[mlir][sparse] introduce sparse_tensor.iterate operation
#88955
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
PeimingLiu
requested review from
aartbik,
yinying-lisa-li and
matthias-springer
as code owners
|
@llvm/pr-subscribers-mlir-sparse Author: Peiming Liu (PeimingLiu) ChangesA Patch is 24.58 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/88955.diff 7 Files Affected:
diff --git a/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensor.h b/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensor.h
index 5e523ec428aefb..081a9b8cad8d62 100644
--- a/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensor.h
+++ b/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensor.h
@@ -17,9 +17,13 @@
#include "mlir/IR/OpDefinition.h"
#include "mlir/IR/OpImplementation.h"
#include "mlir/IR/TensorEncoding.h"
+#include "mlir/Interfaces/ControlFlowInterfaces.h"
#include "mlir/Interfaces/InferTypeOpInterface.h"
+#include "mlir/Interfaces/LoopLikeInterface.h"
#include "mlir/Interfaces/SideEffectInterfaces.h"
+#include "llvm/ADT/bit.h"
+
//===----------------------------------------------------------------------===//
//
// Type aliases to help code be more self-documenting. Unfortunately
@@ -41,6 +45,40 @@ using Level = uint64_t;
/// including the value `ShapedType::kDynamic` (for shapes).
using Size = int64_t;
+/// A simple wrapper to encode a bitset of defined (at most 64) levels.
+class LevelSet {
+ uint64_t bits = 0;
+
+public:
+ LevelSet() = default;
+ explicit LevelSet(uint64_t bits) : bits(bits) {}
+ operator uint64_t() const { return bits; }
+
+ LevelSet &set(unsigned i) {
+ assert(i < 64);
+ bits |= 1 << i;
+ return *this;
+ }
+
+ LevelSet &operator|=(LevelSet lhs) {
+ bits |= static_cast<uint64_t>(lhs);
+ return *this;
+ }
+
+ LevelSet &lshift(unsigned offset) {
+ bits = bits << offset;
+ return *this;
+ }
+
+ bool operator[](unsigned i) const {
+ assert(i < 64);
+ return (bits & (1 << i)) != 0;
+ }
+
+ unsigned count() const { return llvm::popcount(bits); }
+ bool empty() const { return bits == 0; }
+};
+
} // namespace sparse_tensor
} // namespace mlir
diff --git a/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorAttrDefs.td b/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorAttrDefs.td
index 4a9b9169ae4b86..d5398a98f5b171 100644
--- a/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorAttrDefs.td
+++ b/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorAttrDefs.td
@@ -19,6 +19,21 @@ class SparseTensor_Attr<string name,
list<Trait> traits = []>
: AttrDef<SparseTensor_Dialect, name, traits>;
+//===----------------------------------------------------------------------===//
+// A simple bitset attribute wrapped over a single int64_t to encode a set of
+// sparse tensor levels.
+//===----------------------------------------------------------------------===//
+
+def LevelSetAttr :
+ TypedAttrBase<
+ I64, "IntegerAttr",
+ And<[CPred<"::llvm::isa<::mlir::IntegerAttr>($_self)">,
+ CPred<"::llvm::cast<::mlir::IntegerAttr>($_self).getType().isInteger(64)">]>,
+ "LevelSet attribute"> {
+ let returnType = [{::mlir::sparse_tensor::LevelSet}];
+ let convertFromStorage = [{::mlir::sparse_tensor::LevelSet($_self.getValue().getZExtValue())}];
+}
+
//===----------------------------------------------------------------------===//
// These attributes are just like `IndexAttr` except that they clarify whether
// the index refers to a dimension (an axis of the semantic tensor) or a level
diff --git a/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.td b/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.td
index d7121e8320a4bc..7d381c980bd889 100644
--- a/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.td
+++ b/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.td
@@ -15,6 +15,8 @@ include "mlir/Dialect/SparseTensor/IR/SparseTensorTypes.td"
include "mlir/Dialect/SparseTensor/IR/SparseTensorInterfaces.td"
include "mlir/Interfaces/InferTypeOpInterface.td"
include "mlir/Interfaces/SideEffectInterfaces.td"
+include "mlir/Interfaces/ControlFlowInterfaces.td"
+include "mlir/Interfaces/LoopLikeInterface.td"
//===----------------------------------------------------------------------===//
// Base class.
@@ -1277,7 +1279,7 @@ def SparseTensor_SelectOp : SparseTensor_Op<"select", [Pure, SameOperandsAndResu
def SparseTensor_YieldOp : SparseTensor_Op<"yield", [Pure, Terminator,
ParentOneOf<["BinaryOp", "UnaryOp", "ReduceOp", "SelectOp",
- "ForeachOp"]>]>,
+ "ForeachOp", "IterateOp"]>]>,
Arguments<(ins Variadic<AnyType>:$results)> {
let summary = "Yield from sparse_tensor set-like operations";
let description = [{
@@ -1490,6 +1492,103 @@ def ExtractIterSpaceOp : SparseTensor_Op<"extract_iteration_space",
" attr-dict `:` type($tensor) (`,` type($parentIter)^)?";
}
+def IterateOp : SparseTensor_Op<"iterate",
+ [RecursiveMemoryEffects, RecursivelySpeculatable,
+ DeclareOpInterfaceMethods<LoopLikeOpInterface,
+ ["getInitsMutable", "getLoopResults", "getRegionIterArgs",
+ "getYieldedValuesMutable"]>,
+ DeclareOpInterfaceMethods<RegionBranchOpInterface,
+ ["getEntrySuccessorOperands"]>,
+ SingleBlockImplicitTerminator<"sparse_tensor::YieldOp">]> {
+
+ let arguments = (ins AnySparseIterSpace:$iterSpace,
+ Variadic<AnyType>:$initArgs,
+ LevelSetAttr:$crdUsedLvls);
+ let results = (outs Variadic<AnyType>:$results);
+ let regions = (region SizedRegion<1>:$region);
+
+ let summary = "Iterate over a sparse iteration space";
+ let description = [{
+ The `sparse_tensor.iterate` operations represents a loop over the
+ provided iteration space extracted from a specific sparse tensor.
+ The operation defines an SSA value for a sparse iterator that points
+ to the current stored element in the sparse tensor and SSA values
+ for coordinates of the stored element. The coordinates are always
+ converted to `index` type despite of the underlying sparse tensor
+ storage. When coordinates are not used, the SSA values can be skipped
+ by `_` symbols, which usually leads to simpler generated code after
+ sparsification. For example:
+
+ ```mlir
+ // The coordinate for level 0 is not used when iterating over a 2-D
+ // iteration space.
+ %sparse_tensor.iterate %iterator in %space at(_, %crd_1)
+ : !sparse_tensor.iter_space<#CSR, lvls = 0 to 2>
+ ```
+
+ `sparse_tensor.iterate` 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 iterator SSA value and used coordinate SSA values mentioned
+ above. The operation region has an argument for the iterator, variadic
+ arguments for specified (used) coordiates and followed by one argument
+ for each loop-carried variable, representing the value of the variable
+ at the current iteration.
+ The body region must contain exactly one block that terminates with
+ `sparse_tensor.yield`.
+
+ `sparse_tensor.iterate` results hold the final values after the last
+ iteration. If the `sparse_tensor.iterate` defines any values, a yield
+ must be explicitly present.
+ The number and types of the `sparse_tensor.iterate` results must match
+ the initial values in the iter_args binding and the yield operands.
+
+
+ A nested `sparse_tensor.iterate` example that prints all the coordinates
+ stored in the sparse input:
+
+ ```mlir
+ func.func @nested_iterate(%sp : tensor<4x8xf32, #COO>) {
+ // Iterates over the first level of %sp
+ %l1 = sparse_tensor.extract_iteration_space %sp lvls = 0 : tensor<4x8xf32, #COO>
+ %r1 = sparse_tensor.iterate %it1 in %l1 at (%crd0)
+ : !sparse_tensor.iter_space<#COO, lvls = 0 to 1> {
+ // Iterates over the second level of %sp
+ %l2 = sparse_tensor.extract_iteration_space %sp at %it1 lvls = 1
+ : tensor<4x8xf32, #COO>, !sparse_tensor.iterator<#COO, lvls = 0 to 1>
+ %r2 = sparse_tensor.iterate %it2 in %l2 at (crd1)
+ : !sparse_tensor.iter_space<#COO, lvls = 1 to 2> {
+ vector.print %crd0 : index
+ vector.print %crd1 : index
+ }
+ }
+ }
+
+ ```
+ }];
+
+ let extraClassDeclaration = [{
+ unsigned getSpaceDim() {
+ return getIterSpace().getType().getSpaceDim();
+ }
+ BlockArgument getIterator() {
+ return getRegion().getArguments().front();
+ }
+ Block::BlockArgListType getCrds() {
+ // The first block argument is iterator, the remaining arguments are
+ // referenced coordinates.
+ return getRegion().getArguments().slice(1, getCrdUsedLvls().count());
+ }
+ unsigned getNumRegionIterArgs() {
+ return getRegion().getArguments().size() - 1 - getCrdUsedLvls().count();
+ }
+ }];
+
+ let hasVerifier = 1;
+ let hasRegionVerifier = 1;
+ let hasCustomAssemblyFormat = 1;
+}
+
//===----------------------------------------------------------------------===//
// Sparse Tensor Debugging and Test-Only Operations.
//===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/SparseTensor/IR/SparseTensorDialect.cpp b/mlir/lib/Dialect/SparseTensor/IR/SparseTensorDialect.cpp
index 516b0943bdcfac..36908def09f403 100644
--- a/mlir/lib/Dialect/SparseTensor/IR/SparseTensorDialect.cpp
+++ b/mlir/lib/Dialect/SparseTensor/IR/SparseTensorDialect.cpp
@@ -2027,6 +2027,106 @@ static void printLevelRange(OpAsmPrinter &p, Operation *, IntegerAttr lvlLo,
printLevelRange(p, lo, hi);
}
+static ParseResult
+parseSparseSpaceLoop(OpAsmParser &parser, OperationState &state,
+ SmallVectorImpl<OpAsmParser::Argument> &iterators,
+ SmallVectorImpl<OpAsmParser::Argument> &iterArgs) {
+ SmallVector<OpAsmParser::UnresolvedOperand> spaces;
+ SmallVector<OpAsmParser::UnresolvedOperand> initArgs;
+
+ // Parses "%iters, ... in %spaces, ..."
+ if (parser.parseArgumentList(iterators) || parser.parseKeyword("in") ||
+ parser.parseOperandList(spaces))
+ return failure();
+
+ if (iterators.size() != spaces.size())
+ return parser.emitError(
+ parser.getNameLoc(),
+ "mismatch in number of sparse iterators and sparse spaces");
+
+ // Parse "at(%crd0, _, ...)"
+ LevelSet crdUsedLvlSet;
+ bool hasUsedCrds = succeeded(parser.parseOptionalKeyword("at"));
+ unsigned lvlCrdCnt = 0;
+ if (hasUsedCrds) {
+ ParseResult crdList = parser.parseCommaSeparatedList(
+ OpAsmParser::Delimiter::Paren, [&]() -> ParseResult {
+ if (parser.parseOptionalKeyword("_")) {
+ if (parser.parseArgument(iterArgs.emplace_back()))
+ return failure();
+ // Always use IndexType for the coordinate.
+ crdUsedLvlSet.set(lvlCrdCnt);
+ iterArgs.back().type = parser.getBuilder().getIndexType();
+ }
+ lvlCrdCnt += 1;
+ return success();
+ });
+ if (failed(crdList)) {
+ return parser.emitError(
+ parser.getNameLoc(),
+ "expecting SSA value or \"_\" for level coordinates");
+ }
+ }
+ // Set the CrdUsedLvl bitset.
+ state.addAttribute("crdUsedLvls",
+ parser.getBuilder().getI64IntegerAttr(crdUsedLvlSet));
+
+ // Parse "iter_args(%arg = %init, ...)"
+ bool hasIterArgs = succeeded(parser.parseOptionalKeyword("iter_args"));
+ if (hasIterArgs)
+ if (parser.parseAssignmentList(iterArgs, initArgs))
+ return failure();
+
+ SmallVector<Type> iterSpaceTps;
+ // parse ": sparse_tensor.iter_space -> ret"
+ if (parser.parseColon() || parser.parseTypeList(iterSpaceTps))
+ return failure();
+ if (iterSpaceTps.size() != spaces.size())
+ return parser.emitError(parser.getNameLoc(),
+ "mismatch in number of iteration space operands "
+ "and iteration space types");
+
+ for (auto [it, tp] : llvm::zip_equal(iterators, iterSpaceTps)) {
+ IterSpaceType spaceTp = llvm::dyn_cast<IterSpaceType>(tp);
+ if (!spaceTp)
+ return parser.emitError(parser.getNameLoc(),
+ "expected sparse_tensor.iter_space type for "
+ "iteration space operands");
+ if (hasUsedCrds && spaceTp.getSpaceDim() != lvlCrdCnt)
+ return parser.emitError(parser.getNameLoc(),
+ "mismatch in number of iteration space dimension "
+ "and specified coordinates");
+ it.type = spaceTp.getIteratorType();
+ }
+
+ if (hasIterArgs)
+ if (parser.parseArrowTypeList(state.types))
+ return failure();
+
+ // Resolves input operands.
+ if (parser.resolveOperands(spaces, iterSpaceTps, parser.getNameLoc(),
+ state.operands))
+ return failure();
+
+ if (hasIterArgs) {
+ unsigned numCrds = crdUsedLvlSet.count();
+ // Strip off leading args that used for coordinates.
+ MutableArrayRef args = MutableArrayRef(iterArgs).drop_front(numCrds);
+ if (args.size() != initArgs.size() || args.size() != state.types.size()) {
+ return parser.emitError(
+ parser.getNameLoc(),
+ "mismatch in number of iteration arguments and return values");
+ }
+
+ for (auto [it, init, tp] : llvm::zip_equal(args, initArgs, state.types)) {
+ it.type = tp;
+ if (parser.resolveOperand(init, tp, state.operands))
+ return failure();
+ }
+ }
+ return success();
+}
+
LogicalResult ExtractIterSpaceOp::inferReturnTypes(
MLIRContext *ctx, std::optional<Location> loc, ValueRange ops,
DictionaryAttr attr, OpaqueProperties prop, RegionRange region,
@@ -2063,6 +2163,161 @@ LogicalResult ExtractIterSpaceOp::verify() {
return success();
}
+ParseResult IterateOp::parse(OpAsmParser &parser, OperationState &result) {
+ OpAsmParser::Argument iterator;
+ OpAsmParser::UnresolvedOperand iterSpace;
+
+ SmallVector<OpAsmParser::Argument> iters, iterArgs;
+ if (parseSparseSpaceLoop(parser, result, iters, iterArgs))
+ return failure();
+ if (iters.size() != 1)
+ return parser.emitError(parser.getNameLoc(),
+ "expected only one iterator/iteration space");
+
+ iters.append(iterArgs);
+ Region *body = result.addRegion();
+ if (parser.parseRegion(*body, iters))
+ return failure();
+
+ IterateOp::ensureTerminator(*body, parser.getBuilder(), result.location);
+
+ // Parse the optional attribute list.
+ if (parser.parseOptionalAttrDict(result.attributes))
+ return failure();
+
+ return success();
+}
+
+/// Prints the initialization list in the form of
+/// <prefix>(%inner = %outer, %inner2 = %outer2, <...>)
+/// where 'inner' values are assumed to be region arguments and 'outer' values
+/// are regular SSA values.
+static void printInitializationList(OpAsmPrinter &p,
+ Block::BlockArgListType blocksArgs,
+ ValueRange initializers,
+ StringRef prefix = "") {
+ assert(blocksArgs.size() == initializers.size() &&
+ "expected same length of arguments and initializers");
+ if (initializers.empty())
+ return;
+
+ p << prefix << '(';
+ llvm::interleaveComma(llvm::zip(blocksArgs, initializers), p, [&](auto it) {
+ p << std::get<0>(it) << " = " << std::get<1>(it);
+ });
+ p << ")";
+}
+
+static void printUsedCrdsList(OpAsmPrinter &p, unsigned spaceDim,
+ Block::BlockArgListType blocksArgs,
+ LevelSet crdUsedLvls) {
+ if (crdUsedLvls.empty())
+ return;
+
+ p << " at(";
+ for (unsigned i = 0; i < spaceDim; i++) {
+ if (crdUsedLvls[i]) {
+ p << blocksArgs.front();
+ blocksArgs = blocksArgs.drop_front();
+ } else {
+ p << "_";
+ }
+ if (i != spaceDim - 1)
+ p << ", ";
+ }
+ assert(blocksArgs.empty());
+ p << ")";
+}
+
+void IterateOp::print(OpAsmPrinter &p) {
+ p << " " << getIterator() << " in " << getIterSpace();
+ printUsedCrdsList(p, getSpaceDim(), getCrds(), getCrdUsedLvls());
+ printInitializationList(p, getRegionIterArgs(), getInitArgs(), " iter_args");
+
+ p << " : " << getIterSpace().getType() << " ";
+ if (!getInitArgs().empty())
+ p << "-> (" << getInitArgs().getTypes() << ") ";
+
+ p.printRegion(getRegion(), /*printEntryBlockArgs=*/false,
+ /*printBlockTerminators=*/!getInitArgs().empty());
+}
+
+LogicalResult IterateOp::verify() {
+ if (getInitArgs().size() != getNumResults()) {
+ return emitOpError(
+ "mismatch in number of loop-carried values and defined values");
+ }
+ return success();
+}
+
+LogicalResult IterateOp::verifyRegions() {
+ if (getIterator().getType() != getIterSpace().getType().getIteratorType())
+ return emitOpError("mismatch in iterator and iteration space type");
+ if (getNumRegionIterArgs() != getNumResults())
+ return emitOpError(
+ "mismatch in number of basic block args and defined values");
+
+ auto initArgs = getInitArgs();
+ auto iterArgs = getRegionIterArgs();
+ auto yieldVals = getYieldedValues();
+ auto opResults = getResults();
+ if (!llvm::all_equal({initArgs.size(), iterArgs.size(), yieldVals.size(),
+ opResults.size()})) {
+ return emitOpError() << "number mismatch between iter args and results.";
+ }
+
+ for (auto [i, init, iter, yield, ret] :
+ llvm::enumerate(initArgs, iterArgs, yieldVals, opResults)) {
+ if (init.getType() != ret.getType())
+ return emitOpError() << "types mismatch between " << i
+ << "th iter operand and defined value";
+ if (iter.getType() != ret.getType())
+ return emitOpError() << "types mismatch between " << i
+ << "th iter region arg and defined value";
+ if (yield.getType() != ret.getType())
+ return emitOpError() << "types mismatch between " << i
+ << "th yield value and defined value";
+ }
+
+ return success();
+}
+
+/// IterateOp implemented OpInterfaces' methods.
+SmallVector<Region *> IterateOp::getLoopRegions() { return {&getRegion()}; }
+
+MutableArrayRef<OpOperand> IterateOp::getInitsMutable() {
+ return getInitArgsMutable();
+}
+
+Block::BlockArgListType IterateOp::getRegionIterArgs() {
+ return getRegion().getArguments().take_back(getNumRegionIterArgs());
+}
+
+std::optional<MutableArrayRef<OpOperand>> IterateOp::getYieldedValuesMutable() {
+ return cast<sparse_tensor::YieldOp>(
+ getRegion().getBlocks().front().getTerminator())
+ .getResultsMutable();
+}
+
+std::optional<ResultRange> IterateOp::getLoopResults() { return getResults(); }
+
+OperandRange IterateOp::getEntrySuccessorOperands(RegionBranchPoint point) {
+ return getInitArgs();
+}
+
+void IterateOp::getSuccessorRegions(RegionBranchPoint point,
+ SmallVectorImpl<RegionSuccessor> ®ions) {
+ // Both the operation itself and the region may be branching into the body or
+ // back into the operation itself.
+ regions.push_back(RegionSuccessor(&getRegion(), getRegionIterArgs()));
+ // It is possible for loop not to enter the body.
+ regions.push_back(RegionSuccessor(getResults()));
+}
+
+//===----------------------------------------------------------------------===//
+// Sparse Tensor Dialect Setups.
+//===----------------------------------------------------------------------===//
+
/// Materialize a single constant operation from a given attribute value with
/// the desired resultant type.
Operation *SparseTensorDialect::materializeConstant(OpBuilder &builder,
diff --git a/mlir/test/Dialect/SparseTensor/invalid.mlir b/mlir/test/Dialect/SparseTensor/invalid.mlir
index 3fa696e1600a93..b13024cd4ed99d 100644
--- a/mlir/test/Dialect/SparseTensor/invalid.mlir
+++ b/mlir/test/Dialect/SparseTensor/invalid.mlir
@@ -1094,3 +1094,60 @@ func.func @sparse_extract_iter_space(%sp : tensor<4x8xf32, #COO>, %it1 : !sparse
%l1 = sparse_tensor.extract_iteration_space %sp at %it1 lvls = 2 : tensor<4x8xf32, #COO>, !sparse_tensor.iterator<#COO, lvls = 0>
return
}
+
+
+// -----
+
+#COO = #sparse_tensor.encoding<{
+ map = (i, j) -> (
+ i : compressed(nonunique),
+ j : singleton(soa)
+ )
+}>
+
+func.func @sparse_iterate(%sp : tensor<4x8xf32, #COO>, %i : index, %j : index) -> index {
+ %l1 = sparse_tensor.extract_iteration_space %sp lvls = 0 : tensor<4x8xf32, #COO>
+ // expected-error @+1 {{'sparse_tensor.iterate' op different number o...
[truncated]
|
|
@llvm/pr-subscribers-mlir Author: Peiming Liu (PeimingLiu) ChangesA Patch is 24.58 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/88955.diff 7 Files Affected:
diff --git a/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensor.h b/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensor.h
index 5e523ec428aefb..081a9b8cad8d62 100644
--- a/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensor.h
+++ b/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensor.h
@@ -17,9 +17,13 @@
#include "mlir/IR/OpDefinition.h"
#include "mlir/IR/OpImplementation.h"
#include "mlir/IR/TensorEncoding.h"
+#include "mlir/Interfaces/ControlFlowInterfaces.h"
#include "mlir/Interfaces/InferTypeOpInterface.h"
+#include "mlir/Interfaces/LoopLikeInterface.h"
#include "mlir/Interfaces/SideEffectInterfaces.h"
+#include "llvm/ADT/bit.h"
+
//===----------------------------------------------------------------------===//
//
// Type aliases to help code be more self-documenting. Unfortunately
@@ -41,6 +45,40 @@ using Level = uint64_t;
/// including the value `ShapedType::kDynamic` (for shapes).
using Size = int64_t;
+/// A simple wrapper to encode a bitset of defined (at most 64) levels.
+class LevelSet {
+ uint64_t bits = 0;
+
+public:
+ LevelSet() = default;
+ explicit LevelSet(uint64_t bits) : bits(bits) {}
+ operator uint64_t() const { return bits; }
+
+ LevelSet &set(unsigned i) {
+ assert(i < 64);
+ bits |= 1 << i;
+ return *this;
+ }
+
+ LevelSet &operator|=(LevelSet lhs) {
+ bits |= static_cast<uint64_t>(lhs);
+ return *this;
+ }
+
+ LevelSet &lshift(unsigned offset) {
+ bits = bits << offset;
+ return *this;
+ }
+
+ bool operator[](unsigned i) const {
+ assert(i < 64);
+ return (bits & (1 << i)) != 0;
+ }
+
+ unsigned count() const { return llvm::popcount(bits); }
+ bool empty() const { return bits == 0; }
+};
+
} // namespace sparse_tensor
} // namespace mlir
diff --git a/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorAttrDefs.td b/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorAttrDefs.td
index 4a9b9169ae4b86..d5398a98f5b171 100644
--- a/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorAttrDefs.td
+++ b/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorAttrDefs.td
@@ -19,6 +19,21 @@ class SparseTensor_Attr<string name,
list<Trait> traits = []>
: AttrDef<SparseTensor_Dialect, name, traits>;
+//===----------------------------------------------------------------------===//
+// A simple bitset attribute wrapped over a single int64_t to encode a set of
+// sparse tensor levels.
+//===----------------------------------------------------------------------===//
+
+def LevelSetAttr :
+ TypedAttrBase<
+ I64, "IntegerAttr",
+ And<[CPred<"::llvm::isa<::mlir::IntegerAttr>($_self)">,
+ CPred<"::llvm::cast<::mlir::IntegerAttr>($_self).getType().isInteger(64)">]>,
+ "LevelSet attribute"> {
+ let returnType = [{::mlir::sparse_tensor::LevelSet}];
+ let convertFromStorage = [{::mlir::sparse_tensor::LevelSet($_self.getValue().getZExtValue())}];
+}
+
//===----------------------------------------------------------------------===//
// These attributes are just like `IndexAttr` except that they clarify whether
// the index refers to a dimension (an axis of the semantic tensor) or a level
diff --git a/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.td b/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.td
index d7121e8320a4bc..7d381c980bd889 100644
--- a/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.td
+++ b/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.td
@@ -15,6 +15,8 @@ include "mlir/Dialect/SparseTensor/IR/SparseTensorTypes.td"
include "mlir/Dialect/SparseTensor/IR/SparseTensorInterfaces.td"
include "mlir/Interfaces/InferTypeOpInterface.td"
include "mlir/Interfaces/SideEffectInterfaces.td"
+include "mlir/Interfaces/ControlFlowInterfaces.td"
+include "mlir/Interfaces/LoopLikeInterface.td"
//===----------------------------------------------------------------------===//
// Base class.
@@ -1277,7 +1279,7 @@ def SparseTensor_SelectOp : SparseTensor_Op<"select", [Pure, SameOperandsAndResu
def SparseTensor_YieldOp : SparseTensor_Op<"yield", [Pure, Terminator,
ParentOneOf<["BinaryOp", "UnaryOp", "ReduceOp", "SelectOp",
- "ForeachOp"]>]>,
+ "ForeachOp", "IterateOp"]>]>,
Arguments<(ins Variadic<AnyType>:$results)> {
let summary = "Yield from sparse_tensor set-like operations";
let description = [{
@@ -1490,6 +1492,103 @@ def ExtractIterSpaceOp : SparseTensor_Op<"extract_iteration_space",
" attr-dict `:` type($tensor) (`,` type($parentIter)^)?";
}
+def IterateOp : SparseTensor_Op<"iterate",
+ [RecursiveMemoryEffects, RecursivelySpeculatable,
+ DeclareOpInterfaceMethods<LoopLikeOpInterface,
+ ["getInitsMutable", "getLoopResults", "getRegionIterArgs",
+ "getYieldedValuesMutable"]>,
+ DeclareOpInterfaceMethods<RegionBranchOpInterface,
+ ["getEntrySuccessorOperands"]>,
+ SingleBlockImplicitTerminator<"sparse_tensor::YieldOp">]> {
+
+ let arguments = (ins AnySparseIterSpace:$iterSpace,
+ Variadic<AnyType>:$initArgs,
+ LevelSetAttr:$crdUsedLvls);
+ let results = (outs Variadic<AnyType>:$results);
+ let regions = (region SizedRegion<1>:$region);
+
+ let summary = "Iterate over a sparse iteration space";
+ let description = [{
+ The `sparse_tensor.iterate` operations represents a loop over the
+ provided iteration space extracted from a specific sparse tensor.
+ The operation defines an SSA value for a sparse iterator that points
+ to the current stored element in the sparse tensor and SSA values
+ for coordinates of the stored element. The coordinates are always
+ converted to `index` type despite of the underlying sparse tensor
+ storage. When coordinates are not used, the SSA values can be skipped
+ by `_` symbols, which usually leads to simpler generated code after
+ sparsification. For example:
+
+ ```mlir
+ // The coordinate for level 0 is not used when iterating over a 2-D
+ // iteration space.
+ %sparse_tensor.iterate %iterator in %space at(_, %crd_1)
+ : !sparse_tensor.iter_space<#CSR, lvls = 0 to 2>
+ ```
+
+ `sparse_tensor.iterate` 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 iterator SSA value and used coordinate SSA values mentioned
+ above. The operation region has an argument for the iterator, variadic
+ arguments for specified (used) coordiates and followed by one argument
+ for each loop-carried variable, representing the value of the variable
+ at the current iteration.
+ The body region must contain exactly one block that terminates with
+ `sparse_tensor.yield`.
+
+ `sparse_tensor.iterate` results hold the final values after the last
+ iteration. If the `sparse_tensor.iterate` defines any values, a yield
+ must be explicitly present.
+ The number and types of the `sparse_tensor.iterate` results must match
+ the initial values in the iter_args binding and the yield operands.
+
+
+ A nested `sparse_tensor.iterate` example that prints all the coordinates
+ stored in the sparse input:
+
+ ```mlir
+ func.func @nested_iterate(%sp : tensor<4x8xf32, #COO>) {
+ // Iterates over the first level of %sp
+ %l1 = sparse_tensor.extract_iteration_space %sp lvls = 0 : tensor<4x8xf32, #COO>
+ %r1 = sparse_tensor.iterate %it1 in %l1 at (%crd0)
+ : !sparse_tensor.iter_space<#COO, lvls = 0 to 1> {
+ // Iterates over the second level of %sp
+ %l2 = sparse_tensor.extract_iteration_space %sp at %it1 lvls = 1
+ : tensor<4x8xf32, #COO>, !sparse_tensor.iterator<#COO, lvls = 0 to 1>
+ %r2 = sparse_tensor.iterate %it2 in %l2 at (crd1)
+ : !sparse_tensor.iter_space<#COO, lvls = 1 to 2> {
+ vector.print %crd0 : index
+ vector.print %crd1 : index
+ }
+ }
+ }
+
+ ```
+ }];
+
+ let extraClassDeclaration = [{
+ unsigned getSpaceDim() {
+ return getIterSpace().getType().getSpaceDim();
+ }
+ BlockArgument getIterator() {
+ return getRegion().getArguments().front();
+ }
+ Block::BlockArgListType getCrds() {
+ // The first block argument is iterator, the remaining arguments are
+ // referenced coordinates.
+ return getRegion().getArguments().slice(1, getCrdUsedLvls().count());
+ }
+ unsigned getNumRegionIterArgs() {
+ return getRegion().getArguments().size() - 1 - getCrdUsedLvls().count();
+ }
+ }];
+
+ let hasVerifier = 1;
+ let hasRegionVerifier = 1;
+ let hasCustomAssemblyFormat = 1;
+}
+
//===----------------------------------------------------------------------===//
// Sparse Tensor Debugging and Test-Only Operations.
//===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/SparseTensor/IR/SparseTensorDialect.cpp b/mlir/lib/Dialect/SparseTensor/IR/SparseTensorDialect.cpp
index 516b0943bdcfac..36908def09f403 100644
--- a/mlir/lib/Dialect/SparseTensor/IR/SparseTensorDialect.cpp
+++ b/mlir/lib/Dialect/SparseTensor/IR/SparseTensorDialect.cpp
@@ -2027,6 +2027,106 @@ static void printLevelRange(OpAsmPrinter &p, Operation *, IntegerAttr lvlLo,
printLevelRange(p, lo, hi);
}
+static ParseResult
+parseSparseSpaceLoop(OpAsmParser &parser, OperationState &state,
+ SmallVectorImpl<OpAsmParser::Argument> &iterators,
+ SmallVectorImpl<OpAsmParser::Argument> &iterArgs) {
+ SmallVector<OpAsmParser::UnresolvedOperand> spaces;
+ SmallVector<OpAsmParser::UnresolvedOperand> initArgs;
+
+ // Parses "%iters, ... in %spaces, ..."
+ if (parser.parseArgumentList(iterators) || parser.parseKeyword("in") ||
+ parser.parseOperandList(spaces))
+ return failure();
+
+ if (iterators.size() != spaces.size())
+ return parser.emitError(
+ parser.getNameLoc(),
+ "mismatch in number of sparse iterators and sparse spaces");
+
+ // Parse "at(%crd0, _, ...)"
+ LevelSet crdUsedLvlSet;
+ bool hasUsedCrds = succeeded(parser.parseOptionalKeyword("at"));
+ unsigned lvlCrdCnt = 0;
+ if (hasUsedCrds) {
+ ParseResult crdList = parser.parseCommaSeparatedList(
+ OpAsmParser::Delimiter::Paren, [&]() -> ParseResult {
+ if (parser.parseOptionalKeyword("_")) {
+ if (parser.parseArgument(iterArgs.emplace_back()))
+ return failure();
+ // Always use IndexType for the coordinate.
+ crdUsedLvlSet.set(lvlCrdCnt);
+ iterArgs.back().type = parser.getBuilder().getIndexType();
+ }
+ lvlCrdCnt += 1;
+ return success();
+ });
+ if (failed(crdList)) {
+ return parser.emitError(
+ parser.getNameLoc(),
+ "expecting SSA value or \"_\" for level coordinates");
+ }
+ }
+ // Set the CrdUsedLvl bitset.
+ state.addAttribute("crdUsedLvls",
+ parser.getBuilder().getI64IntegerAttr(crdUsedLvlSet));
+
+ // Parse "iter_args(%arg = %init, ...)"
+ bool hasIterArgs = succeeded(parser.parseOptionalKeyword("iter_args"));
+ if (hasIterArgs)
+ if (parser.parseAssignmentList(iterArgs, initArgs))
+ return failure();
+
+ SmallVector<Type> iterSpaceTps;
+ // parse ": sparse_tensor.iter_space -> ret"
+ if (parser.parseColon() || parser.parseTypeList(iterSpaceTps))
+ return failure();
+ if (iterSpaceTps.size() != spaces.size())
+ return parser.emitError(parser.getNameLoc(),
+ "mismatch in number of iteration space operands "
+ "and iteration space types");
+
+ for (auto [it, tp] : llvm::zip_equal(iterators, iterSpaceTps)) {
+ IterSpaceType spaceTp = llvm::dyn_cast<IterSpaceType>(tp);
+ if (!spaceTp)
+ return parser.emitError(parser.getNameLoc(),
+ "expected sparse_tensor.iter_space type for "
+ "iteration space operands");
+ if (hasUsedCrds && spaceTp.getSpaceDim() != lvlCrdCnt)
+ return parser.emitError(parser.getNameLoc(),
+ "mismatch in number of iteration space dimension "
+ "and specified coordinates");
+ it.type = spaceTp.getIteratorType();
+ }
+
+ if (hasIterArgs)
+ if (parser.parseArrowTypeList(state.types))
+ return failure();
+
+ // Resolves input operands.
+ if (parser.resolveOperands(spaces, iterSpaceTps, parser.getNameLoc(),
+ state.operands))
+ return failure();
+
+ if (hasIterArgs) {
+ unsigned numCrds = crdUsedLvlSet.count();
+ // Strip off leading args that used for coordinates.
+ MutableArrayRef args = MutableArrayRef(iterArgs).drop_front(numCrds);
+ if (args.size() != initArgs.size() || args.size() != state.types.size()) {
+ return parser.emitError(
+ parser.getNameLoc(),
+ "mismatch in number of iteration arguments and return values");
+ }
+
+ for (auto [it, init, tp] : llvm::zip_equal(args, initArgs, state.types)) {
+ it.type = tp;
+ if (parser.resolveOperand(init, tp, state.operands))
+ return failure();
+ }
+ }
+ return success();
+}
+
LogicalResult ExtractIterSpaceOp::inferReturnTypes(
MLIRContext *ctx, std::optional<Location> loc, ValueRange ops,
DictionaryAttr attr, OpaqueProperties prop, RegionRange region,
@@ -2063,6 +2163,161 @@ LogicalResult ExtractIterSpaceOp::verify() {
return success();
}
+ParseResult IterateOp::parse(OpAsmParser &parser, OperationState &result) {
+ OpAsmParser::Argument iterator;
+ OpAsmParser::UnresolvedOperand iterSpace;
+
+ SmallVector<OpAsmParser::Argument> iters, iterArgs;
+ if (parseSparseSpaceLoop(parser, result, iters, iterArgs))
+ return failure();
+ if (iters.size() != 1)
+ return parser.emitError(parser.getNameLoc(),
+ "expected only one iterator/iteration space");
+
+ iters.append(iterArgs);
+ Region *body = result.addRegion();
+ if (parser.parseRegion(*body, iters))
+ return failure();
+
+ IterateOp::ensureTerminator(*body, parser.getBuilder(), result.location);
+
+ // Parse the optional attribute list.
+ if (parser.parseOptionalAttrDict(result.attributes))
+ return failure();
+
+ return success();
+}
+
+/// Prints the initialization list in the form of
+/// <prefix>(%inner = %outer, %inner2 = %outer2, <...>)
+/// where 'inner' values are assumed to be region arguments and 'outer' values
+/// are regular SSA values.
+static void printInitializationList(OpAsmPrinter &p,
+ Block::BlockArgListType blocksArgs,
+ ValueRange initializers,
+ StringRef prefix = "") {
+ assert(blocksArgs.size() == initializers.size() &&
+ "expected same length of arguments and initializers");
+ if (initializers.empty())
+ return;
+
+ p << prefix << '(';
+ llvm::interleaveComma(llvm::zip(blocksArgs, initializers), p, [&](auto it) {
+ p << std::get<0>(it) << " = " << std::get<1>(it);
+ });
+ p << ")";
+}
+
+static void printUsedCrdsList(OpAsmPrinter &p, unsigned spaceDim,
+ Block::BlockArgListType blocksArgs,
+ LevelSet crdUsedLvls) {
+ if (crdUsedLvls.empty())
+ return;
+
+ p << " at(";
+ for (unsigned i = 0; i < spaceDim; i++) {
+ if (crdUsedLvls[i]) {
+ p << blocksArgs.front();
+ blocksArgs = blocksArgs.drop_front();
+ } else {
+ p << "_";
+ }
+ if (i != spaceDim - 1)
+ p << ", ";
+ }
+ assert(blocksArgs.empty());
+ p << ")";
+}
+
+void IterateOp::print(OpAsmPrinter &p) {
+ p << " " << getIterator() << " in " << getIterSpace();
+ printUsedCrdsList(p, getSpaceDim(), getCrds(), getCrdUsedLvls());
+ printInitializationList(p, getRegionIterArgs(), getInitArgs(), " iter_args");
+
+ p << " : " << getIterSpace().getType() << " ";
+ if (!getInitArgs().empty())
+ p << "-> (" << getInitArgs().getTypes() << ") ";
+
+ p.printRegion(getRegion(), /*printEntryBlockArgs=*/false,
+ /*printBlockTerminators=*/!getInitArgs().empty());
+}
+
+LogicalResult IterateOp::verify() {
+ if (getInitArgs().size() != getNumResults()) {
+ return emitOpError(
+ "mismatch in number of loop-carried values and defined values");
+ }
+ return success();
+}
+
+LogicalResult IterateOp::verifyRegions() {
+ if (getIterator().getType() != getIterSpace().getType().getIteratorType())
+ return emitOpError("mismatch in iterator and iteration space type");
+ if (getNumRegionIterArgs() != getNumResults())
+ return emitOpError(
+ "mismatch in number of basic block args and defined values");
+
+ auto initArgs = getInitArgs();
+ auto iterArgs = getRegionIterArgs();
+ auto yieldVals = getYieldedValues();
+ auto opResults = getResults();
+ if (!llvm::all_equal({initArgs.size(), iterArgs.size(), yieldVals.size(),
+ opResults.size()})) {
+ return emitOpError() << "number mismatch between iter args and results.";
+ }
+
+ for (auto [i, init, iter, yield, ret] :
+ llvm::enumerate(initArgs, iterArgs, yieldVals, opResults)) {
+ if (init.getType() != ret.getType())
+ return emitOpError() << "types mismatch between " << i
+ << "th iter operand and defined value";
+ if (iter.getType() != ret.getType())
+ return emitOpError() << "types mismatch between " << i
+ << "th iter region arg and defined value";
+ if (yield.getType() != ret.getType())
+ return emitOpError() << "types mismatch between " << i
+ << "th yield value and defined value";
+ }
+
+ return success();
+}
+
+/// IterateOp implemented OpInterfaces' methods.
+SmallVector<Region *> IterateOp::getLoopRegions() { return {&getRegion()}; }
+
+MutableArrayRef<OpOperand> IterateOp::getInitsMutable() {
+ return getInitArgsMutable();
+}
+
+Block::BlockArgListType IterateOp::getRegionIterArgs() {
+ return getRegion().getArguments().take_back(getNumRegionIterArgs());
+}
+
+std::optional<MutableArrayRef<OpOperand>> IterateOp::getYieldedValuesMutable() {
+ return cast<sparse_tensor::YieldOp>(
+ getRegion().getBlocks().front().getTerminator())
+ .getResultsMutable();
+}
+
+std::optional<ResultRange> IterateOp::getLoopResults() { return getResults(); }
+
+OperandRange IterateOp::getEntrySuccessorOperands(RegionBranchPoint point) {
+ return getInitArgs();
+}
+
+void IterateOp::getSuccessorRegions(RegionBranchPoint point,
+ SmallVectorImpl<RegionSuccessor> ®ions) {
+ // Both the operation itself and the region may be branching into the body or
+ // back into the operation itself.
+ regions.push_back(RegionSuccessor(&getRegion(), getRegionIterArgs()));
+ // It is possible for loop not to enter the body.
+ regions.push_back(RegionSuccessor(getResults()));
+}
+
+//===----------------------------------------------------------------------===//
+// Sparse Tensor Dialect Setups.
+//===----------------------------------------------------------------------===//
+
/// Materialize a single constant operation from a given attribute value with
/// the desired resultant type.
Operation *SparseTensorDialect::materializeConstant(OpBuilder &builder,
diff --git a/mlir/test/Dialect/SparseTensor/invalid.mlir b/mlir/test/Dialect/SparseTensor/invalid.mlir
index 3fa696e1600a93..b13024cd4ed99d 100644
--- a/mlir/test/Dialect/SparseTensor/invalid.mlir
+++ b/mlir/test/Dialect/SparseTensor/invalid.mlir
@@ -1094,3 +1094,60 @@ func.func @sparse_extract_iter_space(%sp : tensor<4x8xf32, #COO>, %it1 : !sparse
%l1 = sparse_tensor.extract_iteration_space %sp at %it1 lvls = 2 : tensor<4x8xf32, #COO>, !sparse_tensor.iterator<#COO, lvls = 0>
return
}
+
+
+// -----
+
+#COO = #sparse_tensor.encoding<{
+ map = (i, j) -> (
+ i : compressed(nonunique),
+ j : singleton(soa)
+ )
+}>
+
+func.func @sparse_iterate(%sp : tensor<4x8xf32, #COO>, %i : index, %j : index) -> index {
+ %l1 = sparse_tensor.extract_iteration_space %sp lvls = 0 : tensor<4x8xf32, #COO>
+ // expected-error @+1 {{'sparse_tensor.iterate' op different number o...
[truncated]
|
PeimingLiu
force-pushed
the
cherry-picks-1
branch
from
44c02a1 to
7db93fc
Compare
PeimingLiu
force-pushed
the
cherry-picks-1
branch
from
7db93fc to
51be8a3
Compare
aartbik
reviewed
Apr 17, 2024
mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorAttrDefs.td
Outdated
Show resolved
Hide resolved
aartbik
reviewed
Apr 24, 2024
PeimingLiu
force-pushed
the
cherry-picks-1
branch
3 times, most recently
from
9c4a33f to
2a5fe96
Compare
aartbik
approved these changes
May 24, 2024
PeimingLiu
force-pushed
the
cherry-picks-1
branch
from
2a5fe96 to
08b970d
Compare
|
You're missing some CMake dependency update, the mlir-nvidia bot is broken: |
Thanks! working on it! |
PeimingLiu
pushed a commit
that referenced
this pull request
Jun 10, 2024
|
This is breaking some buildbots! Please revert it if you cannot fix it quickly. |
Should have been fixed by #95034 |
Lukacma
pushed a commit
to Lukacma/llvm-project
that referenced
this pull request
Jun 12, 2024
A `sparse_tensor.iterate` iterates over a sparse iteration space extracted from `sparse_tensor.extract_iteration_space` operation introduced in llvm#88554.
Lukacma
pushed a commit
to Lukacma/llvm-project
that referenced
this pull request
Jun 12, 2024
Closed
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
A
sparse_tensor.iterateiterates over a sparse iteration space extracted fromsparse_tensor.extract_iteration_spaceoperation introduced in #88554.