From 8509422e1e7d629f88d6a5e4e4bded8a682f435b Mon Sep 17 00:00:00 2001
From: Narumol Prangnawarat <narumol.prangnawarat@arm.com>
Date: Fri, 28 Jan 2022 23:26:07 +0000
Subject: IVGCVSW-6700 Enable import aligned host memory in android-nn-driver

!armnn:7025

Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com>
Change-Id: Iaa6112fb52d0f2942e6c52f90d96a8dc96000518
---
 ArmnnPreparedModel.cpp     | 16 ++++++++++++++--
 ArmnnPreparedModel_1_2.cpp |  9 ++++++++-
 ArmnnPreparedModel_1_3.cpp |  9 ++++++++-
 3 files changed, 30 insertions(+), 4 deletions(-)

diff --git a/ArmnnPreparedModel.cpp b/ArmnnPreparedModel.cpp
index 563abd44..ea48c0cc 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>
@@ -305,7 +307,12 @@ void ArmnnPreparedModel<HalVersion>::ExecuteGraph(
         else
         {
             ALOGW("ArmnnPreparedModel::ExecuteGraph m_AsyncModelExecutionEnabled false");
-            status = m_Runtime->EnqueueWorkload(m_NetworkId, inputTensors, outputTensors);
+            std::vector<armnn::ImportedInputId> importedInputIds =
+                m_Runtime->ImportInputs(m_NetworkId, inputTensors, armnn::MemorySource::Malloc);
+            std::vector<armnn::ImportedOutputId> 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)
@@ -386,7 +393,12 @@ bool ArmnnPreparedModel<HalVersion>::ExecuteWithDummyInputs()
         else
         {
             ALOGW("ArmnnPreparedModel::ExecuteGraph m_AsyncModelExecutionEnabled false");
-            status = m_Runtime->EnqueueWorkload(m_NetworkId, inputTensors, outputTensors);
+            std::vector<armnn::ImportedInputId> importedInputIds =
+                m_Runtime->ImportInputs(m_NetworkId, inputTensors, armnn::MemorySource::Malloc);
+            std::vector<armnn::ImportedOutputId> 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)
         {
diff --git a/ArmnnPreparedModel_1_2.cpp b/ArmnnPreparedModel_1_2.cpp
index 29055a70..08790d38 100644
--- a/ArmnnPreparedModel_1_2.cpp
+++ b/ArmnnPreparedModel_1_2.cpp
@@ -9,6 +9,8 @@
 
 #include "Utils.hpp"
 
+#include <armnn/Types.hpp>
+
 #include <log/log.h>
 #include <OperationsUtils.h>
 #include <ExecutionBurstServer.h>
@@ -528,7 +530,12 @@ bool ArmnnPreparedModel_1_2<HalVersion>::ExecuteGraph(
         else
         {
             ALOGW("ArmnnPreparedModel_1_2::ExecuteGraph m_AsyncModelExecutionEnabled false");
-            status = m_Runtime->EnqueueWorkload(m_NetworkId, inputTensors, outputTensors);
+            std::vector<armnn::ImportedInputId> importedInputIds =
+                m_Runtime->ImportInputs(m_NetworkId, inputTensors, armnn::MemorySource::Malloc);
+            std::vector<armnn::ImportedOutputId> importedOutputIds =
+                m_Runtime->ImportOutputs(m_NetworkId, outputTensors, armnn::MemorySource::Malloc);
+            status = m_Runtime->EnqueueWorkload(m_NetworkId, inputTensors, outputTensors,
+                                                importedInputIds, importedOutputIds);
         }
 
         if (cb.ctx.measureTimings == V1_2::MeasureTiming::YES)
diff --git a/ArmnnPreparedModel_1_3.cpp b/ArmnnPreparedModel_1_3.cpp
index abd135e8..236c4f27 100644
--- a/ArmnnPreparedModel_1_3.cpp
+++ b/ArmnnPreparedModel_1_3.cpp
@@ -12,6 +12,8 @@
 #include "ArmnnPreparedModel_1_3.hpp"
 #include "Utils.hpp"
 
+#include <armnn/Types.hpp>
+
 #include <Utils.h>
 #include <android/sync.h>
 #include <log/log.h>
@@ -821,7 +823,12 @@ Return <V1_3::ErrorStatus> ArmnnPreparedModel_1_3<HalVersion>::ExecuteGraph(
         else
         {
             ALOGW("ArmnnPreparedModel_1_3::ExecuteGraph m_AsyncModelExecutionEnabled false");
-            status = m_Runtime->EnqueueWorkload(m_NetworkId, inputTensors, outputTensors);
+            std::vector<armnn::ImportedInputId> importedInputIds =
+                m_Runtime->ImportInputs(m_NetworkId, inputTensors, armnn::MemorySource::Malloc);
+            std::vector<armnn::ImportedOutputId> importedOutputIds =
+                m_Runtime->ImportOutputs(m_NetworkId, outputTensors, armnn::MemorySource::Malloc);
+            status = m_Runtime->EnqueueWorkload(m_NetworkId, inputTensors, outputTensors,
+                                                importedInputIds, importedOutputIds);
         }
 
         if (cb.ctx.measureTimings == V1_2::MeasureTiming::YES)
-- 
cgit v1.2.1