MLIR

Multi-Level IR Compiler Framework

'memref' Dialect

This dialect provides documentation for operations within the MemRef dialect.

Please post an RFC on the forum before adding or changing any operation in this dialect.

Operations 

The memref dialect is intended to hold core memref creation and manipulation ops, which are not strongly associated with any particular other dialect or domain abstraction.

Operation definition 

memref.assume_alignment (::mlir::memref::AssumeAlignmentOp) 

assertion that gives alignment information to the input memref

Syntax:

operation ::= `memref.assume_alignment` $memref `,` $alignment attr-dict `:` type($memref)

The assume_alignment operation takes a memref and an integer of alignment value, and internally annotates the buffer with the given alignment. If the buffer isn’t aligned to the given alignment, the behavior is undefined.

This operation doesn’t affect the semantics of a correct program. It’s for optimization only, and the optimization is best-effort.

Attributes: 

AttributeMLIR TypeDescription
alignment::mlir::IntegerAttr32-bit signless integer attribute whose value is positive

Operands: 

OperandDescription
memrefmemref of any type values

memref.clone (::mlir::memref::CloneOp) 

Syntax:

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

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

Usage:

%arg1 = memref.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.

Operands: 

OperandDescription
inputunranked.memref of any type values or memref of any type values

Results: 

ResultDescription
outputunranked.memref of any type values or memref of any type values

memref.copy (::mlir::memref::CopyOp) 

Syntax:

operation ::= `memref.copy` $source `,` $target attr-dict `:` type($source) `to` type($target)

Copies the data from the source to the destination memref.

Usage:

memref.copy %arg0, %arg1 : memref<?xf32> to memref<?xf32>

Source and destination are expected to have the same element type and shape. Otherwise, the result is undefined. They may have different layouts.

Operands: 

OperandDescription
sourceunranked.memref of any type values or memref of any type values
targetunranked.memref of any type values or memref of any type values

memref.load (::mlir::memref::LoadOp) 

load operation

Syntax:

operation ::= `memref.load` $memref `[` $indices `]` attr-dict `:` type($memref)

The load op reads an element from a memref specified by an index list. The output of load is a new value with the same type as the elements of the memref. The arity of indices is the rank of the memref (i.e., if the memref loaded from is of rank 3, then 3 indices are required for the load following the memref identifier).

In an affine.if or affine.for body, the indices of a load are restricted to SSA values bound to surrounding loop induction variables, symbols , results of a constant operation , or the result of an affine.apply operation that can in turn take as arguments all of the aforementioned SSA values or the recursively result of such an affine.apply operation.

Example:

%1 = affine.apply affine_map<(d0, d1) -> (3*d0)> (%i, %j)
%2 = affine.apply affine_map<(d0, d1) -> (d1+1)> (%i, %j)
%12 = memref.load %A[%1, %2] : memref<8x?xi32, #layout, memspace0>

// Example of an indirect load (treated as non-affine)
%3 = affine.apply affine_map<(d0) -> (2*d0 + 1)>(%12)
%13 = memref.load %A[%3, %2] : memref<4x?xi32, #layout, memspace0>

Context: The load and store operations are specifically crafted to fully resolve a reference to an element of a memref, and (in affine affine.if and affine.for operations) the compiler can follow use-def chains (e.g. through affine.apply operations) to precisely analyze references at compile-time using polyhedral techniques. This is possible because of the restrictions on dimensions and symbols in these contexts.

Operands: 

OperandDescription
memrefmemref of any type values
indicesindex

Results: 

ResultDescription
resultany type

memref.alloc (::mlir::memref::AllocOp) 

memory allocation operation

Syntax:

operation ::= `memref.alloc` `(`$dynamicSizes`)` (`` `[` $symbolOperands^ `]`)? attr-dict `:` type($memref)

The alloc operation allocates a region of memory, as specified by its memref type.

Example:

%0 = memref.alloc() : memref<8x64xf32, 1>

The optional list of dimension operands are bound to the dynamic dimensions specified in its memref type. In the example below, the ssa value ‘%d’ is bound to the second dimension of the memref (which is dynamic).

%0 = memref.alloc(%d) : memref<8x?xf32, 1>

The optional list of symbol operands are bound to the symbols of the memrefs affine map. In the example below, the ssa value ‘%s’ is bound to the symbol ‘s0’ in the affine map specified in the allocs memref type.

%0 = memref.alloc()[%s] : memref<8x64xf32,
                          affine_map<(d0, d1)[s0] -> ((d0 + s0), d1)>, 1>

This operation returns a single ssa value of memref type, which can be used by subsequent load and store operations.

The optional alignment attribute may be specified to ensure that the region of memory that will be indexed is aligned at the specified byte boundary.

%0 = memref.alloc()[%s] {alignment = 8} :
  memref<8x64xf32, affine_map<(d0, d1)[s0] -> ((d0 + s0), d1)>, 1>

Attributes: 

AttributeMLIR TypeDescription
alignment::mlir::IntegerAttr64-bit signless integer attribute whose minimum value is 0

Operands: 

OperandDescription
dynamicSizesindex
symbolOperandsindex

Results: 

ResultDescription
memrefmemref of any type values

memref.alloca (::mlir::memref::AllocaOp) 

stack memory allocation operation

Syntax:

operation ::= `memref.alloca` `(`$dynamicSizes`)` (`` `[` $symbolOperands^ `]`)? attr-dict `:` type($memref)

The alloca operation allocates memory on the stack, to be automatically released when control transfers back from the region of its closest surrounding operation with an AutomaticAllocationScope trait. The amount of memory allocated is specified by its memref and additional operands. For example:

%0 = memref.alloca() : memref<8x64xf32>

The optional list of dimension operands are bound to the dynamic dimensions specified in its memref type. In the example below, the SSA value ‘%d’ is bound to the second dimension of the memref (which is dynamic).

%0 = memref.alloca(%d) : memref<8x?xf32>

The optional list of symbol operands are bound to the symbols of the memref’s affine map. In the example below, the SSA value ‘%s’ is bound to the symbol ‘s0’ in the affine map specified in the allocs memref type.

%0 = memref.alloca()[%s] : memref<8x64xf32,
                           affine_map<(d0, d1)[s0] -> ((d0 + s0), d1)>>

This operation returns a single SSA value of memref type, which can be used by subsequent load and store operations. An optional alignment attribute, if specified, guarantees alignment at least to that boundary. If not specified, an alignment on any convenient boundary compatible with the type will be chosen.

Attributes: 

AttributeMLIR TypeDescription
alignment::mlir::IntegerAttr64-bit signless integer attribute whose minimum value is 0

Operands: 

OperandDescription
dynamicSizesindex
symbolOperandsindex

Results: 

ResultDescription
memrefmemref of any type values

memref.alloca_scope (::mlir::memref::AllocaScopeOp) 

explicitly delimited scope for stack allocation

The memref.alloca_scope operation represents an explicitly-delimited scope for the alloca allocations. Any memref.alloca operations that are used within this scope are going to be cleaned up automatically once the control-flow exits the nested region. For example:

memref.alloca_scope {
  %myalloca = memref.alloca(): memref<4x3xf32>
  ...
}

Here, %myalloca memref is valid within the explicitly delimited scope and is automatically deallocated at the end of the given region. Conceptually, memref.alloca_scope is a passthrough operation with AutomaticAllocationScope that spans the body of the region within the operation.

memref.alloca_scope may also return results that are defined in the nested region. To return a value, one should use memref.alloca_scope.return operation:

%result = memref.alloca_scope {
  ...
  memref.alloca_scope.return %value
}

If memref.alloca_scope returns no value, the memref.alloca_scope.return can be left out, and will be inserted implicitly.

Results: 

ResultDescription
resultsany type

memref.alloca_scope.return (::mlir::memref::AllocaScopeReturnOp) 

terminator for alloca_scope operation

Syntax:

operation ::= `memref.alloca_scope.return` attr-dict ($results^ `:` type($results))?

memref.alloca_scope.return operation returns zero or more SSA values from the region within memref.alloca_scope. If no values are returned, the return operation may be omitted. Otherwise, it has to be present to indicate which values are going to be returned. For example:

memref.alloca_scope.return %value

Operands: 

OperandDescription
resultsany type

memref.buffer_cast (::mlir::memref::BufferCastOp) 

tensor to memref cast operation

Syntax:

operation ::= `memref.buffer_cast` $tensor attr-dict `:` type($memref)

Casts a tensor to a memref.

// Result type is tensor<4x?xf32>
%12 = memref.buffer_cast %10 : memref<4x?xf32, #map0, 42>

Note, that mutating the result of the buffer cast operation leads to undefined behavior.

This operation is a specialized variant of the built-in unrealized_conversion_cast and is intended for use in the context of gradual bufferization.

Operands: 

OperandDescription
tensortensor of any type values

Results: 

ResultDescription
memrefunranked.memref of any type values or memref of any type values

memref.cast (::mlir::memref::CastOp) 

memref cast operation

Syntax:

operation ::= `memref.cast` $source attr-dict `:` type($source) `to` type($dest)

Syntax:

operation ::= ssa-id `=` `memref.cast` ssa-use `:` type `to` type

The memref.cast operation converts a memref from one type to an equivalent type with a compatible shape. The source and destination types are compatible if:

a. Both are ranked memref types with the same element type, address space, and rank and:

  1. Both have the same layout or both have compatible strided layouts.
  2. The individual sizes (resp. offset and strides in the case of strided memrefs) may convert constant dimensions to dynamic dimensions and vice-versa.

If the cast converts any dimensions from an unknown to a known size, then it acts as an assertion that fails at runtime if the dynamic dimensions disagree with resultant destination size.

Example:

// Assert that the input dynamic shape matches the destination static shape.
%2 = memref.cast %1 : memref<?x?xf32> to memref<4x4xf32>
// Erase static shape information, replacing it with dynamic information.
%3 = memref.cast %1 : memref<4xf32> to memref<?xf32>

// The same holds true for offsets and strides.

// Assert that the input dynamic shape matches the destination static stride.
%4 = memref.cast %1 : memref<12x4xf32, offset:?, strides: [?, ?]> to
                      memref<12x4xf32, offset:5, strides: [4, 1]>
// Erase static offset and stride information, replacing it with
// dynamic information.
%5 = memref.cast %1 : memref<12x4xf32, offset:5, strides: [4, 1]> to
                      memref<12x4xf32, offset:?, strides: [?, ?]>

b. Either or both memref types are unranked with the same element type, and address space.

Example:

Cast to concrete shape.
    %4 = memref.cast %1 : memref<*xf32> to memref<4x?xf32>

Erase rank information.
    %5 = memref.cast %1 : memref<4x?xf32> to memref<*xf32>

Operands: 

OperandDescription
sourceunranked.memref of any type values or memref of any type values

Results: 

ResultDescription
destunranked.memref of any type values or memref of any type values

memref.collapse_shape (::mlir::memref::CollapseShapeOp) 

operation to produce a memref with a smaller rank.

The memref.collapse_shape op produces a new view with a smaller rank whose sizes are a reassociation of the original view. Depending on whether or not the reassociated MemRefType is contiguous, the resulting memref may require explicit alloc and copies.

A reassociation is defined as a continuous grouping of dimensions and is represented with an array of I64ArrayAttr attribute.

For now, it is assumed that either:

  1. a reassociation produces and consumes contiguous MemRefType or,
  2. the reshape op will be folded into its consumers (by changing the shape of the computations). All other cases are undefined behavior and a reshape op may not lower to LLVM if it cannot be proven statically that it does not require alloc+copy.

The result memref type of a reshape can be zero-ranked if the operand memref type is statically shaped with all dimensions being unit extent. In such case the reassociation map is empty.

The verification rule is that the reassociation maps are applied to the operand memref with the larger rank to obtain the result memref with the smaller rank.

Examples:

// Dimension collapse (i, j) -> i' and k -> k'
%1 = memref.collapse_shape %0 [[0, 1], [2]] :
  memref<?x?x?xf32, stride_spec> into memref<?x?xf32, stride_spec_2>

Attributes: 

AttributeMLIR TypeDescription
reassociation::mlir::ArrayAttrArray of 64-bit integer array attributes

Operands: 

OperandDescription
srcstrided memref of any type values

Results: 

ResultDescription
resultstrided memref of any type values

memref.dealloc (::mlir::memref::DeallocOp) 

memory deallocation operation

Syntax:

operation ::= `memref.dealloc` $memref attr-dict `:` type($memref)

The dealloc operation frees the region of memory referenced by a memref which was originally created by the alloc operation. The dealloc operation should not be called on memrefs which alias an alloc’d memref (e.g. memrefs returned by view operations).

Example:

%0 = memref.alloc() : memref<8x64xf32, (d0, d1) -> (d0, d1), 1>
memref.dealloc %0 : memref<8x64xf32, (d0, d1) -> (d0, d1), 1>

Operands: 

OperandDescription
memrefunranked.memref of any type values or memref of any type values

memref.dim (::mlir::memref::DimOp) 

dimension index operation

Syntax:

operation ::= `memref.dim` attr-dict $source `,` $index `:` type($source)

The dim operation takes a memref and a dimension operand of type index. It returns the size of the requested dimension of the given memref. If the dimension index is out of bounds the behavior is undefined.

The specified memref type is that of the first operand.

Example:

// Always returns 4, can be constant folded:
%c0 = constant 0 : index
%x = memref.dim %A, %c0 : memref<4 x ? x f32>

// Returns the dynamic dimension of %A.
%c1 = constant 1 : index
%y = memref.dim %A, %c1 : memref<4 x ? x f32>

// Equivalent generic form:
%x = "memref.dim"(%A, %c0) : (memref<4 x ? x f32>, index) -> index
%y = "memref.dim"(%A, %c1) : (memref<4 x ? x f32>, index) -> index

Operands: 

OperandDescription
sourceunranked.memref of any type values or memref of any type values
indexindex

Results: 

ResultDescription
resultindex

memref.expand_shape (::mlir::memref::ExpandShapeOp) 

operation to produce a memref with a higher rank.

The memref.expand_shape op produces a new view with a higher rank whose sizes are a reassociation of the original view. Depending on whether or not the reassociated MemRefType is contiguous, the resulting memref may require explicit alloc and copies.

A reassociation is defined as a continuous grouping of dimensions and is represented with an array of I64ArrayAttr attribute.

For now, it is assumed that either:

  1. a reassociation produces and consumes contiguous MemRefType or,
  2. the reshape op will be folded into its consumers (by changing the shape of the computations). All other cases are undefined behavior and a reshape op may not lower to LLVM if it cannot be proven statically that it does not require alloc+copy.

The operand memref type when dimensions can be zero-ranked if the result memref type is statically shaped with all dimensions being unit extent. In such case the reassociation map is empty.

The verification rule is that the reassociation maps are applied to the result memref with the larger rank to obtain the operand memref with the smaller rank.

Example:

// Dimension expansion i -> (i', j') and (k) -> (k')
%1 = memref.expand_shape %0 [[0, 1], [2]] :
  memref<?x?xf32, stride_spec> into memref<?x?x?xf32, stride_spec_2>

Attributes: 

AttributeMLIR TypeDescription
reassociation::mlir::ArrayAttrArray of 64-bit integer array attributes

Operands: 

OperandDescription
srcstrided memref of any type values

Results: 

ResultDescription
resultstrided memref of any type values

memref.get_global (::mlir::memref::GetGlobalOp) 

get the memref pointing to a global variable

Syntax:

operation ::= `memref.get_global` $name `:` type($result) attr-dict

The memref.get_global operation retrieves the memref pointing to a named global variable. If the global variable is marked constant, writing to the result memref (such as through a memref.store operation) is undefined.

Example:

%x = memref.get_global @foo : memref<2xf32>

Attributes: 

AttributeMLIR TypeDescription
name::mlir::FlatSymbolRefAttrflat symbol reference attribute

Results: 

ResultDescription
resultstatically shaped memref of any type values

memref.global (::mlir::memref::GlobalOp) 

declare or define a global memref variable

Syntax:

operation ::= `memref.global` ($sym_visibility^)?
              (`constant` $constant^)?
              $sym_name `:`
              custom<GlobalMemrefOpTypeAndInitialValue>($type, $initial_value)
              attr-dict

The memref.global operation declares or defines a named global variable. The backing memory for the variable is allocated statically and is described by the type of the variable (which should be a statically shaped memref type). The operation is a declaration if no inital_value is specified, else it is a definition. The initial_value can either be a unit attribute to represent a definition of an uninitialized global variable, or an elements attribute to represent the definition of a global variable with an initial value. The global variable can also be marked constant using the constant unit attribute. Writing to such constant global variables is undefined.

The global variable can be accessed by using the memref.get_global to retrieve the memref for the global variable. Note that the memref for such global variable itself is immutable (i.e., memref.get_global for a given global variable will always return the same memref descriptor).

Example:

// Private variable with an initial value.
memref.global "private" @x : memref<2xf32> = dense<0.0,2.0>

// Declaration of an external variable.
memref.global "private" @y : memref<4xi32>

// Uninitialized externally visible variable.
memref.global @z : memref<3xf16> = uninitialized

// Externally visible constant variable.
memref.global constant @c : memref<2xi32> = dense<1, 4>

Attributes: 

AttributeMLIR TypeDescription
sym_name::mlir::StringAttrstring attribute
sym_visibility::mlir::StringAttrstring attribute
type::mlir::TypeAttrany type attribute
initial_value::mlir::Attributeany attribute
constant::mlir::UnitAttrunit attribute

memref.prefetch (::mlir::memref::PrefetchOp) 

prefetch operation

The “prefetch” op prefetches data from a memref location described with subscript indices similar to memref.load, and with three attributes: a read/write specifier, a locality hint, and a cache type specifier as shown below:

memref.prefetch %0[%i, %j], read, locality<3>, data : memref<400x400xi32>

The read/write specifier is either ‘read’ or ‘write’, the locality hint ranges from locality<0> (no locality) to locality<3> (extremely local keep in cache). The cache type specifier is either ‘data’ or ‘instr’ and specifies whether the prefetch is performed on data cache or on instruction cache.

Attributes: 

AttributeMLIR TypeDescription
isWrite::mlir::BoolAttrbool attribute
localityHint::mlir::IntegerAttr32-bit signless integer attribute whose minimum value is 0 whose maximum value is 3
isDataCache::mlir::BoolAttrbool attribute

Operands: 

OperandDescription
memrefmemref of any type values
indicesindex

memref.reinterpret_cast (::mlir::memref::ReinterpretCastOp) 

memref reinterpret cast operation

Syntax:

operation ::= `memref.reinterpret_cast` $source `to` `offset` `` `:`
              custom<OperandsOrIntegersOffsetsOrStridesList>($offsets, $static_offsets)
              `` `,` `sizes` `` `:`
              custom<OperandsOrIntegersSizesList>($sizes, $static_sizes) `` `,` `strides`
              `` `:`
              custom<OperandsOrIntegersOffsetsOrStridesList>($strides, $static_strides)
              attr-dict `:` type($source) `to` type($result)

Modify offset, sizes and strides of an unranked/ranked memref.

Example:

memref.reinterpret_cast %ranked to
  offset: [0],
  sizes: [%size0, 10],
  strides: [1, %stride1]
: memref<?x?xf32> to memref<?x10xf32, offset: 0, strides: [1, ?]>

memref.reinterpret_cast %unranked to
  offset: [%offset],
  sizes: [%size0, %size1],
  strides: [%stride0, %stride1]
: memref<*xf32> to memref<?x?xf32, offset: ?, strides: [?, ?]>

Attributes: 

AttributeMLIR TypeDescription
static_offsets::mlir::ArrayAttr64-bit integer array attribute
static_sizes::mlir::ArrayAttr64-bit integer array attribute
static_strides::mlir::ArrayAttr64-bit integer array attribute

Operands: 

OperandDescription
sourceunranked.memref of any type values or memref of any type values
offsetsindex
sizesindex
stridesindex

Results: 

ResultDescription
resultmemref of any type values

memref.reshape (::mlir::memref::ReshapeOp) 

memref reshape operation

Syntax:

operation ::= `memref.reshape` $source `(` $shape `)` attr-dict `:` functional-type(operands, results)

The reshape operation converts a memref from one type to an equivalent type with a provided shape. The data is never copied or modified. The source and destination types are compatible if both have the same element type, same number of elements, address space and identity layout map. The following combinations are possible:

a. Source type is ranked or unranked. Shape argument has static size. Result type is ranked.

// Reshape statically-shaped memref.
%dst = memref.reshape %src(%shape)
         : (memref<4x1xf32>, memref<1xi32>) to memref<4xf32>
%dst0 = memref.reshape %src(%shape0)
         : (memref<4x1xf32>, memref<2xi32>) to memref<2x2xf32>
// Flatten unranked memref.
%dst = memref.reshape %src(%shape)
         : (memref<*xf32>, memref<1xi32>) to memref<?xf32>

b. Source type is ranked or unranked. Shape argument has dynamic size. Result type is unranked.

// Reshape dynamically-shaped 1D memref.
%dst = memref.reshape %src(%shape)
         : (memref<?xf32>, memref<?xi32>) to memref<*xf32>
// Reshape unranked memref.
%dst = memref.reshape %src(%shape)
         : (memref<*xf32>, memref<?xi32>) to memref<*xf32>

Operands: 

OperandDescription
sourceunranked.memref of any type values or memref of any type values
shape1D memref of signless integer or index values

Results: 

ResultDescription
resultunranked.memref of any type values or memref of any type values

memref.store (::mlir::memref::StoreOp) 

store operation

Syntax:

operation ::= `memref.store` $value `,` $memref `[` $indices `]` attr-dict `:` type($memref)

Store a value to a memref location given by indices. The value stored should have the same type as the elemental type of the memref. The number of arguments provided within brackets need to match the rank of the memref.

In an affine context, the indices of a store are restricted to SSA values bound to surrounding loop induction variables, symbols , results of a constant operation , or the result of an affine.apply operation that can in turn take as arguments all of the aforementioned SSA values or the recursively result of such an affine.apply operation.

Example:

memref.store %100, %A[%1, 1023] : memref<4x?xf32, #layout, memspace0>

Context: The load and store operations are specifically crafted to fully resolve a reference to an element of a memref, and (in polyhedral affine.if and affine.for operations) the compiler can follow use-def chains (e.g. through affine.apply operations) to precisely analyze references at compile-time using polyhedral techniques. This is possible because of the restrictions on dimensions and symbols in these contexts.

Operands: 

OperandDescription
valueany type
memrefmemref of any type values
indicesindex

memref.transpose (::mlir::memref::TransposeOp) 

transpose produces a new strided memref (metadata-only)

The transpose op produces a strided memref whose sizes and strides are a permutation of the original in memref. This is purely a metadata transformation.

Example:

%1 = memref.transpose %0 (i, j) -> (j, i) : memref<?x?xf32> to memref<?x?xf32, affine_map<(d0, d1)[s0] -> (d1 * s0 + d0)>>

Attributes: 

AttributeMLIR TypeDescription
permutation::mlir::AffineMapAttrAffineMap attribute

Operands: 

OperandDescription
instrided memref of any type values

Results: 

ResultDescription
«unnamed»strided memref of any type values

memref.view (::mlir::memref::ViewOp) 

memref view operation

The “view” operation extracts an N-D contiguous memref with empty layout map with arbitrary element type from a 1-D contiguous memref with empty layout map of i8 element type. The ViewOp supports the following arguments:

  • A single dynamic byte-shift operand must be specified which represents a a shift of the base 1-D memref pointer from which to create the resulting contiguous memref view with identity layout.
  • A dynamic size operand that must be specified for each dynamic dimension in the resulting view memref type.

The “view” operation gives a structured indexing form to a flat 1-D buffer. Unlike “subview” it can perform a type change. The type change behavior requires the op to have special semantics because, e.g. a byte shift of 3 cannot be represented as an offset on f64. For now, a “view” op:

  1. Only takes a contiguous source memref with 0 offset and empty layout.
  2. Must specify a byte_shift operand (in the future, a special integer attribute may be added to support the folded case).
  3. Returns a contiguous memref with 0 offset and empty layout.

Example:

// Allocate a flat 1D/i8 memref.
%0 = memref.alloc() : memref<2048xi8>

// ViewOp with dynamic offset and static sizes.
%1 = memref.view %0[%offset_1024][] : memref<2048xi8> to memref<64x4xf32>

// ViewOp with dynamic offset and two dynamic size.
%2 = memref.view %0[%offset_1024][%size0, %size1] :
  memref<2048xi8> to memref<?x4x?xf32>

Operands: 

OperandDescription
source1D memref of 8-bit signless integer values
byte_shiftindex
sizesindex

Results: 

ResultDescription
«unnamed»memref of any type values

memref.subview (::mlir::memref::SubViewOp) 

memref subview operation

Syntax:

operation ::= `memref.subview` $source ``
              custom<OperandsOrIntegersOffsetsOrStridesList>($offsets, $static_offsets)
              custom<OperandsOrIntegersSizesList>($sizes, $static_sizes)
              custom<OperandsOrIntegersOffsetsOrStridesList>($strides, $static_strides)
              attr-dict `:` type($source) `to` type($result)

The “subview” operation converts a memref type to another memref type which represents a reduced-size view of the original memref as specified by the operation’s offsets, sizes and strides arguments.

The SubView operation supports the following arguments:

  • source: the “base” memref on which to create a “view” memref.
  • offsets: memref-rank number of offsets into the “base” memref at which to create the “view” memref.
  • sizes: memref-rank number of sizes which specify the sizes of the result “view” memref type.
  • strides: memref-rank number of strides that compose multiplicatively with the base memref strides in each dimension.

The representation based on offsets, sizes and strides support a partially-static specification via attributes specified through the static_offsets, static_sizes and static_strides arguments. A special sentinel value ShapedType::kDynamicSize and ShapedType::kDynamicStrideOrOffset encodes that the corresponding entry has a dynamic value.

A subview operation may additionally reduce the rank of the resulting view by removing dimensions that are statically known to be of size 1.

Example 1:

%0 = memref.alloc() : memref<64x4xf32, (d0, d1) -> (d0 * 4 + d1)>

// Create a sub-view of "base" memref '%0' with offset arguments '%c0',
// dynamic sizes for each dimension, and stride arguments '%c1'.
%1 = memref.subview %0[%c0, %c0][%size0, %size1][%c1, %c1]
  : memref<64x4xf32, (d0, d1) -> (d0 * 4 + d1) > to
    memref<?x?xf32, (d0, d1)[s0, s1] -> (d0 * s1 + d1 + s0)>

Example 2:

%0 = memref.alloc() : memref<8x16x4xf32, (d0, d1, d1) -> (d0 * 64 + d1 * 4 + d2)>

// Create a sub-view of "base" memref '%0' with dynamic offsets, sizes,
// and strides.
// Note that dynamic offsets are represented by the linearized dynamic
// offset symbol 's0' in the subview memref layout map, and that the
// dynamic strides operands, after being applied to the base memref
// strides in each dimension, are represented in the view memref layout
// map as symbols 's1', 's2' and 's3'.
%1 = memref.subview %0[%i, %j, %k][%size0, %size1, %size2][%x, %y, %z]
  : memref<8x16x4xf32, (d0, d1, d2) -> (d0 * 64 + d1 * 4 + d2)> to
    memref<?x?x?xf32,
      (d0, d1, d2)[s0, s1, s2, s3] -> (d0 * s1 + d1 * s2 + d2 * s3 + s0)>

Example 3:

%0 = memref.alloc() : memref<8x16x4xf32, (d0, d1, d1) -> (d0 * 64 + d1 * 4 + d2)>

// Subview with constant offsets, sizes and strides.
%1 = memref.subview %0[0, 2, 0][4, 4, 4][64, 4, 1]
  : memref<8x16x4xf32, (d0, d1, d2) -> (d0 * 64 + d1 * 4 + d2)> to
    memref<4x4x4xf32, (d0, d1, d2) -> (d0 * 64 + d1 * 4 + d2 + 8)>

Example 4:

%0 = memref.alloc(%arg0, %arg1) : memref<?x?xf32>

// Subview with constant size, but dynamic offsets and
// strides. The resulting memref has a static shape, but if the
// base memref has an affine map to describe the layout, the result
// memref also uses an affine map to describe the layout. The
// strides of the result memref is computed as follows:
//
// Let #map1 represents the layout of the base memref, and #map2
// represents the layout of the result memref. A #mapsubview can be
// constructed to map an index from the result memref to the base
// memref (note that the description below uses more convenient
// naming for symbols, while in affine maps, symbols are
// represented as unsigned numbers that identify that symbol in the
// given affine map.
//
// #mapsubview = (d0, d1)[o0, o1, t0, t1] -> (d0 * t0 + o0, d1 * t1 + o1)
//
// where, o0, o1, ... are offsets, and t0, t1, ... are strides. Then,
//
// #map2 = #map1.compose(#mapsubview)
//
// If the layout map is represented as
//
// #map1 = (d0, d1)[s0, s1, s2] -> (d0 * s1 + d1 * s2 + s0)
//
// then,
//
// #map2 = (d0, d1)[s0, s1, s2, o0, o1, t0, t1] ->
//              (d0 * s1 * t0 + d1 * s2 * t1 + o0 * s1 + o1 * s2 + s0)
//
// Representing this canonically
//
// #map2 = (d0, d1)[r0, r1, r2] -> (d0 * r1 + d1 * r2 + r0)
//
// where, r0 = o0 * s1 + o1 * s2 + s0, r1 = s1 * t0, r2 = s2 * t1.
%1 = memref.subview %0[%i, %j][4, 4][%x, %y] :
  : memref<?x?xf32, (d0, d1)[s0, s1, s2] -> (d0 * s1 + d1 * s2 + s0)> to
    memref<4x4xf32, (d0, d1)[r0, r1, r2] -> (d0 * r1 + d1 * r2 + r0)>

// Note that the subview op does not guarantee that the result
// memref is "inbounds" w.r.t to base memref. It is upto the client
// to ensure that the subview is accessed in a manner that is
// in-bounds.

Example 5:

// Rank-reducing subview.
%1 = memref.subview %0[0, 0, 0][1, 16, 4][1, 1, 1] :
  memref<8x16x4xf32> to memref<16x4xf32>
%3 = memref.subview %2[3, 4, 2][1, 6, 3][1, 1, 1] :
  memref<8x16x4xf32> to memref<6x3xf32, offset: 210, strides: [4, 1]>

}

Attributes: 

AttributeMLIR TypeDescription
static_offsets::mlir::ArrayAttr64-bit integer array attribute
static_sizes::mlir::ArrayAttr64-bit integer array attribute
static_strides::mlir::ArrayAttr64-bit integer array attribute

Operands: 

OperandDescription
sourcememref of any type values
offsetsindex
sizesindex
stridesindex

Results: 

ResultDescription
resultmemref of any type values

memref.tensor_load (::mlir::memref::TensorLoadOp) 

tensor load operation

Syntax:

operation ::= `memref.tensor_load` $memref attr-dict `:` type($memref)

Create a tensor from a memref, making an independent copy of the element data. The result value is a tensor whose shape and element type match the memref operand.

The opposite of this op is buffer_cast. 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.
%12 = memref.tensor_load %10 : memref<4x?xf32, #layout, memspace0>

If tensor load is used in the bufferization steps, mutating the source buffer after loading leads to undefined behavior.

Operands: 

OperandDescription
memrefunranked.memref of any type values or memref of any type values

Results: 

ResultDescription
resulttensor of any type values

memref.tensor_store (::mlir::memref::TensorStoreOp) 

tensor store operation

Syntax:

operation ::= `memref.tensor_store` $tensor `,` $memref attr-dict `:` type($memref)

Stores the contents of a tensor into a memref. The first operand is a value of tensor type, the second operand is a value of memref type. The shapes and element types of these must match, and are specified by the memref type.

Example:

%9 = dim %8, 1 : tensor<4x?xf32>
%10 = alloc(%9) : memref<4x?xf32, #layout, memspace0>
memref.tensor_store %8, %10 : memref<4x?xf32, #layout, memspace0>

Operands: 

OperandDescription
tensortensor of any type values
memrefunranked.memref of any type values or memref of any type values

‘dma_start’ operation 

Syntax:

operation ::= `dma_start` ssa-use`[`ssa-use-list`]` `,`
               ssa-use`[`ssa-use-list`]` `,` ssa-use `,`
               ssa-use`[`ssa-use-list`]` (`,` ssa-use `,` ssa-use)?
              `:` memref-type `,` memref-type `,` memref-type

Starts a non-blocking DMA operation that transfers data from a source memref to a destination memref. The operands include the source and destination memref’s each followed by its indices, size of the data transfer in terms of the number of elements (of the elemental type of the memref), a tag memref with its indices, and optionally two additional arguments corresponding to the stride (in terms of number of elements) and the number of elements to transfer per stride. The tag location is used by a dma_wait operation to check for completion. The indices of the source memref, destination memref, and the tag memref have the same restrictions as any load/store operation in an affine context (whenever DMA operations appear in an affine context). See restrictions on dimensions and symbols in affine contexts. This allows powerful static analysis and transformations in the presence of such DMAs including rescheduling, pipelining / overlap with computation, and checking for matching start/end operations. The source and destination memref need not be of the same dimensionality, but need to have the same elemental type.

For example, a dma_start operation that transfers 32 vector elements from a memref %src at location [%i, %j] to memref %dst at [%k, %l] would be specified as shown below.

Example:

%size = constant 32 : index
%tag = alloc() : memref<1 x i32, affine_map<(d0) -> (d0)>, 4>
%idx = constant 0 : index
dma_start %src[%i, %j], %dst[%k, %l], %size, %tag[%idx] :
     memref<40 x 8 x vector<16xf32>, affine_map<(d0, d1) -> (d0, d1)>, 0>,
     memref<2 x 4 x vector<16xf32>, affine_map<(d0, d1) -> (d0, d1)>, 2>,
     memref<1 x i32>, affine_map<(d0) -> (d0)>, 4>

‘dma_wait’ operation 

Syntax:

operation ::= `dma_wait` ssa-use`[`ssa-use-list`]` `,` ssa-use `:` memref-type

Blocks until the completion of a DMA operation associated with the tag element specified with a tag memref and its indices. The operands include the tag memref followed by its indices and the number of elements associated with the DMA being waited on. The indices of the tag memref have the same restrictions as load/store indices.

Example:

dma_wait %tag[%idx], %size : memref<1 x i32, affine_map<(d0) -> (d0)>, 4>