IVGCVSW-6936 Add SQRT support to CL
Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: Ib90bade63cd0437329c690b09cf719a2e2bd06a4
diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp
index 09b3f43..6b8cf52 100644
--- a/src/backends/cl/ClLayerSupport.cpp
+++ b/src/backends/cl/ClLayerSupport.cpp
@@ -74,6 +74,7 @@
#include "workloads/ClSpaceToBatchNdWorkload.hpp"
#include "workloads/ClSpaceToDepthWorkload.hpp"
#include "workloads/ClSplitterWorkload.hpp"
+#include "workloads/ClSqrtWorkload.hpp"
#include "workloads/ClStackWorkload.hpp"
#include "workloads/ClStridedSliceWorkload.hpp"
#include "workloads/ClSubtractionWorkload.hpp"
@@ -957,6 +958,11 @@
reasonIfUnsupported,
input,
output);
+ case UnaryOperation::Sqrt:
+ FORWARD_WORKLOAD_VALIDATE_FUNC(ClSqrtWorkloadValidate,
+ reasonIfUnsupported,
+ input,
+ output);
default:
return false;
}
diff --git a/src/backends/cl/backend.mk b/src/backends/cl/backend.mk
index ed0d925..bf9689a 100644
--- a/src/backends/cl/backend.mk
+++ b/src/backends/cl/backend.mk
@@ -82,6 +82,7 @@
workloads/ClSpaceToBatchNdWorkload.cpp \
workloads/ClSpaceToDepthWorkload.cpp \
workloads/ClSplitterWorkload.cpp \
+ workloads/ClSqrtWorkload.cpp \
workloads/ClStackWorkload.cpp \
workloads/ClStridedSliceWorkload.cpp \
workloads/ClSubtractionWorkload.cpp \
diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt
index 52326ae..59e11cd 100644
--- a/src/backends/cl/workloads/CMakeLists.txt
+++ b/src/backends/cl/workloads/CMakeLists.txt
@@ -117,6 +117,8 @@
ClSpaceToDepthWorkload.hpp
ClSplitterWorkload.cpp
ClSplitterWorkload.hpp
+ ClSqrtWorkload.cpp
+ ClSqrtWorkload.hpp
ClStackWorkload.cpp
ClStackWorkload.hpp
ClStridedSliceWorkload.cpp
diff --git a/src/backends/cl/workloads/ClSqrtWorkload.cpp b/src/backends/cl/workloads/ClSqrtWorkload.cpp
new file mode 100644
index 0000000..b78c114
--- /dev/null
+++ b/src/backends/cl/workloads/ClSqrtWorkload.cpp
@@ -0,0 +1,67 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ClSqrtWorkload.hpp"
+
+#include "ClWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+#include <aclCommon/ArmComputeUtils.hpp>
+#include <armnn/utility/PolymorphicDowncast.hpp>
+
+#include <cl/ClTensorHandle.hpp>
+
+namespace armnn
+{
+
+arm_compute::Status ClSqrtWorkloadValidate(const TensorInfo& input, const TensorInfo& output)
+{
+ const arm_compute::TensorInfo aclInput = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+ const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+ ActivationDescriptor descriptor;
+ descriptor.m_Function = ActivationFunction::Sqrt;
+ const arm_compute::ActivationLayerInfo activationLayerInfo =
+ ConvertActivationDescriptorToAclActivationLayerInfo(descriptor);
+
+ return arm_compute::CLActivationLayer::validate(&aclInput, &aclOutput, activationLayerInfo);
+}
+
+ClSqrtWorkload::ClSqrtWorkload(const ElementwiseUnaryQueueDescriptor& descriptor,
+ const WorkloadInfo& info,
+ const arm_compute::CLCompileContext& clCompileContext)
+ : ClBaseWorkload<ElementwiseUnaryQueueDescriptor>(descriptor, info)
+{
+ ARMNN_ASSERT(descriptor.m_Parameters.m_Operation == UnaryOperation::Sqrt);
+
+ // Report Profiling Details
+ ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClSqrtWorkload_Construct",
+ descriptor.m_Parameters,
+ info,
+ this->GetGuid());
+
+ m_Data.ValidateInputsOutputs("ClSqrtWorkload", 1, 1);
+
+ ActivationDescriptor activationDescriptor;
+ activationDescriptor.m_Function = ActivationFunction::Sqrt;
+ const arm_compute::ActivationLayerInfo activationLayerInfo =
+ ConvertActivationDescriptorToAclActivationLayerInfo(activationDescriptor);
+
+ arm_compute::ICLTensor& input = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+ arm_compute::ICLTensor& output = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+ {
+ ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClSqrtWorkload_configure");
+ m_SqrtLayer.configure(clCompileContext, &input, &output, activationLayerInfo);
+ }
+}
+
+void ClSqrtWorkload::Execute() const
+{
+ ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClSqrtWorkload_Execute", this->GetGuid());
+ RunClFunction(m_SqrtLayer, CHECK_LOCATION());
+}
+
+} // namespace armnn
diff --git a/src/backends/cl/workloads/ClSqrtWorkload.hpp b/src/backends/cl/workloads/ClSqrtWorkload.hpp
new file mode 100644
index 0000000..0476800
--- /dev/null
+++ b/src/backends/cl/workloads/ClSqrtWorkload.hpp
@@ -0,0 +1,30 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "ClBaseWorkload.hpp"
+
+#include <arm_compute/core/Error.h>
+#include <arm_compute/runtime/CL/functions/CLActivationLayer.h>
+
+namespace armnn
+{
+
+arm_compute::Status ClSqrtWorkloadValidate(const TensorInfo& input, const TensorInfo& output);
+
+class ClSqrtWorkload : public ClBaseWorkload<ElementwiseUnaryQueueDescriptor>
+{
+public:
+ ClSqrtWorkload(const ElementwiseUnaryQueueDescriptor& descriptor,
+ const WorkloadInfo& info,
+ const arm_compute::CLCompileContext& clCompileContext);
+ virtual void Execute() const override;
+
+private:
+ mutable arm_compute::CLActivationLayer m_SqrtLayer;
+};
+
+} // namespace armnn