IVGCVSW-3185 Add NeonQuantizeWorkload

 * Added NeonQuantizeWorkload to ArmNN

Signed-off-by: Sadik Armagan <sadik.armagan@arm.com>
Change-Id: I41f8707dda35c894841eddae2d3b78f088038c8e

10 files changed, 104 insertions, 6 deletions

diff --git a/src/backends/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp
index e84eb799fc..a85597bbf8 100644
--- a/src/backends/neon/NeonLayerSupport.cpp
+++ b/src/backends/neon/NeonLayerSupport.cpp
@@ -35,6 +35,7 @@
 #include "workloads/NeonPadWorkload.hpp"
 #include "workloads/NeonPermuteWorkload.hpp"
 #include "workloads/NeonPooling2dWorkload.hpp"
+#include "workloads/NeonQuantizeWorkload.hpp"
 #include "workloads/NeonResizeBilinearWorkload.hpp"
 #include "workloads/NeonSoftmaxBaseWorkload.hpp"
 #include "workloads/NeonSplitterWorkload.hpp"
@@ -438,6 +439,16 @@ bool NeonLayerSupport::IsPooling2dSupported(const TensorInfo& input,
     FORWARD_WORKLOAD_VALIDATE_FUNC(NeonPooling2dWorkloadValidate, reasonIfUnsupported, input, output, descriptor);
 }
 
+bool NeonLayerSupport::IsQuantizeSupported(const TensorInfo& input,
+                                           const TensorInfo& output,
+                                           Optional<std::string&> reasonIfUnsupported) const
+{
+    FORWARD_WORKLOAD_VALIDATE_FUNC(NeonQuantizeWorkloadValidate,
+                                   reasonIfUnsupported,
+                                   input,
+                                   output);
+}
+
 bool NeonLayerSupport::IsReshapeSupported(const TensorInfo& input,
                                           const ReshapeDescriptor& descriptor,
                                           Optional<std::string&> reasonIfUnsupported) const
diff --git a/src/backends/neon/NeonLayerSupport.hpp b/src/backends/neon/NeonLayerSupport.hpp
index dd6ed79c9a..b933591a0e 100644
--- a/src/backends/neon/NeonLayerSupport.hpp
+++ b/src/backends/neon/NeonLayerSupport.hpp
@@ -145,6 +145,10 @@ public:
                               const Pooling2dDescriptor& descriptor,
                               Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsQuantizeSupported(const TensorInfo& input,
+                             const TensorInfo& output,
+                             Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsReshapeSupported(const TensorInfo& input,
                             const ReshapeDescriptor& descriptor,
                             Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp
index 4b6225f67b..307aaa56f3 100644
--- a/src/backends/neon/NeonWorkloadFactory.cpp
+++ b/src/backends/neon/NeonWorkloadFactory.cpp
@@ -219,11 +219,10 @@ std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateResizeBilinear(
     return std::make_unique<NeonResizeBilinearWorkload>(descriptor, info);
 }
 
-std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateFakeQuantization(
-    const FakeQuantizationQueueDescriptor& descriptor,
-    const WorkloadInfo& info) const
+std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateQuantize(const QuantizeQueueDescriptor& descriptor,
+                                                                      const WorkloadInfo& info) const
 {
-    return nullptr;
+    return std::make_unique<NeonQuantizeWorkload>(descriptor, info);
 }
 
 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateL2Normalization(const L2NormalizationQueueDescriptor& descriptor,
diff --git a/src/backends/neon/NeonWorkloadFactory.hpp b/src/backends/neon/NeonWorkloadFactory.hpp
index 6a28d12326..c4a2b50c4e 100644
--- a/src/backends/neon/NeonWorkloadFactory.hpp
+++ b/src/backends/neon/NeonWorkloadFactory.hpp
@@ -92,8 +92,8 @@ public:
     std::unique_ptr<IWorkload> CreateResizeBilinear(const ResizeBilinearQueueDescriptor& descriptor,
                                                     const WorkloadInfo& info) const override;
 
-    std::unique_ptr<IWorkload> CreateFakeQuantization(const FakeQuantizationQueueDescriptor& descriptor,
-                                                      const WorkloadInfo& info) const override;
+    std::unique_ptr<IWorkload> CreateQuantize(const QuantizeQueueDescriptor& descriptor,
+                                              const WorkloadInfo& info) const override;
 
     std::unique_ptr<IWorkload> CreateL2Normalization(const L2NormalizationQueueDescriptor& descriptor,
                                                      const WorkloadInfo& info) const override;
diff --git a/src/backends/neon/backend.mk b/src/backends/neon/backend.mk
index 6931bd7325..750118d51b 100644
--- a/src/backends/neon/backend.mk
+++ b/src/backends/neon/backend.mk
@@ -41,6 +41,7 @@ BACKEND_SOURCES := \
         workloads/NeonPadWorkload.cpp \
         workloads/NeonPermuteWorkload.cpp \
         workloads/NeonPooling2dWorkload.cpp \
+        workloads/NeonQuantizeWorkload.cpp \
         workloads/NeonReshapeWorkload.cpp \
         workloads/NeonResizeBilinearWorkload.cpp \
         workloads/NeonSoftmaxBaseWorkload.cpp \
diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp
index d89011d3ae..26677f6e34 100644
--- a/src/backends/neon/test/NeonLayerTests.cpp
+++ b/src/backends/neon/test/NeonLayerTests.cpp
@@ -486,6 +486,10 @@ ARMNN_AUTO_TEST_CASE(ResizeBilinearSqMinNhwc, ResizeBilinearSqMinTest, armnn::Da
 ARMNN_AUTO_TEST_CASE(ResizeBilinearMinNhwc, ResizeBilinearMinTest, armnn::DataLayout::NHWC)
 ARMNN_AUTO_TEST_CASE(ResizeBilinearMagNhwc, ResizeBilinearMagTest, armnn::DataLayout::NHWC)
 
+// Quantize
+ARMNN_AUTO_TEST_CASE(QuantizeSimpleUint8, QuantizeSimpleUint8Test)
+ARMNN_AUTO_TEST_CASE(QuantizeClampUint8, QuantizeClampUint8Test)
+
 // ============================================================================
 // COMPARE tests
 
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index 8b2ad63f45..b7638208d1 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -48,6 +48,8 @@ list(APPEND armnnNeonBackendWorkloads_sources
     NeonPermuteWorkload.hpp
     NeonPooling2dWorkload.cpp
     NeonPooling2dWorkload.hpp
+    NeonQuantizeWorkload.cpp
+    NeonQuantizeWorkload.hpp
     NeonReshapeWorkload.cpp
     NeonReshapeWorkload.hpp
     NeonResizeBilinearWorkload.cpp
diff --git a/src/backends/neon/workloads/NeonQuantizeWorkload.cpp b/src/backends/neon/workloads/NeonQuantizeWorkload.cpp
new file mode 100644
index 0000000000..ef24a7f40b
--- /dev/null
+++ b/src/backends/neon/workloads/NeonQuantizeWorkload.cpp
@@ -0,0 +1,50 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonQuantizeWorkload.hpp"
+#include "NeonWorkloadUtils.hpp"
+
+#include <neon/NeonTensorHandle.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+#include <arm_compute/core/Types.h>
+
+#include <boost/polymorphic_pointer_cast.hpp>
+
+namespace armnn
+{
+using namespace armcomputetensorutils;
+
+arm_compute::Status NeonQuantizeWorkloadValidate(const TensorInfo& input, const TensorInfo& output)
+{
+    const arm_compute::TensorInfo neonInputInfo  = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+    const arm_compute::TensorInfo neonOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+    return arm_compute::NEQuantizationLayer::validate(&neonInputInfo, &neonOutputInfo);
+}
+
+NeonQuantizeWorkload::NeonQuantizeWorkload(const QuantizeQueueDescriptor& descriptor,
+     const WorkloadInfo& workloadInfo)
+     : BaseWorkload<QuantizeQueueDescriptor>(descriptor, workloadInfo)
+{
+    arm_compute::ITensor& input = boost::polymorphic_pointer_downcast<INeonTensorHandle>(
+                                                                      m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ITensor& output = boost::polymorphic_pointer_downcast<INeonTensorHandle>(
+                                                                       m_Data.m_Outputs[0])->GetTensor();
+
+    m_Layer.reset(new arm_compute::NEQuantizationLayer());
+    m_Layer->configure(&input, &output);
+    m_Layer->prepare();
+}
+
+void NeonQuantizeWorkload::Execute() const
+{
+    if (m_Layer)
+    {
+        ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonQuantizeWorkload_Execute");
+        m_Layer->run();
+    }
+}
+
+} // namespace armnn
diff --git a/src/backends/neon/workloads/NeonQuantizeWorkload.hpp b/src/backends/neon/workloads/NeonQuantizeWorkload.hpp
new file mode 100644
index 0000000000..7f5a145371
--- /dev/null
+++ b/src/backends/neon/workloads/NeonQuantizeWorkload.hpp
@@ -0,0 +1,26 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/runtime/NEON/functions/NEQuantizationLayer.h>
+
+namespace armnn {
+
+arm_compute::Status NeonQuantizeWorkloadValidate(const TensorInfo& input, const TensorInfo& output);
+
+class NeonQuantizeWorkload : public BaseWorkload<QuantizeQueueDescriptor>
+{
+public:
+    NeonQuantizeWorkload(const QuantizeQueueDescriptor& descriptor, const WorkloadInfo& workloadInfo);
+    void Execute() const override;
+
+private:
+    mutable std::unique_ptr<arm_compute::NEQuantizationLayer> m_Layer;
+};
+
+} // namespace armnn
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index dc02aecfd7..8ad70d7095 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -26,6 +26,7 @@
 #include "NeonPadWorkload.hpp"
 #include "NeonPermuteWorkload.hpp"
 #include "NeonPooling2dWorkload.hpp"
+#include "NeonQuantizeWorkload.hpp"
 #include "NeonReshapeWorkload.hpp"
 #include "NeonResizeBilinearWorkload.hpp"
 #include "NeonSoftmaxFloatWorkload.hpp"