Merge "Use proper alignment and padding for pointer arguments -- HAL." am: 0a3ca7b629

Original change: https://android-review.googlesource.com/c/platform/hardware/interfaces/+/1702587

Change-Id: I7a91601496932132d9c7d7fe42e265f5efd46a29

neuralnetworks/1.0/utils/src/PreparedModel.cpp

@@ -65,7 +65,8 @@ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Prepare
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared =
             NN_TRY(hal::utils::makeExecutionFailure(hal::utils::convertRequestFromPointerToShared(
-                    &request, &maybeRequestInShared, &relocation)));
+                    &request, nn::kDefaultRequestMemoryAlignment, nn::kMinMemoryPadding,
+                    &maybeRequestInShared, &relocation)));
 
     const auto hidlRequest = NN_TRY(hal::utils::makeExecutionFailure(convert(requestInShared)));
 
@@ -111,7 +112,8 @@ nn::GeneralResult<nn::SharedExecution> PreparedModel::createReusableExecution(
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
-            &request, &maybeRequestInShared, &relocation));
+            &request, nn::kDefaultRequestMemoryAlignment, nn::kMinMemoryPadding,
+            &maybeRequestInShared, &relocation));
 
     auto hidlRequest = NN_TRY(convert(requestInShared));
     return Execution::create(shared_from_this(), std::move(hidlRequest), std::move(relocation));

neuralnetworks/1.2/utils/src/ExecutionBurstController.cpp

@@ -328,7 +328,8 @@ ExecutionBurstController::execute(const nn::Request& request, nn::MeasureTiming
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared =
             NN_TRY(hal::utils::makeExecutionFailure(hal::utils::convertRequestFromPointerToShared(
-                    &request, &maybeRequestInShared, &relocation)));
+                    &request, nn::kDefaultRequestMemoryAlignment, nn::kMinMemoryPadding,
+                    &maybeRequestInShared, &relocation)));
 
     // clear pools field of request, as they will be provided via slots
     const auto requestWithoutPools = nn::Request{
@@ -373,7 +374,8 @@ nn::GeneralResult<nn::SharedExecution> ExecutionBurstController::createReusableE
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
-            &request, &maybeRequestInShared, &relocation));
+            &request, nn::kDefaultRequestMemoryAlignment, nn::kMinMemoryPadding,
+            &maybeRequestInShared, &relocation));
 
     // clear pools field of request, as they will be provided via slots
     const auto requestWithoutPools = nn::Request{

neuralnetworks/1.2/utils/src/PreparedModel.cpp

@@ -97,7 +97,8 @@ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Prepare
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared =
             NN_TRY(hal::utils::makeExecutionFailure(hal::utils::convertRequestFromPointerToShared(
-                    &request, &maybeRequestInShared, &relocation)));
+                    &request, nn::kDefaultRequestMemoryAlignment, nn::kMinMemoryPadding,
+                    &maybeRequestInShared, &relocation)));
 
     const auto hidlRequest = NN_TRY(hal::utils::makeExecutionFailure(convert(requestInShared)));
     const auto hidlMeasure = NN_TRY(hal::utils::makeExecutionFailure(convert(measure)));
@@ -140,7 +141,8 @@ nn::GeneralResult<nn::SharedExecution> PreparedModel::createReusableExecution(
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
-            &request, &maybeRequestInShared, &relocation));
+            &request, nn::kDefaultRequestMemoryAlignment, nn::kMinMemoryPadding,
+            &maybeRequestInShared, &relocation));
 
     auto hidlRequest = NN_TRY(convert(requestInShared));
     auto hidlMeasure = NN_TRY(convert(measure));

neuralnetworks/1.3/utils/src/PreparedModel.cpp

@@ -143,7 +143,8 @@ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Prepare
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared =
             NN_TRY(hal::utils::makeExecutionFailure(hal::utils::convertRequestFromPointerToShared(
-                    &request, &maybeRequestInShared, &relocation)));
+                    &request, nn::kDefaultRequestMemoryAlignment, nn::kMinMemoryPadding,
+                    &maybeRequestInShared, &relocation)));
 
     const auto hidlRequest = NN_TRY(hal::utils::makeExecutionFailure(convert(requestInShared)));
     const auto hidlMeasure = NN_TRY(hal::utils::makeExecutionFailure(convert(measure)));
@@ -184,7 +185,8 @@ PreparedModel::executeFenced(const nn::Request& request, const std::vector<nn::S
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
-            &request, &maybeRequestInShared, &relocation));
+            &request, nn::kDefaultRequestMemoryAlignment, nn::kMinMemoryPadding,
+            &maybeRequestInShared, &relocation));
 
     const auto hidlRequest = NN_TRY(convert(requestInShared));
     const auto hidlWaitFor = NN_TRY(hal::utils::convertSyncFences(waitFor));
@@ -236,7 +238,8 @@ nn::GeneralResult<nn::SharedExecution> PreparedModel::createReusableExecution(
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
-            &request, &maybeRequestInShared, &relocation));
+            &request, nn::kDefaultRequestMemoryAlignment, nn::kMinMemoryPadding,
+            &maybeRequestInShared, &relocation));
 
     auto hidlRequest = NN_TRY(convert(requestInShared));
     auto hidlMeasure = NN_TRY(convert(measure));

neuralnetworks/aidl/utils/src/Burst.cpp

@@ -178,7 +178,8 @@ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Burst::
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared =
             NN_TRY(hal::utils::makeExecutionFailure(hal::utils::convertRequestFromPointerToShared(
-                    &request, &maybeRequestInShared, &relocation)));
+                    &request, nn::kDefaultRequestMemoryAlignment, nn::kDefaultRequestMemoryPadding,
+                    &maybeRequestInShared, &relocation)));
 
     const auto aidlRequest = NN_TRY(hal::utils::makeExecutionFailure(convert(requestInShared)));
     const auto aidlMeasure = NN_TRY(hal::utils::makeExecutionFailure(convert(measure)));
@@ -248,7 +249,8 @@ nn::GeneralResult<nn::SharedExecution> Burst::createReusableExecution(
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
-            &request, &maybeRequestInShared, &relocation));
+            &request, nn::kDefaultRequestMemoryAlignment, nn::kDefaultRequestMemoryPadding,
+            &maybeRequestInShared, &relocation));
 
     auto aidlRequest = NN_TRY(convert(requestInShared));
     const auto aidlMeasure = NN_TRY(convert(measure));

neuralnetworks/aidl/utils/src/PreparedModel.cpp

@@ -80,7 +80,8 @@ nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Prepare
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared =
             NN_TRY(hal::utils::makeExecutionFailure(hal::utils::convertRequestFromPointerToShared(
-                    &request, &maybeRequestInShared, &relocation)));
+                    &request, nn::kDefaultRequestMemoryAlignment, nn::kDefaultRequestMemoryPadding,
+                    &maybeRequestInShared, &relocation)));
 
     const auto aidlRequest = NN_TRY(hal::utils::makeExecutionFailure(convert(requestInShared)));
     const auto aidlMeasure = NN_TRY(hal::utils::makeExecutionFailure(convert(measure)));
@@ -127,7 +128,8 @@ PreparedModel::executeFenced(const nn::Request& request, const std::vector<nn::S
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
-            &request, &maybeRequestInShared, &relocation));
+            &request, nn::kDefaultRequestMemoryAlignment, nn::kDefaultRequestMemoryPadding,
+            &maybeRequestInShared, &relocation));
 
     const auto aidlRequest = NN_TRY(convert(requestInShared));
     const auto aidlWaitFor = NN_TRY(convert(waitFor));
@@ -197,7 +199,8 @@ nn::GeneralResult<nn::SharedExecution> PreparedModel::createReusableExecution(
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
     const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
-            &request, &maybeRequestInShared, &relocation));
+            &request, nn::kDefaultRequestMemoryAlignment, nn::kDefaultRequestMemoryPadding,
+            &maybeRequestInShared, &relocation));
 
     auto aidlRequest = NN_TRY(convert(requestInShared));
     auto aidlMeasure = NN_TRY(convert(measure));

neuralnetworks/utils/common/include/nnapi/hal/CommonUtils.h

@@ -122,8 +122,8 @@ struct RequestRelocation {
 // Unlike `flushDataFromPointerToShared`, this method will not copy the input pointer data to the
 // shared memory pool. Use `relocationOut` to flush the input or output data after the call.
 nn::GeneralResult<std::reference_wrapper<const nn::Request>> convertRequestFromPointerToShared(
-        const nn::Request* request, std::optional<nn::Request>* maybeRequestInSharedOut,
+        const nn::Request* request, uint32_t alignment, uint32_t padding,
-        RequestRelocation* relocationOut);
+        std::optional<nn::Request>* maybeRequestInSharedOut, RequestRelocation* relocationOut);
 
 nn::GeneralResult<std::vector<uint32_t>> countNumberOfConsumers(
         size_t numberOfOperands, const std::vector<nn::Operation>& operations);

neuralnetworks/utils/common/src/CommonUtils.cpp

@@ -220,8 +220,8 @@ void OutputRelocationTracker::flush() const {
 }
 
 nn::GeneralResult<std::reference_wrapper<const nn::Request>> convertRequestFromPointerToShared(
-        const nn::Request* request, std::optional<nn::Request>* maybeRequestInSharedOut,
+        const nn::Request* request, uint32_t alignment, uint32_t padding,
-        RequestRelocation* relocationOut) {
+        std::optional<nn::Request>* maybeRequestInSharedOut, RequestRelocation* relocationOut) {
     CHECK(request != nullptr);
     CHECK(maybeRequestInSharedOut != nullptr);
     CHECK(relocationOut != nullptr);
@@ -249,7 +249,7 @@ nn::GeneralResult<std::reference_wrapper<const nn::Request>> convertRequestFromP
         const void* data = std::visit([](auto ptr) { return static_cast<const void*>(ptr); },
                                       location.pointer);
         CHECK(data != nullptr);
-        input.location = inputBuilder.append(location.length);
+        input.location = inputBuilder.append(location.length, alignment, padding);
         inputRelocationInfos.push_back({data, input.location.length, input.location.offset});
     }
 
@@ -273,7 +273,7 @@ nn::GeneralResult<std::reference_wrapper<const nn::Request>> convertRequestFromP
         output.lifetime = nn::Request::Argument::LifeTime::POOL;
         void* data = std::get<void*>(location.pointer);
         CHECK(data != nullptr);
-        output.location = outputBuilder.append(location.length);
+        output.location = outputBuilder.append(location.length, alignment, padding);
         outputRelocationInfos.push_back({data, output.location.length, output.location.offset});
     }