Vector math benchmarks

[John McCutchan]

Hi,

Following up on my earlier mails about Float32Array/List performance I have done some initial benchmarking of 4x4 matrix multiply, 4x4 matrix inverse, and 4D vector transform by 4x4 matrix. The benchmark compares DartVectorMath, glmatrix.dart, and a new feature to the DartVM I've been working on- Float32List SIMD (SSE) operations:

Each benchmark is averaged over 10 runs of 20,000 iterations. See the raw results below.

Executive summary:

glmatrix.dart performance is _really_ bad, I'm guessing that this is largely a result of the Float32List performance issue I pointed out in the other thread.

DartVectorMath performance is okay. It actually beats the SIMD backend for vector transforms.

SIMD performance is 2x faster for matrix * matrix, 4x faster for matrix inverse, and 50% slower for vector transform.

Thanks,

John

Starting benchmark

Clock frequency: 1000000

=============================================

Matrix Multiplication

=============================================

Avg: 14.59 ms Min: 10.161 ms Max: 22.927 ms (Avg: 14590 Min: 10161 Max: 22927)

=============================================

Matrix Multiplication SIMD

=============================================

Avg: 8.702 ms Min: 8.475 ms Max: 9.217 ms (Avg: 8702 Min: 8475 Max: 9217)

=============================================

Matrix Multiplication glmatrix.dart

=============================================

Avg: 283.353 ms Min: 272.062 ms Max: 287.988 ms (Avg: 283353 Min: 272062 Max: 287988)

=============================================

mat4x4 inverse

=============================================

Avg: 28.289 ms Min: 21.019 ms Max: 34.891 ms (Avg: 28289 Min: 21019 Max: 34891)

=============================================

mat4x4 inverse SIMD

=============================================

Avg: 7.107 ms Min: 6.89 ms Max: 7.754 ms (Avg: 7107 Min: 6890 Max: 7754)

=============================================

mat4x4 glmatrix.dart

=============================================

Avg: 318.909 ms Min: 315.435 ms Max: 325.831 ms (Avg: 318909 Min: 315435 Max: 325831)

=============================================

vector transform

=============================================

Avg: 4.324 ms Min: 2.811 ms Max: 14.859 ms (Avg: 4324 Min: 2811 Max: 14859)

=============================================

vector transform SIMD

=============================================

Avg: 6.415 ms Min: 6.204 ms Max: 7.006 ms (Avg: 6415 Min: 6204 Max: 7006)

=============================================

vector transform glmatrix.dart

=============================================

Avg: 144.431 ms Min: 138.263 ms Max: 153.798 ms (Avg: 144431 Min: 138263 Max: 153798)

--
John McCutchan <jo...@johnmccutchan.com>

[Peter Jakobs]

Thanks for the benchmark, the performance of my glmatrix port is
depressing bad...
Im curiouse how well it performs on the dartium-client with
Float32Array instead of the Float32List.

> John McCutchan <j...@johnmccutchan.com>

[Srinivas JONNALAGADDA]

        What do you think, John, is causing the large sigma in the readings for DartVectorMath?  Or, is it just outlier effect?  Thanks.

                                                            Greetings,

                                                                    JS

____

http://oneofmanyworlds.blogspot.com

[John McCutchan]

Hi Peter,

Don't get discouraged. The Dart VM is still very young and if I'm correct the performance problem can be fixed, bringing your port to parity (maybe faster) with DartVectorMath.

John

--
John McCutchan <jo...@johnmccutchan.com>

[John McCutchan]

Hi JS,

I think the sigma (standard deviation) of DartVectorMath performance comes from the VM running the first couple passes as unoptimized, the dart VM has a call count threshold before optimizing the function. If you pass the disassemble option to the VM you will get both the unoptimized and optimized versions of each function. 

John

--
John McCutchan <jo...@johnmccutchan.com>