'мхло' Диалект

Операции

mhlo.abs (mhlo::AbsOp)

Операция на прессе

Синтаксис:

operation ::= `mhlo.abs` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Выполняет поэлементную операцию абс над тензором operand и создает тензор result .

См.: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#abs .

Пример:

%result = mhlo.abs %operand : tensor<3xi32>

Признаки: AlwaysSpeculatableImplTrait , Elementwise , SameOperandsAndResultShape .

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.add (mhlo::AddOp)

Добавить операцию

Синтаксис:

operation ::= `mhlo.add` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Выполняет поэлементное сложение двух тензоров lhs и rhs и создает result тензор.

См.: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#add .

Пример:

%result = mhlo.add %lhs, %rhs : tensor<2x2xi32>

Признаки: AlwaysSpeculatableImplTrait , Commutative , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.add_dependency (mhlo::AddDependencyOp)

Операция AddDependency

Синтаксис:

operation ::= `mhlo.add_dependency` operands attr-dict `:` functional-type(operands, results)

Эта операция является частной для компилятора XLA, поэтому для нее еще нет спецификации.

Неформально у этой операции два операнда: операнд данных и токен. Результатом операции является операнд данных. При использовании с AfterAll эта операция позволяет упорядочивать операции, не вызывающие побочных эффектов (те, которые не создают значения токенов).

Пример:

%1 = mhlo.add_dependency %arg0, %0 : (tensor<3x4xf32>, !mhlo.token) -> tensor<3x4xf32>

Черты: AlwaysSpeculatableImplTrait

Интерфейсы: ConditionallySpeculatable , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.after_all (mhlo::AfterAllOp)

После операции

Синтаксис:

operation ::= `mhlo.after_all` $inputs attr-dict
              `:` custom<VariadicSameOperandsAndResultType>(ref($inputs), type($inputs), type($result))

Гарантирует, что операции, производящие inputs , выполняются до выполнения любых операций, зависящих от result .

См.: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#after_all .

Пример:

%result = mhlo.after_all %input0, %input1 : !mhlo.token

Черты: AlwaysSpeculatableImplTrait

Интерфейсы: ConditionallySpeculatable , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.all_gather (mhlo::AllGatherOp)

Операция AllGather

Внутри каждой группы процессов в сетке процессов объединяет значения тензора операнда каждого процесса по all_gather_dim и создает результирующий тензор. computation применяются отдельно для каждого операнда в operands , давая один результат для каждого операнда.

См.: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#all_gather .

Пример:

%result = "mhlo.all_gather"(%operand) {
  all_gather_dim = 1 : i64,
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>
  // channel_id = 0
  channel_handle = #mhlo.channel_handle<handle = 0, type = 0>,
  // use_global_device_ids = false
} : (tensor<2x2xf32>) -> tensor<2x4xf32>

Признаки: SameOperandsAndResultElementType

Атрибуты:

Операнды:

Полученные результаты:

mhlo.all_reduce (mhlo::AllReduceOp)

Операция AllReduce

В каждой группе процессов в сетке процессов применяется computation функции сокращения к значениям тензора операнда каждого процесса и создается тензор результата. computation применяются отдельно для каждого операнда в operands , давая один результат для каждого операнда.

См.: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#all_reduce .

Пример:

%result = "mhlo.all_reduce"(%operand) ({
  ^bb0(%arg0: tensor<f32>, %arg1: tensor<f32>):
    %0 = mhlo.add %arg1, %arg2 : tensor<f32>
    mhlo.return %0 : tensor<f32>
}) {
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>
  // channel_id = 0
  channel_handle = #mhlo.channel_handle<handle = 0, type = 0>
  // use_global_device_ids = false
} : (tensor<4xf32>) -> tensor<4xf32>

Признаки: InferTensorType , SingleBlockImplicitTerminator<ReturnOp> , SingleBlock

Интерфейсы: InferShapedTypeOpInterface , InferTypeOpInterface

Атрибуты:

Операнды:

Полученные результаты:

mhlo.all_to_all (mhlo::AllToAllOp)

Операция AllToAll

Внутри каждой группы процессов в сетке процессов разбивает значения тензора operand по split_dimension на части, распределяет разделенные части между процессами, объединяет разбросанные части по concat_dimension и создает result тензор.

См.: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#all_to_all .

Пример:

%result = "mhlo.all_to_all"(%operand) {
  split_dimension = 1 : i64,
  concat_dimension = 0 : i64,
  split_count = 2 : i64,
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>
} : (tensor<2x4xf32>) -> tensor<4x2xf32>

Признаки: AlwaysSpeculatableImplTrait , InferTensorType , SameOperandsElementType , SameOperandsShape , SameVariadicOperandSize

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Атрибуты:

Операнды:

Полученные результаты:

mhlo.and (mhlo::AndOp)

И операция

Синтаксис:

operation ::= `mhlo.and` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Выполняет поэлементное И двух тензоров lhs и rhs и создает result тензор.

См.: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#and

Пример:

%result = mhlo.and %lhs, %rhs : tensor<2x2xi32>

Признаки: AlwaysSpeculatableImplTrait , Commutative , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.async_done (mhlo::AsyncDoneOp)

Операция AsyncDone

Эта операция является частной для компилятора XLA, поэтому для нее еще нет спецификации.

Неформально эта операция блокируется до завершения асинхронных вычислений. Он возвращает окончательный результат асинхронных вычислений.

Дополнительную информацию см. в документации по AsyncStart.

Интерфейсы: InferTypeOpInterface

Атрибуты:

Операнды:

Полученные результаты:

mhlo.async_start (mhlo::AsyncStartOp)

Операция асинхронного запуска

Эта операция является частной для компилятора XLA, поэтому для нее еще нет спецификации.

Неформально эта операция запускает асинхронные вычисления.

Это используется, когда существуют функции, которые содержат как асинхронное ожидание (например, DMA), так и вычисления в потоке. Например, функция может состоять из вычисления, прямого доступа к памяти, другого вычисления, второго прямого доступа к памяти и окончательного вычисления. Это будет представлено как async_start, за которым следуют async_update и async_done. async_start выполнит первые вычисления в потоке, а затем запустит DMA. async_update будет ждать завершения DMA, если оно еще не было выполнено, затем выполнит второе вычисление в функции и запустит второй DMA. Наконец, async_done будет ждать этого последнего DMA, а затем запускает последнее вычисление, которое необходимо выполнить в потоке, и возвращает результат этого окончательного вычисления.

operands передаются в вычисление напрямую. called_computation — это функция, которая будет выполняться асинхронно. execution_thread — это имя потока, в котором она будет выполняться. Основной поток называется «главным». Все темы имеют названия.

Это возвращает все состояние, необходимое между асинхронными операциями. После назначения буфера возвращаемые значения представляют собой пространство, необходимое для хранения входных данных, результатов и любых блокнотов, необходимых или редактируемых асинхронной операцией.

Атрибуты:

Операнды:

Полученные результаты:

mhlo.async_update (mhlo::AsyncUpdateOp)

Операция AsyncUpdate

Эта операция является частной для компилятора XLA, поэтому для нее еще нет спецификации.

Неформально эта операция блокирует асинхронные вычисления до тех пор, пока не возникнет барьер синхронизации. Это возвращает bundle после работы с ним.

Дополнительную информацию см. в документации по AsyncStart.

Интерфейсы: InferTypeOpInterface

Атрибуты:

Операнды:

Полученные результаты:

mhlo.atan2 (mhlo::Atan2Op)

Операция Атан2

Синтаксис:

operation ::= `mhlo.atan2` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Выполняет поэлементную операцию atan2 над тензорами lhs и rhs и создает result тензор.

См.: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#atan2 .

Пример:

%result = mhlo.atan2 %lhs, %rhs : tensor<3xf32>

Признаки: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.batch_norm_grad (mhlo::BatchNormGradOp)

Операция «БатчНормГрад»

Вычисляет градиенты нескольких входных данных BatchNormTrainingOp с обратным распространением ошибки от grad_output и создает тензоры grad_operand , grad_scale и grad_offset .

См.: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#batch_norm_grad .

Пример:

%grad_operand, %grad_scale, %grad_offset =
"mhlo.batch_norm_grad"(%operand, %scale, %mean, %variance, %grad_output) {
  epsilon = 0.0 : f32,
  feature_index = 2 : i64
} : (tensor<2x2x2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
    tensor<2x2x2xf32>) -> (tensor<2x2x2xf32>, tensor<2xf32>, tensor<2xf32>)

Черты: AlwaysSpeculatableImplTrait , InferTensorType

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Атрибуты:

Операнды:

Полученные результаты:

mhlo.batch_norm_inference (mhlo::BatchNormInferenceOp)

Операция BatchNormInference

Нормализует тензор operand по всем измерениям, кроме измерения feature_index , и создает result тензор.

См.: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#batch_norm_inference .

Пример:

%result = "mhlo.batch_norm_inference"(%operand, %scale, %offset, %mean, %variance) {
  epsilon = 0.0 : f32,
  feature_index = 2 : i64
} : (tensor<2x2x2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>) -> tensor<2x2x2xf32>

Черты: AlwaysSpeculatableImplTrait , InferTensorType

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Атрибуты:

Операнды:

Полученные результаты:

mhlo.batch_norm_training (mhlo::BatchNormTrainingOp)

Пакетная нормаОбучающая операция

Вычисляет среднее значение и дисперсию для пакетных и пространственных измерений, нормализует тензор operand для каждого объекта в измерении feature_index и создает output тензоры, batch_mean и batch_var .

См.: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#batch_norm_training .

Пример:

%output, %batch_mean, %batch_var = "mhlo.batch_norm_training"(%operand, %scale, %offset) {
  epsilon = 0.0 : f32,
  feature_index = 2 : i64
} : (tensor<2x2x2xf32>, tensor<2xf32>, tensor<2xf32>) -> (tensor<2x2x2xf32>, tensor<2xf32>, tensor<2xf32>)

Черты: AlwaysSpeculatableImplTrait , InferTensorType

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Атрибуты:

Операнды:

Полученные результаты:

mhlo.bitcast (mhlo::BitcastOp)

Биткаст-операция

Синтаксис:

operation ::= `mhlo.bitcast` operands attr-dict `:` functional-type(operands, results)

Эта операция является частной для компилятора XLA, поэтому для нее еще нет спецификации.

Неформально эта операция меняет форму входных данных таким образом, что физическое расположение элементов остается неизменным.

Для этой операции необходима информация о макете, чтобы понять «физическое расположение элементов», а поддержка макета в MHLO в настоящее время находится в стадии разработки.

Пример:

%0 = mhlo.bitcast %arg0 : (tensor<3x4xf32>) -> tensor<3x4x1xf32>

Черты: AlwaysSpeculatableImplTrait

Интерфейсы: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.bitcast_convert (mhlo::BitcastConvertOp)

Операция BitcastConvert

Синтаксис:

operation ::= `mhlo.bitcast_convert` operands attr-dict `:` functional-type(operands, results)

Выполняет операцию побитового преобразования тензора operand и создает тензор result , в котором биты всего operand переинтерпретируются с использованием типа тензора result .

См.: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#bitcast_convert .

Пример:

%result = mhlo.bitcast_convert %operand : (tensor<2xf32>) -> tensor<2x4xi8>

Черты: AlwaysSpeculatableImplTrait

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.broadcast (mhlo :: froadcastop)

Вещательная операция

Эта операция выходит из stablehlo, поэтому она не включена в спецификацию: https://github.com/openxla/stablehlo/issues/3

Неофициально, эта операция делает то же самое, что и трансляция XLA: https://www.tensorflow.org/xla/operation_semantics#broadcast

Пример:

%result = mhlo.broadcast %operand, sizes = [1, 2] : (tensor<3xi32>) -> tensor<1x2x3xi32>

Черты: AlwaysSpeculatableImplTrait , InferTensorType , SameOperandsAndResultElementType

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Атрибуты:

Операнды:

Полученные результаты:

mhlo.broadcast_in_dim (mhlo :: broadcastNidimop)

РАЗВИТИЯ ВВЕДЕНИЯ

Расширяет размеры и/или ранг входного тензора, дублируя данные в тензоре operand и дает тензор result .

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#broadcast_in_dim

Пример:

%result = mhlo.broadcast_in_dim %operand, dims = [2, 1] : (tensor<1x3xi32>) -> tensor<2x3x2xi32>

Черты: AlwaysSpeculatableImplTrait , HLO_CompatibleOperandsAndResultElementType

Интерфейсы: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Атрибуты:

Операнды:

Полученные результаты:

mhlo.case (mhlo :: caseop)

Дело

Производит выход из выполнения ровно одной function из branches в зависимости от значения index .

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#case

Пример:

%result0, %result1 = "mhlo.case"(%index) ({
  mhlo.return %result_branch0, %result_branch0 : tensor<2xi64>, tensor<2xi64>
}, {
  mhlo.return %result_branch1, %result_branch1 : tensor<2xi64>, tensor<2xi64>
}) : (tensor<i32>) -> (tensor<2xi64>, tensor<2xi64>)

Черты: RecursiveMemoryEffects , SingleBlockImplicitTerminator<ReturnOp> , SingleBlock

Интерфейсы: InferTypeOpInterface

Операнды:

Полученные результаты:

mhlo.cbrt (mhlo :: cbrtop)

Операция CBRT

Синтаксис:

operation ::= `mhlo.cbrt` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Выполняет элементную работу кубического корня на тензоре operand и получает тензор result .

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#cbrt

Пример:

%result = mhlo.cbrt %operand : tensor<4xf32>

Черты: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.ceil (mhlo :: ceilop)

CEIL Operation

Синтаксис:

operation ::= `mhlo.ceil` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Выполняет элементный Ceil из operand Tensor и дает тензор result .

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#ceil

Пример:

%result = mhlo.ceil %operand : tensor<5xf32>

Черты: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.cholesky (mhlo :: chouleskyop)

Чолская операция

Вычисляет размесщение хоузского партии матриц.

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#cholesky

Пример:

%result = mhlo.cholesky %a, lower = true : tensor<3x3xf32>

Черты: AlwaysSpeculatableImplTrait , InferTensorType , SameOperandsAndResultElementType

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Атрибуты:

Операнды:

Полученные результаты:

mhlo.clamp (mhlo :: clampop)

Операция зажима

Синтаксис:

operation ::= `mhlo.clamp` $min `,` $operand `,` $max attr-dict
              `:` custom<SameOperandsAndResultType>(type($min), type($operand), type($max), type($result))

Зажимая каждый элемент тензора operand между минимальным и максимальным значением и дает тензор result .

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#clamp

Пример:

%result = mhlo.clamp %min, %operand, %max : tensor<3xi32>

Черты: AlwaysSpeculatableImplTrait , HLO_BroadcastingElementwise , InferTensorType , SameOperandsAndResultElementType

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.collective_broadcast (mhlo :: collectivebroadcastop)

CollectiveBroadcast Operation

В каждой группе процессов в сетке процесса отправьте значение тензора operand из исходного процесса в целевые процессы и дают тензор result .

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#collective_broadcast

Пример:

%result = "mhlo.collective_broadcast"(%operand) {
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>,
  channel_handle = #mhlo.channel_handle<handle = 0, type = 0>
} : (tensor<1x2xi64>) -> tensor<1x2xi64>

Черты: CompatibleOperandsAndResultType

Интерфейсы: InferShapedTypeOpInterface , InferTypeOpInterface

Атрибуты:

Операнды:

Полученные результаты:

mhlo.collective_permute (mhlo :: collectivepermuteop)

CollectivePermute Operate

В каждой группе процессов в сетке процесса отправляет значение тензора operand из исходного процесса в целевой процесс и дает тензор result .

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#collective_permute

Пример:

%result = "mhlo.collective_permute"(%operand) {
  source_target_pairs = dense<[[0, 1], [1, 2]]> : tensor<2x2xi64>,
  // channel_id = 0
  channel_handle = #mhlo.channel_handle<handle = 0, type = 0>
} : (tensor<4x2xf32>) -> tensor<4x2xf32>

Черты: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Атрибуты:

Операнды:

Полученные результаты:

mhlo.compare (mhlo :: compareop)

Сравните операцию

Синтаксис:

operation ::= `mhlo.compare` $comparison_direction `,` $lhs `,` $rhs (`,` $compare_type^)?
              attr-dict `:` functional-type(operands, results)

Проводит элементное сравнение тензоров lhs и rhs в соответствии с comparison_direction и compare_type и дает result тензора.

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#compare

Пример:

%result = mhlo.compare LT, %lhs, %rhs, FLOAT : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xi1>

Черты: AlwaysSpeculatableImplTrait , Elementwise , InferTensorType , SameOperandsAndResultShape , SameOperandsElementType

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Атрибуты:

Операнды:

Полученные результаты:

mhlo.complex (mhlo :: complexop)

Сложная операция

Синтаксис:

operation ::= `mhlo.complex` operands attr-dict
              `:` custom<ComplexOpType>(type($lhs), type($rhs), type($result))

Выполняет элементное преобразование в сложное значение из пары реальных и воображаемых значений, lhs и rhs , и дает тензор result .

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#complex

Пример:

%result = mhlo.complex %lhs, %rhs : tensor<2xcomplex<f32>>

Черты: AlwaysSpeculatableImplTrait , Elementwise , SameOperandsAndResultShape SameOperandsElementType

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.composite (mhlo :: compositeop)

Композитная операция

Синтаксис:

operation ::= `mhlo.composite` $name $inputs attr-dict `:` functional-type(operands, results)

Инкапсулирует операцию, выполненную (составленную) других операций StableHlo, принимая inputs и composite_attributes и дает results . Семантика OP реализована атрибутом decomposition . composite OP может быть заменен его разложением без изменения семантики программы. В тех случаях, когда внедрение разложения не обеспечивает ту же операционную семантику, предпочитаю использовать custom_call .

Поле version (по умолчанию к 0 ) используется для обозначения при изменении семантики композита.

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#composite

Пример:

%results = mhlo.composite "my.op" %arg0, %arg1 {
  decomposition = @my_op,
  composite_attributes = { my_attribute = "my_value" },
  version = 1 : i32
} : (tensor<f32>, tensor<f32>) -> tensor<f32>

Интерфейсы: SymbolUserOpInterface

Атрибуты:

Операнды:

Полученные результаты:

mhlo.compute_reshape_shape (mhlo :: computershapeShapeop)

ComputerShapeShape Operation

Синтаксис:

operation ::= `mhlo.compute_reshape_shape` operands attr-dict `:` functional-type(operands, results)

Эта операция находится в стадии разработки, поэтому она еще не включена в спецификацию: https://github.com/openxla/stablehlo/issues/8

Неофициально, эта операция вычисляет output_shape для DynamicReshapeop из числа num_elements элементов в операнде DynamicReshapeop и форме dynamic_shape , предоставленной для RESHAPE TF: https://www.tensorflow.org/api_docs/python/tf/reshape

Например, для num_elements = 12 и dynamic_shape = [2, -1] result будет [2, 6] . Если операнды недопустимы (например, если измерения не разделяют количество элементов или если в измерениях есть несколько значений -1), это приводит к неопределенному поведению.

Пример:

%result = mhlo.compute_reshape_shape %num_elements, %dynamic_shape
       : (index, tensor<2xi32>) -> tensor<2xi32>

Черты: AlwaysSpeculatableImplTrait

Интерфейсы: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.concatenate (mhlo :: concatenateop)

Конкатенатная операция

Соглашает варидовое количество тензоров в inputs вдоль измерения dimension в том же порядке, что и заданные аргументы, и дает тензор result .

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#concatenate

Пример:

%result = mhlo.concatenate %input0, %input1, dim = 0 : (tensor<3x2xi64>, tensor<1x2xi64>) -> tensor<4x2xi64>

Черты: AlwaysSpeculatableImplTrait , SameOperandsAndResultElementType

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Атрибуты:

Операнды:

Полученные результаты:

mhlo.constant (mhlo :: constantop)

Постоянная операция

Производит output тензор из постоянного value .

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#constant

Пример:

%output = mhlo.constant dense<[[0.0, 1.0], [2.0, 3.0]]> : tensor<2x2xf32>

Черты: AlwaysSpeculatableImplTrait , ConstantLike

Интерфейсы: ConditionallySpeculatable , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Атрибуты:

Полученные результаты:

mhlo.convert (mhlo :: convertop)

Преобразовать операцию

Синтаксис:

operation ::= `mhlo.convert` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Выполняет элементное преобразование от одного типа элемента в другой на operand тензоре и дает тензор result .

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#convert

Пример:

%result = mhlo.convert %operand : (tensor<3xi32>) -> tensor<3xcomplex<f32>>

Черты: AlwaysSpeculatableImplTrait , Elementwise , SameOperandsAndResultShape

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Операнды:

Полученные результаты:

mhlo.convolution (mhlo :: convolutionop)

Строильная операция

Синтаксис:

operation ::= `mhlo.convolution` `(`operands`)`
              `dim_numbers` `=` custom<ConvolutionDimensions>($dimension_numbers) `,`
              `window` `=` `{` custom<WindowAttributes>($window_strides, $padding,
              $lhs_dilation, $rhs_dilation,
              $window_reversal) `}`
              attr-dict `:` functional-type(operands, results)

Вычисляет точечные продукты между окнами lhs и ломтиками rhs и дает result .

См.: Https://github.com/openxla/stablehlo/blob/main/docs/spec.md#convolution

Пример:

%result = "mhlo.convolution"(%lhs, %rhs) {
  window_strides = dense<4> : tensor<2xi64>,
  padding = dense<0> : tensor<2x2xi64>,
  lhs_dilation = dense<2> : tensor<2xi64>,
  rhs_dilation = dense<1> : tensor<2xi64>,
  window_reversal = dense<false> : tensor<2xi1>,
  dimension_numbers = #mhlo.conv<[b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f]>,
  feature_group_count = 1 : i64,
  batch_group_count = 1 : i64,
  precision_config = [#stablehlo<precision DEFAULT>, #stablehlo<precision DEFAULT>]
} : (tensor<1x4x4x1xi32>, tensor<3x3x1x1xi32>) -> tensor<1x2x2x1xi32>

Черты: AlwaysSpeculatableImplTrait

Интерфейсы: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Атрибуты:

Операнды:

Полученные результаты:

mhlo.copy (mhlo :: copyop)

Копия операции

Синтаксис:

operation ::= `mhlo.copy` operands attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Эта операция является частной для компилятора XLA, поэтому она еще не имеет спецификации.

Неофициально, эта операция копия operand . В зависимости от метаданных, прикрепленных к операции, он может вести себя совершенно иначе от NO-OP.

Пример:

%0 = mhlo.copy %arg0 : tensor<f32>

Черты: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise

Интерфейсы: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Эффекты: MemoryEffects::Effect{}

Атрибуты:

Операнды:

Полученные результаты:

mhlo.cosine (mhlo::CosineOp)

Cosine operation

Синтаксис:

operation ::= `mhlo.cosine` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise cosine operation on operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#cosine

Пример:

%result = mhlo.cosine %operand : tensor<2xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.count_leading_zeros (mhlo::ClzOp)

Clz operation

Синтаксис:

operation ::= `mhlo.count_leading_zeros` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise count of the number of leading zero bits in the operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#count_leading_zeros

Пример:

%result = mhlo.count_leading_zeros %operand : tensor<2x2xi8>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.create_token (mhlo::CreateTokenOp)

CreateToken operation

Синтаксис:

operation ::= `mhlo.create_token` attr-dict `:` type(results)

This operation is on its way out of StableHLO, so it is not included in the specification: https://github.com/openxla/stablehlo/issues/3

Informally, this operation does the same thing as AfterAllOp with 0 inputs: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#after_all

Пример:

%output = mhlo.create_token : !mhlo.token

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Полученные результаты:

mhlo.cross-replica-sum (mhlo::CrossReplicaSumOp)

CrossReplicaSum operation

This operation is on its way out of StableHLO, so it is not included in the specification: https://github.com/openxla/stablehlo/issues/3

Informally, this operation does the same thing as AllReduceOp with channel_id = 0 , use_global_device_ids = false and computation implementing addition: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#all_reduce

Пример:

%result = "mhlo.cross-replica-sum"(%operand) {
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>
} : (tensor<4xf32>) -> tensor<4xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.cstr_reshapable (mhlo::CstrReshapableOp)

CstrReshapable operation

Синтаксис:

operation ::= `mhlo.cstr_reshapable` operands attr-dict `:` functional-type(operands, results)

This operation is a work in progress, so it is not yet included in the specification: https://github.com/openxla/stablehlo/issues/8

Informally, this operation creates a witness on the constraint that ComputeReshapeShape would succeed with the provided operands.

Пример:

%result = mhlo.cstr_reshapable %num_elements, %dynamic_shape
       : (index, tensor<3xi32>) -> !shape.witness

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.custom_call (mhlo::CustomCallOp)

CustomCall operation

Синтаксис:

operation ::= `mhlo.custom_call` custom<CustomCallTarget>($call_target_name) `(` $inputs `)`
              attr-dict `:` functional-type(operands, results)

Encapsulates an implementation-defined operation call_target_name that takes inputs and called_computations and produces results .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#custom_call

Пример:

%results = "mhlo.custom_call"(%input0) {
  call_target_name = "foo",
  has_side_effect = false,
  backend_config = "bar",
  api_version = 1 : i32,
  called_computations = [@foo]
} : (tensor<f32>) -> tensor<f32>

A custom call invokes code external to XLA. The `inputs` are passed to the
external code, and the external code is expected to produce a result of the
given type. The exact mechanism is backend-specific. For example, in the CPU
backend, a call instruction is emitted which targets a symbol with the name
`call_target_name`.

If XLA runtime is enabled for a backend, then custom calls use the runtime
custom call calling convention to call into the external functions. This
calling convention defines an ABI for encoding arguments, attributes and
results.

Depending on the API version there are two ways to pass extra bits of static
information to the external function:

1. For `API_VERSION_TYPED_FFI` custom calls `backend_config` must be a
   dictionary attribute, that will be encoded according to the custom call
   calling convention and passed to the external function as the attributes
   argument. External code is expected to use declarative bindings (see
   `xla/runtime/custom_call.h`) to decode them at run time. These custom
   calls are only supported if XLA uses XLA runtime.

2. For previous API versions it is the user responsibility to encode extra
   bits of static information as a string `backend_config` attribute, and
   decode it at run time.

Interfaces: MemoryEffectOpInterface

Атрибуты:

Operands:

Полученные результаты:

mhlo.divide (mhlo::DivOp)

Div operation

Синтаксис:

operation ::= `mhlo.divide` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Performs element-wise division of dividend lhs and divisor rhs tensors and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#divide

Пример:

%result = mhlo.divide %lhs, %rhs : tensor<4xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.domain (mhlo::DomainOp)

Domain operation

This operation is private to the XLA compiler, so it is does not yet have a specification.

Informally, these operations are used to group instructions with the same DomainMetadata property. ShardingMetadata is the main use case today to group instructions on the same device. Domain instructions provide two major benefits:

• Prevent unintentionally optimizing instructions across domains.
• Automatically assign the metadata of the instructions created in the domain. Without domain instructions, each HLO optimization pass would have to check and propagate the metadata, which would be easy to miss and also adds complexity to the compiler. Since domain instructions connect two different domains, each domain instruction is associated with two DomainMetadata -- one on the operand side and one on the user side of the domain.

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.dot (mhlo::DotOp)

Dot operation

This operation is on its way out of StableHLO, so it is not included in the specification: https://github.com/openxla/stablehlo/issues/3

Informally, this operation does the same thing as XLA's Dot: https://www.tensorflow.org/xla/operation_semantics#dot

Пример:

%0 = mhlo.dot %arg0, %arg1 : (tensor<1x2xi32>, tensor<2x1xi32>) -> tensor<1x1xi32>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.dot_general (mhlo::DotGeneralOp)

DotGeneral operation

Computes dot products between slices of lhs and slices of rhs and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#dot_general

Пример:

%result = "mhlo.dot_general"(%lhs, %rhs) {
  dot_dimension_numbers = #mhlo.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<2x2x2xi32>, tensor<2x2x2xi32>) -> tensor<2x2x2xi32>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.dynamic_broadcast_in_dim (mhlo::DynamicBroadcastInDimOp)

DynamicBroadcastInDim operation

This operation is a work in progress, so it is not yet included in the specification: https://github.com/openxla/stablehlo/issues/8

Informally, this operation does the same thing as BroadcastInDimOp except that the result shape is specified dynamically via output_dimensions : https://github.com/openxla/stablehlo/blob/main/docs/spec.md#broadcast_in_dim

It also accepts optional attributes to express static knowledge about the expanding behavior of dimensions. If not specified, all dimensions are assumed to be possibly expanding. The sets of dimensions that are known to be expanding and the set of dimensions that are known to be non-expanding must be disjoint and they must be a subset of the operand's dimensions.

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.dynamic_conv (mhlo::DynamicConvOp)

DynamicConv operation

This operation is a work in progress, so it is not yet included in the specification: https://github.com/openxla/stablehlo/issues/8

Informally, this operation does the same thing as ConvolutionOp except that padding is specified dynamically via d_padding : https://github.com/openxla/stablehlo/blob/main/docs/spec.md#convolution

Пример:

%result = "mhlo.dynamic_conv"(%lhs, %rhs, %d_padding) {
  window_strides = dense<4> : tensor<2xi64>,
  lhs_dilation = dense<2> : tensor<2xi64>,
  rhs_dilation = dense<1> : tensor<2xi64>,
  window_reversal = dense<false> : tensor<2xi1>,
  dimension_numbers = #mhlo.conv<[b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f]>,
  feature_group_count = 1 : i64,
  batch_group_count = 1 : i64,
  precision_config = [#stablehlo<precision DEFAULT>, #stablehlo<precision DEFAULT>]
} : (tensor<1x4x4x1xi32>, tensor<3x3x1x1xi32>, tensor<2x2xi64>) -> tensor<1x2x2x1xi32>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.dynamic_gather (mhlo::DynamicGatherOp)

DynamicGather operation

This operation is a work in progress, so it is not yet included in the specification: https://github.com/openxla/stablehlo/issues/8

Informally, this operation does the same thing as GatherOp except that slice_sizes are specified dynamically: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#gather

Пример:

%result = "mhlo.dynamic_gather"(%operand, %start_indices, %slice_sizes) {
  dimension_numbers = #mhlo.gather<
    offset_dims = [2, 3],
    collapsed_slice_dims = [0],
    start_index_map = [0, 2],
    index_vector_dim = 2>,
  indices_are_sorted = false
} : (tensor<3x4x2xi32>, tensor<2x3x2xi64>, tensor<3xi64>) -> tensor<2x3x2x2xi32>

Traits: AlwaysSpeculatableImplTrait , InferTensorType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.dynamic_iota (mhlo::DynamicIotaOp)

DynamicIota operation

This operation is a work in progress, so it is not yet included in the specification: https://github.com/openxla/stablehlo/issues/8

Informally, this operation does the same thing as IotaOp except that the result shape is specified dynamically via output_shape : https://github.com/openxla/stablehlo/blob/main/docs/spec.md#iota

Пример:

%0 = mhlo.dynamic_iota %arg0, dim = 0 : (tensor<1xindex>) -> tensor<4xi32>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.dynamic_pad (mhlo::DynamicPadOp)

DynamicPad operation

Синтаксис:

operation ::= `mhlo.dynamic_pad` operands attr-dict `:` functional-type(operands, results)

Dynamically Pads the operand , with amount of padding added at low-end/high-end/interior is passed through input tensors.

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.dynamic_reshape (mhlo::DynamicReshapeOp)

DynamicReshape operation

Синтаксис:

operation ::= `mhlo.dynamic_reshape` operands attr-dict `:` functional-type(operands, results)

This operation is a work in progress, so it is not yet included in the specification: https://github.com/openxla/stablehlo/issues/8

Informally, this operation does the same thing as ReshapeOp except that the result shape is specified dynamically via output_shape : https://github.com/openxla/stablehlo/blob/main/docs/spec.md#reshape

Пример:

%0 = mhlo.dynamic_reshape %arg0, %shape : (tensor<?xf32>, tensor<2xindex>) -> tensor<?x?xf32>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.dynamic_slice (mhlo::DynamicSliceOp)

DynamicSlice operation

Extracts a slice from the operand using dynamically-computed starting indices and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#dynamic_slice

Пример:

%result = mhlo.dynamic_slice %operand, %start_indices0, %start_indices1, sizes = [2, 2]
  : (tensor<4x4xi32>, tensor<i64>, tensor<i64>) -> tensor<2x2xi32>

Traits: AlwaysSpeculatableImplTrait , InferTensorType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.dynamic_update_slice (mhlo::DynamicUpdateSliceOp)

DynamicUpdateSlice operation

Синтаксис:

operation ::= `mhlo.dynamic_update_slice` operands attr-dict `:` functional-type(operands, results)

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 .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#dynamic_update_slice

Пример:

%result = mhlo.dynamic_update_slice %operand, %update, %start_indices0, %start_indices1
  : (tensor<4x4xi32>, tensor<2x2xi32>, tensor<i64>, tensor<i64>) -> tensor<4x4xi32>

Traits: AlwaysSpeculatableImplTrait , InferTensorType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.einsum (mhlo::EinsumOp)

Einsum operation

This operation is on its way out of StableHLO, so it is not included in the specification: https://github.com/openxla/stablehlo/issues/3

Informally, this operation does the same thing as TF's einsum: https://www.tensorflow.org/api_docs/python/tf/einsum

Пример:

%result = "mhlo.einsum"(%lhs, %rhs) {
  einsum_config = "ab,bc->ac"
} : (tensor<4x16xf32>, tensor<16x4xf32>) -> tensor<4x4xf32>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.erf (mhlo::ErfOp)

Erf operation

Синтаксис:

operation ::= `mhlo.erf` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise erf operation on operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#erf

Пример:

%result = mhlo.erf %operand : tensor<2x2xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.exponential (mhlo::ExpOp)

Exp operation

Синтаксис:

operation ::= `mhlo.exponential` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise exponential operation on operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#exponential

Пример:

%result = mhlo.exponential %operand : tensor<2x2xf64>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.exponential_minus_one (mhlo::Expm1Op)

Expm1 operation

Синтаксис:

operation ::= `mhlo.exponential_minus_one` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise exponential minus one operation on operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#exponential_minus_one

Пример:

%result = mhlo.exponential_minus_one %operand : tensor<2xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.fft (mhlo::FftOp)

Fft operation

Performs the forward and inverse Fourier transforms for real and complex inputs/outputs.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#fft

Пример:

%result = mhlo.fft %operand, type = FFT, length = [4] : (tensor<4xcomplex<f32>>) -> tensor<4xcomplex<f32>>

Traits: AlwaysSpeculatableImplTrait , InferTensorType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.floor (mhlo::FloorOp)

Floor operation

Синтаксис:

operation ::= `mhlo.floor` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise floor of operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#floor

Пример:

%result = mhlo.floor %operand : tensor<2xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.fusion (mhlo::FusionOp)

Fusion operation

This operation is private to the XLA compiler, so it is does not yet have a specification.

Informally, this operation consists of a group of basic ops (represented as a region attached to it). It serves as a hint to the backend that it is beneficial to emit the contained ops into a single loop nest or kernel.

Атрибуты:

Operands:

Полученные результаты:

mhlo.gather (mhlo::GatherOp)

Gather operation

Gathers slices from operand tensor from offsets specified in start_indices and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#gather

Пример:

%result = "mhlo.gather"(%operand, %start_indices) {
  dimension_numbers = #mhlo.gather<
    offset_dims = [2, 3],
    collapsed_slice_dims = [0],
    start_index_map = [0, 2],
    index_vector_dim = 2>,
  slice_sizes = dense<[0, 2, 2]> : tensor<3xi64>,
  indices_are_sorted = false
} : (tensor<3x4x2xi32>, tensor<2x3x2xi64>) -> tensor<2x3x2x2xi32>

Traits: AlwaysSpeculatableImplTrait , InferTensorType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.get_dimension_size (mhlo::GetDimensionSizeOp)

GetDimensionSize operation

Produces the size of the given dimension of the operand .

See https://github.com/openxla/stablehlo/blob/main/docs/spec.md#get_dimension_size

Пример:

%result = mhlo.get_dimension_size %operand, dim = 1 : (tensor<2x3xf32>) -> tensor<i32>

Traits: AlwaysSpeculatableImplTrait , InferTensorType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.get_tuple_element (mhlo::GetTupleElementOp)

GetTupleElement operation

Синтаксис:

operation ::= `mhlo.get_tuple_element` $operand `[` $index `]` attr-dict `:` functional-type(operands, results)

Extracts element at index position of the operand tuple and produces a result .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#get_tuple_element

Пример:

%result = mhlo.get_tuple_element %operand[0] : (tuple<tensor<2xf32>, tuple<tensor<i32>>>) -> tensor<2xf32>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.if (mhlo::IfOp)

If operation

Produces the output from executing exactly one branch from true_branch or false_branch depending on the value of pred .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#if

Example: %result = "mhlo.if"(%pred) ({ "mhlo.return"(%result_true_branch) : (tensor ) -> () }, { "mhlo.return"(%result_false_branch) : (tensor ) -> () }) : (tensor ) -> tensor

Traits: RecursiveMemoryEffects , SingleBlockImplicitTerminator<ReturnOp> , SingleBlock

Interfaces: InferTypeOpInterface

Operands:

Полученные результаты:

mhlo.imag (mhlo::ImagOp)

Imag operation

Синтаксис:

operation ::= `mhlo.imag` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Extracts the imaginary part, element-wise, from the operand and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#imag

Пример:

%result = mhlo.imag %operand : (tensor<2xcomplex<f32>>) -> tensor<2xf32>

Traits: AlwaysSpeculatableImplTrait , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.infeed (mhlo::InfeedOp)

Infeed operation

Reads data from the infeed and produces results .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#infeed

Пример:

%results:2 = "mhlo.infeed"(%token) {
  infeed_config = ""
} : (!mhlo.token) -> (tensor<3x3x3xi32>, !mhlo.token)

Атрибуты:

Operands:

Полученные результаты:

mhlo.iota (mhlo::IotaOp)

Iota operation

Fills an output tensor with values in increasing order starting from zero along the iota_dimension dimension.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#iota

Пример:

%output = mhlo.iota dim = 0 : tensor<4x5xi32>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Полученные результаты:

mhlo.is_finite (mhlo::IsFiniteOp)

IsFinite operation

Синтаксис:

operation ::= `mhlo.is_finite` $x attr-dict `:` functional-type(operands, results)

Performs element-wise check whether the value in x is finite (ie is neither +Inf, -Inf, nor NaN) and produces a y tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#is_finite

Пример:

%y = mhlo.is_finite %x : (tensor<7xf32>) -> tensor<7xi1>

Traits: AlwaysSpeculatableImplTrait , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.log (mhlo::LogOp)

Log operation

Синтаксис:

operation ::= `mhlo.log` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise logarithm operation on operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#log

Пример:

%result = mhlo.log %operand : tensor<2x2xf64>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.log_plus_one (mhlo::Log1pOp)

Log1p operation

Синтаксис:

operation ::= `mhlo.log_plus_one` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise logarithm plus one operation on operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#log_plus_one

Пример:

%result = mhlo.log_plus_one %operand : tensor<6xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.logistic (mhlo::LogisticOp)

Logistic operation

Синтаксис:

operation ::= `mhlo.logistic` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise logistic operation on operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#logistic

Пример:

%result = mhlo.logistic %operand : tensor<2x2xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.map (mhlo::MapOp)

Map operation

Applies a map function computation to inputs along the dimensions and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#map

Пример:

%result = "mhlo.map"(%input0, %input1) ({
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>):
    %0 = mhlo.multiply %arg0, %arg1 : tensor<i32>
    mhlo.return %0 : tensor<i32>
}) {
  dimensions = dense<[0, 1]> : tensor<2xi64>
} : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>

Traits: InferTensorType , RecursiveMemoryEffects , SameOperandsAndResultShape , SingleBlockImplicitTerminator<ReturnOp> , SingleBlock

Interfaces: InferShapedTypeOpInterface , InferTypeOpInterface

Атрибуты:

Operands:

Полученные результаты:

mhlo.maximum (mhlo::MaxOp)

Max operation

Синтаксис:

operation ::= `mhlo.maximum` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Performs element-wise max operation on tensors lhs and rhs and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#maximum

Пример:

%result = mhlo.maximum %lhs, %rhs : tensor<4xf32>

Traits: AlwaysSpeculatableImplTrait , Commutative , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.minimum (mhlo::MinOp)

Min operation

Синтаксис:

operation ::= `mhlo.minimum` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Performs element-wise min operation on tensors lhs and rhs and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#minimum

Пример:

%result = mhlo.minimum %lhs, %rhs : tensor<4xf32>

Traits: AlwaysSpeculatableImplTrait , Commutative , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.multiply (mhlo::MulOp)

Mul operation

Синтаксис:

operation ::= `mhlo.multiply` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Performs element-wise product of two tensors lhs and rhs and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#multiply

Пример:

%result = mhlo.multiply %lhs, %rhs : tensor<2xi32>

Traits: AlwaysSpeculatableImplTrait , Commutative , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.negate (mhlo::NegOp)

Neg operation

Синтаксис:

operation ::= `mhlo.negate` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise negation of operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#negate

Пример:

%result = mhlo.negate %operand : tensor<2x3xi32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.not (mhlo::NotOp)

Not operation

Синтаксис:

operation ::= `mhlo.not` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise NOT of tensor operand of type integer and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#not

Пример:

%result = mhlo.not %operand : tensor<5x3x1xi1>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.optimization_barrier (mhlo::OptimizationBarrierOp)

OptimizationBarrier operation

Синтаксис:

operation ::= `mhlo.optimization_barrier` attr-dict ($operand^ `:` custom<PairwiseOpType>(type($operand), type($result))):(`(` `)`)?

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, ie result = operand .

See https://github.com/openxla/stablehlo/blob/main/docs/spec.md#optimization_barrier

Пример:

%result0, %result1 = mhlo.optimization_barrier %operand0, %operand1 : tensor<f32>, tensor<f32>

Traits: AlwaysSpeculatableImplTrait , HLO_PairwiseSameOperandAndResultType

Interfaces: ConditionallySpeculatable , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.or (mhlo::OrOp)

Or operation

Синтаксис:

operation ::= `mhlo.or` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Performs element-wise OR of two tensors lhs and rhs and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#or

Пример:

%result = mhlo.or %lhs, %rhs : tensor<2xi1>

Traits: AlwaysSpeculatableImplTrait , Commutative , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.outfeed (mhlo::OutfeedOp)

Outfeed operation

Writes inputs to the outfeed and produces a result token.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#outfeed

Пример:

%result = "mhlo.outfeed"(%input0, %token) {
  outfeed_config = ""
} : (tensor<3x3x3xi32>, !mhlo.token) -> !mhlo.token

Interfaces: InferTypeOpInterface

Атрибуты:

Operands:

Полученные результаты:

mhlo.pad (mhlo::PadOp)

Pad operation

Expands operand by padding around the tensor as well as between the elements of the tensor with the given padding_value .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#pad

Пример:

%0 = mhlo.pad %arg0, %arg1, low = [0, 1], high = [2, 1], interior = [1, 2]
  : (tensor<2x3xi32>, tensor<i32>) -> tensor<5x9xi32>

Traits: AlwaysSpeculatableImplTrait , SameOperandsAndResultElementType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.partition_id (mhlo::PartitionIdOp)

PartitionId operation

Синтаксис:

operation ::= `mhlo.partition_id` attr-dict `:` type(results)

Produces partition_id of the current process.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#partition_id

Пример:

%result = mhlo.partition_id : tensor<ui32>

Interfaces: InferTypeOpInterface

Полученные результаты:

mhlo.popcnt (mhlo::PopulationCountOp)

PopulationCount operation

Синтаксис:

operation ::= `mhlo.popcnt` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise count of the number of bits set in the operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#popcnt

Пример:

%result = mhlo.popcnt %operand : tensor<4xi8>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.power (mhlo::PowOp)

Pow operation

Синтаксис:

operation ::= `mhlo.power` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Performs element-wise exponentiation of lhs tensor by rhs tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#power

Пример:

%result = mhlo.power %lhs, %rhs : tensor<6xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.real (mhlo::RealOp)

Real operation

Синтаксис:

operation ::= `mhlo.real` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Extracts the real part, element-wise, from the operand and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#real

Пример:

%result = mhlo.real %operand : (tensor<2xcomplex<f32>>) -> tensor<2xf32>

Traits: AlwaysSpeculatableImplTrait , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.real_dynamic_slice (mhlo::RealDynamicSliceOp)

RealDynamicSlice operation

Синтаксис:

operation ::= `mhlo.real_dynamic_slice` operands attr-dict `:` functional-type(operands, results)

This operation is a work in progress, so it is not yet included in the specification: https://github.com/openxla/stablehlo/issues/8

Informally, this operation does the same thing as SliceOp except that start_indices , limit_indices and strides are specified dynamically: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#slice

Пример:

%result = mhlo.real_dynamic_slice %operand,
            %start_indices, %limit_indices, %strides
       : (tensor<256x?xf32>, tensor<2xindex>, tensor<2xindex>, tensor<2xindex>) -> tensor<256x?xf32>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.recv (mhlo::RecvOp)

Recv operation

Receives data from a channel with channel_id and produces results .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#recv

Пример:

%results:2 = "mhlo.recv"(%token) {
  // channel_id = 5 : i64,
  // channel_type = #stablehlo<channel_type HOST_TO_DEVICE>,
  channel_handle = #mhlo.channel_handle<handle = 5, type = 3>,
  is_host_transfer = true
} : (!mhlo.token) -> (tensor<3x4xi32>, !mhlo.token)

Атрибуты:

Operands:

Полученные результаты:

mhlo.reduce (mhlo::ReduceOp)

Reduce operation

Applies a reduction function body to inputs and init_values along the dimensions and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#reduce

Пример:

%result = "mhlo.reduce"(%input, %init_value) ({
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>):
    %0 = "mhlo.add"(%arg0, %arg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
    "mhlo.return"(%0) : (tensor<i32>) -> ()
}) {
  dimensions = dense<1> : tensor<1xi64>
} : (tensor<1x6xi32>, tensor<i32>) -> tensor<1xi32>

Traits: InferTensorType , RecursiveMemoryEffects , SameVariadicOperandSize , SingleBlockImplicitTerminator<ReturnOp> , SingleBlock

Interfaces: InferShapedTypeOpInterface , InferTypeOpInterface

Атрибуты:

Operands:

Полученные результаты:

mhlo.reduce_precision (mhlo::ReducePrecisionOp)

ReducePrecision operation

Синтаксис:

operation ::= `mhlo.reduce_precision` $operand `,` `format` `=` custom<ExponentMantissa>($exponent_bits, $mantissa_bits)
              attr-dict `:` custom<SameOperandsAndResultType>(type($operand), type($output))

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.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#reduce_precision

Пример:

%output = mhlo.reduce_precision %operand, format = e5m2 : tensor<6xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.reduce_scatter (mhlo::ReduceScatterOp)

ReduceScatter operation

Within each process group in the 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 .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#reduce_scatter

Пример:

%result = "mhlo.reduce_scatter"(%operand) ({
  ^bb0(%arg0: tensor<f32>, %arg1: tensor<f32>):
  %0 = mhlo.add %arg0, %arg1 : tensor<f32>
  mhlo.return %0 : tensor<f32>
}) {
  scatter_dimension = 1 : i64,
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>,
  // channel_id = 0
  channel_handle = #mhlo.channel_handle<handle = 0, type = 0>
  // use_global_device_ids = false
} : (tensor<2x4xf32>) -> tensor<2x2xf32>

Атрибуты:

Operands:

Полученные результаты:

mhlo.reduce_window (mhlo::ReduceWindowOp)

ReduceWindow operation

Applies a reduction function body to windows of inputs and init_values and produces results .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#reduce_window

Пример:

%result = "mhlo.reduce_window"(%input, %init_value) ({
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>):
    %0 = mhlo.add %arg0, %arg1 : tensor<i32>
    mhlo.return %0 : tensor<i32>
}) {
  window_dimensions = dense<[2, 1]> : tensor<2xi64>,
  window_strides = dense<[4, 1]> : tensor<2xi64>,
  base_dilations = dense<[2, 1]> : tensor<2xi64>,
  window_dilations = dense<[3, 1]> : tensor<2xi64>,
  padding = dense<[[2, 1], [0, 0]]> : tensor<2x2xi64>
} : (tensor<3x2xi32>, tensor<i32>) -> tensor<2x2xi32>

Traits: InferTensorType , RecursiveMemoryEffects , SameVariadicOperandSize , SingleBlockImplicitTerminator<ReturnOp> , SingleBlock

Interfaces: InferShapedTypeOpInterface , InferTypeOpInterface

Атрибуты:

Operands:

Полученные результаты:

mhlo.remainder (mhlo::RemOp)

Rem operation

Синтаксис:

operation ::= `mhlo.remainder` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Performs element-wise remainder of dividend lhs and divisor rhs tensors and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#remainder

Пример:

%result = mhlo.remainder %lhs, %rhs : tensor<4xi64>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.replica_id (mhlo::ReplicaIdOp)

ReplicaId operation

Синтаксис:

operation ::= `mhlo.replica_id` attr-dict `:` type(results)

Produces replica_id of the current process.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#replica_id

Пример:

%result = mhlo.replica_id : tensor<ui32>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Полученные результаты:

mhlo.reshape (mhlo::ReshapeOp)

Reshape operation

Синтаксис:

operation ::= `mhlo.reshape` operands attr-dict `:` functional-type(operands, results)

Performs reshape of operand tensor to a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#reshape

Пример:

%result = mhlo.reshape %operand : (tensor<2xf32>) -> tensor<1x2xf32>

Traits: AlwaysSpeculatableImplTrait , HLO_CompatibleOperandsAndResultElementType

Interfaces: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.return (mhlo::ReturnOp)

_This operation is a work in progress, so it is not yet included in the specification: https://github.com/openxla/stablehlo/issues/425

Informally, this operation serves as a terminator for regions defined by
the StableHLO ops. Non-StableHLO ops, e.g. `func.func`, have their own
terminators, e.g. `func.return`.

Example:

    ```mlir
    %result = "mhlo.reduce"(%input, %init_value) ({
      ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>):
        %0 = "mhlo.add"(%arg0, %arg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
        "mhlo.return"(%0) : (tensor<i32>) -> ()
    }) {
      dimensions = dense<1> : tensor<1xi64>
    } : (tensor<1x6xi32>, tensor<i32>) -> tensor<1xi32>
    ```_


Syntax:

```

operation ::= mhlo.return $results attr-dict ( : type($results)^)?

Traits: `AlwaysSpeculatableImplTrait`, `Terminator`

Interfaces: `ConditionallySpeculatable`, `NoMemoryEffect (MemoryEffectOpInterface)`

Effects: `MemoryEffects::Effect{}`

#### Operands:

| Operand | Description |
| :-----: | ----------- |
| `results` | variadic of ranked tensor of f8E4M3B11FNUZ type or f8E4M3FN type or f8E4M3FNUZ type or f8E5M2 type or f8E5M2FNUZ type or 16-bit float or 32-bit float or 64-bit float or bfloat16 type or pred (AKA boolean or 1-bit integer) or 4/8/16/32/64-bit signless integer or 4/8/16/32/64-bit unsigned integer or complex type with 32-bit float or 64-bit float elements or 4/8/16/32-bit uniform quantized signed integer or 4/8/16/32-bit uniform quantized unsigned integer or 4/8/16/32-bit uniform quantized per axis signed integer or 4/8/16/32-bit uniform quantized per axis unsigned integer values or token or nested tuple with any combination of ranked tensor of f8E4M3B11FNUZ type or f8E4M3FN type or f8E4M3FNUZ type or f8E5M2 type or f8E5M2FNUZ type or 16-bit float or 32-bit float or 64-bit float or bfloat16 type or pred (AKA boolean or 1-bit integer) or 4/8/16/32/64-bit signless integer or 4/8/16/32/64-bit unsigned integer or complex type with 32-bit float or 64-bit float elements or 4/8/16/32-bit uniform quantized signed integer or 4/8/16/32-bit uniform quantized unsigned integer values or ranked tensor of 4/8/16/32-bit uniform quantized per axis signed integer or 4/8/16/32-bit uniform quantized per axis unsigned integer values or token values


### `mhlo.reverse` (mhlo::ReverseOp)

_Reverse operation_

Reverses the order of elements in the `operand` along the specified
`dimensions` and produces a `result` tensor.

See:
<a href="https://github.com/openxla/stablehlo/blob/main/docs/spec.md#reverse">https://github.com/openxla/stablehlo/blob/main/docs/spec.md#reverse</a>

Example:
```mlir
%result = mhlo.reverse %operand, dims = [1] : tensor<3x2xi32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.rng (mhlo::RngOp)

Rng operation

Generates random numbers using the rng_distribution algorithm and produces a result tensor of a given shape shape .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#rng

Пример:

%result = mhlo.rng %a, %b, %shape, distribution = NORMAL : (tensor<i32>, tensor<i32>, tensor<2xi64>) -> tensor<3x3xi32>

Traits: InferTensorType

Interfaces: InferShapedTypeOpInterface , InferTypeOpInterface

Атрибуты:

Operands:

Полученные результаты:

mhlo.rng_bit_generator (mhlo::RngBitGeneratorOp)

RngBitGenerator operation

Returns an output filled with uniform random data and an updated output state output_state given an initial state initial_state using the pseudorandom number generator algorithm rng_algorithm .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#rng_bit_generator

Пример:

%output_state, %output = mhlo.rng_bit_generator %initial_state, algorithm = THREE_FRY : (tensor<2xui64>) -> (tensor<2xui64>, tensor<2x2xui64>)

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.round_nearest_afz (mhlo::RoundOp)

Round operation

Синтаксис:

operation ::= `mhlo.round_nearest_afz` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise rounding towards the nearest integer, breaking ties away from zero, on the operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#round_nearest_afz

Пример:

%result = mhlo.round_nearest_afz %operand : tensor<5xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.round_nearest_even (mhlo::RoundNearestEvenOp)

RoundNearestEven operation

Синтаксис:

operation ::= `mhlo.round_nearest_even` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise rounding towards the nearest integer, breaking ties towards the even integer, on the operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#round_nearest_even

Пример:

%result = mhlo.round_nearest_even %operand : tensor<5xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.rsqrt (mhlo::RsqrtOp)

Rsqrt operation

Синтаксис:

operation ::= `mhlo.rsqrt` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise reciprocal square root operation on operand tensor and produces a result tensor, implementing the rSqrt operation from the IEEE-754 specification.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#rsqrt

Пример:

%result = mhlo.rsqrt %operand : tensor<2x2xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.scatter (mhlo::ScatterOp)

Scatter operation

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 .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#scatter

Пример:

%result = "mhlo.scatter"(%input, %scatter_indices, %update) ({
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>):
    %0 = mhlo.add %arg0, %arg1 : tensor<i32>
    mhlo.return %0 : tensor<i32>
}) {
  scatter_dimension_numbers = #mhlo.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<3x4x2xi32>, tensor<2x3x2xi64>, tensor<2x3x2x2xi32>) -> tensor<3x4x2xi32>

Traits: RecursiveMemoryEffects , SameVariadicOperandSize

Interfaces: InferTypeOpInterface

Атрибуты:

Operands:

Полученные результаты:

mhlo.select (mhlo::SelectOp)

Select operation

Синтаксис:

operation ::= `mhlo.select` operands attr-dict `:`
              custom<SelectOpType>(type($pred), type($on_true), type($on_false), type($result))

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 .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#select

Пример:

%result = mhlo.select %pred, %on_true, %on_false : tensor<2x2xi1>, tensor<2x2xi32>

Traits: AlwaysSpeculatableImplTrait , HLO_BroadcastingElementwise , InferTensorType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.select_and_scatter (mhlo::SelectAndScatterOp)

SelectAndScatter operation

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.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#select_and_scatter

Пример:

%result = "mhlo.select_and_scatter"(%operand, %source, %init_value) ({
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>):
    %0 = "mhlo.compare"(%arg0, %arg1) {
      comparison_direction = #stablehlo<comparison_direction GE>
    } : (tensor<i32>, tensor<i32>) -> tensor<i1>
    "mhlo.return"(%0) : (tensor<i1>) -> ()
}, {
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>):
    %0 = "mhlo.add"(%arg0, %arg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
    "mhlo.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>

Traits: RecursiveMemoryEffects

Interfaces: InferTypeOpInterface

Атрибуты:

Operands:

Полученные результаты:

mhlo.send (mhlo::SendOp)

Send operation

Sends inputs to a channel channel_id and produces a result token.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#send

Пример:

%result = "mhlo.send"(%operand, %token) {
  // channel_id = 5 : i64,
  // channel_type = #stablehlo<channel_type DEVICE_TO_HOST>,
  channel_handle = #mhlo.channel_handle<handle = 5, type = 2>,
  is_host_transfer = true
} : (tensor<3x4xi32>, !mhlo.token) -> !mhlo.token

Interfaces: InferTypeOpInterface

Атрибуты:

Operands:

Полученные результаты:

mhlo.set_dimension_size (mhlo::SetDimensionSizeOp)

SetDimensionSize operation

This operation is a work in progress, so it is not yet included in the specification: https://github.com/openxla/stablehlo/issues/8

Informally, this operation does the same thing as XLA's SetDimensionSize: https://www.tensorflow.org/xla/operation_semantics#setdimensionsize

Пример:

%0 = mhlo.set_dimension_size %arg0, %arg1, dim = 1 : (tensor<4x2xf32>, tensor<i32>) -> tensor<4x2xf32>

Traits: AlwaysSpeculatableImplTrait , InferTensorType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.shift_left (mhlo::ShiftLeftOp)

ShiftLeft operation

Синтаксис:

operation ::= `mhlo.shift_left` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Performs element-wise left-shift operation on the lhs tensor by rhs number of bits and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#shift_left

Пример:

%result = mhlo.shift_left %lhs, %rhs : tensor<6xi8>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.shift_right_arithmetic (mhlo::ShiftRightArithmeticOp)

ShiftRightArithmetic operation

Синтаксис:

operation ::= `mhlo.shift_right_arithmetic` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Performs element-wise arithmetic right-shift operation on the lhs tensor by rhs number of bits and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#shift_right_arithmetic

Пример:

%result = mhlo.shift_right_arithmetic %lhs, %rhs : tensor<6xi8>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.shift_right_logical (mhlo::ShiftRightLogicalOp)

ShiftRightLogical operation

Синтаксис:

operation ::= `mhlo.shift_right_logical` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Performs element-wise logical right-shift operation on the lhs tensor by rhs number of bits and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#shift_right_logical

Пример:

%result = mhlo.shift_right_logical %lhs, %rhs : tensor<6xi8>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.sign (mhlo::SignOp)

Sign operation

Синтаксис:

operation ::= `mhlo.sign` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Returns the sign of the operand element-wise and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#sign

Пример:

%result = mhlo.sign %operand : tensor<7xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.sine (mhlo::SineOp)

Sine operation

Синтаксис:

operation ::= `mhlo.sine` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise sine operation on operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#sine

Пример:

%result = mhlo.sine %operand : tensor<2xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.slice (mhlo::SliceOp)

Slice operation

Extracts a slice from the operand using statically-computed starting indices and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#slice

Пример:

%result = "mhlo.slice" (%operand) {
  start_indices = dense<[1, 2]> : tensor<2xi64>,
  limit_indices = dense<[3, 4]> : tensor<2xi64>,
  strides = dense<1> : tensor<2xi64>
} : (tensor<3x4xi64>) -> tensor<2x2xi64>

Traits: AlwaysSpeculatableImplTrait , SameOperandsAndResultElementType

Interfaces: ConditionallySpeculatable , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.sort (mhlo::SortOp)

Sort operation

Sorts a variadic number of tensors in inputs together, according to a custom comparator , along the given dimension and produces a variadic number of tensors as results .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#sort

Пример:

%result0, %result1 = "mhlo.sort"(%input0, %input1) ({
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>, %arg2: tensor<i32>, %arg3: tensor<i32>):
    %predicate = "mhlo.compare"(%arg0, %arg1) {
      comparison_direction = #stablehlo<comparison_direction GT>
      } : (tensor<i32>, tensor<i32>) -> tensor<i1>
    "mhlo.return"(%predicate) : (tensor<i1>) -> ()
}) {
  dimension = 0 : i64,
  is_stable = true
} : (tensor<2x3xi32>, tensor<2x3xi32>) -> (tensor<2x3xi32>, tensor<2x3xi32>)

Traits: InferTensorType , RecursiveMemoryEffects , SameOperandsAndResultShape

Interfaces: InferShapedTypeOpInterface , InferTypeOpInterface

Атрибуты:

Operands:

Полученные результаты:

mhlo.sparse_dot (mhlo::SparseDotOp)

Sparse dot operation

Similar to dot_general operation, with one or both of the operands being sparse. An additional argument provides sparsity meta information. Disclaimer: this op is experimental / a work in progress.

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.sqrt (mhlo::SqrtOp)

Sqrt operation

Синтаксис:

operation ::= `mhlo.sqrt` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise square root operation on operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#sqrt

Пример:

%result = mhlo.sqrt %operand : tensor<2x2xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.stochastic_convert (mhlo::StochasticConvertOp)

StochasticConvert operation

This operation is a work in progress, so it is not yet included in the specification: https://github.com/openxla/stablehlo/issues/295

Informally, this operation performs element-wise conversion of values from a bigger type to a smaller one with stochastic rounding using the random number passed in.

Traits: AlwaysSpeculatableImplTrait , Elementwise

Interfaces: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.subtract (mhlo::SubtractOp)

Subtract operation

Синтаксис:

operation ::= `mhlo.subtract` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Performs element-wise subtraction of two tensors lhs and rhs and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#subtract

Пример:

%result = mhlo.subtract %lhs, %rhs : tensor<2xi32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.tan (mhlo::TanOp)

Tan operation

Синтаксис:

operation ::= `mhlo.tan` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

This operation is a work in progress, so it is not yet included in the specification: https://github.com/openxla/stablehlo/issues/954

Informally, this operation returns Tan(operand) element-wise.

Пример:

%0 = mhlo.tan %arg0 : tensor<2xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.tanh (mhlo::TanhOp)

Tanh operation

Синтаксис:

operation ::= `mhlo.tanh` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise hyperbolic tangent operation on operand tensor and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#tanh

Пример:

%result = mhlo.tanh %operand : tensor<2xf32>

Traits: AlwaysSpeculatableImplTrait , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.topk (mhlo::TopKOp)

TopK operation

Синтаксис:

operation ::= `mhlo.topk` `(`$operand `,` `k` `=` $k (`,` `largest` `=` $largest^)? `)` attr-dict `:`
              type($operand) `->` `(`type($values)`,` type($indices)`)`

Returns top k values and their indices, along the last dimension of the operand if largest=true or the bottom k values if largest=false .

See: https://www.tensorflow.org/xla/operation_semantics#top-k

Пример:

%values, %indices = mhlo.topk(%operand, k=5, largest=true)
  : tensor<100xf32> -> (tensor<5xf32>, tensor<5xi32>)

Traits: InferTensorType , RecursiveMemoryEffects

Interfaces: InferShapedTypeOpInterface , InferTypeOpInterface

Атрибуты:

Operands:

Полученные результаты:

mhlo.torch_index_select (mhlo::TorchIndexSelectOp)

TorchIndexSelect operation

This operation is on its way out of StableHLO, so it is not included in the specification: https://github.com/openxla/stablehlo/issues/3

Informally, this operation does the same thing as PyTorch's index_select, augmented with support for batch dimensions: https://pytorch.org/docs/stable/generated/torch.index_select.html

The batch_dims attribute specifies the number of major batch dimensions (0 or more) that act like a multidimensional loop over both the operand and the index.

Пример:

%result = "mhlo.torch_index_select"(%operand, %index) {
  dim = 2 : i64,
  batch_dims = 1 : i64
} : (tensor<8x128x3072x64xf32>, tensor<8x16x1024xi32>) -> tensor<8x128x16x1024x64xf32>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.trace (mhlo::TraceOp)

Trace operation

Синтаксис:

operation ::= `mhlo.trace` $operand `,` $tag attr-dict `:` type($operand)

This operation is on its way out of StableHLO, so it is not included in the specification: https://github.com/openxla/stablehlo/issues/604

It is not used by JAX, PyTorch or TensorFlow, so it looks like we should've classified it as "Private to XLA" and not included it in StableHLO in the first place. With that in mind, its semantics will not be documented here.

Пример:

mhlo.trace %arg0, "In test code." : tensor<5x1x5xi32>

Атрибуты:

Operands:

mhlo.transpose (mhlo::TransposeOp)

Transpose operation

Permutes the dimensions of operand tensor using permutation and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#transpose

Пример:

%0 = mhlo.transpose %arg0, dims = [2, 1, 0] : (tensor<1x2x3xi32>) -> tensor<3x2x1xi32>

Traits: AlwaysSpeculatableImplTrait , HLO_CompatibleOperandsAndResultElementType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.triangular_solve (mhlo::TriangularSolveOp)

TriangularSolve operation

Solves batches of systems of linear equations with lower or upper triangular coefficient matrices.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#triangular_solve

Пример:

%result = "mhlo.triangular_solve"(%a, %b) {
  left_side = true,
  lower = true,
  unit_diagonal = false,
  transpose_a = #stablehlo<transpose NO_TRANSPOSE>
} : (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>

Traits: AlwaysSpeculatableImplTrait , InferTensorType , SameOperandsAndResultElementType

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.tuple (mhlo::TupleOp)

Tuple operation

Синтаксис:

operation ::= `mhlo.tuple` $val attr-dict `:` custom<TupleOpType>(type($val), type($result))

Produces a result tuple from values val .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#tuple

Пример:

%result = mhlo.tuple %val0, %val1 : tuple<tensor<2xf32>, tuple<tensor<i32>>>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.unary_einsum (mhlo::UnaryEinsumOp)

UnaryEinsum operation

This operation is on its way out of StableHLO, so it is not included in the specification: https://github.com/openxla/stablehlo/issues/3

Informally, this operation does the same thing as TF's einsum: https://www.tensorflow.org/api_docs/python/tf/einsum

Пример:

%result = "mhlo.unary_einsum"(%operand) {
  einsum_config = "ab->a"
} : (tensor<4x16xf32>) -> tensor<4xf32>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Атрибуты:

Operands:

Полученные результаты:

mhlo.uniform_dequantize (mhlo::UniformDequantizeOp)

UniformDequantize operation

Синтаксис:

operation ::= `mhlo.uniform_dequantize` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

Performs element-wise conversion of quantized tensor operand to a floating-point tensor result according to the quantization parameters defined by the operand type.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#uniform_dequantize

Пример:

%result = mhlo.uniform_dequantize %operand : (tensor<16x16x!quant.uniform<i8:f32, 34.0:16>>) -> tensor<16x16xf32>

Traits: AlwaysSpeculatableImplTrait , Elementwise , InferTensorType , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.uniform_quantize (mhlo::UniformQuantizeOp)

UniformQuantize operation

Синтаксис:

operation ::= `mhlo.uniform_quantize` $operand attr-dict
              `:` custom<SameOperandsAndResultType>(type($operand), type($result))

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.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#uniform_quantize

Пример:

%result = mhlo.uniform_quantize %operand : (tensor<16x16xf32>) -> tensor<16x16x!quant.uniform<ui8:f32, 34.0:16>>

Traits: AlwaysSpeculatableImplTrait , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

mhlo.while (mhlo::WhileOp)

While operation

Produces the output from executing body function 0 or more times while the cond function outputs true .

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#while

Пример:

%results0, %results1 = "mhlo.while"(%operand0, %operand1) ({
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>):
    %0 = "mhlo.compare"(%arg0, %arg1) {
      comparison_direction = #stablehlo<comparison_direction LT>
    } : (tensor<i32>, tensor<i32>) -> tensor<i1>
    "mhlo.return"(%0) : (tensor<i1>) -> ()
}, {
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>):
    %0 = "mhlo.add"(%arg0, %constant0) : (tensor<i32>, tensor<i32>) -> tensor<i32>
    "mhlo.return"(%0, %arg1) : (tensor<i32>, tensor<i32>) -> ()
}) : (tensor<i32>, tensor<i32>) -> (tensor<i32>, tensor<i32>)

Traits: RecursiveMemoryEffects , SingleBlockImplicitTerminator<ReturnOp> , SingleBlock

Interfaces: InferTypeOpInterface , OpAsmOpInterface

Operands:

Полученные результаты:

mhlo.xla.rng_get_and_update_state (mhlo::XlaRngGetAndUpdateStateOp)

XlaRngGetAndUpdateState operation

Синтаксис:

operation ::= `mhlo.xla.rng_get_and_update_state` attr-dict

This operation is private to the XLA compiler, so it is does not yet have a specification.

Informally, this operation represents the change of the global random number generator state for rng instructions. The global state is incremented by delta and the old state is returned.

The output is currently defined for a single output type. If this changes in the future to support multiple types, lowering to use of a global memref must ensure that a single memref is still used and updated appropriately.

Interfaces: InferTypeOpInterface

Атрибуты:

Полученные результаты:

mhlo.xor (mhlo::XorOp)

Xor operation

Синтаксис:

operation ::= `mhlo.xor` $lhs `,` $rhs attr-dict
              `:` custom<SameOperandsAndResultType>(type($lhs), type($rhs), type($result))

Performs element-wise XOR of two tensors lhs and rhs and produces a result tensor.

See: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#xor

Пример:

%result = mhlo.xor %lhs, %rhs : tensor<2xi32>

Traits: AlwaysSpeculatableImplTrait , Commutative , CompatibleOperandsAndResultType , Elementwise , SameOperandsAndResultShape

Interfaces: ConditionallySpeculatable , InferShapedTypeOpInterface , InferTypeOpInterface , NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Полученные результаты:

Атрибуты

ArgResultAliasAttr

Attribute that models the alias relationship of entry function argument

This attribute captures the alias relationship of an MHLO main function argument to one of the results, denoted by resultIndex . The argTupleIndices and resultTupleIndices are used to index into nested tuples in operand and result respectively. If isMustAlias is true then the operand-result pair must alias.

This is meant to be used as an attribute on a function argument in MHLO. For example, in the following code it expresses that %arg1 may alias 0-th result.

func @main(%arg0: tensor<2xf32>, %arg1: tensor<3xf32> {mhlo.result_alias =
    mhlo.result_alias<result_index = [2], ...>}
  ) -> tensor<2xf32>, tensor<3xf32> {
  // function body ...
}

Параметры:

ChannelHandleAttr

two 64-bit integers 'handle' and 'type'

Синтаксис:

#mhlo.channel_handle<
  int64_t,   # handle
  int64_t   # type
>

Параметры:

ComparisonDirectionAttr

Which comparison operation to perform.

Синтаксис:

#mhlo.comparison_direction<
  ::mlir::mhlo::ComparisonDirection   # value
>

Enum cases:

• EQ ( EQ )
• NE ( NE )
• GE ( GE )
• GT ( GT )
• LE ( LE )
• LT ( LT ) #### Parameters:

ComparisonTypeAttr

Which comparison type to use.

Синтаксис:

#mhlo.comparison_type<
  ::mlir::mhlo::ComparisonType   # value
>

Enum cases:

• NOTYPE ( NOTYPE )
• FLOAT ( FLOAT )
• TOTALORDER ( TOTALORDER )
• SIGNED ( SIGNED )
• UNSIGNED ( UNSIGNED ) #### Parameters:

ConvDimensionNumbersAttr

Structure of dimension information for conv op

Параметры:

CrossProgramPrefetchAttr

Argument that is prefetched from another program

Синтаксис:

#mhlo.cross_program_prefetch<
  int64_t,   # parameter
  ::llvm::ArrayRef<int64_t>,   # indices
  std::optional<int64_t>   # offset
>

This attribute captures an argument that is prefetched from another program. For a given CrossProgramPrefetchAttr , parameter tells us which argument of the main function of the module is prefetched, and indices is a shape index telling us what subshape of that argument is prefetched.

A shape has a subshape iff it is a tuple. In that case, the subshape of the tuple by indices is the shape achieved after indexing by each element of indices in turn. For example, the [1,0] subshape of tuple<tuple<token, token>, tuple<tensor<i32>, token>> is tensor<i32> .

An empty value for indices means the whole shape is prefetched.

Например,

module attributes { mhlo.cross_program_prefetch = [ #mhlo.cross_program_prefetch< parameter = 0, indices = [0]> ]} {
  func.func @copy(%arg0 : tuple<tensor<2x3xi32>, tensor<i32>>) -> tuple<tensor<2x3xi32>, tensor<i32>> {
    %0 = "mhlo.copy"(%arg0) {is_cross_program_prefetch}
    return %0 : tuple<tensor<2x3xi32>, tensor<i32>>
  }
  func.func @main(%arg0 : tuple<tensor<2x3xi32>, tensor<i32>>) -> tuple<tensor<2x3xi32>, tensor<i32>> {
    %1 = "mhlo.async_start"(%arg0) {called_computation=@copy}
    %2 = "mhlo.async_done"(%1) {called_computation=@copy}
    return %2 : tuple<tensor<2x3xi32>, tensor<i32>>
  }
}

The parameter = 0 tells us that the async copy of the 0 th parameter is a cross_program_prefetch , while the index of [0] tells us that the 0 th element of the tuple is prefetched while the other element of the tuple is not.

Параметры:

CustomCallScheduleAttr

Specifies the desired schedule for the custom-call.

Синтаксис:

#mhlo.custom_call_schedule<
  ::mlir::mhlo::CustomCallSchedule   # value
>

Enum cases:

• NONE ( NONE )
• LATEST ( LATEST )
• EARLIEST ( EARLIEST ) #### Parameters:

DequantizeModeAttr

Dequantization mode. Only MIN_COMBINED is supported.

Синтаксис:

#mhlo.dequantize_mode<
  ::mlir::mhlo::DequantizeMode   # value
>

Enum cases:

• MIN_COMBINED ( MIN_COMBINED ) #### Parameters:

DomainKindAttr

Kind of domain metatdata attached to an HLO domain.

Синтаксис:

#mhlo.kind<
  ::mlir::mhlo::DomainKind   # value
>

Enum cases:

• sharding ( sharding ) #### Parameters:

DotDimensionNumbersAttr

Attribute that models the dimension information for dot.

Параметры:

FftTypeAttr

XLA fast fourier transform type.

Синтаксис:

#mhlo.fft_type<
  ::mlir::mhlo::FftType   # value
>

Enum cases:

• FFT ( FFT )
• IFFT ( IFFT )
• RFFT ( RFFT )
• IRFFT ( IRFFT ) #### Parameters:

FusionKindAttr

fusion kind

Синтаксис:

#mhlo.fusion_kind<
  ::mlir::mhlo::FusionKind   # value
>

Enum cases:

• kLoop ( kLoop )
• kInput ( kInput )
• kOutput ( kOutput )
• kCustom ( kCustom ) #### Parameters:

GatherDimensionNumbersAttr

Attribute that models the dimension information for gather

Параметры:

OutputOperandAliasAttr

Attribute that models the alias relationship of output and operand of a CustomCall op

Синтаксис:

#mhlo.output_operand_alias<
  ::llvm::ArrayRef<int64_t>,   # outputTupleIndices
  int64_t,   # operandIndex
  ::llvm::ArrayRef<int64_t>   # operandTupleIndices
>

This attribute captures the alias relationship of the output to one of the operands for a CustomCall op, denoted by operand_index . The output_tuple_indices and operand_tuple_indices are used to index into output and operand types. These indices lists are empty if the corresponding types are not tuple types, and can be arbitrarily long in case of arbitrarily nested tuple types.

See https://www.tensorflow.org/xla/aliasing

Example when used as array with in mhlo.custom-call:

%0 = "mhlo.custom_call"(%arg0, %arg1) {
  // other attributes
  output_operand_alias = [
    #mhlo.output_operand_alias<output_tuple_indices = [0],
                               operand_index = 0,
                               operand_tuple_indices = [1]>
  ]
} : (tuple<tensor<1x1xf32>, tensor<2x3xf32>>, tensor<5x5xf32>) -> tuple<tensor<2x3xf32>>

The output and the 0th operand are both tuples. The aliasing shows the
relationship between the 0th element in output tuple with the 1st element in
the 0th operand. And both of them are of the same type: tensor<2x3xf32>.

Параметры:

PrecisionAttr

XLA precision for an operand. Has backend specific meaning.

Синтаксис:

#mhlo.precision<
  ::mlir::mhlo::Precision   # value
>

Enum cases:

• DEFAULT ( DEFAULT )
• ВЫСОКО HIGH )
• HIGHEST ( HIGHEST )
• PACKED_NIBBLE ( PACKED_NIBBLE ) #### Parameters:

RngAlgorithmAttr

XLA PRNG algorithm to be used.

Синтаксис:

#mhlo.rng_algorithm<
  ::mlir::mhlo::RngAlgorithm   # value
>

Enum cases:

• DEFAULT ( DEFAULT )
• THREE_FRY ( THREE_FRY )
• PHILOX ( PHILOX ) #### Parameters:

RngDistributionAttr

XLA PRNG distribution to be used.

Синтаксис:

#mhlo.rng_distribution<
  ::mlir::mhlo::RngDistribution   # value
>

Enum cases:

• UNIFORM ( UNIFORM )
• NORMAL ( NORMAL ) #### Parameters:

ScatterDimensionNumbersAttr

Attribute that models the dimension information for scatter

Параметры:

SparsityDescriptorAttr

Describes structured (N:M) sparsity configuration

Синтаксис:

#mhlo.sparsity<
  int64_t,   # dimension
  int64_t,   # n
  int64_t   # m
>

This attribute is defined for a sparse dot operation with a structured sparse input tensor. With (N=2,M=4), every 4 consecutive logical elements have exactly 2 non-zero physical elements in the input tensor.

$dimension defines the index of the contracting dimension that is sparse (it has to be the most minor dimension). The additional metadata operand in the sparse dot operation defines which logical elements are zeroed out.

Параметры:

TransposeAttr

Transpose options

Синтаксис:

#mhlo.transpose<
  ::mlir::mhlo::Transpose   # value
>

Enum cases:

• TRANSPOSE_INVALID ( TRANSPOSE_INVALID )
• NO_TRANSPOSE ( NO_TRANSPOSE )
• TRANSPOSE ( TRANSPOSE )
• ADJOINT ( ADJOINT ) #### Parameters:

TypeExtensionsAttr

Attribute that extends tensor type with MHLO type properties.

Синтаксис:

#mhlo.type_extensions<
  ::llvm::ArrayRef<int64_t>   # bounds
>

This attribute is used to extend MLIR tensor type with MHLO tensor specific properties. These properties aren't modeled in the MLIR type. This attribute is set in the encoding field of the tensor type.

See HLO_BoundedAttrInterface for documentation for bounds .

Параметры:

Типы

AsyncBundleType

Opaque collection of other types

Синтаксис:

!mhlo.async_bundle<
  ::llvm::ArrayRef<Type>   # types
>

Параметры: