[IR] Type system stage4: Add some built-in types and type conversion methods

paddle/ir/builtin_type.cc

@@ -0,0 +1,34 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/ir/builtin_type.h"
+
+namespace ir {
+const ir::Type& DenseTensorType::dtype() const { return storage()->dtype_; }
+
+const ir::DenseTensorTypeStorage::Dim& DenseTensorType::dim() const {
+  return storage()->dims_;
+}
+
+const ir::DenseTensorTypeStorage::DataLayout& DenseTensorType::data_layout()
+    const {
+  return storage()->layout_;
+}
+
+const ir::DenseTensorTypeStorage::LoD& DenseTensorType::lod() const {
+  return storage()->lod_;
+}
+
+const size_t& DenseTensorType::offset() const { return storage()->offset_; }
+}  // namespace ir

paddle/ir/builtin_type.h

@@ -14,22 +14,44 @@
 
 #pragma once
 
+#include "paddle/ir/builtin_type_storage.h"
 #include "paddle/ir/type.h"
 
 namespace ir {
 ///
 /// \brief This macro is used to get a list of all built-in types in this file.
+/// The built-in Dialect will use this macro to quickly register all built-in
+/// types.
 ///
-#define GET_BUILT_IN_TYPE_LIST ir::Float32Type, ir::Int32Type
+#define GET_BUILT_IN_TYPE_LIST                                      \
+  ir::Float16Type, ir::Float32Type, ir::Float64Type, ir::Int16Type, \
+      ir::Int32Type, ir::Int64Type, ir::DenseTensorType
 
 ///
-/// \brief Definitions of built-in type classes. The built-in type object get
-/// method is as follows:
+/// \brief Define built-in parameterless types. Please add the necessary
+/// interface functions for built-in types through the macro
+/// DECLARE_TYPE_UTILITY_FUNCTOR.
+///
+/// NOTE(zhangbo9674): If you need to directly
+/// cache the object of this built-in type in IrContext, please overload the get
+/// method, and construct and cache the object in IrContext. For the specific
+/// implementation method, please refer to Float16Type.
+///
+/// The built-in type object get method is as follows:
 /// \code{cpp}
 ///   ir::IrContext *ctx = ir::IrContext::Instance();
 ///   Type fp32 = Float32Type::get(ctx);
 /// \endcode
 ///
+class Float16Type : public ir::Type {
+ public:
+  using Type::Type;
+
+  DECLARE_TYPE_UTILITY_FUNCTOR(Float16Type, ir::TypeStorage);
+
+  static Float16Type get(ir::IrContext *context);
+};
+
 class Float32Type : public ir::Type {
  public:
   using Type::Type;
@@ -39,6 +61,24 @@ class Float32Type : public ir::Type {
   static Float32Type get(ir::IrContext *context);
 };
 
+class Float64Type : public ir::Type {
+ public:
+  using Type::Type;
+
+  DECLARE_TYPE_UTILITY_FUNCTOR(Float64Type, ir::TypeStorage);
+
+  static Float64Type get(ir::IrContext *context);
+};
+
+class Int16Type : public ir::Type {
+ public:
+  using Type::Type;
+
+  DECLARE_TYPE_UTILITY_FUNCTOR(Int16Type, ir::TypeStorage);
+
+  static Int16Type get(ir::IrContext *context);
+};
+
 class Int32Type : public ir::Type {
  public:
   using Type::Type;
@@ -48,4 +88,33 @@ class Int32Type : public ir::Type {
   static Int32Type get(ir::IrContext *context);
 };
 
+class Int64Type : public ir::Type {
+ public:
+  using Type::Type;
+
+  DECLARE_TYPE_UTILITY_FUNCTOR(Int64Type, ir::TypeStorage);
+
+  static Int64Type get(ir::IrContext *context);
+};
+
+///
+/// \brief Define built-in parameteric types.
+///
+class DenseTensorType : public ir::Type {
+ public:
+  using Type::Type;
+
+  DECLARE_TYPE_UTILITY_FUNCTOR(DenseTensorType, DenseTensorTypeStorage);
+
+  const ir::Type &dtype() const;
+
+  const ir::DenseTensorTypeStorage::Dim &dim() const;
+
+  const ir::DenseTensorTypeStorage::DataLayout &data_layout() const;
+
+  const ir::DenseTensorTypeStorage::LoD &lod() const;
+
+  const size_t &offset() const;
+};
+
 }  // namespace ir

paddle/ir/builtin_type_storage.h

@@ -0,0 +1,156 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <type_traits>
+
+#include "paddle/ir/type.h"
+
+namespace std {
+///
+/// \brief Enable hashing std::vector<T> instances.
+///
+template <typename T>
+struct hash<std::vector<T>> {
+  std::size_t operator()(const std::vector<T> &dim) const {
+    std::size_t seed = 0;
+    for (size_t i = 0; i < dim.size(); ++i) {
+      seed ^= std::hash<T>()(dim[i]) + 0x9e3779b9 + (seed << 6) + (seed >> 2);
+    }
+    return seed;
+  }
+};
+
+}  // namespace std
+
+namespace ir {
+///
+/// \brief Define Parameteric TypeStorage for DenseTensorType.
+///
+/// NOTE(zhangbo9674): The derived TypeStorage class needs to implement the
+/// following methods: (1)declare ParamKey, (2)define Construction method,
+/// (3)define HashValue method, (4)overload operator==.
+///
+struct DenseTensorTypeStorage : public ir::TypeStorage {
+  ///
+  /// \brief It is consistent with the DataLayout defined by Phi operator
+  /// library. See the file for details: paddle/phi/common/layout.h.
+  ///
+  enum class DataLayout : unsigned int {
+    UNDEFINED = 0,
+    NHWC,
+    NCHW,
+    NCDHW,
+    NDHWC,
+    ONEDNN,
+    SPARSE_COO,
+    SPARSE_CSR,
+    PSTRING_UNION,
+
+    NUM_DATA_LAYOUTS,
+
+    // See Note [ Why we need ALL in basic kernel key member? ]
+    ALL_LAYOUT = UNDEFINED,
+
+    // Note: Unify phi DataLayout and fluid::framework::DataLayout,
+    // for compatible with fluid DataLayout, here need prefix `k`
+    kNHWC = NHWC,
+    kNCHW = NCHW,
+    kMKLDNN = ONEDNN,  // all layouts supported by ONEDNN internally
+    kNDHWC = NDHWC,
+    kNCDHW = NCDHW,
+  };
+
+  using Dim = std::vector<int64_t>;
+
+  using LoD = std::vector<std::vector<size_t>>;
+
+  ///
+  /// \brief Declare ParamKey according to parameter type.
+  ///
+  using ParamKey = std::tuple<ir::Type, Dim, DataLayout, LoD, size_t>;
+
+  DenseTensorTypeStorage(
+      ir::Type dtype, Dim dims, DataLayout layout, LoD lod, size_t offset)
+      : dtype_(dtype),
+        dims_(dims),
+        layout_(layout),
+        lod_(lod),
+        offset_(offset) {}
+
+  ///
+  /// \brief Each derived TypeStorage must define a Construc method, which
+  /// StorageManager uses to construct a derived TypeStorage.
+  ///
+  static DenseTensorTypeStorage *Construct(ParamKey key) {
+    return new DenseTensorTypeStorage(std::get<0>(key),
+                                      std::get<1>(key),
+                                      std::get<2>(key),
+                                      std::get<3>(key),
+                                      std::get<4>(key));
+  }
+
+  ///
+  /// \brief Each derived TypeStorage must provide a HashValue method.
+  ///
+  static std::size_t HashValue(const ParamKey &key) {
+    std::size_t hash_value = 0;
+    // hash dtype
+    hash_value =
+        hash_combine(hash_value, std::hash<ir::Type>()(std::get<0>(key)));
+    // hash dims
+    hash_value = hash_combine(hash_value, std::hash<Dim>()(std::get<1>(key)));
+    // hash layout
+    hash_value =
+        hash_combine(hash_value,
+                     std::hash<std::underlying_type<DataLayout>::type>()(
+                         static_cast<std::underlying_type<DataLayout>::type>(
+                             std::get<2>(key))));
+    // hash lod
+    hash_value = hash_combine(hash_value, std::hash<LoD>()(std::get<3>(key)));
+    // hash offset
+    hash_value =
+        hash_combine(hash_value, std::hash<size_t>()(std::get<4>(key)));
+    return hash_value;
+  }
+
+  ///
+  /// \brief Each derived TypeStorage needs to overload operator==.
+  ///
+  bool operator==(const ParamKey &key) const {
+    return ParamKey(dtype_, dims_, layout_, lod_, offset_) == key;
+  }
+
+  ParamKey GetAsKey() const {
+    return ParamKey(dtype_, dims_, layout_, lod_, offset_);
+  }
+
+  ///
+  /// \brief DenseTensorTypeStorage include five parameters: dims, dtype,
+  /// layout, lod, offset.
+  ///
+  ir::Type dtype_;
+  Dim dims_;
+  DataLayout layout_;
+  LoD lod_;
+  size_t offset_;
+
+ private:
+  static std::size_t hash_combine(std::size_t lhs, std::size_t rhs) {
+    return lhs ^= rhs + 0x9e3779b9 + (lhs << 6) + (lhs >> 2);
+  }
+};
+
+}  // namespace ir