IVGCVSW-3829 Add CL workload for RSQRT

Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>
Change-Id: Idf40ed0ebf2af3170137bf3a3cb5a9635a203d4a

11 files changed, 151 insertions, 1 deletions

diff --git a/src/armnn/test/CreateWorkload.hpp b/src/armnn/test/CreateWorkload.hpp
index b50135aea5..d6d1ab3c5b 100644
--- a/src/armnn/test/CreateWorkload.hpp
+++ b/src/armnn/test/CreateWorkload.hpp
@@ -129,6 +129,32 @@ std::unique_ptr<WorkloadType> CreateElementwiseWorkloadTest(armnn::IWorkloadFact
     return workload;
 }
 
+template <typename WorkloadType,
+          typename DescriptorType,
+          typename LayerType,
+          armnn::DataType DataType>
+std::unique_ptr<WorkloadType> CreateElementwiseUnaryWorkloadTest(armnn::IWorkloadFactory & factory,
+                                                                 armnn::Graph & graph)
+{
+    Layer* const layer = graph.AddLayer<LayerType>("layer");
+
+    Layer* const input  = graph.AddLayer<InputLayer>(0, "input");
+    Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
+
+    armnn::TensorInfo tensorInfo({ 2, 3 }, DataType);
+    Connect(input, layer, tensorInfo, 0, 0);
+    Connect(layer, output, tensorInfo, 0, 0);
+    CreateTensorHandles(graph, factory);
+
+    auto workload = MakeAndCheckWorkload<WorkloadType>(*layer, graph, factory);
+    DescriptorType queueDescriptor = workload->GetData();
+
+    BOOST_TEST(queueDescriptor.m_Inputs.size()  == 1);
+    BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
+
+    return workload;
+}
+
 template <typename BatchNormalizationWorkloadType, armnn::DataType DataType>
 std::unique_ptr<BatchNormalizationWorkloadType> CreateBatchNormalizationWorkloadTest(
     armnn::IWorkloadFactory& factory, armnn::Graph& graph, DataLayout dataLayout = DataLayout::NCHW)
diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp
index 1fd3aa532e..5905cd990e 100644
--- a/src/backends/cl/ClLayerSupport.cpp
+++ b/src/backends/cl/ClLayerSupport.cpp
@@ -42,6 +42,7 @@
 #include "workloads/ClPooling2dWorkload.hpp"
 #include "workloads/ClPreluWorkload.hpp"
 #include "workloads/ClResizeWorkload.hpp"
+#include "workloads/ClRsqrtWorkload.hpp"
 #include "workloads/ClQuantizedLstmWorkload.hpp"
 #include "workloads/ClQuantizeWorkload.hpp"
 #include "workloads/ClSoftmaxBaseWorkload.hpp"
@@ -601,6 +602,13 @@ bool ClLayerSupport::IsResizeBilinearSupported(const TensorInfo& input,
     return IsResizeSupported(input, output, descriptor, reasonIfUnsupported);
 }
 
+bool ClLayerSupport::IsRsqrtSupported(const TensorInfo& input,
+                                      const TensorInfo& output,
+                                      Optional<std::string&> reasonIfUnsupported) const
+{
+    FORWARD_WORKLOAD_VALIDATE_FUNC(ClRsqrtWorkloadValidate, reasonIfUnsupported, input, output);
+}
+
 bool ClLayerSupport::IsSoftmaxSupported(const TensorInfo& input,
                                         const TensorInfo& output,
                                         const SoftmaxDescriptor& descriptor,
diff --git a/src/backends/cl/ClLayerSupport.hpp b/src/backends/cl/ClLayerSupport.hpp
index b19ddae769..ddc428de07 100644
--- a/src/backends/cl/ClLayerSupport.hpp
+++ b/src/backends/cl/ClLayerSupport.hpp
@@ -201,6 +201,10 @@ public:
                                    const TensorInfo& output,
                                    Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsRsqrtSupported(const TensorInfo& input,
+                          const TensorInfo& output,
+                          Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsSoftmaxSupported(const TensorInfo& input,
                             const TensorInfo& output,
                             const SoftmaxDescriptor& descriptor,
diff --git a/src/backends/cl/ClWorkloadFactory.cpp b/src/backends/cl/ClWorkloadFactory.cpp
index 7e5aecfb58..2c4b843be1 100644
--- a/src/backends/cl/ClWorkloadFactory.cpp
+++ b/src/backends/cl/ClWorkloadFactory.cpp
@@ -422,7 +422,7 @@ std::unique_ptr<IWorkload> ClWorkloadFactory::CreateDebug(const DebugQueueDescri
 std::unique_ptr<IWorkload> ClWorkloadFactory::CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
                                                           const WorkloadInfo& info) const
 {
-    return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info);
+    return MakeWorkload<ClRsqrtWorkload>(descriptor, info);
 }
 
 std::unique_ptr<IWorkload> ClWorkloadFactory::CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor,
diff --git a/src/backends/cl/backend.mk b/src/backends/cl/backend.mk
index 797e085ea1..c417c4f96b 100644
--- a/src/backends/cl/backend.mk
+++ b/src/backends/cl/backend.mk
@@ -53,6 +53,7 @@ BACKEND_SOURCES := \
         workloads/ClQuantizeWorkload.cpp \
         workloads/ClReshapeWorkload.cpp \
         workloads/ClResizeWorkload.cpp \
+        workloads/ClRsqrtWorkload.cpp \
         workloads/ClSoftmaxBaseWorkload.cpp \
         workloads/ClSoftmaxFloatWorkload.cpp \
         workloads/ClSoftmaxUint8Workload.cpp \
diff --git a/src/backends/cl/test/ClCreateWorkloadTests.cpp b/src/backends/cl/test/ClCreateWorkloadTests.cpp
index bb6d041c7c..3c80ea9d49 100644
--- a/src/backends/cl/test/ClCreateWorkloadTests.cpp
+++ b/src/backends/cl/test/ClCreateWorkloadTests.cpp
@@ -146,6 +146,36 @@ BOOST_AUTO_TEST_CASE(CreateDivisionFloat16WorkloadTest)
                                     armnn::DataType::Float16>();
 }
 
+template <typename WorkloadType,
+          typename DescriptorType,
+          typename LayerType,
+          armnn::DataType DataType>
+static void ClCreateElementwiseUnaryWorkloadTest()
+{
+    Graph graph;
+    ClWorkloadFactory factory =
+        ClWorkloadFactoryHelper::GetFactory(ClWorkloadFactoryHelper::GetMemoryManager());
+
+    auto workload = CreateElementwiseUnaryWorkloadTest
+        <WorkloadType, DescriptorType, LayerType, DataType>(factory, graph);
+
+    DescriptorType queueDescriptor = workload->GetData();
+
+    auto inputHandle  = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    auto outputHandle = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Outputs[0]);
+
+    BOOST_TEST(CompareIClTensorHandleShape(inputHandle, {2, 3}));
+    BOOST_TEST(CompareIClTensorHandleShape(outputHandle, {2, 3}));
+}
+
+BOOST_AUTO_TEST_CASE(CreateRsqrtFloat32WorkloadTest)
+{
+    ClCreateElementwiseUnaryWorkloadTest<ClRsqrtWorkload,
+                                         RsqrtQueueDescriptor,
+                                         RsqrtLayer,
+                                         armnn::DataType::Float32>();
+}
+
 template <typename BatchNormalizationWorkloadType, armnn::DataType DataType>
 static void ClCreateBatchNormalizationWorkloadTest(DataLayout dataLayout)
 {
diff --git a/src/backends/cl/test/ClLayerTests.cpp b/src/backends/cl/test/ClLayerTests.cpp
index cc1365711a..99ee4a4b61 100644
--- a/src/backends/cl/test/ClLayerTests.cpp
+++ b/src/backends/cl/test/ClLayerTests.cpp
@@ -640,6 +640,12 @@ ARMNN_AUTO_TEST_CASE(ResizeNearestNeighborMagUint8Nhwc,
                      ResizeNearestNeighborMagTest<DataType::QuantisedAsymm8>,
                      DataLayout::NHWC, 0.1f, 50, 0.1f, 50)
 
+// Rsqrt
+ARMNN_AUTO_TEST_CASE(Rsqrt2d, Rsqrt2dTest<DataType::Float32>)
+ARMNN_AUTO_TEST_CASE(Rsqrt3d, Rsqrt3dTest<DataType::Float32>)
+ARMNN_AUTO_TEST_CASE(RsqrtZero, RsqrtZeroTest<DataType::Float32>)
+ARMNN_AUTO_TEST_CASE(RsqrtNegative, RsqrtNegativeTest<DataType::Float32>)
+
 // Quantize
 ARMNN_AUTO_TEST_CASE(QuantizeSimpleUint8, QuantizeSimpleUint8Test)
 ARMNN_AUTO_TEST_CASE(QuantizeClampUint8, QuantizeClampUint8Test)
diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt
index 3e2ac76266..7578a4029d 100644
--- a/src/backends/cl/workloads/CMakeLists.txt
+++ b/src/backends/cl/workloads/CMakeLists.txt
@@ -66,6 +66,8 @@ list(APPEND armnnClBackendWorkloads_sources
     ClReshapeWorkload.hpp
     ClResizeWorkload.cpp
     ClResizeWorkload.hpp
+    ClRsqrtWorkload.cpp
+    ClRsqrtWorkload.hpp
     ClSoftmaxBaseWorkload.cpp
     ClSoftmaxBaseWorkload.hpp
     ClSoftmaxFloatWorkload.cpp
diff --git a/src/backends/cl/workloads/ClRsqrtWorkload.cpp b/src/backends/cl/workloads/ClRsqrtWorkload.cpp
new file mode 100644
index 0000000000..be687595f7
--- /dev/null
+++ b/src/backends/cl/workloads/ClRsqrtWorkload.cpp
@@ -0,0 +1,44 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ClRsqrtWorkload.hpp"
+
+#include "ClWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <cl/ClTensorHandle.hpp>
+
+#include <boost/cast.hpp>
+
+namespace armnn
+{
+
+arm_compute::Status ClRsqrtWorkloadValidate(const TensorInfo& input, const TensorInfo& output)
+{
+    const arm_compute::TensorInfo aclInput  = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+    const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+    return arm_compute::CLRsqrtLayer::validate(&aclInput, &aclOutput);
+}
+
+ClRsqrtWorkload::ClRsqrtWorkload(const RsqrtQueueDescriptor& descriptor, const WorkloadInfo& info)
+    : BaseWorkload<RsqrtQueueDescriptor>(descriptor, info)
+{
+    m_Data.ValidateInputsOutputs("ClRsqrtWorkload", 1, 1);
+
+    arm_compute::ICLTensor& input  = boost::polymorphic_downcast<ClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ICLTensor& output = boost::polymorphic_downcast<ClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+    m_RsqrtLayer.configure(&input, &output);
+}
+
+void ClRsqrtWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_CL("ClRsqrtWorkload_Execute");
+    RunClFunction(m_RsqrtLayer, CHECK_LOCATION());
+}
+
+} // namespace armnn
diff --git a/src/backends/cl/workloads/ClRsqrtWorkload.hpp b/src/backends/cl/workloads/ClRsqrtWorkload.hpp
new file mode 100644
index 0000000000..8fb6229160
--- /dev/null
+++ b/src/backends/cl/workloads/ClRsqrtWorkload.hpp
@@ -0,0 +1,28 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/core/Error.h>
+#include <arm_compute/runtime/CL/functions/CLElementWiseUnaryLayer.h>
+
+namespace armnn
+{
+
+arm_compute::Status ClRsqrtWorkloadValidate(const TensorInfo& input, const TensorInfo& output);
+
+class ClRsqrtWorkload : public BaseWorkload<RsqrtQueueDescriptor>
+{
+public:
+    ClRsqrtWorkload(const RsqrtQueueDescriptor& descriptor, const WorkloadInfo& info);
+    virtual void Execute() const override;
+
+private:
+    mutable arm_compute::CLRsqrtLayer m_RsqrtLayer;
+};
+
+} // namespace armnn
diff --git a/src/backends/cl/workloads/ClWorkloads.hpp b/src/backends/cl/workloads/ClWorkloads.hpp
index 45030e4f98..7525f84319 100644
--- a/src/backends/cl/workloads/ClWorkloads.hpp
+++ b/src/backends/cl/workloads/ClWorkloads.hpp
@@ -33,6 +33,7 @@
 #include "ClQuantizedLstmWorkload.hpp"
 #include "ClReshapeWorkload.hpp"
 #include "ClResizeWorkload.hpp"
+#include "ClRsqrtWorkload.hpp"
 #include "ClSoftmaxFloatWorkload.hpp"
 #include "ClSoftmaxUint8Workload.hpp"
 #include "ClSpaceToBatchNdWorkload.hpp"