IVGCVSW-2211 Maximum operator support (CL)

   * Added CLMaximumWorkload implementation
   * Added CL unit tests

Change-Id: I922c83fe25d17be21a5d0f4e861038463cb09789

9 files changed, 118 insertions, 1 deletions

diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp
index d02dfb9938..719412bfda 100644
--- a/src/backends/cl/ClLayerSupport.cpp
+++ b/src/backends/cl/ClLayerSupport.cpp
@@ -27,6 +27,7 @@
 #include "workloads/ClFullyConnectedWorkload.hpp"
 #include "workloads/ClL2NormalizationFloatWorkload.hpp"
 #include "workloads/ClLstmFloatWorkload.hpp"
+#include "workloads/ClMaximumWorkload.hpp"
 #include "workloads/ClMeanWorkload.hpp"
 #include "workloads/ClMergerWorkload.hpp"
 #include "workloads/ClMultiplicationWorkload.hpp"
@@ -392,6 +393,18 @@ bool ClLayerSupport::IsLstmSupported(const TensorInfo& input,
                                    cellToOutputWeights);
 }
 
+bool ClLayerSupport::IsMaximumSupported(const TensorInfo& input0,
+                                        const TensorInfo& input1,
+                                        const TensorInfo& output,
+                                        Optional<std::string&> reasonIfUnsupported) const
+{
+    FORWARD_WORKLOAD_VALIDATE_FUNC(ClMaximumWorkloadValidate,
+                                   reasonIfUnsupported,
+                                   input0,
+                                   input1,
+                                   output);
+}
+
 bool ClLayerSupport::IsMeanSupported(const TensorInfo& input,
                                      const TensorInfo& output,
                                      const MeanDescriptor& descriptor,
diff --git a/src/backends/cl/ClLayerSupport.hpp b/src/backends/cl/ClLayerSupport.hpp
index d040a54507..4281dfe325 100644
--- a/src/backends/cl/ClLayerSupport.hpp
+++ b/src/backends/cl/ClLayerSupport.hpp
@@ -126,6 +126,11 @@ public:
                          const TensorInfo* cellToOutputWeights,
                          Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsMaximumSupported(const TensorInfo& input0,
+                            const TensorInfo& input1,
+                            const TensorInfo& output,
+                            Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsMeanSupported(const TensorInfo& input,
                          const TensorInfo& output,
                          const MeanDescriptor& descriptor,
diff --git a/src/backends/cl/ClWorkloadFactory.cpp b/src/backends/cl/ClWorkloadFactory.cpp
index 942defb256..37d9b31f10 100644
--- a/src/backends/cl/ClWorkloadFactory.cpp
+++ b/src/backends/cl/ClWorkloadFactory.cpp
@@ -299,7 +299,7 @@ std::unique_ptr<IWorkload> ClWorkloadFactory::CreateConvertFp32ToFp16(
 std::unique_ptr<IWorkload> ClWorkloadFactory::CreateMaximum(const MaximumQueueDescriptor& descriptor,
                                                             const WorkloadInfo& info) const
 {
-    return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info);
+    return MakeWorkload<ClMaximumWorkload>(descriptor, info);
 }
 
 std::unique_ptr<IWorkload> ClWorkloadFactory::CreateMean(const MeanQueueDescriptor& descriptor,
diff --git a/src/backends/cl/backend.mk b/src/backends/cl/backend.mk
index 71b600830a..ef56356bb0 100644
--- a/src/backends/cl/backend.mk
+++ b/src/backends/cl/backend.mk
@@ -28,6 +28,7 @@ BACKEND_SOURCES := \
         workloads/ClFullyConnectedWorkload.cpp \
         workloads/ClL2NormalizationFloatWorkload.cpp \
         workloads/ClLstmFloatWorkload.cpp \
+        workloads/ClMaximumWorkload.cpp \
         workloads/ClMeanWorkload.cpp \
         workloads/ClMergerWorkload.cpp \
         workloads/ClMultiplicationWorkload.cpp \
diff --git a/src/backends/cl/test/ClLayerTests.cpp b/src/backends/cl/test/ClLayerTests.cpp
index 90b93d70c8..b89fba61fc 100755
--- a/src/backends/cl/test/ClLayerTests.cpp
+++ b/src/backends/cl/test/ClLayerTests.cpp
@@ -336,6 +336,14 @@ ARMNN_AUTO_TEST_CASE(SimpleConvertFp32ToFp16, SimpleConvertFp32ToFp16Test)
 
 ARMNN_AUTO_TEST_CASE(AdditionAfterMaxPool, AdditionAfterMaxPoolTest)
 
+//Max
+ARMNN_AUTO_TEST_CASE(MaximumSimple, MaximumSimpleTest)
+ARMNN_AUTO_TEST_CASE(MaximumBroadcast1Element, MaximumBroadcast1ElementTest)
+ARMNN_AUTO_TEST_CASE(MaximumBroadcast1DVector, MaximumBroadcast1DVectorTest)
+ARMNN_AUTO_TEST_CASE(MaximumUint8, MaximumUint8Test)
+ARMNN_AUTO_TEST_CASE(MaximumBroadcast1ElementUint8, MaximumBroadcast1ElementUint8Test)
+ARMNN_AUTO_TEST_CASE(MaximumBroadcast1DVectorUint8, MaximumBroadcast1DVectorUint8Test)
+
 // Mean
 ARMNN_AUTO_TEST_CASE(MeanUint8Simple, MeanUint8SimpleTest)
 ARMNN_AUTO_TEST_CASE(MeanUint8SimpleAxis, MeanUint8SimpleAxisTest)
diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt
index 11a0fd2db2..d769e577dc 100644
--- a/src/backends/cl/workloads/CMakeLists.txt
+++ b/src/backends/cl/workloads/CMakeLists.txt
@@ -32,6 +32,8 @@ list(APPEND armnnClBackendWorkloads_sources
     ClL2NormalizationFloatWorkload.hpp
     ClLstmFloatWorkload.cpp
     ClLstmFloatWorkload.hpp
+    ClMaximumWorkload.cpp
+    ClMaximumWorkload.hpp
     ClMeanWorkload.cpp
     ClMeanWorkload.hpp
     ClMergerWorkload.cpp
diff --git a/src/backends/cl/workloads/ClMaximumWorkload.cpp b/src/backends/cl/workloads/ClMaximumWorkload.cpp
new file mode 100644
index 0000000000..cd3192d186
--- /dev/null
+++ b/src/backends/cl/workloads/ClMaximumWorkload.cpp
@@ -0,0 +1,58 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ClMaximumWorkload.hpp"
+
+#include "ClWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeUtils.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <backendsCommon/CpuTensorHandle.hpp>
+
+#include <cl/ClLayerSupport.hpp>
+#include <cl/ClTensorHandle.hpp>
+#include <cl/ClLayerSupport.hpp>
+
+namespace armnn
+{
+
+using namespace armcomputetensorutils;
+
+arm_compute::Status ClMaximumWorkloadValidate(const TensorInfo& input0,
+                                              const TensorInfo& input1,
+                                              const TensorInfo& output)
+{
+    const arm_compute::TensorInfo aclInput0Info = BuildArmComputeTensorInfo(input0);
+    const arm_compute::TensorInfo aclInput1Info = BuildArmComputeTensorInfo(input1);
+    const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
+
+    const arm_compute::Status aclStatus = arm_compute::CLElementwiseMax::validate(&aclInput0Info,
+                                                                                  &aclInput1Info,
+                                                                                  &aclOutputInfo);
+
+    return aclStatus;
+}
+
+ClMaximumWorkload::ClMaximumWorkload(const MaximumQueueDescriptor& descriptor,
+                                               const WorkloadInfo& info)
+    : BaseWorkload<MaximumQueueDescriptor>(descriptor, info)
+{
+    m_Data.ValidateInputsOutputs("ClMaximumWorkload", 2, 1);
+
+    arm_compute::ICLTensor& input0 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ICLTensor& input1 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
+    arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+    m_MaximumLayer.configure(&input0, &input1, &output);
+}
+
+void ClMaximumWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_CL("ClMaximumWorkload_Execute");
+    RunClFunction(m_MaximumLayer, CHECK_LOCATION());
+}
+
+} //namespace armnn
diff --git a/src/backends/cl/workloads/ClMaximumWorkload.hpp b/src/backends/cl/workloads/ClMaximumWorkload.hpp
new file mode 100644
index 0000000000..e8cc5ba787
--- /dev/null
+++ b/src/backends/cl/workloads/ClMaximumWorkload.hpp
@@ -0,0 +1,29 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/runtime/CL/CLFunctions.h>
+
+namespace armnn
+{
+
+arm_compute::Status ClMaximumWorkloadValidate(const TensorInfo& input0,
+                                              const TensorInfo& input1,
+                                              const TensorInfo& output);
+
+class ClMaximumWorkload : public BaseWorkload<MaximumQueueDescriptor>
+{
+public:
+    ClMaximumWorkload(const MaximumQueueDescriptor& descriptor, const WorkloadInfo& info);
+    void Execute() const override;
+
+private:
+    mutable arm_compute::CLElementwiseMax m_MaximumLayer;
+};
+
+} //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 9e74327717..b0b78cd9f7 100644
--- a/src/backends/cl/workloads/ClWorkloads.hpp
+++ b/src/backends/cl/workloads/ClWorkloads.hpp
@@ -17,6 +17,7 @@
 #include "ClL2NormalizationFloatWorkload.hpp"
 #include "ClLstmFloatWorkload.hpp"
 #include "ClMergerWorkload.hpp"
+#include "ClMaximumWorkload.hpp"
 #include "ClMeanWorkload.hpp"
 #include "ClMultiplicationWorkload.hpp"
 #include "ClNormalizationFloatWorkload.hpp"