IVGCVSW-2771 Fix SubTensor error in vgg16 ExecuteNetwork NEON
* Add check if Sub-tensors cannot be used, call ACL function
* Add computation of SplitAxis from SplitterDescriptor
* Add NeonSplitterWorkload functions
* Modify IsSplitterSupported to call ACL validate function
if sub-tensor cannot be used
* Also check if quantization parameters match when using sub-tensors
* Add more unit tests for Splitter in TfParser and TfLiteParser
Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com>
Change-Id: I31e4c7d055117c83c65b598c4125442173242226
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index f1c773d..8b2ad63 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -58,6 +58,7 @@
NeonSoftmaxFloatWorkload.hpp
NeonSoftmaxUint8Workload.cpp
NeonSoftmaxUint8Workload.hpp
+ NeonSplitterWorkload.cpp
NeonSplitterWorkload.hpp
NeonSubtractionWorkload.cpp
NeonSubtractionWorkload.hpp
diff --git a/src/backends/neon/workloads/NeonSplitterWorkload.cpp b/src/backends/neon/workloads/NeonSplitterWorkload.cpp
new file mode 100644
index 0000000..bf35939
--- /dev/null
+++ b/src/backends/neon/workloads/NeonSplitterWorkload.cpp
@@ -0,0 +1,112 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonSplitterWorkload.hpp"
+
+#include "NeonWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+#include <aclCommon/ArmComputeUtils.hpp>
+#include <backendsCommon/CpuTensorHandle.hpp>
+#include <neon/NeonTensorHandle.hpp>
+
+
+namespace armnn
+{
+
+using namespace armcomputetensorutils;
+
+namespace
+{
+unsigned int CalcAclAxis(unsigned int numDimensions, unsigned int splitAxis)
+{
+ return (numDimensions - splitAxis) - 1;
+}
+
+} //namespace
+
+arm_compute::Status NeonSplitterWorkloadValidate(const TensorInfo& input,
+ const std::vector<std::reference_wrapper<TensorInfo>>& outputs,
+ unsigned int splitAxis)
+{
+ const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input);
+
+ size_t numOutputs = outputs.size();
+
+ std::vector<arm_compute::TensorInfo> aclOutputs;
+ aclOutputs.reserve(numOutputs);
+
+ std::vector<arm_compute::ITensorInfo*> aclOutputPtr;
+ aclOutputPtr.reserve(numOutputs);
+
+ for (size_t i = 0u; i < outputs.size(); ++i)
+ {
+ aclOutputs.emplace_back(BuildArmComputeTensorInfo(outputs[i]));
+ aclOutputPtr.emplace_back(&aclOutputs.back());
+ }
+
+ unsigned int aclAxis = CalcAclAxis(input.GetNumDimensions(), splitAxis);
+ return arm_compute::NESplit::validate(&aclInputInfo, aclOutputPtr, aclAxis);
+}
+
+NeonSplitterWorkload::NeonSplitterWorkload(const SplitterQueueDescriptor& descriptor, const WorkloadInfo& info)
+ : BaseWorkload<SplitterQueueDescriptor>(descriptor, info)
+{
+ bool allOutputsAreSubtensors = true;
+
+ // Check that all outputs are sub-tensors
+ for (auto output : m_Data.m_Outputs)
+ {
+ if (output && !output->GetParent())
+ {
+ // Non sub-tensor input found so we need to execute the split function
+ allOutputsAreSubtensors = false;
+ break;
+ }
+ }
+
+ if (allOutputsAreSubtensors)
+ {
+ // Can skip configuring the split function since it's not executed
+ return;
+ }
+
+ arm_compute::ITensor& input = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+
+ std::vector<arm_compute::ITensor *> aclOutputs;
+ for (auto output : m_Data.m_Outputs)
+ {
+ arm_compute::ITensor& aclOutput = boost::polymorphic_pointer_downcast<INeonTensorHandle>(output)->GetTensor();
+ aclOutputs.emplace_back(&aclOutput);
+ }
+
+ // Create the layer function
+ m_Layer.reset(new arm_compute::NESplit());
+
+ // Configure input and output tensors
+ std::set<unsigned int> splitAxis = ComputeSplitAxis(descriptor.m_Parameters, m_Data.m_Inputs[0]->GetShape());
+ if (splitAxis.size() != 1)
+ {
+ throw InvalidArgumentException("Cannot derive split axis from SplitterDescriptor");
+ }
+
+ unsigned int aclAxis = CalcAclAxis(descriptor.m_Parameters.GetNumDimensions(), *splitAxis.begin());
+ m_Layer->configure(&input, aclOutputs, aclAxis);
+
+ // Prepare
+ m_Layer->prepare();
+}
+
+void NeonSplitterWorkload::Execute() const
+{
+ if (m_Layer)
+ {
+ ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonSplitterWorkload_Execute");
+ m_Layer->run();
+ }
+}
+
+} //namespace armnn
+
diff --git a/src/backends/neon/workloads/NeonSplitterWorkload.hpp b/src/backends/neon/workloads/NeonSplitterWorkload.hpp
index 2a7ee19..f902566 100644
--- a/src/backends/neon/workloads/NeonSplitterWorkload.hpp
+++ b/src/backends/neon/workloads/NeonSplitterWorkload.hpp
@@ -7,18 +7,26 @@
#include <backendsCommon/Workload.hpp>
+#include <arm_compute/runtime/NEON/NEFunctions.h>
+
+#include <functional>
+
namespace armnn
{
+arm_compute::Status NeonSplitterWorkloadValidate(const TensorInfo& input,
+ const std::vector<std::reference_wrapper<TensorInfo>>& outputs,
+ unsigned int splitAxis);
+
class NeonSplitterWorkload : public BaseWorkload<SplitterQueueDescriptor>
{
public:
- using BaseWorkload<SplitterQueueDescriptor>::BaseWorkload;
+ NeonSplitterWorkload(const SplitterQueueDescriptor& descriptor, const WorkloadInfo& info);
- virtual void Execute() const override
- {
- // With subtensors, splitter is a no-op.
- }
+ void Execute() const override;
+
+private:
+ mutable std::unique_ptr<arm_compute::NESplit> m_Layer;
};
} //namespace armnn