Fix control flow partial CPU fallback bug

Before this change, ExecutionPlan::Controller::mLastStepIndex is only
updated at the beginning of the ExecutionPlan::next() call. When
interpreted control flow is involved, this call might result in several
logical steps being processed until an ExecutionStep is reached.

When an interpreted CF loop terminates successfully but the next step is
an ExecutionStep that fails, ExecutionPlan::fallback() restarts from
mLastStepIndex, which happens to be the index of the WhileStep. This
results in the loop being executed again. This happens forever.

This change fixes the bug by updating mLastStepIndex when
nextCompound(const ExecutionStep* ...) is called.

This change renames mLastStepIndex to mFallbackNextStepIndex to reflect
its usage.

Fix: 152623151
Test: NNT_static
Change-Id: Iff3b68f541adb8facf078e968abac2a6ae8e74c1

2 files changed, 175 insertions, 0 deletions

diff --git a/nn/runtime/test/Android.bp b/nn/runtime/test/Android.bp
index fe14001b8..f25080605 100644
--- a/nn/runtime/test/Android.bp
+++ b/nn/runtime/test/Android.bp
@@ -122,6 +122,7 @@ cc_defaults {
         // "TestOpenmpSettings.cpp",
         "TestCompilationCaching.cpp",
         "TestCompliance.cpp",
+        "TestControlFlowExecution.cpp",
         "TestExecution.cpp",
         "TestExtensions.cpp",
         "TestIntrospectionControl.cpp",
diff --git a/nn/runtime/test/TestControlFlowExecution.cpp b/nn/runtime/test/TestControlFlowExecution.cpp
new file mode 100644
index 000000000..34e4b2923
--- /dev/null
+++ b/nn/runtime/test/TestControlFlowExecution.cpp
@@ -0,0 +1,174 @@
+/*
+ * Copyright (C) 2020 The Android Open Source Project
+ *
+ * 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 <gtest/gtest.h>
+
+#include <memory>
+#include <vector>
+
+#include "CompilationBuilder.h"
+#include "ExecutionPlan.h"
+#include "Manager.h"
+#include "SampleDriverPartial.h"
+#include "TestNeuralNetworksWrapper.h"
+
+namespace android::nn {
+namespace {
+
+using namespace hal;
+using sample_driver::SampleDriverPartial;
+using Result = test_wrapper::Result;
+using WrapperOperandType = test_wrapper::OperandType;
+using WrapperCompilation = test_wrapper::Compilation;
+using WrapperExecution = test_wrapper::Execution;
+using WrapperType = test_wrapper::Type;
+using WrapperModel = test_wrapper::Model;
+
+class EmptyOperationResolver : public IOperationResolver {
+   public:
+    const OperationRegistration* findOperation(OperationType) const override { return nullptr; }
+};
+
+const char* kTestDriverName = "nnapi-test-sqrt-failing";
+
+// A driver that only supports SQRT and fails during execution.
+class FailingTestDriver : public SampleDriverPartial {
+   public:
+    // EmptyOperationResolver causes execution to fail.
+    FailingTestDriver() : SampleDriverPartial(kTestDriverName, &mEmptyOperationResolver) {}
+
+    Return<void> getCapabilities_1_3(getCapabilities_1_3_cb cb) override {
+        cb(V1_3::ErrorStatus::NONE,
+           {.operandPerformance = {{.type = OperandType::TENSOR_FLOAT32,
+                                    .info = {.execTime = 0.1,  // Faster than CPU.
+                                             .powerUsage = 0.1}}}});
+        return Void();
+    }
+
+   private:
+    std::vector<bool> getSupportedOperationsImpl(const Model& model) const override {
+        std::vector<bool> supported(model.main.operations.size());
+        std::transform(
+                model.main.operations.begin(), model.main.operations.end(), supported.begin(),
+                [](const Operation& operation) { return operation.type == OperationType::SQRT; });
+        return supported;
+    }
+
+    const EmptyOperationResolver mEmptyOperationResolver;
+};
+
+class FailingDriverTest : public ::testing::Test {
+    virtual void SetUp() {
+        if (DeviceManager::get()->getUseCpuOnly()) {
+            GTEST_SKIP();
+        }
+        mTestDevice =
+                DeviceManager::forTest_makeDriverDevice(kTestDriverName, new FailingTestDriver());
+        DeviceManager::get()->forTest_setDevices({
+                mTestDevice,
+                DeviceManager::getCpuDevice(),
+        });
+    }
+
+    virtual void TearDown() { DeviceManager::get()->forTest_reInitializeDeviceList(); }
+
+   protected:
+    std::shared_ptr<Device> mTestDevice;
+};
+
+// Regression test for b/152623150.
+TEST_F(FailingDriverTest, FailAfterInterpretedWhile) {
+    // Model:
+    //     f = input0
+    //     b = input1
+    //     while CAST(b):  # Identity cast.
+    //         f = CAST(f)
+    //     # FailingTestDriver fails here. When partial CPU fallback happens,
+    //     # it should not loop forever.
+    //     output0 = SQRT(f)
+
+    WrapperOperandType floatType(WrapperType::TENSOR_FLOAT32, {2});
+    WrapperOperandType boolType(WrapperType::TENSOR_BOOL8, {1});
+
+    WrapperModel conditionModel;
+    {
+        uint32_t f = conditionModel.addOperand(&floatType);
+        uint32_t b = conditionModel.addOperand(&boolType);
+        uint32_t out = conditionModel.addOperand(&boolType);
+        conditionModel.addOperation(ANEURALNETWORKS_CAST, {b}, {out});
+        conditionModel.identifyInputsAndOutputs({f, b}, {out});
+        ASSERT_EQ(conditionModel.finish(), Result::NO_ERROR);
+        ASSERT_TRUE(conditionModel.isValid());
+    }
+
+    WrapperModel bodyModel;
+    {
+        uint32_t f = bodyModel.addOperand(&floatType);
+        uint32_t b = bodyModel.addOperand(&boolType);
+        uint32_t out = bodyModel.addOperand(&floatType);
+        bodyModel.addOperation(ANEURALNETWORKS_CAST, {f}, {out});
+        bodyModel.identifyInputsAndOutputs({f, b}, {out});
+        ASSERT_EQ(bodyModel.finish(), Result::NO_ERROR);
+        ASSERT_TRUE(bodyModel.isValid());
+    }
+
+    WrapperModel model;
+    {
+        uint32_t fInput = model.addOperand(&floatType);
+        uint32_t bInput = model.addOperand(&boolType);
+        uint32_t fTmp = model.addOperand(&floatType);
+        uint32_t fSqrt = model.addOperand(&floatType);
+        uint32_t cond = model.addModelOperand(&conditionModel);
+        uint32_t body = model.addModelOperand(&bodyModel);
+        model.addOperation(ANEURALNETWORKS_WHILE, {cond, body, fInput, bInput}, {fTmp});
+        model.addOperation(ANEURALNETWORKS_SQRT, {fTmp}, {fSqrt});
+        model.identifyInputsAndOutputs({fInput, bInput}, {fSqrt});
+        ASSERT_TRUE(model.isValid());
+        ASSERT_EQ(model.finish(), Result::NO_ERROR);
+    }
+
+    WrapperCompilation compilation(&model);
+    ASSERT_EQ(compilation.finish(), Result::NO_ERROR);
+
+    const CompilationBuilder* compilationBuilder =
+            reinterpret_cast<CompilationBuilder*>(compilation.getHandle());
+    const ExecutionPlan& plan = compilationBuilder->forTest_getExecutionPlan();
+    const std::vector<std::shared_ptr<LogicalStep>>& steps = plan.forTest_compoundGetSteps();
+    ASSERT_EQ(steps.size(), 6u);
+    ASSERT_TRUE(steps[0]->isWhile());
+    ASSERT_TRUE(steps[1]->isExecution());
+    ASSERT_EQ(steps[1]->executionStep()->getDevice(), DeviceManager::getCpuDevice());
+    ASSERT_TRUE(steps[2]->isGoto());
+    ASSERT_TRUE(steps[3]->isExecution());
+    ASSERT_EQ(steps[3]->executionStep()->getDevice(), DeviceManager::getCpuDevice());
+    ASSERT_TRUE(steps[4]->isGoto());
+    ASSERT_TRUE(steps[5]->isExecution());
+    ASSERT_EQ(steps[5]->executionStep()->getDevice(), mTestDevice);
+
+    WrapperExecution execution(&compilation);
+    const float fInput[] = {12 * 12, 5 * 5};
+    const bool8 bInput = false;
+    float fSqrt[] = {0, 0};
+    ASSERT_EQ(execution.setInput(0, &fInput), Result::NO_ERROR);
+    ASSERT_EQ(execution.setInput(1, &bInput), Result::NO_ERROR);
+    ASSERT_EQ(execution.setOutput(0, &fSqrt), Result::NO_ERROR);
+    ASSERT_EQ(execution.compute(), Result::NO_ERROR);
+    ASSERT_EQ(fSqrt[0], 12);
+    ASSERT_EQ(fSqrt[1], 5);
+}
+
+}  // namespace
+}  // namespace android::nn