Introduce reusable execution to canonical interface -- HAL.
This CL modifies the canonical interface for reusable executions:
- Add new interface: IExecution with compute and computeFenced methods
- Add new method IPreparedModel::createExecution
In NNAPI runtime, the new interface IExecution is used to
memoize request-specific execution resources (e.g. converted HAL
request). The expected usage is that, IPreparedModel::createExecution
will be invoked in the first computation of a reusable NDK ANNExecution
object, and IExecution::compute* will be invoked repeatedly.
The IPreparedModel::execute* methods are preserved to avoid redundant
object creation and memoization overhead for a single-time
(non-reusable) execution.
For a vendor implementing the canonical interfaces, only the
IPreparedModel::execute* methods will be called because there is
currently no reusable execution at HAL interface. A DefaultExecution
implementation is provided to reduce the work needed on the vendor side.
Bug: 184073769
Test: NNT_static
Test: neuralnetworks_utils_hal_1_0_test
Test: neuralnetworks_utils_hal_1_1_test
Test: neuralnetworks_utils_hal_1_2_test
Test: neuralnetworks_utils_hal_1_3_test
Test: neuralnetworks_utils_hal_common_test
Test: neuralnetworks_utils_hal_aidl_test
Change-Id: I91790bb5ccf5ae648687fe603f88ffda2c9fd2b2
Merged-In: I91790bb5ccf5ae648687fe603f88ffda2c9fd2b2
(cherry picked from commit 727a7b2104b0962509fedffe720eec508b2ee6de)
diff --git a/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/Execution.h b/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/Execution.h
new file mode 100644
index 0000000..e201e25
--- /dev/null
+++ b/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/Execution.h
@@ -0,0 +1,64 @@
+/*
+ * Copyright (C) 2021 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.
+ */
+
+#ifndef ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_1_0_UTILS_EXECUTION_H
+#define ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_1_0_UTILS_EXECUTION_H
+
+#include <android/hardware/neuralnetworks/1.0/IPreparedModel.h>
+#include <nnapi/IExecution.h>
+#include <nnapi/Result.h>
+#include <nnapi/Types.h>
+#include <nnapi/hal/CommonUtils.h>
+#include <nnapi/hal/ProtectCallback.h>
+
+#include "PreparedModel.h"
+
+#include <memory>
+#include <utility>
+#include <vector>
+
+// See hardware/interfaces/neuralnetworks/utils/README.md for more information on HIDL interface
+// lifetimes across processes and for protecting asynchronous calls across HIDL.
+
+namespace android::hardware::neuralnetworks::V1_0::utils {
+
+class Execution final : public nn::IExecution, public std::enable_shared_from_this<Execution> {
+ struct PrivateConstructorTag {};
+
+ public:
+ static nn::GeneralResult<std::shared_ptr<const Execution>> create(
+ std::shared_ptr<const PreparedModel> preparedModel, Request request,
+ hal::utils::RequestRelocation relocation);
+
+ Execution(PrivateConstructorTag tag, std::shared_ptr<const PreparedModel> preparedModel,
+ Request request, hal::utils::RequestRelocation relocation);
+
+ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> compute(
+ const nn::OptionalTimePoint& deadline) const override;
+
+ nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> computeFenced(
+ const std::vector<nn::SyncFence>& waitFor, const nn::OptionalTimePoint& deadline,
+ const nn::OptionalDuration& timeoutDurationAfterFence) const override;
+
+ private:
+ const std::shared_ptr<const PreparedModel> kPreparedModel;
+ const Request kRequest;
+ const hal::utils::RequestRelocation kRelocation;
+};
+
+} // namespace android::hardware::neuralnetworks::V1_0::utils
+
+#endif // ANDROID_HARDWARE_INTERFACES_NEURALNETWORKS_1_0_UTILS_EXECUTION_H
diff --git a/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/PreparedModel.h b/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/PreparedModel.h
index 8853eea..48be595 100644
--- a/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/PreparedModel.h
+++ b/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/PreparedModel.h
@@ -57,10 +57,17 @@
const nn::OptionalDuration& loopTimeoutDuration,
const nn::OptionalDuration& timeoutDurationAfterFence) const override;
+ nn::GeneralResult<nn::SharedExecution> createReusableExecution(
+ const nn::Request& request, nn::MeasureTiming measure,
+ const nn::OptionalDuration& loopTimeoutDuration) const override;
+
nn::GeneralResult<nn::SharedBurst> configureExecutionBurst() const override;
std::any getUnderlyingResource() const override;
+ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> executeInternal(
+ const V1_0::Request& request, const hal::utils::RequestRelocation& relocation) const;
+
private:
const sp<V1_0::IPreparedModel> kPreparedModel;
const hal::utils::DeathHandler kDeathHandler;
diff --git a/neuralnetworks/1.0/utils/src/Execution.cpp b/neuralnetworks/1.0/utils/src/Execution.cpp
new file mode 100644
index 0000000..7a3216b
--- /dev/null
+++ b/neuralnetworks/1.0/utils/src/Execution.cpp
@@ -0,0 +1,72 @@
+/*
+ * Copyright (C) 2021 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 "Execution.h"
+
+#include "Callbacks.h"
+#include "Conversions.h"
+#include "Utils.h"
+
+#include <android/hardware/neuralnetworks/1.0/IPreparedModel.h>
+#include <android/hardware/neuralnetworks/1.0/types.h>
+#include <nnapi/IExecution.h>
+#include <nnapi/Result.h>
+#include <nnapi/TypeUtils.h>
+#include <nnapi/Types.h>
+#include <nnapi/hal/CommonUtils.h>
+#include <nnapi/hal/HandleError.h>
+#include <nnapi/hal/ProtectCallback.h>
+
+#include <memory>
+#include <utility>
+#include <vector>
+
+// See hardware/interfaces/neuralnetworks/utils/README.md for more information on HIDL interface
+// lifetimes across processes and for protecting asynchronous calls across HIDL.
+
+namespace android::hardware::neuralnetworks::V1_0::utils {
+
+nn::GeneralResult<std::shared_ptr<const Execution>> Execution::create(
+ std::shared_ptr<const PreparedModel> preparedModel, Request request,
+ hal::utils::RequestRelocation relocation) {
+ if (preparedModel == nullptr) {
+ return NN_ERROR() << "V1_0::utils::Execution::create must have non-null preparedModel";
+ }
+
+ return std::make_shared<const Execution>(PrivateConstructorTag{}, std::move(preparedModel),
+ std::move(request), std::move(relocation));
+}
+
+Execution::Execution(PrivateConstructorTag /*tag*/,
+ std::shared_ptr<const PreparedModel> preparedModel, Request request,
+ hal::utils::RequestRelocation relocation)
+ : kPreparedModel(std::move(preparedModel)),
+ kRequest(std::move(request)),
+ kRelocation(std::move(relocation)) {}
+
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Execution::compute(
+ const nn::OptionalTimePoint& /*deadline*/) const {
+ return kPreparedModel->executeInternal(kRequest, kRelocation);
+}
+
+nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> Execution::computeFenced(
+ const std::vector<nn::SyncFence>& /*waitFor*/, const nn::OptionalTimePoint& /*deadline*/,
+ const nn::OptionalDuration& /*timeoutDurationAfterFence*/) const {
+ return NN_ERROR(nn::ErrorStatus::GENERAL_FAILURE)
+ << "IExecution::computeFenced is not supported on 1.0 HAL service";
+}
+
+} // namespace android::hardware::neuralnetworks::V1_0::utils
diff --git a/neuralnetworks/1.0/utils/src/PreparedModel.cpp b/neuralnetworks/1.0/utils/src/PreparedModel.cpp
index 858571d..7987ab4 100644
--- a/neuralnetworks/1.0/utils/src/PreparedModel.cpp
+++ b/neuralnetworks/1.0/utils/src/PreparedModel.cpp
@@ -19,6 +19,7 @@
#include "Burst.h"
#include "Callbacks.h"
#include "Conversions.h"
+#include "Execution.h"
#include "Utils.h"
#include <android/hardware/neuralnetworks/1.0/IPreparedModel.h>
@@ -61,22 +62,34 @@
const nn::OptionalDuration& /*loopTimeoutDuration*/) const {
// Ensure that request is ready for IPC.
std::optional<nn::Request> maybeRequestInShared;
- const nn::Request& requestInShared = NN_TRY(hal::utils::makeExecutionFailure(
- hal::utils::flushDataFromPointerToShared(&request, &maybeRequestInShared)));
+ hal::utils::RequestRelocation relocation;
+ const nn::Request& requestInShared =
+ NN_TRY(hal::utils::makeExecutionFailure(hal::utils::convertRequestFromPointerToShared(
+ &request, &maybeRequestInShared, &relocation)));
const auto hidlRequest = NN_TRY(hal::utils::makeExecutionFailure(convert(requestInShared)));
+ return executeInternal(hidlRequest, relocation);
+}
+
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>>
+PreparedModel::executeInternal(const V1_0::Request& request,
+ const hal::utils::RequestRelocation& relocation) const {
+ if (relocation.input) {
+ relocation.input->flush();
+ }
+
const auto cb = sp<ExecutionCallback>::make();
const auto scoped = kDeathHandler.protectCallback(cb.get());
- const auto ret = kPreparedModel->execute(hidlRequest, cb);
+ const auto ret = kPreparedModel->execute(request, cb);
const auto status = HANDLE_TRANSPORT_FAILURE(ret);
HANDLE_HAL_STATUS(status) << "execution failed with " << toString(status);
auto result = NN_TRY(cb->get());
- NN_TRY(hal::utils::makeExecutionFailure(
- hal::utils::unflushDataFromSharedToPointer(request, maybeRequestInShared)));
-
+ if (relocation.output) {
+ relocation.output->flush();
+ }
return result;
}
@@ -91,6 +104,19 @@
<< "IPreparedModel::executeFenced is not supported on 1.0 HAL service";
}
+nn::GeneralResult<nn::SharedExecution> PreparedModel::createReusableExecution(
+ const nn::Request& request, nn::MeasureTiming /*measure*/,
+ const nn::OptionalDuration& /*loopTimeoutDuration*/) const {
+ // Ensure that request is ready for IPC.
+ std::optional<nn::Request> maybeRequestInShared;
+ hal::utils::RequestRelocation relocation;
+ const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
+ &request, &maybeRequestInShared, &relocation));
+
+ auto hidlRequest = NN_TRY(convert(requestInShared));
+ return Execution::create(shared_from_this(), std::move(hidlRequest), std::move(relocation));
+}
+
nn::GeneralResult<nn::SharedBurst> PreparedModel::configureExecutionBurst() const {
return Burst::create(shared_from_this());
}
diff --git a/neuralnetworks/1.0/utils/test/PreparedModelTest.cpp b/neuralnetworks/1.0/utils/test/PreparedModelTest.cpp
index f19ed77..7820c06 100644
--- a/neuralnetworks/1.0/utils/test/PreparedModelTest.cpp
+++ b/neuralnetworks/1.0/utils/test/PreparedModelTest.cpp
@@ -19,6 +19,7 @@
#include <android/hardware/neuralnetworks/1.0/IExecutionCallback.h>
#include <gmock/gmock.h>
#include <gtest/gtest.h>
+#include <nnapi/IExecution.h>
#include <nnapi/IPreparedModel.h>
#include <nnapi/TypeUtils.h>
#include <nnapi/Types.h>
@@ -224,6 +225,150 @@
EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
}
+TEST(PreparedModelTest, reusableExecute) {
+ // setup call
+ const uint32_t kNumberOfComputations = 2;
+ const auto mockPreparedModel = createMockPreparedModel();
+ const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+ EXPECT_CALL(*mockPreparedModel, execute(_, _))
+ .Times(kNumberOfComputations)
+ .WillRepeatedly(Invoke(makeExecute(V1_0::ErrorStatus::NONE, V1_0::ErrorStatus::NONE)));
+
+ // create execution
+ const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+ ASSERT_TRUE(createResult.has_value())
+ << "Failed with " << createResult.error().code << ": " << createResult.error().message;
+ ASSERT_NE(createResult.value(), nullptr);
+
+ // invoke compute repeatedly
+ for (uint32_t i = 0; i < kNumberOfComputations; i++) {
+ const auto computeResult = createResult.value()->compute({});
+ EXPECT_TRUE(computeResult.has_value()) << "Failed with " << computeResult.error().code
+ << ": " << computeResult.error().message;
+ }
+}
+
+TEST(PreparedModelTest, reusableExecuteLaunchError) {
+ // setup test
+ const auto mockPreparedModel = createMockPreparedModel();
+ const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+ EXPECT_CALL(*mockPreparedModel, execute(_, _))
+ .Times(1)
+ .WillOnce(Invoke(makeExecute(V1_0::ErrorStatus::GENERAL_FAILURE,
+ V1_0::ErrorStatus::GENERAL_FAILURE)));
+
+ // create execution
+ const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+ ASSERT_TRUE(createResult.has_value())
+ << "Failed with " << createResult.error().code << ": " << createResult.error().message;
+ ASSERT_NE(createResult.value(), nullptr);
+
+ // invoke compute
+ const auto computeResult = createResult.value()->compute({});
+ ASSERT_FALSE(computeResult.has_value());
+ EXPECT_EQ(computeResult.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST(PreparedModelTest, reusableExecuteReturnError) {
+ // setup test
+ const auto mockPreparedModel = createMockPreparedModel();
+ const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+ EXPECT_CALL(*mockPreparedModel, execute(_, _))
+ .Times(1)
+ .WillOnce(Invoke(
+ makeExecute(V1_0::ErrorStatus::NONE, V1_0::ErrorStatus::GENERAL_FAILURE)));
+
+ // create execution
+ const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+ ASSERT_TRUE(createResult.has_value())
+ << "Failed with " << createResult.error().code << ": " << createResult.error().message;
+ ASSERT_NE(createResult.value(), nullptr);
+
+ // invoke compute
+ const auto computeResult = createResult.value()->compute({});
+ ASSERT_FALSE(computeResult.has_value());
+ EXPECT_EQ(computeResult.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST(PreparedModelTest, reusableExecuteTransportFailure) {
+ // setup test
+ const auto mockPreparedModel = createMockPreparedModel();
+ const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+ EXPECT_CALL(*mockPreparedModel, execute(_, _))
+ .Times(1)
+ .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
+
+ // create execution
+ const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+ ASSERT_TRUE(createResult.has_value())
+ << "Failed with " << createResult.error().code << ": " << createResult.error().message;
+ ASSERT_NE(createResult.value(), nullptr);
+
+ // invoke compute
+ const auto computeResult = createResult.value()->compute({});
+ ASSERT_FALSE(computeResult.has_value());
+ EXPECT_EQ(computeResult.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST(PreparedModelTest, reusableExecuteDeadObject) {
+ // setup test
+ const auto mockPreparedModel = createMockPreparedModel();
+ const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+ EXPECT_CALL(*mockPreparedModel, execute(_, _))
+ .Times(1)
+ .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
+
+ // create execution
+ const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+ ASSERT_TRUE(createResult.has_value())
+ << "Failed with " << createResult.error().code << ": " << createResult.error().message;
+ ASSERT_NE(createResult.value(), nullptr);
+
+ // invoke compute
+ const auto computeResult = createResult.value()->compute({});
+ ASSERT_FALSE(computeResult.has_value());
+ EXPECT_EQ(computeResult.error().code, nn::ErrorStatus::DEAD_OBJECT);
+}
+
+TEST(PreparedModelTest, reusableExecuteCrash) {
+ // setup test
+ const auto mockPreparedModel = createMockPreparedModel();
+ const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+ const auto ret = [&mockPreparedModel]() -> hardware::Return<V1_0::ErrorStatus> {
+ mockPreparedModel->simulateCrash();
+ return V1_0::ErrorStatus::NONE;
+ };
+ EXPECT_CALL(*mockPreparedModel, execute(_, _)).Times(1).WillOnce(InvokeWithoutArgs(ret));
+
+ // create execution
+ const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+ ASSERT_TRUE(createResult.has_value())
+ << "Failed with " << createResult.error().code << ": " << createResult.error().message;
+ ASSERT_NE(createResult.value(), nullptr);
+
+ // invoke compute
+ const auto computeResult = createResult.value()->compute({});
+ ASSERT_FALSE(computeResult.has_value());
+ EXPECT_EQ(computeResult.error().code, nn::ErrorStatus::DEAD_OBJECT);
+}
+
+TEST(PreparedModelTest, reusableExecuteFencedNotSupported) {
+ // setup test
+ const auto mockPreparedModel = createMockPreparedModel();
+ const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
+
+ // create execution
+ const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+ ASSERT_TRUE(createResult.has_value())
+ << "Failed with " << createResult.error().code << ": " << createResult.error().message;
+ ASSERT_NE(createResult.value(), nullptr);
+
+ // invoke compute
+ const auto computeResult = createResult.value()->computeFenced({}, {}, {});
+ ASSERT_FALSE(computeResult.has_value());
+ EXPECT_EQ(computeResult.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
TEST(PreparedModelTest, configureExecutionBurst) {
// setup test
const auto mockPreparedModel = MockPreparedModel::create();