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>
diff --git a/1.2/ArmnnDriverImpl.cpp b/1.2/ArmnnDriverImpl.cpp
index 3274a8a..1c31384 100644
--- a/1.2/ArmnnDriverImpl.cpp
+++ b/1.2/ArmnnDriverImpl.cpp
@@ -267,7 +267,9 @@
options.GetRequestInputsAndOutputsDumpDir(),
options.IsGpuProfilingEnabled(),
options.isAsyncModelExecutionEnabled(),
- options.getNoOfArmnnThreads()));
+ options.getNoOfArmnnThreads(),
+ options.isImportEnabled(),
+ options.isExportEnabled()));
// 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.
@@ -630,6 +632,8 @@
options.IsGpuProfilingEnabled(),
options.isAsyncModelExecutionEnabled(),
options.getNoOfArmnnThreads(),
+ options.isImportEnabled(),
+ options.isExportEnabled(),
true));
NotifyCallbackAndCheck(cb, V1_0::ErrorStatus::NONE, preparedModel.release());
diff --git a/1.3/ArmnnDriverImpl.cpp b/1.3/ArmnnDriverImpl.cpp
index c8b1d96..474e1c1 100644
--- a/1.3/ArmnnDriverImpl.cpp
+++ b/1.3/ArmnnDriverImpl.cpp
@@ -281,7 +281,9 @@
options.IsGpuProfilingEnabled(),
priority,
options.isAsyncModelExecutionEnabled(),
- options.getNoOfArmnnThreads()));
+ options.getNoOfArmnnThreads(),
+ options.isImportEnabled(),
+ options.isExportEnabled()));
// 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.
@@ -645,6 +647,8 @@
V1_3::Priority::MEDIUM,
options.isAsyncModelExecutionEnabled(),
options.getNoOfArmnnThreads(),
+ options.isImportEnabled(),
+ options.isExportEnabled(),
true));
NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release());
diff --git a/ArmnnDriverImpl.cpp b/ArmnnDriverImpl.cpp
index 0b3b919..89fa54f 100644
--- a/ArmnnDriverImpl.cpp
+++ b/ArmnnDriverImpl.cpp
@@ -202,30 +202,36 @@
options.GetRequestInputsAndOutputsDumpDir(),
options.IsGpuProfilingEnabled(),
options.isAsyncModelExecutionEnabled(),
- options.getNoOfArmnnThreads()));
+ options.getNoOfArmnnThreads(),
+ options.isImportEnabled(),
+ options.isExportEnabled()));
- // 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.
- if (!preparedModel->ExecuteWithDummyInputs())
+ if (std::find(options.GetBackends().begin(),
+ options.GetBackends().end(),
+ armnn::Compute::GpuAcc) != options.GetBackends().end())
{
- return FailPrepareModel(V1_0::ErrorStatus::GENERAL_FAILURE, "Network could not be executed", cb);
- }
-
- if (clTunedParameters &&
- options.GetClTunedParametersMode() == armnn::IGpuAccTunedParameters::Mode::UpdateTunedParameters)
- {
- // Now that we've done one inference the CL kernel parameters will have been tuned, so save the updated file.
- try
+ // 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.
+ if (!preparedModel->ExecuteWithDummyInputs())
{
- clTunedParameters->Save(options.GetClTunedParametersFile().c_str());
+ return FailPrepareModel(V1_0::ErrorStatus::GENERAL_FAILURE, "Network could not be executed", cb);
}
- catch (std::exception& error)
+
+ if (clTunedParameters &&
+ options.GetClTunedParametersMode() == armnn::IGpuAccTunedParameters::Mode::UpdateTunedParameters)
{
- ALOGE("ArmnnDriverImpl::prepareModel: Failed to save CL tuned parameters file '%s': %s",
- options.GetClTunedParametersFile().c_str(), error.what());
+ // Now that we've done one inference the CL kernel parameters will have been tuned, so save the updated file
+ try
+ {
+ clTunedParameters->Save(options.GetClTunedParametersFile().c_str());
+ }
+ catch (std::exception& error)
+ {
+ ALOGE("ArmnnDriverImpl::prepareModel: Failed to save CL tuned parameters file '%s': %s",
+ options.GetClTunedParametersFile().c_str(), error.what());
+ }
}
}
-
NotifyCallbackAndCheck(cb, V1_0::ErrorStatus::NONE, preparedModel);
return V1_0::ErrorStatus::NONE;
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)
{
diff --git a/ArmnnPreparedModel.hpp b/ArmnnPreparedModel.hpp
index 685d950..0d19c07 100644
--- a/ArmnnPreparedModel.hpp
+++ b/ArmnnPreparedModel.hpp
@@ -41,7 +41,9 @@
const std::string& requestInputsAndOutputsDumpDir,
const bool gpuProfilingEnabled,
const bool asyncModelExecutionEnabled = false,
- const unsigned int numberOfThreads = 1);
+ const unsigned int numberOfThreads = 1,
+ const bool importEnabled = false,
+ const bool exportEnabled = true);
virtual ~ArmnnPreparedModel();
@@ -111,6 +113,8 @@
static std::unique_ptr<armnn::Threadpool> m_Threadpool;
std::shared_ptr<armnn::IWorkingMemHandle> m_WorkingMemHandle;
const bool m_AsyncModelExecutionEnabled;
+ const bool m_EnableImport;
+ const bool m_EnableExport;
};
}
diff --git a/ArmnnPreparedModel_1_2.cpp b/ArmnnPreparedModel_1_2.cpp
index 7f35e60..37bc3a4 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>
@@ -151,7 +153,9 @@
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)
@@ -159,6 +163,8 @@
, m_RequestInputsAndOutputsDumpDir(requestInputsAndOutputsDumpDir)
, m_GpuProfilingEnabled(gpuProfilingEnabled)
, m_AsyncModelExecutionEnabled(asyncModelExecutionEnabled)
+ , m_EnableImport(importEnabled)
+ , m_EnableExport(exportEnabled)
, m_PreparedFromCache(false)
{
// Enable profiling if required.
@@ -192,6 +198,8 @@
const bool gpuProfilingEnabled,
const bool asyncModelExecutionEnabled,
const unsigned int numberOfThreads,
+ const bool importEnabled,
+ const bool exportEnabled,
const bool preparedFromCache)
: m_NetworkId(networkId)
, m_Runtime(runtime)
@@ -199,6 +207,8 @@
, m_RequestInputsAndOutputsDumpDir(requestInputsAndOutputsDumpDir)
, m_GpuProfilingEnabled(gpuProfilingEnabled)
, m_AsyncModelExecutionEnabled(asyncModelExecutionEnabled)
+ , m_EnableImport(importEnabled)
+ , m_EnableExport(exportEnabled)
, m_PreparedFromCache(preparedFromCache)
{
// Enable profiling if required.
@@ -531,7 +541,20 @@
else
{
ALOGW("ArmnnPreparedModel_1_2::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 (cb.ctx.measureTimings == V1_2::MeasureTiming::YES)
diff --git a/ArmnnPreparedModel_1_2.hpp b/ArmnnPreparedModel_1_2.hpp
index 255fc18..6afcbbe 100644
--- a/ArmnnPreparedModel_1_2.hpp
+++ b/ArmnnPreparedModel_1_2.hpp
@@ -47,7 +47,9 @@
const std::string& requestInputsAndOutputsDumpDir,
const bool gpuProfilingEnabled,
const bool asyncModelExecutionEnabled = false,
- const unsigned int numberOfThreads = 1);
+ const unsigned int numberOfThreads = 1,
+ const bool importEnabled = false,
+ const bool exportEnabled = true);
ArmnnPreparedModel_1_2(armnn::NetworkId networkId,
armnn::IRuntime* runtime,
@@ -55,6 +57,8 @@
const bool gpuProfilingEnabled,
const bool asyncModelExecutionEnabled = false,
const unsigned int numberOfThreads = 1,
+ const bool importEnabled = false,
+ const bool exportEnabled = true,
const bool preparedFromCache = false);
virtual ~ArmnnPreparedModel_1_2();
@@ -164,6 +168,8 @@
static std::unique_ptr<armnn::Threadpool> m_Threadpool;
std::shared_ptr<IWorkingMemHandle> m_WorkingMemHandle;
const bool m_AsyncModelExecutionEnabled;
+ const bool m_EnableImport;
+ const bool m_EnableExport;
const bool m_PreparedFromCache;
};
diff --git a/ArmnnPreparedModel_1_3.cpp b/ArmnnPreparedModel_1_3.cpp
index 36575b8..34c42ec 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>
@@ -172,7 +174,9 @@
const bool gpuProfilingEnabled,
V1_3::Priority priority,
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)
@@ -181,6 +185,8 @@
, m_GpuProfilingEnabled(gpuProfilingEnabled)
, m_ModelPriority(priority)
, m_AsyncModelExecutionEnabled(asyncModelExecutionEnabled)
+ , m_EnableImport(importEnabled)
+ , m_EnableExport(exportEnabled)
, m_PreparedFromCache(false)
{
// Enable profiling if required.
@@ -215,6 +221,8 @@
V1_3::Priority priority,
const bool asyncModelExecutionEnabled,
const unsigned int numberOfThreads,
+ const bool importEnabled,
+ const bool exportEnabled,
const bool preparedFromCache)
: m_NetworkId(networkId)
, m_Runtime(runtime)
@@ -223,6 +231,8 @@
, m_GpuProfilingEnabled(gpuProfilingEnabled)
, m_ModelPriority(priority)
, m_AsyncModelExecutionEnabled(asyncModelExecutionEnabled)
+ , m_EnableImport(importEnabled)
+ , m_EnableExport(exportEnabled)
, m_PreparedFromCache(preparedFromCache)
{
// Enable profiling if required.
@@ -824,7 +834,19 @@
else
{
ALOGW("ArmnnPreparedModel_1_3::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 (cb.ctx.measureTimings == V1_2::MeasureTiming::YES)
diff --git a/ArmnnPreparedModel_1_3.hpp b/ArmnnPreparedModel_1_3.hpp
index cd5fc0e..cb3c768 100644
--- a/ArmnnPreparedModel_1_3.hpp
+++ b/ArmnnPreparedModel_1_3.hpp
@@ -55,7 +55,9 @@
const bool gpuProfilingEnabled,
V1_3::Priority priority = V1_3::Priority::MEDIUM,
const bool asyncModelExecutionEnabled = false,
- const unsigned int numberOfThreads = 1);
+ const unsigned int numberOfThreads = 1,
+ const bool importEnabled = false,
+ const bool exportEnabled = true);
ArmnnPreparedModel_1_3(armnn::NetworkId networkId,
armnn::IRuntime* runtime,
@@ -64,6 +66,8 @@
V1_3::Priority priority = V1_3::Priority::MEDIUM,
const bool asyncModelExecutionEnabled = false,
const unsigned int numberOfThreads = 1,
+ const bool importEnabled = false,
+ const bool exportEnabled = true,
const bool preparedFromCache = false);
virtual ~ArmnnPreparedModel_1_3();
@@ -201,6 +205,8 @@
static std::unique_ptr<armnn::Threadpool> m_Threadpool;
std::shared_ptr<IWorkingMemHandle> m_WorkingMemHandle;
const bool m_AsyncModelExecutionEnabled;
+ const bool m_EnableImport;
+ const bool m_EnableExport;
const bool m_PreparedFromCache;
};
diff --git a/DriverOptions.cpp b/DriverOptions.cpp
index 8fd5c47..0018d97 100644
--- a/DriverOptions.cpp
+++ b/DriverOptions.cpp
@@ -40,6 +40,8 @@
, m_NumberOfThreads(0)
, m_EnableAsyncModelExecution(false)
, m_ArmnnNumberOfThreads(1)
+ , m_EnableImport(false)
+ , m_EnableExport(true)
{
}
@@ -56,6 +58,8 @@
, m_NumberOfThreads(0)
, m_EnableAsyncModelExecution(false)
, m_ArmnnNumberOfThreads(1)
+ , m_EnableImport(false)
+ , m_EnableExport(true)
{
}
@@ -71,6 +75,8 @@
, m_NumberOfThreads(0)
, m_EnableAsyncModelExecution(false)
, m_ArmnnNumberOfThreads(1)
+ , m_EnableImport(false)
+ , m_EnableExport(true)
{
std::string unsupportedOperationsAsString;
std::string clTunedParametersModeAsString;
@@ -168,7 +174,13 @@
"Assign the number of threads used by ArmNN. "
"Input value must be at least 1. "
"Default is set to 1.",
- cxxopts::value<unsigned int>(m_ArmnnNumberOfThreads)->default_value("1"));
+ cxxopts::value<unsigned int>(m_ArmnnNumberOfThreads)->default_value("1"))
+
+ ("I,enableImport", "Enable Importing of input buffers",
+ cxxopts::value<bool>(m_EnableImport)->default_value("false"))
+
+ ("E,enableExport", "Enable Exporting of output buffers",
+ cxxopts::value<bool>(m_EnableExport)->default_value("true"));
}
catch (const std::exception& e)
{
diff --git a/DriverOptions.hpp b/DriverOptions.hpp
index e1d25c4..ee68a94 100644
--- a/DriverOptions.hpp
+++ b/DriverOptions.hpp
@@ -42,6 +42,8 @@
unsigned int GetNumberOfThreads() const { return m_NumberOfThreads; }
bool isAsyncModelExecutionEnabled() const { return m_EnableAsyncModelExecution; };
unsigned int getNoOfArmnnThreads() const { return m_ArmnnNumberOfThreads; };
+ bool isImportEnabled() const { return m_EnableImport; };
+ bool isExportEnabled() const { return m_EnableExport; };
private:
std::vector<armnn::BackendId> m_Backends;
@@ -63,6 +65,8 @@
unsigned int m_NumberOfThreads;
bool m_EnableAsyncModelExecution;
unsigned int m_ArmnnNumberOfThreads;
+ bool m_EnableImport;
+ bool m_EnableExport;
};
} // namespace armnn_driver
diff --git a/test/Concurrent.cpp b/test/Concurrent.cpp
index 4113a8d..71119cd 100644
--- a/test/Concurrent.cpp
+++ b/test/Concurrent.cpp
@@ -82,15 +82,16 @@
// build the requests
V1_0::Request requests[maxRequests];
+ android::sp<IMemory> inMemory[maxRequests];
android::sp<IMemory> outMemory[maxRequests];
+ float indata[] = {2, 32, 16};
float* outdata[maxRequests];
for (size_t i = 0; i < maxRequests; ++i)
{
requests[i].inputs = hidl_vec<RequestArgument>{input};
requests[i].outputs = hidl_vec<RequestArgument>{output};
// set the input data (matching source test)
- float indata[] = {2, 32, 16};
- AddPoolAndSetData<float>(3, requests[i], indata);
+ inMemory[i] = AddPoolAndSetData<float>(3, requests[i], indata);
// add memory for the output
outMemory[i] = AddPoolAndGetData<float>(1, requests[i]);
outdata[i] = static_cast<float*>(static_cast<void*>(outMemory[i]->getPointer()));
diff --git a/test/DriverTestHelpers.hpp b/test/DriverTestHelpers.hpp
index 383c8fc..98be090 100644
--- a/test/DriverTestHelpers.hpp
+++ b/test/DriverTestHelpers.hpp
@@ -186,13 +186,15 @@
}
template<typename T>
-void AddPoolAndSetData(uint32_t size, V1_0::Request& request, const T* data)
+android::sp<IMemory> AddPoolAndSetData(uint32_t size, V1_0::Request& request, const T* data)
{
android::sp<IMemory> memory = AddPoolAndGetData<T>(size, request);
T* dst = static_cast<T*>(static_cast<void*>(memory->getPointer()));
memcpy(dst, data, size * sizeof(T));
+
+ return memory;
}
template<typename HalPolicy,