MLIR

Multi-Level IR Compiler Framework

'bufferization' Dialect

Bufferization in MLIR is the process of converting the tensor type to the memref type. Simply put, bufferization is the process of converting computations on the mathematical tensor construct to computations on physical memory buffers. The bufferization dialect contains operations/interfaces specific to the bufferization passes.

An overview of the bufferization infrastructure and important conceptual details related to using the MLIR dialect conversion infrastructure can be found in bufferization and buffer deallocation.

Operations 

source

bufferization.alloc_tensor (bufferization::AllocTensorOp) 

Allocate buffer for a tensor

bufferization.alloc_tensor materializes an uninitialized tensor with a given shape (dynamic or static). It always bufferizes to a new buffer allocation of the given shape. The optional copy operand specifies the contents of the tensors. If no copy operand is specified, reading from the result of an alloc_tensor op yields an undefined value.

If copy is specified, no dynamic sizes should be passed, since they are the same as the dynamic sizes of the copy operand.

alloc_tensor is a helper op for bufferization. The operation is provided as an anchor that marks the beginning of a new tensor SSA use-def chain. It can be used to control in-place bufferization decisions during One-Shot Bufferize: The bufferized result of a bufferization.alloc_tensor does not alias with any other buffer, so it can be used to resolve read-after-write conflicts that would have been introduced by the in-place bufferization of another op.

The optional memory_space attribute specifies the memory space when bufferizing this op. The memory space is inferred from copy if specified. If neither copy nor memory_space is specified, the default memory space is used during bufferization.

The optional size_hint operand specifies the number of non-zero elements for sparse tensors. The value of size_hint should be not less than 1 and not larger than the linear size of the corresponding dense tensor type. If this requirement is not met, the behavior of the operator is undefined.

Both dense and sparse tensor types are supported. The result of a bufferization.alloc_tensor is a tensor value that can be used like any other tensor value. In practice, it is often used as the “out” operand of another op. Sparse tensor allocations should always be used in a local construction operation and never escape the function boundary directly.

Example:

%c = bufferization.alloc_tensor(%d1, %d2) : tensor<?x?xf32, #SparseMatrix>
%0 = linalg.matmul
  ins(%a, %b: tensor<?x?xf32, #SparseMatrix>, tensor<?x?xf32, #SparseMatrix>)
  outs(%c: tensor<?x?xf32, #SparseMatrix>) -> tensor<?x?xf32, #SparseMatrix>
return %0 : tensor<?x?xf32, #SparseMatrix>
%c = bufferization.alloc_tensor(%d1, %d2) size_hint = %noe
  : tensor<?x?xf32, #SparseMatrix>

Note: An alloc_tensor with a copy should also be expressed as an alloc_tensor without copy, followed by a copy_tensor.

Traits: AttrSizedOperandSegments

Interfaces: BufferizableOpInterface, ReifyRankedShapedTypeOpInterface

Attributes: 

AttributeMLIR TypeDescription
memory_space::mlir::Attributeany attribute

Operands: 

OperandDescription
dynamic_sizesvariadic of index
copytensor of any type values
size_hintindex

Results: 

ResultDescription
resulttensor of any type values

bufferization.clone (bufferization::CloneOp) 

Clone a memref

Syntax:

operation ::= `bufferization.clone` $input attr-dict `:` type($input) `to` type($output)

Clones the data in the input view into an implicitly defined output view.

Usage:

%arg1 = bufferization.clone %arg0 : memref<?xf32> to memref<?xf32>

Valid implementations of this operation may alias the input and output views or create an actual copy. Mutating the source or result of the clone operation after the clone operation thus leads to undefined behavior.

Interfaces: AllocationOpInterface, CopyOpInterface, MemoryEffectsOpInterface

Operands: 

OperandDescription
inputranked or unranked memref of any type values

Results: 

ResultDescription
outputranked or unranked memref of any type values

bufferization.dealloc (bufferization::DeallocOp) 

Deallocates the given memrefs if no alias is retained

Syntax:

operation ::= `bufferization.dealloc` (` ``(` $memrefs^ `:` type($memrefs) `)` `if` ` ` `(` $conditions `)` )?
              (`retain` ` ` `(` $retained^ `:` type($retained) `)` )? attr-dict

This operation deallocates each of the given memrefs if there is no alias to that memref in the list of retained memrefs and the corresponding condition value is set. This condition can be used to indicate and pass on ownership of memref values (or in other words, the responsibility of deallocating that memref). If two memrefs alias each other, only one will be deallocated to avoid double free situations.

The number of variadic memref operands (the memrefs to be deallocated) must equal the number of variadic condition operands and correspond to each other element-wise.

The memref operands must be the originally allocated memrefs, however, the retained memref operands may be arbitrary memrefs.

This operation returns a variadic number of updatedConditions operands, one updated condition per retained memref. An updated condition indicates the ownership of the respective retained memref. It is computed as the disjunction of all conditions operands where the corresponding to memrefs operand aliases with the retained memref. If the retained memref has no aliases among memrefs, the resulting updated condition is ‘false’. This is because all memrefs that need to be deallocated within one basic block should be added to the same bufferization.dealloc operation at the end of the block; if no aliasing memref is present, then it does not have to be deallocated and thus we don’t need to claim ownership. If the memrefs to be deallocated are split over multiple dealloc operations (e.g., to avoid aliasing checks at runtime between the memref operands), then the results have to be manually combined using an arith.ori operation and all of them still require the same list of retained memref operands unless the (potentially empty) set of aliasing memrefs can be determined statically. In that case, the updatedCondition operand can be replaced accordingly (e.g., by a canonicalizer).

Example:

%0:3 = bufferization.dealloc (%a0, %a1 : memref<2xf32>, memref<4xi32>)
  if (%cond0, %cond1) retain (%r0, %r1, %r2 : memref<?xf32>, memref<f64>,
  memref<2xi32>)

Deallocation will be called on %a0 if %cond0 is ’true’ and neither %r0, %r1, or %r2 are aliases of %a0. %a1 will be deallocated when %cond1 is set to ’true’ and none of %r0, %r1, %r2, and %a0` are aliases.

Note that this can be an expensive operation if there are many operands that cannot be optimized away. The runtime cost of this operation (assuming that nothing is optimized away) is O(|memrefs|^2+|memrefs|*|retained|). The cost in terms of memory space is O(|memrefs|+|retained|). As a result, it is recommended to place it carefully in the IR such that most operands can be optimized away by running the buffer-deallocation-simplification pass.

Traits: AttrSizedOperandSegments

Interfaces: InferTypeOpInterface

Operands: 

OperandDescription
memrefsvariadic of ranked or unranked memref of any type values
conditionsvariadic of 1-bit signless integer
retainedvariadic of ranked or unranked memref of any type values

Results: 

ResultDescription
updatedConditionsvariadic of 1-bit signless integer

bufferization.dealloc_tensor (bufferization::DeallocTensorOp) 

Release underlying storage format of given tensor

Syntax:

operation ::= `bufferization.dealloc_tensor` $tensor attr-dict `:` type($tensor)

bufferization.dealloc_tensor is a buffer deallocation in tensor land. This op can be used for manual buffer deallocation. Some bufferizations (such as One-Shot Bufferize) take care of buffer deallocation, in which case this op is usually not needed. Details can be found in the documentation of the respective bufferization passes.

In case of a dense tensor, this op lowers to a memref.dealloc op during bufferization.

In case of a sparse tensor, this op releases the underlying sparse storage format for a tensor that materialized earlier through a new operation, a convert operation with annotated destination tensor type (unless the convert is folded away), or a bufferization.alloc_tensor operation. The release operation should only be called once for any materialized tensor. After this operation, any subsequent memref querying operation on the tensor returns undefined results.

Example:

bufferization.dealloc_tensor %tensor : tensor<1024x1024xf64, #CSR>

Interfaces: BufferizableOpInterface

Operands: 

OperandDescription
tensortensor of any type values

bufferization.materialize_in_destination (bufferization::MaterializeInDestinationOp) 

Copy a tensor

Syntax:

operation ::= `bufferization.materialize_in_destination` $source `in` (`restrict` $restrict^)? (`writable` $writable^)? $dest
              attr-dict `:` functional-type(operands, results)

This op indicates that the data of the source tensor is guaranteed to materialize in dest, which can be a tensor or a memref. In case of a tensor, source materializes in the future buffer of dest and a the updated destination tensor is returned. If this is not possible, e.g., because the destination tensor is read-only or because its original contents are still read later, the input IR fails to bufferize. In case of a memref, source materializes in dest, which is already a buffer. The op has no results in that case.

source, dest and result (if present) must have the same shape and element type. If the op has a result, the types of result and dest must match exactly (e.g., including any tensor encodings).

By default, this op bufferizes to a memcpy from the future buffer of the source tensor to the future buffer of the dest tensor or to the dest buffer. However, transformations such as “empty tensor elimination” may rewrite IR such that a computation is performed directly in dest and no memcpy is needed.

If dest is a buffer, the writable attribute must be specified and the restrict keyword can be specified. These attributes have the same meaning as the respective attributes of bufferization.to_tensor.

writable indicates that the dest buffer is considered writable. It does not make sense to materialize a computation in a read-only buffer, so writable is required.

restrict indicates that there is no bufferization.to_tensor op and no other bufferization.materialize_in_destination op with dest (or an alias thereof) and “restrict”. Only ops with this attribute are considered for “empty tensor elimination”. As part of empty tensor elimination, a new to_tensor op with dest may be inserted and the restrict attribute is transferred from this op to the new to_tensor op. Having “restrict” on this op guarantees that performing empty tensor elimination would not create invalid IR (i.e., having multiple to_tensor restrict with aliasing buffers).

Note: writable could be removed from this op because it must always be set for memref destinations. This op has that attribute to make clear the requirements on the dest operand in the op assembly format.

Note: If dest is a tensor, tensor.insert_slice could be used for the same purpose, but since tensor dialect ops only indicate what should be computed but not where, it could fold away, causing the computation to materialize in a different buffer.

Interfaces: BufferizableOpInterface, DestinationStyleOpInterface, MemoryEffectOpInterface, ReifyRankedShapedTypeOpInterface, SubsetInsertionOpInterface, SubsetOpInterface

Attributes: 

AttributeMLIR TypeDescription
restrict::mlir::UnitAttrunit attribute
writable::mlir::UnitAttrunit attribute

Operands: 

OperandDescription
sourcetensor of any type values
destshaped of any type values

Results: 

ResultDescription
resulttensor of any type values

bufferization.to_memref (bufferization::ToMemrefOp) 

Cast a tensor to memref

Syntax:

operation ::= `bufferization.to_memref` $tensor (`read_only` $read_only^)? attr-dict `:` type($memref)

An operation that returns the future buffer of a tensor.

// Result type is memref<4x?xf32, #layout, 0>
%m = bufferization.to_memref %t : memref<4x?xf32, #layout, 0>

This operation is a specialized variant of the built-in unrealized_conversion_cast and is used to make sure that the IR stays valid at any point during the bufferization.

The read_only attribute can optionally be set, indicating to the bufferization that the buffer returned by this op (or an alias created from the returned buffer) will not be written to.

Traits: AlwaysSpeculatableImplTrait, SameOperandsAndResultElementType, SameOperandsAndResultShape

Interfaces: BufferizableOpInterface, ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes: 

AttributeMLIR TypeDescription
read_only::mlir::UnitAttrunit attribute

Operands: 

OperandDescription
tensortensor of any type values

Results: 

ResultDescription
memrefranked or unranked memref of any type values

bufferization.to_tensor (bufferization::ToTensorOp) 

Create a tensor from a memref

Syntax:

operation ::= `bufferization.to_tensor` $memref (`restrict` $restrict^)? (`writable` $writable^)? attr-dict
              `:` type($memref)

An operation that creates a tensor from a memref. The result value is a tensor whose shape and element type match the memref operand.

The opposite of this op is to_memref. Together, these two ops are useful for source/target materializations when doing type conversions involving tensors and memrefs.

Example:

// Produces a value of tensor<4x?xf32> type.
%t = bufferization.to_tensor %m : memref<4x?xf32, #layout, 0>

If the writable unit attribute is set, the produced tensor is considered “writable” during bufferization. Otherwise, every OpOperand that bufferizes to a write to the future buffer of the resulting tensor (or an alias thereof) will bufferize out-of-place to prevent emitting any writes to memref during bufferization.

The restrict unit attribute (similar to the C restrict keyword) indicates that the produced tensor result is the only way for the tensor IR to gain access to the memref operand (or an alias thereof). E.g., there must be no other to_tensor op with the same or with an aliasing memref operand.

Note: Only to_tensor ops with the restrict unit attribute are supported by One-Shot Bufferize. Other IR is rejected. (To support to_tensor without restrict, One-Shot Bufferize would have to analyze memref IR.) Ops that have incorrect usage of restrict may bufferize incorrectly.

Example:

%t = bufferization.to_tensor %m restrict writable : memref<4xf32>

// %t is writable, so the tensor.insert may bufferize in-place in the
// absence of other conflicts.
%r = tensor.insert %f into %t[%idx] : tensor<4xf32>

to_tensor ops are not bufferized. They are expected to fold away after bufferization. If there are non-bufferizable ops in the IR and allowUnknownOps is set, they may be part of the resulting IR and not fold away. However, such IR is no longer bufferizable with One-Shot Bufferize.

Traits: SameOperandsAndResultElementType, SameOperandsAndResultShape

Interfaces: BufferizableOpInterface, InferTypeOpInterface

Attributes: 

AttributeMLIR TypeDescription
restrict::mlir::UnitAttrunit attribute
writable::mlir::UnitAttrunit attribute

Operands: 

OperandDescription
memrefranked or unranked memref of any type values

Results: 

ResultDescription
resulttensor of any type values