DPUCADF8H Configuration - 1.0 English

The deep neural network operators and the associated parameters supported by the DPUCADF8H are in the following table.

Table 1. Deep Neural Network Operators and Parameters Supported by DPUCADF8H
Operators	Parameters	Description
Convolution	Kernel Sizes	kernel_w: [1, 16] kernel_h: [1, 16]
	Strides	stride_w: [1, 8] stride_h: [1, 8]
	Padding	pad_w: [0, 15] pad_h: [0, 15]
	Input Size	input_channel: [1, 8192]
	Output Size	Arbitrary
	Activation	ReLU, PReLU, Leaky ReLU, ReLU6
Dilated Convolution	Kernel Sizes	kernel_w: [1, 16] kernel_h: [1, 16]
	Strides	stride_w: [1] stride_h: [1]
	Padding	pad_w: [0, 15] pad_h: [0, 15]
	Input Size	input_channel: [1, 8192]
	Output Size	Arbitrary
	Activation	ReLU, PReLU, Leaky ReLU, ReLU6
Deconvolution	Kernel Sizes	kernel_w: [1, 16] kernel_h: [1, 16]
	Strides	stride_w:1 stride_h: 1
	Padding	pad_w: [0, 15] pad_h: [0, 15]
	Input Size	input_channel: [1, 8192]
	Output Size	Arbitrary
	Activation	ReLU, PReLU, Leaky ReLU, ReLU6
Full Connected	Input Size	input_channel: [1, 1048576]
Pooling	Kernel Sizes	kernel_w: [1, 16] kernel_h: [1, 16]
	Strides	stride_w: [1, 8] stride_h: [1, 8]
	Padding	pad_w: [0, 15] pad_h: [0, 15] constraint: The padded values are counted in any case.
	Input Size	Arbitrary
	Type	Max, Average
Elementwise-sum	Input Size	Arbitrary
Elementwise-sum	Activation	ReLU, Leaky ReLU
Data Reorganization	Kernel Sizes	kernel_w: [1, 16]
	Strides	stride_w: [1, 8]
	Padding	pad_left: [0, 15] pad_right: [0, 15]
	Input Size	input_channel: [1, 32] input_w: [1, 3840] input_h: [1, 2160]

DPUCADF8H Configuration - 1.0 English - PG400