commit 5f6bedb43a7f730789935cb070ecf961b237116d
author Xusong Wang <xusongw@google.com> Wed Mar 03 16:20:37 2021 -0800
committer Xusong Wang <xusongw@google.com> Mon May 10 15:21:36 2021 -0700
tree 1fffae732dff3654f7547be036865da72d04fe67
parent 88b237cfb59aa0edd4a364fb9fa57d768875cb3f

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)

neuralnetworks/aidl/utils/src/Burst.cpp

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);
 }