xref: /llvm-project/llvm/unittests/Analysis/TFUtilsTest.cpp (revision 0cb9746a7d85000fc7bbd8ac5d8557179ca4521c)
1 //===- TFUtilsTest.cpp - test for TFUtils ---------------------------------===//
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 #include "llvm/Analysis/Utils/TFUtils.h"
10 #include "llvm/Analysis/ModelUnderTrainingRunner.h"
11 #include "llvm/Analysis/TensorSpec.h"
12 #include "llvm/AsmParser/Parser.h"
13 #include "llvm/IR/Dominators.h"
14 #include "llvm/IR/Instructions.h"
15 #include "llvm/IR/LLVMContext.h"
16 #include "llvm/IR/Module.h"
17 #include "llvm/Support/Path.h"
18 #include "llvm/Support/SourceMgr.h"
19 #include "llvm/Testing/Support/SupportHelpers.h"
20 #include "gtest/gtest.h"
21 
22 using namespace llvm;
23 
24 extern const char *TestMainArgv0;
25 
26 // NOTE! This test model is currently also used by test/Transforms/Inline/ML tests
27 //- relevant if updating this model.
28 static std::string getModelPath() {
29   SmallString<128> InputsDir = unittest::getInputFileDirectory(TestMainArgv0);
30   llvm::sys::path::append(InputsDir, "ir2native_x86_64_model");
31   return std::string(InputsDir);
32 }
33 
34 // Test observable behavior when no model is provided.
35 TEST(TFUtilsTest, NoModel) {
36   TFModelEvaluator Evaluator("", {}, {});
37   EXPECT_FALSE(Evaluator.isValid());
38 }
39 
40 // Test we can correctly load a savedmodel and evaluate it.
41 TEST(TFUtilsTest, LoadAndExecuteTest) {
42   // We use the ir2native model for test. We know it has one feature of
43   // dimension (1, 214)
44   const static int64_t KnownSize = 214;
45   std::vector<TensorSpec> InputSpecs{TensorSpec::createSpec<int32_t>(
46       "serving_default_input_1", {1, KnownSize})};
47   std::vector<TensorSpec> OutputSpecs{
48       TensorSpec::createSpec<float>("StatefulPartitionedCall", {1})};
49 
50   TFModelEvaluator Evaluator(getModelPath(), InputSpecs, OutputSpecs);
51   EXPECT_TRUE(Evaluator.isValid());
52 
53   int32_t *V = Evaluator.getInput<int32_t>(0);
54   // Fill it up with 1's, we know the output.
55   for (auto I = 0; I < KnownSize; ++I) {
56     V[I] = 1;
57   }
58   {
59     auto ER = Evaluator.evaluate();
60     EXPECT_TRUE(ER.hasValue());
61     float Ret = *ER->getTensorValue<float>(0);
62     EXPECT_EQ(static_cast<int64_t>(Ret), 80);
63     EXPECT_EQ(ER->getUntypedTensorValue(0),
64               reinterpret_cast<const void *>(ER->getTensorValue<float>(0)));
65   }
66   // The input vector should be unchanged
67   for (auto I = 0; I < KnownSize; ++I) {
68     EXPECT_EQ(V[I], 1);
69   }
70   // Zero-out the unused position '0' of the instruction histogram, which is
71   // after the first 9 calculated values. Should the the same result.
72   V[9] = 0;
73   {
74     auto ER = Evaluator.evaluate();
75     EXPECT_TRUE(ER.hasValue());
76     float Ret = *ER->getTensorValue<float>(0);
77     EXPECT_EQ(static_cast<int64_t>(Ret), 80);
78   }
79 }
80 
81 // Test incorrect input setup
82 TEST(TFUtilsTest, EvalError) {
83   // We use the ir2native model for test. We know it has one feature of
84   // dimension (1, 214)
85   const static int64_t KnownSize = 213;
86   std::vector<TensorSpec> InputSpecs{TensorSpec::createSpec<int32_t>(
87       "serving_default_input_1", {1, KnownSize})};
88   std::vector<TensorSpec> OutputSpecs{
89       TensorSpec::createSpec<float>("StatefulPartitionedCall", {1})};
90 
91   TFModelEvaluator Evaluator(getModelPath(), InputSpecs, OutputSpecs);
92   EXPECT_TRUE(Evaluator.isValid());
93 
94   int32_t *V = Evaluator.getInput<int32_t>(0);
95   // Fill it up with 1's, we know the output.
96   for (auto I = 0; I < KnownSize; ++I) {
97     V[I] = 1;
98   }
99   auto ER = Evaluator.evaluate();
100   EXPECT_FALSE(ER.hasValue());
101   EXPECT_FALSE(Evaluator.isValid());
102 }
103 
104 TEST(TFUtilsTest, UnsupportedFeature) {
105   const static int64_t KnownSize = 214;
106   std::vector<TensorSpec> InputSpecs{
107       TensorSpec::createSpec<int32_t>("serving_default_input_1",
108                                       {1, KnownSize}),
109       TensorSpec::createSpec<float>("this_feature_does_not_exist", {2, 5})};
110 
111   LLVMContext Ctx;
112   auto Evaluator = ModelUnderTrainingRunner::createAndEnsureValid(
113       Ctx, getModelPath(), "StatefulPartitionedCall", InputSpecs,
114       {LoggedFeatureSpec{
115           TensorSpec::createSpec<float>("StatefulPartitionedCall", {1}),
116           None}});
117   int32_t *V = Evaluator->getTensor<int32_t>(0);
118   // Fill it up with 1s, we know the output.
119   for (auto I = 0; I < KnownSize; ++I)
120     V[I] = 1;
121 
122   float *F = Evaluator->getTensor<float>(1);
123   for (auto I = 0; I < 2 * 5; ++I)
124     F[I] = 3.14 + I;
125   float Ret = Evaluator->evaluate<float>();
126   EXPECT_EQ(static_cast<int64_t>(Ret), 80);
127   // The input vector should be unchanged
128   for (auto I = 0; I < KnownSize; ++I)
129     EXPECT_EQ(V[I], 1);
130   for (auto I = 0; I < 2 * 5; ++I)
131     EXPECT_FLOAT_EQ(F[I], 3.14 + I);
132 }
133