IVGCVSW-3147 Add CL Dequantization workload

Change-Id: I9baf6af3d98a26005a31075cd9c4e1f40938789b
Signed-off-by: Jim Flynn <jim.flynn@arm.com>

10 files changed, 118 insertions, 0 deletions

diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp
index 48953026a1..5582799ff6 100644
--- a/src/backends/cl/ClLayerSupport.cpp
+++ b/src/backends/cl/ClLayerSupport.cpp
@@ -23,6 +23,7 @@
 #include "workloads/ClConvertFp16ToFp32Workload.hpp"
 #include "workloads/ClConvertFp32ToFp16Workload.hpp"
 #include "workloads/ClConvolution2dWorkload.hpp"
+#include "workloads/ClDequantizeWorkload.hpp"
 #include "workloads/ClDepthwiseConvolutionWorkload.hpp"
 #include "workloads/ClDivisionFloatWorkload.hpp"
 #include "workloads/ClFullyConnectedWorkload.hpp"
@@ -274,6 +275,16 @@ bool ClLayerSupport::IsConvolution2dSupported(const TensorInfo& input,
                                    biases);
 }
 
+bool ClLayerSupport::IsDequantizeSupported(const TensorInfo& input,
+                                           const TensorInfo& output,
+                                           Optional<std::string&> reasonIfUnsupported) const
+{
+    FORWARD_WORKLOAD_VALIDATE_FUNC(ClDequantizeWorkloadValidate,
+                                   reasonIfUnsupported,
+                                   input,
+                                   output);
+}
+
 bool ClLayerSupport::IsDepthwiseConvolutionSupported(const TensorInfo& input,
                                                      const TensorInfo& output,
                                                      const DepthwiseConvolution2dDescriptor& descriptor,
diff --git a/src/backends/cl/ClLayerSupport.hpp b/src/backends/cl/ClLayerSupport.hpp
index f8a9a96f0f..35c00f9176 100644
--- a/src/backends/cl/ClLayerSupport.hpp
+++ b/src/backends/cl/ClLayerSupport.hpp
@@ -59,6 +59,10 @@ 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 IsDepthwiseConvolutionSupported(const TensorInfo& input,
                                          const TensorInfo& output,
                                          const DepthwiseConvolution2dDescriptor& descriptor,
diff --git a/src/backends/cl/ClWorkloadFactory.cpp b/src/backends/cl/ClWorkloadFactory.cpp
index a1f2aff4d9..70a4ddad4d 100644
--- a/src/backends/cl/ClWorkloadFactory.cpp
+++ b/src/backends/cl/ClWorkloadFactory.cpp
@@ -198,6 +198,12 @@ std::unique_ptr<IWorkload> ClWorkloadFactory::CreateDetectionPostProcess(
     return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info);
 }
 
+std::unique_ptr<IWorkload> ClWorkloadFactory::CreateDequantize(const DequantizeQueueDescriptor& descriptor,
+                                                               const WorkloadInfo& info) const
+{
+    return std::make_unique<ClDequantizeWorkload>(descriptor, info);
+}
+
 std::unique_ptr<armnn::IWorkload> ClWorkloadFactory::CreateNormalization(const NormalizationQueueDescriptor& descriptor,
                                                                          const WorkloadInfo&                 info) const
 {
diff --git a/src/backends/cl/ClWorkloadFactory.hpp b/src/backends/cl/ClWorkloadFactory.hpp
index c1216cbac0..c84ce13b8b 100644
--- a/src/backends/cl/ClWorkloadFactory.hpp
+++ b/src/backends/cl/ClWorkloadFactory.hpp
@@ -73,6 +73,9 @@ public:
     std::unique_ptr<IWorkload> CreateDetectionPostProcess(const DetectionPostProcessQueueDescriptor& descriptor,
                                                           const WorkloadInfo& info) const override;
 
+    std::unique_ptr<IWorkload> CreateDequantize(const DequantizeQueueDescriptor& descriptor,
+                                                const WorkloadInfo& info) const override;
+
     std::unique_ptr<IWorkload> CreateNormalization(const NormalizationQueueDescriptor& descriptor,
                                                    const WorkloadInfo& info) const override;
 
diff --git a/src/backends/cl/backend.mk b/src/backends/cl/backend.mk
index f5994c74ad..9bd37cc483 100644
--- a/src/backends/cl/backend.mk
+++ b/src/backends/cl/backend.mk
@@ -30,6 +30,7 @@ BACKEND_SOURCES := \
         workloads/ClConvertFp32ToFp16Workload.cpp \
         workloads/ClConvolution2dWorkload.cpp \
         workloads/ClDepthwiseConvolutionWorkload.cpp \
+        workloads/ClDequantizeWorkload.cpp \
         workloads/ClDivisionFloatWorkload.cpp \
         workloads/ClFloorFloatWorkload.cpp \
         workloads/ClFullyConnectedWorkload.cpp \
diff --git a/src/backends/cl/test/ClLayerTests.cpp b/src/backends/cl/test/ClLayerTests.cpp
index ad87e953c8..76b8175f25 100644
--- a/src/backends/cl/test/ClLayerTests.cpp
+++ b/src/backends/cl/test/ClLayerTests.cpp
@@ -436,6 +436,13 @@ ARMNN_AUTO_TEST_CASE(StridedSlice2DReverseUint8, StridedSlice2DReverseUint8Test)
 ARMNN_AUTO_TEST_CASE(QuantizeSimpleUint8, QuantizeSimpleUint8Test)
 ARMNN_AUTO_TEST_CASE(QuantizeClampUint8, QuantizeClampUint8Test)
 
+// Dequantize
+// NOTE: current clframework (46a49a0a8206f0efa7afd514940e180a88ffd732)
+//       CLDequantizationLayerKernel accepts DataType::QASYMM8 input
+//       and can output DataType::F16 or DataType::F32
+ARMNN_AUTO_TEST_CASE(DequantizeSimpleUint8, DequantizeSimpleUint8Test)
+ARMNN_AUTO_TEST_CASE(DequantizeOffsetUint8, DequantizeOffsetUint8Test)
+
 // ============================================================================
 // COMPARE tests
 
diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt
index 23668c564b..c055244b1f 100644
--- a/src/backends/cl/workloads/CMakeLists.txt
+++ b/src/backends/cl/workloads/CMakeLists.txt
@@ -24,6 +24,8 @@ list(APPEND armnnClBackendWorkloads_sources
     ClConvolution2dWorkload.hpp
     ClDepthwiseConvolutionWorkload.cpp
     ClDepthwiseConvolutionWorkload.hpp
+    ClDequantizeWorkload.cpp
+    ClDequantizeWorkload.hpp
     ClDivisionFloatWorkload.cpp
     ClDivisionFloatWorkload.hpp
     ClFloorFloatWorkload.cpp
diff --git a/src/backends/cl/workloads/ClDequantizeWorkload.cpp b/src/backends/cl/workloads/ClDequantizeWorkload.cpp
new file mode 100644
index 0000000000..67a555a020
--- /dev/null
+++ b/src/backends/cl/workloads/ClDequantizeWorkload.cpp
@@ -0,0 +1,55 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ClDequantizeWorkload.hpp"
+#include "ClWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+#include <backendsCommon/CpuTensorHandle.hpp>
+
+#include <arm_compute/core/Types.h>
+
+#include <cl/ClLayerSupport.hpp>
+#include <cl/ClTensorHandle.hpp>
+
+#include <boost/polymorphic_pointer_cast.hpp>
+
+namespace armnn
+{
+using namespace armcomputetensorutils;
+
+arm_compute::Status ClDequantizeWorkloadValidate(const TensorInfo& input, const TensorInfo& output)
+{
+    const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input);
+    const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
+
+    return arm_compute::CLDequantizationLayer::validate(&aclInputInfo, &aclOutputInfo);
+}
+
+ClDequantizeWorkload::ClDequantizeWorkload(const DequantizeQueueDescriptor& descriptor,
+                                           const WorkloadInfo& workloadInfo)
+                                           : BaseWorkload<DequantizeQueueDescriptor>(descriptor, workloadInfo)
+{
+    arm_compute::ICLTensor& input = boost::polymorphic_pointer_downcast<IClTensorHandle>(
+            m_Data.m_Inputs[0])->GetTensor();
+
+    arm_compute::ICLTensor& output = boost::polymorphic_pointer_downcast<IClTensorHandle>(
+            m_Data.m_Outputs[0])->GetTensor();
+
+    m_Layer.reset(new arm_compute::CLDequantizationLayer());
+    m_Layer->configure(&input, &output);
+    m_Layer->prepare();
+}
+
+void ClDequantizeWorkload::Execute() const
+{
+    if (m_Layer)
+    {
+        ARMNN_SCOPED_PROFILING_EVENT_CL("ClDequantizeWorkload_Execute");
+        m_Layer->run();
+    }
+}
+
+} // namespace armnn
diff --git a/src/backends/cl/workloads/ClDequantizeWorkload.hpp b/src/backends/cl/workloads/ClDequantizeWorkload.hpp
new file mode 100644
index 0000000000..6e61da2ebf
--- /dev/null
+++ b/src/backends/cl/workloads/ClDequantizeWorkload.hpp
@@ -0,0 +1,28 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/runtime/CL/functions/CLDequantizationLayer.h>
+
+namespace armnn
+{
+
+arm_compute::Status ClDequantizeWorkloadValidate(const TensorInfo& input, const TensorInfo& output);
+
+class ClDequantizeWorkload : public BaseWorkload<DequantizeQueueDescriptor>
+{
+public:
+    ClDequantizeWorkload(const DequantizeQueueDescriptor& descriptor, const WorkloadInfo& workloadInfo);
+
+    void Execute() const override;
+
+private:
+    mutable std::unique_ptr<arm_compute::CLDequantizationLayer> m_Layer;
+};
+
+} // namespace armnn
\ No newline at end of file
diff --git a/src/backends/cl/workloads/ClWorkloads.hpp b/src/backends/cl/workloads/ClWorkloads.hpp
index 0060412770..b836781166 100644
--- a/src/backends/cl/workloads/ClWorkloads.hpp
+++ b/src/backends/cl/workloads/ClWorkloads.hpp
@@ -11,6 +11,7 @@
 #include "ClBatchToSpaceNdWorkload.hpp"
 #include "ClConvolution2dWorkload.hpp"
 #include "ClDepthwiseConvolutionWorkload.hpp"
+#include "ClDequantizeWorkload.hpp"
 #include "ClDivisionFloatWorkload.hpp"
 #include "ClFloorFloatWorkload.hpp"
 #include "ClFullyConnectedWorkload.hpp"