'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: ¶

Operands: ¶

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>

Note, that mutating the source or result of the clone operation leads to undefined behavior.

Operands: ¶

Results: ¶

memref.dim (::mlir::memref::DimOp) ¶

dimension index operation

Syntax:

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

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: ¶

Results: ¶

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: ¶

Results: ¶

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: ¶

Operands: ¶

Results: ¶

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: ¶

Operands: ¶

Results: ¶

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.

Operands: ¶

Results: ¶

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: ¶

Results: ¶

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: ¶

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: ¶

Results: ¶

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: ¶

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: ¶

Operands: ¶

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: ¶

Operands: ¶

Results: ¶

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: ¶

Results: ¶

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: ¶

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: ¶

Operands: ¶

Results: ¶

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: ¶

Results: ¶

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: ¶

Operands: ¶

Results: ¶

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: ¶

Results: ¶

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: ¶

‘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>