[Mlir-commits] [mlir] [mlir][sparse] implement sparse space collapse pass. (PR #89003)
Peiming Liu
llvmlistbot at llvm.org
Wed Apr 17 11:27:47 PDT 2024
https://github.com/PeimingLiu updated https://github.com/llvm/llvm-project/pull/89003
>From 7db93fc5e9fab3de3b4805d56066cfdbf74895b4 Mon Sep 17 00:00:00 2001
From: Peiming Liu <peiming at google.com>
Date: Fri, 12 Apr 2024 22:03:06 +0000
Subject: [PATCH 1/2] [mlir][sparse] introduce `sparse_tensor.iterate`
operation
---
.../Dialect/SparseTensor/IR/SparseTensor.h | 38 +++
.../SparseTensor/IR/SparseTensorAttrDefs.td | 15 ++
.../SparseTensor/IR/SparseTensorOps.td | 101 ++++++-
.../SparseTensor/IR/SparseTensorDialect.cpp | 255 ++++++++++++++++++
mlir/test/Dialect/SparseTensor/invalid.mlir | 57 ++++
mlir/test/Dialect/SparseTensor/roundtrip.mlir | 28 ++
.../SparseTensor/sparse_itertion_licm.mlir | 26 ++
7 files changed, 519 insertions(+), 1 deletion(-)
create mode 100644 mlir/test/Dialect/SparseTensor/sparse_itertion_licm.mlir
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 4e4441c640ed95..08140b9d2b6192 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.
@@ -1304,7 +1306,7 @@ def SparseTensor_SelectOp : SparseTensor_Op<"select", [Pure, SameOperandsAndResu
def SparseTensor_YieldOp : SparseTensor_Op<"yield", [Pure, Terminator,
ParentOneOf<["BinaryOp", "UnaryOp", "ReduceOp", "SelectOp",
- "ForeachOp"]>]> {
+ "ForeachOp", "IterateOp"]>]> {
let summary = "Yield from sparse_tensor set-like operations";
let description = [{
Yields a value from within a `binary`, `unary`, `reduce`,
@@ -1513,6 +1515,103 @@ def ExtractIterSpaceOp : SparseTensor_Op<"extract_iteration_space",
let hasVerifier = 1;
}
+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 of region iter_args and yielded values: 2 != 1}}
+ %r1, %r2 = sparse_tensor.iterate %it1 in %l1 at (%crd) iter_args(%si = %i, %sj = %j): !sparse_tensor.iter_space<#COO, lvls = 0> -> (index, index) {
+ sparse_tensor.yield %si : index
+ }
+ return %r1 : index
+}
+
+// -----
+
+#COO = #sparse_tensor.encoding<{
+ map = (i, j) -> (
+ i : compressed(nonunique),
+ j : singleton(soa)
+ )
+}>
+
+// expected-note at +1 {{prior use here}}
+func.func @sparse_iterate(%sp : tensor<4x8xf32, #COO>, %i : index) -> f32 {
+ %l1 = sparse_tensor.extract_iteration_space %sp lvls = 0 : tensor<4x8xf32, #COO>
+ // expected-error @+1 {{use of value '%i' expects different type than prior uses: 'f32' vs 'index'}}
+ %r1 = sparse_tensor.iterate %it1 in %l1 at (%crd) iter_args(%outer = %i): !sparse_tensor.iter_space<#COO, lvls = 0> -> f32 {
+ sparse_tensor.yield %outer : f32
+ }
+ return %r1 : f32
+}
+
+// -----
+
+#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 0-th region iter_arg and 0-th yielded value have different type: 'index' != 'f32'}}
+ %r1 = sparse_tensor.iterate %it1 in %l1 at (%crd) iter_args(%si = %i): !sparse_tensor.iter_space<#COO, lvls = 0> -> index {
+ %y = arith.constant 1.0 : f32
+ sparse_tensor.yield %y : f32
+ }
+ return %r1 : index
+}
diff --git a/mlir/test/Dialect/SparseTensor/roundtrip.mlir b/mlir/test/Dialect/SparseTensor/roundtrip.mlir
index d34071279e5129..e9a898f16b41d2 100644
--- a/mlir/test/Dialect/SparseTensor/roundtrip.mlir
+++ b/mlir/test/Dialect/SparseTensor/roundtrip.mlir
@@ -763,3 +763,31 @@ func.func @sparse_extract_iter_space(%sp : tensor<4x8xf32, #COO>, %it1 : !sparse
%l2 = sparse_tensor.extract_iteration_space %sp at %it1 lvls = 1 : tensor<4x8xf32, #COO>, !sparse_tensor.iterator<#COO, lvls = 0>
return %l1, %l2 : !sparse_tensor.iter_space<#COO, lvls = 0>, !sparse_tensor.iter_space<#COO, lvls = 1>
}
+
+
+// -----
+
+#COO = #sparse_tensor.encoding<{
+ map = (i, j) -> (
+ i : compressed(nonunique),
+ j : singleton(soa)
+ )
+}>
+
+// CHECK-LABEL: func.func @sparse_iterate(
+// CHECK-SAME: %[[VAL_0:.*]]: tensor<4x8xf32, #sparse{{[0-9]*}}>,
+// CHECK-SAME: %[[VAL_1:.*]]: index,
+// CHECK-SAME: %[[VAL_2:.*]]: index) -> index {
+// CHECK: %[[VAL_3:.*]] = sparse_tensor.extract_iteration_space %[[VAL_0]] lvls = 0 : tensor<4x8xf32, #sparse{{[0-9]*}}>
+// CHECK: %[[VAL_4:.*]] = sparse_tensor.iterate %[[VAL_5:.*]] in %[[VAL_3]] at(%[[VAL_6:.*]]) iter_args(%[[VAL_7:.*]] = %[[VAL_1]]) : !sparse_tensor.iter_space<#sparse{{[0-9]*}}, lvls = 0> -> (index) {
+// CHECK: sparse_tensor.yield %[[VAL_7]] : index
+// CHECK: }
+// CHECK: return %[[VAL_4]] : index
+// CHECK: }
+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>
+ %r1 = sparse_tensor.iterate %it1 in %l1 at (%crd) iter_args(%outer = %i): !sparse_tensor.iter_space<#COO, lvls = 0 to 1> -> index {
+ sparse_tensor.yield %outer : index
+ }
+ return %r1 : index
+}
diff --git a/mlir/test/Dialect/SparseTensor/sparse_itertion_licm.mlir b/mlir/test/Dialect/SparseTensor/sparse_itertion_licm.mlir
new file mode 100644
index 00000000000000..e7158d04b37feb
--- /dev/null
+++ b/mlir/test/Dialect/SparseTensor/sparse_itertion_licm.mlir
@@ -0,0 +1,26 @@
+// RUN: mlir-opt %s --loop-invariant-code-motion | FileCheck %s
+
+#CSR = #sparse_tensor.encoding<{
+ map = (i, j) -> (
+ i : dense,
+ j : compressed
+ )
+}>
+
+// Make sure that pure instructions are hoisted outside the loop.
+//
+// CHECK: sparse_tensor.values
+// CHECK: sparse_tensor.positions
+// CHECK: sparse_tensor.coordinate
+// CHECK: sparse_tensor.iterate
+func.func @sparse_iterate(%sp : tensor<?x?xf64, #CSR>) {
+ %l1 = sparse_tensor.extract_iteration_space %sp lvls = 0 : tensor<?x?xf64, #CSR>
+ sparse_tensor.iterate %it1 in %l1 at (%crd) : !sparse_tensor.iter_space<#CSR, lvls = 0> {
+ %0 = sparse_tensor.values %sp : tensor<?x?xf64, #CSR> to memref<?xf64>
+ %1 = sparse_tensor.positions %sp { level = 1 : index } : tensor<?x?xf64, #CSR> to memref<?xindex>
+ %2 = sparse_tensor.coordinates %sp { level = 1 : index } : tensor<?x?xf64, #CSR> to memref<?xindex>
+ "test.op"(%0, %1, %2) : (memref<?xf64>, memref<?xindex>, memref<?xindex>) -> ()
+ }
+
+ return
+}
>From b0f02a64a503283c063f14acb924ab3ce5c93a6e Mon Sep 17 00:00:00 2001
From: Peiming Liu <peiming at google.com>
Date: Tue, 16 Apr 2024 17:00:45 +0000
Subject: [PATCH 2/2] [mlir][sparse] implement sparse space collapse pass.
---
.../Dialect/SparseTensor/Transforms/Passes.h | 6 +
.../Dialect/SparseTensor/Transforms/Passes.td | 16 ++
.../SparseTensor/Transforms/CMakeLists.txt | 1 +
.../Transforms/SparseSpaceCollapse.cpp | 183 ++++++++++++++++++
.../SparseTensor/sparse_space_collapse.mlir | 33 ++++
5 files changed, 239 insertions(+)
create mode 100644 mlir/lib/Dialect/SparseTensor/Transforms/SparseSpaceCollapse.cpp
create mode 100644 mlir/test/Dialect/SparseTensor/sparse_space_collapse.mlir
diff --git a/mlir/include/mlir/Dialect/SparseTensor/Transforms/Passes.h b/mlir/include/mlir/Dialect/SparseTensor/Transforms/Passes.h
index d6d038ef65bdf4..3043a0c4dc4109 100644
--- a/mlir/include/mlir/Dialect/SparseTensor/Transforms/Passes.h
+++ b/mlir/include/mlir/Dialect/SparseTensor/Transforms/Passes.h
@@ -248,6 +248,12 @@ std::unique_ptr<Pass> createSparsificationAndBufferizationPass(
bool enableBufferInitialization, unsigned vectorLength,
bool enableVLAVectorization, bool enableSIMDIndex32, bool enableGPULibgen);
+//===----------------------------------------------------------------------===//
+// Sparse Iteration Transform Passes
+//===----------------------------------------------------------------------===//
+
+std::unique_ptr<Pass> createSparseSpaceCollapsePass();
+
//===----------------------------------------------------------------------===//
// Registration.
//===----------------------------------------------------------------------===//
diff --git a/mlir/include/mlir/Dialect/SparseTensor/Transforms/Passes.td b/mlir/include/mlir/Dialect/SparseTensor/Transforms/Passes.td
index 4706d5ba2f218c..2281807d298ed5 100644
--- a/mlir/include/mlir/Dialect/SparseTensor/Transforms/Passes.td
+++ b/mlir/include/mlir/Dialect/SparseTensor/Transforms/Passes.td
@@ -463,4 +463,20 @@ def SparsificationAndBufferization : Pass<"sparsification-and-bufferization", "M
];
}
+//===----------------------------------------------------------------------===//
+// Sparse Iteration Transform Passes
+//===----------------------------------------------------------------------===//
+
+def SparseSpaceCollapse : Pass<"sparse-space-collapse", "func::FuncOp"> {
+ let summary = "(experimental) sparse space collpasing pass";
+ let description = [{
+ This pass collapse consecutive sparse spaces (extracted from the same tensor)
+ into one multi-dimensional space. The pass is not yet stablized.
+ }];
+ let constructor = "mlir::createSparseSpaceCollapsePass()";
+ let dependentDialects = [
+ "sparse_tensor::SparseTensorDialect",
+ ];
+}
+
#endif // MLIR_DIALECT_SPARSETENSOR_TRANSFORMS_PASSES
diff --git a/mlir/lib/Dialect/SparseTensor/Transforms/CMakeLists.txt b/mlir/lib/Dialect/SparseTensor/Transforms/CMakeLists.txt
index af3a1b48f45af9..2a29ee8a7a87cb 100644
--- a/mlir/lib/Dialect/SparseTensor/Transforms/CMakeLists.txt
+++ b/mlir/lib/Dialect/SparseTensor/Transforms/CMakeLists.txt
@@ -6,6 +6,7 @@ add_mlir_dialect_library(MLIRSparseTensorTransforms
SparseGPUCodegen.cpp
SparseReinterpretMap.cpp
SparseStorageSpecifierToLLVM.cpp
+ SparseSpaceCollapse.cpp
SparseTensorCodegen.cpp
SparseTensorConversion.cpp
SparseTensorPasses.cpp
diff --git a/mlir/lib/Dialect/SparseTensor/Transforms/SparseSpaceCollapse.cpp b/mlir/lib/Dialect/SparseTensor/Transforms/SparseSpaceCollapse.cpp
new file mode 100644
index 00000000000000..bc469992d97103
--- /dev/null
+++ b/mlir/lib/Dialect/SparseTensor/Transforms/SparseSpaceCollapse.cpp
@@ -0,0 +1,183 @@
+//===--------- SparseSpaceCollapse.cpp - Collapse Sparse Space Pass -------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/IR/IRMapping.h"
+#include "mlir/Transforms/Passes.h"
+
+#include "mlir/Dialect/SparseTensor/IR/SparseTensor.h"
+#include "mlir/Dialect/SparseTensor/Transforms/Passes.h"
+
+namespace mlir {
+
+#define GEN_PASS_DEF_SPARSESPACECOLLAPSE
+#include "mlir/Dialect/SparseTensor/Transforms/Passes.h.inc"
+
+namespace sparse_tensor {
+
+struct CollapseSpaceInfo {
+ ExtractIterSpaceOp space;
+ IterateOp loop;
+};
+
+bool isCollapsableLoops(LoopLikeOpInterface parent, LoopLikeOpInterface node) {
+ auto pIterArgs = parent.getRegionIterArgs();
+ auto nInitArgs = node.getInits();
+ if (pIterArgs.size() != nInitArgs.size())
+ return false;
+
+ // Two loops are collapsable if they are perfectly nested.
+ auto pYields = parent.getYieldedValues();
+ auto nResult = node.getLoopResults().value();
+
+ bool yieldEq =
+ llvm::all_of(llvm::zip_equal(pYields, nResult), [](auto zipped) {
+ return std::get<0>(zipped) == std::get<1>(zipped);
+ });
+
+ // Parent iter_args should be passed directly to the node's init_args.
+ bool iterArgEq =
+ llvm::all_of(llvm::zip_equal(pIterArgs, nInitArgs), [](auto zipped) {
+ return std::get<0>(zipped) == std::get<1>(zipped);
+ });
+
+ return yieldEq && iterArgEq;
+}
+
+bool legalToCollapse(SmallVectorImpl<CollapseSpaceInfo> &toCollapse,
+ ExtractIterSpaceOp curSpace) {
+
+ auto getIterateOpOverSpace = [](ExtractIterSpaceOp space) -> IterateOp {
+ Value spaceVal = space.getResultSpace();
+ if (spaceVal.hasOneUse())
+ return llvm::dyn_cast<IterateOp>(*spaceVal.getUsers().begin());
+ return nullptr;
+ };
+
+ if (toCollapse.empty()) {
+ // Collapse root.
+ if (auto itOp = getIterateOpOverSpace(curSpace)) {
+ CollapseSpaceInfo &info = toCollapse.emplace_back();
+ info.space = curSpace;
+ info.loop = itOp;
+ return true;
+ }
+ return false;
+ }
+
+ auto parent = toCollapse.back().space;
+ auto pItOp = toCollapse.back().loop;
+ auto nItOp = getIterateOpOverSpace(curSpace);
+
+ // Can only collapse spaces extracted from the same tensor.
+ if (parent.getTensor() != curSpace.getTensor())
+ return false;
+
+ // Can only collapse consecutive simple iteration on one tensor (i.e., no
+ // coiteration).
+ if (!nItOp || nItOp->getBlock() != curSpace->getBlock() ||
+ pItOp.getIterator() != curSpace.getParentIter() ||
+ curSpace->getParentOp() != pItOp.getOperation())
+ return false;
+
+ if (pItOp && !isCollapsableLoops(pItOp, nItOp))
+ return false;
+
+ CollapseSpaceInfo &info = toCollapse.emplace_back();
+ info.space = curSpace;
+ info.loop = nItOp;
+ return true;
+}
+
+void collapseSparseSpace(MutableArrayRef<CollapseSpaceInfo> toCollapse) {
+ if (toCollapse.size() < 2)
+ return;
+
+ ExtractIterSpaceOp root = toCollapse.front().space;
+ ExtractIterSpaceOp leaf = toCollapse.back().space;
+ Location loc = root.getLoc();
+
+ assert(root->hasOneUse() && leaf->hasOneUse());
+
+ // Insert collapsed operation at the same scope as root operation.
+ OpBuilder builder(root);
+
+ // Construct the collapsed iteration space.
+ auto collapsedSpace = builder.create<ExtractIterSpaceOp>(
+ loc, root.getTensor(), root.getParentIter(), root.getLoLvl(),
+ leaf.getHiLvl());
+
+ auto rItOp = llvm::cast<IterateOp>(*root->getUsers().begin());
+ auto innermost = toCollapse.back().loop;
+
+ IRMapping mapper;
+ mapper.map(leaf, collapsedSpace.getResultSpace());
+ for (auto z : llvm::zip_equal(innermost.getInitArgs(), rItOp.getInitArgs()))
+ mapper.map(std::get<0>(z), std::get<1>(z));
+
+ auto cloned = llvm::cast<IterateOp>(builder.clone(*innermost, mapper));
+ builder.setInsertionPointToStart(cloned.getBody());
+
+ LevelSet crdUsedLvls;
+ unsigned shift = 0, argIdx = 1;
+ for (auto info : toCollapse.drop_back()) {
+ LevelSet set = info.loop.getCrdUsedLvls();
+ crdUsedLvls |= set.lshift(shift);
+ shift += info.loop.getSpaceDim();
+ for (BlockArgument crd : info.loop.getCrds()) {
+ BlockArgument collapsedCrd = cloned.getBody()->insertArgument(
+ argIdx++, builder.getIndexType(), crd.getLoc());
+ crd.replaceAllUsesWith(collapsedCrd);
+ }
+ }
+ crdUsedLvls |= innermost.getCrdUsedLvls().lshift(shift);
+ cloned.getIterator().setType(collapsedSpace.getType().getIteratorType());
+ cloned.setCrdUsedLvls(crdUsedLvls);
+
+ rItOp.replaceAllUsesWith(cloned.getResults());
+ // Erase collapsed loops.
+ rItOp.erase();
+ root.erase();
+}
+
+struct SparseSpaceCollapsePass
+ : public impl::SparseSpaceCollapseBase<SparseSpaceCollapsePass> {
+ SparseSpaceCollapsePass() = default;
+
+ void runOnOperation() override {
+ func::FuncOp func = getOperation();
+
+ // A naive (experimental) implementation to collapse consecutive sparse
+ // spaces. It does NOT handle complex cases where multiple spaces are
+ // extracted in the same basic block. E.g.,
+ //
+ // %space1 = extract_space %t1 ...
+ // %space2 = extract_space %t2 ...
+ // sparse_tensor.iterate(%sp1) ...
+ //
+ SmallVector<CollapseSpaceInfo> toCollapse;
+ func->walk([&](ExtractIterSpaceOp op) {
+ if (!legalToCollapse(toCollapse, op)) {
+ // if not legal to collapse one more space, collapse the existing ones
+ // and clear.
+ collapseSparseSpace(toCollapse);
+ toCollapse.clear();
+ }
+ });
+
+ collapseSparseSpace(toCollapse);
+ }
+};
+
+} // namespace sparse_tensor
+
+std::unique_ptr<Pass> createSparseSpaceCollapsePass() {
+ return std::make_unique<sparse_tensor::SparseSpaceCollapsePass>();
+}
+
+} // namespace mlir
diff --git a/mlir/test/Dialect/SparseTensor/sparse_space_collapse.mlir b/mlir/test/Dialect/SparseTensor/sparse_space_collapse.mlir
new file mode 100644
index 00000000000000..392dfe01884ba8
--- /dev/null
+++ b/mlir/test/Dialect/SparseTensor/sparse_space_collapse.mlir
@@ -0,0 +1,33 @@
+// RUN: mlir-opt %s --sparse-space-collapse | FileCheck %s
+
+#COO = #sparse_tensor.encoding<{
+ map = (i, j) -> (
+ i : compressed(nonunique),
+ j : singleton(soa)
+ )
+}>
+
+// CHECK-LABEL: func.func @sparse_sparse_collapse(
+// CHECK-SAME: %[[VAL_0:.*]]: tensor<4x8xf32, #sparse>,
+// CHECK-SAME: %[[VAL_1:.*]]: index) {
+// CHECK: %[[VAL_3:.*]] = sparse_tensor.extract_iteration_space %[[VAL_0]] lvls = 0 to 2 : tensor<4x8xf32, #sparse>
+// CHECK: %[[VAL_4:.*]] = sparse_tensor.iterate %[[VAL_5:.*]] in %[[VAL_3]] at(%[[VAL_6:.*]], _) iter_args(%[[VAL_7:.*]] = %[[VAL_1]])
+// CHECK: %[[VAL_8:.*]] = "test.op"(%[[VAL_7]]) : (index) -> index
+// CHECK: sparse_tensor.yield %[[VAL_8]] : index
+// CHECK: }
+// CHECK: "test.sink"(%[[VAL_4]]) : (index) -> ()
+// CHECK: return
+// CHECK: }
+func.func @sparse_sparse_collapse(%sp : tensor<4x8xf32, #COO>, %i : index) {
+ %l1 = sparse_tensor.extract_iteration_space %sp lvls = 0 : tensor<4x8xf32, #COO>
+ %r1 = sparse_tensor.iterate %it1 in %l1 at(%crd0) iter_args(%outer = %i): !sparse_tensor.iter_space<#COO, lvls = 0 to 1> -> index {
+ %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 iter_args(%inner = %outer): !sparse_tensor.iter_space<#COO, lvls = 1 to 2> -> index {
+ %k ="test.op"(%inner) : (index) -> index
+ sparse_tensor.yield %k : index
+ }
+ sparse_tensor.yield %r2 : index
+ }
+ "test.sink"(%r1) : (index) -> ()
+ return
+}
More information about the Mlir-commits
mailing list