@@ -15,6 +15,8 @@ include "mlir/Dialect/SparseTensor/IR/SparseTensorTypes.td"
15
15
include "mlir/Dialect/SparseTensor/IR/SparseTensorInterfaces.td"
16
16
include "mlir/Interfaces/InferTypeOpInterface.td"
17
17
include "mlir/Interfaces/SideEffectInterfaces.td"
18
+ include "mlir/Interfaces/ControlFlowInterfaces.td"
19
+ include "mlir/Interfaces/LoopLikeInterface.td"
18
20
19
21
//===----------------------------------------------------------------------===//
20
22
// Base class.
@@ -1277,7 +1279,7 @@ def SparseTensor_SelectOp : SparseTensor_Op<"select", [Pure, SameOperandsAndResu
1277
1279
1278
1280
def SparseTensor_YieldOp : SparseTensor_Op<"yield", [Pure, Terminator,
1279
1281
ParentOneOf<["BinaryOp", "UnaryOp", "ReduceOp", "SelectOp",
1280
- "ForeachOp"]>]>,
1282
+ "ForeachOp", "IterateOp" ]>]>,
1281
1283
Arguments<(ins Variadic<AnyType>:$results)> {
1282
1284
let summary = "Yield from sparse_tensor set-like operations";
1283
1285
let description = [{
@@ -1490,6 +1492,103 @@ def ExtractIterSpaceOp : SparseTensor_Op<"extract_iteration_space",
1490
1492
" attr-dict `:` type($tensor) (`,` type($parentIter)^)?";
1491
1493
}
1492
1494
1495
+ def IterateOp : SparseTensor_Op<"iterate",
1496
+ [RecursiveMemoryEffects, RecursivelySpeculatable,
1497
+ DeclareOpInterfaceMethods<LoopLikeOpInterface,
1498
+ ["getInitsMutable", "getLoopResults", "getRegionIterArgs",
1499
+ "getYieldedValuesMutable"]>,
1500
+ DeclareOpInterfaceMethods<RegionBranchOpInterface,
1501
+ ["getEntrySuccessorOperands"]>,
1502
+ SingleBlockImplicitTerminator<"sparse_tensor::YieldOp">]> {
1503
+
1504
+ let arguments = (ins AnySparseIterSpace:$iterSpace,
1505
+ Variadic<AnyType>:$initArgs,
1506
+ LevelSetAttr:$crdUsedLvls);
1507
+ let results = (outs Variadic<AnyType>:$results);
1508
+ let regions = (region SizedRegion<1>:$region);
1509
+
1510
+ let summary = "Iterate over a sparse iteration space";
1511
+ let description = [{
1512
+ The `sparse_tensor.iterate` operations represents a loop over the
1513
+ provided iteration space extracted from a specific sparse tensor.
1514
+ The operation defines an SSA value for a sparse iterator that points
1515
+ to the current stored element in the sparse tensor and SSA values
1516
+ for coordinates of the stored element. The coordinates are always
1517
+ converted to `index` type despite of the underlying sparse tensor
1518
+ storage. When coordinates are not used, the SSA values can be skipped
1519
+ by `_` symbols, which usually leads to simpler generated code after
1520
+ sparsification. For example:
1521
+
1522
+ ```mlir
1523
+ // The coordinate for level 0 is not used when iterating over a 2-D
1524
+ // iteration space.
1525
+ %sparse_tensor.iterate %iterator in %space at(_, %crd_1)
1526
+ : !sparse_tensor.iter_space<#CSR, lvls = 0 to 2>
1527
+ ```
1528
+
1529
+ `sparse_tensor.iterate` can also operate on loop-carried variables
1530
+ and returns the final values after loop termination.
1531
+ The initial values of the variables are passed as additional SSA operands
1532
+ to the iterator SSA value and used coordinate SSA values mentioned
1533
+ above. The operation region has an argument for the iterator, variadic
1534
+ arguments for specified (used) coordiates and followed by one argument
1535
+ for each loop-carried variable, representing the value of the variable
1536
+ at the current iteration.
1537
+ The body region must contain exactly one block that terminates with
1538
+ `sparse_tensor.yield`.
1539
+
1540
+ `sparse_tensor.iterate` results hold the final values after the last
1541
+ iteration. If the `sparse_tensor.iterate` defines any values, a yield
1542
+ must be explicitly present.
1543
+ The number and types of the `sparse_tensor.iterate` results must match
1544
+ the initial values in the iter_args binding and the yield operands.
1545
+
1546
+
1547
+ A nested `sparse_tensor.iterate` example that prints all the coordinates
1548
+ stored in the sparse input:
1549
+
1550
+ ```mlir
1551
+ func.func @nested_iterate(%sp : tensor<4x8xf32, #COO>) {
1552
+ // Iterates over the first level of %sp
1553
+ %l1 = sparse_tensor.extract_iteration_space %sp lvls = 0 : tensor<4x8xf32, #COO>
1554
+ %r1 = sparse_tensor.iterate %it1 in %l1 at (%crd0)
1555
+ : !sparse_tensor.iter_space<#COO, lvls = 0 to 1> {
1556
+ // Iterates over the second level of %sp
1557
+ %l2 = sparse_tensor.extract_iteration_space %sp at %it1 lvls = 1
1558
+ : tensor<4x8xf32, #COO>, !sparse_tensor.iterator<#COO, lvls = 0 to 1>
1559
+ %r2 = sparse_tensor.iterate %it2 in %l2 at (crd1)
1560
+ : !sparse_tensor.iter_space<#COO, lvls = 1 to 2> {
1561
+ vector.print %crd0 : index
1562
+ vector.print %crd1 : index
1563
+ }
1564
+ }
1565
+ }
1566
+
1567
+ ```
1568
+ }];
1569
+
1570
+ let extraClassDeclaration = [{
1571
+ unsigned getSpaceDim() {
1572
+ return getIterSpace().getType().getSpaceDim();
1573
+ }
1574
+ BlockArgument getIterator() {
1575
+ return getRegion().getArguments().front();
1576
+ }
1577
+ Block::BlockArgListType getCrds() {
1578
+ // The first block argument is iterator, the remaining arguments are
1579
+ // referenced coordinates.
1580
+ return getRegion().getArguments().slice(1, getCrdUsedLvls().count());
1581
+ }
1582
+ unsigned getNumRegionIterArgs() {
1583
+ return getRegion().getArguments().size() - 1 - getCrdUsedLvls().count();
1584
+ }
1585
+ }];
1586
+
1587
+ let hasVerifier = 1;
1588
+ let hasRegionVerifier = 1;
1589
+ let hasCustomAssemblyFormat = 1;
1590
+ }
1591
+
1493
1592
//===----------------------------------------------------------------------===//
1494
1593
// Sparse Tensor Debugging and Test-Only Operations.
1495
1594
//===----------------------------------------------------------------------===//
0 commit comments