[Mlir-commits] [mlir] [mlir][affine] Add pass --affine-raise-from-memref (2) (PR #138004)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Wed Apr 30 11:03:16 PDT 2025
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-mlir
Author: Clément Fournier (oowekyala)
<details>
<summary>Changes</summary>
Note: this is a reopening of #<!-- -->114032, which at the time had been approved but not merged. I inadvertently closed it last month and am just realizing.
CC @<!-- -->ftynse
--
This adds a pass that converts memref.load/store into affine.load/store. This is useful as those memref operators are ignored by passes like --affine-scalrep as they don't implement the Affine[Read/Write]OpInterface. Doing this allows you to put as much of your program in affine form before you apply affine optimization passes.
This also slightly changes the implementation of affine::isValidDim. The previous implementation allowed values from the iter_args of affine loops to be used as valid dims. I think this doesn't make sense and what was meant is just the induction vars. In the real world, there is little reason to find an index in the iter_args, but I wrote that in my tests and found out it was treated as an affine dim, so corrected that.
---
Full diff: https://github.com/llvm/llvm-project/pull/138004.diff
6 Files Affected:
- (modified) mlir/include/mlir/Dialect/Affine/Passes.h (+6)
- (modified) mlir/include/mlir/Dialect/Affine/Passes.td (+12)
- (modified) mlir/lib/Dialect/Affine/IR/AffineOps.cpp (+7-6)
- (modified) mlir/lib/Dialect/Affine/Transforms/CMakeLists.txt (+1)
- (added) mlir/lib/Dialect/Affine/Transforms/RaiseMemrefDialect.cpp (+187)
- (added) mlir/test/Dialect/Affine/raise-memref.mlir (+138)
``````````diff
diff --git a/mlir/include/mlir/Dialect/Affine/Passes.h b/mlir/include/mlir/Dialect/Affine/Passes.h
index e152101236dc7..c1b9c30d302dd 100644
--- a/mlir/include/mlir/Dialect/Affine/Passes.h
+++ b/mlir/include/mlir/Dialect/Affine/Passes.h
@@ -22,6 +22,9 @@ namespace mlir {
namespace func {
class FuncOp;
} // namespace func
+namespace memref {
+class MemRefDialect;
+} // namespace memref
namespace affine {
class AffineForOp;
@@ -48,6 +51,9 @@ createAffineLoopInvariantCodeMotionPass();
/// ops.
std::unique_ptr<OperationPass<func::FuncOp>> createAffineParallelizePass();
+/// Creates a pass that converts some memref operators to affine operators.
+std::unique_ptr<OperationPass<func::FuncOp>> createRaiseMemrefToAffine();
+
/// Apply normalization transformations to affine loop-like ops. If
/// `promoteSingleIter` is true, single iteration loops are promoted (i.e., the
/// loop is replaced by its loop body).
diff --git a/mlir/include/mlir/Dialect/Affine/Passes.td b/mlir/include/mlir/Dialect/Affine/Passes.td
index 77073aa29da73..a77bcac5ed407 100644
--- a/mlir/include/mlir/Dialect/Affine/Passes.td
+++ b/mlir/include/mlir/Dialect/Affine/Passes.td
@@ -397,6 +397,18 @@ def LoopCoalescing : Pass<"affine-loop-coalescing", "func::FuncOp"> {
let dependentDialects = ["affine::AffineDialect","arith::ArithDialect"];
}
+def RaiseMemrefDialect : Pass<"affine-raise-from-memref", "func::FuncOp"> {
+ let summary = "Turn some memref operators to affine operators where supported";
+ let description = [{
+ Raise memref.load and memref.store to affine.store and affine.load, inferring
+ the affine map of those operators if needed. This allows passes like --affine-scalrep
+ to optimize those loads and stores (forwarding them or eliminating them).
+ They can be turned back to memref dialect ops with --lower-affine.
+ }];
+ let constructor = "mlir::affine::createRaiseMemrefToAffine()";
+ let dependentDialects = ["affine::AffineDialect"];
+}
+
def SimplifyAffineStructures : Pass<"affine-simplify-structures", "func::FuncOp"> {
let summary = "Simplify affine expressions in maps/sets and normalize "
"memrefs";
diff --git a/mlir/lib/Dialect/Affine/IR/AffineOps.cpp b/mlir/lib/Dialect/Affine/IR/AffineOps.cpp
index dceebbfec586c..06204188e14e2 100644
--- a/mlir/lib/Dialect/Affine/IR/AffineOps.cpp
+++ b/mlir/lib/Dialect/Affine/IR/AffineOps.cpp
@@ -284,10 +284,12 @@ bool mlir::affine::isValidDim(Value value) {
return isValidDim(value, getAffineScope(defOp));
// This value has to be a block argument for an op that has the
- // `AffineScope` trait or for an affine.for or affine.parallel.
+ // `AffineScope` trait or an induction var of an affine.for or
+ // affine.parallel.
+ if (isAffineInductionVar(value))
+ return true;
auto *parentOp = llvm::cast<BlockArgument>(value).getOwner()->getParentOp();
- return parentOp && (parentOp->hasTrait<OpTrait::AffineScope>() ||
- isa<AffineForOp, AffineParallelOp>(parentOp));
+ return parentOp && parentOp->hasTrait<OpTrait::AffineScope>();
}
// Value can be used as a dimension id iff it meets one of the following
@@ -306,10 +308,9 @@ bool mlir::affine::isValidDim(Value value, Region *region) {
auto *op = value.getDefiningOp();
if (!op) {
- // This value has to be a block argument for an affine.for or an
+ // This value has to be an induction var for an affine.for or an
// affine.parallel.
- auto *parentOp = llvm::cast<BlockArgument>(value).getOwner()->getParentOp();
- return isa<AffineForOp, AffineParallelOp>(parentOp);
+ return isAffineInductionVar(value);
}
// Affine apply operation is ok if all of its operands are ok.
diff --git a/mlir/lib/Dialect/Affine/Transforms/CMakeLists.txt b/mlir/lib/Dialect/Affine/Transforms/CMakeLists.txt
index c42789b01bc9f..1c82822b2bd7f 100644
--- a/mlir/lib/Dialect/Affine/Transforms/CMakeLists.txt
+++ b/mlir/lib/Dialect/Affine/Transforms/CMakeLists.txt
@@ -13,6 +13,7 @@ add_mlir_dialect_library(MLIRAffineTransforms
LoopUnroll.cpp
LoopUnrollAndJam.cpp
PipelineDataTransfer.cpp
+ RaiseMemrefDialect.cpp
ReifyValueBounds.cpp
SuperVectorize.cpp
SimplifyAffineStructures.cpp
diff --git a/mlir/lib/Dialect/Affine/Transforms/RaiseMemrefDialect.cpp b/mlir/lib/Dialect/Affine/Transforms/RaiseMemrefDialect.cpp
new file mode 100644
index 0000000000000..491d2e03c36bc
--- /dev/null
+++ b/mlir/lib/Dialect/Affine/Transforms/RaiseMemrefDialect.cpp
@@ -0,0 +1,187 @@
+//===- RaiseMemrefDialect.cpp - raise memref.store and load to affine ops -===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements functionality to convert memref load and store ops to
+// the corresponding affine ops, inferring the affine map as needed.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Affine/Analysis/Utils.h"
+#include "mlir/Dialect/Affine/Passes.h"
+#include "mlir/Dialect/Affine/Transforms/Transforms.h"
+#include "mlir/Dialect/Affine/Utils.h"
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/IR/AffineExpr.h"
+#include "mlir/IR/Matchers.h"
+#include "mlir/IR/Operation.h"
+#include "mlir/Pass/Pass.h"
+#include "llvm/Support/Casting.h"
+#include "llvm/Support/Debug.h"
+
+namespace mlir {
+namespace affine {
+#define GEN_PASS_DEF_RAISEMEMREFDIALECT
+#include "mlir/Dialect/Affine/Passes.h.inc"
+} // namespace affine
+} // namespace mlir
+
+#define DEBUG_TYPE "raise-memref-to-affine"
+
+using namespace mlir;
+using namespace mlir::affine;
+
+namespace {
+
+/// Find the index of the given value in the `dims` list,
+/// and append it if it was not already in the list. The
+/// dims list is a list of symbols or dimensions of the
+/// affine map. Within the results of an affine map, they
+/// are identified by their index, which is why we need
+/// this function.
+static std::optional<size_t>
+findInListOrAdd(Value value, llvm::SmallVectorImpl<Value> &dims,
+ function_ref<bool(Value)> isValidElement) {
+
+ Value *loopIV = std::find(dims.begin(), dims.end(), value);
+ if (loopIV != dims.end()) {
+ // We found an IV that already has an index, return that index.
+ return {std::distance(dims.begin(), loopIV)};
+ }
+ if (isValidElement(value)) {
+ // This is a valid element for the dim/symbol list, push this as a
+ // parameter.
+ size_t idx = dims.size();
+ dims.push_back(value);
+ return idx;
+ }
+ return std::nullopt;
+}
+
+/// Convert a value to an affine expr if possible. Adds dims and symbols
+/// if needed.
+static AffineExpr toAffineExpr(Value value,
+ llvm::SmallVectorImpl<Value> &affineDims,
+ llvm::SmallVectorImpl<Value> &affineSymbols) {
+ using namespace matchers;
+ IntegerAttr::ValueType cst;
+ if (matchPattern(value, m_ConstantInt(&cst))) {
+ return getAffineConstantExpr(cst.getSExtValue(), value.getContext());
+ }
+
+ Operation *definingOp = value.getDefiningOp();
+ if (llvm::isa_and_nonnull<arith::AddIOp>(definingOp) ||
+ llvm::isa_and_nonnull<arith::MulIOp>(definingOp)) {
+ // TODO: replace recursion with explicit stack.
+ // For the moment this can be tolerated as we only recurse on
+ // arith.addi and arith.muli, so there cannot be any infinite
+ // recursion. The depth of these expressions should be in most
+ // cases very manageable, as affine expressions should be as
+ // simple as `a + b * c`.
+ AffineExpr lhsE =
+ toAffineExpr(definingOp->getOperand(0), affineDims, affineSymbols);
+ AffineExpr rhsE =
+ toAffineExpr(definingOp->getOperand(1), affineDims, affineSymbols);
+
+ if (lhsE && rhsE) {
+ AffineExprKind kind;
+ if (isa<arith::AddIOp>(definingOp)) {
+ kind = mlir::AffineExprKind::Add;
+ } else {
+ kind = mlir::AffineExprKind::Mul;
+
+ if (!lhsE.isSymbolicOrConstant() && !rhsE.isSymbolicOrConstant()) {
+ // This is not an affine expression, give up.
+ return {};
+ }
+ }
+ return getAffineBinaryOpExpr(kind, lhsE, rhsE);
+ }
+ return {};
+ }
+
+ if (auto dimIx = findInListOrAdd(value, affineSymbols, [](Value v) {
+ return affine::isValidSymbol(v);
+ })) {
+ return getAffineSymbolExpr(*dimIx, value.getContext());
+ }
+
+ if (auto dimIx = findInListOrAdd(
+ value, affineDims, [](Value v) { return affine::isValidDim(v); })) {
+
+ return getAffineDimExpr(*dimIx, value.getContext());
+ }
+
+ return {};
+}
+
+static LogicalResult
+computeAffineMapAndArgs(MLIRContext *ctx, ValueRange indices, AffineMap &map,
+ llvm::SmallVectorImpl<Value> &mapArgs) {
+ SmallVector<AffineExpr> results;
+ SmallVector<Value> symbols;
+ SmallVector<Value> dims;
+
+ for (Value indexExpr : indices) {
+ AffineExpr res = toAffineExpr(indexExpr, dims, symbols);
+ if (!res) {
+ return failure();
+ }
+ results.push_back(res);
+ }
+
+ map = AffineMap::get(dims.size(), symbols.size(), results, ctx);
+
+ dims.append(symbols);
+ mapArgs.swap(dims);
+ return success();
+}
+
+struct RaiseMemrefDialect
+ : public affine::impl::RaiseMemrefDialectBase<RaiseMemrefDialect> {
+
+ void runOnOperation() override {
+ auto *ctx = &getContext();
+ Operation *op = getOperation();
+ IRRewriter rewriter(ctx);
+ AffineMap map;
+ SmallVector<Value> mapArgs;
+ op->walk([&](Operation *op) {
+ rewriter.setInsertionPoint(op);
+ if (auto store = llvm::dyn_cast_or_null<memref::StoreOp>(op)) {
+
+ if (succeeded(computeAffineMapAndArgs(ctx, store.getIndices(), map,
+ mapArgs))) {
+ rewriter.replaceOpWithNewOp<AffineStoreOp>(
+ op, store.getValueToStore(), store.getMemRef(), map, mapArgs);
+ return;
+ }
+
+ LLVM_DEBUG(llvm::dbgs()
+ << "[affine] Cannot raise memref op: " << op << "\n");
+
+ } else if (auto load = llvm::dyn_cast_or_null<memref::LoadOp>(op)) {
+ if (succeeded(computeAffineMapAndArgs(ctx, load.getIndices(), map,
+ mapArgs))) {
+ rewriter.replaceOpWithNewOp<AffineLoadOp>(op, load.getMemRef(), map,
+ mapArgs);
+ return;
+ }
+ LLVM_DEBUG(llvm::dbgs()
+ << "[affine] Cannot raise memref op: " << op << "\n");
+ }
+ });
+ }
+};
+
+} // namespace
+
+std::unique_ptr<OperationPass<func::FuncOp>>
+mlir::affine::createRaiseMemrefToAffine() {
+ return std::make_unique<RaiseMemrefDialect>();
+}
diff --git a/mlir/test/Dialect/Affine/raise-memref.mlir b/mlir/test/Dialect/Affine/raise-memref.mlir
new file mode 100644
index 0000000000000..00cc98de1f40f
--- /dev/null
+++ b/mlir/test/Dialect/Affine/raise-memref.mlir
@@ -0,0 +1,138 @@
+// RUN: mlir-opt %s -allow-unregistered-dialect -affine-raise-from-memref --canonicalize | FileCheck %s
+
+// CHECK-LABEL: func @reduce_window_max(
+func.func @reduce_window_max() {
+ %cst = arith.constant 0.000000e+00 : f32
+ %0 = memref.alloc() : memref<1x8x8x64xf32>
+ %1 = memref.alloc() : memref<1x18x18x64xf32>
+ affine.for %arg0 = 0 to 1 {
+ affine.for %arg1 = 0 to 8 {
+ affine.for %arg2 = 0 to 8 {
+ affine.for %arg3 = 0 to 64 {
+ memref.store %cst, %0[%arg0, %arg1, %arg2, %arg3] : memref<1x8x8x64xf32>
+ }
+ }
+ }
+ }
+ affine.for %arg0 = 0 to 1 {
+ affine.for %arg1 = 0 to 8 {
+ affine.for %arg2 = 0 to 8 {
+ affine.for %arg3 = 0 to 64 {
+ affine.for %arg4 = 0 to 1 {
+ affine.for %arg5 = 0 to 3 {
+ affine.for %arg6 = 0 to 3 {
+ affine.for %arg7 = 0 to 1 {
+ %2 = memref.load %0[%arg0, %arg1, %arg2, %arg3] : memref<1x8x8x64xf32>
+ %21 = arith.addi %arg0, %arg4 : index
+ %22 = arith.constant 2 : index
+ %23 = arith.muli %arg1, %22 : index
+ %24 = arith.addi %23, %arg5 : index
+ %25 = arith.muli %arg2, %22 : index
+ %26 = arith.addi %25, %arg6 : index
+ %27 = arith.addi %arg3, %arg7 : index
+ %3 = memref.load %1[%21, %24, %26, %27] : memref<1x18x18x64xf32>
+ %4 = arith.cmpf ogt, %2, %3 : f32
+ %5 = arith.select %4, %2, %3 : f32
+ memref.store %5, %0[%arg0, %arg1, %arg2, %arg3] : memref<1x8x8x64xf32>
+ }
+ }
+ }
+ }
+ }
+ }
+ }
+ }
+ return
+}
+
+// CHECK: %[[cst:.*]] = arith.constant 0
+// CHECK: %[[v0:.*]] = memref.alloc() : memref<1x8x8x64xf32>
+// CHECK: %[[v1:.*]] = memref.alloc() : memref<1x18x18x64xf32>
+// CHECK: affine.for %[[arg0:.*]] =
+// CHECK: affine.for %[[arg1:.*]] =
+// CHECK: affine.for %[[arg2:.*]] =
+// CHECK: affine.for %[[arg3:.*]] =
+// CHECK: affine.store %[[cst]], %[[v0]][%[[arg0]], %[[arg1]], %[[arg2]], %[[arg3]]] :
+// CHECK: affine.for %[[a0:.*]] =
+// CHECK: affine.for %[[a1:.*]] =
+// CHECK: affine.for %[[a2:.*]] =
+// CHECK: affine.for %[[a3:.*]] =
+// CHECK: affine.for %[[a4:.*]] =
+// CHECK: affine.for %[[a5:.*]] =
+// CHECK: affine.for %[[a6:.*]] =
+// CHECK: affine.for %[[a7:.*]] =
+// CHECK: %[[lhs:.*]] = affine.load %[[v0]][%[[a0]], %[[a1]], %[[a2]], %[[a3]]] :
+// CHECK: %[[rhs:.*]] = affine.load %[[v1]][%[[a0]] + %[[a4]], %[[a1]] * 2 + %[[a5]], %[[a2]] * 2 + %[[a6]], %[[a3]] + %[[a7]]] :
+// CHECK: %[[res:.*]] = arith.cmpf ogt, %[[lhs]], %[[rhs]] : f32
+// CHECK: %[[sel:.*]] = arith.select %[[res]], %[[lhs]], %[[rhs]] : f32
+// CHECK: affine.store %[[sel]], %[[v0]][%[[a0]], %[[a1]], %[[a2]], %[[a3]]] :
+
+// CHECK-LABEL: func @symbols(
+func.func @symbols(%N : index) {
+ %0 = memref.alloc() : memref<1024x1024xf32>
+ %1 = memref.alloc() : memref<1024x1024xf32>
+ %2 = memref.alloc() : memref<1024x1024xf32>
+ %cst1 = arith.constant 1 : index
+ %cst2 = arith.constant 2 : index
+ affine.for %i = 0 to %N {
+ affine.for %j = 0 to %N {
+ %7 = memref.load %2[%i, %j] : memref<1024x1024xf32>
+ %10 = affine.for %k = 0 to %N iter_args(%ax = %cst1) -> index {
+ %12 = arith.muli %N, %cst2 : index
+ %13 = arith.addi %12, %cst1 : index
+ %14 = arith.addi %13, %j : index
+ %5 = memref.load %0[%i, %12] : memref<1024x1024xf32>
+ %6 = memref.load %1[%14, %j] : memref<1024x1024xf32>
+ %8 = arith.mulf %5, %6 : f32
+ %9 = arith.addf %7, %8 : f32
+ %4 = arith.addi %N, %cst1 : index
+ %11 = arith.addi %ax, %cst1 : index
+ memref.store %9, %2[%i, %4] : memref<1024x1024xf32> // this uses an expression of the symbol
+ memref.store %9, %2[%i, %11] : memref<1024x1024xf32> // this uses an iter_args and cannot be raised
+ %something = "ab.v"() : () -> index
+ memref.store %9, %2[%i, %something] : memref<1024x1024xf32> // this cannot be raised
+ affine.yield %11 : index
+ }
+ }
+ }
+ return
+}
+
+// CHECK: %[[cst1:.*]] = arith.constant 1 : index
+// CHECK: %[[v0:.*]] = memref.alloc() : memref<
+// CHECK: %[[v1:.*]] = memref.alloc() : memref<
+// CHECK: %[[v2:.*]] = memref.alloc() : memref<
+// CHECK: affine.for %[[a1:.*]] = 0 to %arg0 {
+// CHECK: affine.for %[[a2:.*]] = 0 to %arg0 {
+// CHECK: %[[lhs:.*]] = affine.load %{{.*}}[%[[a1]], %[[a2]]] : memref<1024x1024xf32>
+// CHECK: affine.for %[[a3:.*]] = 0 to %arg0 iter_args(%[[a4:.*]] = %[[cst1]]) -> (index) {
+// CHECK: %[[lhs2:.*]] = affine.load %{{.*}}[%[[a1]], symbol(%arg0) * 2] :
+// CHECK: %[[lhs3:.*]] = affine.load %{{.*}}[%[[a2]] + symbol(%arg0) * 2 + 1, %[[a2]]] :
+// CHECK: %[[lhs4:.*]] = arith.mulf %[[lhs2]], %[[lhs3]]
+// CHECK: %[[lhs5:.*]] = arith.addf %[[lhs]], %[[lhs4]]
+// CHECK: %[[lhs6:.*]] = arith.addi %[[a4]], %[[cst1]]
+// CHECK: affine.store %[[lhs5]], %{{.*}}[%[[a1]], symbol(%arg0) + 1] :
+// CHECK: memref.store %[[lhs5]], %{{.*}}[%[[a1]], %[[lhs6]]] :
+// CHECK: %[[lhs7:.*]] = "ab.v"
+// CHECK: memref.store %[[lhs5]], %{{.*}}[%[[a1]], %[[lhs7]]] :
+// CHECK: affine.yield %[[lhs6]]
+
+
+// CHECK-LABEL: func @non_affine(
+func.func @non_affine(%N : index) {
+ %2 = memref.alloc() : memref<1024x1024xf32>
+ affine.for %i = 0 to %N {
+ affine.for %j = 0 to %N {
+ %ij = arith.muli %i, %j : index
+ %7 = memref.load %2[%i, %ij] : memref<1024x1024xf32>
+ memref.store %7, %2[%ij, %ij] : memref<1024x1024xf32>
+ }
+ }
+ return
+}
+
+// CHECK: affine.for %[[i:.*]] =
+// CHECK: affine.for %[[j:.*]] =
+// CHECK: %[[ij:.*]] = arith.muli %[[i]], %[[j]]
+// CHECK: %[[v:.*]] = memref.load %{{.*}}[%[[i]], %[[ij]]]
+// CHECK: memref.store %[[v]], %{{.*}}[%[[ij]], %[[ij]]]
\ No newline at end of file
``````````
</details>
https://github.com/llvm/llvm-project/pull/138004
More information about the Mlir-commits
mailing list