#HBIR Operator Definition

#hbir.abs (::mlir::hbdk::hbir::AbsOp)

HBIR tensor abs.

Return tensor after the operation, which has the same shape as the input. Traits: CommonVerifier, EltwiseLike, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.add (::mlir::hbdk::hbir::AddOp)

HBIR tensor addition.

Applies addition operator element-wise, $y_i=lhs_i+rhs_i$.

---

Note:

• Our arithmetic operator support broadcast.

---

Prototype: Pytorch add.

Traits: Broadcastable, CommonVerifier, EltwiseLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
lhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type
rhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16, int32, if type is int32, this hbir.add op must be fusible to a Conv op Shape: [*]
rhs	Same as lhs
output	Same as lhs

#J6 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16, int32, if type is int32, this hbir.add op must be fusible to a Conv op Shape: [*]
rhs	Same as lhs
output	Same as lhs

#hbir.asin (::mlir::hbdk::hbir::AsinOp)

HBIR tensor asin.

Return tensor after the operation, which has the same shape as the input. Traits: CommonVerifier, LutLike, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.atan (::mlir::hbdk::hbir::AtanOp)

HBIR tensor atan.

Return tensor after the operation, which has the same shape as the input. Traits: CommonVerifier, LutLike, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.avg_pool2d (::mlir::hbdk::hbir::AvgPool2dOp)

HBIR 2-D average pooling.

Applies a 2D average pooling over input.

In the simplest case, the output value of the operator with input size $(N, H, W, C)$, output $(N, H_{out}, W_{out}, C)$ and kernel size $(ker_{h}, ker_{w})$ can be precisely described as:

$out(N_i, h, w, C_j) = \frac{1} { ker_h *ker_w } \sum_{m = 0} ^ { ker_h - 1 }\sum_{n = 0} ^ { ker_w - 1 } input(N_i, stride[0]\times h + m, stride[1] \times w + n, C_j)$

where $h,w$ respectively represent the size of H and W.

---

Note:

• parameters has the same manner as the Conv2D operator, the same goes for the output size.

• ceilMode controls output's compute is mode of floor or ceil, it's default value is false.

---

Shape:

• Input: $(N, H_{in}, W_{in}, C)$ or $(H_{in}, W_{in}, C)$ or $(*, H_{in}, W_{in})$

• Output: $(N, H_{out}, W_{out}, C)$ or $(H_{out}, W_{out}, C)$ or $(*, H_{out}, W_{out}, C)$, where $*$ represents any number of dimension.

$H_{out} = \lfloor {\frac{H_{in} + padding[0] + padding[2] - kernel[0]} {stride[0]} + 1}\rfloor$

$W_{out} = \lfloor {\frac{W_{in} + padding[1] + padding[3] - kernel[1]} {stride[1]} + 1}\rfloor$

if ceilMode = true, please use ceil replace floor in the above output formula.

---

Prototype: Pytorch avg_pool2d.

Traits: CommonVerifier, PoolLike, StencilLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
kernel	::mlir::ArrayAttr	64-bit integer array attribute with exactly 2 elements
stride	::mlir::ArrayAttr	64-bit integer array attribute with exactly 2 elements
pad	::mlir::ArrayAttr	64-bit integer array attribute with exactly 4 elements
dilation	::mlir::ArrayAttr	64-bit integer array attribute with exactly 2 elements
ceilMode	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*,H,W,C]
output	Same as input
kernel	Shape: [KH,KW] Dim: KH, KW ∈ [1, 256]
stride	Shape: [SH,SW] Dim: SH, SW ∈ [1, 256]
pad	Shape: [PN,PH,PW,PC] PN,PH,PW,PC ∈ [-3, 256]

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*,H,W,C]
output	Same as input
kernel	Shape: [KH,KW] Dim: KH, KW ∈ [1, 256]
stride	Shape: [SH,SW] Dim: SH, SW ∈ [1, 256]
pad	Shape: [PN,PH,PW,PC] PN,PH,PW,PC ∈ [-3, 256]

#hbir.batchnorm (::mlir::hbdk::hbir::BatchNormOp)

Hbir Batch Normalize

Applies Batch Normalization over each dimension of input. This compute can be precisely described as:

$y = \frac{x-mean[x]}{\sqrt{Var[x]+\epsilon}}*weight+bias$

This mean and standard-deviation are calculated per-dimension over the batches and weight and bias are learnable parameter vectors of the input size.

---

Note:

• eps - a value added to the denominator for numerical stability.
• $mean(x)$ and $Var[x]$'s shape are $(C)$.
• weight and bias are learnable scalar.

---

Shape:

• Input: $(N,H,W,C)$ or $(N,M,H,W,C)$ or $(H,W,C)$ or $(*,H,W,C)$, where $*$ reprensent any number of dimension.
• Output: same shape as input.

---

Prototype: Pytorch BatchNorm.

Traits: CommonVerifier, Misc, SameVariadicOperandSize

Interfaces: HBTLExecutable, HbdkExecutorInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, RoiInfer, SchedInterface, SchedTemp, ShapeInference, Tiling

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
eps	::mlir::FloatAttr	64-bit float attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values
mean	1D tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values
var	1D tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values
weight	1D tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values
bias	1D tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*,H,W,C]
mean	Type: f32 Shape: [C]
var	Type: f32 Shape: [C]
weight	Type: f32 Shape: [C]
bias	Type: f32 Shape: [C]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*,H,W,C]
mean	Type: f32 Shape: [C]
var	Type: f32 Shape: [C]
weight	Type: f32 Shape: [C]
bias	Type: f32 Shape: [C]
output	Same as input

#hbir.cast_type (::mlir::hbdk::hbir::CastTypeOp)

elemental type cast operation

Data are actually moved. Traits: CommonVerifier, Misc, NaiveRoiInfer, NaiveTiling, SameOperandsAndResultShape

Interfaces: CastOpInterface, HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), NonBatchAxesInfer, Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
forceSaturate	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16, bool8 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16, bool8 Shape: [*]
output	Same as input

#hbir.clip (::mlir::hbdk::hbir::ClipOp)

HBIR Clip Op.

Clamps all elements in input into the range $[min, max]$.Assume min_value and max_value be min and max, respectively, this performs:

$y_i = min(max(x_i, min\_value_i), max\_value_i)$

---

Note:

• min(min_value): lower-bound of the range to be clamped to
• max(max_value): upper-bound of the range to be clamped to

---

Shape:

• Input: (*), where * means any number of dimensions.
• Output: (*), same shape as the input.

---

Prototype: Pytorch hardtanh.

Traits: CommonVerifier, LutLike, NaiveRoiInfer, NaiveTiling, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
min	::mlir::Attribute	64-bit float attribute or 64-bit signless integer attribute
max	::mlir::Attribute	64-bit float attribute or 64-bit signless integer attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.concat (::mlir::hbdk::hbir::ConcatOp)

Concatenates tensors along one dimension

Concatenates the given sequence of seq tensors in the given dimension. No elemental type conversion.

---

Note:

• dim - the dimension over which the tensors are concatenated.

---

Prototype: Pytorch cat.

Traits: CommonVerifier, MoveLike, NaiveTiling, SameElementType

Interfaces: HBTLExecutable, HbdkExecutorInterface, Layout, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), NonBatchAxesInfer, Perf, PortAccess, Quantizable, RoiInfer, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
dim	::mlir::IntegerAttr	64-bit signless integer attribute

#Operands:

Operand	Description
inputs	variadic of tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
inputs	Arg Number: inputs number ∈ [1, 1024] Dim: all dims < 131072 size < 2G
output	same as inputs

#J6 BPU

Features and Attributes	Constraints
inputs	Arg Number: inputs number ∈ [1, 1024] Dim: all dims < 131072 size < 2G
output	same as inputs

#hbir.constant (::mlir::hbdk::hbir::ConstantOp)

HBIR constant generation op.

Generate a constant with specified type and value Traits: CommonVerifier, Constant

Interfaces: HBTLExecutable, HbdkExecutorInterface, NoMemoryEffect (MemoryEffectOpInterface), NonBatchAxesInfer, Perf, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
values	::mlir::ElementsAttr	constant vector/tensor attribute

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

None

#J6 BPU

None

#hbir.conv2d (::mlir::hbdk::hbir::Conv2dOp)

HBIR 2-D convolution.

Applies a 2D convolution over an input signal composed of several input channels.

In the simplest case, the output value of the layer with input size $(N, H_{in}, W_{in}, C_{in})$ and output $(N, H_{out}, W_{out}, C_{out})$ can be precisely descibed as:

$out(N_i,C_{out_j}) = bias(C_{out_j}) + \sum_{k=0}^{C_{in} - 1}weight(C_{out_j},k) \star input(N_i,k)$

where $\star$ is the valid 2D cross-correlation operation, $N$ is the batch size, $C$ denotes a number of channels, $H$ and $W$ are the size of pixels.

---

Note:

• stride controls the stride for the cross-correlation, an integer array with 2 elements, default value is (1,1).

• padding controls the amount of padding applied to the input, an integer array with 4 elements, the padding sequences is (h_begin,w_begin,h_end,w_end), default value is (0,0,0,0).

• dilation controls the spacing between kernel points, an integer array with 2 elements, default value is (0,0). It's harder to describe, but this link has a nice visualization of what dilation does.

• groups controls the connections between inputs and outputs, an integer variable, default value is 1.

• Weight: $(C_{out}, KH, KW, C_{in})$, bias shape = $C_{out}$ where KW and KH represent kernel's height and width, respectively.

---

Shape:

• Input: $(N,H_{in},W_{in},C_{in})$ or $(H_{in},W_{in},C_{in})$ or $(N,M,H_{in},W_{in},C_{in})$ or $(*,H_{in},W_{in},C_{in})$, where * represent any number of dimension.

• Output: $(N,H_{out},W_{out},C_{out})$ or $(H_{out},W{out},C_{out})$ or $(N,M,H_{out},W_{out},C_{out})$ or $(*,H_{out},W_{out},C_{out})$

$H_{out}=\lfloor \frac{H_{in} + padding[0] + padding[2] - dilation[0]\times(kernel[0]-1)-1}{stride[0]}+1\rfloor$ $W_{out}=\lfloor \frac{W_{in}+padding[1]+padding[3]-dilation[1]\times(kernel[1]-1)-1}{stride[1]}+1\rfloor$

---

Prototype: Pytorch convolution.

Traits: CommonVerifier, ConvLike, StencilLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
stride	::mlir::ArrayAttr	64-bit integer array attribute with exactly 2 elements
pad	::mlir::ArrayAttr	64-bit integer array attribute with exactly 4 elements
dilation	::mlir::ArrayAttr	64-bit integer array attribute with exactly 2 elements
groupNum	::mlir::IntegerAttr	64-bit signless integer attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type
weight	4D tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type
bias	1D tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16; input and weight cannot both be int16 Shape: [*,H,W,C] Dim: * ∈ [1, 4096]; H,W,C ∈ [1, 65536]
weight	Type: int8, int16; input and weight cannot both be int16 Shape: [N,KH,KW,C] Dim: C ∈ [1, 8192]; KH,KW ∈ [1, 31]; N ∈ [1, 65536] if fout is the last layer of conv else [1, 8192] Size: KH*KW*C ∈ [1, 65536]
bias	Type: f32
output	Type: int8, int16, int32 Other constraints: same as fin
stride	Shape: [SH,SW] Dim: SH,SW ∈ [1, 256]; SH,SW ∈ 1 if dilation > 1
pad	Shape: [PN,PH,PW,PC] Dim: PN,PH,PW,PC ∈ [-1, 256]
groupNum	fin.c is divisible by group number
dilation	Shape: [DH,DW] Dim: DH,DW ∈ [1, 18]
others	Don't support even stride when conv is a depthwise conv For each group, fin.c ∈ [1, 8192], KH*KW*fin.c ∈ [1, 65535], fin.c = C when group = 1

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16; input and weight cannot both be int16 Shape: [*,H,W,C] Dim: * ∈ [1, 4096]; H,W,C ∈ [1, 65536]
weight	Type: int8, int16; input and weight cannot both be int16 Shape: [N,KH,KW,C] Dim: C ∈ [1, 8192]; KH,KW ∈ [1, 31]; N ∈ [1, 65536] if fout is the last layer of conv else [1, 8192] Size: KH*KW*C ∈ [1, 65536]
bias	Type: f32
output	Type: int8, int16, int32 Other constraints: same as fin
stride	Shape: [SH,SW] Dim: SH,SW ∈ [1, 256]; SH,SW ∈ 1 if dilation > 1
pad	Shape: [PN,PH,PW,PC] Dim: PN,PH,PW,PC ∈ [-1, 256]
groupNum	fin.c is divisible by group number
dilation	Shape: [DH,DW] Dim: DH,DW ∈ [1, 18]
others	Don't support even stride when conv is a depthwise conv For each group, fin.c ∈ [1, 8192], KH*KW*fin.c ∈ [1, 65535], fin.c = C when group = 1

#hbir.conv2dtranspose (::mlir::hbdk::hbir::Conv2dTransposeOp)

HBIR transposed conv2d op.

Inverse operation of Conv2dTranspose in shape.

---

Parameters:

• input: the input tensor
• weight: the deconvolution kernel
• stride: same as conv2d's stride, [s_h, s_w]
• pad: pad information for clipping output [h_top,w_left,h_bottom,w_right]
• dilation: same as conv2d's dilation, [d_h, d_w]
• group: same as conv2d's group

---

Shape:

• input: $(*, h, w, in\_channel)$
• weight: $(in\_channel, kh, kw, out\_channel / group)$
• output: $(*, ho, wo, out\_channel)$
• bias: $out\_channel$

where:

$ho = (h - 1) * stride[0] - (pad[0] + pad[2]) + dilation[0] * (kh - 1) + 1$ $wo = (w - 1) * stride[1] - (pad[1] + pad[3]) + dilation[1] * (kw - 1) + 1$

Traits: CommonVerifier, ConvLike, StencilLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
stride	::mlir::ArrayAttr	64-bit integer array attribute with exactly 2 elements
pad	::mlir::ArrayAttr	64-bit integer array attribute with exactly 4 elements
dilation	::mlir::ArrayAttr	64-bit integer array attribute with exactly 2 elements
groupNum	::mlir::IntegerAttr	64-bit signless integer attribute
illegalWeight	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type
weight	4D tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type
bias	1D tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16; input and weight cannot both be int16 Shape: [*,H,W,C] Dim: * ∈ [1, 128]; H,W ∈ [1, 65536]; C ∈ [1, 2048]
weight	Type: int8, int16; input and weight cannot both be int16 Shape: [N,KH,KW,C] Dim: N,C ∈ [1, 2048]; KH,KW ∈ [1, 14] Size: KH*KW*C ∈ [1, 65536]
bias	Type: f32
output	Same as input, the type additionally supports int32
stride	Shape: [SH,SW] Dim: SH,SW ∈ [1, 14]; SH < KH; SW < KW;
pad	Shape: [PN,PH,PW,PC] Dim: PN,PH,PW,PC ∈ [0, 256]
dilation	Shape: [DH,DW] Dim: DH,DW ∈ 1

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16; input and weight cannot both be int16 Shape: [*,H,W,C] Dim: * ∈ [1, 128]; H,W ∈ [1, 65536]; C ∈ [1, 2048]
weight	Type: int8, int16; input and weight cannot both be int16 Shape: [N,KH,KW,C] Dim: N,C ∈ [1, 2048]; KH,KW ∈ [1, 14] Size: KH*KW*C ∈ [1, 65536]
bias	Type: f32
output	Same as input, the type additionally supports int32
stride	Shape: [SH,SW] Dim: SH,SW ∈ [1, 14]; SH < KH; SW < KW;
pad	Shape: [PN,PH,PW,PC] Dim: PN,PH,PW,PC ∈ [0, 256]
dilation	Shape: [DH,DW] Dim: DH,DW ∈ 1

#hbir.conv3d (::mlir::hbdk::hbir::Conv3dOp)

HBIR 3-D convolution.

Applies a 3D convolution over an input signal composed of several input planes.

In the simplest case, the output value of the layer with input size $(N, C_{in}, D, H, W)$ and output $(N, C_{out}, D_{out}, H_{out}, W_{out})$ can be precisely described as:

$out(N_i, C_{out_j}) = bias(C_{out_j}) + \sum_{k = 0}^{C_{in} - 1} weight(C_{out_j}, k) \star input(N_i, k)$

where $\star$ is the valid 3D [cross-correlation] operation, $N$ is the batch size, $C$ denotes a number of channels, $D$, $H$ and $W$ are the size of pixels.

---

Note:

• stride controls the stride for the cross-correlation, an integer array with 3 elements, default value is (1,1,1).

• padding controls the amount of padding applied to the input, an integer array with 5 elements, the padding sequences is (d_begin,h_begin,w_begin,d_end,h_end,w_end), default value is (0,0,0,0,0,0).

• dilation controls the spacing between kernel points, an integer array with 3 elements, default value is (0,0,0). It's harder to describe, but this link has a nice visualization of what dilation does.

• groups controls the connections between inputs and outputs, an integer variable, default value is 1.

• Weight: $(C_{out}, KD, KH, KW, C_{in})$, bias shape = $C_{out}$ where KW and KH represent kernel's height and width, respectively.

---

Shape:

• Input: $(N,D_{in},H_{in},W_{in},C_{in})$ or $(D_{in},H_{in},W_{in},C_{in})$ or $(*,D_{in},H_{in},W_{in},C_{in})$, where * represent any number of dimension.

• Output: $(N,D_{out},H_{out},W_{out},C_{out})$ or $(D_{out},H_{out},W{out},C_{out})$ or $(*,D_{out},H_{out},W_{out},C_{out})$.

$D_{out}=\lfloor \frac{D_{in} + padding[0] + padding[3] - dilation[0]\times(kernel[0]-1)-1}{stride[0]}+1\rfloor$

$H_{out}=\lfloor \frac{H_{in} + padding[1] + padding[4] - dilation[1]\times(kernel[1]-1)-1}{stride[1]}+1\rfloor$

$W_{out}=\lfloor \frac{W_{in} + padding[2] + padding[5] - dilation[2]\times(kernel[2]-1)-1}{stride[2]}+1\rfloor$

---

Prototype: Pytorch convolution.

Traits: CommonVerifier, ConvLike, StencilLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
stride	::mlir::ArrayAttr	64-bit integer array attribute with exactly 3 elements
pad	::mlir::ArrayAttr	64-bit integer array attribute with exactly 6 elements
dilation	::mlir::ArrayAttr	64-bit integer array attribute with exactly 3 elements
groupNum	::mlir::IntegerAttr	64-bit signless integer attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type
weight	5D tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type
bias	1D tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
lhs	Type: int8 Shape: [*,D,H,W,C] Dim: * ∈ [1, 128]; D,H,W ∈ [1, 65536]; C ∈ [1, 4096];
rhs	Type: int8 Shape: [N,KD,KH,KW,C] N ∈ [1, 65536]; KD,KH,KW ∈ [1, 9]; Dim: C ∈ [1, 4096]; Size: KD*KH*KW*C ∈ [1, 131072]
bias	Type: f32
output	The same as lhs
stride	Shape: [SD,SH,SW] Dim: SD,SH,SW must be 1 or 2 and equal to each other.
pad	Shape: [PN,PD,PH,PW,PC] Dim: PN,PD,PH,PW,PC ∈ [0, 256]
groupNum	fin.c is divisible by group number
dilation	Shape: [DD,DH,DW] DD,DH,DW = 1 when group number more than 1

#J6 BPU

Features and Attributes	Constraints
lhs	Type: int8 Shape: [*,D,H,W,C] Dim: * ∈ [1, 128]; D,H,W ∈ [1, 65536]; C ∈ [1, 4096];
rhs	Type: int8 Shape: [N,KD,KH,KW,C] N ∈ [1, 65536]; KD,KH,KW ∈ [1, 9]; Dim: C ∈ [1, 4096]; Size: KD*KH*KW*C ∈ [1, 131072]
bias	Type: f32
output	The same as lhs
stride	Shape: [SD,SH,SW] Dim: SD,SH,SW must be 1 or 2 and equal to each other.
pad	Shape: [PN,PD,PH,PW,PC] Dim: PN,PD,PH,PW,PC ∈ [0, 256]
groupNum	fin.c is divisible by group number
dilation	Shape: [DD,DH,DW] DD,DH,DW = 1 when group number more than 1

#hbir.conv (::mlir::hbdk::hbir::ConvOp)

HBIR convolution.

Applies a convolution over an input signal composed of several input planes.

rank in convolution is an integer greater than or equal to 1.

In the simplest case, for 2D convolution, rank=2 the output value of the layer with input size $(N, H_{in}, W_{in}, C_{in})$ and output $(N, H_{out}, W_{out}, C_{out})$ can be precisely descibed as:

$out(N_i,C_{out_j}) = bias(C_{out_j}) + \sum_{k=0}^{C_{in} - 1}weight(C_{out_j},k) \star input(N_i,k)$

where $\star$ is the valid 2D cross-correlation operation, $N$ is the batch size, $C$ denotes a number of channels, $H$ and $W$ are the size of pixels.

---

Note:

• stride controls the stride for the cross-correlation, an integer array with rank elements, default value is (1,1,...).

• padding controls the amount of padding applied to the input, an integer array with 2*rank elements, when rank=2, the padding sequences is (h_begin,w_begin,h_end,w_end), default value is (0,0,0,0), when n=3, the padding sequences is (d_begin,h_begin,w_begin,d_end,h_end,w_end), default value is (0,0,0,0,0,0).

• dilation controls the spacing between kernel points, an integer array with rank elements, default value is (0,0,...). It's harder to describe, but this link has a nice visualization of what dilation does.

• groups controls the connections between inputs and outputs, an integer variable, default value is 1.

• Weight for 2D: $(C_{out}, KH, KW, C_{in})$, bias shape = $C_{out}$ where KW and KH represent kernel's height and width, respectively.

---

Shape for 2D:

• Input: $(N,H_{in},W_{in},C_{in})$ or $(H_{in},W_{in},C_{in})$ or $(N,M,H_{in},W_{in},C_{in})$ or $(*,H_{in},W_{in},C_{in})$, where * represent any number of dimension.

• Output: $(N,H_{out},W_{out},C_{out})$ or $(H_{out},W{out},C_{out})$ or $(N,M,H_{out},W_{out},C_{out})$ or $(*,H_{out},W_{out},C_{out})$

$H_{out}=\lfloor \frac{H_{in} + padding[0] + padding[2] - dilation[0]\times(kernel[0]-1)-1}{stride[0]}+1\rfloor$ $W_{out}=\lfloor \frac{W_{in}+padding[1]+padding[3]-dilation[1]\times(kernel[1]-1)-1}{stride[1]}+1\rfloor$

---

Prototype: Pytorch convolution.

Traits: CommonVerifier, ConvLike, StencilLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
stride	::mlir::ArrayAttr	64-bit integer array attribute
pad	::mlir::ArrayAttr	64-bit integer array attribute
dilation	::mlir::ArrayAttr	64-bit integer array attribute
groupNum	::mlir::IntegerAttr	64-bit signless integer attribute
channelLast	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type
weight	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type
bias	1D tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type

#Constraint:

#J5 BPU

If weight's shape is [N,KD,KH,KW,C] look Conv3d's constraints, If weight's shape is [N,KH,KW,C], look Conv2d's constraints

#J6 BPU

If weight's shape is [N,KD,KH,KW,C] look Conv3d's constraints, If weight's shape is [N,KH,KW,C], look Conv2d's constraints

#hbir.correlation (::mlir::hbdk::hbir::CorrelationOp)

HBIR Correlation.

Applies correlation to inputs. The correlation operator calculate the correlation of the regions of the regions of the two inputs.

---

Note:

• kernel controls correlation kernel size.
• max_d is the max_displacement of patch2 relative to patch1.
• stride1 is stride in lhs, it controls H and W in output.
• stride2 is stride in rhs, it controls channels in output.
• pad_size is the padding size.
• is_multiply controls the operation type, true means multiplication, false means subtraction.

---

Shape:

• Input: lhs's shape is $(*,H_{in},W_{in},C_{in})$, and rhs's shape should be same as lhs, where $*$ represents any number of dimension.

• Output: $(*, H_{out}, W_{out}, C_{out})$, $*$ represents input batch size, $H_{out}$, $W_{out}$ and $C_{out}$ are height, width and channel respectively.

$H_{out} = \frac{H_{in} + 2pad - 2(max\_d + \frac{k - 1} {2})} {stride1}$

$W_{out} =\frac{W_{in} + 2pad - 2(max\_d +\frac{k - 1} {2})} {s1}$

$C_{out} = (2 *\frac{max\_d} {stride2} + 1) ^ 2$

Traits: CommonVerifier, Misc, SameOperandsShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, RoiInfer, SchedInterface, SchedTemp, ShapeInference, Tiling

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
kernel	::mlir::IntegerAttr	64-bit signless integer attribute
max_d	::mlir::IntegerAttr	64-bit signless integer attribute
stride1	::mlir::IntegerAttr	64-bit signless integer attribute
stride2	::mlir::IntegerAttr	64-bit signless integer attribute
pad	::mlir::IntegerAttr	64-bit signless integer attribute
mode	::mlir::hbdk::CorrelationModeAttr	correlation mode

#Operands:

Operand	Description
lhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values
rhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
lhs	Type: int8 Shape: [*,H,W,C] Dim: * ∈ [1, 65535]; H,W ∈ [1, 65536]; C ∈ [1, 1024];
rhs	Same as lhs
kernel	The size of kernel must be odd
others	The H, W and C dim of the lhs and rhs of hbir.correlation must be the same The inputH, inputW, kernel, max_d, stride1, stride2 and pad should satisfy : min(inputH, inputW) + (2 * pad) - (2 * (max_d / (stride2 * stride2))) - kernel + 1 >= stride1
output	Same as lhs

#J6 BPU

Features and Attributes	Constraints
lhs	Type: int8 Shape: [*,H,W,C] Dim: * ∈ [1, 65535]; H,W ∈ [1, 65536]; C ∈ [1, 1024];
rhs	Same as lhs
kernel	The size of kernel must be odd
others	The H, W and C dim of the lhs and rhs of hbir.correlation must be the same The inputH, inputW, kernel, max_d, stride1, stride2 and pad should satisfy : min(inputH, inputW) + (2 * pad) - (2 * (max_d / (stride2 * stride2))) - kernel + 1 >= stride1
output	Same as lhs

#hbir.elu (::mlir::hbdk::hbir::ELUOp)

HBIR ELU Op.

Applies Elu funciton element - wise.Elu function defined as:

$Elu(x) = max(0, x) + alpha * (exp(min(0, x)) - 1)$

---

Note:

• alpha - Controls the value to which an ELU saturates for negative net inputs.

---

Shape:

• Input: (*), where * means any number of dimensions.
• Output: (*), same shape as the input.

---

Prototype: Pytorch elu.

Traits: CommonVerifier, LutLike, NaiveRoiInfer, NaiveTiling, SameElementType, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
alpha	::mlir::FloatAttr	64-bit float attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.empty_like (::mlir::hbdk::hbir::EmptyLikeOp)

Hbir sram allocation operation, but not initial the memory

The hbir.empty_like operation allocates a region of sram on the BPU. It is similar to torch.empty_like op.

Example:

Traits: CommonVerifier, Misc

Interfaces: NoMemoryEffect (MemoryEffectOpInterface), SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
dtype	::mlir::TypeAttr	any type attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

None

#J6 BPU

No limits

#hbir.equal (::mlir::hbdk::hbir::EqualOp)

HBIR tensor equal.

Determines whether two tensors are equal element by element - wise, $y_i = (lhs_i == rhs_i)$. Traits: Broadcastable, CommonVerifier, EltwiseLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
lhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values or none type
rhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values or none type

#Results:

Result	Description
output	tensor of values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16 Shape: [*]
rhs	Same as lhs
output	Type: bool8

#J6 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16 Shape: [*]
rhs	Same as lhs
output	Type: bool8

#hbir.exp (::mlir::hbdk::hbir::ExpOp)

HBIR tensor exp.

Applies exponential operator element - wise, $y=e^{x}$.

---

Note:

• Returns a new tensor with the exponential of the elements of the input tensor input.

Traits: CommonVerifier, LutLike, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.fake_cast (::mlir::hbdk::hbir::FakeCastOp)

fake elemental type cast operation

Cast float32 input to specified dtype, and then cast back to float32. Traits: CommonVerifier, Misc, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
dtype	::mlir::TypeAttr	any type attribute

#Operands:

Operand	Description
input	tensor of 32-bit float values or none type

#Results:

Result	Description
output	tensor of 32-bit float values

#Constraint:

#J5 BPU

Unsupported

#J6 BPU

Unsupported

#hbir.flip (::mlir::hbdk::hbir::FlipOp)

HBIR flip

Reverse the order of the input tensor along given axis in dims.

---

Parametes:

• input: the input tensor.
• dims: axis need to reverse.

---

Shape:

• input: $(*)$, where * represents any dimension.
• output: same as the input.

Traits: CommonVerifier, MoveLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
dims	::mlir::ArrayAttr	64-bit integer array attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
inputs	Type: int8
output	same as inputs

#J6 BPU

Features and Attributes	Constraints
inputs	Type: int8
output	same as inputs

#hbir.gather (::mlir::hbdk::hbir::GatherOp)

HBIR gather op for torch.gather.

HBIR gather op for torch.gather. Traits: CommonVerifier, Misc

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, NoMemoryEffect (MemoryEffectOpInterface), NonBatchAxesInfer, Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
dim	::mlir::IntegerAttr	64-bit signless integer attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values or none type
indices	tensor of 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

Unsupported

#J6 BPU

Unsupported

#hbir.grid_sample (::mlir::hbdk::hbir::GridSampleOp)

HBIR grid_sample.

From the input and a flow-field grid, computes the output using input values and pixel locations from the grid.

---

Shape:

• input: input of shape $(*, H_{in}, W_{in}, C_{in})$, where * represent any number of dimension.
• grid: flow - field of shape $(*, H_{out}, W_{out}, 2)$
• output: $(*, H_{out}, W_{out}, C_{in})$

Traits: CommonVerifier, Expansion, Round, SampleLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
mode	::mlir::hbdk::InterpolationModeAttr	interpolation mode for all march
expansionMode	::mlir::hbdk::ExpansionModeAttr	mode to expand input feature on H/W
alignCorner	::mlir::BoolAttr	bool attribute
padValue	::mlir::Attribute	64-bit float attribute or 64-bit signless integer attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values
grid	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8 Shape: [*,H,W,C] Dim: H,W ∈ [1, 1024]; H*W ≤ 720*1024; other dims ∈ [1, 65536]
grid	Type: int16 Shape: [*,H,W,2]
output	Same as input except Dim constraints

#J6 BPU

Features and Attributes	Constraints
input	Type: int8 Shape: [*,H,W,C] Dim: H,W ∈ [1, 1024]; H*W ≤ 720*1024; other dims ∈ [1, 65536]
grid	Type: int16 Shape: [*,H,W,2]
output	Same as input except Dim constraints

#hbir.leaky_relu (::mlir::hbdk::hbir::LeakyReLUOp)

HBIR Leaky ReLU Op.

Applies LeakyRelu funciton element - wise.LeakyRelu function defined as:

$LeakyRelu(x) = max(0, x) + slope * min(0, x)$

---

Note:

• slope - Controls the angle of the negative slope.

---

Shape:

• Input: (*), where * means any number of dimensions.
• Output: (*), same shape as the input.

---

Prototype: Pytorch leaky_relu.

Traits: CommonVerifier, LutLike, NaiveRoiInfer, NaiveTiling, SameElementType, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
slop	::mlir::FloatAttr	64-bit float attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.linear (::mlir::hbdk::hbir::LinearOp)

HBIR Linear.

Applies a linear transformation to the incoming data: $y = xW ^ T + b$

---

Note:

• Weight's shape is $(C_{in},C_{out})$.

---

Shape:

• Input: $(*, C_{in})$ where $*$ represents any number of dimensions and $C_{in}$ = in_features.
• Output: $(*, C_{out})$ where all but the last dimension are the same shape as the input and $C_{out}$ = out_features.

---

Prototype: pytorch linear.

Traits: CommonVerifier, LinearLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values
weight	2D tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values
bias	1D tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8 Shape: [*,C_in] Dim: *, C_in ∈ [1, 65536]
weight	Type: int8 Shape: [C_out, C_in] Dim: C_in,C_out ∈ [1, 8192]; C_in*C_out ∈ [1, 65536]
bias	Type: f32
output	Type: int8, int16, int32 Other constraints: same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8 Shape: [*,C_in] Dim: *, C_in ∈ [1, 65536]
weight	Type: int8 Shape: [C_out, C_in] Dim: C_in,C_out ∈ [1, 8192]; C_in*C_out ∈ [1, 65536]
bias	Type: f32
output	Type: int8, int16, int32 Other constraints: same as input

#hbir.log (::mlir::hbdk::hbir::LogOp)

HBIR tensor log.

Applies natural logarithm operator element - wise, $y = \log_e{(x)}$.

---

Note:

• Returns a new tensor with the natural logarithm of the elements of input.

Traits: CommonVerifier, LutLike, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.logical_and (::mlir::hbdk::hbir::LogicalAndOp)

HBIR tensor and.

Applies 'logical and' operator element - wise, $y_i = lhs_i \&\& rhs_i$. Traits: Broadcastable, CommonVerifier, EltwiseLike, SameOperandsElementType

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
lhs	tensor of 8-bit signed integer or 16-bit signed integer or or 16-bit float or values or none type
rhs	tensor of 8-bit signed integer or 16-bit signed integer or or 16-bit float or values or none type

#Results:

Result	Description
output	tensor of values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16 Shape: [*]
rhs	Same as lhs
output	Type: bool8

#J6 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16 Shape: [*]
rhs	Same as lhs
output	Type: bool8

#hbir.logical_not (::mlir::hbdk::hbir::LogicalNotOp)

HBIR tensor logical not, output bool.

Traits: CommonVerifier, EltwiseLike, SameOperandsAndResultShape, SameOperandsElementType

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
input	tensor of 8-bit signed integer or 16-bit signed integer or 16-bit float or or values or none type

#Results:

Result	Description
output	tensor of values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Type: bool8

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Type: bool8

#hbir.matmul (::mlir::hbdk::hbir::MatMulOp)

HBIR Matrix Multiplication.

Applies matrix multiplication between two inputs: $C = A \times B$, where $\times$ means matrix multiplication.

---

Note:

• If both tensors are 1-dimensional, the dot product (scalar) is returned
• If both arguments are 2-dimensional, the matrix-matrix product is returned
• If the first argument is 1-dimensional and the second argument is 2-dimensional, the matrix-vector product is returned
• If the first argument is 2-dimensional and the second argument is 1-dimensional, the matrix-vector product is returned
• If both arguments are at least 1-dimensional and at least one argument is N-dimensional (where N > 2), then a batched matrix multiply is returned

---

Shape:

• lhs: $(B_{l}, M, C)$, where $B_{l}$ represent any number of dimension.
• rhs: $(B_{r}, C, N)$, where $B_{r}$ represent any number of dimension.
• output: $(B_{o}, M, N)$, where $B_{o}$ represent represent the result of broadcast between $B_{l}$ and $V_{r}$.

---

Prototype: Pytorch Matmul.

Traits: CommonVerifier, MatmulLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
lhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values
rhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16; lhs and rhs cannot both be int16 Shape: [*,M,C] Dim: * ∈ [1, 4096], M,C ∈ [1, 8192]
rhs	Type: int8, int16; lhs and rhs cannot both be int16 Shape: [*,C,N] Dim: * ∈ [1, 4096]; C,N ∈ [1, 8192]
output	Type: int8, int16, int32 Shape: [*,M,N] Other constraints: same as lhs

#J6 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16; lhs and rhs cannot both be int16 Shape: [*,M,C] Dim: * ∈ [1, 4096], M,C ∈ [1, 8192]
rhs	Type: int8, int16; lhs and rhs cannot both be int16 Shape: [*,C,N] Dim: * ∈ [1, 4096]; C,N ∈ [1, 8192]
output	Type: int8, int16, int32 Shape: [*,M,N] Other constraints: same as lhs

#hbir.max (::mlir::hbdk::hbir::MaxOp)

HBIR tensor max.

Applies maximum operator element-wise, $y_i=max(lhs_i,rhs_i)$.

---

Note:

• Our arithmetic operator support broadcast.

Traits: Broadcastable, CommonVerifier, EltwiseLike, SameElementType

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
lhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type
rhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16 Shape: [*]
rhs	Same as lhs
output	Same as lhs

#J6 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16 Shape: [*]
rhs	Same as lhs
output	Same as lhs

#hbir.max_pool2d (::mlir::hbdk::hbir::MaxPool2dOp)

HBIR 2-D max pooling.

Applies a 2D max Pooling over an input.

In the simplest case, the output value of the operator with input size $(N, H, W, C)$, output $(N, H_{out}, W_{out}, C)$ and kernel size $(ker_{h}, ker_{w})$ can be precisely described as:

$out(N_i, h, w, C_j) = \frac{1} { ker_h *ker_w }\max_{m = 0} ^ { ker_h - 1 }\max_{n = 0} ^ { ker_w - 1 } input(N_i, stride[0] \times h + m, stride[1] \times w + n, C_j)$

where $h,w$ respectively represent the size of H and W.

---

Note:

• parameters has the same manner as the Conv2D operator, the same goes for the output size.
• ceilMode controls output 's compute is mode of floor or ceil, it' s default value is false.

---

Shape:

• Input: $(N, H_{in}, W_{in}, C)$ or $(H_{in}, W_{in}, C)$ or $(*, H_{in}, W_{in})$

• Output: $(N, H_{out}, W_{out}, C)$ or $(H_{out}, W_{out}, C)$ or $(*, H_{out}, W_{out}, C)$, where $*$ represents any number of dimension.

$H_{out} = \lfloor {\frac{H_{in} + padding[0] + padding[2] - kernel[0]} {stride[0]} + 1}\rfloor$

$W_{out} = \lfloor {\frac{W_{in} + padding[1] + padding[3] - kernel[1]} {stride[1]} + 1}\rfloor$

if ceilMode = true, please use ceil replace floor in the above ouput formula.

---

Prototype: Pytorch max_pool2d.

Traits: CommonVerifier, PoolLike, StencilLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
kernel	::mlir::ArrayAttr	64-bit integer array attribute with exactly 2 elements
stride	::mlir::ArrayAttr	64-bit integer array attribute with exactly 2 elements
pad	::mlir::ArrayAttr	64-bit integer array attribute with exactly 4 elements
dilation	::mlir::ArrayAttr	64-bit integer array attribute with exactly 2 elements
ceilMode	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*,H,W,C]
output	Same as input
kernel	Shape: [KH,KW] Dim: KH, KW ∈ [1, 256]
stride	Shape: [SH,SW] Dim: SH, SW ∈ [1, 256]
pad	Shape: [PN,PH,PW,PC] PN,PH,PW,PC ∈ [-3, 256]

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*,H,W,C]
output	Same as input
kernel	Shape: [KH,KW] Dim: KH, KW ∈ [1, 256]
stride	Shape: [SH,SW] Dim: SH, SW ∈ [1, 256]
pad	Shape: [PN,PH,PW,PC] PN,PH,PW,PC ∈ [-3, 256]

#hbir.min (::mlir::hbdk::hbir::MinOp)

HBIR tensor min.

Applies minimum operator element-wise, $y_i=min(lhs_i,rhs_i)$.

---

Note:

• Our arithmetic operator support broadcast.

Traits: Broadcastable, CommonVerifier, EltwiseLike, SameElementType

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
lhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type
rhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16 Shape: [*]
rhs	Same as lhs
output	Same as lhs

#J6 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16 Shape: [*]
rhs	Same as lhs
output	Same as lhs

#hbir.mul (::mlir::hbdk::hbir::MulOp)

HBIR tensor multiplication.

Applies multiplication operator element-wise, $y_i=lhs_i\times rhs_i$.

---

Note:

• Our arithmetic operator support broadcast.

---

Prototype: Pytorch mul.

Traits: Broadcastable, CommonVerifier, EltwiseLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
lhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type
rhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16 Shape: [*]
rhs	Same as lhs
output	Same as lhs

#J6 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16 Shape: [*]
rhs	Same as lhs
output	Same as lhs

#hbir.pad (::mlir::hbdk::hbir::PadOp)

Pad at both edges of Tensor.

Padding at the begin and end position with constant / border value. Traits: CommonVerifier, Expansion, Foldable, MoveLike, SameElementType

Interfaces: HBTLExecutable, HbdkExecutorInterface, Layout, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), NonBatchAxesInfer, Perf, Quantizable, RoiInfer, SchedInterface, SchedTemp, ShapeInference, Tiling

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
begin	::mlir::ArrayAttr	64-bit integer array attribute
end	::mlir::ArrayAttr	64-bit integer array attribute
expansionMode	::mlir::hbdk::ExpansionModeAttr	mode to expand input feature on H/W
padValue	::mlir::Attribute	64-bit float attribute or 64-bit signless integer attribute
foldable	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int64, uint64 and f64 are not supported when expansionMode is 'constant' else no constraints Dim: all dims < 737280 when expansionMode is not 'constant' else no constraints
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int64, uint64 and f64 are not supported when expansionMode is 'constant' else no constraints Dim: all dims < 737280 when expansionMode is not 'constant' else no constraints
output	Same as input

#hbir.point_pillar_preprocess (::mlir::hbdk::hbir::PointPillarPreProcessOp)

HBIR point pillar preprocess op.

HBIR point pillar preprocess.Voxelization and Normalization Traits: CommonVerifier, Misc

Interfaces: HBTLExecutable, HbdkExecutorInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
pcRanges	::mlir::ArrayAttr	64-bit float array attribute
normRanges	::mlir::ArrayAttr	64-bit float array attribute
voxelSizes	::mlir::ArrayAttr	64-bit float array attribute
maxVoxelNum	::mlir::IntegerAttr	64-bit signless integer attribute
maxPointsPerVoxel	::mlir::IntegerAttr	64-bit signless integer attribute
normDims	::mlir::ArrayAttr	64-bit integer array attribute

#Operands:

Operand	Description
points	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Results:

Result	Description
voxels	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values
coords	tensor of 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer values

#Constraint:

#J5 BPU

Unsupported

#J6 BPU

Unsupported

#hbir.point_pillar_scatter (::mlir::hbdk::hbir::PointPillarScatterOp)

HBIR point pillar scatter op.

HBIR point pillar scatter. Traits: CommonVerifier, Misc

Interfaces: HBTLExecutable, HbdkExecutorInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
outShape	::mlir::ArrayAttr	64-bit integer array attribute with exactly 4 elements

#Operands:

Operand	Description
voxels	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values
coords	tensor of 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

Unsupported

#J6 BPU

Unsupported

#hbir.relu (::mlir::hbdk::hbir::ReLUOp)

HBIR ReLU activation.

Applies the rectified linear unit function element - wise.Relu function is defined as:

$ReLU(x) = (x) ^ + = max(0, x)$

---

Shape:

• Input(*), where * means any number of dimensions.
• Output(*), same shapes as the input.

---

Prototype: Pytorch Relu.

Traits: CommonVerifier, LutLike, NaiveRoiInfer, NaiveTiling, SameElementType, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), NonBatchAxesInfer, Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16, int32 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16, int32 Shape: [*]
output	Same as input

#hbir.reciprocal (::mlir::hbdk::hbir::ReciprocalOp)

HBIR tensor reciprocal.

Applies reciprocal operator element - wise, $y = \frac{1}{x}$.

---

Note:

• Returns a new tensor with the reciprocal of the elements of input.

Traits: CommonVerifier, EltwiseLike, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.reduce_argmax (::mlir::hbdk::hbir::ReduceArgmaxOp)

Calculate max on multiple axes and return its index.

Return the indices of the max elements of the input tensor's element along the provided axis.

---

Parameters:

• input: the input tensor.
• dims: dimension to perform reduce argmax on. Accepted range is [-r, r - 1] where r = rank(input).
• keepDim: keep the reduced dimensions or not. Default true means keep reduced dimensions.

---

Shape:

• input: $(*)$, where * represents any dimension.
• output: if keepDim is True, same as input. Otherwise, all reduced dims will be discarded (e.g. if input is of shape 1x2x3x4 and dims=[1], the output shape will be 1x3x4).

---

Prototype: ONNX ReduceArgMax.

Traits: CommonVerifier, ReduceLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, RoiInfer, SchedInterface, SchedTemp, ShapeInference, Tiling

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
dims	::mlir::ArrayAttr	64-bit integer array attribute
keepDim	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Results:

Result	Description
output	tensor of 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input, but output can be of type int32 or int64, as long as the size of the reduced axis can be represented using an int16 number

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input, but output can be of type int32 or int64, as long as the size of the reduced axis can be represented using an int16 number

#hbir.reduce_max (::mlir::hbdk::hbir::ReduceMaxOp)

Calculate max on multiple axes.

Return the max value of all elements in the provided axes of the input tensor.

---

Parameters:

• input: the input tensor.
• dims: dimensions to perform reduce max on. If it's a list, reduce over all of them. Accepted range is [-r, r - 1] where r = rank(input).
• keepDim: keep the reduced dimensions or not. Default true means keep reduced dimensions.

---

Shape:

• input: $(*)$, where * represents any dimension.
• output: if keepDim is True, same as input. Otherwise, all reduced dims will be discarded (e.g. if input is of shape 1x2x3x4 and dims=[1, 2], the output shape will be 1x4).

---

Prototype: ONNX ReduceMax.

Traits: CommonVerifier, ReduceLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, RoiInfer, SchedInterface, SchedTemp, ShapeInference, Tiling

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
dims	::mlir::ArrayAttr	64-bit integer array attribute
keepDim	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.reduce_mean (::mlir::hbdk::hbir::ReduceMeanOp)

Calculate mean on multiple axes.

Return the mean of all elements in the provided axes of the input tensor.

---

Parameters:

• input: the input tensor.
• dims: dimensions to perform reduce mean on. If it's a list, reduce over all of them. Accepted range is [-r, r - 1] where r = rank(input).
• keepDim: keep the reduced dimensions or not. Default true means keep reduced dimensions.

---

Shape:

• input: $(*)$, where * represents any dimension.
• output: if keepDim is True, same as input. Otherwise, all reduced dims will be discarded (e.g. if input is of shape 1x2x3x4 and dims=[1, 2], the output shape will be 1x4).

---

Prototype: ONNX ReduceMean.

Traits: CommonVerifier, ReduceLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, RoiInfer, SchedInterface, SchedTemp, ShapeInference, Tiling

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
dims	::mlir::ArrayAttr	64-bit integer array attribute
keepDim	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*] Dim: reduce axis dim size ∈ [1, 16384]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*] Dim: reduce axis dim size ∈ [1, 16384]
output	Same as input

#hbir.reduce_min (::mlir::hbdk::hbir::ReduceMinOp)

Calculate min on multiple axes.

Return the min value of all elements in the provided axes of the input tensor.

---

Parameters:

• input: the input tensor.
• dims: dimensions to perform reduce min on. If it's a list, reduce over all of them. Accepted range is [-r, r - 1] where r = rank(input).
• keepDim: keep the reduced dimensions or not. Default true means keep reduced dimensions.

---

Shape:

• input: $(*)$, where * represents any dimension.
• output: if keepDim is True, same as input. Otherwise, all reduced dims will be discarded (e.g. if input is of shape 1x2x3x4 and dims=[1, 2], the output shape will be 1x4).

---

Prototype: ONNX ReduceMin

Traits: CommonVerifier, ReduceLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, RoiInfer, SchedInterface, SchedTemp, ShapeInference, Tiling

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
dims	::mlir::ArrayAttr	64-bit integer array attribute
keepDim	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.reduce_sum (::mlir::hbdk::hbir::ReduceSumOp)

Calculate sum on multiple axes.

Return the sum of all elements in the provided axes of the input tensor.

---

Parameters:

• input: the input tensor.
• dims: dimensions to perform reduce sum on. If it's a list, reduce over all of them. Accepted range is [-r, r - 1] where r = rank(input).
• keepDim: keep the reduced dimensions or not. Default true means keep reduced dimensions.

---

Shape:

• input: $(*)$, where * represents any dimension.
• output: if keepDim is True, same as input. Otherwise, all reduced dims will be discarded (e.g. if input is of shape 1x2x3x4 and dims=[1, 2], the output shape will be 1x4).

---

Prototype: ONNX ReduceSum.

Traits: CommonVerifier, ReduceLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, RoiInfer, SchedInterface, SchedTemp, ShapeInference, Tiling

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
dims	::mlir::ArrayAttr	64-bit integer array attribute
keepDim	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*] Dim: reduce axis dim size ∈ [1, 16384]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*] Dim: reduce axis dim size ∈ [1, 16384]
output	Same as input

#hbir.reshape (::mlir::hbdk::hbir::ReshapeOp)

View a tensor as another shape

Returns a tensor with the same data and number of elements as input, but with the specified shape.When possible, the returned tensor will be a view of input.

---

Note:

• shape - the new shape.

---

Prototype: Pytorch reshape.

Traits: CommonVerifier, Foldable, MoveLike, SameElementType

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, Layout, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), NonBatchAxesInfer, Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
shape	::mlir::ArrayAttr	64-bit integer array attribute
foldable	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
inputs	no limits
output	same as inputs

#J6 BPU

Features and Attributes	Constraints
inputs	no limits
output	same as inputs

#hbir.resize2d (::mlir::hbdk::hbir::Resize2dOp)

HBIR 2-D resizing.

Scale the input proportionally.

---

Note:

• ratio controls zoom size.*mode controls interpolation type, it's default value is nearest.

---

Shape:

• Input: $(*, H_{in}, W_{in}, C)$
• Output: $(*, H_{out}, W_{out}, C)$, where

$H_{out} = \lfloor{H_{in} * ratio}\rfloor$

$W_{out} = \lfloor{W_{in} * ratio}\rfloor$

---

Prototype: Pytorch upsample_nearest2d.

Traits: CommonVerifier, Expansion, Round, SampleLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
ratio	::mlir::ArrayAttr	64-bit float array attribute with exactly 2 elements
size	::mlir::ArrayAttr	64-bit integer array attribute with exactly 2 elements
step	::mlir::ArrayAttr	64-bit float array attribute with exactly 2 elements
initialOffset	::mlir::ArrayAttr	64-bit float array attribute with exactly 2 elements
mode	::mlir::hbdk::InterpolationModeAttr	interpolation mode for all march
expansionMode	::mlir::hbdk::ExpansionModeAttr	mode to expand input feature on H/W
padValue	::mlir::Attribute	64-bit float attribute or 64-bit signless integer attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8 Shape: [*,H,W,C]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8 Shape: [*,H,W,C]
output	Same as input

#hbir.roll (::mlir::hbdk::hbir::RollOp)

Roll the tensor along the given dimensions

Roll the tensor along the given dimension.Elements that are shifted beyond the last position are re-introduced at the first position.

---

Note:

• shifts - The number of places by which the elements of the tensor are shifted.
• dims -Axis along which to roll.

---

Prototype: Pytorch roll.

Traits: CommonVerifier, MoveLike, SameElementType, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, RoiInfer, SchedInterface, SchedTemp, ShapeInference, Tiling

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
shifts	::mlir::ArrayAttr	64-bit integer array attribute
dims	::mlir::ArrayAttr	64-bit integer array attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
inputs	no limits
output	same as inputs

#J6 BPU

Features and Attributes	Constraints
inputs	no limits
output	same as inputs

#hbir.sigmoid (::mlir::hbdk::hbir::SigmoidOp)

HBIR Sigmoid activation.

Applies the element - wise function.Sigmoid function is defined as:

$Sigmoid(x) = \frac{1} {1 + exp(-x)}$

---

Shape:

• Input: (*), where * means any number of dimensions.
• Output: (*), same shape as the input.

---

Prototype: Pytorch sigmoid.

Traits: CommonVerifier, LutLike, NaiveRoiInfer, NaiveTiling, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.slice (::mlir::hbdk::hbir::SliceOp)

Slice a tensor out of a tensor

Slicing like python's style means taking elements from one given index to another given index.

---

Note:

• begin -the index start to pick(inclusive).
• end -the index end to pick(exclusive).
• step -the step interval of picking.

---

Prototype: Pytorch slice.

Traits: CommonVerifier, Foldable, MoveLike, SameElementType

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, Layout, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), NonBatchAxesInfer, Perf, Quantizable, RoiInfer, SchedInterface, SchedTemp, ShapeInference, Tiling

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
begin	::mlir::ArrayAttr	64-bit integer array attribute
end	::mlir::ArrayAttr	64-bit integer array attribute
step	::mlir::ArrayAttr	64-bit integer array attribute
foldable	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Dim: all dims < 2097152
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Dim: all dims < 2097152
output	Same as input

#hbir.softmax (::mlir::hbdk::hbir::SoftmaxOp)

HBIR Softmax Op.

Applies the Softmax function to an n - dimensional input Tensor rescaling them so that the elements of the n-dimensional output. Tensor lie in the range $[0, 1]$ and sum to 1.

Softmax function is defined as:

$Softmax(x_i) = \frac{exp(x_i)} {\sum_jexp(x_j)}$

---

Shape:

• Input(*), where $*$ means, any number of additional dimensions.Output(), same shape as the input.

---

Prototype: Pytorch softmax.

Traits: CommonVerifier, Misc, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), NonBatchAxesInfer, Perf, Quantizable, RoiInfer, SchedInterface, SchedTemp, ShapeInference, Tiling

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
dim	::mlir::IntegerAttr	64-bit signless integer attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

Unsupported

#J6 BPU

Unsupported

#hbir.softplus (::mlir::hbdk::hbir::SoftplusOp)

HBIR Softplus Op.

Applies the SoftPlus function element-wise. SoftPlus function defined as:

$SoftPlus(x) = \frac{1}{\beta}*log(1+exp(\beta * x))$

SoftPlus is a smooth approximation to the ReLU function.

---

Note:

• beta - the $\beta$ value for the Softplus formulation.
• max - values is for numerical stability, when $\beta *x > max, SoftPlus(x)=x$.

---

Shape:

• Input: (*), where * means any number of dimensions.
• Output: (*), same shape as the input.

---

Prototype: Pytorch softplus.

Traits: CommonVerifier, LutLike, NaiveRoiInfer, NaiveTiling, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
beta	::mlir::FloatAttr	64-bit float attribute
threshold	::mlir::FloatAttr	64-bit float attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.sort (::mlir::hbdk::hbir::SortOp)

HBIR tensor sort

Sorts the elements of the input tensor along a given dimension in ascending order by value.

---

Parameters:

• input: the input tensor.
• dim: the dimension to sort along.
• descending: controls the sorting order (ascending or descending).
• stable: makes the sorting routine stable, which guarantees that the order of equivalent elements is preserved.

---

Shape:

• Input: $(N, *)$
• Values: $(N, *)$ (same shape as input)
• Indices: $(N, *)$ (same shape as input)

---

Prototype: Pytorch sort.

Traits: CommonVerifier, Misc

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
dim	::mlir::IntegerAttr	64-bit signless integer attribute
descending	::mlir::BoolAttr	bool attribute
stable	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
values	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type
indices	tensor of 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer values or none type

#Constraint:

#J5 BPU

Unsupported

#J6 BPU

Unsupported

#hbir.sqrt (::mlir::hbdk::hbir::SqrtOp)

HBIR tensor sqrt.

Applies square root operator element - wise, $y = \sqrt{x}$.

---

Note:

• Returns a new tensor with the square root of the elements of input. If input is negative, then it will return NaN.

Traits: CommonVerifier, LutLike, SameOperandsAndResultShape

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#J6 BPU

Features and Attributes	Constraints
input	Type: int8, int16 Shape: [*]
output	Same as input

#hbir.sub (::mlir::hbdk::hbir::SubOp)

HBIR tensor substraction.

Applies substraction operator element-wise, $y_i=lhs_i-rhs_i$.

---

Note:

• Our arithmetic operator support broadcast.

Traits: Broadcastable, CommonVerifier, EltwiseLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, MoveTransposeInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Operands:

Operand	Description
lhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type
rhs	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16 Shape: [*]
rhs	Same as lhs
output	Same as lhs

#J6 BPU

Features and Attributes	Constraints
lhs	Type: int8, int16 Shape: [*]
rhs	Same as lhs
output	Same as lhs

#hbir.tile (::mlir::hbdk::hbir::TileOp)

Constructs a tensor by tiling a given tensor.

Traits: CommonVerifier, MoveLike, SameElementType

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
multiplies	::mlir::ArrayAttr	64-bit integer array attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values or none type

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values or none type

#Constraint:

#J5 BPU

Features and Attributes	Constraints
inputs	no limits
output	same as inputs

#J6 BPU

Features and Attributes	Constraints
inputs	no limits
output	same as inputs

#hbir.topk (::mlir::hbdk::hbir::TopkOp)

HBIR tensor topk

Returns the k largest elements of the given input tensor along a given dimension.

If dim is not given, the last dimension of the input is chosen.

If largest is False then the k smallest elements are returned.

values, indices are returned in separate tensors, where the indices are the indices of the elements in the original input tensor.

The boolean option sorted if True, will make sure that the returned k elements are themselves sorted.

Traits: CommonVerifier, Misc

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
k	::mlir::IntegerAttr	64-bit signless integer attribute
dim	::mlir::IntegerAttr	64-bit signless integer attribute
largest	::mlir::BoolAttr	bool attribute
sorted	::mlir::BoolAttr	bool attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type

#Results:

Result	Description
values	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values or none type
indices	tensor of 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer values or none type

#Constraint:

#J5 BPU

Unsupported

#J6 BPU

Unsupported

#hbir.transpose (::mlir::hbdk::hbir::TransposeOp)

Reverse or permute the dims of an array; returns the modified array.

Returns a tensor that is a view of the original tensor input with its dimesions permuted.

---

Note:

• input: the input tensor.
• dims: the desired ordering of dimensions.

---

Prototype: Pytorch permute.

Traits: CommonVerifier, MoveLike, SameElementType

Interfaces: HBTLExecutable, HbdkExecutorInterface, HbdkInferType, Layout, NoMemoryEffect (MemoryEffectOpInterface), NonBatchAxesInfer, Perf, Quantizable, RoiInfer, SchedInterface, SchedTemp, ShapeInference, Tiling

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
dims	::mlir::ArrayAttr	64-bit integer array attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
inputs	no limits
output	same as inputs

#J6 BPU

Features and Attributes	Constraints
inputs	no limits
output	same as inputs

#hbir.triton_call (::mlir::hbdk::hbir::TritonCallOp)

Triton kernel call for custom operator.

The hbir.triton_call operation is similar to hbir.call op, but it is a zero-output operator and has side effect.

Traits: CommonVerifier, Misc

Interfaces: MemoryEffectOpInterface (MemoryEffectOpInterface), SchedTemp, ShapeInference

Effects: MemoryEffects::Effect{MemoryEffects::Read on ::mlir::SideEffects::DefaultResource, MemoryEffects::Write on ::mlir::SideEffects::DefaultResource}

#Attributes:

Attribute	MLIR Type	Description
signature	::mlir::StringAttr	string attribute
inputIndices	::mlir::ArrayAttr	array attribute
outputIndices	::mlir::ArrayAttr	array attribute
inputOutputIndices	::mlir::ArrayAttr	array attribute
parameters	::mlir::ArrayAttr	array attribute
library	::mlir::StringAttr	string attribute

#Operands:

Operand	Description
inputs	variadic of tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

None

#J6 BPU

Unsupported

#hbir.warp (::mlir::hbdk::hbir::WarpOp)

HBIR warp.

From the input, sample(bi - linear interpolation) pixels specified by grid.

---

Shape:

• input: input of shape $(*, H_{in}, W_{in}, C_{in})$, where * represent any number of dimension.
• grid: flow - field of shape $(*, H_{out}, W_{out}, 2)$
• output: $(*, H_{out}, W_{out}, C_{in})$

Traits: CommonVerifier, Expansion, Round, SampleLike

Interfaces: HBTLExecutable, HbdkExecutorInterface, NoMemoryEffect (MemoryEffectOpInterface), Perf, Quantizable, SchedInterface, SchedTemp, ShapeInference

Effects: MemoryEffects::Effect

#Attributes:

Attribute	MLIR Type	Description
mode	::mlir::hbdk::InterpolationModeAttr	interpolation mode for all march
expansionMode	::mlir::hbdk::ExpansionModeAttr	mode to expand input feature on H/W
padValue	::mlir::Attribute	64-bit float attribute or 64-bit signless integer attribute

#Operands:

Operand	Description
input	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values
move	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or values

#Results:

Result	Description
output	tensor of 64-bit float or 32-bit float or 16-bit float or bfloat16 type or 8-bit signed integer or 16-bit signed integer or 32-bit signed integer or 64-bit signed integer or 8-bit unsigned integer or 16-bit unsigned integer or 32-bit unsigned integer or 64-bit unsigned integer or or values

#Constraint:

#J5 BPU

Features and Attributes	Constraints
input	Type: int8 Shape: [*,H,W,C] Dim: H,W ∈ [1, 1024]; H*W ≤ 720*1024; other dims ∈ [1, 65536]
grid	Type: int16 Shape: [*,H,W,2]
output	Same as input except Dim constraints

#J6 BPU

Features and Attributes	Constraints
input	Type: int8 Shape: [*,H,W,C] Dim: H,W ∈ [1, 1024]; H*W ≤ 720*1024; other dims ∈ [1, 65536]
grid	Type: int16 Shape: [*,H,W,2]
output	Same as input except Dim constraints