StableHLO is an operation set for high-level operations (HLO) in machine learning (ML) models. StableHLO works as a portability layer between different ML frameworks and ML compilers: ML frameworks that produce StableHLO programs are compatible with ML compilers that consume StableHLO programs.
Our goal is to simplify and accelerate ML development by creating more interoperability between various ML frameworks (such as TensorFlow, JAX and PyTorch) and ML compilers (such as XLA and IREE). Towards that end, this document provides a specification for the StableHLO programming language.
This specification contains three major sections. First, the Programs section describes the structure of StableHLO programs which consist of StableHLO functions which themselves consist of StableHLO ops. Within that structure, the Ops section specifies the semantics of individual ops. The Execution section provides semantics for all these ops executing together within a program. Finally, the Notation section discusses the notation used throughout the specification.
Program ::= {Func}
StableHLO programs consist of an arbitrary number of StableHLO functions. Below is an example program with a function @main
which has 3 inputs (%image
, %weights
and %bias
) and 1 output. The body of the function has 6 ops.
func.func @main( %image: tensor<28x28xf32>, %weights: tensor<784x10xf32>, %bias: tensor<1x10xf32> ) -> tensor<1x10xf32> { %0 = "stablehlo.reshape"(%image) : (tensor<28x28xf32>) -> tensor<1x784xf32> %1 = "stablehlo.dot"(%0, %weights) : (tensor<1x784xf32>, tensor<784x10xf32>) -> tensor<1x10xf32> %2 = "stablehlo.add"(%1, %bias) : (tensor<1x10xf32>, tensor<1x10xf32>) -> tensor<1x10xf32> %3 = "stablehlo.constant"() {value = dense<0.0> : tensor<1x10xf32>} : () -> tensor<1x10xf32> %4 = "stablehlo.maximum"(%2, %3) : (tensor<1x10xf32>, tensor<1x10xf32>) -> tensor<1x10xf32> "func.return"(%4): (tensor<1x10xf32>) -> () }
Func ::= 'func' '.' 'func' FuncId FuncInputs FuncOutputs '{' FuncBody '}' FuncInputs ::= '(' [FuncInput {',' FuncInput}] `)` FuncInput ::= ValueId ':' ValueType FuncOutputs ::= ['->' FuncOutput, {',' FuncOutput}] FuncOutput ::= ValueType FuncBody ::= {Op}
StableHLO functions (which are also called named functions) have an identifier, inputs/outputs and a body. In the future, we are planning to introduce additional metadata for functions to achieve better compatibility with HLO (#425, #626, #740, #744).
FuncId ::= '@' letter {letter | digit} ValueId ::= '%' digit {digit} | '%' letter {letter | digit} letter ::= 'a' | ... | 'z' | 'A' | ... | 'Z' | '_' digit ::= '0' | ... | '9'
StableHLO identifiers are similar to identifiers in many programming languages, with two peculiarities: 1) all identifiers have sigils which distinguish different kinds of identifiers, 2) value identifiers can be completely numeric to simplify generation of StableHLO programs.
Type ::= ValueType | NonValueType ValueType ::= TensorType | QuantizedTensorType | TokenType | TupleType NonValueType ::= TensorElementType | QuantizedTensorElementType | FunctionType | StringType
StableHLO types are categorized into value types (which are also called first-class types) which represent StableHLO values and non-value types which describe other program elements. StableHLO types are similar to types in many programming languages, with the main peculiarity being StableHLO's domain-specific nature which results in some unusual outcomes (e.g. scalar types are not value types).
TensorType ::= 'tensor' '<' Shape TensorElementType '>' Shape ::= {DimensionSize 'x'} DimensionSize ::= digit {digit} | '?'
Tensor types represent tensors, i.e. multidimensional arrays. They have a shape and an element type, where a shape represents non-negative or unknown dimension sizes in the ascending order of the corresponding dimensions (which are also called axes) numbered from 0
to R-1
. The number of dimensions R
is called rank. For example, tensor<2x3xf32>
is a tensor type with shape 2x3
and element type f32
. It has two dimensions (or, in other words, two axes) - 0th dimension and 1st dimension - whose sizes are 2 and 3. Its rank is 2.
Shapes can be partially or completely unknown (dynamic), e.g. tensor<?x2xf64>
is partially unknown and tensor<?x?xf64>
is completely unknown. Dynamic dimension sizes are represented using a ?
. Shapes cannot be unranked.
In the future, we are planning to explore extending tensor types beyond dimension sizes and element types, for example, to include layouts (#629) and sparsity (#1078).
QuantizedTensorType ::= 'tensor' '<' Shape QuantizedTensorElementType '>' QuantizedTensorElementType ::= '!quant.uniform' '<' QuantizationStorageType ['<' QuantizationStorageMin ':' QuantizationStorageMax '>'] ':' QuantizationExpressedType [':' QuantizationDimension] ',' QuantizationParameters '>' QuantizationStorageType ::= IntegerType QuantizationStorageMin ::= IntegerConstant QuantizationStorageMax ::= IntegerConstant QuantizationExpressedType ::= FloatType QuantizationDimension ::= IntegerConstant QuantizationParameters ::= QuantizationParameter | '{' QuantizationParameter {',' QuantizationParameter} '}' QuantizationParameter ::= QuantizationScale ':' QuantizationZeroPoint QuantizationScale ::= FloatConstant QuantizationZeroPoint ::= IntegerConstant
Name | Type | Constraints |
---|---|---|
storage_type | integer type | (C1-C4), (C9) |
storage_min | integer constant | (C2), (C4), (C8) |
storage_max | integer constant | (C3), (C4), (C8) |
expressed_type | floating-point type | (C1), (C5) |
quantization_dimension | optional integer constant | (C11-C13) |
scales | variadic number of floating-point constants | (C5-C7), (C10), (C11), (C13) |
zero_points | variadic number of integer constants | (C8-C10) |
Quantized element types represent integer values of a storage type in the range from storage_min
to storage_max
(inclusive) that correspond to floating-point values of an expressed type. For a given integer value i
, the corresponding floating-point value f
can be computed as f = (i - zero_point) * scale
, where scale
and zero_point
are called quantization parameters. The storage_min
and storage_max
are optional in the grammar, but have default values of min_value(storage_type)
and max_value(storage_type)
respectively. Quantized element types have the following constraints:
num_bits(storage_type) < num_bits(expressed_type)
.type(storage_min) = storage_type
.type(storage_max) = storage_type
.min_value(storage_type) <= storage_min < storage_max <= max_value(storage_type)
.type(scales...) = expressed_type
.0 < scales
.is_finite(scales...)
.storage_min <= zero_points <= storage_max
.type(zero_points...) = storage_type
.size(scales) = size(zero_points)
.is_empty(quantization_dimension)
, then size(scales) = 1
.0 <= quantization_dimension
.At the moment, QuantizationScale
is a floating-point constant, but there is strong interest in integer-based scales, represented with multipliers and shifts. We are planning to explore this in the near future (#1404).
There is an ongoing discussion on the semantics of QuantizationZeroPoint
, including the type, the values and whether there can be just one or potentially multiple zero points in a quantized tensor type. Based on the results of this discussion, the specification around zero points may change in the future (#1405).
Another ongoing discussion involves the semantics of QuantizationStorageMin
and QuantizationStorageMax
to determine whether any constraints should be imposed on these values and on the values of quantized tensors (#1406).
Finally, we are planning to explore representing unknown scales and zero points, similarly to how we are planning to explore representing unknown dimension sizes (#1407).
Quantized tensor types represent tensors with quantized elements. These tensors are exactly the same as regular tensors, except that their elements have quantized element types, instead of regular element types.
In quantized tensors, quantization can be per-tensor, meaning, having one scale
and zero_point
for the entire tensor or can be per-axis, meaning, having multiple scales
and zero_points
, one pair per slice of a particular dimension quantization_dimension
. More formally, in a tensor t
with per-axis quantization, there are dim(t, quantization_dimension)
slices of the quantization_dimension
: t[:, ..., 0, ..., :], t[:, ..., 1, ..., :]
, etc. All elements in the i
th slice use scales[i]
and zero_points[i]
as their quantization parameters. Quantized tensor types have the following constraints:
quantization_dimension < rank(self)
.dim(self, quantization_dimension) = size(scales)
.TokenType ::= 'token'
Token types represent tokens, i.e. opaque values produced and consumed by some operations. Tokens are used for imposing execution order on operations as described in the Execution section.
TupleType ::= 'tuple' '<' TupleElementTypes '>' TupleElementTypes ::= [ValueType {',' ValueType}]
Tuple types represent tuples, i.e. heterogeneous lists. Tuples are a legacy feature which only exists for compatibility with HLO. In HLO, tuples are used to represent variadic inputs and outputs. In StableHLO, variadic inputs and outputs are supported natively, and the only use of tuples in StableHLO is to comprehensively represent HLO ABI where e.g. T
, tuple<T>
and tuple<tuple<T>>
may be materially different depending on a particular implementation. In the future, we are planning to make changes to HLO ABI which may allow us to remove tuple types from StableHLO (#598).
TensorElementType ::= BooleanType | IntegerType | FloatType | ComplexType BooleanType ::= 'i1' IntegerType ::= SignedIntegerType | UnsignedIntegerType SignedIntegerType ::= 'si4' | 'si8' | 'si16' | 'si32' | 'si64' UnsignedIntegerType ::= 'ui4' | 'ui8' | 'ui16' | 'ui32' | 'ui64' FloatType ::= 'f8E4M3FN' | 'f8E5M2' | 'f8E4M3FNUZ' | 'f8E5M2FNUZ' | 'f8E4M3B11FNUZ' | 'bf16' | 'f16' | 'f32' | 'f64' ComplexType ::= 'complex' '<' ComplexElementType '>' ComplexElementType ::= 'f32' | 'f64'
Element types represent elements of tensor types. Unlike in many programming languages, these types are not first class in StableHLO. This means that StableHLO programs cannot directly represent values of these types (as a result, it is idiomatic to represent scalar values of type T
with 0-dimensional tensor values of type tensor<T>
).
true
and false
.si
) or unsigned (ui
) and have one of the supported bit widths (4
, 8
, 16
, 32
or 64
). Signed siN
types represent integer values from -2^(N-1)
to 2^(N-1)-1
inclusive, and unsigned uiN
types represent integer values from 0
to 2^N-1
inclusive.f8E4M3FN
and f8E5M2
types corresponding to respectively the E4M3
and E5M2
encodings of the FP8 format described in FP8 Formats for Deep Learning.f8E4M3FNUZ
and f8E5M2FNUZ
types corresponding to the E4M3
and E5M2
encodings of the FP8 formats described in 8-bit Numerical Formats for Deep Neural Networks.f8E4M3B11FNUZ
type corresponding to the E4M3
encoding of the FP8 formats described in Hybrid 8-bit Floating Point (HFP8) Training and Inference for Deep Neural Networks.bf16
type corresponding to the bfloat16
format described in BFloat16: The secret to high performance on Cloud TPUs.f16
, f32
and f64
types corresponding to respectively binary16
(“half precision”), binary32
(“single precision”) and binary64
(“double precision”) formats described in the IEEE 754 standard.complex<f32>
(both parts are of type f32
) and complex<f64>
(both parts are of type f64
).FunctionType ::= '(' InputTypes ')' '->' '(' OutputTypes ')' InputTypes ::= [ValueType {',' ValueType}] OutputTypes ::= [ValueType {',' ValueType}]
Function types represent both named and anonymous functions. They have input types (the list of types on the left-hand side of ->
) and output types (the list of types on the right-hand side of ->
). In many programming languages, function types are first class, but not in StableHLO.
StringType ::= 'string'
String type represents sequences of bytes. Unlike in many programming languages, string type is not first class in StableHLO and is only used to specify static metadata for program elements.
StableHLO operations (which are also called ops) represent a closed set of high-level operations in machine learning models. As discussed above, StableHLO syntax is heavily inspired by MLIR, which is not necessarily the most ergonomic alternative, but is arguably the best fit for StableHLO's goal of creating more interoperability between ML frameworks and ML compilers.
Op ::= [OpOutputs] OpName OpInputs ':' OpSignature OpName ::= '"' 'stablehlo' '.' OpMnemonic '"' OpMnemonic ::= 'abs' | 'add' | ...
StableHLO operations (which are also called ops) have a name, inputs/outputs and a signature. The name consists of the stablehlo.
prefix and a mnemonic which uniquely identifies one of the supported ops. See below for a comprehensive list of all supported ops.
At the moment, StableHLO programs in the wild sometimes contain operations that are not described in this document. In the future, we are planning to either absorb these operations into the StableHLO opset or prohibit them from appearing in StableHLO programs. In the meanwhile, here is the list of these operations:
builtin.module
, func.func
, func.call
and func.return
(#425).chlo
operations (#602).broadcast
, create_token
, cross-replica-sum
, dot
, einsum
, torch_index_select
, unary_einsum
(#3), and trace
(#604).compute_reshape_shape
, cstr_reshapable
, dynamic_broadcast_in_dim
, dynamic_conv
, dynamic_gather
, dynamic_iota
, dynamic_pad
, dynamic_reshape
, real_dynamic_slice
, set_dimension_size
(#8).arith
, shape
and tensor
operations (#8).OpInputs ::= OpInputValues OpInputFuncs OpInputAttrs OpInputValues ::= '(' [OpInputValue {',' OpInputValue}] ')' OpInputValue ::= ValueId OpInputFuncs ::= ['(' OpInputFunc {',' OpInputFunc} ')'] OpInputAttrs ::= ['{' OpInputAttr {',' OpInputAttr} '}'] OpOutputs ::= [OpOutput {',' OpOutput} '='] OpOutput ::= ValueId
Ops consume inputs and produce outputs. Inputs are categorized into input values (computed during execution), input functions (provided statically, because in StableHLO functions are not first-class values) and input attributes (also provided statically). The kind of inputs and outputs consumed and produced by an op depends on its mnemonic. For example, the add
op consumes 2 input values and produces 1 output value. In comparison, the select_and_scatter
op consumes 3 input values, 2 input functions and 3 input attributes.
OpInputFunc ::= '{' Unused FuncInputs ':' FuncBody '}' Unused ::= '^' digit {digit} | '^' letter {letter | digit}
Input functions (which are also called anonymous functions) are very similar to named functions except that: 1) they don‘t have an identifier (hence the name “anonymous”), 2) they don’t declare output types (output types are inferred from the return
op within the function).
The syntax for input functions includes a currently unused part (see the Unused
production above) which is there for compatibility with MLIR. In MLIR, there is a more general concept of “regions” which can have multiple “blocks” of ops connected together via jump ops. These blocks have ids which correspond to the Unused
production, so that they can be distinguished from each other. StableHLO doesn't have jump ops, so the corresponding part of MLIR syntax is unused (but is still there).
OpInputAttr ::= OpInputAttrName '=' OpInputAttrValue OpInputAttrName ::= letter {letter | digit} OpInputAttrValue ::= Constant
Input attributes have a name and a value which is one of the supported constants. They are the primary way to specify static metadata for program elements. For example, the concatenate
op uses the attribute dimension
to specify the dimension along which its input values are concatenated. Similarly, the slice
op uses multiple attributes like start_indices
and limit_indices
to specify the bounds that are used to slice the input value.
At the moment, StableHLO programs in the wild sometimes contain attributes which are not described in this document. In the future, we are planning to either absorb these attributes into the StableHLO opset or prohibit them from appearing in StableHLO programs. In the meanwhile, here is the list of these attributes:
layout
(#629).mhlo.frontend_attributes
(#628).mhlo.sharding
(#619).output_operand_aliases
(#740).OpSignature ::= '(' [ValueType {',' ValueType}] ')' '->' '(' [ValueType {',' ValueType}] ')'
Op signature consists of the types of all input values (the list of types on the left-hand side of ->
) and the types of all output values (the list of types on the right-hand side of ->
). Strictly speaking, input types are redundant, and output types are almost always redundant as well (because for most StableHLO ops, output types can be inferred from inputs). Nonetheless, op signature is deliberately part of StableHLO syntax for compatibility with MLIR.
Below is an example op whose mnemonic is select_and_scatter
. It consumes 3 input values (%operand
, %source
and %init_value
), 2 input functions and 3 input attributes (window_dimensions
, window_strides
and padding
). Note how the signature of the op only includes the types of its input values (but not the types of input functions and attributes which are provided inline).
%result = "stablehlo.select_and_scatter"(%operand, %source, %init_value) ({ ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>): %0 = "stablehlo.compare"(%arg0, %arg1) { comparison_direction = #stablehlo<comparison_direction GE> } : (tensor<i32>, tensor<i32>) -> tensor<i1> "stablehlo.return"(%0) : (tensor<i1>) -> () }, { ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>): %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i32>, tensor<i32>) -> tensor<i32> "stablehlo.return"(%0) : (tensor<i32>) -> () }) { window_dimensions = dense<[3, 1]> : tensor<2xi64>, window_strides = dense<[2, 1]> : tensor<2xi64>, padding = dense<[[0, 1], [0, 0]]> : tensor<2x2xi64> } : (tensor<4x2xi32>, tensor<2x2xi32>, tensor<i32>) -> tensor<4x2xi32>
Constant ::= BooleanConstant | IntegerConstant | FloatConstant | ComplexConstant | TensorConstant | QuantizedTensorConstant | StringConstant | EnumConstant
StableHLO constants have a literal and a type which together represent a StableHLO value. Generally, the type is part of the constant syntax, except when it's unambiguous (e.g. a boolean constant unambiguously has type i1
, whereas an integer constant can have multiple possible types).
BooleanConstant ::= BooleanLiteral BooleanLiteral ::= 'true' | 'false'
Boolean constants represent boolean values true
and false
. Boolean constants have type i1
.
IntegerConstant ::= IntegerLiteral ':' IntegerType IntegerLiteral ::= ['-' | '+'] DecimalDigits | ['-' | '+'] '0x' HexadecimalDigits DecimalDigits ::= decimalDigit {decimalDigit} HexadecimalDigits ::= hexadecimalDigit {hexadecimalDigit} decimalDigit ::= '0' | ... | '9' hexadecimalDigit ::= decimalDigit | 'a' | ... | 'f' | 'A' | ... | 'F'
Integer constants represent integer values via strings that use decimal or hexadecimal notation. Other bases, e.g. binary or octal, are not supported. Integer constants have the following constraints:
is_wellformed(integer_literal, integer_type)
.FloatConstant ::= FloatLiteral ':' FloatType FloatLiteral ::= SignPart IntegerPart FractionalPart ScientificPart | '0x' [HexadecimalDigits] SignPart ::= ['-' | '+'] IntegerPart ::= DecimalDigits FractionalPart ::= ['.' [DecimalDigits]] ScientificPart ::= [('e' | 'E') ['-' | '+'] DecimalDigits]
Floating-point constants represent floating-point values via strings that use decimal or scientific notation. Additionally, hexadecimal notation can be used to directly specify the underlying bits in the floating-point format of the corresponding type. Floating-point constants have the following constraints:
is_wellformed(float_literal, float_type)
.size(hexadecimal_digits) = num_bits(float_type) / 4
.ComplexConstant ::= ComplexLiteral ':' ComplexType ComplexLiteral ::= '(' RealPart ',' ImaginaryPart ')' RealPart ::= FloatLiteral ImaginaryPart ::= FloatLiteral
Complex constants represent complex values using lists of a real part (comes first) and an imaginary part (comes second). For example, (1.0, 0.0) : complex<f32>
represents 1.0 + 0.0i
, and (0.0, 1.0) : complex<f32>
represents 0.0 + 1.0i
. The order in which these parts are then stored in memory is implementation-defined. Complex constants have the following constraints:
is_wellformed(real_part, complex_element_type(complex_type))
.is_wellformed(imaginary_part, complex_element_type(complex_type))
.TensorConstant ::= TensorLiteral ':' TensorType TensorLiteral ::= 'dense' '<' (DenseLiteral | ElementLiteral) '>' DenseLiteral ::= DenseDimension | DenseElements DenseDimension ::= '[' [DenseLiteral {',' DenseLiteral}] ']' DenseElements ::= [ElementLiteral {',' ElementLiteral}] ElementLiteral ::= BooleanLiteral | IntegerLiteral | FloatLiteral | ComplexLiteral
Tensor constants represent tensor values using nested lists specified via NumPy notation. For example, dense<[[1, 2, 3], [4, 5, 6]]> : tensor<2x3xi32>
represents a tensor value with the following mapping from indices to elements: {0, 0} => 1
, {0, 1} => 2
, {0, 2} => 3
, {1, 0} => 4
, {1, 1} => 5
, {1, 2} => 6
. The order in which these elements are then stored in memory is implementation-defined. Tensor constants have the following constraints:
has_syntax(tensor_literal, element_type(tensor_type))
, where:has_syntax(element_literal: Syntax, element_type: Type) = is_wellformed(element_literal, type)
.has_syntax(tensor_literal: List, element_type: Type) = has_syntax(tensor_literal..., element_type)
.has_shape(tensor_literal, shape(tensor_type))
, where:has_shape(element_literal: Syntax, []) = true
.has_shape(tensor_literal: List, shape: List) = size(tensor_literal) = shape[0] and has_shape(tensor_literal..., shape[1:])
.false
.QuantizedTensorConstant ::= QuantizedTensorLiteral ':' QuantizedTensorType QuantizedTensorLiteral ::= 'dense' '<' (DenseLiteral | ElementLiteral) '>'
Quantized tensor constants represent quantized tensor values using the same notation as tensor constants, with elements specified as constants of their storage type. Quantized tensor constants have the following constraints:
has_syntax(quantized_tensor_literal, storage_type(quantized_tensor_type))
.has_shape(quantized_tensor_literal, shape(quantized_tensor_type))
.StringConstant ::= StringLiteral StringLiteral ::= '"' {stringCharacter | escapeSequence} '"' stringCharacter ::= all ASCII characters except '\00', '\01', ... '\1f' and '"' escapeSequence ::= '\' ('"' | '\' | 'n' | 't' | (hexadecimalDigit hexadecimalDigit))
String literals consist of bytes specified using ASCII characters and escape sequences. They are encoding-agnostic, so the interpretation of these bytes is implementation-defined. String literals have type string
.
Performs element-wise abs operation on operand
tensor and produces a result
tensor. Depending on the element type, does the following:
abs
from IEEE-754.dequantize_op_quantize(abs, operand, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of signed integer, floating-point, or complex type or per-tensor quantized tensor | (C1-C2) |
Name | Type | Constraints |
---|---|---|
result | tensor of signed integer or floating-point type or per-tensor quantized tensor | (C1-C2) |
shape(result) = shape(operand)
.baseline_element_type(result)
is defined as:complex_element_type(element_type(operand))
if is_complex(operand)
.baseline_element_type(operand)
otherwise.// %operand: [-2, 0, 2] %result = "stablehlo.abs"(%operand) : (tensor<3xi32>) -> tensor<3xi32> // %result: [2, 0, 2]
Performs element-wise addition of two tensors lhs
and rhs
and produces a result
tensor. Depending on the element type, does the following:
addition
from IEEE-754.dequantize_op_quantize(add, lhs, rhs, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor or quantized tensor | (C1-C6) |
(I2) | rhs | tensor or quantized tensor | (C1-C5), (C7) |
Name | Type | Constraints |
---|---|---|
result | tensor or quantized tensor | (C1-C7) |
type(lhs) = type(rhs) = type(result)
.is_quantized(lhs) and is_quantized(rhs) and is_quantized(result)
.storage_type(lhs) = storage_type(rhs) = storage_type(result)
.expressed_type(lhs) = expressed_type(rhs) = expressed_type(result)
.(is_per_axis_quantized(lhs) or is_per_axis_quantized(rhs)) = is_per_axis_quantized(result)
.is_per_axis_quantized(lhs)
, then quantization_dimension(lhs) = quantization_dimension(result)
.is_per_axis_quantized(rhs)
, then quantization_dimension(rhs) = quantization_dimension(result)
.// %lhs: [[1, 2], [3, 4]] // %rhs: [[5, 6], [7, 8]] %result = "stablehlo.add"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32> // %result: [[6, 8], [10, 12]]
Ensures that the operations producing the inputs
are executed before any operations that depend on result
. Execution of this operation does nothing, it only exists to establish data dependencies from result
to inputs
.
Label | Name | Type |
---|---|---|
(I1) | inputs | variadic number of token |
Name | Type |
---|---|
result | token |
// %input0: !stablehlo.token // %input1: !stablehlo.token %result = "stablehlo.after_all"(%input0, %input1) : (!stablehlo.token, !stablehlo.token) -> !stablehlo.token
Within each process group in the StableHLO process grid, concatenates the values of the operand
tensor from each process along all_gather_dim
and produces a result
tensor.
The operation splits the StableHLO process grid into process_groups
which is defined as follows:
cross_replica(replica_groups)
if channel_id <= 0 and use_global_device_ids = false
.cross_replica_and_partition(replica_groups)
if channel_id > 0 and use_global_device_ids = false
.flattened_ids(replica_groups)
if channel_id > 0 and use_global_device_ids = true
.Afterwards, within each process_group
:
operands@receiver = [operand@sender for sender in process_group]
for all receiver
in process_group
.result@process = concatenate(operands@process, all_gather_dim)
for all process
in process_group
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or per-tensor quantized tensor | (C1), (C6) |
(I2) | all_gather_dim | constant of type si64 | (C1), (C6) |
(I3) | replica_groups | 2-dimensional tensor constant of type si64 | (C2-C4) |
(I4) | channel_id | constant of type si64 | (C5) |
(I5) | use_global_device_ids | constant of type i1 | (C5) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C6) |
0 <= all_gather_dim < rank(operand)
.is_unique(replica_groups)
.size(replica_groups)
is defined as:num_replicas
if cross_replica
is used.num_replicas
if cross_replica_and_partition
is used.num_processes
if flattened_ids
is used.0 <= replica_groups < size(replica_groups)
.use_global_device_ids = true
, then channel_id > 0
.type(result) = type(operand)
except:dim(result, all_gather_dim) = dim(operand, all_gather_dim) * dim(process_groups, 1)
.// num_replicas: 2 // num_partitions: 1 // %operand@(0, 0): [[1, 2], [3, 4]] // %operand@(1, 0): [[5, 6], [7, 8]] %result = "stablehlo.all_gather"(%operand) { all_gather_dim = 1 : i64, replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>, // channel_id = 0 channel_handle = #stablehlo.channel_handle<handle = 0, type = 0> // use_global_device_ids = false } : (tensor<2x2xi64>) -> tensor<2x4xi64> // %result@(0, 0): [[1, 2, 5, 6], [3, 4, 7, 8]] // %result@(1, 0): [[1, 2, 5, 6], [3, 4, 7, 8]]
Within each process group in the StableHLO process grid, applies a reduction function computation
to the values of the operand
tensor from each process and produces a result
tensor.
The operation splits the StableHLO process grid into process_groups
which is defined as follows:
cross_replica(replica_groups)
if channel_id <= 0 and use_global_device_ids = false
.cross_replica_and_partition(replica_groups)
if channel_id > 0 and use_global_device_ids = false
.flattened_ids(replica_groups)
if channel_id > 0 and use_global_device_ids = true
.Afterwards, within each process_group
:
result@process[result_index] = exec(schedule)
for some binary tree schedule
where:exec(node)
= computation(exec(node.left), exec(node.right))
.exec(leaf)
= leaf.value
.schedule
is an implementation-defined binary tree whose in-order traversal is to_destination_type(operands@process_group...[result_index], type(func_inputs(computation)[0]))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or per-tensor quantized tensor | (C5), (C6) |
(I2) | replica_groups | variadic number of 1-dimensional tensor constants of type si64 | (C1-C3) |
(I3) | channel_id | constant of type si64 | (C4) |
(I4) | use_global_device_ids | constant of type i1 | (C4) |
(I5) | computation | function | (C5) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C6-C7) |
is_unique(replica_groups)
.size(replica_groups)
is defined as:num_replicas
if cross_replica
is used.num_replicas
if cross_replica_and_partition
is used.num_processes
if flattened_ids
is used.0 <= replica_groups < size(replica_groups)
.use_global_device_ids = true
, then channel_id > 0
.computation
has type (tensor<E>, tensor<E>) -> (tensor<E>)
where is_promotable(element_type(operand), E)
.shape(result) = shape(operand)
.element_type(result) = E
.// num_replicas: 2 // num_partitions: 1 // %operand@(0, 0): [1, 2, 3, 4] // %operand@(1, 0): [5, 6, 7, 8] %result = "stablehlo.all_reduce"(%operand) ({ ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>): %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64> "stablehlo.return"(%0) : (tensor<i64>) -> () }) { replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>, channel_handle = #stablehlo.channel_handle<handle = 0, type = 0> } : (tensor<4xi64>) -> tensor<4xi64> // %result@(0, 0): [6, 8, 10, 12] // %result@(1, 0): [6, 8, 10, 12]
Within each process group in the StableHLO process grid, splits the values of the operand
tensor along split_dimension
into parts, scatters the split parts between the processes, concatenates the scattered parts along concat_dimension
and produces a result
tensor.
The operation splits the StableHLO process grid into process_groups
which is defined as follows:
cross_replica(replica_groups)
if channel_id <= 0
.cross_partition(replica_groups)
if channel_id > 0
.Afterwards, within each process_group
:
split_parts@sender = split(operand@sender, split_count, split_dimension)
for all sender
in process_group
.scattered_parts@receiver = [split_parts@sender[receiver_index] for sender in process_group]
where receiver_index = process_group.index(receiver)
.result@process = concatenate(scattered_parts@process, concat_dimension)
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or per-tensor quantized tensor | (C1-C3), (C9) |
(I2) | split_dimension | constant of type si64 | (C1), (C2), (C9) |
(I3) | concat_dimension | constant of type si64 | (C3), (C9) |
(I4) | split_count | constant of type si64 | (C2), (C4), (C8), (C9) |
(I5) | replica_groups | 2-dimensional tensor constant of type si64 | (C5-C8) |
(I6) | channel_id | constant of type si64 |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C9) |
0 <= split_dimension < rank(operand)
.dim(operand, split_dimension) % split_count = 0
.0 <= concat_dimension < rank(operand)
.0 < split_count
.is_unique(replica_groups)
.size(replica_groups)
is defined as:num_replicas
if cross_replica
is used.num_partitions
if cross_partition
is used.0 <= replica_groups < size(replica_groups)
.dim(replica_groups, 1) = split_count
.type(result) = type(operand)
except:dim(result, split_dimension) = dim(operand, split_dimension) / split_count
.dim(result, concat_dimension) = dim(operand, concat_dimension) * split_count
.// num_replicas: 2 // num_partitions: 1 // %operand@(0, 0): [[1, 2, 3, 4], // [5, 6, 7, 8]] // %operand@(1, 0): [[9, 10, 11, 12], // [13, 14, 15, 16]] %result = "stablehlo.all_to_all"(%operand) { split_dimension = 1 : i64, concat_dimension = 0 : i64, split_count = 2 : i64, replica_groups = dense<[[0, 1]]> : tensor<1x2xi64> } : (tensor<2x4xi64>) -> tensor<4x2xi64> // %result@(0, 0): [[1, 2], // [5, 6], // [9, 10], // [13, 14]] // %result@(1, 0): [[3, 4], // [7, 8], // [11, 12], // [15, 16]]
Performs element-wise AND of two tensors lhs
and rhs
and produces a result
tensor. Depending on the element type, does the following:
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor of boolean or integer type | (C1) |
(I2) | rhs | tensor of boolean or integer type | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of boolean or integer type | (C1) |
type(lhs) = type(rhs) = type(result)
.// %lhs: [[1, 2], [3, 4]] // %rhs: [[5, 6], [7, 8]] %result = "stablehlo.and"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32> // %result: [[1, 2], [3, 0]]
Performs element-wise atan2 operation on lhs
and rhs
tensor and produces a result
tensor. Depending on the element type, does the following:
atan2
from IEEE-754.dequantize_op_quantize(atan2, lhs, rhs, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
(I2) | rhs | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(lhs) = baseline_type(rhs) = baseline_type(result)
.// %lhs: [0.0, 1.0, -1.0] // %rhs: [0.0, 0.0, 0.0] %result = "stablehlo.atan2"(%lhs, %rhs) : (tensor<3xf64>, tensor<3xf64>) -> tensor<3xf64> // %result: [0.0, 1.57079637, -1.57079637] // [0.0, pi/2, -pi/2]
Computes gradients of several inputs of batch_norm_training
backpropagating from grad_output
, and produces grad_operand
, grad_scale
and grad_offset
tensors. More formally, this operation can be expressed as a decomposition to existing StableHLO operations using Python syntax as follows:
def compute_sum(operand, feature_index): (sum,) = reduce( inputs=[operand], init_values=[constant(0, element_type(operand))], dimensions=[i for i in range(rank(operand)) if i != feature_index], body=lambda x, y: add(x, y)) return sum def compute_mean(operand, feature_index): sum = compute_sum(operand, feature_index) divisor = constant(size(operand) / dim(operand, feature_index), element_type(operand)) divisor_bcast = broadcast_in_dim(divisor, [], type(sum)) return divide(sum, divisor_bcast) def batch_norm_grad(operand, scale, mean, variance, grad_output, epsilon, feature_index): # Broadcast inputs to type(operand) scale_bcast = broadcast_in_dim(scale, [feature_index], type(operand)) mean_bcast = broadcast_in_dim(mean, [feature_index], type(operand)) variance_bcast = broadcast_in_dim(variance, [feature_index], type(operand)) epsilon_bcast = broadcast_in_dim(constant(epsilon, element_type(operand)), [], type(operand)) # Perform normalization using the provided `mean` and `variance` # Intermediate values will be useful for computing gradients centered_operand = subtract(operand, mean_bcast) stddev = sqrt(add(variance_bcast, epsilon_bcast)) normalized_operand = divide(centered_operand, stddev) # Use the implementation from batchnorm_expander.cc in XLA # Temporary variables have exactly the same names as in the C++ code elements_per_feature = broadcast_in_dim( constant(divide(size(operand), dim(operand, feature_index)), element_type(grad_output)), [], type(operand)) i1 = multiply(grad_output, elements_per_feature) i2 = broadcast_in_dim( compute_sum(grad_output, feature_index), [feature_index], type(operand)) i3 = broadcast_in_dim( compute_sum(multiply(grad_output, centered_operand), feature_index), [feature_index], type(operand)) i4 = multiply(i3, centered_operand) i5 = divide(i4, add(variance_bcast, epsilon_bcast)) i6 = subtract(subtract(i1, i2), i5) grad_operand = multiply(divide(divide(scale_bcast, stddev), elements_per_feature), i6) grad_scale = compute_sum(multiply(grad_output, normalized_operand), feature_index) grad_offset = compute_sum(grad_output, feature_index) return grad_operand, grad_scale, grad_offset
For quantized types, performs dequantize_batch_norm_grad_or_training_quantize(lambda operand, scale, mean, variance, grad_output: batch_norm_grad(operand, scale, mean, variance, grad_output, epsilon, feature_index), operand, scale, mean, variance, grad_output, type(grad_operand), type(grad_scale), type(feature_index))
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point type or per-tensor quantized tensor | (C1-C3), (C5) |
(I2) | scale | 1-dimensional tensor of floating-point or per-tensor quantized type | (C2), (C4), (C5) |
(I3) | mean | 1-dimensional tensor of floating-point or per-tensor quantized type | (C2), (C4) |
(I4) | variance | 1-dimensional tensor of floating-point or per-tensor quantized type | (C2), (C4) |
(I5) | grad_output | tensor of floating-point type or per-tensor quantized tensor | (C2), (C3) |
(I6) | epsilon | constant of type f32 | |
(I7) | feature_index | constant of type si64 | (C1), (C5) |
Name | Type | Constraints |
---|---|---|
grad_operand | tensor of floating-point type or per-tensor quantized tensor | (C2), (C3) |
grad_scale | 1-dimensional tensor of floating-point or per-tensor quantized type | (C2), (C4) |
grad_offset | 1-dimensional tensor of floating-point or per-tensor quantized type | (C2), (C4) |
0 <= feature_index < rank(operand)
.operand
, scale
, mean
, variance
, grad_output
, grad_operand
, grad_scale
and grad_offset
have the same baseline_element_type
.operand
, grad_output
and grad_operand
have the same shape.scale
, mean
, variance
, grad_scale
and grad_offset
have the same shape.size(scale) = dim(operand, feature_index)
.// %operand: [ // [[1.0, 2.0], [3.0, 4.0]], // [[3.0, 4.0], [1.0, 2.0]] // ] // %scale: [1.0, 1.0] // %mean: [2.0, 3.0] // %variance: [1.0, 1.0] // %grad_output: [ // [[0.1, 0.1], [0.1, 0.1]], // [[0.1, 0.1], [0.1, 0.1]] // ] %grad_operand, %grad_scale, %grad_offset = "stablehlo.batch_norm_grad"(%operand, %scale, %mean, %variance, %grad_output) { epsilon = 0.0 : f32, feature_index = 2 : i64 } : (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>, tensor<2xf64>, tensor<2x2x2xf64>) -> (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>) // %grad_operand: [ // [[0.0, 0.0], [0.0, 0.0]], // [[0.0, 0.0], [0.0, 0.0]] // ] // %grad_scale: [0.0, 0.0] // %grad_offset: [0.4, 0.4]
Normalizes the operand
tensor across all dimensions except for the feature_index
dimension and produces a result
tensor. More formally, this operation can be expressed as a decomposition to existing StableHLO operations using Python syntax as follows:
def batch_norm_inference(operand, scale, offset, mean, variance, epsilon, feature_index): # Broadcast inputs to shape(operand) scale_bcast = broadcast_in_dim(scale, [feature_index], type(operand)) offset_bcast = broadcast_in_dim(offset, [feature_index], type(operand)) mean_bcast = broadcast_in_dim(mean, [feature_index], type(operand)) variance_bcast = broadcast_in_dim(variance, [feature_index], type(operand)) epsilon_bcast = broadcast_in_dim(constant(epsilon, element_type(operand)), [], type(operand)) # Perform normalization using the provided `mean` and `variance` instead of # computing them like `batch_norm_training` does. centered_operand = subtract(operand, mean_bcast) stddev = sqrt(add(variance_bcast, epsilon_bcast)) normalized_operand = divide(centered_operand, stddev) return add(multiply(scale_bcast, normalized_operand), offset_bcast)
For quantized types, performs dequantize_op_quantize(lambda operand, scale, offset, mean, variance: batch_norm_inference(operand, scale, offset, mean, variance, epsilon, feature_index), operand, scale, offset, mean, variance, type(result))
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point type or per-tensor quantized tensor | (C1-C7) |
(I2) | scale | 1-dimensional tensor of floating-point or per-tensor quantized type | (C2), (C3) |
(I3) | offset | 1-dimensional tensor of floating-point or per-tensor quantized type | (C2), (C4) |
(I4) | mean | 1-dimensional tensor of floating-point or per-tensor quantized type | (C5) |
(I5) | variance | 1-dimensional tensor of floating-point or per-tensor quantized type | (C2), (C6) |
(I6) | epsilon | constant of type f32 | |
(I7) | feature_index | constant of type si64 | (C1), (C3-C6) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point type or per-tensor quantized tensor | (C2), (C7) |
0 <= feature_index < rank(operand)
.operand
, scale
, offset
, mean
, variance
and result
have the same baseline_element_type
.size(scale) = dim(operand, feature_index)
.size(offset) = dim(operand, feature_index)
.size(mean) = dim(operand, feature_index)
.size(variance) = dim(operand, feature_index)
.baseline_type(operand) = baseline_type(result)
.// %operand: [ // [[1.0, 2.0], [3.0, 4.0]], // [[3.0, 4.0], [1.0, 2.0]] // ] // %scale: [1.0, 1.0] // %offset: [1.0, 1.0] // %mean: [2.0, 3.0] // %variance: [1.0, 1.0] %result = "stablehlo.batch_norm_inference"(%operand, %scale, %offset, %mean, %variance) { epsilon = 0.0 : f32, feature_index = 2 : i64 } : (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>, tensor<2xf64>, tensor<2xf64>) -> tensor<2x2x2xf64> // %result: [ // [[0.0, 0.0], [2.0, 2.0]], // [[2.0, 2.0], [0.0, 0.0]] // ]
Computes mean and variance across all dimensions except for the feature_index
dimension and normalizes the operand
tensor producing output
, batch_mean
and batch_var
tensors. More formally, this operation can be expressed as a decomposition to existing StableHLO operations using Python syntax as follows:
def compute_mean(operand, feature_index): (sum,) = reduce( inputs=[operand], init_values=[constant(0, element_type(operand))], dimensions=[i for i in range(rank(operand)) if i != feature_index], body=lambda x, y: add(x, y)) divisor = constant(size(operand) / dim(operand, feature_index), element_type(operand)) divisor_bcast = broadcast_in_dim(divisor, [], type(sum)) return divide(sum, divisor_bcast) def compute_variance(operand, feature_index): mean = compute_mean(operand, feature_index) mean_bcast = broadcast_in_dim(mean, [feature_index], type(operand)) centered_operand = subtract(operand, mean_bcast) return compute_mean(mul(centered_operand, centered_operand), feature_index) def batch_norm_training(operand, scale, offset, epsilon, feature_index): mean = compute_mean(operand, feature_index) variance = compute_variance(operand, feature_index) return batch_norm_inference(operand, scale, offset, mean, variance, epsilon, feature_index), mean, variance
For quantized types, performs dequantize_batch_norm_grad_or_training_quantize(lambda operand, scale, offset: batch_norm_training(operand, scale, offset, epsilon, feature_index), operand, scale, offset, type(output), type(batch_mean), type(batch_var))
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point type or per-tensor quantized tensor | (C1) |
(I2) | scale | 1-dimensional tensor of floating-point or per-tensor quantized | (C2), (C3) |
(I3) | offset | 1-dimensional tensor of floating-point or per-tensor quantized | (C2), (C4) |
(I4) | epsilon | constant of type f32 | (C1), (C3-C6) |
(I5) | feature_index | constant of type si64 | (C1), (C3-C6) |
Name | Type | Constraints |
---|---|---|
output | tensor of floating-point type or per-tensor quantized tensor | (C7) |
batch_mean | 1-dimensional tensor of floating-point or per-tensor quantized | (C2), (C5) |
batch_var | 1-dimensional tensor of floating-point or per-tensor quantized | (C2), (C6) |
0 <= feature_index < rank(operand)
.operand
, scale
, offset
, batch_mean
, batch_var
and output
have the same baseline_element_type
.size(scale) = dim(operand, feature_index)
.size(offset) = dim(operand, feature_index)
.size(batch_mean) = dim(operand, feature_index)
.size(batch_var) = dim(operand, feature_index)
.baseline_type(output) = baseline_type(operand)
.// %operand: [ // [[1.0, 2.0], [3.0, 4.0]], // [[3.0, 4.0], [1.0, 2.0]] // ] // %scale: [1.0, 1.0] // %offset: [1.0, 1.0] %output, %batch_mean, %batch_var = "stablehlo.batch_norm_training"(%operand, %scale, %offset) { epsilon = 0.0 : f32, feature_index = 2 : i64 } : (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>) -> (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>) // %output: [ // [[0.0, 0.0], [2.0, 2.0]], // [[2.0, 2.0], [0.0, 0.0]] // ] // %batch_mean: [2.0, 3.0] // %batch_var: [1.0, 1.0]
Performs a bitcast operation on operand
tensor and produces a result
tensor where the bits of the entire operand
tensor are reinterpreted using the type of the result
tensor.
More formally, given E = element_type(operand)
, E' = element_type(result)
, and R = rank(operand)
:
num_bits(E') < num_bits(E)
, bits(result[i0, ..., iR-1, :]) = bits(operand[i0, ..., iR-1])
.num_bits(E') > num_bits(E)
, bits(result[i0, ..., iR-2]) = bits(operand[i0, ..., iR-2, :])
.num_bits(E') = num_bits(E)
, bits(result[i0, ..., iR-1]) = bits(operand[i0, ..., iR-1])
.bits
returns in-memory representation of a given value, and its behavior is implementation-defined because the exact representation of tensors is implementation-defined, and the exact representation of element types is implementation-defined as well.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or quantized tensor | (C1-C2) |
Name | Type | Constraints |
---|---|---|
result | tensor or quantized tensor | (C1-C2) |
E = is_quantized(operand) ? storage_type(operand) : element_type(operand)
, E' = is_quantized(result) ? storage_type(result) : element_type(result)
, and R = rank(operand)
:num_bits(E') = num_bits(E)
, shape(result) = shape(operand)
.num_bits(E') < num_bits(E)
:rank(result) = R + 1
.dim(result, i) = dim(operand, i)
for all 0 <= i < R
.dim(result, R) * num_bits(E') = num_bits(E)
.num_bits(E') > num_bits(E)
:rank(result) = R - 1
.dim(result, i) = dim(operand, i)
for all 0 <= i < R
.dim(operand, R - 1) * num_bits(E) = num_bits(E')
.is_complex(operand) or is_complex(result)
, then is_complex(operand) and is_complex(result)
.// %operand: 0x0123456789ABCDEF %result = "stablehlo.bitcast_convert"(%operand) : (tensor<f64>) -> tensor<4xf16> // %result: [0xCDEF, 0x89AB, 0x4567, 0x0123] // little-endian representation
Expands the dimensions and/or rank of an input tensor by duplicating the data in the operand
tensor and produces a result
tensor. More formally, result[result_index] = operand[operand_index]
where for all d
in axes(operand)
:
operand_index[d] = 0
if dim(operand, d) = 1
.operand_index[d] = result_index[broadcast_dimensions[d]]
otherwise.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or quantized tensor | (C1-C2), (C5-C6) |
(I2) | broadcast_dimensions | 1-dimensional tensor constant of type si64 | (C2-C6) |
Name | Type | Constraints |
---|---|---|
result | tensor or quantized tensor | (C1), (C3), (C5-C6) |
element_type(result)
is given by:element_type(operand)
, if !is_per_axis_quantized(operand)
.element_type(operand)
except that quantization_dimension(operand)
, scales(operand)
, and zero_points(operand)
may differ from quantization_dimension(result)
, scales(result)
, and zero_points(result)
resp., otherwise.size(broadcast_dimensions) = rank(operand)
.0 <= broadcast_dimensions < rank(result)
.is_unique(broadcast_dimensions)
.d
in axes(operand)
:dim(operand, d) = 1
ordim(operand, d) = dim(result, broadcast_dimensions[d])
.is_per_axis_quantized(result)
:quantization_dimension(result) = broadcast_dimensions[quantization_dimension(operand)]
.dim(operand, quantization_dimension(operand)) = 1
, then scales(result)[i] = scales(operand)[0] and zero_points(result)[i] = zero_points(operand)[0] for i in range(dim(result, quantization_dimension(result)))
.// %operand: [ // [1, 2, 3] // ] %result = "stablehlo.broadcast_in_dim"(%operand) { broadcast_dimensions = array<i64: 2, 1> } : (tensor<1x3xi32>) -> tensor<2x3x2xi32> // %result: [ // [ // [1, 1], // [2, 2], // [3, 3] // ], // [ // [1, 1], // [2, 2], // [3, 3] // ] // ]
Produces the output from executing exactly one function from branches
depending on the value of index
. More formally, result = selected_branch()
where:
selected_branch = branches[index]
if 0 <= index < size(branches)
.selected_branch = branches[-1]
otherwise.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | index | 0-dimensional tensor of type si32 | |
(I2) | branches | variadic number of functions | (C1-C4) |
Name | Type | Constraints |
---|---|---|
results | variadic number of tensors, quantized tensors or tokens | (C4) |
0 < size(branches)
.input_types(branches...) = []
.same(output_types(branches...))
.type(results...) = output_types(branches[0])
.// %index: -1 // %result_branch0: [0, 0] // %result_branch1: [1, 1] %result0, %result1 = "stablehlo.case"(%index) ({ "stablehlo.return"(%result_branch0, %result_branch0) : (tensor<2xi64>, tensor<2xi64>) -> () }, { "stablehlo.return"(%result_branch1, %result_branch1) : (tensor<2xi64>, tensor<2xi64>) -> () }) : (tensor<i32>) -> (tensor<2xi64>, tensor<2xi64>) // %result0: [1, 1] // %result1: [1, 1]
Performs element-wise cubic root operation on operand
tensor and produces a result
tensor. Depending on the element type, does the following:
rootn(x, 3)
from IEEE-754.dequantize_op_quantize(cbrt, operand, type(result))
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand: [0.0, 1.0, 8.0, 27.0] %result = "stablehlo.cbrt"(%operand) : (tensor<4xf64>) -> tensor<4xf64> // %result: [0.0, 1.0, 2.0, 3.0]
Performs element-wise ceil of operand
tensor and produces a result
tensor. Implements the roundToIntegralTowardPositive
operation from the IEEE-754 specification. For quantized types, performs dequantize_op_quantize(ceil, operand, type(result))
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand: [-0.8166, -0.2530, 0.2530, 0.8166, 2.0] %result = "stablehlo.ceil"(%operand) : (tensor<5xf32>) -> tensor<5xf32> // %result: [-0.0, -0.0, 1.0, 1.0, 2.0]
Computes the Cholesky decomposition of a batch of matrices.
More formally, for all i
in index_space(result)
, result[i0, ..., iR-3, :, :]
is a Cholesky decomposition of a[i0, ..., iR-3, :, :]
, in the form of either of a lower-triangular (if lower
is true
) or upper-triangular (if lower
is false
) matrix. The output values in the opposite triangle, i.e. the strict upper triangle or strict lower triangle correspondingly, are implementation-defined.
If there exists i
where the input matrix is not an Hermitian positive-definite matrix, then the behavior is undefined.
For quantized types, performs dequantize_op_quantize(lambda operand: cholesky(operand, lower), a, type(result))
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | a | tensor of floating-point or complex type or per-tensor quantized tensor | (C1-C3) |
(I2) | lower | 0-dimensional tensor constant of type i1 |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(a) = baseline_type(result)
.2 <= rank(a)
.dim(a, -2) = dim(a, -1)
.// %a: [ // [1.0, 2.0, 3.0], // [2.0, 20.0, 26.0], // [3.0, 26.0, 70.0] // ] %result = "stablehlo.cholesky"(%a) { lower = true } : (tensor<3x3xf32>) -> tensor<3x3xf64> // %result: [ // [1.0, 0.0, 0.0], // [2.0, 4.0, 0.0], // [3.0, 5.0, 6.0] // ]
Clamps every element of the operand
tensor between a minimum and maximum value and produces a result
tensor. More formally, result[result_index] = minimum(maximum(operand[result_index], min_element), max_element)
, where min_element = rank(min) = 0 ? min[] : min[result_index]
, max_element = rank(max) = 0 ? max[] : max[result_index]
. For quantized types, performs dequantize_op_quantize(clamp, min, operand, max, type(result))
.
Imposing an ordering on complex numbers involves surprising semantics, so in the future we are planning to remove support for complex numbers for this operation (#560).
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | min | tensor or per-tensor quantized tensor | (C1), (C3) |
(I2) | operand | tensor or per-tensor quantized tensor | (C1-C4) |
(I3) | max | tensor or per-tensor quantized tensor | (C2), (C3) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C4) |
rank(min) = 0 or shape(min) = shape(operand)
.rank(max) = 0 or shape(max) = shape(operand)
.baseline_element_type(min) = baseline_element_type(operand) = baseline_element_type(max)
.baseline_type(operand) = baseline_type(result)
.// %min: [5, 10, 15] // %operand: [3, 13, 23] // %max: [10, 15, 20] %result = "stablehlo.clamp"(%min, %operand, %max) : (tensor<3xi32>, tensor<3xi32>, tensor<3xi32>) -> tensor<3xi32> // %result: [5, 13, 20]
Within each process group in the StableHLO process grid, send the value of the operand
tensor from the source process to the target processes and produce a result
tensor.
The operation splits the StableHLO process grid into process_groups
which is defined as follows:
cross_replica(replica_groups)
if channel_id <= 0
.cross_partition(replica_groups)
if channel_id > 0
.Afterwards, result@process
is given by:
operand@process_groups[i, 0]
if there exists an i
such that the process is in process_groups[i]
.broadcast_in_dim(constant(is_quantized(result) ? quantize(0, element_type(result)) : 0, element_type(result)), [], type(result))
otherwise.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or per-tensor quantized tensor | (C3) |
(I2) | replica_groups | variadic number of 1-dimensional tensor constants of type si64 | (C1), (C2) |
(I3) | channel_id | constant of type si64 |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C3) |
is_unique(replica_groups)
.0 <= replica_groups < N
where N
is defined as:num_replicas
if cross_replica
is used.num_partitions
if cross_partition
is used.type(result) = type(operand)
.// num_replicas: 4 // num_partitions: 1 // %operand@(0, 0): [[1, 2]] // %operand@(1, 0): [[3, 4]] // %operand@(2, 0): [[5, 6]] // %operand@(3, 0): [[7, 8]] %result = "stablehlo.collective_broadcast"(%operand) { replica_groups = dense<[[2, 1]]> : tensor<1x2xi64>, channel_handle = #stablehlo.channel_handle<handle = 0, type = 0> } : (tensor1x2xi64>) -> tensor<1x2xi64> // %result@(0, 0): [[0, 0]] // %result@(1, 0): [[5, 6]] // %result@(2, 0): [[5, 6]] // %result@(3, 0): [[0, 0]]
Within each process group in the StableHLO process grid, sends the value of the operand
tensor from the source process to the target process and produces a result
tensor.
The operation splits the StableHLO process grid into process_groups
which is defined as follows:
cross_replica(source_target_pairs)
if channel_id <= 0
.cross_partition(source_target_pairs)
if channel_id > 0
.Afterwards, result@process
is given by:
operand@process_groups[i, 0]
, if there exists an i
such that process_groups[i, 1] = process
.broadcast_in_dim(constant(is_quantized(result) ? quantize(0, element_type(result)) : 0, element_type(result)), [], type(result))
otherwise.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or per-tensor quantized tensor | (C5) |
(I2) | source_target_pairs | 2-dimensional tensor constant of type si64 | (C1-C4) |
(I3) | channel_id | constant of type si64 |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C1) |
dim(source_target_pairs, 1) = 2
.is_unique(source_target_pairs[:, 0])
.is_unique(source_target_pairs[:, 1])
.0 <= source_target_pairs < N
, where N
is defined as:num_replicas
if cross_replica
is used.num_partitions
if cross_partition
is used.type(result) = type(operand)
.// num_replicas: 3 // num_partitions: 1 // %operand@(0, 0): [[1, 2], [3, 4]] // %operand@(1, 0): [[5, 6], [7, 8]] // %operand@(2, 0): [[9, 10], [11, 12]] %result = "stablehlo.collective_permute"(%operand) { source_target_pairs = dense<[[0, 1], [1, 2]]> : tensor<2x2xi64>, channel_handle = #stablehlo.channel_handle<handle = 0, type = 0> } : (tensor<2x2xi64>) -> tensor<2x2xi64> // // %result@(0, 0): [[0, 0], [0, 0]] // %result@(1, 0): [[1, 2], [3, 4]] // %result@(2, 0): [[5, 6], [7, 8]]
Performs element-wise comparison of lhs
and rhs
tensors according to comparison_direction
and compare_type
, and produces a result
tensor.
The values of comparison_direction
and compare_type
have the following semantics:
For boolean and integer element types:
EQ
: lhs = rhs
.NE
: lhs != rhs
.GE
: lhs >= rhs
.GT
: lhs > rhs
.LE
: lhs <= rhs
.LT
: lhs < rhs
.For floating-point element types with compare_type = FLOAT
, the op implements the following IEEE-754 operations:
EQ
: compareQuietEqual
.NE
: compareQuietNotEqual
.GE
: compareQuietGreaterEqual
.GT
: compareQuietGreater
.LE
: compareQuietLessEqual
.LT
: compareQuietLess
.For floating-point element types with compare_type = TOTALORDER
, the op uses the combination of totalOrder
and compareQuietEqual
operations from IEEE-754. This feature appears to be unused, so in the future, we are planning to remove it (#584).
For complex element types, lexicographic comparison of (real, imag)
pairs is performed using the provided comparison_direction
and compare_type
. Imposing an ordering on complex numbers involves surprising semantics, so in the future we are planning to remove support for complex numbers when comparison_direction
is GE
, GT
, LE
or LT
(#560).
For quantized types. performs dequantize_compare(lhs, rhs, comparison_direction)
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor or per-tensor quantized tensor | (C1-C3) |
(I2) | rhs | tensor or per-tensor quantized tensor | (C1-C2) |
(I3) | comparison_direction | enum of EQ , NE , GE , GT , LE , and LT | |
(I4) | compare_type | enum of FLOAT , TOTALORDER , SIGNED , and UNSIGNED | (C3) |
Name | Type | Constraints |
---|---|---|
result | tensor of boolean type | (C2) |
baseline_element_type(lhs) = baseline_element_type(rhs)
.shape(lhs) = shape(rhs) = shape(result)
.compare_type
is defined as:SIGNED
if is_signed_integer(element_type(lhs))
.UNSIGNED
if is_unsigned_integer(element_type(lhs)) or is_boolean(element_type(lhs))
.FLOAT
or TOTALORDER
if is_float(element_type(lhs))
.FLOAT
if is_complex(element_type(lhs))
.// %lhs: [1.0, 3.0] // %rhs: [1.1, 2.9] %result = "stablehlo.compare"(%lhs, %rhs) { comparison_direction = #stablehlo<comparison_direction LT>, compare_type = #stablehlo<comparison_type FLOAT> } : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xi1> // %result: [true, false]
Performs element-wise conversion to a complex value from a pair of real and imaginary values, lhs
and rhs
, and produces a result
tensor.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor of type f32 or f64 | (C1-C3) |
(I2) | rhs | tensor of type f32 or f64 | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of complex type | (C2), (C3) |
type(lhs) = type(rhs)
.shape(result) = shape(lhs)
.element_type(result)
has type complex<E>
where E = element_type(lhs)
.// %lhs: [1.0, 3.0] // %rhs: [2.0, 4.0] %result = "stablehlo.complex"(%lhs, %rhs) : (tensor<2xf64>, tensor<2xf64>) -> tensor<2xcomplex<f64>> // %result: [(1.0, 2.0), (3.0, 4.0)]
Encapsulates an operation made up (composed) of other StableHLO operations, taking inputs
and composite_attributes
and producing results
. The semantics of the op are implemented by the decomposition
attribute. The composite
op can be replaced with its decomposition without changing program semantics. In cases where inlining the decomposition does not provide the same op semantics, prefer using custom_call
.
The version
field (defaults to 0
) is used to denote when a composite's semantics change.
Label | Name | Type |
---|---|---|
(I1) | inputs | variadic number of values |
(I2) | name | constant of type string |
(I3) | composite_attributes | attribute dictionary |
(I4) | decomposition | constant of type string |
(I5) | version | constant of type si32 |
Name | Type |
---|---|
results | variadic number of values |
is_namespaced_op_name(name)
is_defined_in_parent_scope(decomposition)
types(inputs...) == input_types(decomposition)
types(results...) == output_types(decomposition)
%results = "stablehlo.composite"(%input0, %input1) { name = "my_namespace.my_op", composite_attributes = { my_attribute = "my_value" }, decomposition = @my_op, version = 1 : i32 } : (tensor<f32>, tensor<f32>) -> tensor<f32>
Concatenates inputs
along dimension
dimension in the same order as the given arguments and produces a result
tensor. More formally, result[i0, ..., id, ..., iR-1] = inputs[k][i0, ..., kd, ..., iR-1]
, where:
id = d0 + ... + dk-1 + kd
.d
is equal to dimension
, and d0
, ... are d
th dimension sizes of inputs
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | inputs | variadic number of tensors or per-tensor quantized tensors | (C1-C6) |
(I2) | dimension | constant of type si64 | (C2), (C4), (C6) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C5-C6) |
same(element_type(inputs...))
.same(shape(inputs...))
except for dim(inputs..., dimension)
.0 < size(inputs)
.0 <= dimension < rank(inputs[0])
.element_type(result) = element_type(inputs[0])
.shape(result) = shape(inputs[0])
except for:dim(result, dimension) = dim(inputs[0], dimension) + ...
.// %input0: [[1, 2], [3, 4], [5, 6]] // %input1: [[7, 8]] %result = "stablehlo.concatenate"(%input0, %input1) { dimension = 0 : i64 } : (tensor<3x2xi64>, tensor<1x2xi64>) -> tensor<4x2xi64> // %result: [[1, 2], [3, 4], [5, 6], [7, 8]]
Produces an output
tensor from a constant value
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | value | constant | (C1) |
Name | Type | Constraints |
---|---|---|
output | tensor or quantized tensor | (C1) |
type(value) = type(output)
.%output = "stablehlo.constant"() { value = dense<[[0.0, 1.0], [2.0, 3.0]]> : tensor<2x2xf32> } : () -> tensor<2x2xf32> // %output: [[0.0, 1.0], [2.0, 3.0]]
Performs an element-wise conversion from one element type to another on operand
tensor and produces a result
tensor.
For boolean-to-any-supported-type conversions, the value false
is converted to zero, and the value true
is converted to one. For any-supported-type-to-boolean conversions, a zero value is converted to false
, and non-zero values are converted to true
. See below for how this work for complex types.
For conversions involving integer-to-integer, integer-to-floating-point or floating-point-to-floating-point, if the source value can be exactly represented in the destination type, the result value is that exact representation. Otherwise, the behavior is TBD (#180).
For conversions involving floating-point-to-integer, the fractional part is truncated. If the truncated value cannot be represented in the destination type, the behavior is TBD (#180).
Conversion involving complex-to-complex follow the same behavior of floating-point-to-floating-point conversions for converting real and imaginary parts.
For complex-to-any-other-type and any-other-type-to-complex conversions, the source imaginary value is ignored or the destination imaginary value is zeroed, respectively. The conversion of the real part follows the floating-point conversions.
In principle, this operation could express dequantization (conversion from quantized tensors to regular tensors), quantization (conversion from regular tensors to quantized tensors) and requantization (conversion between quantized tensors), but at the moment we have dedicated operations for that - uniform_dequantize
for the first use case and uniform_quantize
for the second and the third use cases. In the future, these two ops may be merged into convert
(#1576).
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor | (C1) |
shape(operand) = shape(result)
.// %operand: [-1, 0, 1] %result = "stablehlo.convert"(%operand) : (tensor<3xi64>) -> tensor<3xcomplex<f64>> // %result: [(-1.0, 0.0), (0.0, 0.0), (1.0, 0.0)]
Computes dot products between windows of lhs
and slices of rhs
and produces result
. The following diagram shows how elements in result
are computed from lhs
and rhs
using a concrete example.
More formally, consider the following reframing of the inputs in terms of lhs
in order to be able to express windows of lhs
:
lhs_window_dimensions = lhs_shape(dim(lhs, input_batch_dimension), dim(rhs, kernel_spatial_dimensions), dim(lhs, input_feature_dimension))
.lhs_window_strides = lhs_shape(1, window_strides, 1)
.lhs_padding = lhs_shape([0, 0], padding, [0, 0])
.lhs_base_dilations = lhs_shape(1, lhs_dilation, 1)
.lhs_window_dilations = lhs_shape(1, rhs_dilation, 1)
.This reframing uses the following helper functions:
lhs_shape(n, hw, c) = permute([n] + hw + [c], [input_batch_dimension] + input_spatial_dimensions + [input_feature_dimension])
.result_shape(n1, hw, c1) = permute([n1] + hw + [c1], [output_batch_dimension] + output_spatial_dimensions + [output_feature_dimension])
.permute([j0, j1, ..., jR-1], permutation) = [i0, i1, ..., iR-1]
where j[d] = i[permutation[d]]
.If feature_group_count = 1
and batch_group_count = 1
, then for all output_spatial_index
in index_space(dim(result, output_spatial_dimensions...))
, result[result_shape(:, output_spatial_index, :)] = dot_product
where:
padding_value = constant(0, element_type(lhs))
.padded_lhs = pad(lhs, padding_value, lhs_padding[:, 0], lhs_padding[:, 1], lhs_base_dilations - 1)
.lhs_window_start = lhs_shape(0, output_spatial_index, 0) * lhs_window_strides
.lhs_window = slice(padded_lhs, lhs_window_start, lhs_window_start + lhs_window_dimensions, lhs_window_dilations)
.reversed_lhs_window = reverse(lhs_window, [input_spatial_dimensions[dim] for dim in range(size(window_reversal)) if window_reversal[dim] = true])
. This feature appears to be unused, so in the future we are planning to remove it (#1181).dot_product = dot_general(reversed_lhs_window, rhs, lhs_batching_dimensions=[], lhs_contracting_dimensions=input_spatial_dimensions + [input_feature_dimension], rhs_batching_dimensions=[], rhs_contracting_dimensions=kernel_spatial_dimensions + [kernel_input_feature_dimension])
.If feature_group_count > 1
:
lhses = split(lhs, feature_group_count, input_feature_dimension)
.rhses = split(rhs, feature_group_count, kernel_output_feature_dimension)
.results... = convolution(lhses..., rhses..., ..., feature_group_count=1, ...)
.result = concatenate(results, output_feature_dimension)
.If batch_group_count > 1
:
lhses = split(lhs, batch_group_count, input_batch_dimension)
.rhses = split(rhs, batch_group_count, kernel_output_feature_dimension)
.results... = convolution(lhses..., rhses..., ..., batch_group_count=1, ...)
.result = concatenate(results, output_feature_dimension)
.For quantized types, performs dequantize_op_quantize( lambda lhs, rhs: convolution(lhs, rhs, window_strides, padding, lhs_dilation, rhs_dilation, window_reversal, input_batch_dimension, input_feature_dimension, input_spatial_dimensions, kernel_input_feature_dimension, kernel_output_feature_dimension, kernel_spatial_dimensions, output_batch_dimension, output_feature_dimension, output_spatial_dimensions, feature_group_count, batch_group_count, precision_config), lhs, rhs, type(result))
.
For hybrid quantized types, performs hybrid_dequantize_then_op( lambda lhs, rhs: convolution(lhs, rhs, window_strides, padding, lhs_dilation, rhs_dilation, window_reversal, input_batch_dimension, input_feature_dimension, input_spatial_dimensions, kernel_input_feature_dimension, kernel_output_feature_dimension, kernel_spatial_dimensions, output_batch_dimension, output_feature_dimension, output_spatial_dimensions, feature_group_count, batch_group_count, precision_config), lhs, rhs)
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor or per-tensor quantized tensor | (C1), (C10-C11), (C14) (C25), (C27-C30) |
(I2) | rhs | tensor or quantized tensor | (C1), (C14-C16), (C25), (C27-C32) |
(I3) | window_strides | 1-dimensional tensor constant of type si64 | (C2-C3), (C25) |
(I4) | padding | 2-dimensional tensor constant of type si64 | (C4), (C25) |
(I5) | lhs_dilation | 1-dimensional tensor constant of type si64 | (C5-C6), (C25) |
(I6) | rhs_dilation | 1-dimensional tensor constant of type si64 | (C7-C8), (C25) |
(I7) | window_reversal | 1-dimensional tensor constant of type i1 | (C9) |
(I8) | input_batch_dimension | constant of type si64 | (C10), (C13), (C25) |
(I9) | input_feature_dimension | constant of type si64 | (C11), (C13-C14) |
(I10) | input_spatial_dimensions | 1-dimensional tensor constant of type si64 | (C12), (C13), (C25) |
(I11) | kernel_input_feature_dimension | constant of type si64 | (C14), (C18) |
(I12) | kernel_output_feature_dimension | constant of type si64 | (C15-C16), (C18), (C25), (C32) |
(I13) | kernel_spatial_dimensions | 1-dimensional tensor constant of type si64 | (C17-C18), (C25) |
(I14) | output_batch_dimension | constant of type si64 | (C20), (C25) |
(I15) | output_feature_dimension | constant of type si64 | (C20), (C25), (C33) |
(I16) | output_spatial_dimensions | 1-dimensional tensor constant of type si64 | (C19-C20), (C25) |
(I17) | feature_group_count | constant of type si64 | (C11), (C14), (C16), (C21), (C23) |
(I18) | batch_group_count | constant of type si64 | (C10), (C15), (C22), (C23), (C25) |
(I19) | precision_config | variadic number of enums of DEFAULT , HIGH , and HIGHEST | (C24) |
Name | Type | Constraints |
---|---|---|
result | tensor or quantized tensor | (C25-C28), (C30-C31), (C33) |
N = rank(lhs) = rank(rhs)
.size(window_strides) = N - 2
.0 < window_strides
.shape(padding) = [N - 2, 2]
.size(lhs_dilation) = N - 2
.0 < lhs_dilation
.size(rhs_dilation) = N - 2
.0 < rhs_dilation
.size(window_reversal) = N - 2
.dim(lhs, input_batch_dimension) % batch_group_count = 0
.dim(lhs, input_feature_dimension) % feature_group_count = 0
.size(input_spatial_dimensions) = N - 2
.input_dimensions = [input_batch_dimension] + input_spatial_dimensions + [input_feature_dimension]
:is_unique(input_dimensions)
.0 <= input_dimensions < N
.dim(rhs, kernel_input_feature_dimension) = dim(lhs, input_feature_dimension) / feature_group_count
.dim(rhs, kernel_output_feature_dimension) % batch_group_count = 0
.dim(rhs, kernel_output_feature_dimension) % feature_group_count = 0
.size(kernel_spatial_dimensions) = N - 2
.kernel_dimensions = kernel_spatial_dimensions + [kernel_input_feature_dimension] + [kernel_output_feature_dimension]
:is_unique(kernel_dimensions)
.0 <= kernel_dimensions < N
.size(output_spatial_dimensions) = N - 2
.output_dimensions = [output_batch_dimension] + output_spatial_dimensions + [output_feature_dimension]
:is_unique(output_dimensions)
.0 <= output_dimensions < N
.0 < feature_group_count
.0 < batch_group_count
.feature_group_count = 1 or batch_group_count = 1
.size(precision_config) = 2
.dim(result, result_dim)
is defined as:dim(lhs, input_batch_dimension) / batch_group_count
if result_dim = output_batch_dimension
.dim(rhs, kernel_output_feature_dimension)
if result_dim = output_feature_dimension
.num_windows
otherwise, where:output_spatial_dimensions[spatial_dim] = result_dim
.lhs_dim = input_spatial_dimensions[spatial_dim]
.rhs_dim = kernel_spatial_dimensions[spatial_dim]
.dilated_input_shape[lhs_dim] = dim(lhs, lhs_dim) = 0 ? 0 : (dim(lhs, lhs_dim) - 1) * lhs_dilation[spatial_dim] + 1
.padded_input_shape[lhs_dim] = padding[spatial_dim, 0] + dilated_input_shape[lhs_dim] + padding[spatial_dim, 1]
.dilated_window_shape[lhs_dim] = dim(rhs, rhs_dim) = 0 ? 0 : (dim(rhs, rhs_dim) - 1) * rhs_dilation[spatial_dim] + 1
.is_empty_window[lhs_dim] = padded_input_shape[lhs_dim] = 0 || dilated_window_shape[lhs_dim] > padded_input_shape[lhs_dim]
.num_windows = is_empty_window[lhs_dim] ? 0 : floor((padded_input_shape[lhs_dim] - dilated_window_shape[lhs_dim]) / window_strides[spatial_dim]) + 1
.rank(result) = N
.element_type(lhs) = element_type(rhs) = element_type(result)
.is_quantized(lhs) = is_quantized(result) and is_quantized(rhs)
.is_per_axis_quantized(rhs)
, then quantization_dimension(rhs) = kernel_output_feature_dimension
.is_per_axis_quantized(result)
, then quantization_dimension(result) = output_feature_dimension
.is_quantized(lhs)
:storage_type(lhs) = storage_type(rhs)
.expressed_type(lhs) = expressed_type(rhs) = expressed_type(result)
.is_per_tensor_quantized(rhs)
, then is_per_tensor_quantized(result)
.!is_quantized(lhs)
:element_type(lhs) = expressed_type(rhs) = element_type(result)
.// %lhs: [[ // [ // [1], [2], [5], [6] // ], // [ // [3], [4], [7], [8] // ], // [ // [10], [11], [14], [15] // ], // [ // [12], [13], [16], [17] // ] // ]] // // %rhs : [ // [[[1]], [[1]], [[1]]], // [[[1]], [[1]], [[1]]], // [[[1]], [[1]], [[1]]] // ] %result = "stablehlo.convolution"(%lhs, %rhs) { window_strides = array<i64: 4, 4>, padding = dense<0> : tensor<2x2xi64>, lhs_dilation = array<i64: 2, 2>, rhs_dilation = array<i64: 1, 1>, window_reversal = array<i1: false, false>, // In the StableHLO dialect, dimension numbers are encoded via: // `[<input dimensions>]x[<kernel dimensions>]->[output dimensions]`. // "b" is batch dimension, "f" is feature dimension, // "i" is input feature dimension, "o" is output feature dimension, // "0/1/etc" are spatial dimensions. dimension_numbers = #stablehlo.conv<[b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f]>, batch_group_count = 1 : i64, feature_group_count = 1 : i64, precision_config = [#stablehlo<precision DEFAULT>, #stablehlo<precision DEFAULT>] } : (tensor<1x4x4x1xi64>, tensor<3x3x1x1xi64>) -> tensor<1x2x2x1xi64> // %result: [[ // [[10], [26]], // [[46], [62]] // ]]
Performs element-wise cosine operation on operand
tensor and produces a result
tensor. Depending on the element type, does the following:
cos
from IEEE-754.dequantize_op_quantize(cosine, operand, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand: [ // [0.0, 1.57079632], // [0, pi/2] // [3.14159265, 4.71238898] // [pi, 3pi/2] // ] %result = "stablehlo.cosine"(%operand) : (tensor<2x2xf32>) -> tensor<2x2xf32> // %result: [[1.0, 0.0], [-1.0, 0.0]]
Performs element-wise count of the number of leading zero bits in the operand
tensor and produces a result
tensor.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of integer type | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of integer type | (C1) |
type(operand) = type(result)
.// %operand: [[0, 1], [128, -1]] %result = "stablehlo.count_leading_zeros"(%operand) : (tensor<2x2xi64>) -> tensor<2x2xi64> // %result: [[64, 63], [56, 0]]
Encapsulates an implementation-defined operation call_target_name
that takes inputs
and called_computations
and produces results
. has_side_effect
, backend_config
and api_version
may be used to provide additional implementation-defined metadata.
At the moment, this operation contains a fairly disorganized collection of metadata which reflects organic evolution of its counterpart operation in the XLA compiler. In the future, we are planning to unify this metadata (#741).
Label | Name | Type |
---|---|---|
(I1) | inputs | variadic number of values |
(I2) | call_target_name | constant of type string |
(I3) | has_side_effect | constant of type i1 |
(I4) | backend_config | constant of type string |
(I5) | api_version | constant of type si32 |
(I6) | called_computations | variadic number of constants of type string |
Name | Type |
---|---|
results | variadic number of values |
%results = "stablehlo.custom_call"(%input0) { call_target_name = "foo", has_side_effect = false, backend_config = "bar", api_version = 1 : i32, called_computations = [@foo] } : (tensor<f64>) -> tensor<f64>
Performs element-wise division of dividend lhs
and divisor rhs
tensors and produces a result
tensor. Depending on the element type, does the following:
division
from IEEE-754.dequantize_op_quantize(divide, lhs, rhs, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor of integer, floating-point or complex type or per-tensor quantized tensor | (C1) |
(I2) | rhs | tensor of integer, floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of integer, floating-point, or complex type or per-tensor quantized tensor | (C1) |
baseline_type(lhs) = baseline_type(rhs) = baseline_type(result)
.// %lhs: [17.1, -17.1, 17.1, -17.1] // %rhs: [3.0, 3.0, -3.0, -3.0] %result = "stablehlo.divide"(%lhs, %rhs) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> // %result: [5.66666651, -5.66666651, -5.66666651, 5.66666651]
Computes dot products between slices of lhs
and slices of rhs
and produces a result
tensor.
More formally, result[result_index] = dot_product
, where:
lhs_result_dimensions = [d for d in axes(lhs) and d not in lhs_batching_dimensions and d not in lhs_contracting_dimensions]
.rhs_result_dimensions = [d for d in axes(rhs) and d not in rhs_batching_dimensions and d not in rhs_contracting_dimensions]
.result_batching_index + result_lhs_index + result_rhs_index = result_index
where size(result_batching_index) = size(lhs_batching_dimensions)
, size(result_lhs_index) = size(lhs_result_dimensions)
and size(result_rhs_index) = size(rhs_result_dimensions)
.transposed_lhs = transpose(lhs, lhs_batching_dimensions + lhs_result_dimensions + lhs_contracting_dimensions)
.transposed_lhs_slice = slice(transposed_lhs, result_batching_index + result_lhs_index + [:, ..., :])
.reshaped_lhs_slice = reshape(transposed_lhs_slice, dims(lhs, lhs_contracting_dimensions))
.transposed_rhs = transpose(rhs, rhs_batching_dimensions + rhs_result_dimensions + rhs_contracting_dimensions)
.transposed_rhs_slice = slice(transposed_rhs, result_batching_index + result_rhs_index + [:, ..., :])
.reshaped_rhs_slice = reshape(transposed_rhs_slice, dims(rhs, rhs_contracting_dimensions))
.dot_product = reduce( inputs=[multiply(reshaped_lhs_slice, reshaped_rhs_slice)], init_values=[constant(0, element_type(result))], dimensions=range(size(lhs_contracting_dimensions)), body=lambda x, y: add(x, y))
.For quantized types, performs dequantize_op_quantize( lambda lhs, rhs: dot_general(lhs, rhs, lhs_batching_dimensions, rhs_batching_dimensions, lhs_contracting_dimensions, rhs_contracting_dimensions, precision_config), lhs, rhs, type(result))
.
This only specifies semantics for per-tensor quantization. Per-axis quantization is work in progress (#1574). Also, in the future we may consider adding support for hybrid quantization (#1575).
precision_config
controls the tradeoff between speed and accuracy for computations on accelerator backends. This can be one of the following (at the moment, the semantics of these enum values is underspecified, but we are planning to address this in #755):
DEFAULT
: Fastest calculation, but least accurate approximation to the original number.HIGH
: Slower calculation, but more accurate approximation to the original number.HIGHEST
: Slowest calculation, but most accurate approximation to the original number.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor or per-tensor quantized tensor | (C5-C6), (C9-C10), (C12-C16) |
(I2) | rhs | tensor or quantized tensor | (C7-C10), (C12), (C18-C19) |
(I3) | lhs_batching_dimensions | 1-dimensional tensor constant of type si64 | (C1), (C3), (C5), (C9), (C12) |
(I4) | rhs_batching_dimensions | 1-dimensional tensor constant of type si64 | (C1), (C4), (C7), (C9) |
(I5) | lhs_contracting_dimensions | 1-dimensional tensor constant of type si64 | (C2), (C3), (C6), (C10) |
(I6) | rhs_contracting_dimensions | 1-dimensional tensor constant of type si64 | (C2), (C4), (C8), (C10), (C19) |
(I7) | precision_config | variadic number of enums of DEFAULT , HIGH , and HIGHEST | (C11) |
Name | Type | Constraints |
---|---|---|
result | tensor or quantized tensor | (C12), (C14), (C16), (C18) |
size(lhs_batching_dimensions) = size(rhs_batching_dimensions)
.size(lhs_contracting_dimensions) = size(rhs_contracting_dimensions)
.is_unique(lhs_batching_dimensions + lhs_contracting_dimensions)
.is_unique(rhs_batching_dimensions + rhs_contracting_dimensions)
.0 <= lhs_batching_dimensions < rank(lhs)
.0 <= lhs_contracting_dimensions < rank(lhs)
.0 <= rhs_batching_dimensions < rank(rhs)
.0 <= rhs_contracting_dimensions < rank(rhs)
.dim(lhs, lhs_batching_dimensions...) = dim(rhs, rhs_batching_dimensions...)
.dim(lhs, lhs_contracting_dimensions...) = dim(rhs, rhs_contracting_dimensions...)
.size(precision_config) = 2
.shape(result) = dim(lhs, lhs_batching_dimensions) + dim(lhs, lhs_result_dimensions) + dim(rhs, rhs_result_dimensions)
.element_type(lhs) = element_type(rhs)
.is_quantized(lhs) and is_quantized(rhs) and is_quantized(result)
.storage_type(lhs) = storage_type(rhs)
.expressed_type(lhs) = expressed_type(rhs) = expressed_type(result)
.zero_points(rhs) = 0
.is_per_tensor_quantized(rhs)
, then is_per_tensor_quantized(result)
.is_per_axis_quantized(rhs)
, then quantization_dimension(rhs)
not in rhs_contracting_dimensions
.// %lhs: [ // [[1, 2], // [3, 4]], // [[5, 6], // [7, 8]] // ] // %rhs: [ // [[1, 0], // [0, 1]], // [[1, 0], // [0, 1]] // ] %result = "stablehlo.dot_general"(%lhs, %rhs) { dot_dimension_numbers = #stablehlo.dot< lhs_batching_dimensions = [0], rhs_batching_dimensions = [0], lhs_contracting_dimensions = [2], rhs_contracting_dimensions = [1] >, precision_config = [#stablehlo<precision DEFAULT>, #stablehlo<precision DEFAULT>] } : (tensor<2x2x2xi64>, tensor<2x2x2xi64>) -> tensor<2x2x2xi64> // %result: [ // [[1, 2], // [3, 4]], // [[5, 6], // [7, 8]] // ]
Extracts a slice from the operand
using dynamically-computed starting indices and produces a result
tensor. start_indices
contain the starting indices of the slice for each dimension subject to potential adjustment, and slice_sizes
contain the sizes of the slice for each dimension. More formally, result[result_index] = operand[operand_index]
where:
adjusted_start_indices = clamp(0, start_indices, shape(operand) - slice_sizes)
.operand_index = adjusted_start_indices + result_index
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or per-tensor quantized tensor | (C1), (C2), (C4) |
(I2) | start_indices | variadic number of 0-dimensional tensors of integer type | (C2), (C3) |
(I3) | slice_sizes | 1-dimensional tensor constant of type si64 | (C2), (C4), (C5) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C1), (C5) |
element_type(operand) = element_type(result)
.size(start_indices) = size(slice_sizes) = rank(operand)
.same(type(start_indices...))
.0 <= slice_sizes <= shape(operand)
.shape(result) = slice_sizes
.// %operand: [ // [0, 0, 1, 1], // [0, 0, 1, 1], // [0, 0, 0, 0], // [0, 0, 0, 0] // ] // %start_indices0: -1 // %start_indices1: 3 %result = "stablehlo.dynamic_slice"(%operand, %start_indices0, %start_indices1) { slice_sizes = array<i64: 2, 2> } : (tensor<4x4xi32>, tensor<i64>, tensor<i64>) -> tensor<2x2xi32> // %result: [ // [1, 1], // [1, 1] // ]
Produces a result
tensor which is equal to the operand
tensor except that the slice starting at start_indices
is updated with the values in update
. More formally, result[result_index]
is defined as:
update[update_index]
if 0 <= update_index < shape(update)
where:adjusted_start_indices = clamp(0, start_indices, shape(operand) - shape(update))
.update_index = result_index - adjusted_start_indices
.operand[result_index]
otherwise.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or per-tensor quantized tensor | (C1-C4), (C6) |
(I2) | update | tensor or per-tensor quantized tensor | (C2), (C3), (C6) |
(I3) | start_indices | variadic number of 0-dimensional tensors of integer type | (C4), (C5) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C1) |
type(operand) = type(result)
.element_type(update) = element_type(operand)
.rank(update) = rank(operand)
.size(start_indices) = rank(operand)
.same(type(start_indices...))
.0 <= shape(update) <= shape(operand)
.// %operand: [ // [1, 1, 0, 0], // [1, 1, 0, 0], // [1, 1, 1, 1], // [1, 1, 1, 1] // ] // %update: [ // [1, 1], // [1, 1] // ] // %start_indices0: -1 // %start_indices1: 3 %result = "stablehlo.dynamic_update_slice"(%operand, %update, %start_indices0, %start_indices1) : (tensor<4x4xi32>, tensor<2x2xi32>, tensor<i64>, tensor<i64>) -> tensor<4x4xi32> // %result: [ // [1, 1, 1, 1], // [1, 1, 1, 1], // [1, 1, 1, 1], // [1, 1, 1, 1] // ]
Performs element-wise exponential operation on operand
tensor and produces a result
tensor. Depending on the element type, does the following:
exp
from IEEE-754.dequantize_op_quantize(exponential, operand, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand: [[0.0, 1.0], [2.0, 3.0]] %result = "stablehlo.exponential"(%operand) : (tensor<2x2xf64>) -> tensor<2x2xf64> // %result: [[1.0, 2.7182818284590451], [7.3890560989306504, 20.085536923187668]]
Performs element-wise exponential minus one operation on operand
tensor and produces a result
tensor. Depending on the element type, does the following:
expm1
from IEEE-754.dequantize_op_quantize(exponential_minus_one, operand, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand: [0.0, 1.0] %result = "stablehlo.exponential_minus_one"(%operand) : (tensor<2xf64>) -> tensor<2xf64> // %result: [0.0, 1.71828187]
Performs the forward and inverse Fourier transforms for real and complex inputs/outputs.
fft_type
is one of the following:
FFT
: Forward complex-to-complex FFT.IFFT
: Inverse complex-to-complex FFT.RFFT
: Forward real-to-complex FFT.IRFFT
: Inverse real-to-complex FFT (i.e. takes complex, returns real).More formally, given the function fft
which takes 1-dimensional tensors of complex types as input, produces 1-dimensional tensors of same types as output and computes the discrete Fourier transform:
For fft_type = FFT
, result
is defined as the final result of a series of L computations where L = size(fft_length)
. For example, for L = 3
:
result1[i0, ..., :] = fft(operand[i0, ..., :])
.result2[i0, ..., :, iR-1] = fft(result1[i0, ..., :, iR-1])
.result[i0, ..., :, iR-2, iR-1] = fft(result2[i0, ..., :, iR-2, iR-1])
.Furthermore, given the function ifft
which has the same type signature and computes the inverse of fft
:
For fft_type = IFFT
, result
is defined as the inverse of the computations for fft_type = FFT
. For example, for L = 3
:
result1[i0, ..., :, iR-2, iR-1] = ifft(operand[i0, ..., :, iR-2, iR-1])
.result2[i0, ..., :, iR-1] = ifft(result1[i0, ..., :, iR-1])
.result[i0, ..., :] = ifft(result2[i0, ..., :])
.Furthermore, given the function rfft
which takes 1-dimensional tensors of floating-point types, produces 1-dimensional tensors of complex types of the same floating-point semantics and works as follows:
rfft(real_operand) = truncated_result
wherecomplex_operand... = (real_operand..., 0.0)
.complex_result = fft(complex_operand)
.truncated_result = complex_result[:(rank(complex_result) / 2 + 1)]
.(When the discrete Fourier transform is computed for real operands, the first N/2 + 1
elements of the result unambiguously define the rest of the result, so the result of rfft
is truncated to avoid computing redundant elements).
For fft_type = RFFT
, result
is defined as the final result of a series of L computations where L = size(fft_length)
. For example, for L = 3
:
result1[i0, ..., :] = rfft(operand[i0, ..., :])
.result2[i0, ..., :, iR-1] = fft(result1[i0, ..., :, iR-1])
.result[i0, ..., :, iR-2, iR-1] = fft(result2[i0, ..., :, iR-2, iR-1])
.Finally, given the function irfft
which has the same type signature and computes the inverse of rfft
:
For fft_type = IRFFT
, result
is defined as the inverse of the computations for fft_type = RFFT
. For example, for L = 3
:
result1[i0, ..., :, iR-2, iR-1] = ifft(operand[i0, ..., :, iR-2, iR-1])
.result2[i0, ..., :, iR-1] = ifft(result1[i0, ..., :, iR-1])
.result[i0, ..., :] = irfft(result2[i0, ..., :])
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type | (C1), (C2), (C4), (C5) |
(I2) | fft_type | enum of FFT , IFFT , RFFT , and IRFFT | (C2), (C5) |
(I3) | fft_length | 1-dimensional tensor constant of type si64 | (C1), (C3), (C4) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type | (C2), (C4), (C5) |
size(fft_length) <= rank(operand)
.operand
and result
element types varies:fft_type = FFT
, element_type(operand)
and element_type(result)
have the same complex type.fft_type = IFFT
, element_type(operand)
and element_type(result)
have the same complex type.fft_type = RFFT
, element_type(operand)
is a floating-point type and element_type(result)
is a complex type of the same floating-point semantics.fft_type = IRFFT
, element_type(operand)
is a complex type and element_type(result)
is a floating-point type of the same floating-point semantics.1 <= size(fft_length) <= 3
.operand
and result
, there is a tensor real
of a floating-point type, then shape(real)[-size(fft_length):] = fft_length
.shape(result) = shape(operand)
except for:fft_type = RFFT
, dim(result, -1) = dim(operand, -1) = 0 ? 0 : dim(operand, -1) / 2 + 1
.fft_type = IRFFT
, dim(operand, -1) = dim(result, -1) = 0 ? 0 : dim(result, -1) / 2 + 1
.// %operand: [(1.0, 0.0), (0.0, 0.0), (0.0, 0.0), (0.0, 0.0)] %result = "stablehlo.fft"(%operand) { fft_type = #stablehlo<fft_type FFT>, fft_length = array<i64: 4> } : (tensor<4xcomplex<f32>>) -> tensor<4xcomplex<f32>> // %result: [(1.0, 0.0), (1.0, 0.0), (1.0, 0.0), (1.0, 0.0)]
Performs element-wise floor of operand
tensor and produces a result
tensor. Implements the roundToIntegralTowardNegative
operation from the IEEE-754 specification. For quantized types, performs dequantize_op_quantize(floor, operand, type(result))
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand: [-0.8166, -0.2530, 0.2530, 0.8166, 2.0] %result = "stablehlo.floor"(%operand) : (tensor<5xf32>) -> tensor<5xf32> // %result: [-1.0, -1.0, 0.0, 0.0, 2.0]
Gathers slices from operand
tensor from offsets specified in start_indices
and produces a result
tensor.
The following diagram shows how elements in result
map on elements in operand
using a concrete example. The diagram picks a few example result
indices and explains in detail which operand
indices they correspond to.
More formally, result[result_index] = operand[operand_index]
where:
batch_dims = [d for d in axes(result) and d not in offset_dims]
.batch_index = result_index[batch_dims...]
.start_index
is defined as:start_indices[bi0, ..., :, ..., biN]
where bi
are individual elements in batch_index
and :
is inserted at the index_vector_dim
index, if index_vector_dim
< rank(start_indices)
.[start_indices[batch_index]]
otherwise.d_operand
in axes(operand)
,full_start_index[d_operand] = clamp(start_index[d_start], 0, dim(operand, d_operand) - slice_sizes[d_operand])
if d_operand = start_index_map[d_start]
.full_start_index[d_operand] = 0
otherwise.offset_index = result_index[offset_dims...]
.full_offset_index = [oi0, ..., 0, ..., oiN]
where oi
are individual elements in offset_index
, and 0
is inserted at indices from collapsed_slice_dims
.operand_index = full_start_index + full_offset_index
.If indices_are_sorted
is true
then the implementation can assume that start_indices
are sorted with respect to start_index_map
, otherwise the behavior is undefined. More formally, for all i1 < i2
from indices(result)
, full_start_index(i1) <= full_start_index(i2)
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or per-tensor quantized tensor | (C1), (C7), (C10-C12), (C14) |
(I2) | start_indices | tensor of integer type | (C2), (C3), (C13) |
(I3) | offset_dims | 1-dimensional tensor constant of type si64 | (C1), (C4-C5), (C13) |
(I4) | collapsed_slice_dims | 1-dimensional tensor constant of type si64 | (C1), (C6-C8), (C13) |
(I5) | start_index_map | 1-dimensional tensor constant of type si64 | (C3), (C9), (C10) |
(I6) | index_vector_dim | constant of type si64 | (C2), (C3), (C13) |
(I7) | slice_sizes | 1-dimensional tensor constant of type si64 | (C8), (C11-C13) |
(I8) | indices_are_sorted | constant of type i1 |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C5), (C13-C14) |
rank(operand) = size(offset_dims) + size(collapsed_slice_dims)
.0 <= index_vector_dim <= rank(start_indices)
.size(start_index_map) = index_vector_dim < rank(start_indices) ? dim(start_indices, index_vector_dim) : 1
.is_unique(offset_dims) and is_sorted(offset_dims)
.0 <= offset_dims < rank(result)
.is_unique(collapsed_slice_dims) and is_sorted(collapsed_slice_dims)
.0 <= collapsed_slice_dims < rank(operand)
.slice_sizes[collapsed_slice_dims...] <= 1
.is_unique(start_index_map)
.0 <= start_index_map < rank(operand)
.size(slice_sizes) = rank(operand)
.0 <= slice_sizes <= shape(operand)
.shape(result) = combine(batch_dim_sizes, offset_dim_sizes)
where:batch_dim_sizes = shape(start_indices)
except that the dimension size of start_indices
corresponding to index_vector_dim
is not included.offset_dim_sizes = shape(slice_sizes)
except that the dimension sizes in slice_sizes
corresponding to collapsed_slice_dims
are not included.combine
puts batch_dim_sizes
at axes corresponding to batch_dims
and offset_dim_sizes
at axes corresponding to offset_dims
.element_type(operand) = element_type(result)
.// %operand: [ // [[1, 2], [3, 4], [5, 6], [7, 8]], // [[9, 10],[11, 12], [13, 14], [15, 16]], // [[17, 18], [19, 20], [21, 22], [23, 24]] // ] // %start_indices: [ // [[0, 0], [1, 0], [2, 1]], // [[0, 1], [1, 1], [0, 2]] // ] %result = "stablehlo.gather"(%operand, %start_indices) { dimension_numbers = #stablehlo.gather< offset_dims = [2, 3], collapsed_slice_dims = [0], start_index_map = [1, 0], index_vector_dim = 2>, slice_sizes = array<i64: 1, 2, 2>, indices_are_sorted = false } : (tensor<3x4x2xi32>, tensor<2x3x2xi64>) -> tensor<2x3x2x2xi32> // %result: [ // [ // [[1, 2], [3, 4]], // [[3, 4], [5, 6]], // [[13, 14], [15, 16]] // ], // [ // [[9, 10], [11, 12]], // [[11, 12], [13, 14]], // [[17, 18], [19, 20]] // ] // ]
Produces the size of the given dimension
of the operand
. More formally, result = dim(operand, dimension)
. The Semantics concerns only with the shape component of the type. The element-type could be anything.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or quantized tensor | (C1) |
(I2) | dimension | constant of type si64 | (C1) |
Name | Type |
---|---|
result | 0-dimensional tensor of type si32 |
0 <= dimension < rank(operand)
.// %operand: [[1, 2, 3], [4, 5, 6]] %result = "stablehlo.get_dimension_size"(%operand) { dimension = 1 : i64 } : (tensor<2x3xi64>) -> tensor<i32> // %result: 3
Extracts element at index
position of the operand
tuple and produces a result
. More formally, result = operand[index]
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tuple | (C1), (C2) |
(I2) | index | constant of type si32 | (C1), (C2) |
Name | Type | Constraints |
---|---|---|
result | any supported type | (C2) |
0 <= index < size(operand)
.type(result) = tuple_element_types(operand)[index]
.// %operand: ([1.0, 2.0], (3)) %result = "stablehlo.get_tuple_element"(%operand) { index = 0 : i32 } : (tuple<tensor<2xf32>, tuple<tensor<i32>>>) -> tensor<2xf32> // %result: [1.0, 2.0]
Produces the output from executing exactly one function from true_branch
or false_branch
depending on the value of pred
. More formally, result = pred ? true_branch() : false_branch()
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | pred | 0-dimensional tensor of type i1 | |
(I2) | true_branch | function | (C1-C3) |
(I3) | false_branch | function | (C1), (C2) |
Name | Type | Constraints |
---|---|---|
results | variadic number of tensors, quantized tensors or tokens | (C3) |
input_types(true_branch) = input_types(false_branch) = []
.output_types(true_branch) = output_types(false_branch)
.type(results...) = output_types(true_branch)
.// %result_true_branch: 10 // %result_false_branch: 11 // %pred: true %result = "stablehlo.if"(%pred) ({ "stablehlo.return"(%result_true_branch) : (tensor<i32>) -> () }, { "stablehlo.return"(%result_false_branch) : (tensor<i32>) -> () }) : (tensor<i1>) -> tensor<i32> // %result: 10
Extracts the imaginary part, element-wise, from the operand
and produces a result
tensor. More formally, for each element x
: imag(x) = is_complex(x) ? imaginary_part(x) : constant(0, element_type(result))
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type | (C1), (C2) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point type | (C1), (C2) |
shape(result) = shape(operand)
.element_type(result)
is defined as:complex_element_type(element_type(operand))
if is_complex(operand)
.element_type(operand)
otherwise.// %operand: [(1.0, 2.0), (3.0, 4.0)] %result = "stablehlo.imag"(%operand) : (tensor<2xcomplex<f32>>) -> tensor<2xf32> // %result: [2.0, 4.0]
Reads data from the infeed and produces results
.
Semantics of infeed_config
is implementation-defined.
results
consist of payload values which come first and a token which comes last. In the future, we are planning to split the payload and the token into two separate outputs to improve clarity (#670).
Label | Name | Type |
---|---|---|
(I1) | token | token |
(I2) | infeed_config | constant of type string |
Name | Type | Constraints |
---|---|---|
results | variadic number of tensors, quantized tensors or tokens | (C1-C3) |
0 < size(results)
.is_empty(result[:-1])
or is_tensor(type(results[:-1]))
.is_token(type(results[-1]))
.// %token: !stablehlo.token // infeed_queue[0]: [[1, 2], [3, 4]] // infeed_queue[1]: [[5, 6], [7, 8]] %results0:2 = "stablehlo.infeed"(%token) { infeed_config = "" } : (!stablehlo.token) -> (tensor<2x2xi64>, !stablehlo.token) // results0#0: [[1, 2], [3, 4]] %results1:2 = "stablehlo.infeed"(%token) { infeed_config = "" } : (!stablehlo.token) -> (tensor<2x2xi64>, !stablehlo.token) // results1#0: [[5, 6], [7, 8]]
Fills an output
tensor with values in increasing order starting from zero along the iota_dimension
dimension. More formally,
output[result_index] = constant(is_quantized(output) ? quantize(result_index[iota_dimension], element_type(output)) : result_index[iota_dimension], element_type(output))
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | iota_dimension | si64 | (C1) |
Name | Type | Constraints |
---|---|---|
output | tensor of integer, floating-point or complex type or per-tensor quantized tensor | (C1) |
0 <= iota_dimension < rank(output)
.%output = "stablehlo.iota"() { iota_dimension = 0 : i64 } : () -> tensor<4x5xi32> // %output: [ // [0, 0, 0, 0, 0], // [1, 1, 1, 1, 1], // [2, 2, 2, 2, 2], // [3, 3, 3, 3, 3] // ] %output = "stablehlo.iota"() { iota_dimension = 1 : i64 } : () -> tensor<4x5xi32> // %output: [ // [0, 1, 2, 3, 4], // [0, 1, 2, 3, 4], // [0, 1, 2, 3, 4], // [0, 1, 2, 3, 4] // ]
Performs element-wise check whether the value in x
is finite (i.e. is neither +Inf, -Inf, nor NaN) and produces a y
tensor. Implements the isFinite
operation from the IEEE-754 specification. For quantized types, the result is always true
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | x | tensor of floating-point type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
y | tensor of boolean type | (C1) |
shape(x) = shape(y)
.// Logical values: -Inf, +Inf, NaN, ... // %x: [0xFFF0000000000000, 0x7FF0000000000000, 0x7FF8000000000000, -10.0, -0.0, 0.0, 10.0] %y = "stablehlo.is_finite"(%x) : (tensor<7xf64) -> tensor<7xi1> // %y: [false, false, false, true, true, true, true]
Performs element-wise logarithm operation on operand
tensor and produces a result
tensor. Depending on the element type, does the following:
log
from IEEE-754.dequantize_op_quantize(log, operand, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand: [[1.0, 2.0], [3.0, 4.0]] %result = "stablehlo.log"(%operand) : (tensor<2x2xf64>) -> tensor<2x2xf64> // %result: [[0.0, 0.69314718055994529], [1.0986122886681098, 1.3862943611198906]]
Performs element-wise logarithm plus one operation on operand
tensor and produces a result
tensor. Depending on the element type, does the following:
logp1
from IEEE-754.dequantize_op_quantize(log_plus_one, operand, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand: [0.0, -0.999, 7.0, 6.38905621, 15.0] %result = "stablehlo.log_plus_one"(%operand) : (tensor<5xf64>) -> tensor<5xf64> // %result: [0.0, -6.90776825, 2.07944155, 2.0, 2.77258873]
Performs element-wise logistic operation on operand
tensor and produces a result
tensor. Depending on the element type, does the following:
division(1, addition(1, exp(-x)))
from IEEE-754.dequantize_op_quantize(logistic, operand, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand: [[0.0, 1.0], [2.0, 3.0]] %result = "stablehlo.logistic"(%operand) : (tensor<2x2xf64>) -> tensor<2x2xf64> // %result: [[0.5, 0.73105858], [0.88079708, 0.95257413]]
Applies a map function computation
to inputs
along the dimensions
and produces a result
tensor.
More formally, result[result_index] = computation(inputs...[result_index])
. Note that dimensions
are currently unused and will likely be removed in the future (#487).
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | inputs | variadic number of tensors or per-tensor quantized tensors | (C1-C4) |
(I2) | dimensions | 1-dimensional tensor constant of type si64 | (C3) |
(I3) | computation | function | (C4) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C1), (C4) |
shape(inputs...) = shape(result)
.0 < size(inputs) = N
.dimensions = range(rank(inputs[0]))
.computation
has type (tensor<E0>, ..., tensor<EN-1>) -> tensor<E'>
where Ei = element_type(inputs[i])
and E' = element_type(result)
.// %input0: [[0, 1], [2, 3]] // %input1: [[4, 5], [6, 7]] %result = "stablehlo.map"(%input0, %input1) ({ ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>): %0 = stablehlo.multiply %arg0, %arg1 : tensor<i64> stablehlo.return %0 : tensor<i64> }) { dimensions = array<i64: 0, 1> } : (tensor<2x2xi64>, tensor<2x2xi64>) -> tensor<2x2xi64> // %result: [[0, 5], [12, 21]]
Performs element-wise max operation on tensors lhs
and rhs
and produces a result
tensor. Depending on the element type, does the following:
maximum
from IEEE-754.(real, imaginary)
pair. Imposing an ordering on complex numbers involves surprising semantics, so in the future we are planning to remove support for complex numbers for this operation (#560).dequantize_op_quantize(maximum, lhs, rhs, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor or per-tensor quantized tensor | (C1) |
(I2) | rhs | tensor or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C1) |
baseline_type(lhs) = baseline_type(rhs) = baseline_type(result)
.// %lhs: [[1, 2], [7, 8]] // %rhs: [[5, 6], [3, 4]] %result = "stablehlo.maximum"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32> // %result: [[5, 6], [7, 8]]
Performs element-wise min operation on tensors lhs
and rhs
and produces a result
tensor. Depending on the element type, does the following:
minimum
from IEEE-754.(real, imaginary)
pair. Imposing an ordering on complex numbers involves surprising semantics, so in the future we are planning to remove support for complex numbers for this operation (#560).dequantize_op_quantize(minimum, lhs, rhs, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor or per-tensor quantized tensor | (C1) |
(I2) | rhs | tensor or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C1) |
baseline_type(lhs) = baseline_type(rhs) = baseline_type(result)
.// %lhs: [[1, 2], [7, 8]] // %rhs: [[5, 6], [3, 4]] %result = "stablehlo.minimum"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32> // %result: [[1, 2], [3, 4]]
Performs element-wise product of two tensors lhs
and rhs
and produces a result
tensor. Depending on the element type, does the following:
multiplication
from IEEE-754.dequantize_op_quantize(multiply, lhs, rhs, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor or per-tensor quantized tensor | (C1) |
(I2) | rhs | tensor or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %lhs: [[1, 2], [3, 4]] // %rhs: [[5, 6], [7, 8]] %result = "stablehlo.multiply"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32> // %result: [[5, 12], [21, 32]]
Performs element-wise negation of operand
tensor and produces a result
tensor. Depending on the element type, does the following:
negate
from IEEE-754.dequantize_op_quantize(negate, operand, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of integer, floating-point, or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of integer, floating-point, or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// Negation operation with integer Tensors // %operand: [0, -2] %result = "stablehlo.negate"(%operand) : (tensor<2xi32>) -> tensor<2xi32> // %result: [0, 2] // Negation operation with with complex tensors // %operand: (2.5, 0.0) %result = "stablehlo.negate"(%operand) : (tensor<1xcomplex<f32>>) -> tensor<1xcomplex<f32>> // %result: [-2.5, -0.0]
Performs element-wise NOT of tensor operand
and produces a result
tensor. Depending on the element type, does the following:
Name | Type | Constraints |
---|---|---|
operand | tensor of boolean or integer type | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of boolean or integer type | (C1) |
type(operand) = type(result)
.// Bitwise operation with with integer tensors // %operand: [[1, 2], [3, 4]] %result = "stablehlo.not"(%operand) : (tensor<2x2xi32>) -> tensor<2x2xi32> // %result: [[-2, -3], [-4, -5]] // Bitwise operation with with boolean tensors // %operand: [true, false] %result = "stablehlo.not"(%operand) : (tensor<2xi1>) -> tensor<2xi1> // %result: [false, true]
Ensures that the operations that produce the operand
are executed before any operations that depend on the result
and prevents compiler transformations from moving operations across the barrier. Other than that, the operation is an identity, i.e. result = operand
.
Name | Type | Constraints |
---|---|---|
operand | variadic number of tensors, per-tensor quantized tensors or tokens | (C1) |
Name | Type | Constraints |
---|---|---|
result | variadic number of tensors, per-tensor quantized tensors or tokens | (C1) |
type(operand...) = type(result...)
.// %operand0: 0.0 // %operand1: 1.0 %result0, %result1 = "stablehlo.optimization_barrier"(%operand0, %operand1) : (tensor<f32>, tensor<f32>) -> (tensor<f32>, tensor<f32>) // %result0: 0.0 // %result1: 1.0
Performs element-wise OR of two tensors lhs
and rhs
and produces a result
tensor. Depending on the element type, does the following:
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor of integer or boolean type | (C1) |
(I2) | rhs | tensor of integer or boolean type | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of integer or boolean type | (C1) |
type(lhs) = type(rhs) = type(result)
.// Bitwise operation with with integer tensors // %lhs: [[1, 2], [3, 4]] // %rhs: [[5, 6], [7, 8]] %result = "stablehlo.or"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32> // %result: [[5, 6], [7, 12]] // Logical operation with with boolean tensors // %lhs: [[false, false], [true, true]] // %rhs: [[false, true], [false, true]] %result = "stablehlo.or"(%lhs, %rhs) : (tensor<2x2xi1>, tensor<2x2xi1>) -> tensor<2x2xi1> // %result: [[false, true], [true, true]]
Writes inputs
to the outfeed and produces a result
token.
Semantics of outfeed_config
is implementation-defined.
Label | Name | Type |
---|---|---|
(I1) | inputs | variadic number of tensors or quantized tensors |
(I2) | token | token |
(I3) | outfeed_config | constant of type string |
Name | Type |
---|---|
result | token |
%result = "stablehlo.outfeed"(%input0, %token) { outfeed_config = "" } : (tensor<2x2x2xi64>, !stablehlo.token) -> !stablehlo.token
Expands operand
by padding around the tensor as well as between the elements of the tensor with the given padding_value
.
edge_padding_low
and edge_padding_high
specify the amount of padding added at the low-end (next to index 0) and the high-end (next to the highest index) of each dimension respectively. The amount of padding can be negative, where the absolute value of negative padding indicates the number of elements to remove from the specified dimension.
interior_padding
specifies the amount of padding added between any two elements in each dimension which may not be negative. Interior padding occurs before edge padding such that negative edge padding will remove elements from the interior-padded operand.
More formally, result[result_index]
is defined as:
operand[operand_index]
if result_index = edge_padding_low + operand_index * (interior_padding + 1)
.padding_value
otherwise.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or per-tensor quantized tensor | (C1), (C2), (C4) |
(I2) | padding_value | 0-dimensional tensor or per-tensor quantized tensor | (C1) |
(I3) | edge_padding_low | 1-dimensional tensor constant of type si64 | (C1), (C4) |
(I4) | edge_padding_high | 1-dimensional tensor constant of type si64 | (C1), (C4) |
(I5) | interior_padding | 1-dimensional tensor constant of type si64 | (C2-C4) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C3-C6) |
element_type(operand) = element_type(padding_value) = element_type(result)
.size(edge_padding_low) = size(edge_padding_high) = size(interior_padding) = rank(operand)
.0 <= interior_padding
.shape(result) = shape(operand) + edge_padding_low + max(shape(operand) - 1, 0) * interior_padding + edge_padding_high
.// %operand: [ // [1, 2, 3], // [4, 5, 6] // ] // %padding_value: 0 %result = "stablehlo.pad"(%operand, %padding_value) { edge_padding_low = array<i64: 0, 1>, edge_padding_high = array<i64: 2, 1>, interior_padding = array<i64: 1, 2> } : (tensor<2x3xi32>, tensor<i32>) -> tensor<5x9xi32> // %result: [ // [0, 1, 0, 0, 2, 0, 0, 3, 0], // [0, 0, 0, 0, 0, 0, 0, 0, 0], // [0, 4, 0, 0, 5, 0, 0, 6, 0], // [0, 0, 0, 0, 0, 0, 0, 0, 0], // [0, 0, 0, 0, 0, 0, 0, 0, 0] // ]
Produces partition_id
of the current process.
Name | Type |
---|---|
result | 0-dimensional tensor of type ui32 |
%result = "stablehlo.partition_id"() : () -> tensor<ui32>
Performs element-wise count of the number of bits set in the operand
tensor and produces a result
tensor.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of integer type | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of integer type | (C1) |
type(operand) = type(result)
.// %operand: [0, 1, 2, 127] %result = "stablehlo.popcnt"(%operand) : (tensor<4xi64>) -> tensor<4xi64> // %result: [0, 1, 1, 7]
Performs element-wise exponentiation of lhs
tensor by rhs
tensor and produces a result
tensor. Depending on the element type, does the following:
pow
from IEEE-754.dequantize_op_quantize(power, lhs, rhs, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor of integer, floating-point, or complex type or per-tensor quantized tensor | (C1) |
(I2) | rhs | tensor of integer, floating-point, or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of integer, floating-point, or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %lhs: [-2.0, -0.0, -36.0, 5.0, 3.0, 10000.0] // %rhs: [2.0, 2.0, 1.1, 2.0, -1.0, 10.0] %result = "stablehlo.power"(%lhs, %rhs) : (tensor<6xf64>, tensor<6xf64>) -> tensor<6xf64> // %result: [4.0, 0.0, -nan, 25.0, 0.333333343, inf]
Extracts the real part, element-wise, from the operand
and produces a result
tensor. More formally, for each element x
: real(x) = is_complex(x) ? real_part(x) : x
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type | (C1), (C2) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point type | (C1), (C2) |
shape(result) = shape(operand)
.element_type(result)
is defined as:complex_element_type(element_type(operand))
if is_complex(operand)
.element_type(operand)
otherwise.// %operand: [(1.0, 2.0), (3.0, 4.0)] %result = "stablehlo.real"(%operand) : (tensor<2xcomplex<f32>>) -> tensor<2xf32> // %result: [1.0, 3.0]
Receives data from a channel with channel_id
and produces results
.
If is_host_transfer
is true
, then the operation transfers data from the host. Otherwise, it transfers data from another device. What this means is implementation-defined. This flag duplicates the information provided in channel_type
, so in the future we are planning to only keep one of them (#666).
results
consist of payload values which come first and a token which comes last. In the future, we are planning to split the payload and the token into two separate outputs to improve clarity (#670).
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | token | token | (C4) |
(I2) | channel_id | constant of type si64 | |
(I3) | channel_type | enum of DEVICE_TO_DEVICE and HOST_TO_DEVICE | (C1) |
(I4) | is_host_transfer | constant of type i1 | (C1) |
Name | Type | Constraints |
---|---|---|
results | variadic number of tensors, quantized tensors or tokens | (C2-C4) |
channel_type
is defined as:HOST_TO_DEVICE
if is_host_transfer = true
,DEVICE_TO_DEVICE
otherwise.0 < size(results)
.is_empty(result[:-1])
or is_tensor(type(results[:-1]))
.is_token(type(results[-1]))
.%results0, %results1 = "stablehlo.recv"(%token) { channel_handle = #stablehlo.channel_handle<handle = 1, type = 3>, is_host_transfer = true } : (!stablehlo.token) -> (tensor<2x2xi64>, !stablehlo.token)
Applies a reduction function body
to inputs
and init_values
along the dimensions
and produces results
tensors.
The order of reductions is implementation-defined, which means that body
and init_values
must form a monoid to guarantee that the operation produces the same results for all inputs on all implementations. However, this condition doesn‘t hold for many popular reductions. E.g. floating-point addition for body
and zero for init_values
don’t actually form a monoid because floating-point addition is not associative.
More formally, results...[j0, ..., jR-1] = reduce(input_slices_converted)
where:
input_slices = inputs...[j0, ..., :, ..., jR-1]
, where :
are inserted at dimensions
.input_slices_converted = to_destination_type(input_slices..., type(func_inputs(body)[:len(func_inputs(body))//2])...)
.init_values_converted = to_destination_type(init_values..., type(func_inputs(body)[len(func_inputs(body))//2:])...)
.reduce(input_slices_converted) = exec(schedule)
for some binary tree schedule
where:exec(node) = body(exec(node.left), exec(node.right))
.exec(leaf) = leaf.value
.schedule
is an implementation-defined full binary tree whose in-order traversal consists of:input_slices_converted...[index]
values, for all index
in index_space(input_slices_converted)
in the ascending lexicographic order of index
.init_values_converted
at implementation-defined positions.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | inputs | variadic number of tensors or per-tensor quantized tensors | (C1-C4), (C6), (C7) |
(I2) | init_values | variadic number of 0-dimensional tensors or per-tensor quantized tensors | (C2), (C3) |
(I3) | dimensions | 1-dimensional tensor constant of type si64 | (C4), (C5), (C7) |
(I4) | body | function | (C6) |
Name | Type | Constraints |
---|---|---|
results | variadic number of tensors or per-tensor quantized tensors | (C3), (C7), (C8) |
same(shape(inputs...))
.element_type(inputs...) = element_type(init_values...)
.0 < size(inputs) = size(init_values) = size(results) = N
.0 <= dimensions < rank(inputs[0])
.is_unique(dimensions)
.body
has type (tensor<E0>, ..., tensor<EN-1>, tensor<E0>, ...,
tensor<EN-1>) -> (tensor<E0>, ..., tensor<EN-1>)
where is_promotable(element_type(inputs[i]), Ei)
.shape(results...) = shape(inputs...)
except that the dimension sizes of inputs...
corresponding to dimensions
are not included.element_type(results[i]) = Ei
for all i
in [0,N)
.// %input = [[0, 1, 2, 3, 4, 5]] // %init_value = 0 %result = "stablehlo.reduce"(%input, %init_value) ({ ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>): %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64> "stablehlo.return"(%0) : (tensor<i64>) -> () }) { dimensions = array<i64: 1> } : (tensor<1x6xi64>, tensor<i64>) -> tensor<1xi64> // %result = [15]
Performs element-wise conversion of operand
to another floating-point type that uses exponent_bits
and mantissa_bits
and back to the original floating-point type and produces an output
tensor.
More formally:
mantissa_bits
using roundToIntegralTiesToEven
semantics.mantissa_bits
are smaller than the number of mantissa bits of the original value, the mantissa bits are truncated to mantissa_bits
.exponent_bits
, the intermediate result overflows to infinity using the original sign or underflows to zero using the original sign.dequantize_op_quantize( lambda operand: reduce_precision(operand, exponent_bits, mantissa_bits), operand, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point type or per-tensor quantized tensor | (C1) |
(I2) | exponent_bits | constant of type si32 | (C2) |
(I3) | mantissa_bits | constant of type si32 | (C3) |
Name | Type | Constraints |
---|---|---|
output | tensor of floating-point type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(output)
.1 <= exponent_bits
.0 <= mantissa_bits
.// Logical values: +Inf, NaN, +Denormal, 0.0, 65519.0, 65520.0 // %operand: [0x7FF0000000000000, 0x7FFFFFFFFFFFFFFF, 0x0000000000000001, 0.0, 65519.0, 65520.0] %output = "stablehlo.reduce_precision"(%operand) { exponent_bits = 5 : i32, mantissa_bits = 10 : i32 } : (tensor<6xf64>) -> tensor<6xf64> // Logical values: +Inf, NaN, 0.0, 0.0, 65504.0, +Inf // %output: [0x7FF0000000000000, 0x7FFFFFFFFFFFFFFF, 0.0, 0.0, 65504.0, 0x7FF0000000000000]
Within each process group in the StableHLO process grid, performs reduction, using computations
, over the values of the operand
tensor from each process, splits the reduction result along scatter_dimension
into parts, and scatters the split parts between the processes to produce the result
.
The operation splits the StableHLO process grid into process_groups
which is defined as follows:
cross_replica(replica_groups)
if channel_id <= 0 and use_global_device_ids = false
.cross_replica_and_partition(replica_groups)
if channel_id > 0 and use_global_device_ids = false
.flattened_ids(replica_groups)
if channel_id > 0 and use_global_device_ids = true
.Afterwards, within each process_group
:
reduced_value = all_reduce(operand, replica_groups, channel_id, use_global_device_ids, computation)
.parts@sender = split(reduced_value@sender, dim(process_groups, 1), scatter_dimension)
.result@receiver = parts@sender[receiver_index]
for all sender
in process_group
, where receiver_index = process_group.index(receiver)
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or per-tensor quantized tensor | (C1), (C2), (C7), (C8) |
(I2) | scatter_dimension | constant of type si64 | (C1), (C2), (C8) |
(I3) | replica_groups | 2-dimensional tensor constant of type si64 | (C3-C5) |
(I4) | channel_id | constant of type si64 | (C6) |
(I5) | use_global_device_ids | constant of type i1 | (C6) |
(I6) | computation | function | (C7) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C8-C9) |
dim(operand, scatter_dimension) % dim(process_groups, 1) = 0
.0 <= scatter_dimension < rank(operand)
.is_unique(replica_groups)
.size(replica_groups)
is defined as:num_replicas
if cross_replica
is used.num_replicas
if cross_replica_and_partition
is used.num_processes
if flattened_ids
is used.0 <= replica_groups < size(replica_groups)
.use_global_device_ids = true
, then channel_id > 0
.computation
has type (tensor<E>, tensor<E>) -> (tensor<E>)
where is_promotable(element_type(operand), E)
.shape(result) = shape(operand)
except:dim(result, scatter_dimension) = dim(operand, scatter_dimension) / dim(process_groups, 1)
.element_type(result) = E
.// num_replicas: 2 // num_partitions: 1 // %operand@(0, 0): [[1, 2, 3, 4], // [5, 6, 7, 8]] // %operand@(1, 0): [[9, 10, 11, 12], // [13, 14, 15, 16]] %result = "stablehlo.reduce_scatter"(%operand) ({ ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>): %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64> "stablehlo.return"(%0) : (tensor<i64>) -> () }) { scatter_dimension = 1 : i64, replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>, channel_handle = #stablehlo.channel_handle<handle = 0, type = 0> } : (tensor<2x4xi64>) -> tensor<2x2xi64> // // %result@(0, 0): [[10, 12], // [18, 20]] // %result@(1, 0): [[14, 16], // [22, 24]]
Applies a reduction function body
to windows of inputs
and init_values
and produces results
.
The following diagram shows how elements in results...
are computed from inputs...
using a concrete example.
More formally, results...[result_index] = reduce(windows, init_values, axes(inputs...), body)
(see reduce) where:
padded_inputs = pad(inputs..., init_values..., padding[:, 0], padding[:, 1], base_dilations - 1)
.window_start = result_index * window_strides
.window_end = window_start + (window_dimensions - 1) * window_dilations + 1
.windows = slice(padded_inputs..., window_start, window_end, window_dilations)
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | inputs | variadic number of tensors or per-tensor quantized tensors | (C1-C4), (C6), (C8), (C10), (C12), (C13), (C15) |
(I2) | init_values | variadic number of 0-dimensional tensors or per-tensor quantized tensors | (C1), (C13) |
(I3) | window_dimensions | 1-dimensional tensor constant of type si64 | (C4), (C5), (C15) |
(I4) | window_strides | 1-dimensional tensor constant of type si64 | (C6), (C7), (C15) |
(I5) | base_dilations | 1-dimensional tensor constant of type si64 | (C8), (C9), (C15) |
(I6) | window_dilations | 1-dimensional tensor constant of type si64 | (C10), (C11), (C15) |
(I7) | padding | 2-dimensional tensor constant of type si64 | (C12), (C15) |
(I8) | body | function | (C13) |
Name | Type | Constraints |
---|---|---|
results | variadic number of tensors or per-tensor quantized tensors | (C1), (C14-C16) |
0 < size(inputs) = size(init_values) = size(results) = N
.same(shape(inputs...))
.element_type(inputs...) = element_type(init_values...)
.size(window_dimensions) = rank(inputs[0])
.0 < window_dimensions
.size(window_strides) = rank(inputs[0])
.0 < window_strides
.size(base_dilations) = rank(inputs[0])
.0 < base_dilations
.size(window_dilations) = rank(inputs[0])
.0 < window_dilations
.shape(padding) = [rank(inputs[0]), 2]
.body
has type (tensor<E0>, ..., tensor<EN-1>, tensor<E0>, ...,
tensor<EN-1>) -> (tensor<E0>, ..., tensor<EN-1>)
where is_promotable(element_type(inputs[i]), Ei)
.same(shape(results...))
.shape(results[0]) = num_windows
where:dilated_input_shape = shape(inputs[0]) = 0 ? 0 : (shape(inputs[0]) - 1) * base_dilations + 1
.padded_input_shape = padding[:, 0] + dilated_input_shape + padding[:, 1]
.dilated_window_shape = (window_dimensions - 1) * window_dilations + 1
.is_empty_window = padded_input_shape = 0 || dilated_window_shape > padded_input_shape
.num_windows = is_empty_window ? 0 : floor((padded_input_shape - dilated_window_shape) / window_strides) + 1
.element_type(results[i]) = Ei
for all i
in [0,N)
.// %input = [[1, 2], [3, 4], [5, 6]] // %init_value = 0 %result = "stablehlo.reduce_window"(%input, %init_value) ({ ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>): %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64> "stablehlo.return"(%0) : (tensor<i64>) -> () }) { window_dimensions = array<i64: 2, 1>, window_strides = array<i64: 4, 1>, base_dilations = array<i64: 2, 1>, window_dilations = array<i64: 3, 1>, padding = dense<[[2, 1], [0, 0]]> : tensor<2x2xi64> } : (tensor<3x2xi64>, tensor<i64>) -> tensor<2x2xi64> // %result = [[0, 0], [3, 4]]
Performs element-wise remainder of dividend lhs
and divisor rhs
tensors and produces a result
tensor.
More formally, the sign of the result is taken from the dividend, and the absolute value of the result is always less than the divisor's absolute value. The remainder is calculated as lhs - d * rhs
, where d
is given by:
stablehlo.divide(lhs, rhs)
.division(lhs, rhs)
from IEEE-754 with rounding attribute roundTowardZero
.dequantize_op_quantize(remainder, lhs, rhs, type(result))
.For floating-point element types, this operation is in contrast with the remainder
operation from IEEE-754 specification where d
is an integral value nearest to the exact value of lhs/rhs
with ties to even.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor of integer, floating-point or complex type or per-tensor quantized tensor | (C1) |
(I2) | rhs | tensor of integer, floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of integer, floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %lhs: [17, -17, 17, -17] // %rhs: [3, 3, -3, -3] %result = "stablehlo.remainder"(%lhs, %rhs) : (tensor<4xi64>, tensor<4xi64>) -> tensor<4xi64> // %result: [2, -2, 2, -2]
Produces replica_id
of the current process.
Name | Type |
---|---|
result | 0-dimensional tensor of type ui32 |
%result = "stablehlo.replica_id"() : () -> tensor<ui32>
Performs reshape of operand
tensor to a result
tensor. Conceptually, it amounts to keeping the same canonical representation but potentially changing the shape, e.g. from tensor<2x3xf32>
to tensor<3x2xf32>
or tensor<6xf32>
.
More formally, result[result_index] = operand[operand_index]
where result_index
and operand_index
have the same position in the lexicographic ordering of index_space(result)
and index_space(operand)
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or quantized tensor | (C1-C3) |
Name | Type | Constraints |
---|---|---|
result | tensor or quantized tensor | (C1-C3) |
element_type(result)
is given by:element_type(operand)
, if !is_per_axis_quantized(operand)
.element_type(operand)
except that quantization_dimension(operand)
and quantization_dimension(result)
may differ, otherwise.size(operand) = size(result)
.is_per_axis_quantized(operand)
:reduce(dims(operand, [0, 1, ..., quantization_dimension(operand) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) = reduce(dims(result, [0, 1, ..., quantization_dimension(result) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y)
.dim(operand, quantization_dimension(operand)) = dim(result, quantization_dimension(result))
.reduce(dims(operand, [quantization_dimension(operand) + 1, ..., rank(operand) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) = reduce(dims(result, [quantization_dimension(result) + 1, ..., rank(result) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y)
.// %operand: [[1, 2, 3], [4, 5, 6]] %result = "stablehlo.reshape"(%operand) : (tensor<2x3xi32>) -> tensor<3x2xi32> // %result: [[1, 2], [3, 4], [5, 6]]
Reverses the order of elements in the operand
along the specified dimensions
and produces a result
tensor. More formally, result[result_index] = operand[operand_index]
where:
operand_index[d] = dim(result, d) - result_index[d] - 1
if d
in dimensions
.operand_index[d] = result_index[d]
otherwise.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or per-tensor quantized tensor | (C1), (C3) |
(I2) | dimensions | 1-dimensional tensor constant of type si64 | (C2), (C3) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C1), (C3) |
type(operand) = type(result)
.is_unique(dimensions)
.0 <= dimensions < rank(result)
.// %operand = [[1, 2], [3, 4], [5, 6]] %result = "stablehlo.reverse"(%operand) { dimensions = array<i64: 1> } : (tensor<3x2xi32>) -> tensor<3x2xi32> // %result: [[2, 1], [4, 3], [6, 5]]
Generates random numbers using the rng_distribution
algorithm and produces a result
tensor of a given shape shape
.
If rng_distribution = UNIFORM
, then the random numbers are generated following the uniform distribution over the interval [a, b)
. If a >= b
, the behavior is undefined.
If rng_distribution = NORMAL
, then the random numbers are generated following the normal distribution with mean = a
and standard deviation = b
. If b < 0
, the behavior is undefined.
The exact way how random numbers are generated is implementation-defined. For example, they may or may not be deterministic, and they may or may not use hidden state.
In conversations with many stakeholders, this op has come up as effectively deprecated, so in the future we are planning to explore removing it (#597).
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | a | 0-dimensional tensor of integer, boolean, or floating-point type | (C1), (C2) |
(I2) | b | 0-dimensional tensor of integer, boolean, or floating-point type | (C1), (C2) |
(I3) | shape | 1-dimensional tensor constant of type si64 | (C3) |
(I4) | rng_distribution | enum of UNIFORM and NORMAL | (C2) |
Name | Type | Constraints |
---|---|---|
result | tensor of integer, boolean, or floating-point type | (C1-C3) |
element_type(a) = element_type(b) = element_type(result)
.rng_distribution = NORMAL
, then is_float(a)
.shape(result) = shape
.// %a = 0 // %b = 2 // %shape = [3, 3] %result = "stablehlo.rng"(%a, %b, %shape) { rng_distribution = #stablehlo<rng_distribution UNIFORM> } : (tensor<i32>, tensor<i32>, tensor<2xi64>) -> tensor<3x3xi32> // %result: [ // [1, 0, 1], // [1, 1, 1], // [0, 0, 0] // ]
Returns an output
filled with uniform random bits and an updated output state output_state
using the pseudorandom number generator algorithm rng_algorithm
given an initial state initial_state
. The output is guaranteed to be deterministic function of initial_state
, but it is not guaranteed to be deterministic between implementations.
rng_algorithm
is one of the following:
DEFAULT
: Implementation-defined algorithm.THREE_FRY
: Implementation-defined variant of the Threefry algorithm.*PHILOX
: Implementation-defined variant of the Philox algorithm.** See: Salmon et al. SC 2011. Parallel random numbers: as easy as 1, 2, 3.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | rng_algorithm | enum of DEFAULT , THREE_FRY , and PHILOX | (C2) |
(I2) | initial_state | 1-dimensional tensor of type ui64 | (C1), (C2) |
Name | Type | Constraints |
---|---|---|
output_state | 1-dimensional tensor of type ui64 | (C1) |
output | tensor of integer or floating-point type |
type(initial_state) = type(output_state)
.size(initial_state)
is defined as:rng_algorithm = DEFAULT
.2
if rng_algorithm = THREE_FRY
.2
or 3
if rng_algorithm = PHILOX
.// %initial_state: [1, 2] %output_state, %output = "stablehlo.rng_bit_generator"(%initial_state) { rng_algorithm = #stablehlo<rng_algorithm THREE_FRY> } : (tensor<2xui64>) -> (tensor<2xui64>, tensor<2x2xui64>) // %output_state: [1, 6] // %output: [ // [9236835810183407956, 16087790271692313299], // [18212823393184779219, 2658481902456610144] // ]
Performs element-wise rounding towards the nearest integer, breaking ties away from zero, on the operand
tensor and produces a result
tensor. Implements the roundToIntegralTiesToAway
operation from the IEEE-754 specification. For quantized types, performs dequantize_op_quantize(round_nearest_afz, operand, type(result))
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand = [-2.5, 0.4, 0.5, 0.6, 2.5] %result = "stablehlo.round_nearest_afz"(%operand) : (tensor<5xf64>) -> tensor<5xf64> // %result: [-3.0, 0.0, 1.0, 1.0, 3.0]
Performs element-wise rounding towards the nearest integer, breaking ties towards the even integer, on the operand
tensor and produces a result
tensor. Implements the roundToIntegralTiesToEven
operation from the IEEE-754 specification. For quantized types, performs dequantize_op_quantize(round_nearest_even, operand, type(result))
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand = [-2.5, 0.4, 0.5, 0.6, 2.5] %result = "stablehlo.round_nearest_even"(%operand) : (tensor<5xf64>) -> tensor<5xf64> // %result: [-2.0, 0.0, 0.0, 1.0, 2.0]
Performs element-wise reciprocal square root operation on operand
tensor and produces a result
tensor. Depending on the element type, does the following:
rSqrt
from IEEE-754.dequantize_op_quantize(rsqrt, operand, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand: [[1.0, 4.0], [9.0, 25.0]] %result = "stablehlo.rsqrt"(%operand) : (tensor<2x2xf32>) -> tensor<2x2xf32> // %result: [[1.0, 0.5], [0.33333343, 0.2]]
Produces results
tensors which are equal to inputs
tensors except that several slices specified by scatter_indices
are updated with the values updates
using update_computation
.
The following diagram shows how elements in updates...
map on elements in results...
using a concrete example. The diagram picks a few example updates...
indices and explains in detail which results...
indices they correspond to.
More formally, for all update_index
in index_space(updates[0])
:
update_scatter_dims = [d for d in axes(updates[0]) and d not in update_window_dims]
.update_scatter_index = update_index[update_scatter_dims...]
.start_index
is defined as:scatter_indices[si0, ..., :, ..., siN]
where si
are individual elements in update_scatter_index
and :
is inserted at the index_vector_dim
index, if index_vector_dim
< rank(scatter_indices)
.[scatter_indices[update_scatter_index]]
otherwise.d_input
in axes(inputs[0])
,full_start_index[d_input] = start_index[d_start]
if d_input = scatter_dims_to_operand_dims[d_start]
.full_start_index[d_input] = 0
otherwise.update_window_index = update_index[update_window_dims...]
.full_window_index = [wi0, ..., 0, ..., wiN]
where wi
are individual elements in update_window_index
, and 0
is inserted at indices from inserted_window_dims
.result_index = full_start_index + full_window_index
.Given that, results = exec(schedule, inputs)
, where:
schedule
is an implementation-defined permutation of index_space(updates[0])
.exec([update_index, ...], results) = exec([...], updated_results)
where:result_index
is in bounds for shape(results...)
updates_converted = to_destination_type( updates...[update_index], type(func_inputs(update_computation) [len(func_inputs(update_computation))//2:])... )
updated_values = update_computation(results...[result_index], updates_converted)
updated_results
is a copy of results
with results...[result_index]
set to updated_values...
.updated_results = results
.exec([], results) = results
.If indices_are_sorted
is true
then the implementation can assume that scatter_indices
are sorted with respect to scatter_dims_to_operand_dims
, otherwise the behavior is undefined. More formally, for all i1 < i2
from indices(result)
, full_start_index(i1)
<= full_start_index(i2)
.
If unique_indices
is true
then the implementation can assume that all result_index
indices being scattered to are unique. If unique_indices
is true
but the indices being scattered to are not unique then the behavior is undefined.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | inputs | variadic number of tensors or per-tensor quantized tensors | (C1), (C2), (C4-C6), (C10), (C13), (C15-C16) |
(I2) | scatter_indices | tensor of integer type | (C4), (C11), (C14) |
(I3) | updates | variadic number of tensors or per-tensor quantized tensors | (C3-C6), (C8) |
(I4) | update_window_dims | 1-dimensional tensor constant of type si64 | (C2), (C4), (C7), (C8) |
(I5) | inserted_window_dims | 1-dimensional tensor constant of type si64 | (C2), (C4), (C9), (C10) |
(I6) | scatter_dims_to_operand_dims | 1-dimensional tensor constant of type si64 | (C11-C13) |
(I7) | index_vector_dim | constant of type si64 | (C4), (C11), (C14) |
(I8) | indices_are_sorted | constant of type i1 | |
(I9) | unique_indices | constant of type i1 | |
(I10) | update_computation | function | (C15) |
Name | Type | Constraints |
---|---|---|
results | variadic number of tensors or per-tensor quantized tensors | (C15-C17) |
same(shape(inputs...))
.rank(inputs[0]) = size(update_window_dims) + size(inserted_window_dims)
.same(shape(updates...))
.shape(updates[0]) = combine(update_scatter_dim_sizes, update_window_dim_sizes)
where:update_scatter_dim_sizes = shape(scatter_indices)
except that the dimension size of scatter_indices
corresponding to index_vector_dim
is not included.update_window_dim_sizes <= shape(inputs[0])
except that the dimension sizes in inputs[0]
corresponding to inserted_window_dims
are not included.combine
puts update_scatter_dim_sizes
at axes corresponding to update_scatter_dims
and update_window_dim_sizes
at axes corresponding to update_window_dims
.0 < size(inputs) = size(updates) = N
.element_type(updates...) = element_type(inputs...)
.is_unique(update_window_dims) and is_sorted(update_window_dims)
.0 <= update_window_dims < rank(updates[0])
.is_unique(inserted_window_dims) and is_sorted(update_window_dims)
.0 <= inserted_window_dims < rank(inputs[0])
.size(scatter_dims_to_operand_dims) = index_vector_dim < rank(scatter_indices) ? dim(scatter_indices, index_vector_dim) : 1
.is_unique(scatter_dims_to_operand_dims)
.0 <= scatter_dims_to_operand_dims < rank(inputs[0])
.0 <= index_vector_dim <= rank(scatter_indices)
.update_computation
has type (tensor<E0>, ..., tensor<EN-1>, tensor<E0>, ..., tensor<EN-1>) -> (tensor<E0>, ..., tensor<EN-1>)
, where is_promotable(element_type(inputs[i]), Ei)
.shape(inputs...) = shape(results...)
.element_type(results[i]) = Ei
for all i
in [0,N)
.// %input: [ // [[1, 2], [3, 4], [5, 6], [7, 8]], // [[9, 10], [11, 12], [13, 14], [15, 16]], // [[17, 18], [19, 20], [21, 22], [23, 24]] // ] // %scatter_indices: [[[0, 2], [1, 0], [2, 1]], [[0, 1], [1, 0], [0, 9]]] // %update: [ // [[[1, 1], [1, 1]], [[1, 1], [1, 1]], [[1, 1], [1, 1]]], // [[[1, 1], [1, 1]], [[1, 1], [1, 1]], [[1, 1], [1, 1]]] // ] %result = "stablehlo.scatter"(%input, %scatter_indices, %update) ({ ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>): %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64> "stablehlo.return"(%0) : (tensor<i64>) -> () }) { scatter_dimension_numbers = #stablehlo.scatter< update_window_dims = [2, 3], inserted_window_dims = [0], scatter_dims_to_operand_dims = [1, 0], index_vector_dim = 2>, indices_are_sorted = false, unique_indices = false } : (tensor<3x4x2xi64>, tensor<2x3x2xi64>, tensor<2x3x2x2xi64>) -> tensor<3x4x2xi64> // %result: [ // [[1, 2], [5, 6], [7, 8], [7, 8]], // [[10, 11], [12, 13], [14, 15], [16, 17]], // [[18, 19], [20, 21], [21, 22], [23, 24]] // ]
Produces a result
tensor where each element is selected from on_true
or on_false
tensor based on the value of the corresponding element of pred
. More formally, result[result_index] = pred_element ? on_true[result_index] : on_false[result_index]
, where pred_element = rank(pred) = 0 ? pred[] : pred[result_index]
. For quantized types, performs dequantize_select_quantize(pred, on_true, on_false, type(result))
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | pred | tensor of type i1 | (C1) |
(I2) | on_true | tensor or per-tensor quantized tensor | (C1-C2) |
(I3) | on_false | tensor or per-tensor quantized tensor | (C2) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C2) |
rank(pred) = 0 or shape(pred) = shape(on_true)
.baseline_type(on_true) = baseline_type(on_false) = baseline_type(result)
.// %pred: [[false, true], [true, false]] // %on_true: [[1, 2], [3, 4]] // %on_false: [[5, 6], [7, 8]] %result = "stablehlo.select"(%pred, %on_true, %on_false) : (tensor<2x2xi1>, tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32> // %result: [[5, 2], [3, 8]]
Scatters the values from the source
tensor using scatter
based on the outcome of reduce_window
of the input
tensor using select
and produces a result
tensor.
The following diagram shows how elements in result
are computed from operand
and source
using a concrete example.
More formally:
selected_values = reduce_window_without_init(...)
with the following inputs:
inputs = [operand].
window_dimensions
, window_strides
, and padding
which are used as is.base_dilations = windows_dilations = 1
.body
is defined as:def body(arg0: tensor<E>, arg1: tensor<E>) -> tensor<E>: return select(arg0, arg1) ? arg0 : arg1;
where E = element_type(operand)
, and reduce_window_without_init
works exactly like reduce_window
, except that the schedule
of the underlying reduce
(see reduce) doesn‘t include init values. It is currently unspecified what happens if the corresponding window doesn’t have values (#731).
result[result_index] = reduce([source_values], [init_value], [0], scatter)
where:
source_values = [source[source_index] for source_index in source_indices]
.selected_index(source_index) = operand_index
if selected_values[source_index]
has the operand
element from operand_index
.source_indices = [source_index for source_index in indices(source) if selected_index(source_index) = result_index]
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or per-tensor quantized tensor | (C1-C4), (C6), (C8-C11) |
(I2) | source | tensor or per-tensor quantized tensor | (C1), (C2) |
(I3) | init_value | 0-dimensional tensor or per-tensor quantized tensor | (C3) |
(I4) | window_dimensions | 1-dimensional tensor constant of type si64 | (C2), (C4), (C5) |
(I5) | window_strides | 1-dimensional tensor constant of type si64 | (C2), (C6), (C7) |
(I6) | padding | 2-dimensional tensor constant of type si64 | (C2), (C8) |
(I7) | select | function | (C9) |
(I8) | scatter | function | (C10) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C11-C12) |
element_type(operand) = element_type(source)
.shape(source) = num_windows
where:padded_operand_shape = padding[:, 0] + shape(operand) + padding[:, 1]
.is_empty_window = padded_operand_shape = 0 || window_dimensions > padded_operand_shape
.num_windows = is_empty_window ? 0 : floor((padded_operand_shape - window_dimensions) / window_strides) + 1
.element_type(init_value) = element_type(operand)
.size(window_dimensions) = rank(operand)
.0 < window_dimensions
.size(window_strides) = rank(operand)
.0 < window_strides
.shape(padding) = [rank(operand), 2]
.select
has type (tensor<E>, tensor<E>) -> tensor<i1>
where E = element_type(operand)
.scatter
has type (tensor<E>, tensor<E>) -> tensor<E>
where is_promotable(element_type(operand), E)
.shape(operand) = shape(result)
.element_type(result) = E
.// %operand: [[1, 5], [2, 5], [3, 6], [4, 4]] // %source: [[5, 6], [7, 8]] // %init_value: 0 %result = "stablehlo.select_and_scatter"(%operand, %source, %init_value) ({ ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>): %0 = "stablehlo.compare"(%arg0, %arg1) { comparison_direction = #stablehlo<comparison_direction GE> } : (tensor<i64>, tensor<i64>) -> tensor<i1> "stablehlo.return"(%0) : (tensor<i1>) -> () }, { ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>): %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64> "stablehlo.return"(%0) : (tensor<i64>) -> () }) { window_dimensions = array<i64: 3, 1>, window_strides = array<i64: 2, 1>, padding = dense<[[0, 1], [0, 0]]> : tensor<2x2xi64> } : (tensor<4x2xi64>, tensor<2x2xi64>, tensor<i64>) -> tensor<4x2xi64> // %result: [[0, 0], [0, 0], [5, 14], [7, 0]]
Sends inputs
to a channel channel_id
and produces a result
token.
If is_host_transfer
is true
, then the operation transfers data to the host. Otherwise, it transfers data to another device. What this means is implementation-defined. This flag duplicates the information provided in channel_type
, so in the future we are planning to only keep one of them (#666).
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | inputs | variadic number of tensors or quantized tensors | |
(I2) | token | token | |
(I3) | channel_id | constant of type si64 | |
(I4) | channel_type | enum of DEVICE_TO_DEVICE and DEVICE_TO_HOST | (C1) |
(I5) | is_host_transfer | constant of type i1 | (C1) |
Name | Type |
---|---|
result | token |
channel_type
is defined as:DEVICE_TO_HOST
if is_host_transfer = true
,DEVICE_TO_DEVICE
otherwise.%result = "stablehlo.send"(%operand, %token) { channel_handle = #stablehlo.channel_handle<handle = 1, type = 2>, is_host_transfer = true } : (tensor<2x2xi64>, !stablehlo.token) -> !stablehlo.token
Performs element-wise left-shift operation on the lhs
tensor by rhs
number of bits and produces a result
tensor.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor of integer type | (C1) |
(I2) | rhs | tensor of integer type | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of integer type | (C1) |
type(lhs) = type(rhs) = type(result)
.// %lhs: [-1, 0, 1] // %rhs: [1, 2, 3] %result = "stablehlo.shift_left"(%lhs, %rhs): (tensor<3xi64>, tensor<3xi64>) -> tensor<3xi64> // %result: [-2, 0, 8]
Performs element-wise arithmetic right-shift operation on the lhs
tensor by rhs
number of bits and produces a result
tensor.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor of integer type | (C1) |
(I2) | rhs | tensor of integer type | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of integer type | (C1) |
type(lhs) = type(rhs) = type(result)
.// %lhs: [-1, 0, 8] // %rhs: [1, 2, 3] %result = "stablehlo.shift_right_arithmetic"(%lhs, %rhs): (tensor<3xi64>, tensor<3xi64>) -> tensor<3xi64> // %result: [-1, 0, 1]
Performs element-wise logical right-shift operation on the lhs
tensor by rhs
number of bits and produces a result
tensor.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor of integer type | (C1) |
(I2) | rhs | tensor of integer type | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of integer type | (C1) |
type(lhs) = type(rhs) = type(result)
.// %lhs: [-1, 0, 8] // %rhs: [1, 2, 3] %result = "stablehlo.shift_right_logical"(%lhs, %rhs): (tensor<3xi64>, tensor<3xi64>) -> tensor<3xi64> // %result: [9223372036854775807, 0, 1]
Returns the sign of the operand
element-wise and produces a result
tensor. More formally, for each element x
, the semantics can be expressed using Python syntax as follows:
def sign(x): if is_integer(x): if compare(x, 0, LT, SIGNED): return -1 if compare(x, 0, EQ, SIGNED): return 0 return 1 elif is_float(x): if is_nan(x): return NaN if compare(x, -0.0, EQ, FLOAT): return -0.0 if compare(x, +0.0, EQ, FLOAT): return +0.0 if compare(x, 0.0, LT, FLOAT): return -1.0 return 1.0 elif is_complex(x): if is_nan(real(x)) or is_nan(imag(x)): return (NaN, NaN) if compare(x, (0.0, 0.0), EQ, FLOAT): return (0.0, 0.0) return divide(x, convert(abs(x), type(x)))
For quantized types, performs dequantize_op_quantize(sign, operand, type(result))
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of signed integer, floating-point, or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of signed integer, floating-point, or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// Logical values: +NaN, -1.0, -0.0, +0.0, 1.0 // operand: [0x7FFFFFFFFFFFFFFF, -1.0, -0.0, 0.0, 1.0] %result = "stablehlo.sign"(%operand) : (tensor<5xf64>) -> tensor<5xf64> // Logical values: +NaN, -1.0, -0.0, +0.0, 1.0 // %result: [0x7FFFFFFFFFFFFFFF, -1.0, -0.0, 0.0, 1.0]
Performs element-wise sine operation on operand
tensor and produces a result
tensor. Depending on the element type, does the following:
sin
from IEEE-754.dequantize_op_quantize(sine, operand, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand: [ // [0.0, 1.57079632], // [0, pi/2] // [3.14159265, 4.71238898] // [pi, 3pi/2] // ] %result = "stablehlo.sine"(%operand) : (tensor<2x2xf32>) -> tensor<2x2xf32> // %result: [[0.0, 1.0], [0.0, -1.0]]
Extracts a slice from the operand
using statically-computed starting indices and produces a result
tensor. start_indices
contain the starting indices of the slice for each dimension, limit_indices
contain the ending indices (exclusive) for the slice for each dimension, and strides
contain the strides for each dimension.
More formally, result[result_index] = operand[operand_index]
where operand_index = start_indices + result_index * strides
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or per-tensor quantized tensor | (C1-C3), (C5) |
(I2) | start_indices | 1-dimensional tensor constant of type si64 | (C2), (C3), (C5) |
(I3) | limit_indices | 1-dimensional tensor constant of type si64 | (C2), (C3), (C5) |
(I4) | strides | 1-dimensional tensor constant of type si64 | (C2), (C4) |
Name | Type | Constraints |
---|---|---|
result | tensor or per-tensor quantized tensor | (C1), (C5) |
element_type(operand) = element_type(result)
.size(start_indices) = size(limit_indices) = size(strides) = rank(operand)
.0 <= start_indices <= limit_indices <= shape(operand)
.0 < strides
.shape(result) = ceil((limit_indices - start_indices) / strides)
.// %operand: [ // [0, 0, 0, 0], // [0, 0, 1, 1], // [0, 0, 1, 1] // ] %result = "stablehlo.slice"(%operand) { start_indices = array<i64: 1, 2>, limit_indices = array<i64: 3, 4>, strides = array<i64: 1, 1> } : (tensor<3x4xi64>) -> tensor<2x2xi64> // % result: [ // [1, 1], // [1, 1] // ]
Sorts 1-dimensional slices of inputs
along the dimension dimension
together, according to a comparator
and produces results
.
Unlike similar inputs in other operations, dimension
allows negative values, with the semantics described below. In the future, this may be disallowed for consistency reasons (#1377).
If is_stable
is true, then the sorting is stable, that is, relative order of elements considered to be equal by the comparator is preserved. For the case where there is a single input, two elements e1
and e2
are considered to be equal by the comparator if and only if comparator(e1, e2) = comparator(e2, e1) = false
. See the formalization below for how this generalizes to multiple inputs.
More formally, for all result_index
in index_space(results[0])
:
adjusted_dimension = dimension >= 0 ? dimension : rank(inputs[0]) + dimension
.
result_slice = [ri0, ..., :, ..., riR-1]
where riN
are individual elements in result_index
, and :
is inserted at adjusted_dimension
.
inputs_together = (inputs[0]..., ..., inputs[N-1]...)
.
results_together[result_slice] = sort(inputs_together[result_slice], comparator_together)
.
where sort
sorts a 1-dimensional slice in non-descending order expecting that comparator_together
returns true
if the left-hand side argument is less than the right-hand second argument.
def comparator_together(lhs_together, rhs_together): args = [] for (lhs_el, rhs_el) in zip(lhs_together, rhs_together): args.append(lhs_el) args.append(rhs_el) return comparator(*args)
(results[0]..., ..., results[N-1]...) = results_together
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | inputs | variadic number of tensors or per-tensor quantized tensors | (C1-C5) |
(I2) | dimension | constant of type si64 | (C4) |
(I3) | is_stable | constant of type i1 | |
(I4) | comparator | function | (C5) |
Name | Type | Constraints |
---|---|---|
results | variadic number of tensors or per-tensor quantized tensors | (C2), (C3) |
0 < size(inputs)
.type(inputs...) = type(results...)
.same(shape(inputs...) + shape(results...))
.-R <= dimension < R
, where R = rank(inputs[0])
.comparator
has type (tensor<E1>, tensor<E1>, ..., tensor<EN-1>, tensor<EN-1>) -> tensor<i1>
, where Ei = element_type(inputs[i])
.// %input0 = [[1, 2, 3], [3, 2, 1]] // %input1 = [[3, 2, 1], [1, 2, 3]] %result0, %result1 = "stablehlo.sort"(%input0, %input1) ({ ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>, %arg2: tensor<i64>, %arg3: tensor<i64>): %predicate = "stablehlo.compare"(%arg0, %arg1) { comparison_direction = #stablehlo<comparison_direction GT> } : (tensor<i64>, tensor<i64>) -> tensor<i1> "stablehlo.return"(%predicate) : (tensor<i1>) -> () }) { dimension = 0 : i64, is_stable = true } : (tensor<2x3xi64>, tensor<2x3xi64>) -> (tensor<2x3xi64>, tensor<2x3xi64>) // %result0 = [[3, 2, 3], [1, 2, 1]] // %result1 = [[1, 2, 1], [3, 2, 3]]
Performs element-wise square root operation on operand
tensor and produces a result
tensor. Depending on the element type, does the following:
squareRoot
from IEEE-754.dequantize_op_quantize(sqrt, operand, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand: [[0.0, 1.0], [4.0, 9.0]] %result = "stablehlo.sqrt"(%operand) : (tensor<2x2xf32>) -> tensor<2x2xf32> // %result: [[0.0, 1.0], [2.0, 3.0]]
Performs element-wise subtraction of two tensors lhs
and rhs
and produces a result
tensor. Depending on the element type, does the following:
subtraction
from IEEE-754.dequantize_op_quantize(subtract, lhs, rhs, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor of integer, floating-point, or complex type or per-tensor quantized tensor | (C1) |
(I2) | rhs | tensor of integer, floating-point, or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of integer, floating-point, or complex type or per-tensor quantized tensor | (C1) |
baseline_type(lhs) = baseline_type(rhs) = baseline_type(result)
.// %lhs: [[6, 8], [10, 12]] // %rhs: [[5, 6], [7, 8]] %result = "stablehlo.subtract"(%lhs, %rhs) : (tensor<2x2xf32>, tensor<2x2xf32>) -> (tensor<2x2xf32>) // %result: [[1, 2], [3, 4]]
Performs element-wise hyperbolic tangent operation on operand
tensor and produces a result
tensor. Depending on the element type, does the following:
tanh
from IEEE-754.dequantize_op_quantize(tanh, operand, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_type(operand) = baseline_type(result)
.// %operand: [-1.0, 0.0, 1.0] %result = "stablehlo.tanh"(%operand) : (tensor<3xf32>) -> tensor<3xf32> // %result: [-0.76159416, 0.0, 0.76159416]
Permutes the dimensions of operand
tensor using permutation
and produces a result
tensor. More formally, result[result_index] = operand[operand_index]
where result_index[d] = operand_index[permutation[d]]
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor or quantized tensor | (C1-C4) |
(I2) | permutation | 1-dimensional tensor constant of type si64 | (C2-C4) |
Name | Type | Constraints |
---|---|---|
result | tensor or quantized tensor | (C1), (C3-C4) |
element_type(result)
is given by:element_type(operand)
, if !is_per_axis_quantized(operand)
.element_type(operand)
except that quantization_dimension(operand)
and quantization_dimension(result)
may differ, otherwise.permutation
is a permutation of range(rank(operand))
.shape(result) = dim(operand, permutation...)
.is_per_axis_quantized(result)
, then quantization_dimension(operand) = permutation(quantization_dimension(result))
.// %operand: [ // [[1,2], [3,4], [5,6]], // [[7,8], [9,10], [11,12]] // ] %result = "stablehlo.transpose"(%operand) { permutation = array<i64: 2, 1, 0> } : (tensor<2x3x2xi32>) -> tensor<2x3x2xi32> // %result: [ // [[1,7], [3,9], [5,11]], // [[2,8], [4,10], [6,12]] // ]
Solves batches of systems of linear equations with lower or upper triangular coefficient matrices.
More formally, given a
and b
, result[i0, ..., iR-3, :, :]
is the solution to op(a[i0, ..., iR-3, :, :]) * x = b[i0, ..., iR-3, :, :]
when left_side
is true
or x * op(a[i0, ..., iR-3, :, :]) = b[i0, ..., iR-3, :, :]
when left_side
is false
, solving for the variable x
where op(a)
is determined by transpose_a
, which can be one of the following:
NO_TRANSPOSE
: Perform operation using a
as-is.TRANSPOSE
: Perform operation on transpose of a
.ADJOINT
: Perform operation on conjugate transpose of a
.Input data is read only from the lower triangle of a
, if lower
is true
or upper triangle of a
, otherwise. Output data is returned in the same triangle; the values in the other triangle are implementation-defined.
If unit_diagonal
is true, then the implementation can assume that the diagonal elements of a
are equal to 1, otherwise the behavior is undefined.
For quantized types, performs dequantize_op_quantize(lambda x, y: triangular_solve(x, y, left_side, lower, unit_diagonal, transpose_a), a, b, type(result))
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | a | tensor of floating-point or complex type or per-tensor quantized tensor | (C1-C3) |
(I2) | b | tensor of floating-point or complex type or per-tensor quantized tensor | (C1-C4) |
(I3) | left_side | constant of type i1 | (C3) |
(I4) | lower | constant of type i1 | |
(I5) | unit_diagonal | constant of type i1 | |
(I6) | transpose_a | enum of NO_TRANSPOSE , TRANSPOSE , and ADJOINT |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point or complex type or per-tensor quantized tensor | (C1) |
baseline_element_type(a) = baseline_element_type(b)
.2 <= rank(a) = rank(b) = R
.shape(a)
and shape(b)
is defined as follows:shape(a)[:-3] = shape(b)[:-3]
.dim(a, -2) = dim(a, -1) = dim(b, left_side ? -2 : -1)
.baseline_type(b) = baseline_type(result)
.// %a = [ // [1.0, 0.0, 0.0], // [2.0, 4.0, 0.0], // [3.0, 5.0, 6.0] // ] // %b = [ // [2.0, 0.0, 0.0], // [4.0, 8.0, 0.0], // [6.0, 10.0, 12.0] // ] %result = "stablehlo.triangular_solve"(%a, %b) { left_side = true, lower = true, unit_diagonal = false, transpose_a = #stablehlo<transpose NO_TRANSPOSE> } : (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32> // %result: [ // [2.0, 0.0, 0.0], // [0.0, 2.0, 0.0], // [0.0, 0.0, 2.0] // ]
Produces a result
tuple from values val
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | val | variadic number of values | (C1) |
Name | Type | Constraints |
---|---|---|
result | tuple | (C1) |
result
has type tuple<E0, ..., EN-1>
where Ei = type(val[i])
.// %val0: [1.0, 2.0] // %val1: (3) %result = "stablehlo.tuple"(%val0, %val1) : (tensor<2xf32>, tuple<tensor<i32>>) -> tuple<tensor<2xf32>, tuple<tensor<i32>>> // %result: ([1.0, 2.0], (3))
Performs element-wise conversion of quantized tensor operand
to a floating-point tensor result
according to the quantization parameters defined by the operand
type.
More formally, result = dequantize(operand)
.
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | quantized tensor | (C1), (C2) |
Name | Type | Constraints |
---|---|---|
result | tensor of floating-point type | (C1), (C2) |
shape(operand) = shape(result)
.element_type(result) = expressed_type(operand)
.// %operand: [10, 10] %result = "stablehlo.uniform_dequantize"(%operand) : (tensor<2x!quant.uniform<i8:f32:0, {0.1:-30,0.5:-20}>>) -> tensor<2xf32> // %result: [4.0, 15.0]
Performs element-wise conversion of floating-point tensor or quantized tensor operand
to a quantized tensor result
according to the quantization parameters defined by the result
type.
More formally,
is_float(operand)
:result = quantize(operand, type(result))
.is_quantized(operand)
:float_result = dequantize(operand)
.result = quantize(float_result, type(result))
.Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | tensor of floating-point or quantized type | (C1), (C2) |
Name | Type | Constraints |
---|---|---|
result | quantized tensor | (C1), (C2) |
shape(operand) = shape(result)
.expressed_type(result) = is_float(operand) ? element_type(operand) : expressed_type(operand)
.// %operand: [4.0, 15.0] %result = "stablehlo.uniform_quantize"(%operand) : (tensor<2xf32>) -> tensor<2x!quant.uniform<i8:f32:0, {0.1:-30,0.5:-20}>> // %result: [10, 10] // %operand: [10, 10] %result = "stablehlo.uniform_quantize"(%operand) : (tensor<2x!quant.uniform<i8:f32:0, {0.1:-30,0.5:-20}>>) -> tensor<2x!quant.uniform<i8:f32:0, {0.1:-20,0.2:-30}>> // %result: [20, 45]
Produces the output from executing body
function 0 or more times while the cond
function outputs true
. More formally, the semantics can be expressed using Python syntax as follows:
internal_state = operand while cond(*internal_state): internal_state = body(*internal_state) results = internal_state
The behavior of an infinite loop is TBD (#383).
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | operand | variadic number of tensors, quantized tensors or tokens | (C1-C3) |
(I2) | cond | function | (C1) |
(I3) | body | function | (C2) |
Name | Type | Constraints |
---|---|---|
results | variadic number of tensors, quantized tensors or tokens | (C3) |
cond
has type (T0, ..., TN-1) -> tensor<i1>
, where Ti = type(operand[i])
.body
has type (T0, ..., TN-1) -> (T0, ..., TN-1)
, where Ti = type(operand[i])
.type(results...) = type(operand...)
.// %init_i: 1 // %init_sum: 0 // %one: 1 // %ten: 10 %results0, %results1 = "stablehlo.while"(%init_i, %init_sum) ({ ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>): %cond = "stablehlo.compare"(%arg0, %ten) { comparison_direction = #stablehlo<comparison_direction LT> } : (tensor<i64>, tensor<i64>) -> tensor<i1> stablehlo.return %cond : tensor<i1> }, { ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>): %new_sum = stablehlo.add %arg1, %one : tensor<i64> %new_i = stablehlo.add %arg0, %one : tensor<i64> stablehlo.return %new_i, %new_sum : tensor<i64>, tensor<i64> }) : (tensor<i64>, tensor<i64>) -> (tensor<i64>, tensor<i64>) // %results0: 10 // %results1: 10
Performs element-wise XOR of two tensors lhs
and rhs
and produces a result
tensor. Depending on the element type, does the following:
Label | Name | Type | Constraints |
---|---|---|---|
(I1) | lhs | tensor of boolean or integer type | (C1) |
(I2) | rhs | tensor of boolean or integer type | (C1) |
Name | Type | Constraints |
---|---|---|
result | tensor of boolean or integer type | (C1) |
type(lhs) = type(rhs) = type(result)
.// Bitwise operation with with integer tensors // %lhs: [[1, 2], [3, 4]] // %rhs: [[5, 6], [7, 8]] %result = "stablehlo.xor"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32> // %result: [[4, 4], [4, 12]] // Logical operation with with boolean tensors // %lhs: [[false, false], [true, true]] // %rhs: [[false, true], [false, true]] %result = "stablehlo.xor"(%lhs, %rhs) : (tensor<2x2xi1>, tensor<2x2xi1>) -> tensor<2x2xi1> // %result: [[false, true], [true, false]]
A StableHLO program is executed by providing input values to the main
function and computing output values. Output values of a function are computed by executing the graph of ops rooted in the corresponding return
op.
The execution order is implementation-defined as long as it is aligned with dataflow, i.e. if ops are executed before their uses. In StableHLO, all side-effecting ops consume one token and produce one token (multiple tokens can be multiplexed into one token via after_all
), so the execution order of side effects is also aligned with dataflow. For example, in the below program there are two possible execution orders: %0
→ %1
→ %2
→ return
and %1
→ %0
→ %2
→ return
.
func.func @main() -> tensor<f64> { %0 = stablehlo.constant dense<1.0> : tensor<f64> %1 = stablehlo.constant dense<2.0> : tensor<f64> %2 = stablehlo.add %0, %1 : tensor<f64> return %2 : tensor<f64> }
More formally, a StableHLO process is a combination of: 1) a StableHLO program, 2) operation statuses (not executed yet, already executed), and 3) intermediate values that the process is working on. The process starts with input values to the main
function, progresses through the graph of ops updating operation statuses and intermediate values and finishes with output values. Further formalization is TBD (#484).
StableHLO programs can be executed in parallel, organized into a 2D process grid of num_replicas
by num_partitions
which both have type ui32
.
In the StableHLO process grid, num_replicas * num_partitions
of StableHLO processes are executing at the same time. Each process has a unique process_id = (replica_id, partition_id)
, where replica_id
in replica_ids = range(num_replicas)
and partition_id
in partition_ids = range(num_partitions)
which both have type ui32
.
The size of the process grid is known statically for every program (in the future, we are planning to make it an explicit part of StableHLO programs #650), and the position within the process grid is known statically for every process. Each process has access to its position within the process grid via the replica_id
and partition_id
ops.
Within the process grid, the programs can all be the same (in the “Single Program, Multiple Data” style), can all be different (in the “Multiple Program, Multiple Data” style) or something in between. In the future, we are planning to introduce support for other idioms of defining parallel StableHLO programs, including GSPMD (#619).
Within the process grid, the processes are mostly independent from each other - they have separate operation statuses, separate input/intermediate/output values and most of the ops are executed separately between processes, with the exception of a small number of collective ops described below.
Given that execution of most of the ops is only using values from the same process, it is usually unambiguous to refer to these values by their names. However, when describing semantics of collective ops, that is insufficient, and that gives rise to the notation name@process_id
to refer to the value name
within a particular process. (From that perspective, unqualified name
can be viewed as a shorthand for name@(replica_id(), partition_id())
).
The execution order across processes is implementation-defined, except for the synchronization introduced by point-to-point communication and collective ops as described below.
StableHLO processes can communicate with each other through StableHLO channels. A channel is represented by a positive id of type si64
. Through various ops, it is possible to send values to channels and receive them from channels.
Further formalization, e.g. where these channel ids are coming from, how processes programs become aware of them and what kind of synchronization is introduced by them, is TBD (#484).
Every StableHLO process has access to two streaming interfaces:
Unlike channels, which are used to communicate between processes and therefore have processes at both of their ends, infeeds and outfeeds have their other end implementation-defined.
Further formalization, e.g. how streaming communication influences execution order and what kind of synchronization is introduced by it, is TBD (#484).
There are six collective ops in StableHLO: all_gather
, all_reduce
, all_to_all
, collective_broadcast
, collective_permute
, and reduce_scatter
. All these ops split the processes in the StableHLO process grid into StableHLO process groups and execute a joint computation within each process group, independently from other process groups.
Within each process group, collective ops may introduce a synchronization barrier. Further formalization, e.g. elaborating on when exactly this synchronization happens, how exactly the processes arrive at this barrier, and what happens if they don't, is TBD (#484).
If the process group involves cross-partition communication, i.e. there are processes in the process group whose partition ids are different, then execution of the collective op needs a channel, and the collective op must provide a positive channel_id
of type si64
. Cross-replica communication doesn't need channels.
The computations performed by the collective ops are specific to individual ops and are described in individual op sections above. However, the strategies by which the process grid is split into process groups are shared between these ops and are described in this section. More formally, StableHLO supports the following four strategies.
Only cross-replica communications happen within each process group. This strategy takes replica_groups
- a list of lists of replica ids - and computes a Cartesian product of replica_groups
by partition_ids
. replica_groups
must have unique elements and cover all replica_ids
. More formally, using Python syntax:
def cross_replica(replica_groups: List[List[ReplicaId]]) -> List[List[ProcessId]]: for replica_group in replica_groups: for partition_id in partition_ids: process_group = [] for replica_id in replica_group: process_group.append((replica_id, partition_id)) yield process_group
For example, for replica_groups = [[0, 1], [2, 3]]
and num_partitions = 2
, cross_replica
will produce [[(0, 0), (1, 0)], [(0, 1), (1, 1)], [(2, 0), (3, 0)], [(2, 1), (3, 1)]]
.
Only cross-partition communications happen within each process group. This strategy takes partition_groups
- a list of lists of partition ids - and computes a Cartesian product of partition_groups
by replica_ids
. partition_groups
must have unique elements and cover all partition_ids
. More formally, using Python syntax:
def cross_partition(partition_groups: List[List[PartitionId]]) -> List[List[ProcessId]]: for partition_group in partition_groups: for replica_id in replica_ids: process_group = [] for partition_id in partition_group: process_group.append((replica_id, partition_id)) yield process_group
For example, for partition_groups = [[0, 1]]
and num_replicas = 4
, cross_partition
will produce [[(0, 0), (0, 1)], [(1, 0), (1, 1)], [(2, 0), (2, 1)], [(3, 0), (3, 1)]]
.
Both cross-replica and cross-partition communications may happen within each process group. This strategy takes replica_groups
- a list of lists of replica ids - and computes Cartesian products of each replica_group
by partition_ids
. replica_groups
must have unique elements and cover all replica_ids
. More formally, using Python syntax:
def cross_replica_and_partition(replica_groups: List[List[ReplicaId]]) -> List[List[ProcessId]]: for replica_group in replica_groups: process_group = [] for partition_id in partition_ids: for replica_id in replica_group: process_group.append((replica_id, partition_id)) yield process_group
For example, for replica_groups = [[0, 1], [2, 3]]
and num_partitions = 2
, cross_replica_and_partition
will produce [[(0, 0), (1, 0), (0, 1), (1, 1)], [(2, 0), (3, 0), (2, 1), (3, 1)]]
.
This strategy takes flattened_id_groups
- a list of lists of “flattened” process ids in the form of replica_id * num_partitions + partition_id
- and turns them into process ids. flattened_id_groups
must have unique elements and cover all process_ids
. More formally, using Python syntax:
def flattened_ids(flattened_id_groups: List[List[ui32]]) -> List[List[ProcessId]]: for flattened_id_group in flattened_id_groups: process_group = [] for flattened_id in flattened_id_group: replica_id = flattened_id // num_partitions partition_id = flattened_id % num_partitions process_group.append((replica_id, partition_id)) yield process_group
For example, for flattened_id_groups = [[0, 1, 2, 3], [4, 5, 6, 7]]
, num_replicas = 4
and num_partitions = 2
, flattened_ids
will produce [[(0, 0), (0, 1), (1, 0), (1, 1)], [(2, 0), (2, 1), (3, 0), (3, 1)]]
.
At the moment, StableHLO does not provide guarantees about numerical accuracy, but this may change in the future (#1156).
The interpretation of quantized StableHLO operations may vary depending on the hardware requirements and capabilities. For instance, some hardware may opt to interpret quantized operations using a “dequantize, perform floating-point operation, and finally quantize” strategy. Others may perform the entire computation with integer arithmetic. Consequently, the interpretation of quantized StableHLO operations is exclusively determined by the specific implementation. The interpretation of hybrid quantization (#1575) should be based on the it's semantics as prescribed in the specification (via 1792).
StableHLO programs are validated through an extensive set of constraints for individual ops, which rules out many classes of errors prior to run time. However, error conditions are still possible, e.g. through integer overflows, out-of-bounds accesses, etc. Unless explicitly called out, all these errors result in implementation-defined behavior, but this may change in the future (#1157).
As an exception to this rule, floating-point exceptions in StableHLO programs have well-defined behavior. Operations which result in exceptions defined by the IEEE-754 standard (invalid operation, division-by-zero, overflow, underflow, or inexact exceptions) produce default results (as defined in the standard) and continue execution without raising the corresponding status flag; similar to raiseNoFlag
exception handling from the standard. Exceptions for nonstandard operations (e.g. complex arithmetic and certain transcendental functions) are implementation-defined.
StableHLO supports dynamically-shaped tensors. However, shapes have to agree at runtime, otherwise the behavior is undefined. StableHLO does not explicitly provide an op that can assert that a tensor has a given shape at runtime. Generating correct code is the responsibility of the producer.
As a specific example, the below program is valid. However, at runtime, the exact shapes of %arg0
and %arg1
will have to be the same, otherwise the behavior of the program is undefined:
func.func @foo(%arg0: tensor<?xi32>, %arg1: tensor<?xi32>) -> tensor<?xi32> { %0 = stablehlo.add %arg0, %arg1 : tensor<?xi32> return %0 : tensor<?xi32> }
For describing syntax, this document is using the modified ISO flavor of EBNF syntax (ISO/IEC 14977:1996, Wikipedia), with two modifications: 1) rules are defined using ::=
rather than =
, 2) concatenation is expressed using juxtaposition rather than ,
.
For describing semantics (i.e. within “Types”, “Constants” and “Ops” sections), we are using formulas which are based on Python syntax extended with support for concisely expressing array operations as described below. This works well for small snippets of code, but in rare cases when larger snippets of code are needed, we use vanilla Python syntax which is always introduced explicitly.
Let's explore how formulas work based on an example from the dot_general
specification. One of the constraints for this operation looks as follows: dim(lhs, lhs_batching_dimensions...) = dim(rhs, rhs_batching_dimensions...)
.
The names used in this formula come from two sources: 1) global functions, i.e. dim
, 2) member definitions of the corresponding program element, i.e. lhs
, lhs_batching_dimensions
, rhs
and rhs_batching_dimensions
inputs defined in the “Inputs” section of dot_general
.
As mentioned above, the syntax of this formula is Python-based with some conciseness-oriented extensions. To make sense of the formula, let's transform it into vanilla Python syntax.
A) In these formulas, we are using =
to represent equality, so the first step towards obtaining Python syntax is replacing =
with ==
, as follows: dim(lhs, lhs_batching_dimensions...) == dim(rhs, rhs_batching_dimensions...)
.
B) Also, these formulas support ellipses (...
) which turn scalar expressions into tensor expressions. In a nutshell, f(xs...)
roughly means “for each scalar x
in the tensor xs
, compute a scalar f(x)
and then return all these scalar results together as a tensor result”. In vanilla Python syntax, our example formula turns into: [dim(lhs, dim1) for dim1 in lhs_batching_dimensions] == [dim(rhs, dim2) for dim2 in rhs_batching_dimensions]
.
Thanks to ellipses, it is often possible to avoid working at the level of individual scalars. However, in some tricky cases, lower-level semi-informal syntax may be used like in the start_indices[bi0, ..., :, ..., biN]
formula from the gather
specification. In the service of conciseness, we don‘t provide an exact formalism for translating such syntax to vanilla Python, in hopes that it is still intuitively understandable on case-by-case basis. Please let us know if some specific formulas look opaque, and we’ll try to improve them.
Also, you will notice that formulas use ellipses to expand all sorts of lists, including tensors, lists of tensors (which e.g. can arise from a variadic number of tensors), etc. This is another area where we don't provide an exact formalism (e.g. lists are not even part of the StableHLO type system) and instead rely on intuitive understandability.
C) The final noteworthy notational vehicle that we employ is implicit broadcasting. While the StableHLO opset doesn't support implicit broadcasting, the formulas do, also in the service of conciseness. In a nutshell, if a scalar is used in a context where a tensor is expected, the scalar is broadcasted to the expected shape.
To continue the dot_general
example, here's another constraint: 0 <= lhs_batching_dimensions < rank(lhs)
. As defined in the dot_general
specification, lhs_batching_dimensions
is a tensor, however both 0
and rank(lhs)
are scalars. After we apply implicit broadcasting, the formula will become [0, ..., 0] <= lhs_batching_dimensions < [rank(lhs), ..., rank(lhs)]
.
When applied to a particular dot_general
operation, this formula will evaluate to a tensor of booleans. When formulas are used as constraints, the constraint holds if the formula evaluates to either true
or to a tensor which only has true
elements.
In formulas, lexical scope includes: 1) global functions, 2) member definitions, 3) local definitions. The list of global functions is provided below. The list of element definitions depends on the program element that the notation is applied to:
IntegerLiteral
=> integer_literal
), but sometimes names get abbreviated in the process (e.g. QuantizationStorageType
=> storage_type
) in which case the names are introduced explicitly similarly to “Inputs” / “Outputs” sections in operation specifications.self
to refer to the corresponding program element.When formulas are evaluated, they work with the following types of values: 1) Value
(actual values, e.g. dense<[[1, 2], [3, 4]]> : tensor<2x2xi32>
; they always know their types), 2) Placeholder
(future values, e.g. lhs
, rhs
or result
; their actual values are not known yet, only their types are known), 3) Type
(types as defined in the “Types” section), 4) Function
(global functions as defined in the “Functions” section).
Depending on the context, names may be referring to different values. More specifically, the “Semantics” section for ops (and equivalents for other program elements) defines runtime logic, so all inputs are available as Value
. In contrast, the “Constraints” section for ops (and equivalents) defines “compile-time” logic, i.e. something that is typically executed before runtime, so only constant inputs are available as Value
and other inputs are available only as Placeholder
.
Names | In “Semantics” | In “Constraints” |
---|---|---|
Global functions | Function | Function |
Constant inputs | Value | Value |
Non-constant inputs | Value | Placeholder |
Outputs | Value | Placeholder |
Local definitions | Depends on the definition | Depends on the definition |
Let's consider an example transpose
operation:
%result = "stablehlo.transpose"(%operand) { permutation = dense<[2, 1, 0]> : tensor<3xi64> } : (tensor<2x3x2xi32>) -> tensor<2x3x2xi32>
For this operation, permutation
is a constant, so it's available as a Value
in both semantics and constraints. In contrast, operand
and result
are available as a Value
in semantics but only as a Placeholder
in constraints.
There are no functions that can be used to construct types. Instead, we directly use type syntax because it's typically more concise. E.g. (tensor<E>, tensor<E>) -> (tensor<E>)
rather than function_type( [tensor_type([], E), tensor_type([], E)], [tensor_type([], E)])
.
element_type
is defined on tensor types and quantized tensor types and returns, respectively, the TensorElementType
or QuantizedTensorElementType
part of the corresponding TensorType
or QuantizedTensorType
.def element_type(x: Value | Placeholder | Type): if type(x) == TensorType: return tensor_element_type(x) if type(x) == QuantizedTensorType: return quantized_tensor_element_type(x) if type(x) is not Type: return element_type(type(x))
is_per_axis_quantized(x: Value | Placeholder | Type) -> Value
is a shortcut for is_quantized(x) and quantization_dimension(x) is not None
.
is_per_tensor_quantized(x: Value | Placeholder | Type) -> Value
is a shortcut for is_quantized(x) and quantization_dimension(x) is None
.
is_promotable(x: Type, y: Type) -> bool
checks if type x
can be promoted to type y
. When x
and y
are QuantizedTensorElementType
s, the promotion is applied only to the storage_type
. This specific version of promotion is currently used in context of reduction computation (refer to RFC for more details).
def is_promotable(x: Type, y: Type) -> Value: is_same_type = (is_bool(x) and is_bool(y)) or (is_integer(x) and is_integer(y)) or (is_float(x) and is_float(y)) or (is_complex(x) and is_complex(y)) or (is_quantized(x) and is_quantized(y) and expressed_type(x) = expressed_type(y)) if is_same_type == False: return False if is_integer(x) or is_float(x): return bitwidth(x) <= bitwidth(y) if is_complex(x): return bitwidth(element_type(x)) <= bitwidth(element_type(y)) if is_quantized(x): return bitwidth(storage_type(x)) <= bitwidth(storage_type(y)) return false
is_quantized(x: Value | Placeholder | Type) -> Value
is a shortcut for is_quantized_tensor_element_type(x)
.
is_type_name(x: Value | Placeholder | Type) -> Value
. Available for all types. For example, is_float(x)
returns true
if x
is a FloatType
. If x
is a value or placeholder, this function is a shortcut for is_type_name(type(x))
.
max_value(x: Type) -> Value
returns the maximum value of an TensorElementType
. If x
is not an TensorElementType
, returns None
.
min_value(x: Type) -> Value
returns the minimum possible value of an TensorElementType
. If x
is not an TensorElementType
, returns None
.
member_name(x: Value | Placeholder | Type) -> Any
. Available for all member definitions member_name
of all types. For example, tensor_element_type(x)
returns the TensorElementType
part of a corresponding TensorType
. If x
is a value or placeholder, this function is a shortcut for member_name(type(x))
. If x
is not a type that has an appropriate member, or a value or a placeholder of such a type, returns None
.
operation_name(*xs: Value | Type) -> Value
. Available for all operations. For example, add(lhs, rhs)
takes two tensor values lhs
and rhs
and returns the output of evaluating the add
operation with these inputs. For some operations e.g. broadcast_in_dim
, types of their outputs are “load-bearing”, i.e. needed to evaluate an operation. In this case, the function takes these types as arguments.All Python's operators and functions are available. E.g. both subscription and slicing notations from Python are available to index into tensors, quantized tensors and tuples.
to_destination_type(x: Value, destination_type: Type) -> Value
is defined on tensors and returns the converted value of x
based on the type(x)
and destination_type
as follows:
def to_destination_type(x: Value, destination_type: Type) -> Value: if type(x) == destination_type: return x if is_quantized(destination_type): if is_quantized(type(x)): return quantize(x, destination_type) assert is_float(type(x)) return quantize(x, destination_type) if is_quantized(type(x)): assert destination_type = expressed_type(type(x)) return dequantize(type(x)) return convert(x, destination_type)
There is early discussion on merging convert
, uniform_quantize
and uniform_dequantize
operations (#1576). After the merge we do not need the above function and can use the operation name for convert
instead.
is_nan(x: Value) -> Value
is defined on tensors and returns true
if all elements of x
are NaN
or false
otherwise. If x
is not a tensor, returns None
.
is_sorted(x: Value) -> Value
is defined on tensors and returns true
if elements of x
are sorted in ascending order with respect to the ascending lexicographical order of their indices or false
otherwise. If x
is not a tensor, returns None
.
is_unique(x: Value) -> Value
is defined on tensors and returns true
if x
doesn't have duplicate elements or false
otherwise. If x
is not a tensor, returns None
.
member_name(x: Value) -> Any
is defined for all member definitions member_name
of all values. For example, real_part(x)
returns the RealPart
part of a corresponding ComplexConstant
. If x
is not a value that has an appropriate member, returns None
.
same(x: Value) -> Value
is defined on tensors and returns true
if elements of x
are all equal to each other or false
otherwise. If the tensor doesn't have elements, that counts as “all equal to each other”, i.e. the function returns true
. If x
is not a tensor, returns None
.
split(x: Value, num_results: Value, axis: Value) -> Value
is defined on tensors and returns num_results
slices of x
along the axis axis
. If x
is not a tensor or dim(x, axis) % num_results != 0
, returns None
.
is_defined_in_parent_scope(x: Value) -> Value
is defined on strings and returns true
if x
is the name of a function defined in the same scope as the parent function of the relevant op.
is_namespaced_op_name(x: Value) -> Value
is defined on strings and returns true
if x
is a valid op name, that is it respects the following regular expression: [a-zA-Z][a-zA-Z0-9_]*([.][a-zA-Z0-9_$]+)+
axes(x: Value | Placeholder | Type) -> Value
is a shortcut for range(rank(x))
.
dim(x: Value | Placeholder | Type, axis: Value) -> Value
is a shortcut for shape(x)[axis]
.
dims(x: Value | Placeholder | Type, axes: List) -> List
is a shortcut for list(map(lambda axis: dim(x, axis), axes))
.
index_space(x: Value | Placeholder | Type) -> Value
is defined on tensors and returns size(x)
indices for the corresponding TensorType
sorted in ascending lexicographical order, i.e. [0, ..., 0]
, [0, ..., 1]
, ..., shape(x) - 1
. If x
is not a tensor type, a quantized tensor type, or a value or a placeholder of one of these types, returns None
.
rank(x: Value | Placeholder | Type) -> Value
is a shortcut for size(shape(x))
.
shape(x: Value | Placeholder | Type) -> Value
is defined in the “Functions on types” section via member_name
.
size(x: Value | Placeholder | Type) -> Value
is a shortcut for reduce(lambda x, y: x * y, shape(x))
.
def baseline_element_type(x: Value | Placeholder | Type) -> Type
is a shortcut for element_type(baseline_type(x))
.
baseline_type
is defined on tensor types and quantized tensor types and transforms them to a “baseline”, i.e. a type with the same shape but with the quantization parameters of the element type reset to default values. This is used as a handy trick to compare both tensor and quantized tensor types uniformly, which is needed quite often. For quantized types, this enables comparing types ignoring the quantization parameters, that is, shape
, storage_type
, expressed_type
, storage_min
, storage_max
, and quantization_dimension
(for per-axis quantized type) must all match, but scales
and zero points
may differ.
def baseline_type(x: Value | Placeholder | Type) -> Type: if type(x) == TensorType: return x if type(x) == QuantizedTensorType: element_type = quantized_tensor_element_type(x) baseline_element_type = QuantizedTensorElementType( storage_type = storage_type(element_type), storage_min = storage_min(element_type), storage_max = storage_max(element_type), expressed_type = expressed_type(element_type), quantization_dimension = quantization_dimension(element_type), scales = [constant(1.0, expressed_type(element_type))] * dim(x, quantization_dimension(element_type)), zero_points = [constant(0, storage_type(element_type))] * dim(x, quantization_dimension(element_type))) return QuantizedTensorType(shape(x), baseline_element_type) if type(x) is not Type: return baseline_element_type(type(x))
dequantize
is defined on quantized tensor types and turns them into floating-point tensor types. This happens via converting quantized elements which represent integer values of the storage type into corresponding floating-point values of the expressed type using the zero point and scale associated with the quantized element type.def compute_zero_points(quantized_type, result_type): if is_per_tensor_quantized(quantized_type): return broadcast_in_dim(constant(zero_point(quantized_type), storage_type(quantized_type)), [], result_type) if is_per_axis_quantized(quantized_type): for i in index_space(result_type): d = quantization_dimension(quantized_type) zero_points[i] = zero_points(quantized_type)[i[d]] return zero_points def compute_scales(quantized_type, result_type): if is_per_tensor_quantized(quantized_type): return broadcast_in_dim(constant(scale(quantized_type), expressed_type(quantized_type)), [], type(result_type)) if is_per_axis_quantized(quantized_type): for i in index_space(result_type): d = quantization_dimension(quantized_type) scales[i] = scales(quantized_type)[i[d]] return scales def dequantize(x: Value) -> Value: assert is_quantized(x) x_storage = bitcast_convert(x, storage_type(x)) x_storage_sub = x_storage - compute_zero_points(type(x), type(x_storage)) x_expressed_sub = convert(x_storage_sub, expressed_type(x)) return x_expressed_sub * compute_scales(type(x), type(x_expressed_sub))
quantize
is defined on floating-point tensor types and turns them into quantized tensor types. This happens via converting floating-point values of the expressed type into corresponding integer values of the storage type using the zero point and scale associated with the quantized element type.def quantize(x: Value, type: Type) -> Value: assert is_float(x) and is_quantized(type) x_expressed_rounded = round_nearest_even(x / compute_scales(type, type(x))) x_storage_rounded = convert(x_expressed_rounded, storage_type(type)) x_storage_add = x_storage_rounded + compute_zero_points(type, type(x_storage_rounded)) x_storage = clamp(storage_min(type), x_storage_add, storage_max(type)) return bitcast_convert(x_storage, type)
dequantize_op_quantize
is used to specify element-wise computations on quantized tensors. It dequantizes, i.e. turns quantized elements into their expressed types, then performs an operation, and then quantizes, i.e. turns the results back into their storage types. At the moment, this function only works for per-tensor quantization. Per-axis quantization is work in progress (#1574).def dequantize_op_quantize(op, *inputs_and_output_type): inputs = inputs_and_output_type[:-1] output_type = inputs_and_output_type[-1] float_inputs = map(dequantize, inputs) float_result = op(*float_inputs) return quantize(float_result, output_type) def dequantize_batch_norm_grad_or_training_quantize(op, *inputs_and_output_types): inputs = inputs_and_output_type[:-3] float_inputs = map(dequantize, inputs) float_results = op(*float_inputs) return map(quantize, float_results, inputs_and_output_type[-3:]) def dequantize_compare(lhs, rhs, comparison_direction): float_lhs = dequantize(lhs) float_rhs = dequantize(rhs) return compare(float_lhs, float_rhs, comparison_direction, FLOAT) def dequantize_select_quantize(pred, on_true, on_false, output_type): float_on_true = dequantize(on_true) float_on_false = dequantize(on_false) float_result = select(pred, float_on_true, float_on_false) return quantize(float_result, output_type)
hybrid_dequantize_then_op
is used to specify weight-only quantization for hybrid op which accepts lhs in floating-point and rhs in quantized types. It dequantizes quantized inputs into their expressed types and performs computation in float. Element type of float lhs tensor and expressed type of quantized rhs tensor should be identical.def hybrid_dequantize_then_op(op, lhs, rhs): assert(is_float(lhs) and is_quantized(rhs) and element_type(lhs) == expressed_type(rhs)) return op(lhs, dequantize(rhs))
cross_partition(replica_groups: Value) -> Value
. See the “cross_replica” section above.
cross_replica(replica_groups: Value) -> Value
. See the “cross_replica” section above.
cross_replica_and_partition(replica_groups: Value) -> Value
. See the “cross_replica_and_partition” section above.
flattened_ids(replica_groups: Value) -> Value
. See the “flattened_ids” section above.
StableHLO values can have dynamic dimension sizes, e.g. tensor<?xi64>
. However, StableHLO values cannot have a dynamic number of dimensions (unranked dynamism, e.g. tensor<*xi64>
). Operands and results are allowed to use dynamic dimension sizes, even if there are constraints on the sizes. Constraints will be verified statically if possible, otherwise they are deferred to runtime and mismatches will result in undefined behavior. See below for examples.
Consider the following toy program:
func.func @foo(%arg0: tensor<?xf64>) { %0 = stablehlo.abs %arg0 : (tensor<?xf64>) -> tensor<2xf64> return }
Such a program is unusual, because it is not common to know the shape of the result but not the shape of the input. Nonetheless, this is a valid StableHLO program. It is not possible to statically validate the abs
operation in this program, because the exact shape of the operand is unknown. However, the shapes are certainly compatible, and this can be checked statically: ?
could turn out to be 2
at runtime, and there would be no issue. However, ?
could also turn out to be some other integer, in which case the behavior is undefined.
Note that if a dimension size is dynamic in the result, there cannot be undefined behavior. Indeed, there is no “expected” size, so there cannot be a mismatch.
Consider the following toy program:
func.func @foo(%arg0: tensor<?xf64>, %arg1: tensor<?xf64>) { %0 = stablehlo.add %arg0, %arg0 : (tensor<?xf64>, tensor<?xf64>) -> tensor<?xf64> return }
When it comes to binary elementwise operations, the shapes of the inputs and the result must agree at runtime. At compile time, static dimensions must be equal, otherwise they merely need to be compatible. If any dimension is dynamic in the inputs, then there could be undefined behavior at runtime, because the dynamic size may not match the corresponding size in the other operand (be it static or dynamic). If all the inputs are static, then whether the result is dynamic or not does not matter: statically known dimensions will be checked statically, and dynamic dimensions do not impose any constraints.
Consider the following toy program:
func.func @foo(%arg0: tensor<2xi32>) { %0 = stablehlo.dynamic_iota %arg0, dim = 0 : (tensor<2xi32>) -> tensor<3x4xi64> return }
The values in the shape operand at runtime must match the shape of the result, otherwise the behavior is undefined. That is, at runtime %arg0
must have a value of dense<[3, 4]> : tensor<2xi32>
. If the shape operand is constant, this can be verified statically. If the result shape is fully dynamic, then there cannot be a mismatch.