GPU accelerated Javascript. Numerical computing in your browser with performance comparable to native.
Currently includes hundreds of unit tests, which verify correctness on hundreds of millions of data points.
Operations
Our focus is on numerical operations useful for neural networks and machine learning. So far, we've got 32-bit versions of each of these:
- sscal - Matrix (and Vector) Scale (with addition)
- sgemm - Matrix Multiply
- sdwns - Matrix (and Image) Downsample (for Max Pooling)
- sclmp - Matrix clamp (for ReLU)
Don't see what you need? Give a
Usage
First, include the weblas.js
file (from a release or the dist
directory).
<script type="text/javascript" src="weblas.js"></script>
Then use it like this.
<script>
// create Matrices as arrays
var height_A = 1024, width_A = 1024,
height_B = 1024, width_B = 1024;
var A = new Float32Array(height_A * width_A);
var B = new Float32Array(height_B * width_B);
// fill A and B with science
var M = height_A,
N = width_B,
K = height_B; // must match width_A
var alpha = 1.0;
var beta = 0.0;
var C = new Float32Array(width_B) // specialized for neural net bias calculation
// result will contain matrix multiply of A x B (times alpha)
result = weblas.sgemm(M, N, K, alpha, A, B, beta, C);
</script>
Pipeline Mode
Pipeline mode gives (sometimes very large) increases in performance by leaving data in GPU memory. A demo illustrating performance on a deep neural net can be found here.
Here's a basic example:
// create Tensor containers for interacting directly with GPU memory
var t0 = new weblas.pipeline.Tensor([M, K], data0);
// second matrix must be transposed
var t1 = new weblas.pipeline.Tensor([N, K], weblas.util.transpose(K, N, data1));
var t2 = new weblas.pipeline.Tensor([1, N], data2);
var alpha = 1.0;
var beta = 0.5;
/* NOTE: pipeline.sgemm takes a transpose matrix in the
second slot (t1 here)
(this requirement allows for improved performance)
*/
var t3 = weblas.pipeline.sgemm(alpha, t0, t1, beta, t2);
// result is a Float32Array
var result = t3.transfer();
More information can be found on the wiki Pipeline page.
Testing
Unit tests and benchmarks both require browserify
and testling
.
Install with:
npm install -g browserify
npm install -g testling
Unit Tests
All operations have unit test coverage. Unit tests use data generated outside
the browser (to verify correctness). Generating the data requires python
and
the modules in requirements.txt
.
With pip
installed run:
pip install -r requirements.txt
Then, to generate the data, run:
npm run data
Then, run the unit tests with:
npm test
OS Setup
If the tests won't run, try this (it restores the default npm browser setting)
OSX
npm config set browser open
Linux
npm config set browser xdg-open
Windows
npm config set browser start
Benchmarks
After installing browserify
and testling
, run the benchmarks with:
npm run benchmark
results
TAP version 13
ok 1 sgemm: 128x128 . 128x128
# 1.032 GFlops/sec ±3.71% n = 50 µ = 4ms
ok 2 sgemm: 128x256 . 256x128
# 1.745 GFlops/sec ±2.89% n = 44 µ = 5ms
ok 3 sgemm: 256x256 . 256x256
# 5.061 GFlops/sec ±2.89% n = 42 µ = 7ms
ok 4 sgemm: 512x256 . 256x512
# 15.454 GFlops/sec ±3.86% n = 51 µ = 9ms
ok 5 sgemm: 256x512 . 512x256
# 10.262 GFlops/sec ±2.76% n = 47 µ = 7ms
ok 6 sgemm: 512x512 . 512x512
# 22.231 GFlops/sec ±3.54% n = 50 µ = 12ms
ok 7 sgemm: 513x513 . 513x513
# 14.474 GFlops/sec ±4.51% n = 43 µ = 19ms
ok 8 sgemm: 1024x512 . 512x1024
# 41.859 GFlops/sec ±3.38% n = 43 µ = 26ms
ok 9 sgemm: 512x1024 . 1024x512
# 31.353 GFlops/sec ±2.60% n = 46 µ = 17ms
ok 10 sgemm: 1024x1024 . 1024x1024
# 45.545 GFlops/sec ±3.99% n = 31 µ = 47ms
ok 11 sgemm: 2048x2048 . 2048x2048
# 62.159 GFlops/sec ±28.88% n = 13 µ = 276ms
1..11
# tests 11
# pass 11
# ok
more information about benchmarks (including test configuration) can be found on the wiki.
Donations
Want to see more happen here? Contribute on