Releases: waylonflinn/weblas
The Pipeline Release
The Pipeline Release™️
This release is all about performance. It adds a new version of all operations that allow data to be reused without removing it from GPU memory. This means long term items, like network weights and biases, can stay in GPU memory between operations. The results of previous computations can also be used as inputs. This can significantly increase performance.
More information (and examples) can be found on the wiki Pipeline page.
This release also includes compatibility improvements. The library should now work in most major browsers on most major platforms (including internet explorer). Pipeline features are likely to fail on mobile (as they require the ability to render to floating point textures)
The Fractal of Features Release
The Fractal of Features Release™️
Includes most of the stuff you need for building a convolutional neural network.
- downsampling (
sdwns) - Max Pooling - clamping (
sclmp) - Rectified Linear Units (ReLU) - scaling (
sscal) - Image Staging and Normalization
Performance is good, but memory transfers are still a bottleneck for realtime image classification. The next release (The Pipeline Release™️) should slay that beast. I should also add some examples, so everyone in the world can build a deep neural network for their browser.
The Nice Release
The Nice Release™️
- Up to four times faster than previous releases
- Up to fifty percent memory savings
- Functional source map
- No annoying warnings in the console
Alpha - For the Second Time
- Improved precision on non-square matrices
- Reduced file size
- Fixed a bug on Intel GPUs that caused values near 128 to have reduced precision
Alpha
Basic (but still very fast) matrix multiplication functionality.
100xspeedup (over javascript)- matrices up to
4096 x 4096 Float32Array
Tests done with with: GeForce GTX 970 vs Core i7 CPU 860 @ 2.80GHz on 1024 x 1024 (speedup is even larger on larger matrices). In our tests, results were comparable to OpenBLAS on 8 cores (0.053 seconds)