IVGCVSW-3293 Add CL backend support for PReLU

Change-Id: I1bb187db89bb3eb883b8f0aca4c3439d82b56583
Signed-off-by: Nikhil Raj <nikhil.raj@arm.com>

11 files changed, 149 insertions, 0 deletions

diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp
index d79f6126a7..a6baa3c38f 100644
--- a/src/backends/cl/ClLayerSupport.cpp
+++ b/src/backends/cl/ClLayerSupport.cpp
@@ -39,6 +39,7 @@
 #include "workloads/ClPadWorkload.hpp"
 #include "workloads/ClPermuteWorkload.hpp"
 #include "workloads/ClPooling2dWorkload.hpp"
+#include "workloads/ClPreluWorkload.hpp"
 #include "workloads/ClQuantizeWorkload.hpp"
 #include "workloads/ClSoftmaxBaseWorkload.hpp"
 #include "workloads/ClSpaceToBatchNdWorkload.hpp"
@@ -571,6 +572,14 @@ bool ClLayerSupport::IsPooling2dSupported(const TensorInfo& input,
     FORWARD_WORKLOAD_VALIDATE_FUNC(ClPooling2dWorkloadValidate, reasonIfUnsupported, input, output, descriptor);
 }
 
+bool ClLayerSupport::IsPreluSupported(const armnn::TensorInfo &input,
+                                      const armnn::TensorInfo &alpha,
+                                      const armnn::TensorInfo &output,
+                                      armnn::Optional<std::string &> reasonIfUnsupported) const
+{
+    FORWARD_WORKLOAD_VALIDATE_FUNC(ClPreluWorkloadValidate, reasonIfUnsupported, input, alpha, output);
+}
+
 bool ClLayerSupport::IsQuantizeSupported(const TensorInfo& input,
                                          const TensorInfo& output,
                                          Optional<std::string&> reasonIfUnsupported) const
diff --git a/src/backends/cl/ClLayerSupport.hpp b/src/backends/cl/ClLayerSupport.hpp
index 1461f41691..f77889bef6 100644
--- a/src/backends/cl/ClLayerSupport.hpp
+++ b/src/backends/cl/ClLayerSupport.hpp
@@ -190,6 +190,11 @@ public:
                               const Pooling2dDescriptor& descriptor,
                               Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsPreluSupported(const TensorInfo& input,
+                          const TensorInfo& alpha,
+                          const TensorInfo& output,
+                          Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsQuantizeSupported(const TensorInfo& input,
                              const TensorInfo& output,
                              Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
diff --git a/src/backends/cl/ClWorkloadFactory.cpp b/src/backends/cl/ClWorkloadFactory.cpp
index c662a9db29..307f954bae 100644
--- a/src/backends/cl/ClWorkloadFactory.cpp
+++ b/src/backends/cl/ClWorkloadFactory.cpp
@@ -180,6 +180,12 @@ std::unique_ptr<armnn::IWorkload> ClWorkloadFactory::CreatePooling2d(const Pooli
     return MakeWorkload<ClPooling2dWorkload>(descriptor, info);
 }
 
+std::unique_ptr<armnn::IWorkload> ClWorkloadFactory::CreatePrelu(const armnn::PreluQueueDescriptor &descriptor,
+                                                                 const armnn::WorkloadInfo &info) const
+{
+    return MakeWorkload<ClPreluWorkload>(descriptor, info);
+}
+
 std::unique_ptr<armnn::IWorkload> ClWorkloadFactory::CreateConvolution2d(const Convolution2dQueueDescriptor& descriptor,
                                                                          const WorkloadInfo&               info) const
 {
diff --git a/src/backends/cl/ClWorkloadFactory.hpp b/src/backends/cl/ClWorkloadFactory.hpp
index 32925f7f95..5448504737 100644
--- a/src/backends/cl/ClWorkloadFactory.hpp
+++ b/src/backends/cl/ClWorkloadFactory.hpp
@@ -64,6 +64,9 @@ public:
     std::unique_ptr<IWorkload> CreatePooling2d(const Pooling2dQueueDescriptor& descriptor,
                                                const WorkloadInfo& info) const override;
 
+    std::unique_ptr<IWorkload> CreatePrelu(const PreluQueueDescriptor& descriptor,
+                                           const WorkloadInfo& info) const override;
+
     std::unique_ptr<IWorkload> CreateConvolution2d(const Convolution2dQueueDescriptor& descriptor,
                                                    const WorkloadInfo& info) const override;
 
diff --git a/src/backends/cl/backend.mk b/src/backends/cl/backend.mk
index 137cf7654b..87964cdc7a 100644
--- a/src/backends/cl/backend.mk
+++ b/src/backends/cl/backend.mk
@@ -45,6 +45,7 @@ BACKEND_SOURCES := \
         workloads/ClPadWorkload.cpp \
         workloads/ClPermuteWorkload.cpp \
         workloads/ClPooling2dWorkload.cpp \
+        workloads/ClPreluWorkload.cpp \
         workloads/ClQuantizeWorkload.cpp \
         workloads/ClReshapeWorkload.cpp \
         workloads/ClResizeBilinearFloatWorkload.cpp \
diff --git a/src/backends/cl/test/ClCreateWorkloadTests.cpp b/src/backends/cl/test/ClCreateWorkloadTests.cpp
index aa1393f407..9e4f85b46c 100644
--- a/src/backends/cl/test/ClCreateWorkloadTests.cpp
+++ b/src/backends/cl/test/ClCreateWorkloadTests.cpp
@@ -452,6 +452,48 @@ BOOST_AUTO_TEST_CASE(CreatePooling2dFloat16NhwcWorkload)
     ClPooling2dWorkloadTest<armnn::DataType::Float16>(DataLayout::NHWC);
 }
 
+static void ClCreatePreluWorkloadTest(const armnn::TensorShape& inputShape,
+                                      const armnn::TensorShape& alphaShape,
+                                      const armnn::TensorShape& outputShape,
+                                      armnn::DataType dataType)
+{
+    Graph graph;
+    ClWorkloadFactory factory =
+            ClWorkloadFactoryHelper::GetFactory(ClWorkloadFactoryHelper::GetMemoryManager());
+
+    auto workload = CreatePreluWorkloadTest<ClPreluWorkload>(factory,
+                                                             graph,
+                                                             inputShape,
+                                                             alphaShape,
+                                                             outputShape,
+                                                             dataType);
+
+    // Checks that outputs and inputs are as we expect them (see definition of CreatePreluWorkloadTest).
+    PreluQueueDescriptor queueDescriptor = workload->GetData();
+    auto inputHandle = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    auto alphaHandle = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Inputs[1]);
+    auto outputHandle = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Outputs[0]);
+
+    BOOST_TEST((inputHandle->GetShape() == inputShape));
+    BOOST_TEST((alphaHandle->GetShape() == alphaShape));
+    BOOST_TEST((outputHandle->GetShape() == outputShape));
+}
+
+BOOST_AUTO_TEST_CASE(CreatePreluFloat16Workload)
+{
+    ClCreatePreluWorkloadTest({ 1, 4, 1, 2 }, { 5, 4, 3, 1 }, { 5, 4, 3, 2 }, DataType::Float16);
+}
+
+BOOST_AUTO_TEST_CASE(CreatePreluFloatWorkload)
+{
+    ClCreatePreluWorkloadTest({ 1, 4, 1, 2 }, { 5, 4, 3, 1 }, { 5, 4, 3, 2 }, DataType::Float32);
+}
+
+BOOST_AUTO_TEST_CASE(CreatePreluUint8Workload)
+{
+    ClCreatePreluWorkloadTest({ 1, 4, 1, 2 }, { 5, 4, 3, 1 }, { 5, 4, 3, 2 }, DataType::QuantisedAsymm8);
+}
+
 template <typename armnn::DataType DataType>
 static void ClCreateReshapeWorkloadTest()
 {
diff --git a/src/backends/cl/test/ClLayerTests.cpp b/src/backends/cl/test/ClLayerTests.cpp
index 3618adc70d..9ac244e505 100644
--- a/src/backends/cl/test/ClLayerTests.cpp
+++ b/src/backends/cl/test/ClLayerTests.cpp
@@ -331,6 +331,10 @@ ARMNN_AUTO_TEST_CASE(PadFloat322d, PadFloat322dTest)
 ARMNN_AUTO_TEST_CASE(PadFloat323d, PadFloat323dTest)
 ARMNN_AUTO_TEST_CASE(PadFloat324d, PadFloat324dTest)
 
+// PReLU
+ARMNN_AUTO_TEST_CASE(PreluFloat32, PreluTest<armnn::DataType::Float32>)
+ARMNN_AUTO_TEST_CASE(PreluUint8,   PreluTest<armnn::DataType::QuantisedAsymm8>)
+
 // Permute
 ARMNN_AUTO_TEST_CASE(SimplePermuteFloat32, SimplePermuteFloat32Test)
 ARMNN_AUTO_TEST_CASE(SimplePermuteUint8, SimplePermuteUint8Test)
diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt
index 1b4c85b947..40c5f766e4 100644
--- a/src/backends/cl/workloads/CMakeLists.txt
+++ b/src/backends/cl/workloads/CMakeLists.txt
@@ -54,6 +54,8 @@ list(APPEND armnnClBackendWorkloads_sources
     ClPermuteWorkload.hpp
     ClPooling2dWorkload.cpp
     ClPooling2dWorkload.hpp
+    ClPreluWorkload.cpp
+    ClPreluWorkload.hpp
     ClQuantizeWorkload.cpp
     ClQuantizeWorkload.hpp
     ClReshapeWorkload.cpp
diff --git a/src/backends/cl/workloads/ClPreluWorkload.cpp b/src/backends/cl/workloads/ClPreluWorkload.cpp
new file mode 100644
index 0000000000..1813105c21
--- /dev/null
+++ b/src/backends/cl/workloads/ClPreluWorkload.cpp
@@ -0,0 +1,48 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ClPreluWorkload.hpp"
+#include "ClWorkloadUtils.hpp"
+#include <backendsCommon/CpuTensorHandle.hpp>
+#include <aclCommon/ArmComputeUtils.hpp>
+#include <cl/ClLayerSupport.hpp>
+#include <cl/ClTensorHandle.hpp>
+
+namespace armnn
+{
+
+arm_compute::Status ClPreluWorkloadValidate(const TensorInfo& input,
+                                            const TensorInfo& alpha,
+                                            const TensorInfo& output)
+{
+    const arm_compute::TensorInfo aclInput = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+    const arm_compute::TensorInfo aclAlpha = armcomputetensorutils::BuildArmComputeTensorInfo(alpha);
+    const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+    return arm_compute::CLPReluLayer::validate(&aclInput,
+                                               &aclAlpha,
+                                               &aclOutput);
+}
+
+ClPreluWorkload::ClPreluWorkload(const PreluQueueDescriptor& descriptor,
+                                 const WorkloadInfo& info)
+    : BaseWorkload<PreluQueueDescriptor>(descriptor, info)
+{
+    m_Data.ValidateInputsOutputs("ClPreluWorkload", 1, 1);
+
+    arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ICLTensor& alpha = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
+    arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+    m_PreluLayer.configure(&input, &alpha, &output);
+}
+
+void ClPreluWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_CL("ClPreluWorkload_Execute");
+    RunClFunction(m_PreluLayer, CHECK_LOCATION());
+}
+
+} //namespace armnn
diff --git a/src/backends/cl/workloads/ClPreluWorkload.hpp b/src/backends/cl/workloads/ClPreluWorkload.hpp
new file mode 100644
index 0000000000..6ffe4ca7c0
--- /dev/null
+++ b/src/backends/cl/workloads/ClPreluWorkload.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/CLFunctions.h>
+
+namespace armnn
+{
+arm_compute::Status ClPreluWorkloadValidate(const TensorInfo& input,
+                                            const TensorInfo& alpha,
+                                            const TensorInfo& output);
+
+class ClPreluWorkload : public BaseWorkload<PreluQueueDescriptor>
+{
+public:
+    ClPreluWorkload(const PreluQueueDescriptor& descriptor, const WorkloadInfo& info);
+    void Execute() const override;
+
+private:
+    mutable arm_compute::CLPReluLayer m_PreluLayer;
+};
+
+} //namespace armnn
diff --git a/src/backends/cl/workloads/ClWorkloads.hpp b/src/backends/cl/workloads/ClWorkloads.hpp
index e82afc54c3..6ba8138780 100644
--- a/src/backends/cl/workloads/ClWorkloads.hpp
+++ b/src/backends/cl/workloads/ClWorkloads.hpp
@@ -27,6 +27,7 @@
 #include "ClPermuteWorkload.hpp"
 #include "ClPadWorkload.hpp"
 #include "ClPooling2dWorkload.hpp"
+#include "ClPreluWorkload.hpp"
 #include "ClQuantizeWorkload.hpp"
 #include "ClReshapeWorkload.hpp"
 #include "ClResizeBilinearFloatWorkload.hpp"