **contains(k)**: The time taken to check if a key is in the filter 312 | * **false+ prob.**: False positive probability, the probability that a containment check will erroneously return true for a key that has not actually been added to the filter. 313 | * **bits per entry**: The amount of memory in bits the filter uses to store a single entry. 314 | * **peak populate**: Amount of memory consumed during filter population, excluding keys themselves (8 bytes * num_keys.) 315 | * **filter total**: Amount of memory consumed for filter itself in total (bits per entry * entries.) 316 | 317 |
318 | 319 | ## Related readings 320 | 321 | * Blog post by Daniel Lemire: [Xor Filters: Faster and Smaller Than Bloom Filters](https://lemire.me/blog/2019/12/19/xor-filters-faster-and-smaller-than-bloom-filters) 322 | * Fuse Filters ([arxiv paper](https://arxiv.org/abs/1907.04749)), as described [by @jbapple](https://github.com/FastFilter/xor_singleheader/pull/11#issue-356508475) (note these are not to be confused with _binary fuse filters_.) 323 | 324 | ## Special thanks 325 | 326 | * [**Thomas Mueller Graf**](https://github.com/thomasmueller) and [**Daniel Lemire**](https://github.com/lemire) - _for their excellent research into xor filters, xor+ filters, their C implementation, and more._ 327 | * [**Martin Dietzfelbinger**](https://arxiv.org/search/cs?searchtype=author&query=Dietzfelbinger%2C+M) and [**Stefan Walzer**](https://arxiv.org/search/cs?searchtype=author&query=Walzer%2C+S) - _for their excellent research into fuse filters._ 328 | * [**Jim Apple**](https://github.com/jbapple) - _for their C implementation[[1]](https://github.com/FastFilter/xor_singleheader/pull/11) of fuse filters_ 329 | * [**@Andoryuuta**](https://github.com/Andoryuuta) - _for providing substantial help in debugging several issues in the Zig implementation._ 330 | 331 | If it was not for the above people, I ([@emidoots](https://github.com/emidoots)) would not have been able to write this implementation and learn from the excellent [C implementation](https://github.com/FastFilter/xor_singleheader). Please credit the above people if you use this library. 332 | -------------------------------------------------------------------------------- /build.zig: -------------------------------------------------------------------------------- 1 | const std = @import("std"); 2 | 3 | pub fn build(b: *std.Build) void { 4 | const optimize = b.standardOptimizeOption(.{}); 5 | const target = b.standardTargetOptions(.{}); 6 | 7 | _ = b.addModule("fastfilter", .{ 8 | .root_source_file = b.path("src/main.zig"), 9 | }); 10 | 11 | const main_tests = b.addTest(.{ 12 | .name = "fastfilter-tests", 13 | .root_source_file = b.path("src/main.zig"), 14 | .target = target, 15 | .optimize = optimize, 16 | }); 17 | const run_main_tests = b.addRunArtifact(main_tests); 18 | 19 | const test_step = b.step("test", "Run library tests"); 20 | test_step.dependOn(&run_main_tests.step); 21 | } 22 | -------------------------------------------------------------------------------- /src/MeasuredAllocator.zig: -------------------------------------------------------------------------------- 1 | /// This allocator takes an existing allocator, wraps it, and provides measurements about the 2 | /// allocations such as peak memory usage. 3 | const std = @import("std"); 4 | const assert = std.debug.assert; 5 | const mem = std.mem; 6 | const Allocator = std.mem.Allocator; 7 | 8 | const MeasuredAllocator = @This(); 9 | 10 | parent_allocator: Allocator, 11 | state: State, 12 | 13 | /// Inner state of MeasuredAllocator. Can be stored rather than the entire MeasuredAllocator as a 14 | /// memory-saving optimization. 15 | pub const State = struct { 16 | peak_memory_usage_bytes: usize = 0, 17 | current_memory_usage_bytes: usize = 0, 18 | 19 | pub fn promote(self: State, parent_allocator: Allocator) MeasuredAllocator { 20 | return .{ 21 | .parent_allocator = parent_allocator, 22 | .state = self, 23 | }; 24 | } 25 | }; 26 | 27 | const BufNode = std.SinglyLinkedList([]u8).Node; 28 | 29 | pub fn init(parent_allocator: Allocator) MeasuredAllocator { 30 | return (State{}).promote(parent_allocator); 31 | } 32 | 33 | pub fn allocator(self: *MeasuredAllocator) Allocator { 34 | return .{ 35 | .ptr = self, 36 | .vtable = &.{ 37 | .alloc = allocFn, 38 | .resize = resizeFn, 39 | .free = freeFn, 40 | }, 41 | }; 42 | } 43 | 44 | fn allocFn(ptr: *anyopaque, len: usize, ptr_align: u8, ret_addr: usize) ?[*]u8 { 45 | const self = @as(*MeasuredAllocator, @ptrCast(@alignCast(ptr))); 46 | const result = self.parent_allocator.rawAlloc(len, ptr_align, ret_addr); 47 | if (result) |_| { 48 | self.state.current_memory_usage_bytes += len; 49 | if (self.state.current_memory_usage_bytes > self.state.peak_memory_usage_bytes) self.state.peak_memory_usage_bytes = self.state.current_memory_usage_bytes; 50 | } 51 | return result; 52 | } 53 | 54 | fn resizeFn(ptr: *anyopaque, buf: []u8, buf_align: u8, new_len: usize, ret_addr: usize) bool { 55 | const self = @as(*MeasuredAllocator, @ptrCast(@alignCast(ptr))); 56 | if (self.parent_allocator.rawResize(buf, buf_align, new_len, ret_addr)) { 57 | self.state.current_memory_usage_bytes -= buf.len - new_len; 58 | if (self.state.current_memory_usage_bytes > self.state.peak_memory_usage_bytes) self.state.peak_memory_usage_bytes = self.state.current_memory_usage_bytes; 59 | return true; 60 | } 61 | return false; 62 | } 63 | 64 | fn freeFn(ptr: *anyopaque, buf: []u8, buf_align: u8, ret_addr: usize) void { 65 | const self = @as(*MeasuredAllocator, @ptrCast(@alignCast(ptr))); 66 | self.parent_allocator.rawFree(buf, buf_align, ret_addr); 67 | self.state.current_memory_usage_bytes -= buf.len; 68 | } 69 | -------------------------------------------------------------------------------- /src/benchmark.zig: -------------------------------------------------------------------------------- 1 | // zig run -O ReleaseFast src/benchmark.zig 2 | 3 | const std = @import("std"); 4 | const time = std.time; 5 | const Timer = time.Timer; 6 | const xorfilter = @import("main.zig"); 7 | const MeasuredAllocator = @import("MeasuredAllocator.zig"); 8 | 9 | fn formatTime(writer: anytype, comptime spec: []const u8, start: u64, end: u64, division: usize) !void { 10 | const ns = @as(f64, @floatFromInt((end - start) / division)); 11 | if (ns <= time.ns_per_ms) { 12 | try std.fmt.format(writer, spec, .{ ns, "ns " }); 13 | return; 14 | } 15 | if (ns <= time.ns_per_s) { 16 | try std.fmt.format(writer, spec, .{ ns / @as(f64, @floatFromInt(time.ns_per_ms)), "ms " }); 17 | return; 18 | } 19 | if (ns <= time.ns_per_min) { 20 | try std.fmt.format(writer, spec, .{ ns / @as(f64, @floatFromInt(time.ns_per_s)), "s " }); 21 | return; 22 | } 23 | try std.fmt.format(writer, spec, .{ ns / @as(f64, @floatFromInt(time.ns_per_min)), "min" }); 24 | return; 25 | } 26 | 27 | fn formatBytes(writer: anytype, comptime spec: []const u8, bytes: u64) !void { 28 | const kib = 1024; 29 | const mib = 1024 * kib; 30 | const gib = 1024 * mib; 31 | if (bytes < kib) { 32 | try std.fmt.format(writer, spec, .{ bytes, "B " }); 33 | } 34 | if (bytes < mib) { 35 | try std.fmt.format(writer, spec, .{ bytes / kib, "KiB" }); 36 | return; 37 | } 38 | if (bytes < gib) { 39 | try std.fmt.format(writer, spec, .{ bytes / mib, "MiB" }); 40 | return; 41 | } 42 | try std.fmt.format(writer, spec, .{ bytes / gib, "GiB" }); 43 | return; 44 | } 45 | 46 | fn bench(algorithm: []const u8, Filter: anytype, size: usize, trials: usize) !void { 47 | const allocator = std.heap.page_allocator; 48 | 49 | var filterMA = MeasuredAllocator.init(allocator); 50 | const filterAllocator = filterMA.allocator(); 51 | 52 | var buildMA = MeasuredAllocator.init(allocator); 53 | const buildAllocator = buildMA.allocator(); 54 | 55 | const stdout = std.io.getStdOut().writer(); 56 | var timer = try Timer.start(); 57 | 58 | // Initialize filter. 59 | var filter = try Filter.init(filterAllocator, size); 60 | defer filter.deinit(allocator); 61 | 62 | // Generate keys. 63 | var keys = try allocator.alloc(u64, size); 64 | defer allocator.free(keys); 65 | for (keys, 0..) |_, i| { 66 | keys[i] = i; 67 | } 68 | 69 | // Populate filter. 70 | timer.reset(); 71 | const populateTimeStart = timer.lap(); 72 | try filter.populate(buildAllocator, keys[0..]); 73 | const populateTimeEnd = timer.read(); 74 | 75 | // Perform random matches. 76 | var random_matches: u64 = 0; 77 | var i: u64 = 0; 78 | var rng = std.Random.DefaultPrng.init(0); 79 | const random = rng.random(); 80 | timer.reset(); 81 | const randomMatchesTimeStart = timer.lap(); 82 | while (i < trials) : (i += 1) { 83 | const random_key: u64 = random.uintAtMost(u64, std.math.maxInt(u64)); 84 | if (filter.contain(random_key)) { 85 | if (random_key >= keys.len) { 86 | random_matches += 1; 87 | } 88 | } 89 | } 90 | const randomMatchesTimeEnd = timer.read(); 91 | 92 | const fpp = @as(f64, @floatFromInt(random_matches)) * 1.0 / @as(f64, @floatFromInt(trials)); 93 | 94 | const bitsPerEntry = @as(f64, @floatFromInt(filter.sizeInBytes())) * 8.0 / @as(f64, @floatFromInt(size)); 95 | const filterBitsPerEntry = @as(f64, @floatFromInt(filterMA.state.peak_memory_usage_bytes)) * 8.0 / @as(f64, @floatFromInt(size)); 96 | if (!std.math.approxEqAbs(f64, filterBitsPerEntry, bitsPerEntry, 0.001)) { 97 | @panic("sizeInBytes reporting wrong numbers?"); 98 | } 99 | 100 | try stdout.print("| {s: <12} ", .{algorithm}); 101 | try stdout.print("| {: <10} ", .{keys.len}); 102 | try stdout.print("| ", .{}); 103 | try formatTime(stdout, "{d: >7.1}{s}", populateTimeStart, populateTimeEnd, 1); 104 | try stdout.print(" | ", .{}); 105 | try formatTime(stdout, "{d: >8.1}{s}", randomMatchesTimeStart, randomMatchesTimeEnd, trials); 106 | try stdout.print(" | {d: >12} ", .{fpp}); 107 | try stdout.print("| {d: >14.2} ", .{bitsPerEntry}); 108 | try stdout.print("| ", .{}); 109 | try formatBytes(stdout, "{: >9} {s}", buildMA.state.peak_memory_usage_bytes); 110 | try formatBytes(stdout, " | {: >8} {s}", filterMA.state.peak_memory_usage_bytes); 111 | try stdout.print(" |\n", .{}); 112 | } 113 | 114 | fn usage() void { 115 | std.log.warn( 116 | \\benchmark [options] 117 | \\ 118 | \\Options: 119 | \\ --num-trials [int=10000000] number of trials / containment checks to perform 120 | \\ --help 121 | \\ 122 | , .{}); 123 | } 124 | pub fn main() !void { 125 | var buffer: [1024]u8 = undefined; 126 | var fixed = std.heap.FixedBufferAllocator.init(buffer[0..]); 127 | const args = try std.process.argsAlloc(fixed.allocator()); 128 | 129 | var num_trials: usize = 100_000_000; 130 | var i: usize = 1; 131 | while (i < args.len) : (i += 1) { 132 | if (std.mem.eql(u8, args[i], "--num-trials")) { 133 | i += 1; 134 | if (i == args.len) { 135 | usage(); 136 | std.process.exit(1); 137 | } 138 | num_trials = try std.fmt.parseUnsigned(usize, args[i], 10); 139 | } 140 | } 141 | 142 | const stdout = std.io.getStdOut().writer(); 143 | try stdout.print("| Algorithm | # of keys | populate | contains(k) | false+ prob. | bits per entry | peak populate | filter total |\n", .{}); 144 | try stdout.print("|--------------|------------|------------|-------------|--------------|----------------|---------------|--------------|\n", .{}); 145 | try bench("binaryfuse8", xorfilter.BinaryFuse(u8), 1_000_000, num_trials); 146 | try bench("binaryfuse16", xorfilter.BinaryFuse(u16), 1_000_000, num_trials); 147 | try bench("binaryfuse32", xorfilter.BinaryFuse(u32), 1_000_000, num_trials); 148 | try bench("xor2", xorfilter.Xor(u2), 1_000_000, num_trials); 149 | try bench("xor4", xorfilter.Xor(u4), 1_000_000, num_trials); 150 | try bench("xor8", xorfilter.Xor(u8), 1_000_000, num_trials); 151 | try bench("xor16", xorfilter.Xor(u16), 1_000_000, num_trials); 152 | try bench("xor32", xorfilter.Xor(u32), 1_000_000, num_trials); 153 | try stdout.print("| | | | | | | | |\n", .{}); 154 | try bench("binaryfuse8", xorfilter.BinaryFuse(u8), 10_000_000, num_trials / 10); 155 | try bench("binaryfuse16", xorfilter.BinaryFuse(u16), 10_000_000, num_trials / 10); 156 | try bench("binaryfuse32", xorfilter.BinaryFuse(u32), 10_000_000, num_trials / 10); 157 | try bench("xor2", xorfilter.Xor(u2), 10_000_000, num_trials / 10); 158 | try bench("xor4", xorfilter.Xor(u4), 10_000_000, num_trials / 10); 159 | try bench("xor8", xorfilter.Xor(u8), 10_000_000, num_trials / 10); 160 | try bench("xor16", xorfilter.Xor(u16), 10_000_000, num_trials / 10); 161 | try bench("xor32", xorfilter.Xor(u32), 10_000_000, num_trials / 10); 162 | try stdout.print("| | | | | | | | |\n", .{}); 163 | try bench("binaryfuse8", xorfilter.BinaryFuse(u8), 100_000_000, num_trials / 100); 164 | try bench("binaryfuse16", xorfilter.BinaryFuse(u16), 100_000_000, num_trials / 100); 165 | try bench("binaryfuse32", xorfilter.BinaryFuse(u32), 100_000_000, num_trials / 100); 166 | try bench("xor2", xorfilter.Xor(u2), 100_000_000, num_trials / 100); 167 | try bench("xor4", xorfilter.Xor(u4), 100_000_000, num_trials / 100); 168 | try bench("xor8", xorfilter.Xor(u8), 100_000_000, num_trials / 100); 169 | try bench("xor16", xorfilter.Xor(u16), 100_000_000, num_trials / 100); 170 | try bench("xor32", xorfilter.Xor(u32), 100_000_000, num_trials / 100); 171 | try stdout.print("| | | | | | | | |\n", .{}); 172 | try stdout.print("\n", .{}); 173 | try stdout.print("Legend:\n\n", .{}); 174 | try stdout.print("* **contains(k)**: The time taken to check if a key is in the filter\n", .{}); 175 | try stdout.print("* **false+ prob.**: False positive probability, the probability that a containment check will erroneously return true for a key that has not actually been added to the filter.\n", .{}); 176 | try stdout.print("* **bits per entry**: The amount of memory in bits the filter uses to store a single entry.\n", .{}); 177 | try stdout.print("* **peak populate**: Amount of memory consumed during filter population, excluding keys themselves (8 bytes * num_keys.)\n", .{}); 178 | try stdout.print("* **filter total**: Amount of memory consumed for filter itself in total (bits per entry * entries.)\n", .{}); 179 | } 180 | -------------------------------------------------------------------------------- /src/binaryfusefilter.zig: -------------------------------------------------------------------------------- 1 | const std = @import("std"); 2 | const Allocator = std.mem.Allocator; 3 | const math = std.math; 4 | const testing = std.testing; 5 | 6 | const util = @import("util.zig"); 7 | const Error = util.Error; 8 | 9 | const builtin = @import("builtin"); 10 | const is_debug = builtin.mode == .Debug; 11 | 12 | /// BinaryFuse8 provides a binary fuse filter with 8-bit fingerprints. 13 | /// 14 | /// See `BinaryFuse` for more details. 15 | pub const BinaryFuse8 = BinaryFuse(u8); 16 | 17 | /// probability of success should always be > 0.5 so 100 iterations is highly unlikely (read: 18 | /// statistically unlikely to happen. If it does, probably we need to increase max_iterations or 19 | /// contact the fastfilter authors.) 20 | const max_iterations: usize = 100; 21 | 22 | /// A binary fuse filter. This is an extension of fuse filters: 23 | /// 24 | /// Dietzfelbinger & Walzer's fuse filters, described in "Dense Peelable Random Uniform Hypergraphs", 25 | /// https://arxiv.org/abs/1907.04749, can accomodate fill factors up to 87.9% full, rather than 26 | /// 1 / 1.23 = 81.3%. In the 8-bit case, this reduces the memory usage from 9.84 bits per entry to 27 | /// 9.1 bits. 28 | /// 29 | /// An issue with traditional fuse filters is that the algorithm requires a large number of unique 30 | /// keys in order for population to succeed, see [FastFilter/xor_singleheader#21](https://github.com/FastFilter/xor_singleheader/issues/21). 31 | /// If you have few (<~125k consecutive) keys, fuse filter creation would fail. 32 | /// 33 | /// By contrast, binary fuse filters, a revision of fuse filters made by Thomas Mueller Graf & 34 | /// Daniel Lemire do not suffer from this issue. See https://github.com/FastFilter/xor_singleheader/issues/21 35 | /// 36 | /// Note: We assume that you have a large set of 64-bit integers and you want a data structure to 37 | /// do membership tests using no more than ~8 or ~16 bits per key. If your initial set is made of 38 | /// strings or other types, you first need to hash them to a 64-bit integer. 39 | pub fn BinaryFuse(comptime T: type) type { 40 | return struct { 41 | seed: u64, 42 | segment_length: u32, 43 | segment_length_mask: u32, 44 | segment_count: u32, 45 | segment_count_length: u32, 46 | fingerprints: []T, 47 | 48 | const Self = @This(); 49 | 50 | /// initializes a binary fuse filter with enough capacity for a set containing up to `size` 51 | /// elements. 52 | /// 53 | /// `deinit()` must be called by the caller to free the memory. 54 | pub fn init(allocator: Allocator, size: usize) !Self { 55 | const arity: u32 = 3; 56 | var segment_length = calculateSegmentLength(arity, size); 57 | if (segment_length > 262144) { 58 | segment_length = 262144; 59 | } 60 | const segment_length_mask = segment_length - 1; 61 | const size_factor: f64 = if (size == 0) 4 else calculateSizeFactor(arity, size); 62 | const capacity = if (size <= 1) 0 else @as(u32, @intFromFloat(@round(@as(f64, @floatFromInt(size)) * size_factor))); 63 | const init_segment_count: u32 = (capacity + segment_length - 1) / segment_length -% (arity - 1); 64 | var slice_length = (init_segment_count +% arity - 1) * segment_length; 65 | var segment_count = (slice_length + segment_length - 1) / segment_length; 66 | if (segment_count <= arity - 1) { 67 | segment_count = 1; 68 | } else { 69 | segment_count = segment_count - (arity - 1); 70 | } 71 | slice_length = (segment_count + arity - 1) * segment_length; 72 | const segment_count_length = segment_count * segment_length; 73 | 74 | return Self{ 75 | .seed = undefined, 76 | .segment_length = segment_length, 77 | .segment_length_mask = segment_length_mask, 78 | .segment_count = segment_count, 79 | .segment_count_length = segment_count_length, 80 | .fingerprints = try allocator.alloc(T, slice_length), 81 | }; 82 | } 83 | 84 | pub inline fn deinit(self: *const Self, allocator: Allocator) void { 85 | allocator.free(self.fingerprints); 86 | } 87 | 88 | /// reports the size in bytes of the filter. 89 | pub inline fn sizeInBytes(self: *const Self) usize { 90 | return self.fingerprints.len * @sizeOf(T) + @sizeOf(Self); 91 | } 92 | 93 | /// populates the filter with the given keys. 94 | /// 95 | /// The function could return an error after too many iterations, but it is statistically 96 | /// unlikely and you probably don't need to worry about it. 97 | /// 98 | /// The provided allocator will be used for creating temporary buffers that do not outlive the 99 | /// function call. 100 | pub fn populate(self: *Self, allocator: Allocator, keys: []u64) Error!void { 101 | var iter = util.SliceIterator(u64).init(keys); 102 | return self.populateIter(allocator, &iter); 103 | } 104 | 105 | /// Identical to populate, except it takes an iterator of keys so you need not store them 106 | /// in-memory. 107 | /// 108 | /// `keys.next()` must return `?u64`, the next key or none if the end of the list has been 109 | /// reached. The iterator must reset after hitting the end of the list, such that the `next()` 110 | /// call leads to the first element again. 111 | /// 112 | /// `keys.len()` must return the `usize` length. 113 | pub fn populateIter(self: *Self, allocator: Allocator, keys: anytype) Error!void { 114 | if (keys.len() == 0) { 115 | return; 116 | } 117 | var rng_counter: u64 = 0x726b2b9d438b9d4d; 118 | self.seed = util.rngSplitMix64(&rng_counter); 119 | 120 | var size = keys.len(); 121 | const reverse_order = try allocator.alloc(u64, size + 1); 122 | defer allocator.free(reverse_order); 123 | @memset(reverse_order, 0); 124 | 125 | const capacity = self.fingerprints.len; 126 | const alone = try allocator.alloc(u32, capacity); 127 | defer allocator.free(alone); 128 | 129 | const t2count = try allocator.alloc(T, capacity); 130 | defer allocator.free(t2count); 131 | @memset(t2count, 0); 132 | 133 | const reverse_h = try allocator.alloc(T, size); 134 | defer allocator.free(reverse_h); 135 | 136 | const t2hash = try allocator.alloc(u64, capacity); 137 | defer allocator.free(t2hash); 138 | @memset(t2hash, 0); 139 | 140 | var block_bits: u5 = 1; 141 | while ((@as(u32, 1) << block_bits) < self.segment_count) { 142 | block_bits += 1; 143 | } 144 | const block: u32 = @as(u32, 1) << block_bits; 145 | 146 | const start_pos = try allocator.alloc(u32, @as(usize, 1) << block_bits); 147 | defer allocator.free(start_pos); 148 | 149 | var expect_num_keys: ?usize = null; 150 | 151 | var h012: [5]u32 = undefined; 152 | 153 | reverse_order[size] = 1; 154 | var loop: usize = 0; 155 | while (true) : (loop += 1) { 156 | if (loop + 1 > max_iterations) { 157 | // too many iterations, this is statistically unlikely to happen. If it does, 158 | // probably we need to increase max_iterations or contact the fastfilter authors 159 | return Error.KeysLikelyNotUnique; 160 | } 161 | 162 | var i: u32 = 0; 163 | while (i < block) : (i += 1) { 164 | // important : i * size would overflow as a 32-bit number in some 165 | // cases. 166 | start_pos[i] = @as(u32, @truncate((@as(u64, @intCast(i)) * size) >> block_bits)); 167 | } 168 | 169 | const mask_block: u64 = block - 1; 170 | var got_num_keys: usize = 0; 171 | while (keys.next()) |key| { 172 | if (is_debug) got_num_keys += 1; 173 | const sum: u64 = key +% self.seed; 174 | const hash: u64 = util.murmur64(sum); 175 | 176 | const shift_count = @as(usize, 64) - @as(usize, block_bits); 177 | var segment_index: u64 = if (shift_count >= 63) 0 else hash >> @as(u6, @truncate(shift_count)); 178 | while (reverse_order[start_pos[segment_index]] != 0) { 179 | segment_index += 1; 180 | segment_index &= mask_block; 181 | } 182 | reverse_order[start_pos[segment_index]] = hash; 183 | start_pos[segment_index] += 1; 184 | } 185 | if (is_debug) { 186 | if (expect_num_keys) |expect| { 187 | if (expect != got_num_keys) @panic("fastfilter: iterator illegal: does not wrap around"); 188 | } 189 | expect_num_keys = got_num_keys; 190 | } 191 | 192 | var err = false; 193 | var duplicates: u32 = 0; 194 | i = 0; 195 | while (i < size) : (i += 1) { 196 | const hash = reverse_order[i]; 197 | const h0 = self.fuseHash(0, hash); 198 | const h1 = self.fuseHash(1, hash); 199 | const h2 = self.fuseHash(2, hash); 200 | t2count[h0] +%= 4; 201 | t2hash[h0] ^= hash; 202 | t2count[h1] +%= 4; 203 | t2count[h1] ^= 1; 204 | t2hash[h1] ^= hash; 205 | t2count[h2] +%= 4; 206 | t2count[h2] ^= 2; 207 | t2hash[h2] ^= hash; 208 | // If we have duplicated hash values, then it is likely that the next comparison 209 | // is true 210 | if (t2hash[h0] & t2hash[h1] & t2hash[h2] == 0) { 211 | // next we do the actual test 212 | if (((t2hash[h0] == 0) and (t2count[h0] == 8)) or ((t2hash[h1] == 0) and (t2count[h1] == 8)) or ((t2hash[h2] == 0) and (t2count[h2] == 8))) { 213 | duplicates += 1; 214 | t2count[h0] -%= 4; 215 | t2hash[h0] ^= hash; 216 | t2count[h1] -%= 4; 217 | t2count[h1] ^= 1; 218 | t2hash[h1] ^= hash; 219 | t2count[h2] -%= 4; 220 | t2count[h2] ^= 2; 221 | t2hash[h2] ^= hash; 222 | } 223 | } 224 | err = (t2count[h0] < 4) or err; 225 | err = (t2count[h1] < 4) or err; 226 | err = (t2count[h2] < 4) or err; 227 | } 228 | if (err) { 229 | i = 0; 230 | while (i < size) : (i += 1) { 231 | reverse_order[i] = 0; 232 | } 233 | i = 0; 234 | while (i < capacity) : (i += 1) { 235 | t2count[i] = 0; 236 | t2hash[i] = 0; 237 | } 238 | self.seed = util.rngSplitMix64(&rng_counter); 239 | continue; 240 | } 241 | 242 | // End of key addition 243 | var Qsize: u32 = 0; 244 | // Add sets with one key to the queue. 245 | i = 0; 246 | while (i < capacity) : (i += 1) { 247 | alone[Qsize] = i; 248 | Qsize += if ((t2count[i] >> 2) == 1) @as(u32, 1) else @as(u32, 0); 249 | } 250 | var stacksize: u32 = 0; 251 | while (Qsize > 0) { 252 | Qsize -= 1; 253 | const index: u32 = alone[Qsize]; 254 | if ((t2count[index] >> 2) == 1) { 255 | const hash = t2hash[index]; 256 | 257 | //h012[0] = self.fuseHash(0, hash); 258 | h012[1] = self.fuseHash(1, hash); 259 | h012[2] = self.fuseHash(2, hash); 260 | h012[3] = self.fuseHash(0, hash); // == h012[0]; 261 | h012[4] = h012[1]; 262 | const found = t2count[index] & 3; 263 | reverse_h[stacksize] = found; 264 | reverse_order[stacksize] = hash; 265 | stacksize += 1; 266 | const other_index1 = h012[found + 1]; 267 | alone[Qsize] = other_index1; 268 | Qsize += if ((t2count[other_index1] >> 2) == 2) @as(u32, 1) else @as(u32, 0); 269 | 270 | t2count[other_index1] -= 4; 271 | t2count[other_index1] ^= fuseMod3(T, found + 1); 272 | t2hash[other_index1] ^= hash; 273 | 274 | const other_index2 = h012[found + 2]; 275 | alone[Qsize] = other_index2; 276 | Qsize += if ((t2count[other_index2] >> 2) == 2) @as(u32, 1) else @as(u32, 0); 277 | t2count[other_index2] -= 4; 278 | t2count[other_index2] ^= fuseMod3(T, found + 2); 279 | t2hash[other_index2] ^= hash; 280 | } 281 | } 282 | if (stacksize + duplicates == size) { 283 | // success 284 | size = stacksize; 285 | break; 286 | } 287 | @memset(reverse_order[0..size], 0); 288 | @memset(t2count[0..capacity], 0); 289 | @memset(t2hash[0..capacity], 0); 290 | self.seed = util.rngSplitMix64(&rng_counter); 291 | } 292 | if (size == 0) return; 293 | 294 | var i: u32 = @as(u32, @truncate(size - 1)); 295 | while (i < size) : (i -%= 1) { 296 | // the hash of the key we insert next 297 | const hash: u64 = reverse_order[i]; 298 | const xor2: T = @as(T, @truncate(util.fingerprint(hash))); 299 | const found: T = reverse_h[i]; 300 | h012[0] = self.fuseHash(0, hash); 301 | h012[1] = self.fuseHash(1, hash); 302 | h012[2] = self.fuseHash(2, hash); 303 | h012[3] = h012[0]; 304 | h012[4] = h012[1]; 305 | self.fingerprints[h012[found]] = xor2 ^ self.fingerprints[h012[found + 1]] ^ self.fingerprints[h012[found + 2]]; 306 | } 307 | } 308 | 309 | /// reports if the specified key is within the set with false-positive rate. 310 | pub inline fn contain(self: *const Self, key: u64) bool { 311 | const hash = util.mixSplit(key, self.seed); 312 | var f = @as(T, @truncate(util.fingerprint(hash))); 313 | const hashes = self.fuseHashBatch(hash); 314 | f ^= self.fingerprints[hashes.h0] ^ self.fingerprints[hashes.h1] ^ self.fingerprints[hashes.h2]; 315 | return f == 0; 316 | } 317 | 318 | inline fn fuseHashBatch(self: *const Self, hash: u64) Hashes { 319 | const hi: u64 = mulhi(hash, self.segment_count_length); 320 | var ans: Hashes = undefined; 321 | ans.h0 = @as(u32, @truncate(hi)); 322 | ans.h1 = ans.h0 + self.segment_length; 323 | ans.h2 = ans.h1 + self.segment_length; 324 | ans.h1 ^= @as(u32, @truncate(hash >> 18)) & self.segment_length_mask; 325 | ans.h2 ^= @as(u32, @truncate(hash)) & self.segment_length_mask; 326 | return ans; 327 | } 328 | 329 | inline fn fuseHash(self: *Self, index: usize, hash: u64) u32 { 330 | var h = mulhi(hash, self.segment_count_length); 331 | h +%= index * self.segment_length; 332 | // keep the lower 36 bits 333 | const hh: u64 = hash & ((@as(u64, 1) << 36) - 1); 334 | // index 0: right shift by 36; index 1: right shift by 18; index 2: no shift 335 | // 336 | // NOTE: using u64 here instead of size_it as in upstream C implementation; I think 337 | // that size_t may be incorrect for 32-bit platforms? 338 | const shift_count = (36 - 18 * index); 339 | if (shift_count >= 63) { 340 | h ^= 0 & self.segment_length_mask; 341 | } else { 342 | h ^= (hh >> @as(u6, @truncate(shift_count))) & self.segment_length_mask; 343 | } 344 | return @as(u32, @truncate(h)); 345 | } 346 | }; 347 | } 348 | 349 | inline fn mulhi(a: u64, b: u64) u64 { 350 | return @as(u64, @truncate((@as(u128, @intCast(a)) *% @as(u128, @intCast(b))) >> 64)); 351 | } 352 | 353 | const Hashes = struct { 354 | h0: u32, 355 | h1: u32, 356 | h2: u32, 357 | }; 358 | 359 | inline fn calculateSegmentLength(arity: u32, size: usize) u32 { 360 | // These parameters are very sensitive. Replacing `floor` by `round` can substantially affect 361 | // the construction time. 362 | if (size == 0) return 4; 363 | if (arity == 3) { 364 | const shift_count = @as(u32, @truncate(relaxedFloatToInt(usize, @floor(math.log(f64, math.e, @as(f64, @floatFromInt(size))) / math.log(f64, math.e, 3.33) + 2.25)))); 365 | return if (shift_count >= 31) 0 else @as(u32, 1) << @as(u5, @truncate(shift_count)); 366 | } else if (arity == 4) { 367 | const shift_count = @as(u32, @truncate(relaxedFloatToInt(usize, @floor(math.log(f64, math.e, @as(f64, @floatFromInt(size))) / math.log(f64, math.e, 2.91) - 0.5)))); 368 | return if (shift_count >= 31) 0 else @as(u32, 1) << @as(u5, @truncate(shift_count)); 369 | } 370 | return 65536; 371 | } 372 | 373 | inline fn relaxedFloatToInt(comptime DestType: type, float: anytype) DestType { 374 | if (math.isInf(float) or math.isNegativeInf(float) or math.isNan(float)) { 375 | return 1 << @bitSizeOf(DestType) - 1; 376 | } 377 | return @as(DestType, @intFromFloat(float)); 378 | } 379 | 380 | inline fn max(a: f64, b: f64) f64 { 381 | return if (a < b) b else a; 382 | } 383 | 384 | inline fn calculateSizeFactor(arity: u32, size: usize) f64 { 385 | if (arity == 3) { 386 | return max(1.125, 0.875 + 0.25 * math.log(f64, math.e, 1000000.0) / math.log(f64, math.e, @as(f64, @floatFromInt(size)))); 387 | } else if (arity == 4) { 388 | return max(1.075, 0.77 + 0.305 * math.log(f64, math.e, 600000.0) / math.log(f64, math.e, @as(f64, @floatFromInt(size)))); 389 | } 390 | return 2.0; 391 | } 392 | 393 | inline fn fuseMod3(comptime T: type, x: T) T { 394 | return if (x > 2) x - 3 else x; 395 | } 396 | 397 | const special_size_duplicates = 1337; 398 | 399 | fn binaryFuseTest(T: anytype, size: usize, size_in_bytes: usize) !void { 400 | const allocator = std.heap.page_allocator; 401 | var filter = try BinaryFuse(T).init(allocator, size); 402 | defer filter.deinit(allocator); 403 | 404 | var keys: []u64 = undefined; 405 | if (size == special_size_duplicates) { 406 | const duplicate_keys: [6]u64 = .{ 303, 1, 77, 31, 241, 303 }; 407 | keys = try allocator.alloc(u64, duplicate_keys.len); 408 | for (keys, 0..) |_, i| { 409 | keys[i] = duplicate_keys[i]; 410 | } 411 | } else { 412 | keys = try allocator.alloc(u64, size); 413 | for (keys, 0..) |_, i| { 414 | keys[i] = i; 415 | } 416 | } 417 | defer allocator.free(keys); 418 | 419 | try filter.populate(allocator, keys[0..]); 420 | 421 | if (size != special_size_duplicates) { 422 | if (size == 0) { 423 | try testing.expect(!filter.contain(0)); 424 | try testing.expect(!filter.contain(1)); 425 | } 426 | if (size > 0) try testing.expect(filter.contain(0)); 427 | if (size > 1) try testing.expect(filter.contain(1)); 428 | if (size > 9) { 429 | try testing.expect(filter.contain(1) == true); 430 | try testing.expect(filter.contain(5) == true); 431 | try testing.expect(filter.contain(9) == true); 432 | } 433 | if (size > 1234) try testing.expect(filter.contain(1234) == true); 434 | } 435 | try testing.expectEqual(@as(usize, size_in_bytes), filter.sizeInBytes()); 436 | 437 | for (keys) |key| { 438 | try testing.expect(filter.contain(key) == true); 439 | } 440 | 441 | var random_matches: u64 = 0; 442 | const trials = 10000000; 443 | var i: u64 = 0; 444 | var rng = std.Random.DefaultPrng.init(0); 445 | const random = rng.random(); 446 | while (i < trials) : (i += 1) { 447 | const random_key: u64 = random.uintAtMost(u64, std.math.maxInt(u64)); 448 | if (filter.contain(random_key)) { 449 | if (random_key >= keys.len) { 450 | random_matches += 1; 451 | } 452 | } 453 | } 454 | 455 | // std.debug.print("fpp {d:3.10} (estimated) \n", .{@floatFromInt(f64, random_matches) * 1.0 / trials}); 456 | // std.debug.print("bits per entry {d:3.1}\n", .{@floatFromInt(f64, filter.sizeInBytes()) * 8.0 / @floatFromInt(f64, size)}); 457 | } 458 | 459 | test "binaryFuse8_small_input_edge_cases" { 460 | // See https://github.com/FastFilter/xor_singleheader/issues/26 461 | try binaryFuseTest(u8, 0, 52); 462 | try binaryFuseTest(u8, 1, 52); 463 | try binaryFuseTest(u8, 2, 52); 464 | try binaryFuseTest(u8, 3, 64); 465 | } 466 | 467 | test "binaryFuse8_zero" { 468 | try binaryFuseTest(u8, 0, 52); 469 | } 470 | 471 | test "binaryFuse8_1" { 472 | try binaryFuseTest(u8, 1, 52); 473 | } 474 | 475 | test "binaryFuse8_10" { 476 | try binaryFuseTest(u8, 10, 88); 477 | } 478 | 479 | test "binaryFuse8" { 480 | try binaryFuseTest(u8, 1_000_000, 1130536); 481 | } 482 | 483 | test "binaryFuse8_2m" { 484 | try binaryFuseTest(u8, 2_000_000, 2261032); 485 | } 486 | 487 | test "binaryFuse8_5m" { 488 | try binaryFuseTest(u8, 5_000_000, 5636136); 489 | } 490 | 491 | test "binaryFuse16" { 492 | try binaryFuseTest(u16, 1_000_000, 2261032); 493 | } 494 | 495 | test "binaryFuse32" { 496 | try binaryFuseTest(u32, 1_000_000, 4522024); 497 | } 498 | 499 | test "binaryFuse8_duplicate_keys" { 500 | try binaryFuseTest(u8, special_size_duplicates, 2088); 501 | } 502 | 503 | test "binaryFuse8_mid_num_keys" { 504 | try binaryFuseTest(u8, 11500, 14376); 505 | } 506 | -------------------------------------------------------------------------------- /src/fusefilter.zig: -------------------------------------------------------------------------------- 1 | const std = @import("std"); 2 | const util = @import("util.zig"); 3 | const Allocator = std.mem.Allocator; 4 | const testing = std.testing; 5 | const Error = util.Error; 6 | 7 | const FUSE_ARITY = 3; 8 | const FUSE_SEGMENT_COUNT = 100; 9 | const FUSE_SLOTS = FUSE_SEGMENT_COUNT + FUSE_ARITY - 1; 10 | 11 | /// Fuse8 provides a fuse filter with 8-bit fingerprints. 12 | /// 13 | /// See `Fuse` for more details. 14 | pub const Fuse8 = Fuse(u8); 15 | 16 | /// DEPRECATED: Consider using binary fuse filters instead, they are less prone to creation failure 17 | /// (i.e. the algorithm works with small sets) and are generally all around better. 18 | /// 19 | /// Dietzfelbinger & Walzer's fuse filters, described in "Dense Peelable Random Uniform Hypergraphs", 20 | /// https://arxiv.org/abs/1907.04749, can accomodate fill factors up to 87.9% full, rather than 21 | /// 1 / 1.23 = 81.3%. In the 8-bit case, this reduces the memory usage from 9.84 bits per entry to 22 | /// 9.1 bits. 23 | /// 24 | /// We assume that you have a large set of 64-bit integers and you want a data structure to do 25 | /// membership tests using no more than ~8 or ~16 bits per key. If your initial set is made of 26 | /// strings or other types, you first need to hash them to a 64-bit integer. 27 | pub fn Fuse(comptime T: type) type { 28 | return struct { 29 | seed: u64, 30 | segmentLength: u64, // == slotCount / FUSE_SLOTS 31 | fingerprints: []T, // has room for 3*segmentLength values 32 | 33 | /// probability of success should always be > 0.5 so 100 iterations is highly unlikely 34 | max_iterations: usize = 100, 35 | 36 | const Self = @This(); 37 | 38 | /// initializes a fuse filter with enough capacity for a set containing up to `size` elements. 39 | /// 40 | /// `deinit()` must be called by the caller to free the memory. 41 | pub fn init(allocator: Allocator, size: usize) !Self { 42 | var capacity = @as(usize, @intFromFloat((1.0 / 0.879) * @as(f64, @floatFromInt(size)))); 43 | capacity = capacity / FUSE_SLOTS * FUSE_SLOTS; 44 | return Self{ 45 | .seed = 0, 46 | .fingerprints = try allocator.alloc(T, capacity), 47 | .segmentLength = capacity / FUSE_SLOTS, 48 | }; 49 | } 50 | 51 | pub inline fn deinit(self: *const Self, allocator: Allocator) void { 52 | allocator.free(self.fingerprints); 53 | } 54 | 55 | /// reports if the specified key is within the set with false-positive rate. 56 | pub inline fn contain(self: *const Self, key: u64) bool { 57 | const hash = util.mixSplit(key, self.seed); 58 | const f = @as(T, @truncate(util.fingerprint(hash))); 59 | const r0 = @as(u32, @truncate(hash)); 60 | const r1 = @as(u32, @truncate(util.rotl64(hash, 21))); 61 | const r2 = @as(u32, @truncate(util.rotl64(hash, 42))); 62 | const r3 = @as(u32, @truncate((0xBF58476D1CE4E5B9 *% hash) >> 32)); 63 | const seg = util.reduce(r0, FUSE_SEGMENT_COUNT); 64 | const sl: u64 = self.segmentLength; 65 | const h0 = (seg + 0) * sl + util.reduce(r1, @as(u32, @truncate(sl))); 66 | const h1 = (seg + 1) * sl + util.reduce(r2, @as(u32, @truncate(sl))); 67 | const h2 = (seg + 2) * sl + util.reduce(r3, @as(u32, @truncate(sl))); 68 | return f == (self.fingerprints[h0] ^ self.fingerprints[h1] ^ self.fingerprints[h2]); 69 | } 70 | 71 | /// reports the size in bytes of the filter. 72 | pub inline fn sizeInBytes(self: *const Self) usize { 73 | return FUSE_SLOTS * self.segmentLength * @sizeOf(T) + @sizeOf(Self); 74 | } 75 | 76 | /// populates the filter with the given keys. 77 | /// 78 | /// The caller is responsible for ensuring that there are no duplicated keys. 79 | /// 80 | /// The inner loop will run up to max_iterations times (default 100) and will never fail, 81 | /// except if there are duplicated keys. 82 | /// 83 | /// The provided allocator will be used for creating temporary buffers that do not outlive the 84 | /// function call. 85 | pub fn populate(self: *Self, allocator: Allocator, keys: []u64) Error!void { 86 | var iter = util.SliceIterator(u64).init(keys); 87 | return self.populateIter(allocator, &iter); 88 | } 89 | 90 | /// Identical to populate, except it takes an iterator of keys so you need not store them 91 | /// in-memory. 92 | /// 93 | /// `keys.next()` must return `?u64`, the next key or none if the end of the list has been 94 | /// reached. The iterator must reset after hitting the end of the list, such that the `next()` 95 | /// call leads to the first element again. 96 | /// 97 | /// `keys.len()` must return the `usize` length. 98 | pub fn populateIter(self: *Self, allocator: Allocator, keys: anytype) Error!void { 99 | var rng_counter: u64 = 1; 100 | self.seed = util.rngSplitMix64(&rng_counter); 101 | 102 | var sets = try allocator.alloc(Set, self.segmentLength * FUSE_SLOTS); 103 | defer allocator.free(sets); 104 | 105 | var Q = try allocator.alloc(Keyindex, sets.len); 106 | defer allocator.free(Q); 107 | 108 | var stack = try allocator.alloc(Keyindex, keys.len()); 109 | defer allocator.free(stack); 110 | 111 | var loop: usize = 0; 112 | while (true) : (loop += 1) { 113 | if (loop + 1 > self.max_iterations) { 114 | return Error.KeysLikelyNotUnique; // too many iterations, keys are not unique. 115 | } 116 | for (sets[0..sets.len]) |*b| b.* = std.mem.zeroes(Set); 117 | 118 | while (keys.next()) |key| { 119 | const hs = getH0H1H2(self, key); 120 | sets[hs.h0].fusemask ^= hs.h; 121 | sets[hs.h0].count += 1; 122 | sets[hs.h1].fusemask ^= hs.h; 123 | sets[hs.h1].count += 1; 124 | sets[hs.h2].fusemask ^= hs.h; 125 | sets[hs.h2].count += 1; 126 | } 127 | 128 | // TODO(upstream): the flush should be sync with the detection that follows scan values 129 | // with a count of one. 130 | var Qsize: usize = 0; 131 | for (sets, 0..) |set, i| { 132 | if (set.count == 1) { 133 | Q[Qsize].index = @as(u32, @intCast(i)); 134 | Q[Qsize].hash = sets[i].fusemask; 135 | Qsize += 1; 136 | } 137 | } 138 | 139 | var stack_size: usize = 0; 140 | while (Qsize > 0) { 141 | Qsize -= 1; 142 | const keyindex = Q[Qsize]; 143 | const index = keyindex.index; 144 | if (sets[index].count == 0) { 145 | continue; // not actually possible after the initial scan. 146 | } 147 | const hash = keyindex.hash; 148 | const hs = getJustH0H1H2(self, hash); 149 | 150 | stack[stack_size] = keyindex; 151 | stack_size += 1; 152 | 153 | sets[hs.h0].fusemask ^= hash; 154 | sets[hs.h0].count -= 1; 155 | if (sets[hs.h0].count == 1) { 156 | Q[Qsize].index = hs.h0; 157 | Q[Qsize].hash = sets[hs.h0].fusemask; 158 | Qsize += 1; 159 | } 160 | 161 | sets[hs.h1].fusemask ^= hash; 162 | sets[hs.h1].count -= 1; 163 | if (sets[hs.h1].count == 1) { 164 | Q[Qsize].index = hs.h1; 165 | Q[Qsize].hash = sets[hs.h1].fusemask; 166 | Qsize += 1; 167 | } 168 | 169 | sets[hs.h2].fusemask ^= hash; 170 | sets[hs.h2].count -= 1; 171 | if (sets[hs.h2].count == 1) { 172 | Q[Qsize].index = hs.h2; 173 | Q[Qsize].hash = sets[hs.h2].fusemask; 174 | Qsize += 1; 175 | } 176 | } 177 | if (stack_size == keys.len()) { 178 | // success 179 | break; 180 | } 181 | self.seed = util.rngSplitMix64(&rng_counter); 182 | } 183 | 184 | var stack_size = keys.len(); 185 | while (stack_size > 0) { 186 | stack_size -= 1; 187 | const ki = stack[stack_size]; 188 | const hs = getJustH0H1H2(self, ki.hash); 189 | var hsh: T = @as(T, @truncate(util.fingerprint(ki.hash))); 190 | if (ki.index == hs.h0) { 191 | hsh ^= self.fingerprints[hs.h1] ^ self.fingerprints[hs.h2]; 192 | } else if (ki.index == hs.h1) { 193 | hsh ^= self.fingerprints[hs.h0] ^ self.fingerprints[hs.h2]; 194 | } else { 195 | hsh ^= self.fingerprints[hs.h0] ^ self.fingerprints[hs.h1]; 196 | } 197 | self.fingerprints[ki.index] = hsh; 198 | } 199 | return; 200 | } 201 | 202 | inline fn getH0H1H2(self: *Self, k: u64) Hashes { 203 | const hash = util.mixSplit(k, self.seed); 204 | const r0 = @as(u32, @truncate(hash)); 205 | const r1 = @as(u32, @truncate(util.rotl64(hash, 21))); 206 | const r2 = @as(u32, @truncate(util.rotl64(hash, 42))); 207 | const r3 = @as(u32, @truncate((0xBF58476D1CE4E5B9 *% hash) >> 32)); 208 | const seg = util.reduce(r0, FUSE_SEGMENT_COUNT); 209 | const sl = self.segmentLength; 210 | return Hashes{ 211 | .h = hash, 212 | .h0 = @as(u32, @truncate(@as(u64, @intCast((seg + 0))) * sl + @as(u64, @intCast(util.reduce(r1, @as(u32, @truncate(sl))))))), 213 | .h1 = @as(u32, @truncate(@as(u64, @intCast((seg + 1))) * sl + @as(u64, @intCast(util.reduce(r2, @as(u32, @truncate(sl))))))), 214 | .h2 = @as(u32, @truncate(@as(u64, @intCast((seg + 2))) * sl + @as(u64, @intCast(util.reduce(r3, @as(u32, @truncate(sl))))))), 215 | }; 216 | } 217 | 218 | inline fn getJustH0H1H2(self: *Self, hash: u64) H0h1h2 { 219 | const r0 = @as(u32, @truncate(hash)); 220 | const r1 = @as(u32, @truncate(util.rotl64(hash, 21))); 221 | const r2 = @as(u32, @truncate(util.rotl64(hash, 42))); 222 | const r3 = @as(u32, @truncate((0xBF58476D1CE4E5B9 *% hash) >> 32)); 223 | const seg = util.reduce(r0, FUSE_SEGMENT_COUNT); 224 | const sl = self.segmentLength; 225 | return H0h1h2{ 226 | .h0 = @as(u32, @truncate(@as(u64, @intCast((seg + 0))) * sl + @as(u64, @intCast(util.reduce(r1, @as(u32, @truncate(sl))))))), 227 | .h1 = @as(u32, @truncate(@as(u64, @intCast((seg + 1))) * sl + @as(u64, @intCast(util.reduce(r2, @as(u32, @truncate(sl))))))), 228 | .h2 = @as(u32, @truncate(@as(u64, @intCast((seg + 2))) * sl + @as(u64, @intCast(util.reduce(r3, @as(u32, @truncate(sl))))))), 229 | }; 230 | } 231 | }; 232 | } 233 | 234 | const Set = struct { 235 | fusemask: u64, 236 | count: u32, 237 | }; 238 | 239 | const Hashes = struct { 240 | h: u64, 241 | h0: u32, 242 | h1: u32, 243 | h2: u32, 244 | }; 245 | 246 | const H0h1h2 = struct { 247 | h0: u32, 248 | h1: u32, 249 | h2: u32, 250 | }; 251 | 252 | const Keyindex = struct { 253 | hash: u64, 254 | index: u32, 255 | }; 256 | 257 | fn fuseTest(T: anytype, size: usize, size_in_bytes: usize) !void { 258 | const allocator = std.heap.page_allocator; 259 | var filter = try Fuse(T).init(allocator, size); 260 | defer filter.deinit(allocator); 261 | 262 | var keys = try allocator.alloc(u64, size); 263 | defer allocator.free(keys); 264 | for (keys, 0..) |_, i| { 265 | keys[i] = i; 266 | } 267 | 268 | try filter.populate(allocator, keys[0..]); 269 | 270 | try testing.expect(filter.contain(1) == true); 271 | try testing.expect(filter.contain(5) == true); 272 | try testing.expect(filter.contain(9) == true); 273 | try testing.expect(filter.contain(1234) == true); 274 | try testing.expectEqual(@as(usize, size_in_bytes), filter.sizeInBytes()); 275 | 276 | for (keys) |key| { 277 | try testing.expect(filter.contain(key) == true); 278 | } 279 | 280 | var random_matches: u64 = 0; 281 | const trials = 10000000; 282 | var i: u64 = 0; 283 | var rng = std.Random.DefaultPrng.init(0); 284 | const random = rng.random(); 285 | while (i < trials) : (i += 1) { 286 | const random_key: u64 = random.uintAtMost(u64, std.math.maxInt(u64)); 287 | if (filter.contain(random_key)) { 288 | if (random_key >= keys.len) { 289 | random_matches += 1; 290 | } 291 | } 292 | } 293 | 294 | // std.debug.print("fpp {d:3.10} (estimated) \n", .{@floatFromInt(f64, random_matches) * 1.0 / trials}); 295 | // std.debug.print("bits per entry {d:3.1}\n", .{@floatFromInt(f64, filter.sizeInBytes()) * 8.0 / @floatFromInt(f64, size)}); 296 | } 297 | 298 | test "fuse4" { 299 | try fuseTest(u4, 1000000 / 2, 568792); 300 | } 301 | 302 | test "fuse8" { 303 | try fuseTest(u8, 1000000, 1137646); 304 | } 305 | 306 | test "fuse16" { 307 | try fuseTest(u16, 1000000, 2275252); 308 | } 309 | -------------------------------------------------------------------------------- /src/main.zig: -------------------------------------------------------------------------------- 1 | const std = @import("std"); 2 | const testing = std.testing; 3 | 4 | pub const BinaryFuse = @import("binaryfusefilter.zig").BinaryFuse; 5 | pub const BinaryFuse8 = @import("binaryfusefilter.zig").BinaryFuse8; 6 | pub const Xor = @import("xorfilter.zig").Xor; 7 | pub const Xor8 = @import("xorfilter.zig").Xor8; 8 | pub const Xor16 = @import("xorfilter.zig").Xor16; 9 | pub const AutoUnique = @import("unique.zig").AutoUnique; 10 | pub const Unique = @import("unique.zig").Unique; 11 | pub const Error = @import("util.zig").Error; 12 | pub const SliceIterator = @import("util.zig").SliceIterator; 13 | pub const MeasuredAllocator = @import("MeasuredAllocator.zig"); 14 | 15 | test "exports" { 16 | const allocator = std.heap.page_allocator; 17 | 18 | var binaryFuse8Filter = try BinaryFuse8.init(allocator, 100); 19 | defer binaryFuse8Filter.deinit(allocator); 20 | 21 | var xorFilter = try Xor(u8).init(allocator, 1); 22 | defer xorFilter.deinit(allocator); 23 | 24 | var xor8Filter = try Xor8.init(allocator, 1); 25 | defer xor8Filter.deinit(allocator); 26 | 27 | var xor16Filter = try Xor16.init(allocator, 1); 28 | defer xor16Filter.deinit(allocator); 29 | 30 | var array = [_]i32{ 1, 2, 2 }; 31 | _ = AutoUnique(i32, void)({}, array[0..]); 32 | 33 | _ = SliceIterator; 34 | _ = MeasuredAllocator; 35 | 36 | testing.refAllDecls(@import("fusefilter.zig")); 37 | testing.refAllDecls(@import("benchmark.zig")); 38 | } 39 | -------------------------------------------------------------------------------- /src/unique.zig: -------------------------------------------------------------------------------- 1 | const std = @import("std"); 2 | const testing = std.testing; 3 | 4 | // The same as `AutoUnique`, but accepts custom hash and equality functions. 5 | pub fn Unique( 6 | comptime T: type, 7 | comptime Context: type, 8 | comptime hash: fn ( 9 | ctx: Context, 10 | key: T, 11 | ) u64, 12 | comptime eql: fn (ctx: Context, a: T, b: T) bool, 13 | ) fn (Context, []T) []T { 14 | return struct { 15 | pub fn inPlace(ctx: Context, data: []T) []T { 16 | return doInPlace(ctx, data, 0); 17 | } 18 | 19 | fn swap(data: []T, i: usize, j: usize) void { 20 | const tmp = data[i]; 21 | data[i] = data[j]; 22 | data[j] = tmp; 23 | } 24 | 25 | fn doInPlace(ctx: Context, dataIn: []T, start: usize) []T { 26 | if (dataIn.len - start < 2) { 27 | return dataIn; 28 | } 29 | 30 | const sentinel = dataIn[start]; 31 | const data = dataIn[start + 1 .. dataIn.len]; 32 | 33 | var index: usize = 0; 34 | while (index < data.len) { 35 | if (eql(ctx, data[index], sentinel)) { 36 | index += 1; 37 | continue; 38 | } 39 | 40 | const hsh = hash(ctx, data[index]) % data.len; 41 | if (index == hsh) { 42 | index += 1; 43 | continue; 44 | } 45 | 46 | if (eql(ctx, data[index], data[hsh])) { 47 | data[index] = sentinel; 48 | index += 1; 49 | continue; 50 | } 51 | 52 | if (eql(ctx, data[hsh], sentinel)) { 53 | swap(data, hsh, index); 54 | index += 1; 55 | continue; 56 | } 57 | 58 | const hashHash = hash(ctx, data[hsh]) % data.len; 59 | if (hashHash != hsh) { 60 | swap(data, index, hsh); 61 | if (hsh < index) { 62 | index += 1; 63 | } 64 | } else { 65 | index += 1; 66 | } 67 | } 68 | 69 | var swapPos: usize = 0; 70 | for (data, 0..) |v, i| { 71 | if ((!eql(ctx, v, sentinel)) and (i == (hash(ctx, v) % data.len))) { 72 | swap(data, i, swapPos); 73 | swapPos += 1; 74 | } 75 | } 76 | 77 | var sentinelPos: usize = data.len; 78 | var i = swapPos; 79 | while (i < sentinelPos) { 80 | if (eql(ctx, data[i], sentinel)) { 81 | sentinelPos -= 1; 82 | swap(data, i, sentinelPos); 83 | } else { 84 | i += 1; 85 | } 86 | } 87 | 88 | return doInPlace(ctx, dataIn[0 .. sentinelPos + start + 1], start + swapPos + 1); 89 | } 90 | }.inPlace; 91 | } 92 | 93 | /// Modifies a slice in-place to contain only unique values. 94 | /// 95 | /// The input slice is modified, with element positions being swapped. The returned slice is an 96 | /// equal length or subset with duplicate entries omitted. 97 | /// 98 | /// It is guaranteed to be O(1) in axillary space, and is typically O(N) time complexity. 99 | /// Order preservation is not guaranteed. 100 | /// 101 | /// The algorithm is described here: https://stackoverflow.com/a/1533667 102 | /// 103 | /// To hash T and perform equality checks of T, `std.hash_map.getAutoHashFn` and 104 | /// `std.hash_map.getAutoEqlFn` are used, which support most data types. Use `Unique` if you need 105 | /// to use your own hash and equality functions. 106 | pub fn AutoUnique(comptime T: type, comptime Context: type) fn (void, []T) []T { 107 | return comptime Unique( 108 | T, 109 | Context, 110 | std.hash_map.getAutoHashFn(T, Context), 111 | std.hash_map.getAutoEqlFn(T, Context), 112 | ); 113 | } 114 | 115 | test "AutoUnique_simple" { 116 | var array = [_]i32{ 1, 2, 2, 3, 3, 4, 2, 1, 4, 1, 2, 3, 4, 4, 3, 2, 1 }; 117 | const unique = AutoUnique(i32, void)({}, array[0..]); 118 | const expected = &[_]i32{ 1, 2, 3, 4 }; 119 | try testing.expectEqualSlices(i32, expected, unique); 120 | } 121 | 122 | test "AutoUnique_complex" { 123 | // Produce an array with 3x duplicated keys 124 | const allocator = std.heap.page_allocator; 125 | const size = 10; 126 | var keys = try allocator.alloc(u64, size * 3); 127 | defer allocator.free(keys); 128 | 129 | var i: usize = 0; 130 | while (i < size * 3) : (i += 3) { 131 | keys[i] = i % size; 132 | keys[i + 1] = i % size; 133 | keys[i + 2] = i % size; 134 | } 135 | 136 | const unique = AutoUnique(u64, void)({}, keys[0..]); 137 | try testing.expect(unique.len == size); 138 | } 139 | -------------------------------------------------------------------------------- /src/util.zig: -------------------------------------------------------------------------------- 1 | const std = @import("std"); 2 | const testing = std.testing; 3 | const Allocator = std.mem.Allocator; 4 | 5 | pub const Error = error{ 6 | KeysLikelyNotUnique, 7 | OutOfMemory, 8 | }; 9 | 10 | pub inline fn murmur64(h: u64) u64 { 11 | var v = h; 12 | v ^= v >> 33; 13 | v *%= 0xff51afd7ed558ccd; 14 | v ^= v >> 33; 15 | v *%= 0xc4ceb9fe1a85ec53; 16 | v ^= v >> 33; 17 | return v; 18 | } 19 | 20 | pub inline fn mixSplit(key: u64, seed: u64) u64 { 21 | return murmur64(key +% seed); 22 | } 23 | 24 | pub inline fn rotl64(n: u64, c: usize) u64 { 25 | return (n << @as(u6, @intCast(c & 63))) | (n >> @as(u6, @intCast((-%c) & 63))); 26 | } 27 | 28 | pub inline fn reduce(hash: u32, n: u32) u32 { 29 | // http://lemire.me/blog/2016/06/27/a-fast-alternative-to-the-modulo-reduction 30 | return @as(u32, @truncate((@as(u64, @intCast(hash)) *% @as(u64, @intCast(n))) >> 32)); 31 | } 32 | 33 | pub inline fn fingerprint(hash: u64) u64 { 34 | return hash ^ (hash >> 32); 35 | } 36 | 37 | pub fn SliceIterator(comptime T: type) type { 38 | return struct { 39 | slice: []const T, 40 | i: usize = 0, 41 | 42 | const Self = @This(); 43 | 44 | pub inline fn init(slice: []const T) Self { 45 | return .{ .slice = slice }; 46 | } 47 | 48 | pub inline fn next(self: *Self) ?T { 49 | if (self.i >= self.slice.len) { 50 | self.i = 0; 51 | return null; 52 | } 53 | const v = self.slice[self.i]; 54 | self.i += 1; 55 | return v; 56 | } 57 | 58 | pub inline fn len(self: *Self) usize { 59 | return self.slice.len; 60 | } 61 | }; 62 | } 63 | 64 | // returns random number, modifies the seed. 65 | pub inline fn rngSplitMix64(seed: *u64) u64 { 66 | seed.* = seed.* +% 0x9E3779B97F4A7C15; 67 | var z = seed.*; 68 | z = (z ^ (z >> 30)) *% 0xBF58476D1CE4E5B9; 69 | z = (z ^ (z >> 27)) *% 0x94D049BB133111EB; 70 | return z ^ (z >> 31); 71 | } 72 | 73 | test "murmur64" { 74 | // Arbitrarily chosen inputs for validation. 75 | try testing.expectEqual(@as(u64, 11156705658460211942), murmur64(0xF + 5)); 76 | try testing.expectEqual(@as(u64, 9276143743022464963), murmur64(0xFF + 123)); 77 | try testing.expectEqual(@as(u64, 9951468085874665196), murmur64(0xFFF + 1337)); 78 | try testing.expectEqual(@as(u64, 4797998644499646477), murmur64(0xFFFF + 143)); 79 | try testing.expectEqual(@as(u64, 4139256335519489731), murmur64(0xFFFFFF + 918273987)); 80 | } 81 | 82 | test "mixSplit" { 83 | // Arbitrarily chosen inputs for validation. 84 | try testing.expectEqual(@as(u64, 11156705658460211942), mixSplit(0xF, 5)); 85 | } 86 | 87 | test "rotl64" { 88 | // Arbitrarily chosen inputs for validation. 89 | try testing.expectEqual(@as(u64, 193654783976931328), rotl64(43, 52)); 90 | } 91 | 92 | test "reduce" { 93 | // Arbitrarily chosen inputs for validation. 94 | try testing.expectEqual(@as(u32, 8752776), reduce(1936547838, 19412321)); 95 | } 96 | 97 | test "fingerprint" { 98 | // Arbitrarily chosen inputs for validation. 99 | try testing.expectEqual(@as(u64, 1936547838), fingerprint(1936547838)); 100 | } 101 | 102 | test "rngSplitMix64" { 103 | var seed: u64 = 13337; 104 | var r = rngSplitMix64(&seed); 105 | try testing.expectEqual(@as(u64, 8862613829200693549), r); 106 | r = rngSplitMix64(&seed); 107 | try testing.expectEqual(@as(u64, 1009918040199880802), r); 108 | r = rngSplitMix64(&seed); 109 | try testing.expectEqual(@as(u64, 8603670078971061766), r); 110 | } 111 | -------------------------------------------------------------------------------- /src/xorfilter.zig: -------------------------------------------------------------------------------- 1 | const std = @import("std"); 2 | const util = @import("util.zig"); 3 | const Allocator = std.mem.Allocator; 4 | const testing = std.testing; 5 | const Error = util.Error; 6 | 7 | /// Xor8 is the recommended default, no more than a 0.3% false-positive probability. 8 | /// 9 | /// See `Xor` for more details. 10 | pub const Xor8 = Xor(u8); 11 | 12 | /// Xor16 provides a xor filter with 16-bit fingerprints. 13 | /// 14 | /// See `Xor` for more details. 15 | pub const Xor16 = Xor(u16); 16 | 17 | /// Xor returns a xor filter with the specified base type, usually u8 or u16, for which the 18 | /// helpers Xor8 and Xor16 may be used. 19 | /// 20 | /// We assume that you have a large set of 64-bit integers and you want a data structure to do 21 | /// membership tests using no more than ~8 or ~16 bits per key. If your initial set is made of 22 | /// strings or other types, you first need to hash them to a 64-bit integer. 23 | /// 24 | /// Xor8 is the recommended default, no more than a 0.3% false-positive probability. 25 | pub fn Xor(comptime T: type) type { 26 | return struct { 27 | seed: u64, 28 | blockLength: u64, 29 | fingerprints: []T, // has room for 3*blockLength values 30 | 31 | /// probability of success should always be > 0.5 so 100 iterations is highly unlikely 32 | max_iterations: usize = 100, 33 | 34 | const Self = @This(); 35 | 36 | /// initializes a Xor filter with enough capacity for a set containing up to `size` elements. 37 | /// 38 | /// `deinit()` must be called by the caller to free the memory. 39 | pub fn init(allocator: Allocator, size: usize) !Self { 40 | var capacity = @as(usize, @intFromFloat(32 + 1.23 * @as(f64, @floatFromInt(size)))); 41 | capacity = capacity / 3 * 3; 42 | return Self{ 43 | .seed = 0, 44 | .fingerprints = try allocator.alloc(T, capacity), 45 | .blockLength = capacity / 3, 46 | }; 47 | } 48 | 49 | pub inline fn deinit(self: *const Self, allocator: Allocator) void { 50 | allocator.free(self.fingerprints); 51 | } 52 | 53 | /// reports if the specified key is within the set with false-positive rate. 54 | pub inline fn contain(self: *const Self, key: u64) bool { 55 | const hash = util.mixSplit(key, self.seed); 56 | const f = @as(T, @truncate(util.fingerprint(hash))); 57 | const r0 = @as(u32, @truncate(hash)); 58 | const r1 = @as(u32, @truncate(util.rotl64(hash, 21))); 59 | const r2 = @as(u32, @truncate(util.rotl64(hash, 42))); 60 | const bl = @as(u32, @truncate(self.blockLength)); 61 | const h0: u32 = util.reduce(r0, bl); 62 | const h1: u32 = util.reduce(r1, bl) + bl; 63 | const h2: u32 = util.reduce(r2, bl) + 2 * bl; 64 | return f == (self.fingerprints[h0] ^ self.fingerprints[h1] ^ self.fingerprints[h2]); 65 | } 66 | 67 | /// reports the size in bytes of the filter. 68 | pub inline fn sizeInBytes(self: *const Self) usize { 69 | return 3 * self.blockLength * @sizeOf(T) + @sizeOf(Self); 70 | } 71 | 72 | /// populates the filter with the given keys. 73 | /// 74 | /// The caller is responsible for ensuring that there are no duplicated keys. 75 | /// 76 | /// The inner loop will run up to max_iterations (default 100) and should never fail, 77 | /// except if there are duplicated keys. 78 | /// 79 | /// The provided allocator will be used for creating temporary buffers that do not outlive the 80 | /// function call. 81 | pub fn populate(self: *Self, allocator: Allocator, keys: []u64) Error!void { 82 | var iter = util.SliceIterator(u64).init(keys); 83 | return self.populateIter(allocator, &iter); 84 | } 85 | 86 | /// Identical to populate, except it takes an iterator of keys so you need not store them 87 | /// in-memory. 88 | /// 89 | /// `keys.next()` must return `?u64`, the next key or none if the end of the list has been 90 | /// reached. The iterator must reset after hitting the end of the list, such that the `next()` 91 | /// call leads to the first element again. 92 | /// 93 | /// `keys.len()` must return the `usize` length. 94 | pub fn populateIter(self: *Self, allocator: Allocator, keys: anytype) Error!void { 95 | var rng_counter: u64 = 1; 96 | self.seed = util.rngSplitMix64(&rng_counter); 97 | 98 | var sets = try allocator.alloc(Set, self.blockLength * 3); 99 | defer allocator.free(sets); 100 | 101 | var Q = try allocator.alloc(Keyindex, sets.len); 102 | defer allocator.free(Q); 103 | 104 | var stack = try allocator.alloc(Keyindex, keys.len()); 105 | defer allocator.free(stack); 106 | 107 | var sets0 = sets; 108 | var sets1 = sets[self.blockLength..]; 109 | var sets2 = sets[2 * self.blockLength ..]; 110 | var Q0 = Q; 111 | var Q1 = Q[self.blockLength..]; 112 | var Q2 = Q[2 * self.blockLength ..]; 113 | 114 | var loop: usize = 0; 115 | while (true) : (loop += 1) { 116 | if (loop + 1 > self.max_iterations) { 117 | return Error.KeysLikelyNotUnique; // too many iterations, keys are not unique. 118 | } 119 | for (sets[0..sets.len]) |*b| b.* = std.mem.zeroes(Set); 120 | 121 | while (keys.next()) |key| { 122 | const hs = self.getH0H1H2(key); 123 | sets0[hs.h0].xormask ^= hs.h; 124 | sets0[hs.h0].count += 1; 125 | sets1[hs.h1].xormask ^= hs.h; 126 | sets1[hs.h1].count += 1; 127 | sets2[hs.h2].xormask ^= hs.h; 128 | sets2[hs.h2].count += 1; 129 | } 130 | 131 | // TODO(upstream): the flush should be sync with the detection that follows scan values 132 | // with a count of one. 133 | var Q0size: usize = 0; 134 | var Q1size: usize = 0; 135 | var Q2size: usize = 0; 136 | { 137 | var i: usize = 0; 138 | while (i < self.blockLength) : (i += 1) { 139 | if (sets0[i].count == 1) { 140 | Q0[Q0size].index = @as(u32, @intCast(i)); 141 | Q0[Q0size].hash = sets0[i].xormask; 142 | Q0size += 1; 143 | } 144 | } 145 | } 146 | { 147 | var i: usize = 0; 148 | while (i < self.blockLength) : (i += 1) { 149 | if (sets1[i].count == 1) { 150 | Q1[Q1size].index = @as(u32, @intCast(i)); 151 | Q1[Q1size].hash = sets1[i].xormask; 152 | Q1size += 1; 153 | } 154 | } 155 | } 156 | { 157 | var i: usize = 0; 158 | while (i < self.blockLength) : (i += 1) { 159 | if (sets2[i].count == 1) { 160 | Q2[Q2size].index = @as(u32, @intCast(i)); 161 | Q2[Q2size].hash = sets2[i].xormask; 162 | Q2size += 1; 163 | } 164 | } 165 | } 166 | 167 | var stack_size: usize = 0; 168 | while (Q0size + Q1size + Q2size > 0) { 169 | while (Q0size > 0) { 170 | Q0size -%= 1; 171 | const keyindex = Q0[Q0size]; 172 | const index = keyindex.index; 173 | if (sets0[index].count == 0) { 174 | continue; // not actually possible after the initial scan. 175 | } 176 | const hash = keyindex.hash; 177 | const h1 = self.getH1(hash); 178 | const h2 = self.getH2(hash); 179 | 180 | stack[stack_size] = keyindex; 181 | stack_size += 1; 182 | sets1[h1].xormask ^= hash; 183 | sets1[h1].count -%= 1; 184 | if (sets1[h1].count == 1) { 185 | Q1[Q1size].index = h1; 186 | Q1[Q1size].hash = sets1[h1].xormask; 187 | Q1size += 1; 188 | } 189 | sets2[h2].xormask ^= hash; 190 | sets2[h2].count -%= 1; 191 | if (sets2[h2].count == 1) { 192 | Q2[Q2size].index = h2; 193 | Q2[Q2size].hash = sets2[h2].xormask; 194 | Q2size += 1; 195 | } 196 | } 197 | while (Q1size > 0) { 198 | Q1size -%= 1; 199 | var keyindex = Q1[Q1size]; 200 | const index = keyindex.index; 201 | if (sets1[index].count == 0) { 202 | continue; // not actually possible after the initial scan. 203 | } 204 | const hash = keyindex.hash; 205 | const h0 = self.getH0(hash); 206 | const h2 = self.getH2(hash); 207 | keyindex.index += @as(u32, @truncate(self.blockLength)); 208 | 209 | stack[stack_size] = keyindex; 210 | stack_size += 1; 211 | sets0[h0].xormask ^= hash; 212 | sets0[h0].count -%= 1; 213 | if (sets0[h0].count == 1) { 214 | Q0[Q0size].index = h0; 215 | Q0[Q0size].hash = sets0[h0].xormask; 216 | Q0size += 1; 217 | } 218 | sets2[h2].xormask ^= hash; 219 | sets2[h2].count -%= 1; 220 | if (sets2[h2].count == 1) { 221 | Q2[Q2size].index = h2; 222 | Q2[Q2size].hash = sets2[h2].xormask; 223 | Q2size += 1; 224 | } 225 | } 226 | while (Q2size > 0) { 227 | Q2size -%= 1; 228 | var keyindex = Q2[Q2size]; 229 | const index = keyindex.index; 230 | if (sets2[index].count == 0) { 231 | continue; // not actually possible after the initial scan. 232 | } 233 | const hash = keyindex.hash; 234 | const h0 = self.getH0(hash); 235 | const h1 = self.getH1(hash); 236 | keyindex.index += @as(u32, @truncate(2 * @as(u64, @intCast(self.blockLength)))); 237 | 238 | stack[stack_size] = keyindex; 239 | stack_size += 1; 240 | sets0[h0].xormask ^= hash; 241 | sets0[h0].count -%= 1; 242 | if (sets0[h0].count == 1) { 243 | Q0[Q0size].index = h0; 244 | Q0[Q0size].hash = sets0[h0].xormask; 245 | Q0size += 1; 246 | } 247 | sets1[h1].xormask ^= hash; 248 | sets1[h1].count -%= 1; 249 | if (sets1[h1].count == 1) { 250 | Q1[Q1size].index = h1; 251 | Q1[Q1size].hash = sets1[h1].xormask; 252 | Q1size += 1; 253 | } 254 | } 255 | } 256 | if (stack_size == keys.len()) { 257 | // success 258 | break; 259 | } 260 | self.seed = util.rngSplitMix64(&rng_counter); 261 | } 262 | 263 | const fingerprints0: []T = self.fingerprints; 264 | const fingerprints1: []T = self.fingerprints[self.blockLength..]; 265 | const fingerprints2: []T = self.fingerprints[2 * self.blockLength ..]; 266 | 267 | var stack_size = keys.len(); 268 | while (stack_size > 0) { 269 | stack_size -= 1; 270 | const ki = stack[stack_size]; 271 | var val: u64 = util.fingerprint(ki.hash); 272 | if (ki.index < @as(u32, @truncate(self.blockLength))) { 273 | val ^= fingerprints1[self.getH1(ki.hash)] ^ fingerprints2[self.getH2(ki.hash)]; 274 | } else if (ki.index < 2 * @as(u32, @truncate(self.blockLength))) { 275 | val ^= fingerprints0[self.getH0(ki.hash)] ^ fingerprints2[self.getH2(ki.hash)]; 276 | } else { 277 | val ^= fingerprints0[self.getH0(ki.hash)] ^ fingerprints1[self.getH1(ki.hash)]; 278 | } 279 | self.fingerprints[ki.index] = @as(T, @truncate(val)); 280 | } 281 | return; 282 | } 283 | 284 | inline fn getH0H1H2(self: *Self, k: u64) Hashes { 285 | const hash = util.mixSplit(k, self.seed); 286 | const r0 = @as(u32, @truncate(hash)); 287 | const r1 = @as(u32, @truncate(util.rotl64(hash, 21))); 288 | const r2 = @as(u32, @truncate(util.rotl64(hash, 42))); 289 | return Hashes{ 290 | .h = hash, 291 | .h0 = util.reduce(r0, @as(u32, @truncate(self.blockLength))), 292 | .h1 = util.reduce(r1, @as(u32, @truncate(self.blockLength))), 293 | .h2 = util.reduce(r2, @as(u32, @truncate(self.blockLength))), 294 | }; 295 | } 296 | 297 | inline fn getH0(self: *Self, hash: u64) u32 { 298 | const r0 = @as(u32, @truncate(hash)); 299 | return util.reduce(r0, @as(u32, @truncate(self.blockLength))); 300 | } 301 | 302 | inline fn getH1(self: *Self, hash: u64) u32 { 303 | const r1 = @as(u32, @truncate(util.rotl64(hash, 21))); 304 | return util.reduce(r1, @as(u32, @truncate(self.blockLength))); 305 | } 306 | 307 | inline fn getH2(self: *Self, hash: u64) u32 { 308 | const r2 = @as(u32, @truncate(util.rotl64(hash, 42))); 309 | return util.reduce(r2, @as(u32, @truncate(self.blockLength))); 310 | } 311 | }; 312 | } 313 | 314 | const Set = struct { 315 | xormask: u64, 316 | count: u32, 317 | }; 318 | 319 | const Hashes = struct { 320 | h: u64, 321 | h0: u32, 322 | h1: u32, 323 | h2: u32, 324 | }; 325 | 326 | const H0h1h2 = struct { 327 | h0: u32, 328 | h1: u32, 329 | h2: u32, 330 | }; 331 | 332 | const Keyindex = struct { 333 | hash: u64, 334 | index: u32, 335 | }; 336 | 337 | fn xorTest(T: anytype, size: usize, size_in_bytes: usize) !void { 338 | const allocator = std.heap.page_allocator; 339 | var filter = try Xor(T).init(allocator, size); 340 | defer filter.deinit(allocator); 341 | 342 | var keys = try allocator.alloc(u64, size); 343 | defer allocator.free(keys); 344 | for (keys, 0..) |_, i| { 345 | keys[i] = i; 346 | } 347 | 348 | try filter.populate(allocator, keys[0..]); 349 | 350 | try testing.expect(filter.contain(1) == true); 351 | try testing.expect(filter.contain(5) == true); 352 | try testing.expect(filter.contain(9) == true); 353 | try testing.expect(filter.contain(1234) == true); 354 | try testing.expectEqual(@as(usize, size_in_bytes), filter.sizeInBytes()); 355 | 356 | for (keys) |key| { 357 | try testing.expect(filter.contain(key) == true); 358 | } 359 | 360 | var random_matches: u64 = 0; 361 | const trials = 10000000; 362 | var i: u64 = 0; 363 | var rng = std.Random.DefaultPrng.init(0); 364 | const random = rng.random(); 365 | while (i < trials) : (i += 1) { 366 | const random_key: u64 = random.uintAtMost(u64, std.math.maxInt(u64)); 367 | if (filter.contain(random_key)) { 368 | if (random_key >= keys.len) { 369 | random_matches += 1; 370 | } 371 | } 372 | } 373 | 374 | // const fpp = @floatFromInt(f64, random_matches) * 1.0 / trials; 375 | // std.debug.print("fpp {d:3.10} (estimated)\n", .{fpp}); 376 | // std.debug.print("\t(keys={}, random_matches={}, trials={})\n", .{ size, random_matches, trials }); 377 | // std.debug.print("\tbits per entry {d:3.1}\n", .{@floatFromInt(f64, filter.sizeInBytes()) * 8.0 / @floatFromInt(f64, size)}); 378 | } 379 | 380 | test "xor8" { 381 | try xorTest(u8, 10000, 12370); 382 | } 383 | 384 | test "xor16" { 385 | try xorTest(u16, 10000, 24700); 386 | } 387 | 388 | test "xor20" { 389 | try xorTest(u20, 10000, 49360); 390 | } 391 | 392 | test "xor32" { 393 | // NOTE: We only use 1m keys here to keep the test running fast. With 100 million keys, the 394 | // test can take a minute or two on a 2020 Macbook and requires ~6.3 GiB of memory. Still, 395 | // estimated fpp is 0 - I leave it to the reader to estimate the fpp of xor32/xor64. 396 | // 397 | // If you have a really beefy machine, it would be cool to try this test with a huge amount of 398 | // keys and higher `trials` in `xorTest`. 399 | try xorTest(u32, 1000000, 4920160); 400 | } 401 | 402 | test "xor64" { 403 | // NOTE: We only use 1m keys here to keep the test running fast. With 100 million keys, the 404 | // test can take a minute or two on a 2020 Macbook and requires ~6.3 GiB of memory. Still, 405 | // estimated fpp is 0 - I leave it to the reader to estimate the fpp of xor32/xor64. 406 | // 407 | // If you have a really beefy machine, it would be cool to try this test with a huge amount of 408 | // keys and higher `trials` in `xorTest`. 409 | try xorTest(u64, 1000000, 9840280); 410 | } 411 | --------------------------------------------------------------------------------