1 //===- TensorSpec.h - type descriptor for a tensor --------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 #ifndef LLVM_ANALYSIS_TENSORSPEC_H
10 #define LLVM_ANALYSIS_TENSORSPEC_H
11 
12 #include "llvm/Config/llvm-config.h"
13 
14 #include "llvm/ADT/StringMap.h"
15 #include "llvm/IR/LLVMContext.h"
16 #include "llvm/Support/JSON.h"
17 
18 #include <memory>
19 #include <optional>
20 #include <vector>
21 
22 namespace llvm {
23 /// TensorSpec encapsulates the specification of a tensor: its dimensions, or
24 /// "shape" (row-major), its type (see TensorSpec::getDataType specializations
25 /// for supported types), its name and port (see "TensorFlow: Large-Scale
26 /// Machine Learning on Heterogeneous Distributed Systems", section 4.2, para 2:
27 /// https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45166.pdf)
28 ///
29 /// Known tensor types. The left part is the C type, the right is a name we
30 /// can use to identify the type (to implement TensorSpec equality checks), and
31 /// to use, if needed, when mapping to an underlying evaluator's type system.
32 /// The main requirement is that the C type we use has the same size and
33 /// encoding (e.g. endian-ness) as the one used by the evaluator.
34 #define SUPPORTED_TENSOR_TYPES(M)                                              \
35   M(float, Float)                                                              \
36   M(double, Double)                                                            \
37   M(int8_t, Int8)                                                              \
38   M(uint8_t, UInt8)                                                            \
39   M(int16_t, Int16)                                                            \
40   M(uint16_t, UInt16)                                                          \
41   M(int32_t, Int32)                                                            \
42   M(uint32_t, UInt32)                                                          \
43   M(int64_t, Int64)                                                            \
44   M(uint64_t, UInt64)
45 
46 enum class TensorType {
47   Invalid,
48 #define _TENSOR_TYPE_ENUM_MEMBERS(_, Name) Name,
49   SUPPORTED_TENSOR_TYPES(_TENSOR_TYPE_ENUM_MEMBERS)
50 #undef _TENSOR_TYPE_ENUM_MEMBERS
51       Total
52 };
53 
54 class TensorSpec final {
55 public:
56   template <typename T>
57   static TensorSpec createSpec(const std::string &Name,
58                                const std::vector<int64_t> &Shape,
59                                int Port = 0) {
60     return TensorSpec(Name, Port, getDataType<T>(), sizeof(T), Shape);
61   }
62 
63   const std::string &name() const { return Name; }
64   int port() const { return Port; }
65   TensorType type() const { return Type; }
66   const std::vector<int64_t> &shape() const { return Shape; }
67 
68   bool operator==(const TensorSpec &Other) const {
69     return Name == Other.Name && Port == Other.Port && Type == Other.Type &&
70            Shape == Other.Shape;
71   }
72 
73   bool operator!=(const TensorSpec &Other) const { return !(*this == Other); }
74 
75   /// Get the number of elements in a tensor with this shape.
76   size_t getElementCount() const { return ElementCount; }
77   /// Get the size, in bytes, of one element.
78   size_t getElementByteSize() const { return ElementSize; }
79   /// Get the total size of a memory buffer needed to store the whole tensor.
80   size_t getTotalTensorBufferSize() const { return ElementCount * ElementSize; }
81 
82   template <typename T> bool isElementType() const {
83     return getDataType<T>() == Type;
84   }
85 
86   TensorSpec(const std::string &NewName, const TensorSpec &Other)
87       : TensorSpec(NewName, Other.Port, Other.Type, Other.ElementSize,
88                    Other.Shape) {}
89 
90   void toJSON(json::OStream &OS) const;
91 
92 private:
93   TensorSpec(const std::string &Name, int Port, TensorType Type,
94              size_t ElementSize, const std::vector<int64_t> &Shape);
95 
96   template <typename T> static TensorType getDataType();
97 
98   std::string Name;
99   int Port = 0;
100   TensorType Type = TensorType::Invalid;
101   std::vector<int64_t> Shape;
102   size_t ElementCount = 0;
103   size_t ElementSize = 0;
104 };
105 
106 /// For debugging.
107 std::string tensorValueToString(const char *Buffer, const TensorSpec &Spec);
108 
109 /// Construct a TensorSpec from a JSON dictionary of the form:
110 /// { "name": <string>,
111 ///   "port": <int>,
112 ///   "type": <string. Use LLVM's types, e.g. float, double, int64_t>,
113 ///   "shape": <array of ints> }
114 /// For the "type" field, see the C++ primitive types used in
115 /// TFUTILS_SUPPORTED_TYPES.
116 std::optional<TensorSpec> getTensorSpecFromJSON(LLVMContext &Ctx,
117                                                 const json::Value &Value);
118 
119 #define TFUTILS_GETDATATYPE_DEF(T, Name)                                       \
120   template <> TensorType TensorSpec::getDataType<T>();
121 SUPPORTED_TENSOR_TYPES(TFUTILS_GETDATATYPE_DEF)
122 
123 #undef TFUTILS_GETDATATYPE_DEF
124 } // namespace llvm
125 
126 #endif // LLVM_ANALYSIS_TENSORSPEC_H
127