IVGCVSW-3148 Add Neon backend support for Dequantize

 * Add NeonDequantizeWorkload
 * Add IsDequantizeSupported to call validate from ACL function
 * Add CreateDequantize to NeonWorkloadFactory
 * Unit tests

Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com>
Change-Id: I96a216ef78cc3f6a57aa439a16ae6aafd783ff93

10 files changed, 111 insertions, 0 deletions

diff --git a/src/backends/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp
index a85597bbf8..32027d46e7 100644
--- a/src/backends/neon/NeonLayerSupport.cpp
+++ b/src/backends/neon/NeonLayerSupport.cpp
@@ -23,6 +23,7 @@
 #include "workloads/NeonBatchNormalizationWorkload.hpp"
 #include "workloads/NeonConvolution2dWorkload.hpp"
 #include "workloads/NeonDepthwiseConvolutionWorkload.hpp"
+#include "workloads/NeonDequantizeWorkload.hpp"
 #include "workloads/NeonGreaterWorkload.hpp"
 #include "workloads/NeonL2NormalizationFloatWorkload.hpp"
 #include "workloads/NeonMaximumWorkload.hpp"
@@ -245,6 +246,16 @@ bool NeonLayerSupport::IsDepthwiseConvolutionSupported(const TensorInfo& input,
                                    biases);
 }
 
+bool NeonLayerSupport::IsDequantizeSupported(const TensorInfo& input,
+                                             const TensorInfo& output,
+                                             Optional<std::string&> reasonIfUnsupported) const
+{
+    FORWARD_WORKLOAD_VALIDATE_FUNC(NeonDequantizeWorkloadValidate,
+                                   reasonIfUnsupported,
+                                   input,
+                                   output);
+}
+
 bool NeonLayerSupport::IsDilatedDepthwiseConvolutionSupported(const TensorInfo& input,
                                                               const TensorInfo& output,
                                                               const DepthwiseConvolution2dDescriptor& descriptor,
diff --git a/src/backends/neon/NeonLayerSupport.hpp b/src/backends/neon/NeonLayerSupport.hpp
index b933591a0e..1539ffea93 100644
--- a/src/backends/neon/NeonLayerSupport.hpp
+++ b/src/backends/neon/NeonLayerSupport.hpp
@@ -61,6 +61,10 @@ public:
                                          const Optional<TensorInfo>& biases,
                                          Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsDequantizeSupported(const TensorInfo& input,
+                               const TensorInfo& output,
+                               Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsDilatedDepthwiseConvolutionSupported(const TensorInfo& input,
                                                 const TensorInfo& output,
                                                 const DepthwiseConvolution2dDescriptor& descriptor,
diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp
index 307aaa56f3..86fb404637 100644
--- a/src/backends/neon/NeonWorkloadFactory.cpp
+++ b/src/backends/neon/NeonWorkloadFactory.cpp
@@ -158,6 +158,12 @@ std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDepthwiseConvolution2d(
     return std::make_unique<NeonDepthwiseConvolutionWorkload>(descriptor, info);
 }
 
+std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDequantize(const DequantizeQueueDescriptor& descriptor,
+                                                                 const WorkloadInfo& info) const
+{
+    return std::make_unique<NeonDequantizeWorkload>(descriptor, info);
+}
+
 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDetectionPostProcess(
     const armnn::DetectionPostProcessQueueDescriptor& descriptor, const armnn::WorkloadInfo& info) const
 {
diff --git a/src/backends/neon/NeonWorkloadFactory.hpp b/src/backends/neon/NeonWorkloadFactory.hpp
index c4a2b50c4e..1b1305c175 100644
--- a/src/backends/neon/NeonWorkloadFactory.hpp
+++ b/src/backends/neon/NeonWorkloadFactory.hpp
@@ -71,6 +71,9 @@ public:
     std::unique_ptr<IWorkload> CreateDepthwiseConvolution2d(const DepthwiseConvolution2dQueueDescriptor& descriptor,
                                                             const WorkloadInfo& info) const override;
 
+    std::unique_ptr<IWorkload> CreateDequantize(const DequantizeQueueDescriptor& descriptor,
+                                                const WorkloadInfo& info) const override;
+
     std::unique_ptr<IWorkload> CreateDetectionPostProcess(const DetectionPostProcessQueueDescriptor& descriptor,
                                                           const WorkloadInfo& info) const override;
 
diff --git a/src/backends/neon/backend.mk b/src/backends/neon/backend.mk
index 750118d51b..7b037d27d6 100644
--- a/src/backends/neon/backend.mk
+++ b/src/backends/neon/backend.mk
@@ -28,6 +28,7 @@ BACKEND_SOURCES := \
         workloads/NeonConvertFp32ToFp16Workload.cpp \
         workloads/NeonConvolution2dWorkload.cpp \
         workloads/NeonDepthwiseConvolutionWorkload.cpp \
+        workloads/NeonDequantizeWorkload.cpp \
         workloads/NeonFloorFloatWorkload.cpp \
         workloads/NeonFullyConnectedWorkload.cpp \
         workloads/NeonGreaterWorkload.cpp \
diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp
index 26677f6e34..08e463ada2 100644
--- a/src/backends/neon/test/NeonLayerTests.cpp
+++ b/src/backends/neon/test/NeonLayerTests.cpp
@@ -218,6 +218,10 @@ BOOST_AUTO_TEST_CASE(DepthwiseConv2dUtils)
                                                             weightsInfo3x3, biasesInfo));
 }
 
+// Dequantize
+ARMNN_AUTO_TEST_CASE(DequantizeSimpleUint8, DequantizeSimpleUint8Test)
+ARMNN_AUTO_TEST_CASE(DequantizeOffsetUint8, DequantizeOffsetUint8Test)
+
 // Pooling
 ARMNN_AUTO_TEST_CASE(SimpleMaxPooling2dSize3x3Stride2x4, SimpleMaxPooling2dSize3x3Stride2x4Test, true)
 ARMNN_AUTO_TEST_CASE(SimpleMaxPooling2dSize3x3Stride2x4Uint8, SimpleMaxPooling2dSize3x3Stride2x4Uint8Test, true)
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index b7638208d1..f6568e1f7e 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -22,6 +22,8 @@ list(APPEND armnnNeonBackendWorkloads_sources
     NeonConvolution2dWorkload.hpp
     NeonDepthwiseConvolutionWorkload.cpp
     NeonDepthwiseConvolutionWorkload.hpp
+    NeonDequantizeWorkload.cpp
+    NeonDequantizeWorkload.hpp
     NeonFloorFloatWorkload.cpp
     NeonFloorFloatWorkload.hpp
     NeonFullyConnectedWorkload.cpp
diff --git a/src/backends/neon/workloads/NeonDequantizeWorkload.cpp b/src/backends/neon/workloads/NeonDequantizeWorkload.cpp
new file mode 100644
index 0000000000..9840b48dd2
--- /dev/null
+++ b/src/backends/neon/workloads/NeonDequantizeWorkload.cpp
@@ -0,0 +1,48 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonDequantizeWorkload.hpp"
+
+#include "NeonWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+#include <backendsCommon/CpuTensorHandle.hpp>
+#include <neon/NeonTensorHandle.hpp>
+
+namespace armnn
+{
+
+using namespace armcomputetensorutils;
+
+arm_compute::Status NeonDequantizeWorkloadValidate(const TensorInfo& input,
+                                                   const TensorInfo& output)
+{
+    const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input);
+    const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output);
+
+    return arm_compute::NEDequantizationLayer::validate(&aclInput, &aclOutput);
+}
+
+NeonDequantizeWorkload::NeonDequantizeWorkload(const DequantizeQueueDescriptor& descriptor, const WorkloadInfo& info)
+        : BaseWorkload<DequantizeQueueDescriptor>(descriptor, info)
+{
+    m_Data.ValidateInputsOutputs("NeonDequantizeWorkload", 1, 1);
+
+    arm_compute::ITensor& input = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ITensor& output = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+    m_Layer.reset(new arm_compute::NEDequantizationLayer());
+    m_Layer->configure(&input, &output);
+    m_Layer->prepare();
+}
+
+void NeonDequantizeWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonDequantizeWorkload_Execute");
+    m_Layer->run();
+}
+
+} //namespace armnn
+
diff --git a/src/backends/neon/workloads/NeonDequantizeWorkload.hpp b/src/backends/neon/workloads/NeonDequantizeWorkload.hpp
new file mode 100644
index 0000000000..a6616376e7
--- /dev/null
+++ b/src/backends/neon/workloads/NeonDequantizeWorkload.hpp
@@ -0,0 +1,31 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/runtime/NEON/NEFunctions.h>
+
+#include <functional>
+
+namespace armnn
+{
+
+arm_compute::Status NeonDequantizeWorkloadValidate(const TensorInfo& input,
+                                                   const TensorInfo& output);
+
+class NeonDequantizeWorkload : public BaseWorkload<DequantizeQueueDescriptor>
+{
+public:
+    NeonDequantizeWorkload(const DequantizeQueueDescriptor& descriptor, const WorkloadInfo& info);
+
+    void Execute() const override;
+
+private:
+    mutable std::unique_ptr<arm_compute::NEDequantizationLayer> m_Layer;
+};
+
+} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index 8ad70d7095..ecd50c9d02 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -12,6 +12,7 @@
 #include "NeonConvertFp32ToFp16Workload.hpp"
 #include "NeonConvolution2dWorkload.hpp"
 #include "NeonDepthwiseConvolutionWorkload.hpp"
+#include "NeonDequantizeWorkload.hpp"
 #include "NeonFloorFloatWorkload.hpp"
 #include "NeonFullyConnectedWorkload.hpp"
 #include "NeonGreaterWorkload.hpp"