From cad4e91027a29a62c210d422ce1c9130e46f2199 Mon Sep 17 00:00:00 2001 From: Narumol Prangnawarat Date: Tue, 2 Jun 2020 12:07:43 +0100 Subject: 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 Change-Id: Ic5f4309249b744c2a8f625c986eede381a26028b --- 1.3/ArmnnDriverImpl.cpp | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) (limited to '1.3/ArmnnDriverImpl.cpp') 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 ArmnnDriverImpl::prepareArmnnModel_1_3( const DriverOptions& options, const V1_3::Model& model, const sp& cb, - bool float32ToFloat16) + bool float32ToFloat16, + V1_3::Priority priority) { ALOGV("ArmnnDriverImpl::prepareArmnnModel_1_3()"); @@ -204,7 +205,8 @@ Return 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. -- cgit v1.2.1