1 //===- ReleaseModeModelRunner.h - Fast, precompiled model runner  ---------===//
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 // This file implements a model runner wrapping an AOT compiled ML model.
10 // Only inference is supported.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef LLVM_ANALYSIS_RELEASEMODEMODELRUNNER_H
15 #define LLVM_ANALYSIS_RELEASEMODEMODELRUNNER_H
16 
17 #include "llvm/Analysis/MLModelRunner.h"
18 #include "llvm/Analysis/TensorSpec.h"
19 #include "llvm/Support/ErrorHandling.h"
20 
21 #include <memory>
22 #include <vector>
23 
24 namespace llvm {
25 
26 /// ReleaseModeModelRunner - production mode implementation of the
27 /// MLModelRunner. It uses an AOT-compiled SavedModel for efficient execution.
28 template <class TGen>
29 class ReleaseModeModelRunner final : public MLModelRunner {
30 public:
31   /// FeatureNames' type should be an indexed collection of std::string, like
32   /// std::array or std::vector, that has a size() method.
33   template <class FType>
34   ReleaseModeModelRunner(LLVMContext &Ctx, const FType &InputSpec,
35                          StringRef DecisionName, StringRef FeedPrefix = "feed_",
36                          StringRef FetchPrefix = "fetch_")
37       : MLModelRunner(Ctx, MLModelRunner::Kind::Release, InputSpec.size()),
38         CompiledModel(std::make_unique<TGen>()) {
39     assert(CompiledModel && "The CompiledModel should be valid");
40 
41     for (size_t I = 0; I < InputSpec.size(); ++I) {
42       const int Index =
43           CompiledModel->LookupArgIndex(FeedPrefix.str() + InputSpec[I].name());
44       void *Buffer = nullptr;
45       if (Index >= 0)
46         Buffer = CompiledModel->arg_data(Index);
47       setUpBufferForTensor(I, InputSpec[I], Buffer);
48     }
49 
50     ResultIndex = CompiledModel->LookupResultIndex(FetchPrefix.str() +
51                                                    DecisionName.str());
52     assert(ResultIndex >= 0 && "Cannot find DecisionName in inlining model");
53   }
54 
55   virtual ~ReleaseModeModelRunner() = default;
56 
classof(const MLModelRunner * R)57   static bool classof(const MLModelRunner *R) {
58     return R->getKind() == MLModelRunner::Kind::Release;
59   }
60 
61 private:
evaluateUntyped()62   void *evaluateUntyped() override {
63     CompiledModel->Run();
64     return CompiledModel->result_data(ResultIndex);
65   }
66 
67   int32_t ResultIndex = -1;
68   std::unique_ptr<TGen> CompiledModel;
69 };
70 
71 /// A mock class satisfying the interface expected by ReleaseModeModelRunner for
72 /// its `TGen` parameter. Useful to avoid conditional compilation complexity, as
73 /// a compile-time replacement for a real AOT-ed model.
74 class NoopSavedModelImpl final {
75 #define NOOP_MODEL_ERRMSG                                                      \
76   "The mock AOT-ed saved model is a compile-time stub and should not be "      \
77   "called."
78 
79 public:
80   NoopSavedModelImpl() = default;
LookupArgIndex(const std::string &)81   int LookupArgIndex(const std::string &) { llvm_unreachable(NOOP_MODEL_ERRMSG); }
LookupResultIndex(const std::string &)82   int LookupResultIndex(const std::string &) { llvm_unreachable(NOOP_MODEL_ERRMSG); }
Run()83   void Run() { llvm_unreachable(NOOP_MODEL_ERRMSG); }
result_data(int)84   void *result_data(int) { llvm_unreachable(NOOP_MODEL_ERRMSG); }
arg_data(int)85   void *arg_data(int) { llvm_unreachable(NOOP_MODEL_ERRMSG); }
86 #undef NOOP_MODEL_ERRMSG
87 };
88 
isEmbeddedModelEvaluatorValid()89 template <class T> bool isEmbeddedModelEvaluatorValid() { return true; }
90 
91 template <> inline bool isEmbeddedModelEvaluatorValid<NoopSavedModelImpl>() {
92   return false;
93 }
94 } // namespace llvm
95 
96 #endif // LLVM_ANALYSIS_RELEASEMODEMODELRUNNER_H
97