IVGCVSW-1806 Refactored Android-NN-Driver, added common "getCapabilities",

"getSupportedOperations" and "prepareModel" implementations

 * Added common base ArmnnDriverImpl class
 * Added common template implementation of the driver's "getCapabilities",
   "getSupportedOperations" and "prepareModel" methods
 * Refactored ArmnnPreparedModel and RequestThread to support HAL
   v1.1 models
 * Moved "getStatus" to the common base class, as it is shared by both
   HAL implementations
 * Refactored the code where necessary

Change-Id: I747334730026d63b4002662523fb93608f67c899

1 files changed, 288 insertions, 0 deletions

diff --git a/ArmnnDriverImpl.cpp b/ArmnnDriverImpl.cpp
new file mode 100644
index 00000000..0298f3b7
--- /dev/null
+++ b/ArmnnDriverImpl.cpp
@@ -0,0 +1,288 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// See LICENSE file in the project root for full license information.
+//
+
+#define LOG_TAG "ArmnnDriver"
+
+#include "ArmnnDriverImpl.hpp"
+#include "ModelToINetworkConverter.hpp"
+#include "ArmnnPreparedModel.hpp"
+#include "SystemPropertiesUtils.hpp"
+
+#if defined(ARMNN_ANDROID_P)
+// The headers of the ML framework have changed between Android O and Android P.
+// The validation functions have been moved into their own header, ValidateHal.h.
+#include <ValidateHal.h>
+#endif
+
+#include <log/log.h>
+
+using namespace std;
+using namespace android;
+using namespace android::nn;
+using namespace android::hardware;
+
+namespace
+{
+
+const char *g_Float32PerformanceExecTimeName = "ArmNN.float32Performance.execTime";
+const char *g_Float32PerformancePowerUsageName = "ArmNN.float32Performance.powerUsage";
+const char *g_Quantized8PerformanceExecTimeName = "ArmNN.quantized8Performance.execTime";
+const char *g_Quantized8PerformancePowerUsageName = "ArmNN.quantized8Performance.powerUsage";
+
+void NotifyCallbackAndCheck(const sp<IPreparedModelCallback>& callback,
+                            ErrorStatus errorStatus,
+                            const sp<IPreparedModel>& preparedModelPtr)
+{
+    Return<void> returned = callback->notify(errorStatus, preparedModelPtr);
+    // This check is required, if the callback fails and it isn't checked it will bring down the service
+    if (!returned.isOk())
+    {
+        ALOGE("ArmnnDriverImpl::prepareModel: hidl callback failed to return properly: %s ",
+            returned.description().c_str());
+    }
+}
+
+Return<ErrorStatus> FailPrepareModel(ErrorStatus error,
+                                     const string& message,
+                                     const sp<IPreparedModelCallback>& callback)
+{
+    ALOGW("ArmnnDriverImpl::prepareModel: %s", message.c_str());
+    NotifyCallbackAndCheck(callback, error, nullptr);
+    return error;
+}
+
+} // namespace
+
+namespace armnn_driver
+{
+
+template <typename HalVersion>
+Return<void> ArmnnDriverImpl<HalVersion>::getCapabilities(
+        const armnn::IRuntimePtr& runtime,
+        HalGetCapabilities_cb cb)
+{
+    ALOGV("ArmnnDriverImpl::getCapabilities()");
+
+    HalCapabilities capabilities;
+    if (runtime)
+    {
+        capabilities.float32Performance.execTime =
+            ParseSystemProperty(g_Float32PerformanceExecTimeName, .1f);
+
+        capabilities.float32Performance.powerUsage =
+            ParseSystemProperty(g_Float32PerformancePowerUsageName, .1f);
+
+        capabilities.quantized8Performance.execTime =
+            ParseSystemProperty(g_Quantized8PerformanceExecTimeName, .1f);
+
+        capabilities.quantized8Performance.powerUsage =
+            ParseSystemProperty(g_Quantized8PerformancePowerUsageName, .1f);
+
+        cb(ErrorStatus::NONE, capabilities);
+    }
+    else
+    {
+        capabilities.float32Performance.execTime = 0;
+        capabilities.float32Performance.powerUsage = 0;
+        capabilities.quantized8Performance.execTime = 0;
+        capabilities.quantized8Performance.powerUsage = 0;
+
+        cb(ErrorStatus::DEVICE_UNAVAILABLE, capabilities);
+    }
+
+    return Void();
+}
+
+template <typename HalVersion>
+Return<void> ArmnnDriverImpl<HalVersion>::getSupportedOperations(
+        const armnn::IRuntimePtr& runtime,
+        const DriverOptions& options,
+        const HalModel& model,
+        HalGetSupportedOperations_cb cb)
+{
+    ALOGV("ArmnnDriverImpl::getSupportedOperations()");
+
+    vector<bool> result;
+
+    if (!runtime)
+    {
+        cb(ErrorStatus::DEVICE_UNAVAILABLE, result);
+        return Void();
+    }
+
+    // Run general model validation, if this doesn't pass we shouldn't analyse the model anyway.
+    if (!android::nn::validateModel(model))
+    {
+        cb(ErrorStatus::INVALID_ARGUMENT, result);
+        return Void();
+    }
+
+    // Attempt to convert the model to an ArmNN input network (INetwork).
+    ModelToINetworkConverter<HalVersion> modelConverter(options.GetComputeDevice(),
+                                                        model,
+                                                        options.GetForcedUnsupportedOperations());
+
+    if (modelConverter.GetConversionResult() != ConversionResult::Success
+            && modelConverter.GetConversionResult() != ConversionResult::UnsupportedFeature)
+    {
+        cb(ErrorStatus::GENERAL_FAILURE, result);
+        return Void();
+    }
+
+    // Check each operation if it was converted successfully and copy the flags
+    // into the result (vector<bool>) that we need to return to Android.
+    result.reserve(model.operations.size());
+    for (uint32_t operationIdx = 0; operationIdx < model.operations.size(); operationIdx++)
+    {
+        bool operationSupported = modelConverter.IsOperationSupported(operationIdx);
+        result.push_back(operationSupported);
+    }
+
+    cb(ErrorStatus::NONE, result);
+    return Void();
+}
+
+template <typename HalVersion>
+Return<ErrorStatus> ArmnnDriverImpl<HalVersion>::prepareModel(
+        const armnn::IRuntimePtr& runtime,
+        const armnn::IGpuAccTunedParametersPtr& clTunedParameters,
+        const DriverOptions& options,
+        const HalModel& model,
+        const sp<IPreparedModelCallback>& cb,
+        bool float32ToFloat16)
+{
+    ALOGV("ArmnnDriverImpl::prepareModel()");
+
+    if (cb.get() == nullptr)
+    {
+        ALOGW("ArmnnDriverImpl::prepareModel: Invalid callback passed to prepareModel");
+        return ErrorStatus::INVALID_ARGUMENT;
+    }
+
+    if (!runtime)
+    {
+        return FailPrepareModel(ErrorStatus::DEVICE_UNAVAILABLE,
+                                "ArmnnDriverImpl::prepareModel: Device unavailable", cb);
+    }
+
+    if (!android::nn::validateModel(model))
+    {
+        return FailPrepareModel(ErrorStatus::INVALID_ARGUMENT,
+                                "ArmnnDriverImpl::prepareModel: Invalid model passed as input", cb);
+    }
+
+    // Deliberately ignore any unsupported operations requested by the options -
+    // at this point we're being asked to prepare a model that we've already declared support for
+    // and the operation indices may be different to those in getSupportedOperations anyway.
+    set<unsigned int> unsupportedOperations;
+    ModelToINetworkConverter<HalVersion> modelConverter(options.GetComputeDevice(),
+                                                        model,
+                                                        unsupportedOperations);
+
+    if (modelConverter.GetConversionResult() != ConversionResult::Success)
+    {
+        FailPrepareModel(ErrorStatus::GENERAL_FAILURE,
+                         "ArmnnDriverImpl::prepareModel: ModelToINetworkConverter failed", cb);
+        return ErrorStatus::NONE;
+    }
+
+    // Optimize the network
+    armnn::IOptimizedNetworkPtr optNet(nullptr, nullptr);
+    armnn::OptimizerOptions OptOptions;
+    OptOptions.m_ReduceFp32ToFp16 = float32ToFloat16;
+
+    try
+    {
+        optNet = armnn::Optimize(*modelConverter.GetINetwork(),
+                                 {options.GetComputeDevice()},
+                                 runtime->GetDeviceSpec(),
+                                 OptOptions);
+    }
+    catch (armnn::Exception &e)
+    {
+        stringstream message;
+        message << "ArmnnDriverImpl::prepareModel: armnn::Exception (" << e.what() << ") caught from optimize.";
+        FailPrepareModel(ErrorStatus::GENERAL_FAILURE, message.str(), cb);
+        return ErrorStatus::NONE;
+    }
+
+    // Check that the optimized network is valid.
+    if (!optNet)
+    {
+        FailPrepareModel(ErrorStatus::GENERAL_FAILURE,
+                         "ArmnnDriverImpl::prepareModel: Invalid optimized network", cb);
+        return ErrorStatus::NONE;
+    }
+
+    // Export the optimized network graph to a dot file if an output dump directory
+    // has been specified in the drivers' arguments.
+    ExportNetworkGraphToDotFile<HalModel>(*optNet, options.GetRequestInputsAndOutputsDumpDir(), model);
+
+    // Load it into the runtime.
+    armnn::NetworkId netId = 0;
+    try
+    {
+        if (runtime->LoadNetwork(netId, move(optNet)) != armnn::Status::Success)
+        {
+            return FailPrepareModel(ErrorStatus::GENERAL_FAILURE,
+                                    "ArmnnDriverImpl::prepareModel: Network could not be loaded", cb);
+        }
+    }
+    catch (armnn::Exception& e)
+    {
+        stringstream message;
+        message << "ArmnnDriverImpl::prepareModel: armnn::Exception (" << e.what()<< ") caught from LoadNetwork.";
+        FailPrepareModel(ErrorStatus::GENERAL_FAILURE, message.str(), cb);
+        return ErrorStatus::NONE;
+    }
+
+    unique_ptr<ArmnnPreparedModel<HalVersion>> preparedModel(
+                new ArmnnPreparedModel<HalVersion>(
+                    netId,
+                    runtime.get(),
+                    model,
+                    options.GetRequestInputsAndOutputsDumpDir(),
+                    options.IsGpuProfilingEnabled()));
+
+    // Run a single 'dummy' inference of the model. This means that CL kernels will get compiled (and tuned if
+    // this is enabled) before the first 'real' inference which removes the overhead of the first inference.
+    preparedModel->ExecuteWithDummyInputs();
+
+    if (clTunedParameters &&
+        options.GetClTunedParametersMode() == armnn::IGpuAccTunedParameters::Mode::UpdateTunedParameters)
+    {
+        // Now that we've done one inference the CL kernel parameters will have been tuned, so save the updated file.
+        try
+        {
+            clTunedParameters->Save(options.GetClTunedParametersFile().c_str());
+        }
+        catch (const armnn::Exception& error)
+        {
+            ALOGE("ArmnnDriverImpl::prepareModel: Failed to save CL tuned parameters file '%s': %s",
+                options.GetClTunedParametersFile().c_str(), error.what());
+        }
+    }
+
+    NotifyCallbackAndCheck(cb, ErrorStatus::NONE, preparedModel.release());
+
+    return ErrorStatus::NONE;
+}
+
+template <typename HalVersion>
+Return<DeviceStatus> ArmnnDriverImpl<HalVersion>::getStatus()
+{
+    ALOGV("ArmnnDriver::getStatus()");
+
+    return DeviceStatus::AVAILABLE;
+}
+
+// Class template specializations
+template class ArmnnDriverImpl<HalVersion_1_0>;
+
+#ifdef ARMNN_ANDROID_NN_V1_1
+template class ArmnnDriverImpl<HalVersion_1_1>;
+#endif
+
+} // namespace armnn_driver