From 42477c1d3e7ddf74863e84ab79dbe6f42e4a0ba3 Mon Sep 17 00:00:00 2001
From: Kevin May <kevin.may@arm.com>
Date: Thu, 26 Mar 2020 13:34:14 +0000
Subject: IVGCVSW-4447 Add Hal 1_3 Support

* Add new 1.3 files HalPolicy, ArmnnDriver, ArmnnDriverImpl
* Add new .rc file for 1.3 service
* Add ArmnnPreparedModel_1_3 and implement new functions
* Update Android.mk with 1.3 driver and service
* Refactor ifdef to include ARMNN_ANDROID_NN_V1_3
* Create Utils getMainModel for new 1.3 Model Main Subgraph
* Use android Utils to convertToV1_X in ArmnnPrepapredModel_1_3
* Refactor HAL 1.2 convert functions into ConversionUtils_1_2.hpp
* Replace ArmnnBurstExecutorWithCache with call to ExecutionBurstServer

Signed-off-by: Kevin May <kevin.may@arm.com>
Change-Id: I514069e9e1b16bcd1c4abfb5d563d25ac22d02e3
---
 ArmnnPreparedModel_1_3.hpp | 137 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 137 insertions(+)
 create mode 100644 ArmnnPreparedModel_1_3.hpp

(limited to 'ArmnnPreparedModel_1_3.hpp')

diff --git a/ArmnnPreparedModel_1_3.hpp b/ArmnnPreparedModel_1_3.hpp
new file mode 100644
index 00000000..247149c8
--- /dev/null
+++ b/ArmnnPreparedModel_1_3.hpp
@@ -0,0 +1,137 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "ArmnnDriver.hpp"
+#include "ArmnnDriverImpl.hpp"
+#include "RequestThread.hpp"
+#include "ModelToINetworkConverter.hpp"
+
+#include <NeuralNetworks.h>
+#include <armnn/ArmNN.hpp>
+
+#include <string>
+#include <vector>
+
+namespace armnn_driver
+{
+using CallbackAsync_1_3 = std::function<
+                                void(V1_3::ErrorStatus errorStatus,
+                                std::vector<::android::hardware::neuralnetworks::V1_2::OutputShape> outputShapes,
+                                const ::android::hardware::neuralnetworks::V1_2::Timing& timing,
+                                std::string callingFunction)>;
+
+struct ExecutionContext_1_3
+{
+    ::android::hardware::neuralnetworks::V1_2::MeasureTiming    measureTimings =
+        ::android::hardware::neuralnetworks::V1_2::MeasureTiming::NO;
+    TimePoint driverStart;
+};
+
+using CallbackContext_1_3 = CallbackContext<CallbackAsync_1_3, ExecutionContext_1_3>;
+
+using executeFenced_cb = std::function<void(::android::hardware::neuralnetworks::V1_3::ErrorStatus status,
+    const ::android::hardware::hidl_handle& syncFence,
+    const ::android::sp<::android::hardware::neuralnetworks::V1_3::IFencedExecutionCallback>& callback)>;
+
+template <typename HalVersion>
+class ArmnnPreparedModel_1_3 : public V1_3::IPreparedModel
+{
+public:
+    using HalModel = typename V1_3::Model;
+
+    ArmnnPreparedModel_1_3(armnn::NetworkId networkId,
+                           armnn::IRuntime* runtime,
+                           const HalModel& model,
+                           const std::string& requestInputsAndOutputsDumpDir,
+                           const bool gpuProfilingEnabled);
+
+    virtual ~ArmnnPreparedModel_1_3();
+
+    Return<V1_0::ErrorStatus> execute(const V1_0::Request& request,
+                                      const sp<V1_0::IExecutionCallback>& callback) override;
+
+    Return<V1_0::ErrorStatus> execute_1_2(const V1_0::Request& request, MeasureTiming measure,
+                                          const sp<V1_2::IExecutionCallback>& callback) override;
+
+    Return<V1_3::ErrorStatus> execute_1_3(const V1_3::Request& request,
+                                          V1_2::MeasureTiming measure,
+                                          const V1_3::OptionalTimePoint&,
+                                          const sp<V1_3::IExecutionCallback>& callback) override;
+
+    Return<void> executeSynchronously(const V1_0::Request &request,
+                                      MeasureTiming measure,
+                                      V1_3::IPreparedModel::executeSynchronously_cb cb) override;
+
+    Return<void> executeSynchronously_1_3(const V1_3::Request &request,
+                                          MeasureTiming measure,
+                                          const V1_3::OptionalTimePoint& deadline,
+                                          V1_3::IPreparedModel::executeSynchronously_1_3_cb cb) override;
+
+    Return<void> executeFenced(const V1_3::Request& request,
+                               const android::hardware::hidl_vec<android::hardware::hidl_handle>& wait_for,
+                               MeasureTiming measure,
+                               const V1_3::OptionalTimePoint& deadline,
+                               const V1_3::OptionalTimeoutDuration& duration,
+                               executeFenced_cb callback) override;
+
+    Return<void> configureExecutionBurst(
+            const sp<V1_2::IBurstCallback>& callback,
+            const android::hardware::MQDescriptorSync<V1_2::FmqRequestDatum>& requestChannel,
+            const android::hardware::MQDescriptorSync<V1_2::FmqResultDatum>& resultChannel,
+            configureExecutionBurst_cb cb) override;
+
+    template<typename CallbackContext>
+    Return<void> ExecuteSynchronously(const V1_3::Request& request, CallbackContext cbCtx);
+
+    /// execute the graph prepared from the request
+    template<typename CallbackContext>
+    bool ExecuteGraph(std::shared_ptr<std::vector<::android::nn::RunTimePoolInfo>>& pMemPools,
+                      armnn::InputTensors& inputTensors,
+                      armnn::OutputTensors& outputTensors,
+                      CallbackContext callback);
+
+    /// Executes this model with dummy inputs (e.g. all zeroes).
+    /// \return false on failure, otherwise true
+    bool ExecuteWithDummyInputs();
+
+private:
+    Return <V1_3::ErrorStatus> Execute(const V1_3::Request& request,
+                                       MeasureTiming measureTiming,
+                                       CallbackAsync_1_3 callback);
+
+    Return<V1_3::ErrorStatus> PrepareMemoryForInputs(
+        armnn::InputTensors& inputs,
+        const V1_3::Request& request,
+        const std::vector<android::nn::RunTimePoolInfo>& memPools);
+
+    Return<V1_3::ErrorStatus> PrepareMemoryForOutputs(
+        armnn::OutputTensors& outputs,
+        std::vector<OutputShape> &outputShapes,
+        const V1_3::Request& request,
+        const std::vector<android::nn::RunTimePoolInfo>& memPools);
+
+    std::tuple<V1_3::ErrorStatus, hidl_vec<OutputShape>, Timing, std::string> PrepareMemoryForIO(
+        armnn::InputTensors& inputs,
+        armnn::OutputTensors& outputs,
+        std::vector<android::nn::RunTimePoolInfo>& memPools,
+        const V1_3::Request& request);
+
+    template <typename TensorBindingCollection>
+    void DumpTensorsIfRequired(char const* tensorNamePrefix, const TensorBindingCollection& tensorBindings);
+
+    armnn::NetworkId                                                            m_NetworkId;
+    armnn::IRuntime*                                                            m_Runtime;
+    V1_3::Model                                                                 m_Model;
+    // There must be a single RequestThread for all ArmnnPreparedModel objects to ensure serial execution of workloads
+    // It is specific to this class, so it is declared as static here
+    static RequestThread<ArmnnPreparedModel_1_3, HalVersion, CallbackContext_1_3> m_RequestThread;
+    uint32_t                                                                    m_RequestCount;
+    const std::string&                                                          m_RequestInputsAndOutputsDumpDir;
+    const bool                                                                  m_GpuProfilingEnabled;
+};
+
+}
-- 
cgit v1.2.1