_____ __ __ _ _ _ ____
/ ____| \/ | | | | | | |___ \
| (___ | \ / | |__| | __ _ ___| |__ ___ _ __ __) |
\___ \| |\/| | __ |/ _` / __| '_ \ / _ \ '__|__ <
____) | | | | | | | (_| \__ \ | | | __/ | ___) |
|_____/|_| |_|_| |_|\__,_|___/_| |_|\___|_| |____/
=======================================================
If you are interested in the latest hash test results
(currently from SMHasher3 SMHasher3 beta3-c6b9cc18
), they are in the
results/
directory.
There are a few hashes which do not pass all the tests, but are very close
to doing so. These are: chaskey-12
are MeowHash
. These failures are
false positives. The failure thresholds have been adjusted, and these
hashes are expected to be listed as "passing" in the next round of results.
SMHasher3 is a tool for testing the quality of hash functions in terms of their distribution, collision, and performance properties. It constructs sets of hash keys, passes them through the hash function to test, and analyzes their outputs in numerous ways. It also does some performance testing of the hash function.
SMHasher3 is based on the SMHasher fork maintained by Reini Urban, which is in turn based on the original SMHasher by Austin Appleby. The commit history of both of those codebases up to their respective fork points is contained in this repository.
The major differences from rurban's fork are:
- Fix several critical bugs
- Several new tests and test methods added
- Significant performance increases
- Report on p-values for all supported tests
- Detailed reporting on hashes when test failures occur
- Better statistical foundations for some tests
- Overhauled all hash implementations to be more consistent
Additional significant changes include:
- Many fixes to threaded testing and hashing
- More consistent testing across systems and configurations
- More consistent and human-friendlier reporting formats
- Common framework code explicitly sharable across all hashes
- Flexible metadata system for both hashes and their implementations
- Major progress towards full big-endian support
- Support of more hash seed methods (64-bit seeds and ctx pointers)
- Ability to supply a global seed value for testing
- Test of varying alignments and buffer tail sizes during speed tests
- Refactored code to improve maintainability and rebuild times
- Reorganized code layout to improve readability
- Compilation-based platform probing and configuration
- Consistent code formatting
- More explicit license handling
- Fully C++11-based implementation
As of 2023-12-12, SMHasher3 beta3 has been released.
I have hopes to release a 1.0 version sometime before the end of September 2024. The current code should still be very useful in evaluating hashes, as most remaining features are either additional tests, tweaks to specific hashes, or quality-of-life usage enhancements.
This code is compiled and run successfully on Linux x64, arm, and powerpc using gcc and clang quite often. Importantly, I do not have the ability to test on Mac or Windows environments. It has been compiled successfully using MSVC and clang-cl in the past; efforts are made to ensure this remains the case, but some things may slip through. The goal is to support all of the above, and while the CMake files Should Just Work(tm), MSVC in particular has its own ideas about some corners of the various specs. So reports of success or failure are appreciated, as are patches to make things work.
mkdir build
cd build
cmake ..
orC=mycc CXX=mycxx CXXFLAGS="-foo bar" cmake ..
as needed for your systemmake -j4
ormake -j4 all test
./SMHasher3 --tests
will show all available test suites./SMHasher3 --list
will show all available hashes and their descriptions./SMHasher3 <hashname>
will test the given hash with the default set of test suites (which is called "All" and is most but not literally all of them)./SMHasher3 <hashname> --extra --notest=Speed,Hashmap
will test the given hash with the default set of test suites excluding the Speed and Hashmap tests, with each run test suite using an extended set of tests./SMHasher3 <hashname> --ncpu=1
will test the given hash with the default set of test suites, using only a single thread./SMHasher3 --help
will show many other usage options
Note that a hashname specified on the command-line is looked up via case-insensitive search, so you do not have to precisely match the names given from the list of available hashes. Even fuzzier name matching is planned for future releases.
If SMHasher3 found a usable threading implementation during the build, then
the default is to assume --ncpu=4
, which uses up to 4 threads to speed up
testing. Not all test suites use threading. While all included hashes are
thread-safe as of this writing, if a non-thread safe hash is detected then
threading will be disabled and a warning will be given. If no usable
threading library was found, then a warning will be given if a --ncpu=
value above 1 was used.
To add a new hash function to be tested, either add the implementation to an existing
file in hashes/
(if related hashes are already there), or copy hashes/EXAMPLE.cpp
to a new filename and then add it to the list of files in hashes/Hashsrc.cmake
.
Many more details can be found in hashes/README.addinghashes.md
.
This section has been placed near the front of the README because it is the most important and most visible new feature for existing SMHasher users.
The tests in the base SMHasher code had a variety of metrics for reporting results, and those metrics often did not take the test parameters into account well (or at all), leading to results that were questionable and hard to interpret. For example, the Avalanche test reports on the worst result over all test buckets, but tests with longer inputs have more buckets. This was not part of the result calculation, and so longer inputs naturally get higher percentage biases (on average) even with truly random hashes. In other words, a bias of "0.75%" on a 32-bit input was not the same as a bias of "0.75%" on a 1024-bit input. This is not to call out the Avalanche test specifically; many tests exhibited some variation of this problem.
To address these issues, SMHasher3 tests compare aspects of the distribution of hash values from the hash function against those from a hypothetical true random number generator, and summarizes the result in the form of a p-value.
P-values are probabilities: they are numbers between 0 and 1. Their values are approximately the probability of a true RNG producing a test result that was at least as bad as the observed result from the hash function. Smaller p-values would indicate worse hash results.
However, these p-values quite often end up being very small values near zero, even in cases of good results. Reporting them in their decimal form, or even in scientific notation, would probably not be very useful, and could be very difficult to compare or interpret just by looking at them.
In SMhasher3, these p-values are reported by a caret symbol (^) followed by the p-value expressed in negative powers of two. For example, if it is determined that a true RNG would be expected to produce the same or a worse result with a probability of 0.075, then SMHasher3 would compute that that p-value is about 2^-3.737. It would then round the exponent towards zero, simply discard the sign (since probabilities are never greater than 1, the exponent is always negative), and finally report the p-value as "^ 3".
Therefore, smaller p-values (which indicate worse test results) result in larger numbers when reported using caret notation. You can think of the values in caret notation as indicating how improbable, and thus worse, the test result was. For example, "^50" could be interpreted as "there is, at best, only a 1 in 2^50 chance that an RNG would have produced a result as bad as the hash did".
The p-value computations only care about the likelihood of bad results (e.g. more collisions than an RNG would produce). Test results that are better than a typical RNG result but would still be outliers from a purely statistical point-of-view, such as seeing no or very few collisions when at least some would be expected, do not produce extreme p-values. In statistics terms, the p-values are one-tailed when appropriate, instead of always being two-tailed.
The p-value computations also take into account how many tests are being summarized, which can lead to unintuitive results. As an example, here are some lines from a single batch of test keys:
Keyset 'Sparse' - 256-bit keys with up to 3 bits set - 2796417 keys
Testing all collisions (high 32-bit) - Expected 910.2, actual 989 (1.087x) (^ 7)
Testing all collisions (high 20..38 bits) - Worst is 32 bits: 989/910 (1.087x) (^ 3)
The middle line reports ^7 for seeing 989 collisions when 910 were expected, and the last line reports ^3 for what seems like the same result. This is due to the fact that the middle line is reporting that as the result of a single test, and the last line is reporting that as the worst result over 19 tests. It's much more likely to see a result at least that bad if you have 19 tries to get it than if you just had 1 try, and so the improbability is much lower. Indeed, 19 is around 2^4, and the first reported result is about 4 powers of 2 worse than the second (7 - 3), as expected.
A true RNG would generally have about twice as many ^4 results as ^5 results, and twice as many ^3 results as ^4 results, and so on. However, many of the statistical formulas used by SMHasher3 only produce bounds on the result probabilities, and sometimes those bounds are not very tight and/or get significantly worse for higher-likelihood results. The formulas used were typically chosen for greater accuracy in failure / long-tail cases. Further, some tests are very unlikely to get even a single "hit", and so a result of zero hits can't really give a precise p-value. For those reasons and more, you should expect to see more lower numbers than the power-of-2 relationship would imply, and you will see many more ^0 results than you would expect mathematically.
All non-deprecated tests in SMHasher3 support p-value reporting. These p-value results are the only numbers used by SMHasher3 to determine pass/warn/fail status for tests. And since the really, truly most important result of testing is "does a hash pass or fail", and perhaps noting how close to the line it is, the precise p-value is not very important. The reported values are always lower bounds on the actual p-value exponents (the "true" result could be worse than reported but never better), so any failures reported should be genuine.
The precise cutoffs for test warnings and failures can be found at the top
of util/Reporting.cpp, in the variables
FAILURE_PBOUND
and WARNING_PBOUND
. As of this writing, a warning is
given at ^16, and a failure is given at ^20. Those bounds might seem
surprisingly high, but that is because there are so many tests. Since a
typical full SMHasher3 test run consists of about 16,000 tests, even
testing a cryptographic-quality hash function is expected to produce a ^14
event every run on average (-log2(1/16,000) =~ 13.966) (this calculation
overstates things because test failures are far from independent). The
failure threshold was chosen to correspond to be less than a 1% chance of
false test failure, and the warning threshold was arbitrarily chosen to
make them about 16 times as likely as failures.
For the statistics folks, this is a correction for the Multiple comparisons problem, and the correction used is slightly weaker than what the Bonferroni correction would call for. This should be OK since SMHasher3 uses a large number of tests, and failures are positively correlated. Maybe the Harmonic mean p-value procedure will be used in the future.
A final summary table of p-values in caret notation is currently produced after a full run. This table can be useful to see a summary of how close to the pass/fail line a particular hash is, or to see if some suspicious patterns (e.g. many warning values) exist. It is important to remember that this table should NOT be used to compare hashes. SMHasher3 focuses on broad testing to find classes of bad behavior in hashes. It doesn't do nearly the depth of testing to fairly compare the quality of hash outputs across candidate functions, regardless of any particular definition of "quality", which may also vary across perspectives.
A number of significant performance improvements have been made over the base SMHasher code. Here are some runtime comparisons on my system (AMD Ryzen 9 3950X, 1 or 4 isolated CPUs, all with boost disabled and pinned to 3500 MHz for timing consistency, gcc 9.3, Slackware 14.2, SMHasher3 beta1, smhasher-rurban b116571):
Test Name | SMHasher | SMHasher3 | Delta | SMHasher | SMHasher3 | Delta |
---|---|---|---|---|---|---|
BadSeeds | 996s | 311s | -69% | 1194s | 78s | -93% |
Window | 935s | 341s | -64% | |||
Avalanche | 720s | 92s | -87% | 810s | 23s | -97% |
Sparse | 478s | 151s | -68% | |||
TwoBytes | 292s | 79s | -73% | |||
Diff | 263s | 171s | -34% | 263s | 43s | -84% |
Permutation | 256s | 98s | -62% | |||
Popcount | 163s | 20s | -87% | 90s | 5s | -94% |
BIC | 152s | 9s | -94% | 152s | 3s | -98% |
Text | 65s | 20s | -69% | |||
PerlinNoise | 47s | 30s | -36% | |||
Prng | 33s | 8s | -76% | |||
DiffDist | 11s | 14s | +27% | |||
Cyclic | 8s | 2s | -75% | |||
Zeroes | 5s | 5s | 0% | |||
Seed | 5s | 7s | +40% | |||
Sanity | 3s | <1s | -90% |
Since Gitlab's flavor of markdown only supports one header row, the first 3 columns of numbers are for 1 CPU, and the last 3 columns are for 4 CPUs on tests which support threading in SMHasher3.
The SMHasher results are somewhat confusing. The BadSeeds test is threaded but takes more wall clock time than the unthreaded version. I attribute this to a large amount of system CPU time that the threaded version takes that the unthreaded version doesn't. I don't see any obvious synchronization primitives being used or intentional data sharing across threads, so I am unsure of its source. The Avalanche test is not threaded in SMHasher, but it takes more wall clock time than the unthreaded version regardless, which I have no good hypothesis about. Both of these results are repeatable and consistent, though, so I am keeping them in the table.
After beta1, the test methodology in SMHasher3 diverges quite a lot from SMHasher, and so direct performance comparisons are less meaningful. But to give one data point, on the above system a complete run of SMHasher3 beta2 on a fast hash function (wyhash) with --extra but without BadSeed testing takes 1578 seconds and finds 40 failing tests, while smhasher-rurban takes 2860 seconds and finds 2 failing tests.
For beta3, or shortly thereafter, I plan on publishing some explicit data from performance profiling, to show the places that I think are the best to look for more performance gains, or at least would have the highest impact.
Right now, there are 3 places that already have work underway. First, I have more ways to improve blobsort() performance significantly, I'm pretty sure. Second, another contributor is working on improving the bit correlation histogram code (util/Histogram.h and the related code in the BIC and SeedBIC tests). Third, a number of tests can be augmented to use threads. I have a number of thoughts on this issue, so please reach out if you decide to start on that.
The priority of SMHasher3 is to be the best test utility for hash function quality. Other important goals are:
- Support as many platforms as practical
- Have identical results on those platforms
- Test all valid variations on a hash function
- Be as fast as possible, given those constraints
- Allow for comparisons between hash function cores, both within and across hash function families, to facilitate learning about hash internals
- Provide tools for more quickly iterating on hash functions
- Be easily expandable, both for tests and hash function infrastructure
- Have a consistent, readable code base
- Be explicit about code licensing
- Have human-friendly reporting
SMHasher3 also does performance testing, and this is expected to be worked on and expanded in the future. However, performance testing fidelity will always come second to functional testing. The goal of the performance testing in SMHasher3 will be to provide effective general comparisons of hash functions, not absolute performance ratings or measurements.
There are some other things that SMHasher3 is explicitly NOT trying to do:
- Be an authoritative or complete repository of hash implementations
- Use hash authors' implementation code with as few changes as possible
- Provide a numeric score for overall hash quality
- Be a final arbiter of which hash functions are "best"
See Changelog.md
for a detailed list of the differences going from the
forked copy of SMHasher to SMHasher3.
One of the goals for SMHasher3 is full support of both big- and little-endian systems. Currently this is, in some sense, a little bit more than half complete. Every hash implementation computes results for both endiannesses, regardless of system endianness. Most of the testing code is not yet endian-independent, however, and so test results will currently vary greatly depending on the system.
For hash authors, this represents a tiny bit of extra work, but it can be put off until late in hash development, and is not very difficult to add.
Hashes which have explicit specifications of endian-independent hash values (mostly cryptographic hashes; e.g. SHA-1 needs to return the same result for the same inputs no matter the system) have either what is referred to in SMHasher3 as their "canonical" endianness (aka "CE", which matches their spec) and their "non-canonical" endianness (aka "NE", which doesn't).
Most non-cryptographic hashes have no such requirement, and are more interested in the speed of not having to do byteswapping on their input and output data than they are in hash result consistency across platforms. These hashes have what SMHasher3 refers to as "little-endian" and "big-endian" results.
The user can request that hashes compute results for big- or little-endianness
explicitly, or they can request system-native or system-nonnative endianness, or they
can request the "default" endianness (which is "canonical" if that exists or
"system-native" if not) or "nondefault" (whichever one "default" isn't). This is done
through the --endian
command-line option.
Each hash itself is usually not concerned with this, though! It is only concerned with "do inputs and outputs have to byteswapped, or not?", and the SMHasher3 framework will use the appropriate implementation based on a combination of the user's request, the hash's metadata, and the detected endianness of the system.
While I am very happy with this goal, and mostly happy with the hash-side implementation of things, I'm less sure that the user-side is good enough. Please let me know if you find something confusing.
In beta2 of SMHasher3, the algorithm for computing hash verification codes is
unchanged from the base SMHasher. Most hashes' verification codes are also unchanged,
but a number have changed for various reasons, and some have been added or removed;
see Changelog.md
for specifics. This should help verify that the hash
implementations didn't change unexpectedly when they were ported.
Since SMHasher3 supports 64-bit seeds and the current algorithm for computing verification codes does not exercise even all of the low 32 seed bits, I expect that the algorithm will change in the near future.
A stand-alone vanilla C99 program for computing hash verification codes outside of
SMHasher3 is in misc/hashverify.c
.
The original SMHasher code base had some infrastructure for what it called VCodes, although they were not implemented. The intention seemed to be that they would be short signatures of all of the hash inputs and outputs as well as the test results. SMHasher3 has implemented this, though it uses a somewhat different framework internally.
I've kept the name "VCode", but it is important not to confuse these with the "verification codes" from the previous section. They are unrelated features, except for the fact that they both are intended for human use to quickly verify that SMHasher3 is functioning the same way in different configurations.
If the --vcode
command-line option is used, then these signatures are computed and
reported on. A final summary "verification value" is computed from these 3 component
VCode signatures, and is reported on the last line of output. They are an easy way to
compare complete operation across runs and/or platforms, without having to compare
result-by-result. There is a small but noticable performance hit when this is
enabled. If it is not enabled, then the VCode component output line is no longer
emitted, but the final summary code is (with a value of 1 to indicate it was not
computed), to keep output lengths consistent.
The performance-oriented tests do not contribute to VCode calculations. Note that even input VCodes will vary by hash width.
Most of the tests in SMHasher3 don't vary the hash seed values, and so use a global
seed value which defaults to 0. A different value can be specified via the --seed=
command-line option. This allows for a single hash to be tested multiple times, which
can show things like if a given result is a fluke or is consistent across runs.
Tests which do vary the seed values generally ignore this global seed value.
If a hash implementation only takes a 32-bit seed and the given global seed value exceeds 32-bit representable values, then the seed value is truncated and a warning is emitted.
There are 2 main parts of the codebase, and they are largely independent of each other.
The first is Hashlib, which is a collection of hash function implementations and their metadata, some shared routines for things which hash functions frequently do, as well as some code for querying and managing the collection.
The second is Testlib, which is a collection of code and utility functions to generate sets of keys to be hashed and to analyze the lists of hash results.
There is also main.cpp, which is the main SMHasher3 program and ties the whole thing together.
Hashlib hash implementations are under hashes/. There is one .cpp file, starting with
a lowercase letter, per family of hash functions. There may also be a directory with
the same name as the base name of the .cpp file it belongs to
(e.g. hashes/blake2.cpp
has a hashes/blake2/
directory). This directory can
contain whatever other data or code that hash family needs. The most common use of
that directory is to contain a set of .h files which each contain a
separately-optimized version of the same core hash routines (e.g. a generic version,
an SSE2+ version, an AVX+ version, and an ARM version) which are included as needed
in the .cpp file.. More details about hash naming and reasoning on certain
implementation aspects can be found in hashes/README.md
.
The rest of the Hashlib code is under lib/
, and the interface header files are
under include/
. include/hashlib/
files are only exposed to the hash functions,
and include/common/
files are exposed globally.
Testlib code lives under tests/
, where each test suite has one .cpp file starting
with an uppercase letter. Common testing code lives under util/
, where each group
of routines has one .h file and possibly one .cpp file, both also starting with an
uppercase letter.
There is some build-related code in hashes/
in files starting with uppercase
letters, and quite a lot in the platform/
directory. The misc/
directory contains
other code which is related to SMHasher3 but is not part of the main program.
It is very common in C++ to have considerable amounts of code in header (.h)
files. While there can be good reasons for this, I strongly prefer having most code
in .cpp files where it will be compiled only once. SMHasher3's test code is generally
templated based on the output width of the hash function being tested. Ordinarily
this would mean that the code would need to be in header files, so that it can be
instantiated based on the way it is called. However, since the complete list of hash
output types is known a priori, SMHasher3 can keep the code in .cpp files by
explicitly instantiating the appropriate versions there. See util/Instantiate.h
for
the gory details.
While SMHasher3 keeps the CMake foundation, the build system is very different from the base SMHasher code. The short version is that feature detection is done during the CMake configuration step by attempting to compile possible implementations of platform-specific features and then caching the results. Header files containing the detection results are generated in the build directory, and these are included by the rest of the code.
The primary file generated this way is build/include/Platform.h
. You can see an
example of what a rendered Platform.h
file might look like in
platform/Platform.h.EXAMPLE
. build/include/Timing.h
is the other generated header
file, and similarly you can see an example of what a rendered Timing.h
file might
look like in platform/Timing.h.EXAMPLE
.
In the worst case, since all data is kept in CMake's cache, detection state can be reset by removing the build directory contents, or by building in a new, clean directory.
More technical details and some discussion of the rationale behind this are in
platform/README.md
.
Files under hashes/, include/, lib/, and misc/ may have a variety of different different licenses, such as BSD 2, zlib, GPL3, public domain/unlicense/CC0, or MIT. Each file should have its license terms clearly stated at the top.
Files under results/ have the Creative Commons Attribution 4.0 International license (CC BY 4.0).
Files under util/parallel_hashmap/ have the Apache 2 license.
All other files in this distribution (including, but not limited to, main.cpp and files in platform/, tests/, and util/) are licensed under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
I would prefer to have the above information in LICENSE, but Gitlab
offers no way to manually set an advertised license, and so I need
to rely on its auto-detection to find "GPL3" (which is at least
closest to reality), so it lives here instead. :-/
The original SMHasher's README says:
SMHasher is released under the MIT license.
although there is no LICENSE file, nor are there any per-file license or copyright notices in the test code.
rurban's SMHasher's LICENSE file says:
smhasher itself is MIT licensed.
Plus copyright by the individual contributors. Note that the individual hash function all have different licenses, from Apache 2, BSD 2, zlib, SSL, GPL, public domain, ... to MIT.
and again there are no per-file license or copyright notices in the test code.
One of the goals of this fork is to be much more clear and explicit about code licensing. This proved to be trickier than expected. After considering several options and consulting with an attorney, I decided that the best option was to explicitly distribute the SMHasher3 testing code under the GPL.
Since some of the code linked in (at least 5 different hash implementations) to SMHasher in rurban's fork is under the GPL, the whole of the project must also be distributable under the GPL. Further, the MIT license explicitly allows sublicensing. Having the test code be explicitly under the GPL also should increase the odds that any new hash implementation being added to this project would not require any relicensing of the test code.
This decision was not taken lightly, as I would prefer to keep the original authors' license when possible, as was done with the modifications made to the hash implmentations. I believe this to have been the least bad option to get the improvements in SMHasher3 out to the world.
The LICENSE file of this project has been updated to reflect these terms. I have added the GPL license text to many of the files that are covered by it, and I have added the text of the original MIT license, as well as a list of contibutor copyrights, explicitly to much code that is being distributed here under the GPL, in order to comply with the MIT license terms of the originals. If I have somehow missed an attribution for some of the forked code, please do not hesitate to reach out so I can fix it!
Some code in SMHasher which seemed to be incompatible with GPL3 was not forked. Finally, other code files which are being distributed under non-GPL licenses will have their license added to them, to help remove confusion.
- Bit Twiddling Hacks [https://graphics.stanford.edu/~seander/bithacks.html]
- GNU parallel [https://www.gnu.org/software/parallel/]
- Hedley [https://nemequ.github.io/hedley/]
- Portable snippets [https://github.com/nemequ/portable-snippets]
- FFT [https://github.com/NFFT/nfft]
- Uncrustify [https://github.com/uncrustify/uncrustify]
- Valgrind [https://valgrind.org/]
- kcachegrind [https://github.com/KDE/kcachegrind]
- Parallel Hashmap [https://github.com/greg7mdp/parallel-hashmap/]
- MPFR (Multiple Precision Floating-Point Reliable Library) [https://www.mpfr.org/]
- MPFI (Multiple Precision Floating-Point Interval Library) [https://directory.fsf.org/wiki/MPFI]
- YAWL (Yet Another Word List) [https://github.com/elasticdog/yawl]
- Slackware Linux [http://www.slackware.com/]
- RMSBolt [https://gitlab.com/jgkamat/rmsbolt]