ECBC: Efficient Convolution via Blocked Columnizing

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Direct convolution methods avoid storage issues but require complex data formatting.
• This formatting leads to increased time and memory usage.
• Indirect convolution methods, like im2col, can be memory-intensive if not optimized.

Purpose of the Study:

• To present an optimized indirect convolution algorithm that achieves high performance.
• To demonstrate that indirect convolution can avoid substantial memory overhead when implemented properly.
• To introduce the efficient convolution via blocked columnizing (ECBC) algorithm.

Main Methods:

• ECBC is inspired by the im2col algorithm and block matrix multiplication.
• Convolution computation is performed blockwise.
• The tensor-to-matrix transformation (im2col) is also done blockwise, reducing memory requirements.

Main Results:

• ECBC achieves high computation performance using optimized matrix multiplication subroutines.
• The blockwise approach minimizes memory overhead during tensor-to-matrix transformation.
• Experiments show ECBC's effectiveness across various platforms and networks.

Conclusions:

• ECBC offers a superior alternative to existing industrial-level convolution algorithms.
• Properly implemented indirect convolution, like ECBC, balances performance and memory efficiency.
• The blockwise strategy is key to ECBC's success in optimizing convolution computations.