SUPPORTED_LAYERS.md

CHaiDNN-v2
Analysis and Eval
	Supported Layers		Performance/Resource Utilization
	Performance Eval
Design and Development
	API Reference	Quantization User Guide for CHaiDNN	Model Zoo	Running Inference on new Network
	Creating SDx GUI Project	Configurable Parameters	Custom Platform Generation	Software Layer Plugin
	SDSoC Environment User Guide		Hardware-Software Partitioning for Performance

DNN Layer Support

Supported Layers
Convolution	BatchNorm	Power	Scale
Deconvolution*	ReLU	Pooling(Max, Avg)	InnerProduct
Dropout	Softmax	Crop	Concat
Permute	Normalize(L2 Norm)	Argmax	Flatten
PriorBox	Reshape	NMS	Eltwise
CReLU**	Depthwise Separable Convolution	Software Layer Plugin***	Input/ Data
Dilated Convolution

^{* It performs combined Deconvolution+Argmax operation.}

^{** refers CReLU supported as a composition operation, i.e., Concat(Convolution, Power(Convolution, -1)), where Power(Convolution, -1) is expected to perform invert operation by multiplying input with -1.}

^{*** refers to CHai Software-layer-Plugin}

Hardware Accelerated Layers

The following table describes the hardware accelerated layers.

Layer Name	Hardware Kernel	Notes/Restrictions
Convolution	Convolution	Filter sizes: 1x1, 3x3, 5x5, 7x7, 11x11. Horizontal and vertical strides must be same. Number of Input and output feature maps must be less than 4096.
Dilated Convolution	Convolution	Dilation factor: 6
Batch Normalization	Convolution	Number of Input and output feature maps must be less than 2048.
Scale and Bias	Convolution	Number of Input and output feature maps must be less than 2048.
Element-wise addition	Convolution
Pooling (Max, Average)	Convolution/Pool	Number of Input and output feature maps must be less than 4096.
Deconvolution	Deconvolution	16-bit only. It performs only Deconvolution + Argmax combined. Standalone deconvolution output won't be available.
Depthwise Separable Convolution	Convolution	Number of Input and output feature maps must be less than 4096.
ReLU	Convolution	ReLU operation is performed in-line with other supported operations. The fusion of ReLU is supported for the below Layers: Convolution, Dilated Convolution, Batch Normalization, Scale and Bias, 3D separable Convolution, Element-wise Addition

Software Optimized Layers

The following table describes the software optimized layers.

Layer Name	Software Kernel	Notes/Restrictions
L2-Normalization	Norm	This layer works if it lies between two Hardware convolution layers (as present in VGGSSD network).
Permute	Permute	Input is in packed format. This works for the order=[0,2,3,1] only (as present in VGGSSD network).
Inner Product	CBLAS GEMV	Using CBLAS library function.
Softmax	Softmax
NMS	NMS	Max box count 200

---

^{Copyright© 2018 Xilinx}