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/aidl/utils/src/Burst.cpp b/neuralnetworks/aidl/utils/src/Burst.cpp
index 0b475bc..b20f6ae 100644
--- a/neuralnetworks/aidl/utils/src/Burst.cpp
+++ b/neuralnetworks/aidl/utils/src/Burst.cpp
@@ -148,8 +148,10 @@
 
     // 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 aidlRequest = NN_TRY(hal::utils::makeExecutionFailure(convert(requestInShared)));
     const auto aidlMeasure = NN_TRY(hal::utils::makeExecutionFailure(convert(measure)));
@@ -174,6 +176,10 @@
     }
     CHECK_EQ(request.pools.size(), memoryIdentifierTokens.size());
 
+    if (relocation.input) {
+        relocation.input->flush();
+    }
+
     ExecutionResult executionResult;
     const auto ret =
             kBurst->executeSynchronously(aidlRequest, memoryIdentifierTokens, aidlMeasure,
@@ -188,9 +194,9 @@
     auto [outputShapes, timing] = NN_TRY(hal::utils::makeExecutionFailure(
             convertExecutionResults(executionResult.outputShapes, executionResult.timing)));
 
-    NN_TRY(hal::utils::makeExecutionFailure(
-            hal::utils::unflushDataFromSharedToPointer(request, maybeRequestInShared)));
-
+    if (relocation.output) {
+        relocation.output->flush();
+    }
     return std::make_pair(std::move(outputShapes), timing);
 }
 
diff --git a/neuralnetworks/aidl/utils/src/Execution.cpp b/neuralnetworks/aidl/utils/src/Execution.cpp
new file mode 100644
index 0000000..2aee8a6
--- /dev/null
+++ b/neuralnetworks/aidl/utils/src/Execution.cpp
@@ -0,0 +1,79 @@
+/*
+ * 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 "Conversions.h"
+#include "PreparedModel.h"
+#include "Utils.h"
+
+#include <aidl/android/hardware/neuralnetworks/Request.h>
+#include <nnapi/IExecution.h>
+#include <nnapi/Result.h>
+#include <nnapi/Types.h>
+#include <nnapi/hal/CommonUtils.h>
+#include <nnapi/hal/HandleError.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 aidl::android::hardware::neuralnetworks::utils {
+
+nn::GeneralResult<std::shared_ptr<const Execution>> Execution::create(
+        std::shared_ptr<const PreparedModel> preparedModel, Request request,
+        hal::utils::RequestRelocation relocation, bool measure, int64_t loopTimeoutDuration) {
+    if (preparedModel == nullptr) {
+        return NN_ERROR() << "aidl::utils::Execution::create must have non-null preparedModel";
+    }
+
+    return std::make_shared<const Execution>(PrivateConstructorTag{}, std::move(preparedModel),
+                                             std::move(request), std::move(relocation), measure,
+                                             loopTimeoutDuration);
+}
+
+Execution::Execution(PrivateConstructorTag /*tag*/,
+                     std::shared_ptr<const PreparedModel> preparedModel, Request request,
+                     hal::utils::RequestRelocation relocation, bool measure,
+                     int64_t loopTimeoutDuration)
+    : kPreparedModel(std::move(preparedModel)),
+      kRequest(std::move(request)),
+      kRelocation(std::move(relocation)),
+      kMeasure(measure),
+      kLoopTimeoutDuration(loopTimeoutDuration) {}
+
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Execution::compute(
+        const nn::OptionalTimePoint& deadline) const {
+    const auto aidlDeadline = NN_TRY(hal::utils::makeExecutionFailure(convert(deadline)));
+    return kPreparedModel->executeInternal(kRequest, kMeasure, aidlDeadline, kLoopTimeoutDuration,
+                                           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 {
+    const auto aidlWaitFor = NN_TRY(convert(waitFor));
+    const auto aidlDeadline = NN_TRY(convert(deadline));
+    const auto aidlTimeoutDurationAfterFence = NN_TRY(convert(timeoutDurationAfterFence));
+    return kPreparedModel->executeFencedInternal(kRequest, aidlWaitFor, kMeasure, aidlDeadline,
+                                                 kLoopTimeoutDuration,
+                                                 aidlTimeoutDurationAfterFence, kRelocation);
+}
+
+}  // namespace aidl::android::hardware::neuralnetworks::utils
diff --git a/neuralnetworks/aidl/utils/src/PreparedModel.cpp b/neuralnetworks/aidl/utils/src/PreparedModel.cpp
index 003965b..1915607 100644
--- a/neuralnetworks/aidl/utils/src/PreparedModel.cpp
+++ b/neuralnetworks/aidl/utils/src/PreparedModel.cpp
@@ -19,8 +19,11 @@
 #include "Burst.h"
 #include "Callbacks.h"
 #include "Conversions.h"
+#include "Execution.h"
+#include "ProtectCallback.h"
 #include "Utils.h"
 
+#include <aidl/android/hardware/neuralnetworks/Request.h>
 #include <android/binder_auto_utils.h>
 #include <nnapi/IPreparedModel.h>
 #include <nnapi/Result.h>
@@ -74,18 +77,31 @@
         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 aidlRequest = NN_TRY(hal::utils::makeExecutionFailure(convert(requestInShared)));
     const auto aidlMeasure = NN_TRY(hal::utils::makeExecutionFailure(convert(measure)));
     const auto aidlDeadline = NN_TRY(hal::utils::makeExecutionFailure(convert(deadline)));
     const auto aidlLoopTimeoutDuration =
             NN_TRY(hal::utils::makeExecutionFailure(convert(loopTimeoutDuration)));
+    return executeInternal(aidlRequest, aidlMeasure, aidlDeadline, aidlLoopTimeoutDuration,
+                           relocation);
+}
+
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>>
+PreparedModel::executeInternal(const Request& request, bool measure, int64_t deadline,
+                               int64_t loopTimeoutDuration,
+                               const hal::utils::RequestRelocation& relocation) const {
+    if (relocation.input) {
+        relocation.input->flush();
+    }
 
     ExecutionResult executionResult;
-    const auto ret = kPreparedModel->executeSynchronously(
-            aidlRequest, aidlMeasure, aidlDeadline, aidlLoopTimeoutDuration, &executionResult);
+    const auto ret = kPreparedModel->executeSynchronously(request, measure, deadline,
+                                                          loopTimeoutDuration, &executionResult);
     HANDLE_ASTATUS(ret) << "executeSynchronously failed";
     if (!executionResult.outputSufficientSize) {
         auto canonicalOutputShapes =
@@ -96,9 +112,9 @@
     auto [outputShapes, timing] = NN_TRY(hal::utils::makeExecutionFailure(
             convertExecutionResults(executionResult.outputShapes, executionResult.timing)));
 
-    NN_TRY(hal::utils::makeExecutionFailure(
-            hal::utils::unflushDataFromSharedToPointer(request, maybeRequestInShared)));
-
+    if (relocation.output) {
+        relocation.output->flush();
+    }
     return std::make_pair(std::move(outputShapes), timing);
 }
 
@@ -109,8 +125,9 @@
                              const nn::OptionalDuration& timeoutDurationAfterFence) const {
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
-    const nn::Request& requestInShared =
-            NN_TRY(hal::utils::flushDataFromPointerToShared(&request, &maybeRequestInShared));
+    hal::utils::RequestRelocation relocation;
+    const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
+            &request, &maybeRequestInShared, &relocation));
 
     const auto aidlRequest = NN_TRY(convert(requestInShared));
     const auto aidlWaitFor = NN_TRY(convert(waitFor));
@@ -118,11 +135,25 @@
     const auto aidlDeadline = NN_TRY(convert(deadline));
     const auto aidlLoopTimeoutDuration = NN_TRY(convert(loopTimeoutDuration));
     const auto aidlTimeoutDurationAfterFence = NN_TRY(convert(timeoutDurationAfterFence));
+    return executeFencedInternal(aidlRequest, aidlWaitFor, aidlMeasure, aidlDeadline,
+                                 aidlLoopTimeoutDuration, aidlTimeoutDurationAfterFence,
+                                 relocation);
+}
+
+nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
+PreparedModel::executeFencedInternal(const Request& request,
+                                     const std::vector<ndk::ScopedFileDescriptor>& waitFor,
+                                     bool measure, int64_t deadline, int64_t loopTimeoutDuration,
+                                     int64_t timeoutDurationAfterFence,
+                                     const hal::utils::RequestRelocation& relocation) const {
+    if (relocation.input) {
+        relocation.input->flush();
+    }
 
     FencedExecutionResult result;
-    const auto ret = kPreparedModel->executeFenced(aidlRequest, aidlWaitFor, aidlMeasure,
-                                                   aidlDeadline, aidlLoopTimeoutDuration,
-                                                   aidlTimeoutDurationAfterFence, &result);
+    const auto ret =
+            kPreparedModel->executeFenced(request, waitFor, measure, deadline, loopTimeoutDuration,
+                                          timeoutDurationAfterFence, &result);
     HANDLE_ASTATUS(ret) << "executeFenced failed";
 
     auto resultSyncFence = nn::SyncFence::createAsSignaled();
@@ -137,12 +168,12 @@
 
     // If executeFenced required the request memory to be moved into shared memory, block here until
     // the fenced execution has completed and flush the memory back.
-    if (maybeRequestInShared.has_value()) {
+    if (relocation.output) {
         const auto state = resultSyncFence.syncWait({});
         if (state != nn::SyncFence::FenceState::SIGNALED) {
             return NN_ERROR() << "syncWait failed with " << state;
         }
-        NN_TRY(hal::utils::unflushDataFromSharedToPointer(request, maybeRequestInShared));
+        relocation.output->flush();
     }
 
     // Create callback which can be used to retrieve the execution error status and timings.
@@ -159,6 +190,22 @@
     return std::make_pair(std::move(resultSyncFence), std::move(resultCallback));
 }
 
+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 aidlRequest = NN_TRY(convert(requestInShared));
+    auto aidlMeasure = NN_TRY(convert(measure));
+    auto aidlLoopTimeoutDuration = NN_TRY(convert(loopTimeoutDuration));
+    return Execution::create(shared_from_this(), std::move(aidlRequest), std::move(relocation),
+                             aidlMeasure, aidlLoopTimeoutDuration);
+}
+
 nn::GeneralResult<nn::SharedBurst> PreparedModel::configureExecutionBurst() const {
     std::shared_ptr<IBurst> burst;
     const auto ret = kPreparedModel->configureExecutionBurst(&burst);