Convolutional Neural Networks (CNN) kernels

Introduction

This module consists of kernels to implement the core computations occurring in the context of convolutional neural networks.

• CNN Module is partitioned into
  • Row Convolution, CNN convolution operation is performed with continuous rows of input feature map processing. This is targeted for dense convolution
  • Column Convolution, CNN convolution operation is performed with partial rows but all columns processed. This is targeted for depthwise convolution
• Fully connected layer
• CNN Deconvolution operation post processing operation on the convolution output
• Input feature map data is 8, 16 bit signed and unsigned
• Coefficient data is 8, 16 bit signed
• Output feature map data is 8, 16 bit signed and unsigned

Sub Modules
	MMALIB_CNN_convolve_col_smallNo_ixX_ixX_oxX
	Kernel for computing CNN-style 2D convolution using column major data ordering on the input and output feature maps. This approach computes more quickly if filter grouping is chosen such that Ni=No=1, or if filter grouping is chosen such that NiFrFc < MMA_SIZE, otherwise use regular convolution method MMALIB_CNN_convolve_row_ixX_ixX_oxX. This kernel is also referred to as depth-wise convolution.
	MMALIB_CNN_convolve_row_ixX_ixX_oxX
	Kernel for computing dense CNN convolution with row based processing and matrix multiplication.
	MMALIB_CNN_deconvolve_row_ixX_ixX_oxX
	Kernel for computing dense CNN deconvolution with row-based processing and matrix-matrix multiplication.
	MMALIB_CNN_fullyConnected_ixX_ixX_oxX
	Kernel provides compute functionality of Fully Connected Layer: \( Y^T = X^T \times H^T \).
	MMALIB_CNN_pixelShuffle_row_ixX_ixX_oxX
	Kernel for computing dense CNN convolution with row based processing and matrix multiplication followed by column interleaving, which results in a partial output of final form in pixel shuffle operator.
	MMALIB_CNN_tensor_convert_ixX_oxX
	Kernel for converting tensors of various datatypes and formats.