[Mlir-commits] [mlir] [mlir]Fix compose subview (PR #80551)

llvmlistbot at llvm.org llvmlistbot at llvm.org
Sat Feb 3 08:33:43 PST 2024


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir

Author: lonely eagle (linuxlonelyeagle)

<details>
<summary>Changes</summary>

I found a bug in `test-compose-subview`,You can see the example I gave.
```
#map = affine_map<() -> ()>
module {
  func.func private @<!-- -->fun(%arg0: memref<10x10xf32>, %arg1: memref<5x5xf32>) -> memref<5x5xf32> {
    %c0 = arith.constant 0 : index
    %c5 = arith.constant 5 : index
    %c1 = arith.constant 1 : index
    %subview = memref.subview %arg0[0, 0] [5, 5] [1, 1] : memref<10x10xf32> to memref<5x5xf32, strided<[10, 1]>>
    %alloc = memref.alloc() : memref<5x5xf32>
    scf.for %arg2 = %c0 to %c5 step %c1 {
      scf.for %arg3 = %c0 to %c5 step %c1 {
        %subview_0 = memref.subview %subview[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32, strided<[10, 1]>> to memref<f32, strided<[], offset: ?>>
        %subview_1 = memref.subview %arg1[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32> to memref<f32, strided<[], offset: ?>>
        %alloc_2 = memref.alloc() : memref<f32>
        linalg.generic {indexing_maps = [#map, #map, #map], iterator_types = []} ins(%subview_0, %subview_1 : memref<f32, strided<[], offset: ?>>, memref<f32, strided<[], offset: ?>>) outs(%alloc_2 : memref<f32>) {
        ^bb0(%in: f32, %in_4: f32, %out: f32):
          %0 = arith.addf %in, %in_4 : f32
          linalg.yield %0 : f32
        }
        %subview_3 = memref.subview %alloc[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32> to memref<f32, strided<[], offset: ?>>
        memref.copy %alloc_2, %subview_3 : memref<f32> to memref<f32, strided<[], offset: ?>>
      }
    }
    return %alloc : memref<5x5xf32>
  }
  func.func @<!-- -->test(%arg0: memref<10x10xf32>, %arg1: memref<5x5xf32>) -> memref<5x5xf32> {
    %0 = call @<!-- -->fun(%arg0, %arg1) : (memref<10x10xf32>, memref<5x5xf32>) -> memref<5x5xf32>
    return %0 : memref<5x5xf32>
  }
}
```
When I run `mlir-opt test.mlir ---test-compose-subview`.
```
test.mlir:14:9: error: 'linalg.generic' op expected operand rank (2) to match the result rank of indexing_map #<!-- -->0 (0)
        linalg.generic {indexing_maps = [#map, #map, #map], iterator_types = []} ins(%subview_0, %subview_1 : memref<f32, strided<[], offset: ?>>, memref<f32, strided<[], offset: ?>>) outs(%alloc_2 : memref<f32>) {
        ^
test1.mlir:14:9: note: see current operation: 
"linalg.generic"(%4, %5, %6) <{indexing_maps = [affine_map<() -> ()>, affine_map<() -> ()>, affine_map<() -> ()>], iterator_types = [], operandSegmentSizes = array<i32: 2, 1>}> ({
^bb0(%arg4: f32, %arg5: f32, %arg6: f32):
  %8 = "arith.addf"(%arg4, %arg5) <{fastmath = #arith.fastmath<none>}> : (f32, f32) -> f32
  "linalg.yield"(%8) : (f32) -> ()
}) : (memref<1x1xf32, strided<[10, 1], offset: ?>>, memref<f32, strided<[], offset: ?>>, memref<f32>) -> ()
```
This PR fixes that.In the meantime I've extended this PR to handle cases where stride is greater than 1.
```
func.func private @<!-- -->Unknown0(%arg0: memref<10x10xf32>, %arg1: memref<5x5xf32>) -> memref<5x5xf32> {
  %c0 = arith.constant 0 : index
  %c5 = arith.constant 5 : index
  %c1 = arith.constant 1 : index
  %subview = memref.subview %arg0[0, 0] [5, 5] [2, 2] : memref<10x10xf32> to memref<5x5xf32, strided<[20, 2]>>
  %alloc = memref.alloc() : memref<5x5xf32>
  scf.for %arg2 = %c0 to %c5 step %c1 {
    scf.for %arg3 = %c0 to %c5 step %c1 {
      %subview_0 = memref.subview %subview[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32, strided<[20, 2]>> to memref<f32, strided<[], offset: ?>>
      %subview_1 = memref.subview %arg1[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32> to memref<f32, strided<[], offset: ?>>
      %alloc_2 = memref.alloc() : memref<f32>
      linalg.generic {indexing_maps = [affine_map<() -> ()>, affine_map<() -> ()>, affine_map<() -> ()>], iterator_types = []} ins(%subview_0, %subview_1 : memref<f32, strided<[], offset: ?>>, memref<f32, strided<[], offset: ?>>) outs(%alloc_2 : memref<f32>) {
      ^bb0(%in: f32, %in_4: f32, %out: f32):
        %0 = arith.addf %in, %in_4 : f32
        linalg.yield %0 : f32
      }
      %subview_3 = memref.subview %alloc[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32> to memref<f32, strided<[], offset: ?>>
      memref.copy %alloc_2, %subview_3 : memref<f32> to memref<f32, strided<[], offset: ?>>
    }
  }
  return %alloc : memref<5x5xf32>
}
$ mlir-opt test.mlir -test-compose-subview
#map = affine_map<()[s0] -> (s0 * 2)>
#map1 = affine_map<() -> ()>
module {
  func.func private @<!-- -->Unknown0(%arg0: memref<10x10xf32>, %arg1: memref<5x5xf32>) -> memref<5x5xf32>  {
    %c0 = arith.constant 0 : index
    %c5 = arith.constant 5 : index
    %c1 = arith.constant 1 : index
    %alloc = memref.alloc() : memref<5x5xf32>
    scf.for %arg2 = %c0 to %c5 step %c1 {
      scf.for %arg3 = %c0 to %c5 step %c1 {
        %0 = affine.apply #map()[%arg2]
        %1 = affine.apply #map()[%arg3]
        %subview = memref.subview %arg0[%0, %1] [1, 1] [2, 2] : memref<10x10xf32> to memref<f32, strided<[], offset: ?>>
        %subview_0 = memref.subview %arg1[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32> to memref<f32, strided<[], offset: ?>>
        %alloc_1 = memref.alloc() : memref<f32>
        linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = []} ins(%subview, %subview_0 : memref<f32, strided<[], offset: ?>>, memref<f32, strided<[], offset: ?>>) outs(%alloc_1 : memref<f32>) {
        ^bb0(%in: f32, %in_3: f32, %out: f32):
          %2 = arith.addf %in, %in_3 : f32
          linalg.yield %2 : f32
        }
        %subview_2 = memref.subview %alloc[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32> to memref<f32, strided<[], offset: ?>>
        memref.copy %alloc_1, %subview_2 : memref<f32> to memref<f32, strided<[], offset: ?>>
      }
    }
    return %alloc : memref<5x5xf32>
  }
}
```

---
Full diff: https://github.com/llvm/llvm-project/pull/80551.diff


2 Files Affected:

- (modified) mlir/lib/Dialect/MemRef/Transforms/ComposeSubView.cpp (+52-37) 
- (modified) mlir/test/Transforms/compose-subview.mlir (+48) 


``````````diff
diff --git a/mlir/lib/Dialect/MemRef/Transforms/ComposeSubView.cpp b/mlir/lib/Dialect/MemRef/Transforms/ComposeSubView.cpp
index 431d270b0a2cb..dd10865544a50 100644
--- a/mlir/lib/Dialect/MemRef/Transforms/ComposeSubView.cpp
+++ b/mlir/lib/Dialect/MemRef/Transforms/ComposeSubView.cpp
@@ -24,14 +24,14 @@ using namespace mlir;
 
 namespace {
 
-// Replaces a subview of a subview with a single subview. Only supports subview
-// ops with static sizes and static strides of 1 (both static and dynamic
+// Replaces a subview of a subview with a single subview(both static and dynamic
 // offsets are supported).
 struct ComposeSubViewOpPattern : public OpRewritePattern<memref::SubViewOp> {
   using OpRewritePattern::OpRewritePattern;
 
   LogicalResult matchAndRewrite(memref::SubViewOp op,
                                 PatternRewriter &rewriter) const override {
+
     // 'op' is the 'SubViewOp' we're rewriting. 'sourceOp' is the op that
     // produces the input of the op we're rewriting (for 'SubViewOp' the input
     // is called the "source" value). We can only combine them if both 'op' and
@@ -52,66 +52,81 @@ struct ComposeSubViewOpPattern : public OpRewritePattern<memref::SubViewOp> {
 
     // Offsets, sizes and strides OpFoldResult for the combined 'SubViewOp'.
     SmallVector<OpFoldResult> offsets, sizes, strides;
-
-    // Because we only support input strides of 1, the output stride is also
-    // always 1.
-    if (llvm::all_of(strides, [](OpFoldResult &valueOrAttr) {
-          Attribute attr = llvm::dyn_cast_if_present<Attribute>(valueOrAttr);
-          return attr && cast<IntegerAttr>(attr).getInt() == 1;
-        })) {
-      strides = SmallVector<OpFoldResult>(sourceOp.getMixedStrides().size(),
-                                          rewriter.getI64IntegerAttr(1));
-    } else {
-      return failure();
+    auto opStrides = op.getMixedStrides();
+    auto sourceStrides = sourceOp.getMixedStrides();
+
+    // The output stride in each dimension is equal to the product of the
+    // dimensions corresponding to source and op.
+    for (auto [opStride, sourceStride] : llvm::zip(opStrides, sourceStrides)) {
+      Attribute opStrideAttr = dyn_cast_if_present<Attribute>(opStride);
+      Attribute sourceStrideAttr = dyn_cast_if_present<Attribute>(sourceStride);
+      if (!opStrideAttr || !sourceStrideAttr)
+        return failure();
+      strides.push_back(rewriter.getI64IntegerAttr(
+          cast<IntegerAttr>(opStrideAttr).getInt() *
+          cast<IntegerAttr>(sourceStrideAttr).getInt()));
     }
 
     // The rules for calculating the new offsets and sizes are:
     // * Multiple subview offsets for a given dimension compose additively.
-    //   ("Offset by m" followed by "Offset by n" == "Offset by m + n")
+    //   ("Offset by m and Stride by k" followed by "Offset by n" == "Offset by
+    //   m + n * k")
     // * Multiple sizes for a given dimension compose by taking the size of the
     //   final subview and ignoring the rest. ("Take m values" followed by "Take
     //   n values" == "Take n values") This size must also be the smallest one
     //   by definition (a subview needs to be the same size as or smaller than
     //   its source along each dimension; presumably subviews that are larger
     //   than their sources are disallowed by validation).
-    for (auto it : llvm::zip(op.getMixedOffsets(), sourceOp.getMixedOffsets(),
-                             op.getMixedSizes())) {
-      auto opOffset = std::get<0>(it);
-      auto sourceOffset = std::get<1>(it);
-      auto opSize = std::get<2>(it);
-
+    for (auto [opOffset, sourceOffset, sourceStride, opSize] :
+         llvm::zip(op.getMixedOffsets(), sourceOp.getMixedOffsets(),
+                   sourceOp.getMixedStrides(), op.getMixedSizes())) {
       // We only support static sizes.
       if (opSize.is<Value>()) {
         return failure();
       }
-
       sizes.push_back(opSize);
       Attribute opOffsetAttr = llvm::dyn_cast_if_present<Attribute>(opOffset),
                 sourceOffsetAttr =
-                    llvm::dyn_cast_if_present<Attribute>(sourceOffset);
-
+                    llvm::dyn_cast_if_present<Attribute>(sourceOffset),
+                sourceStrideAttr =
+                    llvm::dyn_cast_if_present<Attribute>(sourceStride);
       if (opOffsetAttr && sourceOffsetAttr) {
+
         // If both offsets are static we can simply calculate the combined
         // offset statically.
         offsets.push_back(rewriter.getI64IntegerAttr(
-            cast<IntegerAttr>(opOffsetAttr).getInt() +
+            cast<IntegerAttr>(opOffsetAttr).getInt() *
+                cast<IntegerAttr>(sourceStrideAttr).getInt() +
             cast<IntegerAttr>(sourceOffsetAttr).getInt()));
       } else {
-        // When either offset is dynamic, we must emit an additional affine
-        // transformation to add the two offsets together dynamically.
-        AffineExpr expr = rewriter.getAffineConstantExpr(0);
+        AffineExpr expr0 = rewriter.getAffineConstantExpr(0);
+        AffineExpr expr1 = rewriter.getAffineConstantExpr(0);
         SmallVector<Value> affineApplyOperands;
-        for (auto valueOrAttr : {opOffset, sourceOffset}) {
-          if (auto attr = llvm::dyn_cast_if_present<Attribute>(valueOrAttr)) {
-            expr = expr + cast<IntegerAttr>(attr).getInt();
+        SmallVector<OpFoldResult> opOffsets{sourceOffset, opOffset};
+        for (auto [idx, offset] : llvm::enumerate(opOffsets)) {
+          if (auto attr = llvm::dyn_cast_if_present<Attribute>(offset)) {
+            if (idx == 0) {
+              expr0 = expr0 + cast<IntegerAttr>(attr).getInt();
+            } else if (idx == 1) {
+              expr1 = expr1 + cast<IntegerAttr>(attr).getInt();
+              expr1 = expr1 * cast<IntegerAttr>(sourceStrideAttr).getInt();
+              expr0 = expr0 + expr1;
+            }
           } else {
-            expr =
-                expr + rewriter.getAffineSymbolExpr(affineApplyOperands.size());
-            affineApplyOperands.push_back(valueOrAttr.get<Value>());
+            if (idx == 0) {
+              expr0 = expr0 +
+                      rewriter.getAffineSymbolExpr(affineApplyOperands.size());
+              affineApplyOperands.push_back(offset.get<Value>());
+            } else if (idx == 1) {
+              expr1 = expr1 +
+                      rewriter.getAffineSymbolExpr(affineApplyOperands.size());
+              affineApplyOperands.push_back(offset.get<Value>());
+              expr1 = expr1 * cast<IntegerAttr>(sourceStrideAttr).getInt();
+              expr0 = expr0 + expr1;
+            }
           }
         }
-
-        AffineMap map = AffineMap::get(0, affineApplyOperands.size(), expr);
+        AffineMap map = AffineMap::get(0, affineApplyOperands.size(), expr0);
         Value result = rewriter.create<affine::AffineApplyOp>(
             op.getLoc(), map, affineApplyOperands);
         offsets.push_back(result);
@@ -120,8 +135,8 @@ struct ComposeSubViewOpPattern : public OpRewritePattern<memref::SubViewOp> {
 
     // This replaces 'op' but leaves 'sourceOp' alone; if it no longer has any
     // uses it can be removed by a (separate) dead code elimination pass.
-    rewriter.replaceOpWithNewOp<memref::SubViewOp>(op, sourceOp.getSource(),
-                                                   offsets, sizes, strides);
+    rewriter.replaceOpWithNewOp<memref::SubViewOp>(
+        op, op.getType(), sourceOp.getSource(), offsets, sizes, strides);
     return success();
   }
 };
diff --git a/mlir/test/Transforms/compose-subview.mlir b/mlir/test/Transforms/compose-subview.mlir
index cb8ebcb2bf9e9..2fa551c218d42 100644
--- a/mlir/test/Transforms/compose-subview.mlir
+++ b/mlir/test/Transforms/compose-subview.mlir
@@ -45,3 +45,51 @@ func.func @main(%input: memref<4x1024xf32>) -> memref<1x128xf32, strided<[1024,
   %1 = memref.subview %0[1, %cst_128] [1, 128] [1, 1] : memref<2x256xf32, strided<[1024, 1], offset: ?>> to memref<1x128xf32, strided<[1024, 1], offset: ?>>
   return %1 : memref<1x128xf32, strided<[1024, 1], offset: ?>>
 }
+
+// -----
+
+func.func @main(%input: memref<4x1024xf32>) -> memref<1x64xf32, strided<[4096, 4], offset: 4480>> {
+  //      CHECK: subview %arg0[4, 384] [1, 64] [4, 4] 
+  // CHECK-SAME: memref<4x1024xf32> to memref<1x64xf32, strided<[4096, 4], offset: 4480>>
+  %0 = memref.subview %input[2, 256] [2, 256] [2, 2] : memref<4x1024xf32> to memref<2x256xf32, strided<[2048, 2], offset: 2304>>
+  %1 = memref.subview %0[1, 64] [1, 64] [2, 2] : memref<2x256xf32, strided<[2048, 2], offset: 2304>> to memref<1x64xf32, strided<[4096, 4], offset: 4480>>
+  return %1 : memref<1x64xf32, strided<[4096, 4], offset: 4480>>
+}
+
+// -----
+
+func.func @main(%input: memref<30x30xf32>) -> memref<2x2xf32, strided<[240, 8], offset: 217>> {
+  //      CHECK: subview %arg0[7, 7] [2, 2] [8, 8]
+  // CHECK-SAME: memref<30x30xf32> to memref<2x2xf32, strided<[240, 8], offset: 217>>
+  %0 = memref.subview %input[1, 1] [12, 12] [2, 2] : memref<30x30xf32> to memref<12x12xf32, strided<[60, 2], offset: 31>>
+  %1 = memref.subview %0[1, 1] [5, 5] [2, 2] : memref<12x12xf32, strided<[60, 2], offset: 31>> to memref<5x5xf32, strided<[120, 4], offset: 93>>
+  %2 = memref.subview %1[1, 1] [2, 2] [2, 2] : memref<5x5xf32, strided<[120, 4], offset: 93>> to memref<2x2xf32, strided<[240, 8], offset: 217>>
+  return %2 : memref<2x2xf32, strided<[240, 8], offset: 217>> 
+}
+
+// -----
+
+func.func @main(%input: memref<4x1024xf32>) -> memref<1x64xf32, strided<[4096, 4], offset: ?>> {
+  //      CHECK:%[[VAL_1:.*]] = arith.constant 4 : index
+  %cst_2 = arith.constant 2 : index
+  //      CHECK:%[[VAL_2:.*]] = arith.constant 384 : index
+  %cst_64 = arith.constant 64 : index
+  //      CHECK: subview %arg0{{\[}}%[[VAL_1]], %[[VAL_2]]] [1, 64] [4, 4]
+  // CHECK-SAME: memref<4x1024xf32> to memref<1x64xf32, strided<[4096, 4], offset: ?>>
+  %0 = memref.subview %input[%cst_2, 256] [2, 256] [2, 2] : memref<4x1024xf32> to memref<2x256xf32, strided<[2048, 2], offset: ?>>
+  %1 = memref.subview %0[1, %cst_64] [1, 64] [2, 2] : memref<2x256xf32, strided<[2048, 2], offset: ?>> to memref<1x64xf32, strided<[4096, 4], offset: ?>>
+  return %1 : memref<1x64xf32, strided<[4096, 4], offset: ?>>
+}
+
+// -----
+
+func.func @main(%input: memref<4x1024xf32>) -> memref<1x64xf32, strided<[4096, 4], offset: ?>> {
+  //      CHECK: %[[VAL_1:.*]] = arith.constant 4 : index
+  %cst_1 = arith.constant 1 : index
+  %cst_2 = arith.constant 2 : index
+  //      CHECK: subview %arg0{{\[}}%[[VAL_1]], 384] [1, 64] [4, 4]
+  // CHECK-SAME: memref<4x1024xf32> to memref<1x64xf32, strided<[4096, 4], offset: ?>>
+  %0 = memref.subview %input[%cst_2, 256] [2, 256] [2, 2] : memref<4x1024xf32> to memref<2x256xf32, strided<[2048, 2], offset: ?>>
+  %1 = memref.subview %0[%cst_1, 64] [1, 64] [2, 2] : memref<2x256xf32, strided<[2048, 2], offset: ?>> to memref<1x64xf32, strided<[4096, 4], offset: ?>>
+  return %1 : memref<1x64xf32, strided<[4096, 4], offset: ?>>
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/80551


More information about the Mlir-commits mailing list