[MLIR] Fix lowering of affine operations with return values.

[pr4tgpt]

Summary

This commit addresses the issue of lowering affine.for and affine.parallel having return values. Relevant test cases are also added.
Signed-off-by: Prateek Gupta <[e-mail]>

[bondhugula]

mlir/lib/Conversion/AffineToStandard/AffineToStandard.cpp

369

Why is its argument optional? This will just crash in the switch check if it doesn't have a value!

388

Likewise.

[bondhugula]

mlir/lib/Conversion/AffineToStandard/AffineToStandard.cpp

400

Nit: terminate comment with a full stop.

mlir/test/Conversion/AffineToStandard/lower-affine.mlir

45

No need of the %{{.*}} = part unless you are capturing and matching it.

[dcaballe]

Thanks for working on this! Just some minor comments from my side, some of them unrelated to the patch, for my understanding :)

mlir/lib/Conversion/AffineToStandard/AffineToStandard.cpp

368

value of identity value -> identity value?
a Atomic -> an Atomic?

384

Wouldn't this return be unreachable?

467

nit: please, compute the upper bound only once: for (unsigned i = 0, end = reductions.size(); i < end...

mlir/test/Conversion/AffineToStandard/lower-affine.mlir

34

Unrelated question: is affine analysis able to reason about reductions in this form? I remember a discussion about using mem2reg <--> reg2mem. Has there been any progress on that from?

51

Sorry, a bit unrelated but, while at this... Could we modernize this code and add the special string // ----- for separation and enable -split-input-file?

[pr4tgpt]

First update referring to the comments. Corrected the Optional arguments issue. Also fixed minor issues related to code comments and other optimizations.

[pr4tgpt]

mlir/lib/Conversion/AffineToStandard/AffineToStandard.cpp

384

Changed the Optional arguments.

[bondhugula]

Mostly minor comments.

mlir/lib/Conversion/AffineToStandard/AffineToStandard.cpp

387–388

Nit: This comment could be made more accurate.

Return the value obtained by applying the reduction operation kind associated with a binary AtomicRMWKind op` to lhs and rhs`
?

457

Nit: Operation -> operation

462

Use /* ... =*/nullptr for the last argument.

464

Rephrase -> "Copy the body of the affine.parallel op"?

472

You don't need mlir::.

481

Nit: resultReduction -> reductionResult ?

mlir/test/Conversion/AffineToStandard/lower-affine.mlir

34

UNRELATED: Not really. The affine passes haven't been reviewed to handle reduction values where appropriate.

51

This I suspect may require some non-trivial updates to the test cases. Let's do this in an NFC patch to not confuse this history.

706

These test cases don't have return values. I assume you just added them for additional safety so that the changes don't break these cases?

[pr4tgpt]

Updated few comments and minor changes as suggested.

[pr4tgpt]

mlir/test/Conversion/AffineToStandard/lower-affine.mlir

706

Yes!

[pr4tgpt]

Updated minor comments.

[bondhugula]

mlir/lib/Conversion/AffineToStandard/AffineToStandard.cpp

387

Triple /// comments please.

448

Nit: having -> with

[pr4tgpt]

Update minor comments.

mlir/lib/Conversion/AffineToStandard/AffineToStandard.cpp

 
 /// Affine yields ops are removed.
 class AffineYieldOpLowering : public OpRewritePattern<AffineYieldOp> {
 public:
   using OpRewritePattern<AffineYieldOp>::OpRewritePattern;
 
   LogicalResult matchAndRewrite(AffineYieldOp op,
                                 PatternRewriter &rewriter) const override {
+    if (isa<scf::ParallelOp>(op.getParentOp())) {
+      // scf.parallel does not yield any values via its terminator scf.yield but
+      // models reductions differently using additional ops in its region.
       rewriter.replaceOpWithNewOp<scf::YieldOp>(op);
       return success();
     }
+    rewriter.replaceOpWithNewOp<scf::YieldOp>(op, op.operands());
+    return success();
+  }
 };
 
 class AffineForLowering : public OpRewritePattern<AffineForOp> {
 public:
   using OpRewritePattern<AffineForOp>::OpRewritePattern;
 
   LogicalResult matchAndRewrite(AffineForOp op,
                                 PatternRewriter &rewriter) const override {
     Location loc = op.getLoc();
     Value lowerBound = lowerAffineLowerBound(op, rewriter);
     Value upperBound = lowerAffineUpperBound(op, rewriter);
     Value step = rewriter.create<ConstantIndexOp>(loc, op.getStep());
-    auto f = rewriter.create<scf::ForOp>(loc, lowerBound, upperBound, step);
-    rewriter.eraseBlock(f.getBody());
-    rewriter.inlineRegionBefore(op.region(), f.region(), f.region().end());
-    rewriter.eraseOp(op);
+    auto scfForOp = rewriter.create<scf::ForOp>(loc, lowerBound, upperBound,
+                                                step, op.getIterOperands());
+    rewriter.eraseBlock(scfForOp.getBody());
+    rewriter.inlineRegionBefore(op.region(), scfForOp.region(),
+                                scfForOp.region().end());
+    rewriter.replaceOp(op, scfForOp.results());
     return success();
   }
 };
 
+/// Returns the identity value associated with an AtomicRMWKind op.
+static Value getIdentityValue(AtomicRMWKind op, OpBuilder &builder,
+                              Location loc) {
+  switch (op) {
+  case AtomicRMWKind::addf:
+    return builder.create<ConstantOp>(loc, builder.getF32FloatAttr(0));
+  case AtomicRMWKind::addi:
+    return builder.create<ConstantOp>(loc, builder.getI32IntegerAttr(0));
+  case AtomicRMWKind::mulf:
+    return builder.create<ConstantOp>(loc, builder.getF32FloatAttr(1));
+  case AtomicRMWKind::muli:
+    return builder.create<ConstantOp>(loc, builder.getI32IntegerAttr(1));
+  // TODO: Add remaining reduction operations.
+  default:
+    emitOptionalError(loc, "Reduction operation type not supported");
+  }
+  return nullptr;
+}
+
+/// Return the value obtained by applying the reduction operation kind
+/// associated with a binary AtomicRMWKind op to `lhs` and `rhs`.
+static Value getReductionOp(AtomicRMWKind op, OpBuilder &builder, Location loc,
+                            Value lhs, Value rhs) {
+  switch (op) {
+  case AtomicRMWKind::addf:
+    return builder.create<AddFOp>(loc, lhs, rhs);
+  case AtomicRMWKind::addi:
+    return builder.create<AddIOp>(loc, lhs, rhs);
+  case AtomicRMWKind::mulf:
+    return builder.create<MulFOp>(loc, lhs, rhs);
+  case AtomicRMWKind::muli:
+    return builder.create<MulIOp>(loc, lhs, rhs);
+  // TODO: Add remaining reduction operations.
+  default:
+    emitOptionalError(loc, "Reduction operation type not supported");
+  }
+  return nullptr;
+}
+
 /// Convert an `affine.parallel` (loop nest) operation into a `scf.parallel`
 /// operation.
 class AffineParallelLowering : public OpRewritePattern<AffineParallelOp> {
 public:
   using OpRewritePattern<AffineParallelOp>::OpRewritePattern;
 
   LogicalResult matchAndRewrite(AffineParallelOp op,
                                 PatternRewriter &rewriter) const override {
     Location loc = op.getLoc();
     SmallVector<Value, 8> steps;
     SmallVector<Value, 8> upperBoundTuple;
     SmallVector<Value, 8> lowerBoundTuple;
+    SmallVector<Value, 8> identityVals;
     // Finding lower and upper bound by expanding the map expression.
     // Checking if expandAffineMap is not giving NULL.
-    Optional<SmallVector<Value, 8>> upperBound = expandAffineMap(
-        rewriter, loc, op.upperBoundsMap(), op.getUpperBoundsOperands());
     Optional<SmallVector<Value, 8>> lowerBound = expandAffineMap(
         rewriter, loc, op.lowerBoundsMap(), op.getLowerBoundsOperands());
+    Optional<SmallVector<Value, 8>> upperBound = expandAffineMap(
+        rewriter, loc, op.upperBoundsMap(), op.getUpperBoundsOperands());
     if (!lowerBound || !upperBound)
       return failure();
     upperBoundTuple = *upperBound;
     lowerBoundTuple = *lowerBound;
     steps.reserve(op.steps().size());
     for (Attribute step : op.steps())
       steps.push_back(rewriter.create<ConstantIndexOp>(
           loc, step.cast<IntegerAttr>().getInt()));
-    // Creating empty scf.parallel op body with appropriate bounds.
-    auto parallelOp = rewriter.create<scf::ParallelOp>(loc, lowerBoundTuple,
-                                                       upperBoundTuple, steps);
-    rewriter.eraseBlock(parallelOp.getBody());
-    rewriter.inlineRegionBefore(op.region(), parallelOp.region(),
-                                parallelOp.region().end());
-    rewriter.eraseOp(op);
+    // Get the terminator op.
+    Operation *affineParOpTerminator = op.getBody()->getTerminator();
+    scf::ParallelOp parOp;
+    if (op.results().empty()) {
+      // Case with no reduction operations/return values.
+      parOp = rewriter.create<scf::ParallelOp>(loc, lowerBoundTuple,
+                                               upperBoundTuple, steps,
+                                               /*bodyBuilderFn=*/nullptr);
+      rewriter.eraseBlock(parOp.getBody());
+      rewriter.inlineRegionBefore(op.region(), parOp.region(),
+                                  parOp.region().end());
+      rewriter.replaceOp(op, parOp.results());
+      return success();
+    }
+    // Case with affine.parallel with reduction operations/return values.
+    // scf.parallel handles the reduction operation differently unlike
+    // affine.parallel.
+    ArrayRef<Attribute> reductions = op.reductions().getValue();
+    for (Attribute reduction : reductions) {
+      // For each of the reduction operations get the identity values for
+      // initialization of the result values.
+      Optional<AtomicRMWKind> reductionOp = symbolizeAtomicRMWKind(
+          static_cast<uint64_t>(reduction.cast<IntegerAttr>().getInt()));
+      assert(reductionOp.hasValue() &&
+             "Reduction operation cannot be of None Type");
+      AtomicRMWKind reductionOpValue = reductionOp.getValue();
+      identityVals.push_back(getIdentityValue(reductionOpValue, rewriter, loc));
+    }
+    parOp = rewriter.create<scf::ParallelOp>(
+        loc, lowerBoundTuple, upperBoundTuple, steps, identityVals,
+        /*bodyBuilderFn=*/nullptr);
+
+    //  Copy the body of the affine.parallel op.
+    rewriter.eraseBlock(parOp.getBody());
+    rewriter.inlineRegionBefore(op.region(), parOp.region(),
+                                parOp.region().end());
+    assert(reductions.size() == affineParOpTerminator->getNumOperands() &&
+           "Unequal number of reductions and operands.");
+    for (unsigned i = 0, end = reductions.size(); i < end; i++) {
+      // For each of the reduction operations get the respective mlir::Value.
+      Optional<AtomicRMWKind> reductionOp =
+          symbolizeAtomicRMWKind(reductions[i].cast<IntegerAttr>().getInt());
+      assert(reductionOp.hasValue() &&
+             "Reduction Operation cannot be of None Type");
+      AtomicRMWKind reductionOpValue = reductionOp.getValue();
+      rewriter.setInsertionPoint(&parOp.getBody()->back());
+      auto reduceOp = rewriter.create<scf::ReduceOp>(
+          loc, affineParOpTerminator->getOperand(i));
+      rewriter.setInsertionPointToEnd(&reduceOp.reductionOperator().front());
+      Value reductionResult =
+          getReductionOp(reductionOpValue, rewriter, loc,
+                         reduceOp.reductionOperator().front().getArgument(0),
+                         reduceOp.reductionOperator().front().getArgument(1));
+      rewriter.create<scf::ReduceReturnOp>(loc, reductionResult);
+    }
+    rewriter.replaceOp(op, parOp.results());
     return success();
   }
 };
 
 class AffineIfLowering : public OpRewritePattern<AffineIfOp> {
 public:
   using OpRewritePattern<AffineIfOp>::OpRewritePattern;
 

mlir/test/Conversion/AffineToStandard/lower-affine.mlir

func @simple_loop() {
   affine.for %i = 1 to 42 {
     call @body(%i) : (index) -> ()
   }
   return
 }
 
 /////////////////////////////////////////////////////////////////////
 
+func @for_with_yield(%buffer: memref<1024xf32>) -> (f32) {
+  %sum_0 = constant 0.0 : f32
+  %sum = affine.for %i = 0 to 10 step 2 iter_args(%sum_iter = %sum_0) -> (f32) {
+    %t = affine.load %buffer[%i] : memref<1024xf32>
+    %sum_next = addf %sum_iter, %t : f32
+    affine.yield %sum_next : f32
+  }
+  return %sum : f32
+}
+
+// CHECK-LABEL: func @for_with_yield
+// CHECK:         %[[INIT_SUM:.*]] = constant 0.000000e+00 : f32
+// CHECK-NEXT:    %[[LOWER:.*]] = constant 0 : index
+// CHECK-NEXT:    %[[UPPER:.*]] = constant 10 : index
+// CHECK-NEXT:    %[[STEP:.*]] = constant 2 : index
+// CHECK-NEXT:    %[[SUM:.*]] = scf.for %[[IV:.*]] = %[[LOWER]] to %[[UPPER]] step %[[STEP]] iter_args(%[[SUM_ITER:.*]] = %[[INIT_SUM]]) -> (f32) {
+// CHECK-NEXT:      load    
+// CHECK-NEXT:      %[[SUM_NEXT:.*]] = addf
+// CHECK-NEXT:      scf.yield %[[SUM_NEXT]] : f32
+// CHECK-NEXT:    }
+// CHECK-NEXT:    return %[[SUM]] : f32
+
+/////////////////////////////////////////////////////////////////////
+
 func private @pre(index) -> ()
 func private @body2(index, index) -> ()
 func private @post(index) -> ()
 
 // CHECK-LABEL: func @imperfectly_nested_loops
 // CHECK-NEXT:   %[[c0:.*]] = constant 0 : index
 // CHECK-NEXT:   %[[c42:.*]] = constant 42 : index
 // CHECK-NEXT:   %[[c1:.*]] = constant 1 : index

 // CHECK-DAG:       %[[C1:.*]] = constant 1
 // CHECK-DAG:       %[[C1_9:.*]] = constant 1
 // CHECK-DAG:       %[[C1_10:.*]] = constant 1
 // CHECK:           scf.parallel (%[[arg6:.*]], %[[arg7:.*]], %[[arg8:.*]]) = (%[[arg3]], %[[arg4]], %[[arg5]]) to (%[[A0]], %[[A1]], %[[A2]]) step (%[[C1]], %[[C1_9]], %[[C1_10]]) {
 // CHECK:             %[[A3:.*]] = load %[[ARG1]][%[[arg6]], %[[arg8]]] : memref<100x100xf32>
 // CHECK:             %[[A4:.*]] = load %[[ARG2]][%[[arg8]], %[[arg7]]] : memref<100x100xf32>
 // CHECK:             mulf %[[A3]], %[[A4]] : f32
 // CHECK:             scf.yield
+
+/////////////////////////////////////////////////////////////////////
+
+func @affine_parallel_simple(%arg0: memref<3x3xf32>, %arg1: memref<3x3xf32>) -> (memref<3x3xf32>) {
+  %O = alloc() : memref<3x3xf32>
+  affine.parallel (%kx, %ky) = (0, 0) to (2, 2) {
+      %1 = affine.load %arg0[%kx, %ky] : memref<3x3xf32>
+      %2 = affine.load %arg1[%kx, %ky] : memref<3x3xf32>
+      %3 = mulf %1, %2 : f32
+      affine.store %3, %O[%kx, %ky] : memref<3x3xf32>
+  }
+  return %O : memref<3x3xf32>
+}
+// CHECK-LABEL: func @affine_parallel_simple
+// CHECK:         %[[LOWER_1:.*]] = constant 0 : index
+// CHECK-NEXT:    %[[LOWER_2:.*]] = constant 0 : index
+// CHECK-NEXT:    %[[UPPER_1:.*]] = constant 2 : index
+// CHECK-NEXT:    %[[UPPER_2:.*]] = constant 2 : index
+// CHECK-NEXT:    %[[STEP_1:.*]] = constant 1 : index
+// CHECK-NEXT:    %[[STEP_2:.*]] = constant 1 : index
+// CHECK-NEXT:    scf.parallel (%[[I:.*]], %[[J:.*]]) = (%[[LOWER_1]], %[[LOWER_2]]) to (%[[UPPER_1]], %[[UPPER_2]]) step (%[[STEP_1]], %[[STEP_2]]) {
+// CHECK-NEXT:      %[[VAL_1:.*]] = load
+// CHECK-NEXT:      %[[VAL_2:.*]] = load
+// CHECK-NEXT:      %[[PRODUCT:.*]] = mulf
+// CHECK-NEXT:      store
+// CHECK-NEXT:      scf.yield
+// CHECK-NEXT:    }
+// CHECK-NEXT:    return
+// CHECK-NEXT:  }
+
+/////////////////////////////////////////////////////////////////////
+
+func @affine_parallel_simple_dynamic_bounds(%arg0: memref<?x?xf32>, %arg1: memref<?x?xf32>, %arg2: memref<?x?xf32>) {
+  %c_0 = constant 0 : index
+  %output_dim = dim %arg0, %c_0 : memref<?x?xf32>
+  affine.parallel (%kx, %ky) = (%c_0, %c_0) to (%output_dim, %output_dim) {
+      %1 = affine.load %arg0[%kx, %ky] : memref<?x?xf32>
+      %2 = affine.load %arg1[%kx, %ky] : memref<?x?xf32>
+      %3 = mulf %1, %2 : f32
+      affine.store %3, %arg2[%kx, %ky] : memref<?x?xf32>
+  }
+  return
+}
+// CHECK-LABEL: func @affine_parallel_simple_dynamic_bounds
+// CHECK-SAME:  %[[ARG_0:.*]]: memref<?x?xf32>, %[[ARG_1:.*]]: memref<?x?xf32>, %[[ARG_2:.*]]: memref<?x?xf32>
+// CHECK:         %[[DIM_INDEX:.*]] = constant 0 : index
+// CHECK-NEXT:    %[[UPPER:.*]] = dim %[[ARG_0]], %[[DIM_INDEX]] : memref<?x?xf32>
+// CHECK-NEXT:    %[[LOWER_1:.*]] = constant 0 : index
+// CHECK-NEXT:    %[[LOWER_2:.*]] = constant 0 : index
+// CHECK-NEXT:    %[[STEP_1:.*]] = constant 1 : index
+// CHECK-NEXT:    %[[STEP_2:.*]] = constant 1 : index
+// CHECK-NEXT:    scf.parallel (%[[I:.*]], %[[J:.*]]) = (%[[LOWER_1]], %[[LOWER_2]]) to (%[[UPPER]], %[[UPPER]]) step (%[[STEP_1]], %[[STEP_2]]) {
+// CHECK-NEXT:      %[[VAL_1:.*]] = load
+// CHECK-NEXT:      %[[VAL_2:.*]] = load
+// CHECK-NEXT:      %[[PRODUCT:.*]] = mulf
+// CHECK-NEXT:      store
+// CHECK-NEXT:      scf.yield
+// CHECK-NEXT:    }
+// CHECK-NEXT:    return
+// CHECK-NEXT:  }
+
+/////////////////////////////////////////////////////////////////////
+
+func @affine_parallel_with_reductions(%arg0: memref<3x3xf32>, %arg1: memref<3x3xf32>) -> (f32, f32) {
+  %0:2 = affine.parallel (%kx, %ky) = (0, 0) to (2, 2) reduce ("addf", "mulf") -> (f32, f32) {
+            %1 = affine.load %arg0[%kx, %ky] : memref<3x3xf32>
+            %2 = affine.load %arg1[%kx, %ky] : memref<3x3xf32>
+            %3 = mulf %1, %2 : f32
+            %4 = addf %1, %2 : f32
+            affine.yield %3, %4 : f32, f32
+          }
+  return %0#0, %0#1 : f32, f32
+}
+// CHECK-LABEL: func @affine_parallel_with_reductions
+// CHECK:         %[[LOWER_1:.*]] = constant 0 : index
+// CHECK-NEXT:    %[[LOWER_2:.*]] = constant 0 : index
+// CHECK-NEXT:    %[[UPPER_1:.*]] = constant 2 : index
+// CHECK-NEXT:    %[[UPPER_2:.*]] = constant 2 : index
+// CHECK-NEXT:    %[[STEP_1:.*]] = constant 1 : index
+// CHECK-NEXT:    %[[STEP_2:.*]] = constant 1 : index
+// CHECK-NEXT:    %[[INIT_1:.*]] = constant 0.000000e+00 : f32
+// CHECK-NEXT:    %[[INIT_2:.*]] = constant 1.000000e+00 : f32
+// CHECK-NEXT:    %[[RES:.*]] = scf.parallel (%[[I:.*]], %[[J:.*]]) = (%[[LOWER_1]], %[[LOWER_2]]) to (%[[UPPER_1]], %[[UPPER_2]]) step (%[[STEP_1]], %[[STEP_2]]) init (%[[INIT_1]], %[[INIT_2]]) -> (f32, f32) {
+// CHECK-NEXT:      %[[VAL_1:.*]] = load
+// CHECK-NEXT:      %[[VAL_2:.*]] = load
+// CHECK-NEXT:      %[[PRODUCT:.*]] = mulf
+// CHECK-NEXT:      %[[SUM:.*]] = addf
+// CHECK-NEXT:      scf.reduce(%[[PRODUCT]]) : f32 {
+// CHECK-NEXT:      ^bb0(%[[LHS:.*]]: f32, %[[RHS:.*]]: f32):
+// CHECK-NEXT:        %[[RES:.*]] = addf
+// CHECK-NEXT:        scf.reduce.return %[[RES]] : f32
+// CHECK-NEXT:      }
+// CHECK-NEXT:      scf.reduce(%[[SUM]]) : f32 {
+// CHECK-NEXT:      ^bb0(%[[LHS:.*]]: f32, %[[RHS:.*]]: f32):
+// CHECK-NEXT:        %[[RES:.*]] = mulf
+// CHECK-NEXT:        scf.reduce.return %[[RES]] : f32
+// CHECK-NEXT:      }
+// CHECK-NEXT:      scf.yield
+// CHECK-NEXT:    }
+// CHECK-NEXT:    return
+// CHECK-NEXT:  }