ElementWise

Computes a per-element binary operation between two input tensors to produce an output tensor. When applicable, broadcasting is used (refer to Shape Information for more information).

Attributes

operation ElementWise operation can be one of:

• SUM \(output=input1+input2\)

• PROD \(output=input1*input2\)

• MAX \(output=max(input1,input2)\)

• MIN \(output=min(input1,input2)\)

• SUB \(output=input1-input2\)

• DIV \(output=\frac{input1}{input2}\)

• POWER \(output=input1^{input2}\)

• FLOOR_DIV \(output=\lfloor\frac{a}{b}\rfloor\)

• AND \(output=and(input1,input2)\)

• OR \(output=or(input1,input2)\)

• XOR \(output=xor(input1,input2)\)

• EQUAL \(output=(input1==input2)\)

• GREATER \(output=(input1>input2)\)

• LESS \(output=(input1<input2)\)

Inputs

input1: tensor of type T1

input2: tensor of type T1

Outputs

output: tensor of type T2

Data Types

When not specified explicitly, T1==T2

Operation	T1	T2
SUM	int8, int32, int64, float16, float32, bfloat16
PROD	int8, int32, int64, float16, float32, bfloat16
MAX	int8, int32, int64, float16, float32, bfloat16
MIN	int8, int32, int64, float16, float32, bfloat16
SUB	int8, int32, int64, float16, float32, bfloat16
DIV	int8, int32, int64, float16, float32, bfloat16
POWER	int8, float16, float32, bfloat16
FLOOR_DIV	int8, int32, int64, float16, float32, bfloat16
AND	bool
OR	bool
XOR	bool
EQUAL	int32, int64, float16, float32, bfloat16		bool
GREATER	int32, int64, float16, float32, bfloat16		bool
LESS	int32, int64, float16, float32, bfloat16		bool

Shape Information

Inputs must have the same rank. For each dimension, their lengths must match, or one of them must be equal to 1. In the latter case, the tensor is broadcast along that axis.

The output has the same rank as the inputs. For each output dimension, its length is equal to the lengths of the corresponding input dimensions if they match, otherwise it is equal to the length that is not 1.

DLA Support

DLA supports the following operation types under specific conditions:

• POW where exactly one of the inputs is a constant

• DIV where at most one of the inputs is a constant

• EQUAL where at most one of the inputs is a constant

• GREATER where at most one of the inputs is a constant

• LESS where at most one of the inputs is a constant

Comparison operations (EQUAL, GREATER, and LESS) support broadcasting. However, to be considered broadcast compatible on DLA, it is necessary for one shape to have one of the following configurations: NCHW (i.e. shapes equal), NC11 (i.e. N and C equal, H and W are 1), or N111 (i.e. N equal, C, H, and W are 1).

DLA also supports “greater or equal” and “less or equal” when both GREATER/LESS and EQUAL have the same input.

Examples

ElementWise With Broadacast

in1 = network.add_input("input1", dtype=trt.float32, shape=(2, 3))
in2 = network.add_input("input2", dtype=trt.float32, shape=(1, 3))
layer = network.add_elementwise(in1, in2, op=trt.ElementWiseOperation.PROD)
network.mark_output(layer.get_output(0))

inputs[in1.name] = np.array([[-3.0, -2.0, -1.0], [0.0, 1.0, 2.0]])
inputs[in2.name] = np.array([[4.0, 5.0, 6.0]])

outputs[layer.get_output(0).name] = layer.get_output(0).shape

expected[layer.get_output(0).name] = np.array([[-12.0, -10.0, -6.0], [0.0, 5.0, 12.0]])

C++ API

For more information about the C++ IElementWiseLayer operator, refer to the C++ IElementWiseLayer documentation.

Python API

For more information about the Python IElementWiseLayer operator, refer to the Python IElementWiseLayer documentation.