[Mlir-commits] [mlir] [mlir][sparse] implement sparse space collapse pass. (PR #89003)

[Ingo Müller]

================
@@ -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);
----------------
ingomueller-net wrote:

I am wondering whether this couldn't be a pattern. I guess the answer is: currently no because we want to collect all collapsible loops in one go.

Eventually, you might want to be able to merge an already collapsed iteration with the next collapsible level. In that case, you could solve the special case of only non-collapsed ops by repeatedly applying the previous logic. Maybe *that* could work as patterns then...?

https://github.com/llvm/llvm-project/pull/89003