IVGCVSW-4780 Add QoS to AndroidNNDriver

 * Add model priority to ArmnnPreparedModel_1_3
 * Add RequestThread_1_3 to allow execution based on priority
 * Add RETIRE_RATE to Android.mk to be able to configure the retire rate

Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com>
Change-Id: Ic5f4309249b744c2a8f625c986eede381a26028b

3 files changed, 8 insertions, 4 deletions

diff --git a/1.3/ArmnnDriver.hpp b/1.3/ArmnnDriver.hpp
index 798c4381..b6b55fae 100644
--- a/1.3/ArmnnDriver.hpp
+++ b/1.3/ArmnnDriver.hpp
@@ -206,7 +206,8 @@ public:
                                                       model,
                                                       cb,
                                                       model.relaxComputationFloat32toFloat16
-                                                      && m_Options.GetFp16Enabled());
+                                                      && m_Options.GetFp16Enabled(),
+                                                      priority);
     }
 
     Return<void> getSupportedExtensions(getSupportedExtensions_cb cb)
diff --git a/1.3/ArmnnDriverImpl.cpp b/1.3/ArmnnDriverImpl.cpp
index 4b2ff148..6168c9d0 100644
--- a/1.3/ArmnnDriverImpl.cpp
+++ b/1.3/ArmnnDriverImpl.cpp
@@ -101,7 +101,8 @@ Return<V1_3::ErrorStatus> ArmnnDriverImpl::prepareArmnnModel_1_3(
        const DriverOptions& options,
        const V1_3::Model& model,
        const sp<V1_3::IPreparedModelCallback>& cb,
-       bool float32ToFloat16)
+       bool float32ToFloat16,
+       V1_3::Priority priority)
 {
     ALOGV("ArmnnDriverImpl::prepareArmnnModel_1_3()");
 
@@ -204,7 +205,8 @@ Return<V1_3::ErrorStatus> ArmnnDriverImpl::prepareArmnnModel_1_3(
                     runtime.get(),
                     model,
                     options.GetRequestInputsAndOutputsDumpDir(),
-                    options.IsGpuProfilingEnabled()));
+                    options.IsGpuProfilingEnabled(),
+                    priority));
 
     // 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.
diff --git a/1.3/ArmnnDriverImpl.hpp b/1.3/ArmnnDriverImpl.hpp
index 8a665ea5..2b39d4e0 100644
--- a/1.3/ArmnnDriverImpl.hpp
+++ b/1.3/ArmnnDriverImpl.hpp
@@ -30,7 +30,8 @@ public:
                                                            const DriverOptions& options,
                                                            const V1_3::Model& model,
                                                            const android::sp<V1_3::IPreparedModelCallback>& cb,
-                                                           bool float32ToFloat16 = false);
+                                                           bool float32ToFloat16 = false,
+                                                           V1_3::Priority priority = V1_3::Priority::MEDIUM);
 
     static Return<void> getCapabilities_1_3(const armnn::IRuntimePtr& runtime,
                                             V1_3::IDevice::getCapabilities_1_3_cb cb);