[mlir][bufferization] Support custom types at function boundaries (#159766)

Support custom types (3/N): allow custom tensor and buffer types in
function signatures and at call-sites. This is one of the major building
blocks to move in the direction of module-level one-shot-bufferization
support.

To achieve this, `BufferizationOptions::FunctionArgTypeConverterFn`
callback is converted to work with tensor-like and buffer-like types,
instead of the builtin counterparts. The default behavior for builtins
remains unchanged, while custom types by default go through
`TensorLikeType::getBufferType()` which is a general conversion
interface.

Author: andrey-golubev

mlir/include/mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h

@@ -260,12 +260,12 @@ struct BufferizationOptions {
       std::function<LogicalResult(OpBuilder &, Location, Value, Value)>;
   /// Initializer function for analysis state.
   using AnalysisStateInitFn = std::function<void(AnalysisState &)>;
-  /// Tensor -> MemRef type converter.
-  /// Parameters: tensor type, memory space, func op, bufferization options
+  /// Tensor-like -> Buffer-like type conversion.
+  /// Parameters: tensor-like type, memory space, func op, bufferization options
   using FunctionArgTypeConverterFn =
-      std::function<BaseMemRefType(TensorType, Attribute memorySpace,
+      std::function<BufferLikeType(TensorLikeType, Attribute memorySpace,
                                    func::FuncOp, const BufferizationOptions &)>;
-  /// Tensor -> MemRef type converter.
+  /// Tensor -> MemRef type conversion.
   /// Parameters: tensor type, memory space, bufferization options
   using UnknownTypeConverterFn = std::function<BaseMemRefType(
       TensorType, Attribute memorySpace, const BufferizationOptions &)>;
@@ -345,10 +345,12 @@ struct BufferizationOptions {
   /// predictable.
   void setFunctionBoundaryTypeConversion(LayoutMapOption layoutMapOption);
 
-  /// Type converter from tensors to memrefs. This type converter is used to
-  /// determine bufferized function argument and result types. By default, a
-  /// type converter that returns a memref type with a fully dynamic layout map
-  /// is used.
+  /// Type conversion from tensors to buffers. This type conversion is used to
+  /// determine bufferized function argument and result types.
+  ///
+  /// By default, if tensor is a (builtin) tensor type, it is converted to a
+  /// memref type with a fully dynamic layout map; if tensor is a (generic)
+  /// tensor-like type, it is converted using TensorLikeType::getBufferType().
   ///
   /// If `bufferizeFunctionBoundaries` is not set, this function isn't used.
   FunctionArgTypeConverterFn functionArgTypeConverterFn = nullptr;
@@ -360,10 +362,9 @@ struct BufferizationOptions {
   /// If `bufferizeFunctionBoundaries` is not set, this flag has no effect.
   bool inferFunctionResultLayout = true;
 
-  /// Type converter from tensors to memrefs. This type converter is used if no
-  /// memref type could be inferred during bufferization. By default, a type
-  /// converter that returns a memref type with a fully dynamic layout map is
-  /// used.
+  /// Type conversion from tensors to memrefs. This type conversion is used if
+  /// no memref type could be inferred during bufferization. By default, returns
+  /// a memref type with a fully dynamic layout map.
   UnknownTypeConverterFn unknownTypeConverterFn = nullptr;
 
   // Use during type conversion to determine the memory space for memref based

mlir/lib/Dialect/Bufferization/IR/BufferizableOpInterface.cpp

@@ -314,11 +314,21 @@ bool OpFilter::isOpAllowed(Operation *op) const {
 namespace {
 
 /// Default function arg type converter: Use a fully dynamic layout map.
-BaseMemRefType
-defaultFunctionArgTypeConverter(TensorType type, Attribute memorySpace,
+BufferLikeType
+defaultFunctionArgTypeConverter(TensorLikeType type, Attribute memorySpace,
                                 func::FuncOp funcOp,
                                 const BufferizationOptions &options) {
-  return getMemRefTypeWithFullyDynamicLayout(type, memorySpace);
+  if (auto tensorType = mlir::dyn_cast<TensorType>(type)) {
+    return cast<BufferLikeType>(
+        getMemRefTypeWithFullyDynamicLayout(tensorType, memorySpace));
+  }
+
+  // If not builtin, fallback to TensorLikeType::getBufferType()
+  auto bufferType =
+      type.getBufferType(options, [&]() { return funcOp->emitError(); });
+  assert(succeeded(bufferType) &&
+         "a valid buffer is always expected at function boundary");
+  return *bufferType;
 }
 /// Default unknown type converter: Use a fully dynamic layout map.
 BaseMemRefType
@@ -361,14 +371,25 @@ BufferizationOptions::dynCastBufferizableOp(Value value) const {
 
 void BufferizationOptions::setFunctionBoundaryTypeConversion(
     LayoutMapOption layoutMapOption) {
-  functionArgTypeConverterFn = [=](TensorType tensorType, Attribute memorySpace,
+  functionArgTypeConverterFn = [=](TensorLikeType type, Attribute memorySpace,
                                    func::FuncOp funcOp,
                                    const BufferizationOptions &options) {
-    if (layoutMapOption == LayoutMapOption::IdentityLayoutMap)
-      return bufferization::getMemRefTypeWithStaticIdentityLayout(tensorType,
-                                                                  memorySpace);
-    return bufferization::getMemRefTypeWithFullyDynamicLayout(tensorType,
-                                                              memorySpace);
+    if (auto tensorType = mlir::dyn_cast<TensorType>(type)) {
+      if (layoutMapOption == LayoutMapOption::IdentityLayoutMap)
+        return cast<BufferLikeType>(
+            bufferization::getMemRefTypeWithStaticIdentityLayout(tensorType,
+                                                                 memorySpace));
+      return cast<BufferLikeType>(
+          bufferization::getMemRefTypeWithFullyDynamicLayout(tensorType,
+                                                             memorySpace));
+    }
+
+    // If not builtin, fallback to TensorLikeType::getBufferType()
+    auto bufferType =
+        type.getBufferType(options, [&]() { return funcOp->emitError(); });
+    assert(succeeded(bufferType) &&
+           "a valid buffer is always expected at function boundary");
+    return *bufferType;
   };
   inferFunctionResultLayout =
       layoutMapOption == LayoutMapOption::InferLayoutMap;

mlir/lib/Dialect/Bufferization/Transforms/Bufferize.cpp

@@ -425,7 +425,7 @@ bufferization::bufferizeBlockSignature(Block *block, RewriterBase &rewriter,
   // Compute the new signature.
   SmallVector<Type> newTypes;
   for (BlockArgument &bbArg : block->getArguments()) {
-    auto tensorType = dyn_cast<TensorType>(bbArg.getType());
+    auto tensorType = dyn_cast<TensorLikeType>(bbArg.getType());
     if (!tensorType) {
       newTypes.push_back(bbArg.getType());
       continue;

mlir/lib/Dialect/Bufferization/Transforms/FuncBufferizableOpInterfaceImpl.cpp

@@ -50,29 +50,47 @@ void FuncAnalysisState::startFunctionAnalysis(FuncOp funcOp) {
 #endif // NDEBUG
 }
 
+// Note: this is a local adaptor to unify TensorType and TensorLikeType code
+// paths that both work with BufferizationOptions.
+static mlir::Attribute
+getDefaultMemorySpace(const BufferizationOptions &options,
+                      TensorLikeType type) {
+  if (auto tensorType = dyn_cast<TensorType>(type)) {
+    return *options.defaultMemorySpaceFn(tensorType);
+  }
+  return nullptr;
+}
+
 /// Return the index-th bufferized function argument type. This assumes that the
 /// specified argument is a tensor. If the tensor is ranked, a layout map may be
 /// specified by the user (as per `options.functionArgTypeConverterFn`).
-static BaseMemRefType
+static BufferLikeType
 getBufferizedFunctionArgType(FuncOp funcOp, int64_t index,
                              const BufferizationOptions &options) {
-  auto tensorType =
-      dyn_cast<TensorType>(funcOp.getFunctionType().getInput(index));
-  assert(tensorType && "expected TensorType");
-
-  BaseMemRefType memrefType = options.functionArgTypeConverterFn(
-      tensorType, *options.defaultMemorySpaceFn(tensorType), funcOp, options);
-
-  auto layoutAttr = funcOp.getArgAttrOfType<AffineMapAttr>(
-      index, BufferizationDialect::kBufferLayoutAttrName);
-  if (!layoutAttr)
-    return memrefType;
-
-  auto rankedMemrefType = dyn_cast<MemRefType>(memrefType);
-  assert(rankedMemrefType && "buffer layout not supported on unranked tensors");
-  return MemRefType::get(
-      rankedMemrefType.getShape(), rankedMemrefType.getElementType(),
-      layoutAttr.getValue(), rankedMemrefType.getMemorySpace());
+  auto type =
+      dyn_cast<TensorLikeType>(funcOp.getFunctionType().getInput(index));
+  assert(type && "expected TensorLikeType");
+
+  // Note: For builtin tensors there is additional logic related to layout.
+  if (auto tensorType = dyn_cast<TensorType>(type)) {
+    BufferLikeType memrefType = options.functionArgTypeConverterFn(
+        type, *options.defaultMemorySpaceFn(tensorType), funcOp, options);
+
+    auto layoutAttr = funcOp.getArgAttrOfType<MemRefLayoutAttrInterface>(
+        index, BufferizationDialect::kBufferLayoutAttrName);
+    if (!layoutAttr)
+      return memrefType;
+
+    auto rankedMemrefType = dyn_cast<MemRefType>(memrefType);
+    assert(rankedMemrefType &&
+           "buffer layout not supported on unranked tensors");
+    return cast<BufferLikeType>(MemRefType::get(
+        rankedMemrefType.getShape(), rankedMemrefType.getElementType(),
+        layoutAttr, rankedMemrefType.getMemorySpace()));
+  }
+
+  return options.functionArgTypeConverterFn(type, /*memSpace=*/nullptr, funcOp,
+                                            options);
 }
 
 /// Return the FuncOp called by `callOp`.
@@ -207,13 +225,13 @@ struct CallOpInterface
     FunctionType funcType = funcOp.getFunctionType();
     Type resultType =
         funcType.getResult(cast<OpResult>(value).getResultNumber());
-    if (auto bufferizedType = dyn_cast<BaseMemRefType>(resultType))
-      return cast<BufferLikeType>(bufferizedType);
+    if (auto bufferizedType = dyn_cast<BufferLikeType>(resultType))
+      return bufferizedType;
 
     // Otherwise, call the type converter to compute the bufferized type.
-    auto tensorType = cast<TensorType>(resultType);
+    auto tensorType = cast<TensorLikeType>(resultType);
     return cast<BufferLikeType>(options.functionArgTypeConverterFn(
-        tensorType, *options.defaultMemorySpaceFn(tensorType), funcOp,
+        tensorType, getDefaultMemorySpace(options, tensorType), funcOp,
         options));
   }
 
@@ -227,7 +245,7 @@ struct CallOpInterface
     SmallVector<Type> resultTypes;
     for (Value result : callOp.getResults()) {
       Type returnType = result.getType();
-      if (!isa<TensorType>(returnType)) {
+      if (!isa<TensorLikeType>(returnType)) {
         // Non-tensor values are returned.
         resultTypes.push_back(returnType);
         continue;
@@ -250,7 +268,7 @@ struct CallOpInterface
 
     for (OpOperand &opOperand : callOp->getOpOperands()) {
       // Non-tensor operands are just copied.
-      if (!isa<TensorType>(opOperand.get().getType())) {
+      if (!isa<TensorLikeType>(opOperand.get().getType())) {
         newOperands.push_back(opOperand.get());
         continue;
       }
@@ -263,8 +281,8 @@ struct CallOpInterface
       Value buffer = *maybeBuffer;
 
       // Caller / callee type mismatch is handled with castOrReallocMemRefValue.
-      auto memRefType = funcType.getInput(opOperand.getOperandNumber());
-      if (!isa<BaseMemRefType>(memRefType)) {
+      auto bufferType = funcType.getInput(opOperand.getOperandNumber());
+      if (!isa<BufferLikeType>(bufferType)) {
         // The called function was not bufferized yet. This can happen when
         // there cycles in the function call graph. Compute the bufferized
         // result type.
@@ -273,7 +291,7 @@ struct CallOpInterface
                 funcOp.getArgument(opOperand.getOperandNumber()), options);
         if (failed(maybeBufferType))
           return failure();
-        memRefType = *maybeBufferType;
+        bufferType = *maybeBufferType;
       }
 
       // Since we don't yet have a clear layout story, to_buffer may
@@ -282,8 +300,8 @@ struct CallOpInterface
       // that will either canonicalize away or fail compilation until we can do
       // something better. Insert a reallocation + copy if it cannot be
       // statically guaranteed that a direct cast would be valid.
-      if (buffer.getType() != memRefType) {
-        auto memrefDstType = dyn_cast<MemRefType>(memRefType);
+      if (buffer.getType() != bufferType) {
+        auto memrefDstType = dyn_cast<MemRefType>(bufferType);
         assert(memrefDstType &&
                "buffer layout not supported on unranked tensors");
         FailureOr<Value> replacement = bufferization::castOrReallocMemRefValue(
@@ -345,7 +363,7 @@ struct FuncOpInterface
   static bool supportsUnstructuredControlFlow() { return true; }
 
   bool hasTensorSemantics(Operation *op) const {
-    auto isaTensor = llvm::IsaPred<TensorType>;
+    auto isaTensor = llvm::IsaPred<TensorLikeType>;
 
     // A function has tensor semantics if it has tensor arguments/results.
     auto funcOp = cast<FuncOp>(op);
@@ -380,8 +398,8 @@ struct FuncOpInterface
 
     // Function arguments are special.
     if (bbArg.getOwner() == &funcOp.getBody().front())
-      return cast<BufferLikeType>(
-          getBufferizedFunctionArgType(funcOp, bbArg.getArgNumber(), options));
+      return getBufferizedFunctionArgType(funcOp, bbArg.getArgNumber(),
+                                          options);
 
     return OpWithUnstructuredControlFlowBufferizableOpInterfaceExternalModel::
         getBufferType(op, value, options, invocationStack);
@@ -403,7 +421,7 @@ struct FuncOpInterface
     SmallVector<Type> argTypes;
     for (const auto &it : llvm::enumerate(funcType.getInputs())) {
       Type argType = it.value();
-      if (isa<TensorType>(argType)) {
+      if (isa<TensorLikeType>(argType)) {
         argTypes.push_back(
             getBufferizedFunctionArgType(funcOp, it.index(), options));
         continue;
@@ -414,9 +432,9 @@ struct FuncOpInterface
     // Compute the result types.
     SmallVector<Type> retTypes;
     for (Type resultType : funcType.getResults()) {
-      if (auto tensorType = dyn_cast<TensorType>(resultType)) {
-        BaseMemRefType resultType = options.functionArgTypeConverterFn(
-            tensorType, *options.defaultMemorySpaceFn(tensorType), funcOp,
+      if (auto tensorType = dyn_cast<TensorLikeType>(resultType)) {
+        BufferLikeType resultType = options.functionArgTypeConverterFn(
+            tensorType, getDefaultMemorySpace(options, tensorType), funcOp,
             options);
         retTypes.push_back(resultType);
         continue;
@@ -446,7 +464,7 @@ struct FuncOpInterface
       SmallVector<Value> returnValues;
       for (auto [returnVal, bufferizedType] :
            llvm::zip_equal(returnOp->getOperands(), retTypes)) {
-        auto tensorType = dyn_cast<TensorType>(returnVal.getType());
+        auto tensorType = dyn_cast<TensorLikeType>(returnVal.getType());
         rewriter.setInsertionPoint(returnOp);
 
         // If not a tensor type just forward it.

mlir/test/Dialect/Bufferization/Transforms/one-shot-module-bufferize.mlir

@@ -796,3 +796,58 @@ func.func @result_type_mismatch(%c: i1) -> tensor<5xf32> {
   return %1 : tensor<5xf32>
 }
 
+// -----
+
+// CHECK:   func.func @custom_types(
+// CHECK-SAME:    %[[arg:.*]]: !test.test_memref<[4, 4], f64>
+// CHECK-SAME:  ) -> (!test.test_memref<[4, 8], f64>,
+// CHECK-SAME:        !test.test_memref<[4, 8], f64>)
+func.func @custom_types(%arg: !test.test_tensor<[4, 4], f64>)
+    -> (!test.test_tensor<[4, 8], f64>, !test.test_tensor<[4, 8], f64>) {
+  // CHECK: %[[out1:.*]] = "test.dummy_memref_op"(%[[arg]]) :
+  // CHECK-SAME: (!test.test_memref<[4, 4], f64>) -> !test.test_memref<[4, 8], f64>
+  %out1 = "test.dummy_tensor_op"(%arg) : (!test.test_tensor<[4, 4], f64>)
+    -> !test.test_tensor<[4, 8], f64>
+
+  // CHECK: %[[alloc:.*]] = "test.create_memref_op"
+  // CHECK: %[[out2:.*]] = "test.dummy_memref_op"(%[[alloc]])
+  // CHECK-SAME: (!test.test_memref<[4, 4], f64>) -> !test.test_memref<[4, 8], f64>
+  %alloc = "test.create_tensor_op"() : () -> !test.test_tensor<[4, 4], f64>
+  %out2 = "test.dummy_tensor_op"(%alloc) : (!test.test_tensor<[4, 4], f64>)
+    -> !test.test_tensor<[4, 8], f64>
+
+  // CHECK: return %[[out1]], %[[out2]]
+  return %out1, %out2 :
+    !test.test_tensor<[4, 8], f64>, !test.test_tensor<[4, 8], f64>
+}
+
+// -----
+
+// CHECK:   func.func @custom_types_foo(
+// CHECK-SAME:    %[[arg:.*]]: !test.test_memref<[4, 4], f64>
+// CHECK-SAME:  ) -> !test.test_memref<[4, 4], f64>
+func.func @custom_types_foo(%arg: !test.test_tensor<[4, 4], f64>)
+    -> !test.test_tensor<[4, 4], f64> {
+  // CHECK: %[[out:.*]] = "test.dummy_memref_op"(%[[arg]])
+  %out = "test.dummy_tensor_op"(%arg) : (!test.test_tensor<[4, 4], f64>)
+    -> !test.test_tensor<[4, 4], f64>
+  // CHECK: return %[[out]]
+  return %out : !test.test_tensor<[4, 4], f64>
+}
+
+// CHECK:   func.func @custom_types_bar(
+// CHECK-SAME:    %[[arg:.*]]: !test.test_memref<[4, 4], f64>
+// CHECK-SAME:  ) -> !test.test_memref<[4, 8], f64>
+func.func @custom_types_bar(%arg: !test.test_tensor<[4, 4], f64>)
+    -> !test.test_tensor<[4, 8], f64> {
+  // CHECK: %[[call:.*]] = call @custom_types_foo(%[[arg]])
+  %call = func.call @custom_types_foo(%arg) : (!test.test_tensor<[4, 4], f64>)
+    -> !test.test_tensor<[4, 4], f64>
+
+  // CHECK: %[[out:.*]] = "test.dummy_memref_op"(%[[call]])
+  %out = "test.dummy_tensor_op"(%call) : (!test.test_tensor<[4, 4], f64>)
+    -> !test.test_tensor<[4, 8], f64>
+
+  // CHECK: return %[[out]]
+  return %out : !test.test_tensor<[4, 8], f64>
+}