commit d1a947fd487eb7674271f759e43507d10e932ab0
author Narumol Prangnawarat <narumol.prangnawarat@arm.com> Mon Feb 07 13:12:24 2022 +0000
committer David Monahan <david.monahan@arm.com> Wed Feb 16 14:42:00 2022 +0000
tree 17ad41e29935a657c5025485173b4ded85c4a47c
parent 29520c4c785fff938586ec231cbf5e51dadbf614

Revert "Revert "IVGCVSW-6700 Enable import aligned host memory in android-nn-driver""

This reverts commit 8069603dc44b7673b356f66517cd8b25af8080f0.

 * Reason for revert: Try reenable import aligned host memory in android-nn-driver
 * Added a check to ArmNNDriverImpl.cpp to not call ExecuteWithDummyInputs with GpuAcc
 * Added new android-nn-driver driver options to enable / disable Import and Export
 * Import is disabled by default for now due to conv2d issues
 * Export is enabled by default

!armnn:7147

Change-Id: I91110c58ebb3931d1c458e3774944e55c1250dd8
Signed-off-by: David Monahan <David.Monahan@arm.com>

ArmnnPreparedModel.cpp

diff --git a/ArmnnPreparedModel.cpp b/ArmnnPreparedModel.cpp
index 38f1bc2..326351c 100644
--- a/ArmnnPreparedModel.cpp
+++ b/ArmnnPreparedModel.cpp
@@ -8,6 +8,8 @@
 #include "ArmnnPreparedModel.hpp"
 #include "Utils.hpp"
 
+#include <armnn/Types.hpp>
+
 #include <log/log.h>
 #include <OperationsUtils.h>
 #include <ValidateHal.h>
@@ -116,7 +118,9 @@
                                                    const std::string& requestInputsAndOutputsDumpDir,
                                                    const bool gpuProfilingEnabled,
                                                    const bool asyncModelExecutionEnabled,
-                                                   const unsigned int numberOfThreads)
+                                                   const unsigned int numberOfThreads,
+                                                   const bool importEnabled,
+                                                   const bool exportEnabled)
     : m_NetworkId(networkId)
     , m_Runtime(runtime)
     , m_Model(model)
@@ -124,6 +128,8 @@
     , m_RequestInputsAndOutputsDumpDir(requestInputsAndOutputsDumpDir)
     , m_GpuProfilingEnabled(gpuProfilingEnabled)
     , m_AsyncModelExecutionEnabled(asyncModelExecutionEnabled)
+    , m_EnableImport(importEnabled)
+    , m_EnableExport(exportEnabled)
 {
     // Enable profiling if required.
     m_Runtime->GetProfiler(m_NetworkId)->EnableProfiling(m_GpuProfilingEnabled);
@@ -308,7 +314,19 @@
         else
         {
             ALOGW("ArmnnPreparedModel::ExecuteGraph m_AsyncModelExecutionEnabled false");
-            status = m_Runtime->EnqueueWorkload(m_NetworkId, inputTensors, outputTensors);
+            // Create a vector of Input and Output Ids which can be imported. An empty vector means all will be copied.
+            std::vector<armnn::ImportedInputId> importedInputIds;
+            if (m_EnableImport)
+            {
+                importedInputIds =  m_Runtime->ImportInputs(m_NetworkId, inputTensors, armnn::MemorySource::Malloc);
+            }
+            std::vector<armnn::ImportedOutputId> importedOutputIds;
+            if (m_EnableExport)
+            {
+                importedOutputIds = m_Runtime->ImportOutputs(m_NetworkId, outputTensors, armnn::MemorySource::Malloc);
+            }
+            status = m_Runtime->EnqueueWorkload(m_NetworkId, inputTensors, outputTensors,
+                                                importedInputIds, importedOutputIds);
         }
 
         if (status != armnn::Status::Success)
@@ -389,7 +407,19 @@
         else
         {
             ALOGW("ArmnnPreparedModel::ExecuteGraph m_AsyncModelExecutionEnabled false");
-            status = m_Runtime->EnqueueWorkload(m_NetworkId, inputTensors, outputTensors);
+            // Create a vector of Input and Output Ids which can be imported. An empty vector means all will be copied.
+            std::vector<armnn::ImportedInputId> importedInputIds;
+            if (m_EnableImport)
+            {
+                importedInputIds =  m_Runtime->ImportInputs(m_NetworkId, inputTensors, armnn::MemorySource::Malloc);
+            }
+            std::vector<armnn::ImportedOutputId> importedOutputIds;
+            if (m_EnableExport)
+            {
+                importedOutputIds = m_Runtime->ImportOutputs(m_NetworkId, outputTensors, armnn::MemorySource::Malloc);
+            }
+            status = m_Runtime->EnqueueWorkload(m_NetworkId, inputTensors, outputTensors,
+                                                importedInputIds, importedOutputIds);
         }
         if (status != armnn::Status::Success)
         {