IVGCVSW-3732 Add CL workload for DepthToSpace

Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>
Change-Id: I577c45099eda1d04214e7a1f75fa10d1352efc80

10 files changed, 170 insertions, 1 deletions

diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp
index 5905cd990e..450391581e 100644
--- a/src/backends/cl/ClLayerSupport.cpp
+++ b/src/backends/cl/ClLayerSupport.cpp
@@ -24,8 +24,9 @@
 #include "workloads/ClConvertFp16ToFp32Workload.hpp"
 #include "workloads/ClConvertFp32ToFp16Workload.hpp"
 #include "workloads/ClConvolution2dWorkload.hpp"
-#include "workloads/ClDequantizeWorkload.hpp"
+#include "workloads/ClDepthToSpaceWorkload.hpp"
 #include "workloads/ClDepthwiseConvolutionWorkload.hpp"
+#include "workloads/ClDequantizeWorkload.hpp"
 #include "workloads/ClDivisionFloatWorkload.hpp"
 #include "workloads/ClFullyConnectedWorkload.hpp"
 #include "workloads/ClGreaterWorkload.hpp"
@@ -303,6 +304,18 @@ bool ClLayerSupport::IsDequantizeSupported(const TensorInfo& input,
                                    output);
 }
 
+bool ClLayerSupport::IsDepthToSpaceSupported(const TensorInfo& input,
+                                             const TensorInfo& output,
+                                             const DepthToSpaceDescriptor& descriptor,
+                                             Optional<std::string&> reasonIfUnsupported) const
+{
+    FORWARD_WORKLOAD_VALIDATE_FUNC(ClDepthToSpaceWorkloadValidate,
+                                   reasonIfUnsupported,
+                                   input,
+                                   output,
+                                   descriptor);
+}
+
 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 ddc428de07..1a37315d1d 100644
--- a/src/backends/cl/ClLayerSupport.hpp
+++ b/src/backends/cl/ClLayerSupport.hpp
@@ -67,6 +67,11 @@ public:
                                const TensorInfo& output,
                                Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsDepthToSpaceSupported(const TensorInfo& input,
+                                 const TensorInfo& output,
+                                 const DepthToSpaceDescriptor& descriptor,
+                                 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 536d4dd058..ea3c27ebf2 100644
--- a/src/backends/cl/ClWorkloadFactory.cpp
+++ b/src/backends/cl/ClWorkloadFactory.cpp
@@ -463,4 +463,10 @@ std::unique_ptr<IWorkload> ClWorkloadFactory::CreateStack(const StackQueueDescri
     return MakeWorkload<ClStackWorkload>(descriptor, info);
 }
 
+std::unique_ptr<IWorkload> ClWorkloadFactory::CreateDepthToSpace(const DepthToSpaceQueueDescriptor& descriptor,
+                                                                 const WorkloadInfo& info) const
+{
+    return MakeWorkload<ClDepthToSpaceWorkload>(descriptor, info);
+}
+
 } // namespace armnn
diff --git a/src/backends/cl/ClWorkloadFactory.hpp b/src/backends/cl/ClWorkloadFactory.hpp
index c8d58dbb5c..7ea7f261a1 100644
--- a/src/backends/cl/ClWorkloadFactory.hpp
+++ b/src/backends/cl/ClWorkloadFactory.hpp
@@ -75,6 +75,9 @@ public:
     std::unique_ptr<IWorkload> CreateConvolution2d(const Convolution2dQueueDescriptor& descriptor,
                                                    const WorkloadInfo& info) const override;
 
+    std::unique_ptr<IWorkload> CreateDepthToSpace(const DepthToSpaceQueueDescriptor& descriptor,
+                                                  const WorkloadInfo& info) const override;
+
     std::unique_ptr<IWorkload> CreateDepthwiseConvolution2d(const DepthwiseConvolution2dQueueDescriptor& descriptor,
                                                             const WorkloadInfo& info) const override;
 
diff --git a/src/backends/cl/backend.mk b/src/backends/cl/backend.mk
index c417c4f96b..20ece76953 100644
--- a/src/backends/cl/backend.mk
+++ b/src/backends/cl/backend.mk
@@ -32,6 +32,7 @@ BACKEND_SOURCES := \
         workloads/ClConvertFp16ToFp32Workload.cpp \
         workloads/ClConvertFp32ToFp16Workload.cpp \
         workloads/ClConvolution2dWorkload.cpp \
+        workloads/ClDepthToSpaceWorkload.cpp \
         workloads/ClDepthwiseConvolutionWorkload.cpp \
         workloads/ClDequantizeWorkload.cpp \
         workloads/ClDivisionFloatWorkload.cpp \
diff --git a/src/backends/cl/test/ClLayerTests.cpp b/src/backends/cl/test/ClLayerTests.cpp
index 6a26a3d144..76e5061ce2 100644
--- a/src/backends/cl/test/ClLayerTests.cpp
+++ b/src/backends/cl/test/ClLayerTests.cpp
@@ -356,6 +356,47 @@ ARMNN_AUTO_TEST_CASE(Concat4dDiffShapeDim0Uint8, Concat4dDiffShapeDim0Uint8Test)
 ARMNN_AUTO_TEST_CASE(Concat4dDiffShapeDim1Uint8, Concat4dDiffShapeDim1Uint8Test)
 ARMNN_AUTO_TEST_CASE(Concat4dDiffShapeDim3Uint8, Concat4dDiffShapeDim3Uint8Test, false)
 
+// DepthToSpace
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat32_1, DepthToSpaceTest1<DataType::Float32>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat32_2, DepthToSpaceTest2<DataType::Float32>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat32_3, DepthToSpaceTest3<DataType::Float32>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat32_4, DepthToSpaceTest4<DataType::Float32>, DataLayout::NCHW);
+
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat16_1, DepthToSpaceTest1<DataType::Float16>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat16_2, DepthToSpaceTest2<DataType::Float16>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat16_3, DepthToSpaceTest3<DataType::Float16>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat16_4, DepthToSpaceTest4<DataType::Float16>, DataLayout::NCHW);
+
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwUint8_1, DepthToSpaceTest1<DataType::QuantisedAsymm8>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwUint8_2, DepthToSpaceTest2<DataType::QuantisedAsymm8>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwUint8_3, DepthToSpaceTest3<DataType::QuantisedAsymm8>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwUint8_4, DepthToSpaceTest4<DataType::QuantisedAsymm8>, DataLayout::NCHW);
+
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwInt16_1, DepthToSpaceTest1<DataType::QuantisedSymm16>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwInt16_2, DepthToSpaceTest2<DataType::QuantisedSymm16>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwInt16_3, DepthToSpaceTest3<DataType::QuantisedSymm16>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwInt16_4, DepthToSpaceTest4<DataType::QuantisedSymm16>, DataLayout::NCHW);
+
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcFloat32_1, DepthToSpaceTest1<DataType::Float32>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcFloat32_2, DepthToSpaceTest2<DataType::Float32>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcFloat32_3, DepthToSpaceTest3<DataType::Float32>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcFloat32_4, DepthToSpaceTest4<DataType::Float32>, DataLayout::NHWC);
+
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcFloat16_1, DepthToSpaceTest1<DataType::Float16>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcFloat16_2, DepthToSpaceTest2<DataType::Float16>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcFloat16_3, DepthToSpaceTest3<DataType::Float16>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcFloat16_4, DepthToSpaceTest4<DataType::Float16>, DataLayout::NHWC);
+
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcUint8_1, DepthToSpaceTest1<DataType::QuantisedAsymm8>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcUint8_2, DepthToSpaceTest2<DataType::QuantisedAsymm8>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcUint8_3, DepthToSpaceTest3<DataType::QuantisedAsymm8>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcUint8_4, DepthToSpaceTest4<DataType::QuantisedAsymm8>, DataLayout::NHWC);
+
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcInt16_1, DepthToSpaceTest1<DataType::QuantisedSymm16>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcInt16_2, DepthToSpaceTest2<DataType::QuantisedSymm16>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcInt16_3, DepthToSpaceTest3<DataType::QuantisedSymm16>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcInt16_4, DepthToSpaceTest4<DataType::QuantisedSymm16>, DataLayout::NHWC);
+
 // Floor
 ARMNN_AUTO_TEST_CASE(SimpleFloor, SimpleFloorTest<DataType::Float32>)
 
diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt
index 7578a4029d..affe869c89 100644
--- a/src/backends/cl/workloads/CMakeLists.txt
+++ b/src/backends/cl/workloads/CMakeLists.txt
@@ -24,6 +24,8 @@ list(APPEND armnnClBackendWorkloads_sources
     ClConvertFp32ToFp16Workload.hpp
     ClConvolution2dWorkload.cpp
     ClConvolution2dWorkload.hpp
+    ClDepthToSpaceWorkload.cpp
+    ClDepthToSpaceWorkload.hpp
     ClDepthwiseConvolutionWorkload.cpp
     ClDepthwiseConvolutionWorkload.hpp
     ClDequantizeWorkload.cpp
diff --git a/src/backends/cl/workloads/ClDepthToSpaceWorkload.cpp b/src/backends/cl/workloads/ClDepthToSpaceWorkload.cpp
new file mode 100644
index 0000000000..800a98409e
--- /dev/null
+++ b/src/backends/cl/workloads/ClDepthToSpaceWorkload.cpp
@@ -0,0 +1,66 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ClDepthToSpaceWorkload.hpp"
+
+#include "ClWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <cl/ClTensorHandle.hpp>
+
+#include <boost/numeric/conversion/cast.hpp>
+#include <boost/polymorphic_pointer_cast.hpp>
+
+namespace armnn
+{
+
+using namespace armcomputetensorutils;
+
+arm_compute::Status ClDepthToSpaceWorkloadValidate(const TensorInfo& input,
+                                                   const TensorInfo& output,
+                                                   const DepthToSpaceDescriptor& desc)
+{
+    DataLayout dataLayout = desc.m_DataLayout;
+    const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, dataLayout);
+
+    int32_t blockSize = boost::numeric_cast<int32_t>(desc.m_BlockSize);
+
+    const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, dataLayout);
+
+    const arm_compute::Status aclStatus = arm_compute::CLDepthToSpaceLayer::validate(&aclInputInfo,
+                                                                                     &aclOutputInfo,
+                                                                                     blockSize);
+    return aclStatus;
+}
+
+ClDepthToSpaceWorkload::ClDepthToSpaceWorkload(const DepthToSpaceQueueDescriptor& desc,
+                                               const WorkloadInfo& info)
+    : BaseWorkload<DepthToSpaceQueueDescriptor>(desc, info)
+{
+    m_Data.ValidateInputsOutputs("ClDepthToSpaceWorkload", 1, 1);
+
+    arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
+
+    arm_compute::ICLTensor& input =
+        boost::polymorphic_pointer_downcast<IClTensorHandle>(m_Data.m_Inputs[0])->GetTensor();
+    input.info()->set_data_layout(aclDataLayout);
+
+    int32_t blockSize = boost::numeric_cast<int32_t>(desc.m_Parameters.m_BlockSize);
+
+    arm_compute::ICLTensor& output =
+        boost::polymorphic_pointer_downcast<IClTensorHandle>(m_Data.m_Outputs[0])->GetTensor();
+    output.info()->set_data_layout(aclDataLayout);
+
+    m_Layer.configure(&input, &output, blockSize);
+}
+
+void ClDepthToSpaceWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_CL("ClDepthToSpaceWorkload_Execute");
+    RunClFunction(m_Layer, CHECK_LOCATION());
+}
+
+} // namespace armnn
diff --git a/src/backends/cl/workloads/ClDepthToSpaceWorkload.hpp b/src/backends/cl/workloads/ClDepthToSpaceWorkload.hpp
new file mode 100644
index 0000000000..de8b496669
--- /dev/null
+++ b/src/backends/cl/workloads/ClDepthToSpaceWorkload.hpp
@@ -0,0 +1,31 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <armnn/Descriptors.hpp>
+
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/runtime/CL/functions/CLDepthToSpaceLayer.h>
+
+namespace armnn
+{
+
+arm_compute::Status ClDepthToSpaceWorkloadValidate(const TensorInfo& input,
+                                                   const TensorInfo& output,
+                                                   const DepthToSpaceDescriptor& desc);
+
+class ClDepthToSpaceWorkload : public BaseWorkload<DepthToSpaceQueueDescriptor>
+{
+public:
+    ClDepthToSpaceWorkload(const DepthToSpaceQueueDescriptor& descriptor, const WorkloadInfo& info);
+    void Execute() const override;
+
+private:
+    mutable arm_compute::CLDepthToSpaceLayer m_Layer;
+};
+
+} // namespace armnn
diff --git a/src/backends/cl/workloads/ClWorkloads.hpp b/src/backends/cl/workloads/ClWorkloads.hpp
index 7525f84319..5c42e764b8 100644
--- a/src/backends/cl/workloads/ClWorkloads.hpp
+++ b/src/backends/cl/workloads/ClWorkloads.hpp
@@ -11,6 +11,7 @@
 #include "ClBatchNormalizationFloatWorkload.hpp"
 #include "ClBatchToSpaceNdWorkload.hpp"
 #include "ClConvolution2dWorkload.hpp"
+#include "ClDepthToSpaceWorkload.hpp"
 #include "ClDepthwiseConvolutionWorkload.hpp"
 #include "ClDequantizeWorkload.hpp"
 #include "ClDivisionFloatWorkload.hpp"