experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Disallow use of NNAPI CPU implementation by default on Android 10 and later. TFLite's own implementation typically performs better.

Browse Source 
PiperOrigin-RevId: 339195098
Change-Id: If0d09d60c9e6fb7288abcf0b4f3f8b732c58e724
This commit is contained in:
A. Unique TensorFlower 2020-10-27 00:13:05 -07:00 committed by TensorFlower Gardener
parent 51e1e188a2
commit 9c38a2aeb7
10 changed files with 42 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
tensorflow/lite/delegates/nnapi/java/src/main/java/org/tensorflow/lite/nnapi/NnApiDelegate.java
View File

@ -76,8 +76,8 @@ public class NnApiDelegate implements Delegate, AutoCloseable {
}
/**
* Configure the location to be used to store model compilation cache entries. If either
* {@code cacheDir} or {@code modelToken} parameters are unset NNAPI caching will be disabled.
* Configure the location to be used to store model compilation cache entries. If either {@code
* cacheDir} or {@code modelToken} parameters are unset NNAPI caching will be disabled.
*
* <p>Only effective on Android 10 (API level 29) and above.
*/
@ -151,7 +151,7 @@ public class NnApiDelegate implements Delegate, AutoCloseable {
/*overrideDisallowCpu=*/ options.useNnapiCpu != null,
/*disallowCpuValue=*/ options.useNnapiCpu != null
? !options.useNnapiCpu.booleanValue()
: false,
: true,
options.allowFp16 != null ? options.allowFp16 : false);
}

14
tensorflow/lite/delegates/nnapi/nnapi_delegate.cc
View File

@ -565,14 +565,18 @@ TfLiteStatus GetTargetSdkVersion(
// If exclude_nnapi_reference is true this method will return false if the
// accelerator_name in the delegate options is equal to "nnapi-reference"
bool ShouldUseTargetDevices(StatefulNnApiDelegate::Options delegate_options,
const NnApi* nnapi,
bool exclude_nnapi_reference = false) {
const char* device_name_ptr = delegate_options.accelerator_name;
std::string nnapi_cpu("nnapi-reference");
bool has_selected_accelerator = device_name_ptr != nullptr;
if (exclude_nnapi_reference && has_selected_accelerator) {
has_selected_accelerator = nnapi_cpu != device_name_ptr;
if (nnapi_cpu == device_name_ptr) return false;
}
return (delegate_options.disallow_nnapi_cpu) || has_selected_accelerator;
return (delegate_options.disallow_nnapi_cpu &&
nnapi->android_sdk_version >=
delegate::nnapi::kMinSdkVersionForNNAPI12) ||
has_selected_accelerator;
}
// Fills the given result vector with the list of devices the given delegate
@ -3479,7 +3483,7 @@ TfLiteStatus NNAPIDelegateKernel::Init(TfLiteContext* context,
const auto delegate_options =
StatefulNnApiDelegate::GetOptions(params->delegate);
if (nnapi_->android_sdk_version >= kMinSdkVersionForNNAPI12 &&
ShouldUseTargetDevices(delegate_options)) {
ShouldUseTargetDevices(delegate_options, nnapi_)) {
TF_LITE_ENSURE_STATUS(GetTargetDevices(context, params->delegate, nnapi_,
nnapi_errno, &nnapi_devices_));
@ -4809,7 +4813,7 @@ TfLiteStatus StatefulNnApiDelegate::DoPrepare(TfLiteContext* context,
// If not, don't delegate to NNAPI's CPU reference implementation unless
// it has been specified as target accelerator.
if (nnapi->android_sdk_version >= kMinSdkVersionForNNAPI12) {
if (ShouldUseTargetDevices(delegate_options)) {
if (ShouldUseTargetDevices(delegate_options, nnapi)) {
std::vector<ANeuralNetworksDevice*> devices;
TF_LITE_ENSURE_STATUS(
GetTargetDevices(context, delegate, nnapi, nnapi_errno, &devices));
@ -4849,7 +4853,7 @@ TfLiteStatus StatefulNnApiDelegate::DoPrepare(TfLiteContext* context,
// Check for every node if it is supported
const bool is_accelerator_specified = ShouldUseTargetDevices(
delegate_options, /*exclude_nnapi_reference=*/true);
delegate_options, nnapi, /*exclude_nnapi_reference=*/true);
for (int node_index : TfLiteIntArrayView(plan)) {
TfLiteNode* node;
TfLiteRegistration* registration;

2
tensorflow/lite/delegates/nnapi/nnapi_delegate.h
View File

@ -79,7 +79,7 @@ class StatefulNnApiDelegate : public TfLiteDelegate {
// above. The NNAPI CPU typically performs less well than built-in TfLite
// kernels, but allowing CPU allows partial acceleration of models. If this
// is set to true, NNAPI is only used if the whole model is accelerated.
bool disallow_nnapi_cpu = false;
bool disallow_nnapi_cpu = true;
// Specifies the max number of partitions to delegate. A value <= 0 means
// no limit.

3
tensorflow/lite/delegates/nnapi/nnapi_delegate_device_selection_test.cc
View File

@ -114,7 +114,7 @@ struct NnApiDeviceSelectionTest
FloatAddOpModel m;
};
TEST_F(NnApiDeviceSelectionTest, DoesntSetDevicesWithoutFlags) {
TEST_F(NnApiDeviceSelectionTest, DoesntSetDevicesWhenCpuAllowed) {
nnapi_mock_->StubCompilationCreateForDevicesWith(
[](ANeuralNetworksModel* model,
const ANeuralNetworksDevice* const* devices, uint32_t numDevices,
@ -124,6 +124,7 @@ TEST_F(NnApiDeviceSelectionTest, DoesntSetDevicesWithoutFlags) {
});
tflite::StatefulNnApiDelegate::Options options;
options.disallow_nnapi_cpu = false;
InitWithOptions(options);
m.Invoke();
EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk);

5
tensorflow/lite/delegates/nnapi/nnapi_delegate_errno_test.cc
View File

@ -16,6 +16,7 @@ limitations under the License.
#include <gtest/gtest.h>
#include "tensorflow/lite/c/common.h"
#include "tensorflow/lite/core/subgraph.h"
#include "tensorflow/lite/delegates/nnapi/nnapi_delegate.h"
#include "tensorflow/lite/delegates/nnapi/nnapi_delegate_mock_test.h"
#include "tensorflow/lite/interpreter.h"
@ -30,7 +31,8 @@ namespace {
class SingleOpModelWithNNAPI : public SingleOpModel {
public:
explicit SingleOpModelWithNNAPI(const NnApi* nnapi) {
stateful_delegate_.reset(new StatefulNnApiDelegate(nnapi));
options_.disallow_nnapi_cpu = false;
stateful_delegate_.reset(new StatefulNnApiDelegate(nnapi, options_));
this->SetDelegate(stateful_delegate_.get());
}
@ -42,6 +44,7 @@ class SingleOpModelWithNNAPI : public SingleOpModel {
private:
std::unique_ptr<StatefulNnApiDelegate> stateful_delegate_;
StatefulNnApiDelegate::Options options_;
};
class FloatAddOpModel : public SingleOpModelWithNNAPI {

4
tensorflow/lite/delegates/nnapi/nnapi_delegate_signed_quantization_test.cc
View File

@ -41,7 +41,8 @@ class SingleOpModelWithNNAPI : public SingleOpModel {
public:
SingleOpModelWithNNAPI() = default;
void Init(const NnApi* nnapi) {
stateful_delegate_.reset(new StatefulNnApiDelegate(nnapi));
options_.disallow_nnapi_cpu = false;
stateful_delegate_.reset(new StatefulNnApiDelegate(nnapi, options_));
SetDelegate(stateful_delegate_.get());
}
@ -54,6 +55,7 @@ class SingleOpModelWithNNAPI : public SingleOpModel {
protected:
std::unique_ptr<StatefulNnApiDelegate> stateful_delegate_;
StatefulNnApiDelegate::Options options_;
TfLiteStatus compilation_status_;
};

11
tensorflow/lite/delegates/nnapi/nnapi_delegate_test.cc
View File

@ -47,11 +47,17 @@ MATCHER(QuantizedNear, "") {
class SingleOpModelWithNNAPI : public SingleOpModel {
public:
SingleOpModelWithNNAPI() { SetDelegate(NnApiDelegate()); }
SingleOpModelWithNNAPI() {
options_.disallow_nnapi_cpu = false;
stateful_delegate_.reset(new StatefulNnApiDelegate(options_));
SetDelegate(stateful_delegate_.get());
}
explicit SingleOpModelWithNNAPI(
const StatefulNnApiDelegate::Options& options) {
stateful_delegate_.reset(new StatefulNnApiDelegate(options));
options_ = options;
options_.disallow_nnapi_cpu = false;
stateful_delegate_.reset(new StatefulNnApiDelegate(options_));
SetDelegate(stateful_delegate_.get());
}
@ -107,6 +113,7 @@ class SingleOpModelWithNNAPI : public SingleOpModel {
private:
// Stateful NNAPI delegate. This is valid only if the state-ful constructor is
// used.
StatefulNnApiDelegate::Options options_;
std::unique_ptr<StatefulNnApiDelegate> stateful_delegate_;
};

1
tensorflow/lite/kernels/test_main.cc
View File

@ -34,6 +34,7 @@ void InitKernelTest(int* argc, char** argv) {
auto* params = delegate_providers->MutableParams();
if (!params->HasValueSet<std::string>("nnapi_accelerator_name")) {
params->Set<std::string>("nnapi_accelerator_name", "nnapi-reference");
params->Set("disable_nnapi_cpu", false);
}
}
}

7
tensorflow/lite/kernels/test_util.cc
View File

@ -230,6 +230,13 @@ TfLiteStatus SingleOpModel::ApplyDelegate() {
++num_applied_delegates_;
} else {
auto* delegate_providers = tflite::KernelTestDelegateProviders::Get();
// Most TFLite NNAPI delegation tests have been written to run against the
// NNAPI CPU path. We'll enable that for tests. However, need to first check
// if the parameter is present - it will not be if the NNAPI delegate
// provider is not linked into the test.
if (delegate_providers->ConstParams().HasParam("disable_nnapi_cpu")) {
delegate_providers->MutableParams()->Set("disable_nnapi_cpu", false);
}
for (auto& one : delegate_providers->CreateAllDelegates()) {
// The raw ptr always points to the actual TfLiteDegate object.
auto* delegate_raw_ptr = one.get();

6
tensorflow/lite/tools/delegates/nnapi_delegate_provider.cc
View File

@ -33,7 +33,7 @@ class NnapiDelegateProvider : public DelegateProvider {
default_params_.AddParam("nnapi_accelerator_name",
ToolParam::Create<std::string>(""));
default_params_.AddParam("disable_nnapi_cpu",
ToolParam::Create<bool>(false));
ToolParam::Create<bool>(true));
default_params_.AddParam("nnapi_allow_fp16",
ToolParam::Create<bool>(false));
}
@ -104,8 +104,8 @@ TfLiteDelegatePtr NnapiDelegateProvider::CreateTfLiteDelegate(
params.Get<std::string>("nnapi_accelerator_name");
if (!accelerator_name.empty()) {
options.accelerator_name = accelerator_name.c_str();
} else if (params.Get<bool>("disable_nnapi_cpu")) {
options.disallow_nnapi_cpu = true;
} else {
options.disallow_nnapi_cpu = params.Get<bool>("disable_nnapi_cpu");
}
if (params.Get<bool>("nnapi_allow_fp16")) {