This subrepository demonstrates the binary matrix decomposition described in the article:
AAAI-19: Multi-Precision Quantized Neural Networks via Encoding Decomposition of {-1,+1}.

• run:

nvcc BMD-2-2bit.cu -std=c++11 -lcublas && ./a.out

This benchmark performs 8192x8192x8192 matrix multiplications with our method and baseline kernel. We implement the multiplication of two matrixes via our {-1, +1} encoding scheme on GTX 1080 GPU.

The matrix multiplication after encoded by 2-bit has obtained at most ∼15.89× speedup ratio than baseline kernel.

• Nvidia GPU
• CUDA

BinaryNet: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1.

Code: BinaryNet