NNPACK - AshokBhat/ml GitHub Wiki
About
- An acceleration package for neural network computations
- https://github.com/Maratyszcza/NNPACK
- No longer actively developed
Features
- Multiple algorithms for convolutional layers:
- Fast convolution based on Fourier transform (for kernels up to 16x16 without stride)
- Fast convolution based on Winograd transform (for 3x3 kernels without stride)
- Implicit matrix-matrix multiplication algorithm (no limitations)
- Direct convolution algorithm (for 1x1 kernels without stride)
- Multi-threaded SIMD-aware implementations of neural network layers
Development status
- No longer actively developed
- Originally developed by Marat Dukhan of Georgia Tech with guidance from Facebook AI Research
- Marat Dukhan now is one of the main contributors to XNNPACK
A comment in GitHub -
NNPACK is not actively developed anymore and XNNPACK is the successor and is in my experience faster (as well as having a better and more complete interface)
Usage
Frameworks
- PyTorch - On mobile for inference in convolutional layers.
- TVM - for inference in convolutional layers
- MXNet - for inference in convolutional layers, fully-connected, and max-pooling layers.
End users
- Facebook - in production.
- Prisma - in the mobile app