CPU Guidelines

General Restrictions

Data layout supports NHWC only.
Dynamic weights on Conv2D and FullyConnected layers are unsupported.
Dynamic tensor shape is unsupported.

Supported OPs Specification

OP Name	TFLite OP	NNAPI	Data Type	Restrictions
Abs	ABS	ABS	FP32 FP16
ArgMax ArgMin	ARG_MAX ARG_MIN	ARGMAX ARGMIN	FP32 FP16 UINT8
AvgPooling	AVERAGE_POOL_2D	AVERAGE_POOL_2D	FP32 UINT8
BatchToSpace	BATCH_TO_SPACE_ND	BATCH_TO_SPACE_ND	FP32 FP16 UINT8
Concat	CONCATENATION	CONCATENATION	FP32 UINT8
Conv2D	CONV_2D	CONV_2D	FP32 FP16 UINT8
DepthToSpace	DEPTH_TO_SPACE	DEPTH_TO_SPACE	FP32 FP16 UINT8	Block size should be greater than 1. Requant is unsupported.
DepthwiseConv2D	DEPTHWISE_CONV_2D	DEPTHWISE_CONV_2D	FP32 FP16 UINT8	Requires constant filter and bias.
Dequantize	DEQUANTIZE	DEQUANTIZE	Input - UINT8 Output - FP32	Per channel quantization is unsupported.
ElementWiseAdd	ADD	ADD	FP32 FP16 UINT8
ElementWiseDiv	DIV	DIV	FP32 FP16 UINT8
ElementWiseMul	MUL	MUL	FP32 FP16 UINT8
ElementWiseSub	SUB	SUB	FP32 FP16 UINT8
Elu	ELU	ELU	FP32 FP16
Equal	EQUAL	EQUAL	FP32 FP16 UINT8 INT32
FullyConnected	FULLY_CONNECTED	FULLY_CONNECTED	FP32 FP16 UINT8	Requires constant filter and bias. Per channel quantization is unsupported.
Greater	GREATER	GREATER	FP32 FP16 UINT8 INT32
GreaterEqual	GREATER_EQUAL	GREATER_EQUAL	FP32 FP16 UINT8 INT32
GroupConv2D	Composite pattern of CONV_2D	GROUPED_CONV_2D	FP32 FP16 UINT8
HardSwish	HARD_SWISH	HARD_SWISH	FP32 FP16 UINT8
InstanceNorm		INSTANCE_NORMALIZATION	FP32	Only support 4D.
Less	LESS	LESS	FP32 FP16 UINT8 INT32
LessEqual	LESS_EQUAL	LESS_EQUAL	FP32 FP16 UINT8 INT32
LogicalAnd LogicalOr LogicalNot	LOGICAL_AND LOGICAL_OR LOGICAL_NOT	LOGICAL_AND LOGICAL_OR LOGICAL_NOT	BOOL
L2Norm	L2_NORMALIZATION	L2_NORMALIZATION	FP32	Only support 4D. Only the last dimension of the input tensor is supported as the axis.
L2Pooling	L2_POOL_2D	L2_POOL_2D	FP32
MaxPooling	MAX_POOL_2D	MAX_POOL_2D	FP32 UINT8
Maximum	MAXIMUM	MAXIMUM	FP32 FP16 UINT8
Mean	MEAN	MEAN	FP32 FP16 UINT8
Minimum	MINIMUM	MINIMUM	FP32 FP16 UINT8
Neg	NEG	NEG	FP32 FP16
NotEqual	NOT_EQUAL	NOT_EQUAL	FP32 FP16 UINT8 INT32
Pad	PAD PADV2	PAD PAD_V2	FP32 FP16 UINT8	Support pad with non-zero value.
PRelu	PRELU	PRELU	FP32 FP16 UINT8	Per channel quantization is unsupported.
QLSTM	LSTM	QUANTIZED_16BIT_LSTM	UINT8	Only support constant weight.
Quantize	QUANTIZE	QUANTIZE	Input - FP32 Output - UINT8
ReLU ReLU1 ReLU6	RELU RELU_N1_TO_1 RELU6	RELU RELU1 RELU6	FP32 FP16 UINT8
Reshape	RESHAPE	RESHAPE	FP32 FP16 UINT8	Requantization is unsupported.
Resize::BILINEAR Resize::NEAREST	RESIZE_BILINEAR RESIZE_NEAREST_NEIGHBOR	RESIZE_BILINEAR RESIZE_NEAREST_NEIGHBOR	FP32 FP16 UINT8	Requantization is unsupported. Unsupported align corners and half pixel centers.
RSqrt	RSQRT	RSQRT	FP32 FP16
Sigmoid	LOGISTIC	LOGISTIC	FP32 FP16 UINT8
SoftMax	SOFTMAX	SOFTMAX	FP32 UINT8	Only the last dimension of the input tensor is supported as the axis.
SpaceToBatch	SPACE_TO_BATCH_ND	SPACE_TO_BATCH_ND	FP32 FP16 UINT8	Block size should be greater than 1.
SpaceToDepth	SPACE_TO_DEPTH	SPACE_TO_DEPTH	FP32 FP16 UINT8	Block size should be greater than 1.
Sqrt	SQRT	SQRT	FP32 FP16
StridedSlice	STRIDED_SLICE	STRIDED_SLICE	FP32 FP16 UINT8	Rank greater than 4 is not supported.
Tanh	TANH	TANH	FP32 FP16 UINT8
Transpose	TRANSPOSE	TRANSPOSE	FP32 FP16 UINT8	Rank 4 input tensors supported only.
TransposeConv2D	TRANSPOSE_CONV	TRANSPOSE_CONV_2D	FP32 FP16 UINT8	Requires constant filter and bias.