-
Notifications
You must be signed in to change notification settings - Fork 7
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
* Simplified the in-place native methods' signatures * Added a check whether JNI "get critical" methods returned a copy * Added a benchmark JAR task * Simplified the native methods' signatures * Removed F64FlatArray.transpose * Introduced unroll mechanics for F64Array * Serialization uses isFlattenable instead of isDense * Changed native signature of logAddExp to dst-src syntax * Simplified logRescale signature * Simplified cumSum signature * logAddExp now deals with positive infinities and NaNs * Corrected guessShape implementation * Correct JNI copy processing * Changed Viewer safety mode to publication * Added extensive Markdown documentation * Reorganized benchmarks * Removed weightedSum / Mean, since they were never used * Fixed argMin / argMax logic * Added benchmarking data and description thereof to documentation * Fixed Travis builds * Bumped version to 1.0.0
- Loading branch information
Showing
79 changed files
with
2,611 additions
and
1,532 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -24,3 +24,5 @@ cache: | |
| directories: | ||
| - $HOME/.gradle/caches/ | ||
| - $HOME/.gradle/wrapper/ | ||
|
|
||
| dist: trusty | ||
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,120 @@ | ||
| # viktor Benchmarks | ||
|
|
||
| We designed a series of microbenchmarks to compare `viktor`'s | ||
| native SIMD optimization efficiency to that of a simple Kotlin/Java loop. | ||
|
|
||
| The microbenchmark code can be found in `src/jmh` folder. To run the benchmarks, run | ||
| the following commands from `viktor`'s root folder: | ||
| ```bash | ||
| $ ./gradlew clean assemble benchmarkJar | ||
| $ java -jar ./build/libs/viktor-benchmark.jar | ||
| ``` | ||
| You can add the usual [JMH](https://openjdk.java.net/projects/code-tools/jmh/) | ||
| command line arguments to the latter command, e.g. `-o` to specify | ||
| the output file or `-t` to control the number of threads. | ||
|
|
||
| ## Benchmark Environment | ||
|
|
||
| We conducted the benchmarking on two machines: | ||
| a laptop running on Intel Core i7-6820HQ CPU at 2.70GHz | ||
| and a server running on Intel Xeon E5-2690 at 2.90GHz. | ||
| The following table summarizes the main features of both: | ||
|
|
||
| machine | laptop | server | ||
| --------|--------|------- | ||
| CPU | Intel Core i7-6820HQ | Intel Xeon E5-2690 | ||
| frequency | 2.70GHz | 2.90 GHz | ||
| cores[^cores] | 8 | 32 | ||
| architecture | `amd64` | `amd64` | ||
| highest SIMD extension[^simd] | `AVX` | `SSE2` | ||
| OS | Ubuntu 18.04.3 LTS | CentOS Linux 7 (Core) | ||
|
|
||
| [^cores]: The number of cores shouldn't matter since all benchmarks ran in a single thread. | ||
|
|
||
| [^simd]: The most advanced extension that was used by `viktor`. In reality, the laptop | ||
| had `AVX2` and the server had `SSE4.2` as the highest extension. | ||
|
|
||
| ## Benchmark Results | ||
|
|
||
| The following data is provided for informational purposes only. | ||
| In each benchmark, we considered arrays of size `1000`, `100_000` and `10_000_000`. | ||
| We investigate two metrics: | ||
| * `Array ops/s` is the number of times the operation was performed on the entire array | ||
| per second. Shown on a logarithmic scale. | ||
| * `FLOPS` is the number of equivalent scalar operations performed per second. It is equal | ||
| to `Array ops/s` multiplied by `Array size`. Shown on a linear scale. | ||
|
|
||
| ### Math Benchmarks | ||
|
|
||
| The task here was to calculate exponent (logarithm, `expm1`, `log1p` respectively) | ||
| of all the elements in a double array. It was done either with a simple loop, | ||
| or with a dedicated `viktor` array method (e.g. `expInPlace`). We also evaluated the | ||
| efficiency of | ||
| [`FastMath`](https://commons.apache.org/proper/commons-math/javadocs/api-3.3/org/apache/commons/math3/util/FastMath.html) | ||
| methods compared to Java's built-in `Math`. | ||
|
|
||
| We also measured the performance of `logAddExp` method for adding | ||
| two logarithmically stored arrays. It was compared with the scalar `logAddExp` | ||
| defined in `viktor`. | ||
|
|
||
| #### Laptop | ||
|
|
||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|
|
||
| #### Server | ||
|
|
||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|
|
||
| ### Statistics Benchmarks | ||
|
|
||
| We tested the `sum()`, `sd()` and `logSumExp()` methods here. We also measured | ||
| the throughput of a dot product of two arrays (`dot()` method). All these benchmarks | ||
| (except for `logSumExp`) don't have a loop-based `FastMath` version since they only | ||
| use arithmetic operations in the loop. | ||
|
|
||
| #### Laptop | ||
|
|
||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|
|
||
| #### Server | ||
|
|
||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|  | ||
|
|
||
| ## Cautious Conclusions | ||
|
|
||
| `viktor` seems to perform up to three times better than the | ||
| regular scalar computation approach. The only exception to that seems to be | ||
| `FastMath.exp()` which is on par `viktor`'s `exp()` method on `AVX` | ||
| (and faster than that on `SSE2`) . |
Oops, something went wrong.
1efb244There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'll probably never know why GitHub decided to do a merge commit even though I clicked "Squash and Rebase".