MLIR

Multi-Level IR Compiler Framework

'scf' Dialect

Operation definition ¶

scf.condition (::mlir::scf::ConditionOp) ¶

loop continuation condition

Syntax:

operation ::= scf.condition ( $condition ) attr-dict ($args^ : type($args))? This operation accepts the continuation (i.e., inverse of exit) condition of the scf.while construct. If its first argument is true, the “after” region of scf.while is executed, with the remaining arguments forwarded to the entry block of the region. Otherwise, the loop terminates. Traits: AlwaysSpeculatableImplTrait, HasParent, Terminator Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface), RegionBranchTerminatorOpInterface Effects: MemoryEffects::Effect{} Operands: ¶ OperandDescription condition1-bit signless integer argsany type scf.execute_region (::mlir::scf::ExecuteRegionOp) ¶ operation that executes its region exactly once The execute_region operation is used to allow multiple blocks within SCF and other operations which can hold only one block. The execute_region operation executes the region held exactly once and cannot have any operands. As such, its region has no arguments. All SSA values that dominate the op can be accessed inside the op. The op’s region can have multiple blocks and the blocks can have multiple distinct terminators. Values returned from this op’s region define the op’s results. Example: scf.for %i = 0 to 128 step %c1 { %y = scf.execute_region -> i32 { %x = load %A[%i] : memref<128xi32> scf.yield %x : i32 } } affine.for %i = 0 to 100 { "foo"() : () -> () %v = scf.execute_region -> i64 { cf.cond_br %cond, ^bb1, ^bb2 ^bb1: %c1 = arith.constant 1 : i64 cf.br ^bb3(%c1 : i64) ^bb2: %c2 = arith.constant 2 : i64 cf.br ^bb3(%c2 : i64) ^bb3(%x : i64): scf.yield %x : i64 } "bar"(%v) : (i64) -> () } Interfaces: RegionBranchOpInterface Results: ¶ ResultDescription «unnamed»any type scf.for (::mlir::scf::ForOp) ¶ for operation The “scf.for” operation represents a loop taking 3 SSA value as operands that represent the lower bound, upper bound and step respectively. The operation defines an SSA value for its induction variable. It has one region capturing the loop body. The induction variable is represented as an argument of this region. This SSA value always has type index, which is the size of the machine word. The step is a value of type index, required to be positive. The lower and upper bounds specify a half-open range: the range includes the lower bound but does not include the upper bound. The body region must contain exactly one block that terminates with “scf.yield”. Calling ForOp::build will create such a region and insert the terminator implicitly if none is defined, so will the parsing even in cases when it is absent from the custom format. For example: scf.for %iv = %lb to %ub step %step { ... // body } scf.for can also operate on loop-carried variables and returns the final values after loop termination. The initial values of the variables are passed as additional SSA operands to the “scf.for” following the 3 loop control SSA values mentioned above (lower bound, upper bound and step). The operation region has an argument for the induction variable, followed by one argument for each loop-carried variable, representing the value of the variable at the current iteration. The region must terminate with a “scf.yield” that passes the current values of all loop-carried variables to the next iteration, or to the “scf.for” result, if at the last iteration. The static type of a loop-carried variable may not change with iterations; its runtime type is allowed to change. Note, that when the loop-carried variables are present, calling ForOp::build will not insert the terminator implicitly. The caller must insert “scf.yield” in that case. “scf.for” results hold the final values after the last iteration. For example, to sum-reduce a memref: func.func @reduce(%buffer: memref<1024xf32>, %lb: index, %ub: index, %step: index) -> (f32) { // Initial sum set to 0. %sum_0 = arith.constant 0.0 : f32 // iter_args binds initial values to the loop's region arguments. %sum = scf.for %iv = %lb to %ub step %step iter_args(%sum_iter = %sum_0) -> (f32) { %t = load %buffer[%iv] : memref<1024xf32> %sum_next = arith.addf %sum_iter, %t : f32 // Yield current iteration sum to next iteration %sum_iter or to %sum // if final iteration. scf.yield %sum_next : f32 } return %sum : f32 } If the “scf.for” defines any values, a yield must be explicitly present. The number and types of the “scf.for” results must match the initial values in the “iter_args” binding and the yield operands. Another example with a nested “scf.if” (see “scf.if” for details) to perform conditional reduction: func.func @conditional_reduce(%buffer: memref<1024xf32>, %lb: index, %ub: index, %step: index) -> (f32) { %sum_0 = arith.constant 0.0 : f32 %c0 = arith.constant 0.0 : f32 %sum = scf.for %iv = %lb to %ub step %step iter_args(%sum_iter = %sum_0) -> (f32) { %t = load %buffer[%iv] : memref<1024xf32> %cond = arith.cmpf "ugt", %t, %c0 : f32 %sum_next = scf.if %cond -> (f32) { %new_sum = arith.addf %sum_iter, %t : f32 scf.yield %new_sum : f32 } else { scf.yield %sum_iter : f32 } scf.yield %sum_next : f32 } return %sum : f32 } Traits: AutomaticAllocationScope, RecursiveMemoryEffects, SingleBlockImplicitTerminatorscf::YieldOp Interfaces: ConditionallySpeculatable, LoopLikeOpInterface, RegionBranchOpInterface Operands: ¶ OperandDescription lowerBoundindex upperBoundindex stepindex initArgsany type Results: ¶ ResultDescription resultsany type scf.foreach_thread (::mlir::scf::ForeachThreadOp) ¶ evaluate a block multiple times in parallel scf.foreach_thread is a target-independent multi-dimensional parallel region application operation. It has exactly one block that represents the parallel body and it takes index operands that indicate how many parallel instances of that function are created. The op also takes a variadic number of tensor operands (shared_outs). The future buffers corresponding to these tensors are shared among all threads. Shared tensors should be accessed via their corresponding block arguments. If multiple threads write to a shared buffer in a racy fashion, these writes will execute in some unspecified order. Tensors that are not shared can be used inside the body (i.e., the op is not isolated from above); however, if a use of such a tensor bufferizes to a memory write, the tensor is privatized, i.e., a thread-local copy of the tensor is used. This ensures that memory side effects of a thread are not visible to other threads (or in the parent body), apart from explicitly shared tensors. The name “thread” conveys the fact that the parallel execution is mapped (i.e. distributed) to a set of virtual threads of execution, one function application per thread. Further lowerings are responsible for specifying how this is materialized on concrete hardware resources. An optional mapping is an attribute array that specifies processing units with their dimension, how it remaps 1-1 to a set of concrete processing element resources (e.g. a CUDA grid dimension or a level of concrete nested async parallelism). It is expressed via any attribute that implements the device mapping interface. It is the reponsibility of the lowering mechanism to interpret the mapping attributes in the context of the concrete target the op is lowered to, or to ignore it when the specification is ill-formed or unsupported for a particular target. The only allowed terminator is scf.foreach_thread.perform_concurrently. scf.foreach_thread returns one value per shared_out operand. The actions of the perform_concurrently terminators specify how to combine the partial results of all parallel invocations into a full value, in some unspecified order. The “destination” of each such op must be a shared_out block argument of the scf.foreach_thread op. The actions involved in constructing the return values are further described by tensor.parallel_insert_slice. scf.foreach_thread acts as an implicit synchronization point. When the parallel function body has side effects, their order is unspecified across threads. Example: // // Sequential context. // %matmul_and_pointwise:2 = scf.foreach_thread (%thread_id_1, %thread_id_2) in (%num_threads_1, %numthread_id_2) shared_outs(%o1 = %C, %o2 = %pointwise) -> (tensor<?x?xT>, tensor<?xT>) { // // Parallel context, each thread with id = (%thread_id_1, %thread_id_2) // runs its version of the code. // %sA = tensor.extract_slice %A[f((%thread_id_1, %thread_id_2))]: tensor<?x?xT> to tensor<?x?xT> %sB = tensor.extract_slice %B[g((%thread_id_1, %thread_id_2))]: tensor<?x?xT> to tensor<?x?xT> %sC = tensor.extract_slice %o1[h((%thread_id_1, %thread_id_2))]: tensor<?x?xT> to tensor<?x?xT> %sD = matmul ins(%sA, %sB) outs(%sC) %spointwise = subtensor %o2[i((%thread_id_1, %thread_id_2))]: tensor<?xT> to tensor<?xT> %sE = add ins(%spointwise) outs(%sD) scf.foreach_thread.perform_concurrently { scf.foreach_thread.parallel_insert_slice %sD into %o1[h((%thread_id_1, %thread_id_2))]: tensor<?x?xT> into tensor<?x?xT> scf.foreach_thread.parallel_insert_slice %spointwise into %o2[i((%thread_id_1, %thread_id_2))]: tensor<?xT> into tensor<?xT> } } // Implicit synchronization point. // Sequential context. // Example with mapping attribute: // // Sequential context. Here mapping is expressed as GPU thread mapping // attributes // %matmul_and_pointwise:2 = scf.foreach_thread (%thread_id_1, %thread_id_2) in (%num_threads_1, %numthread_id_2) shared_outs(...) -> (tensor<?x?xT>, tensor<?xT>) { // // Parallel context, each thread with id = **(%thread_id_2, %thread_id_1)** // runs its version of the code. // scf.foreach_thread.perform_concurrently { ... } } { mapping = [#gpu.thread<y>, #gpu.thread<x>] } // Implicit synchronization point. // Sequential context. // Example with privatized tensors: %t0 = ... %t1 = ... %r = scf.foreach_thread ... shared_outs(%o = t0) -> tensor<?xf32> { // %t0 and %t1 are privatized. %t0 is definitely copied for each thread // because the scf.foreach_thread op's %t0 use bufferizes to a memory // write. In the absence of other conflicts, %t1 is copied only if there // are uses of %t1 in the body that bufferize to a memory read and to a // memory write. "some_use"(%t0) "some_use"(%t1) } Traits: AttrSizedOperandSegments, AutomaticAllocationScope, RecursiveMemoryEffects, SingleBlockImplicitTerminatorscf::PerformConcurrentlyOp Attributes: ¶ AttributeMLIR TypeDescription mapping::mlir::ArrayAttrDevice Mapping array attribute Operands: ¶ OperandDescription num_threadsindex outputsranked tensor of any type values Results: ¶ ResultDescription resultsany type scf.if (::mlir::scf::IfOp) ¶ if-then-else operation The scf.if operation represents an if-then-else construct for conditionally executing two regions of code. The operand to an if operation is a boolean value. For example: scf.if %b { ... } else { ... } scf.if may also produce results. Which values are returned depends on which execution path is taken. Example: %x, %y = scf.if %b -> (f32, f32) { %x_true = ... %y_true = ... scf.yield %x_true, %y_true : f32, f32 } else { %x_false = ... %y_false = ... scf.yield %x_false, %y_false : f32, f32 } The “then” region has exactly 1 block. The “else” region may have 0 or 1 block. In case the scf.if produces results, the “else” region must also have exactly 1 block. The blocks are always terminated with scf.yield. If scf.if defines no values, the scf.yield can be left out, and will be inserted implicitly. Otherwise, it must be explicit. Example: scf.if %b { ... } The types of the yielded values must match the result types of the scf.if. Traits: NoRegionArguments, RecursiveMemoryEffects, SingleBlockImplicitTerminatorscf::YieldOp Interfaces: InferTypeOpInterface, RegionBranchOpInterface Operands: ¶ OperandDescription condition1-bit signless integer Results: ¶ ResultDescription resultsany type scf.index_switch (::mlir::scf::IndexSwitchOp) ¶ switch-case operation on an index argument Syntax: operation ::= scf.index_switch$arg attr-dict (-> type($results)^)? custom<SwitchCases>($cases, $caseRegions) \n  default$defaultRegion

The scf.index_switch is a control-flow operation that branches to one of the given regions based on the values of the argument and the cases. The argument is always of type index.

The operation always has a “default” region and any number of case regions denoted by integer constants. Control-flow transfers to the case region whose constant value equals the value of the argument. If the argument does not equal any of the case values, control-flow transfer to the “default” region.

Example:

%0 = scf.index_switch %arg0 : index -> i32
case 2 {
%1 = arith.constant 10 : i32
scf.yield %1 : i32
}
case 5 {
%2 = arith.constant 20 : i32
scf.yield %2 : i32
}
default {
%3 = arith.constant 30 : i32
scf.yield %3 : i32
}

Traits: RecursiveMemoryEffects, SingleBlockImplicitTerminatorscf::YieldOp

Interfaces: RegionBranchOpInterface

Attributes: ¶

AttributeMLIR TypeDescription
cases::mlir::DenseI64ArrayAttri64 dense array attribute

Operands: ¶

OperandDescription
argindex

Results: ¶

ResultDescription
resultsany type

scf.parallel (::mlir::scf::ParallelOp) ¶

parallel for operation

The “scf.parallel” operation represents a loop nest taking 4 groups of SSA values as operands that represent the lower bounds, upper bounds, steps and initial values, respectively. The operation defines a variadic number of SSA values for its induction variables. It has one region capturing the loop body. The induction variables are represented as an argument of this region. These SSA values always have type index, which is the size of the machine word. The steps are values of type index, required to be positive. The lower and upper bounds specify a half-open range: the range includes the lower bound but does not include the upper bound. The initial values have the same types as results of “scf.parallel”. If there are no results, the keyword init can be omitted.

Semantically we require that the iteration space can be iterated in any order, and the loop body can be executed in parallel. If there are data races, the behavior is undefined.

The parallel loop operation supports reduction of values produced by individual iterations into a single result. This is modeled using the scf.reduce operation (see scf.reduce for details). Each result of a scf.parallel operation is associated with an initial value operand and reduce operation that is an immediate child. Reductions are matched to result and initial values in order of their appearance in the body. Consequently, we require that the body region has the same number of results and initial values as it has reduce operations.

The body region must contain exactly one block that terminates with “scf.yield” without operands. Parsing ParallelOp will create such a region and insert the terminator when it is absent from the custom format.

Example:

%init = arith.constant 0.0 : f32
scf.parallel (%iv) = (%lb) to (%ub) step (%step) init (%init) -> f32 {
%elem_to_reduce = load %buffer[%iv] : memref<100xf32>
scf.reduce(%elem_to_reduce) : f32 {
^bb0(%lhs : f32, %rhs: f32):
%res = arith.addf %lhs, %rhs : f32
scf.reduce.return %res : f32
}
}

Traits: AttrSizedOperandSegments, AutomaticAllocationScope, RecursiveMemoryEffects, SingleBlockImplicitTerminatorscf::YieldOp

Interfaces: LoopLikeOpInterface

Operands: ¶

OperandDescription
lowerBoundindex
upperBoundindex
stepindex
initValsany type

Results: ¶

ResultDescription
resultsany type

It has a single region with a single block that contains a flat list of ops. Each such op participates in the aggregate formation of a single result of the enclosing scf.foreach_thread. The result number corresponds to the position of the op in the terminator.

Traits: AlwaysSpeculatableImplTrait, HasOnlyGraphRegion, HasParent, NoTerminator, SingleBlock, Terminator

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface), ParallelCombiningOpInterface, RegionKindInterface

Effects: MemoryEffects::Effect{}

scf.reduce (::mlir::scf::ReduceOp) ¶

reduce operation for parallel for

“scf.reduce” is an operation occurring inside “scf.parallel” operations. It consists of one block with two arguments which have the same type as the operand of “scf.reduce”.

“scf.reduce” is used to model the value for reduction computations of a “scf.parallel” operation. It has to appear as an immediate child of a “scf.parallel” and is associated with a result value of its parent operation.

Association is in the order of appearance in the body where the first result of a parallel loop operation corresponds to the first “scf.reduce” in the operation’s body region. The reduce operation takes a single operand, which is the value to be used in the reduction.

The reduce operation contains a region whose entry block expects two arguments of the same type as the operand. As the iteration order of the parallel loop and hence reduction order is unspecified, the result of reduction may be non-deterministic unless the operation is associative and commutative.

The result of the reduce operation’s body must have the same type as the operands and associated result value of the parallel loop operation. Example:

%operand = arith.constant 1.0 : f32
scf.reduce(%operand) : f32 {
^bb0(%lhs : f32, %rhs: f32):
%res = arith.addf %lhs, %rhs : f32
scf.reduce.return %res : f32
}

Traits: HasParent

Operands: ¶

OperandDescription
operandany type

scf.reduce.return (::mlir::scf::ReduceReturnOp) ¶

terminator for reduce operation

Syntax:

operation ::= scf.reduce.return $result attr-dict : type($result)

“scf.reduce.return” is a special terminator operation for the block inside “scf.reduce”. It terminates the region. It should have the same type as the operand of “scf.reduce”. Example for the custom format:

scf.reduce.return %res : f32

Traits: AlwaysSpeculatableImplTrait, HasParent, Terminator

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands: ¶

OperandDescription
resultany type

scf.while (::mlir::scf::WhileOp) ¶

a generic ‘while’ loop

This operation represents a generic “while”/“do-while” loop that keeps iterating as long as a condition is satisfied. There is no restriction on the complexity of the condition. It consists of two regions (with single block each): “before” region and “after” region. The names of regions indicates whether they execute before or after the condition check. Therefore, if the main loop payload is located in the “before” region, the operation is a “do-while” loop. Otherwise, it is a “while” loop.

The “before” region terminates with a special operation, scf.condition, that accepts as its first operand an i1 value indicating whether to proceed to the “after” region (value is true) or not. The two regions communicate by means of region arguments. Initially, the “before” region accepts as arguments the operands of the scf.while operation and uses them to evaluate the condition. It forwards the trailing, non-condition operands of the scf.condition terminator either to the “after” region if the control flow is transferred there or to results of the scf.while operation otherwise. The “after” region takes as arguments the values produced by the “before” region and uses scf.yield to supply new arguments for the “before” region, into which it transfers the control flow unconditionally.

A simple “while” loop can be represented as follows.

%res = scf.while (%arg1 = %init1) : (f32) -> f32 {
// "Before" region.
// In a "while" loop, this region computes the condition.
%condition = call @evaluate_condition(%arg1) : (f32) -> i1

// Forward the argument (as result or "after" region argument).
scf.condition(%condition) %arg1 : f32

} do {
^bb0(%arg2: f32):
// "After" region.
// In a "while" loop, this region is the loop body.
%next = call @payload(%arg2) : (f32) -> f32

// Forward the new value to the "before" region.
// The operand types must match the types of the scf.while operands.
scf.yield %next : f32
}

A simple “do-while” loop can be represented by reducing the “after” block to a simple forwarder.

%res = scf.while (%arg1 = %init1) : (f32) -> f32 {
// "Before" region.
// In a "do-while" loop, this region contains the loop body.
%next = call @payload(%arg1) : (f32) -> f32

// And also evaluates the condition.
%condition = call @evaluate_condition(%arg1) : (f32) -> i1

// Loop through the "after" region.
scf.condition(%condition) %next : f32

} do {
^bb0(%arg2: f32):
// "After" region.
// Forwards the values back to "before" region unmodified.
scf.yield %arg2 : f32
}

Note that the types of region arguments need not to match with each other. The op expects the operand types to match with argument types of the “before” region; the result types to match with the trailing operand types of the terminator of the “before” region, and with the argument types of the “after” region. The following scheme can be used to share the results of some operations executed in the “before” region with the “after” region, avoiding the need to recompute them.

%res = scf.while (%arg1 = %init1) : (f32) -> i64 {
// One can perform some computations, e.g., necessary to evaluate the
// condition, in the "before" region and forward their results to the
// "after" region.
%shared = call @shared_compute(%arg1) : (f32) -> i64

// Evaluate the condition.
%condition = call @evaluate_condition(%arg1, %shared) : (f32, i64) -> i1

// Forward the result of the shared computation to the "after" region.
// The types must match the arguments of the "after" region as well as
// those of the scf.while results.
scf.condition(%condition) %shared : i64

} do {
^bb0(%arg2: i64) {
// Use the partial result to compute the rest of the payload in the
// "after" region.
%res = call @payload(%arg2) : (i64) -> f32

// Forward the new value to the "before" region.
// The operand types must match the types of the scf.while operands.
scf.yield %res : f32
}

The custom syntax for this operation is as follows.

op ::= scf.while assignments : function-type region do region
attributes attribute-dict
initializer ::= /* empty */ | ( assignment-list )
assignment-list ::= assignment | assignment , assignment-list
assignment ::= ssa-value = ssa-value

Traits: RecursiveMemoryEffects

Interfaces: RegionBranchOpInterface

Operands: ¶

OperandDescription
initsany type

Results: ¶

ResultDescription
resultsany type

scf.yield (::mlir::scf::YieldOp) ¶

loop yield and termination operation

Syntax:

operation ::= scf.yield attr-dict ($results^ : type($results))?

“scf.yield” yields an SSA value from the SCF dialect op region and terminates the regions. The semantics of how the values are yielded is defined by the parent operation. If “scf.yield” has any operands, the operands must match the parent operation’s results. If the parent operation defines no values, then the “scf.yield” may be left out in the custom syntax and the builders will insert one implicitly. Otherwise, it has to be present in the syntax to indicate which values are yielded.

Traits: AlwaysSpeculatableImplTrait, HasParent<ExecuteRegionOp, ForOp, IfOp, IndexSwitchOp, ParallelOp, WhileOp>, ReturnLike, Terminator

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands: ¶

OperandDescription
resultsany type