Improve support for krnl.dim (#317)

* Reorganize main function.

* Follow review comments.

* Emit constants are globals in Krnl and LLVM dialects.

* Make krnl dim more robust.

* Format.

* Update comments.

* Change pass name.

src/InitOMPasses.hpp

@@ -77,7 +77,8 @@ void initOMPasses() {
         return mlir::createPackKrnlGlobalConstantsPass();
       });
 
-  mlir::registerPass("disconnect-dims", "Disconnect dims from allocs.",
+  mlir::registerPass("lower-krnl-shape-to-std",
+      "Lowers krnl shape-related operations.",
       []() -> std::unique_ptr<mlir::Pass> {
         return mlir::createDisconnectKrnlDimFromAllocPass();
       });

src/Transform/DisconnectKrnlDimFromAlloc.cpp

@@ -27,13 +27,37 @@ namespace {
 /*!
  *  RewritePattern that replaces:
  *    %0 = alloc(%d) : memref<?x10x<type>, #map>
- *    %1 = krnl.dim(%0, 0) : memref<?x10x<type>>
+ *    %1 = krnl.dim(%0, 0) : (memref<?x10x<type>, #map>, index) -> index
- *    %2 = krnl.dim(%0, 1) : memref<?x10x<type>>
+ *    %2 = krnl.dim(%0, 1) : (memref<?x10x<type>, #map>, index) -> index
  *    %3 = add %1, %2
  *  with:
  *    %0 = alloc(%d) : memref<?x10x<type>, #map>
  *    %2 = constant 10 : index
  *    %3 = add %d, %2
+ *
+ *  When the first argument of the krnl.dim is an input argument
+ * i.e. it is not the output of an alloc operation, we emit either
+ * the constant or the strandard dim operation depending on whether
+ * the dimension is static or dynamic.
+ *
+ *  function(%arg0 : memref<?x10x<type>>) {
+ *    %0 = krnl.dim(%arg0, 0) : (memref<?x10x<type>>, index) -> index
+ *    %1 = krnl.dim(%arg0, 1) : memref<?x10x<type>>
+ *  }
+ *
+ *
+ *  becomes:
+ *
+ *  function(%arg0 : memref<?x10x<type>>) {
+ *    %0 = dim %arg0, 0 : (memref<?x10x<type>>, index) -> index
+ *    %1 = constant 10 : index
+ *  }
+ *
+ *  The following case is not supported:
+ *
+ *  function(%arg0 : memref<?x10x<type>, #map>) {
+ *    %0 = krnl.dim(%arg0, 0) : (memref<?x10x<type>, #map>, index) -> index
+ *  }
  */
 
 class DisconnectKrnlDimFromAlloc : public OpRewritePattern<KrnlDimOp> {
@@ -53,25 +77,40 @@ public:
     // Get the integer value of the index.
     int64_t index = indexOp.getAttrOfType<IntegerAttr>("value").getInt();
 
-    // Get defining operation for the MemRef argument.
+    // Get the shape of the MemRef argument.
-    AllocOp allocOp = dyn_cast<AllocOp>(krnlDimOp.alloc().getDefiningOp());
+    auto memRefType = convertToMemRefType(krnlDimOp.alloc().getType());
-    auto memRefShape =
+    auto memRefShape = memRefType.getShape();
-        convertToMemRefType(allocOp.getResult().getType()).getShape();
     auto rank = memRefShape.size();
     assert(index >= 0 && index < rank && "Index must be in bounds");
 
+    // Get the defining operation of the first argument of krnl.dim.
+    // If this operation is not an alloc, and the value comes from the
+    // list of input arguments, the support is limited to MemRefs without
+    // maps.
+    auto firstArgDefOp = krnlDimOp.alloc().getDefiningOp();
+
     Value result;
     if (memRefShape[index] > -1) {
       // If dimension is static, then we can just emit the constant value.
       result = rewriter.create<ConstantOp>(loc,
           rewriter.getIntegerAttr(rewriter.getIndexType(), memRefShape[index]));
-    } else {
+    } else if (firstArgDefOp && isa<AllocOp>(firstArgDefOp)) {
+      // Get defining operation for the MemRef argument.
+      AllocOp allocOp = dyn_cast<AllocOp>(krnlDimOp.alloc().getDefiningOp());
+
       // If dimension is dynamic we need to return the input alloc Value which
       // corresponds to it.
       int64_t dynDimIdx = getAllocArgIndex(allocOp, index);
       assert(dynDimIdx >= 0 && dynDimIdx < allocOp.getOperands().size() &&
              "Dynamic index outside range of alloc argument list.");
       result = allocOp.getOperands()[dynDimIdx];
+    } else if (memRefType.getAffineMaps().empty()) {
+      // Use a standard DimOp since no map is present.
+      result =
+          rewriter.create<DimOp>(loc, krnlDimOp.alloc(), krnlDimOp.index());
+    } else {
+      llvm_unreachable(
+          "dynamic sized MemRef with map must be defined by an AllocOp");
     }
 
     rewriter.replaceOp(krnlDimOp, result);

test/mlir/krnl/krnl_disconnect_dim_from_alloc.mlir

@@ -1,5 +1,6 @@
-// RUN: onnx-mlir-opt --disconnect-dims %s -split-input-file | FileCheck %s
+// RUN: onnx-mlir-opt --lower-krnl-shape-to-std %s -split-input-file | FileCheck %s
 
+/// Lower krnl.dim when input MemRef does not have an affine map.
 func @test_krnl_dim_lowering(%arg0: memref<?x?xf32>) -> index {
   %c1 = constant 1 : index
   %c0 = constant 0 : index
@@ -21,6 +22,7 @@ func @test_krnl_dim_lowering(%arg0: memref<?x?xf32>) -> index {
 
 // -----
 
+/// Lower krnl.dim when input MemRef has an affine map.
 #map = affine_map<(d0, d1) -> (d1, d0)>
 func @test_krnl_dim_lowering_with_map(%arg0: memref<?x?xf32>) -> index {
   %c1 = constant 1 : index
@@ -41,3 +43,32 @@ func @test_krnl_dim_lowering_with_map(%arg0: memref<?x?xf32>) -> index {
   // CHECK: [[SUM:%.+]] = addi [[DIM]], [[CONST10]] : index
   // CHECK: return [[SUM]] : index
 }
+
+// -----
+
+/// Lower krnl.dim to constant when first argument of krnl.dim is an input arg
+/// and the dimensions is static.
+func @test_krnl_dim_lowering_with_const_arg(%arg0: memref<10x20xf32>) -> index {
+  %c0 = constant 0 : index
+  %0 = "krnl.dim"(%arg0, %c0) : (memref<10x20xf32>, index) -> index
+  return %0 : index
+
+  // CHECK-LABEL: test_krnl_dim_lowering_with_const_arg
+  // CHECK: [[CONST10:%.+]] = constant 10 : index
+  // CHECK: return [[CONST10]] : index
+}
+
+// -----
+
+/// Lower krnl.dim to a standard dim operation when first argument of krnl.dim
+/// is an input arg and the dimensions is dynamic.
+func @test_krnl_dim_lowering_with_dynamic_arg(%arg0: memref<10x?xf32>) -> index {
+  %c0 = constant 1 : index
+  %0 = "krnl.dim"(%arg0, %c0) : (memref<10x?xf32>, index) -> index
+  return %0 : index
+
+  // CHECK-LABEL: test_krnl_dim_lowering_with_dynamic_arg
+  // CHECK: [[CONST1:%.+]] = constant 1 : index
+  // CHECK: [[DIM:%.+]] = dim %arg0, [[CONST1]] : memref<10x?xf32>
+  // CHECK: return [[DIM]] : index
+}