4 ms·

GLMW – WebAssembly Powered Matrix and Vector Library

32 points by indescions_2018 9y ago

bhouston 9y ago
Amazing! I'm one of the main contributors to the Three.JS math library and I am always looking for speed improvements. Can you help me understand the benchmarks because there is some weirdness in them. Usually I can make sense of JS benchmarks, either it is memory allocations, deoptimization, function overhead, implicit conversions, etc. Specifically these results: https://maierfelix.github.io/glmw/mat4/ https://maierfelix.github.io/glmw/mat4/ mat4.create: GLM: 0.3000000142492354ms GLMW: 0.2999999560415745ms mat4.copy: GLM: 4.399999976158142ms GLMW: 0.5000000237487257ms mat4.set: GLM: 6.999999983236194ms GLMW: 6.800000031944364ms mat4.identity: GLM: 2.800000016577542ms GLMW: 0.4000000189989805ms Each of the above operations appears to be a series of 16 assignments in order, and some of the above have an additional allocation first. These options sometimes it takes the same amount of time in JS and WA and sometimes WA is way faster. Any idea as to why? The one that is weirdest above is that mat4.create is faster than mat4.identity, but mat4.create does more, specifically an allocation? I am struggling to make a mental model of why that is? Also why is perspective so slow in WA: mat4.perspective: GLM: 4.799999995157123ms GLMW: 4.499999980907887ms It is a tiny bit of math, with a tan and then 16 assignments. Is tan that slow? I sort of think there is something wrong with the benchmarking.
- bhouston 9y ago
  I see that you used a custom benchmarking library. I think that is the problem and why I can not make sense of the results. I suggest https://benchmarkjs.com/ https://benchmarkjs.com/ It is what we use for Three.JS benchmarking.
  - beiller 9y ago
    Im just thinking of ways to patch this in to THREE.Matrix4 :) Do you think that would speed things up significantly? Like I have large many-bones skeletons that can take time to animate.
    - bhouston 9y ago
      Until I see reliable benchmarks and understand why Matrix4 is slow I do not know. I think that the speed difference may be mostly related to avoiding allocations as in WA it is mostly just views rather than new ArrayBuffers if I am not mistaken. I thought everything else should be fast. But I need to see the benchmarks accurately to know.
johnsonjo 9y ago
Sorry if I’m piggybacking this thread my comment might be too off topic. This seems interesting, although it’s probably too low level for my current needs (I don’t have a background in linear algebra and numerical computing). I need to rewrite some Fortran code into JavaScript for a school project (a neural network) and I need to do operations in the BLAS subroutines. I think I’m going to end up using SciJS [1]. It seems promising at least feature wise, but if anyone has any further suggestions/recommendations I’d like to look into it. [1]: http://scijs.net/packages/ http://scijs.net/packages/

Amazing! I'm one of the main contributors to the Three.JS math library and I am always looking for speed improvements. Can you help me understand the benchmarks because there is some weirdness in them. Usually I can make sense of JS benchmarks, either it is memory allocations, deoptimization, function overhead, implicit conversions, etc. Specifically these results: https://maierfelix.github.io/glmw/mat4/ https://maierfelix.github.io/glmw/mat4/ mat4.create: GLM: 0.3000000142492354ms GLMW: 0.2999999560415745ms mat4.copy: GLM: 4.399999976158142ms GLMW: 0.5000000237487257ms mat4.set: GLM: 6.999999983236194ms GLMW: 6.800000031944364ms mat4.identity: GLM: 2.800000016577542ms GLMW: 0.4000000189989805ms Each of the above operations appears to be a series of 16 assignments in order, and some of the above have an additional allocation first. These options sometimes it takes the same amount of time in JS and WA and sometimes WA is way faster. Any idea as to why? The one that is weirdest above is that mat4.create is faster than mat4.identity, but mat4.create does more, specifically an allocation? I am struggling to make a mental model of why that is? Also why is perspective so slow in WA: mat4.perspective: GLM: 4.799999995157123ms GLMW: 4.499999980907887ms It is a tiny bit of math, with a tan and then 16 assignments. Is tan that slow? I sort of think there is something wrong with the benchmarking.

bhouston 9y ago
I see that you used a custom benchmarking library. I think that is the problem and why I can not make sense of the results. I suggest https://benchmarkjs.com/ https://benchmarkjs.com/ It is what we use for Three.JS benchmarking.
- beiller 9y ago
  Im just thinking of ways to patch this in to THREE.Matrix4 :) Do you think that would speed things up significantly? Like I have large many-bones skeletons that can take time to animate.
  - bhouston 9y ago
    Until I see reliable benchmarks and understand why Matrix4 is slow I do not know. I think that the speed difference may be mostly related to avoiding allocations as in WA it is mostly just views rather than new ArrayBuffers if I am not mistaken. I thought everything else should be fast. But I need to see the benchmarks accurately to know.

johnsonjo 9y ago

Sorry if I’m piggybacking this thread my comment might be too off topic. This seems interesting, although it’s probably too low level for my current needs (I don’t have a background in linear algebra and numerical computing). I need to rewrite some Fortran code into JavaScript for a school project (a neural network) and I need to do operations in the BLAS subroutines. I think I’m going to end up using SciJS [1]. It seems promising at least feature wise, but if anyone has any further suggestions/recommendations I’d like to look into it. [1]: http://scijs.net/packages/ http://scijs.net/packages/