Supported Operators

This document provides a comprehensive list of operators supported by the Synaptics backend within the IREE compiler stack. It details which operations are implemented and verified across various input/output types, data layouts, and tiling strategies. The table serves as a reference for users to understand backend coverage and integration status for TOSA and Linalg-based models targeting Synaptics NPU.

Activation related Operators

Op	Input	TorqHL	in_t	weight_t	out_t	Tiling	Notes
Abs	Tosa	Act(ABS)	bf16	n/a	bf16	generic
			f32	n/a	f32	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
			i8	n/a	i8	generic
Cast	Tosa	Act(ACT)	bf16	n/a	f32	generic
		Act(F2I)	bf16	n/a	i16	generic
		Act(F2I)	bf16	n/a	i32	generic
		Act(F2I)	bf16	n/a	i8	generic
		Act(ACT)	f32	n/a	bf16	generic
		Act(F2I)	f32	n/a	i16	generic
		Act(F2I)	f32	n/a	i32	generic
		Act(F2I)	f32	n/a	i8	generic
		Act(I2F)	i16	n/a	bf16	generic
		Act(I2F)	i16	n/a	f32	generic
		Act(ACT)	i16	n/a	i32	generic
		Act(ACT)	i16	n/a	i8	generic
		Act(I2F)	i32	n/a	bf16	generic
		Act(I2F)	i32	n/a	f32	generic
		Act(I2F)	i32	n/a	f32	generic
		Act(I2I)	i32	n/a	i16	generic
		Act(I2I)	i32	n/a	i8	generic
		Act(I2F)	i8	n/a	bf16	generic
		Act(I2F)	i8	n/a	f32	generic
		Act(I2I)	i8	n/a	i16	generic
		Act(I2I)	i8	n/a	i32	generic
		Act(I2I)	i1	n/a	i16	generic
Ceil	Tosa	Act(CEL)	bf16	n/a	bf16	generic
			f32	n/a	f32	generic
Clamp	Linalg	Act(CLP)	bf16	n/a	bf16	generic
			f32	n/a	f32	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
			i8	n/a	i8	generic
Clz	Clz	Act(CLZ)	i32	n/a	int32	generic
Floor	Tosa	Act(FLR)	bf16	n/a	bf16	generic
			f32	n/a	f32	generic
Negate	Tosa	Act(NEG)	bf16	n/a	bf16	generic
			f32	n/a	f32	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
			i8	n/a	i8	generic

Element-Wise Operators

Op	Input	TorqHL	in_t	weight_t	out_t	Tiling	Notes
ElementWiseBinary	Arith	EltwiseBin(BAND)	i8	n/a	i8	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
		EltwiseBin(BOR)	i8	n/a	i8	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
		EltwiseBin(BXOR)	i8	n/a	i8	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
		EltwiseBin(AND)	bool	n/a	bool	generic
		EltwiseBin(OR)	bool	n/a	bool	generic
		EltwiseBin(XOR)	bool	n/a	bool	generic
		EltwiseBin(EQ)	i8	n/a	bool	generic
			i16	n/a	bool	generic
			i32	n/a	bool	generic
			bf16	n/a	bool	generic
		EltwiseBin(GT)	i8	n/a	bool	generic
			i16	n/a	bool	generic
			i32	n/a	bool	generic
			bf16	n/a	bool	generic
		EltwiseBin(GT)	i16	n/a	bool	generic
		EltwiseBin(GT)	i16	n/a	bool	generic
		EltwiseBin(GTEQ)	i8	n/a	bool	generic
			i16	n/a	bool	generic
			i32	n/a	bool	generic
			bf16	n/a	bool	generic
		EltwiseBin(MAX)	i8	n/a	i8	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
			bf16	n/a	bf16	generic
		EltwisBin(MIN)	i8	n/a	i8	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
			bf16	n/a	bf16	generic
ElementWiseUnary	Arith	EltwiseUnary(BNOT)	i8	n/a	i8	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
		EltwiseUnary(NOT)	bool	n/a	bool	generic

Composite Operators

Op	Input	TorqHL	in_t	weight_t	out_t	Tiling	Notes
Reciprocal	Linalg		bf16	n/a	bf16	generic
Div	Arith		bf16	n/a	bf16	generic
Erf	Math		bf16	n/a	bf16	generic
QuantizedBatchMatMul	Linalg		i8	n/a	i32	generic

Other Operators

Op	Input	TorqHL	in_t	weight_t	out_t	Tiling	Notes
Add	Linalg	Add	i8	n/a	i8	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
			bf16	n/a	bf16	generic
Bitcast	Arith	Identity	i16	n/a	bf16	No
			bf16	n/a	i16	No
Sub	Linalg	Add	i8	n/a	i8	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
			bf16	n/a	bf16	generic
ArgMax	TOSA	ArgMax	i8	i8	i8	No
AvgPool	TOSA	AvgPool	i8	i8	i8	No
Conv1D	Linalg	Conv1D	bf16	bf16	bf16	custom
Conv2D	Linalg	Conv2D	i8	i8	i8	custom
	Linalg	Conv2D	i16	i8	i32	custom	TOSA specifies i48 output
	Linalg	Conv2D	bf16	bf16	bf16	custom
DepthwiseConv2D	TOSA	DW	i8	i8	i8	custom
Fill	Linalg	Fill	any	n/a	any	generic
FullyConnected	Linalg	FC	i8	i8	i8	custom
Gather	TOSA	Gather	i8	i8	i8	No
		Gather	i16	i16	i16	No
		Gather	i32	i32	i32	No
	Linalg	Gather	i16	i16	i16	No
Identity	TOSA	Identity	any	n/a	any	No
MatMul	Linalg	MatMul	i8	i8	i32	generic
	Linalg	MatMul	i16	i16	i32	generic	TOSA specifies i48 output
	Linalg	MatMul	bf16	bf16	f32	generic
Dot	Linalg	MatMul	i8	i8	i16	generic
	Linalg	MatMul	i16	i16	i16	generic
	Linalg	MatMul	bf16	bf16	f32	generic
MatVec	Linalg	MatMul	i8	i8	i16	generic
	Linalg	MatMul	i16	i16	i16	generic
	Linalg	MatMul	bf16	bf16	f32	generic
BatchMatMul	Linalg	MatMul	i8	i8	i32	generic
MaxPool2D	Linalg	MaxPool2D	i8	i8	i8	custom
	Linalg	MaxPool2D	i16	i8	i16	custom
Mul	Linalg	Mul	i8	n/a	i8	generic
	Linalg	Mul	i16	n/a	i16	generic
	Linalg	Mul	bf16	n/a	bf16	generic
Reduce	Linalg	Reduce(all)	i8	n/a	i8	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
		Reduce(any)	i8	n/a	i8	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
		Reduce(Max)	i8	n/a	i8	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
			bf16	n/a	bf16	generic
		Reduce(Min)	i8	n/a	i8	generic
			i16	n/a	i16	generic
			i32	n/a	i32	generic
			bf16	n/a	bf16	generic
		Reduce(Sum)	i32	n/a	i32	generic
		Reduce(Mul)	bf16	n/a	bf16	generic
		Reduce(XOR)	i16	n/a	i16	generic
			i32	n/a	i32	generic
Reshape-collapse	Linalg	Identity	i8	n/a	i8	No
Reshape-expand	Linalg	Identity	i8	n/a	i8	No
Rescale	Linalg	FMA	ui8	ui8	i8	generic
	Linalg	FMA	i8	i8	ui8	generic
	Linalg	FMA	i8	i16	i16	generic
	Linalg	FMA	i8	i16	i32	generic
	Linalg	FMA	i16	i16	i8	generic
	Linalg	FMA	i32	i8	i8	generic
	Linalg	FMA	i32	i8	i16	generic
	Linalg	FMA	i32	i8	i32	generic
	Linalg	FMA	i16	i16	i32	generic
Resize	TOSA	ResizeNearestNeighbor	i8	i16	i8	No	block size of 2
Scale-and-Add	Linalg	Add	i8	i16	i8	custom
		Add	i16	i16	i16	custom
Scale-and-Sub	Linalg	Add	i8	i16	i8	custom
		Add	i16	i16	i16	custom
Scatter	TOSA	SynpHL	i8	i8	i8	No
Table	TOSA	Table	i8	n/a	i8	No	Default is Table based
Table	TOSA	Table	i16	n/a	i32	No
Transpose	Linalg	Transpose	i8	i8	i8	No
Broadcast	Linalg	Broadcast	*	n/a	*	generic	*: i8/i16/i32/bf16/f32 supported