IVGCVSW-7570 GpuFsa Op: Add ElemenWiseBinary Operators available

* Refactor to generalize
* Add MUL

Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: I2ee273d50d3a8b114b5a41abc8ee7585b15e3308

7 files changed, 169 insertions, 202 deletions

diff --git a/src/backends/gpuFsa/layers/CMakeLists.txt b/src/backends/gpuFsa/layers/CMakeLists.txt
index 9ea36b6f2c..3fe4bdcbc6 100644
--- a/src/backends/gpuFsa/layers/CMakeLists.txt
+++ b/src/backends/gpuFsa/layers/CMakeLists.txt
@@ -10,10 +10,8 @@ list(APPEND armnnGpuFsaBackendLayers_sources
         GpuFsaConvolution2d.hpp
         GpuFsaDepthwiseConvolution2d.cpp
         GpuFsaDepthwiseConvolution2d.hpp
-        GpuFsaElementwiseBinaryAdd.cpp
-        GpuFsaElementwiseBinaryAdd.hpp
-        GpuFsaElementwiseBinarySub.cpp
-        GpuFsaElementwiseBinarySub.hpp
+        GpuFsaElementwiseBinary.cpp
+        GpuFsaElementwiseBinary.hpp
         GpuFsaPooling2d.cpp
         GpuFsaPooling2d.hpp
         GpuFsaResize.cpp
diff --git a/src/backends/gpuFsa/layers/GpuFsaElementwiseBinary.cpp b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinary.cpp
new file mode 100644
index 0000000000..7c3760a4ef
--- /dev/null
+++ b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinary.cpp
@@ -0,0 +1,145 @@
+//
+// Copyright © 2024 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "GpuFsaElementwiseBinary.hpp"
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadContext.h>
+#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h>
+#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuAdd.h>
+#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuAdd.h>
+#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuMul.h>
+#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSub.h>
+#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h>
+
+using namespace arm_compute::experimental::dynamic_fusion;
+using namespace armnn::armcomputetensorutils;
+
+namespace armnn
+{
+
+arm_compute::Status GpuFsaElementwiseBinaryValidate(const TensorInfo& input0,
+                                                    const TensorInfo& input1,
+                                                    const ElementwiseBinaryDescriptor& descriptor)
+{
+    // Create a new workload sketch, for validation purposes
+    auto compileCtx         = arm_compute::CLKernelLibrary::get().get_compile_context();
+    auto workloadContext    = GpuWorkloadContext(&compileCtx);
+    GpuWorkloadSketch sketch{ &workloadContext };
+
+    arm_compute::TensorInfo aclInput0Info = BuildArmComputeTensorInfo(input0, input0.GetNumDimensions());
+    arm_compute::TensorInfo aclInput1Info = BuildArmComputeTensorInfo(input1, input1.GetNumDimensions());
+
+    aclInput0Info.set_are_values_constant(input0.IsConstant());
+    aclInput1Info.set_are_values_constant(input1.IsConstant());
+
+    arm_compute::ITensorInfo*  inputInfo0 = workloadContext.create_tensor_info(aclInput0Info);
+    arm_compute::ITensorInfo*  inputInfo1 = workloadContext.create_tensor_info(aclInput1Info);
+
+    switch (descriptor.m_Operation)
+    {
+        case BinaryOperation::Add:
+        {
+            return GpuAdd::validate_op(sketch, inputInfo0, inputInfo1);
+        }
+        case BinaryOperation::Mul:
+        {
+            return GpuMul::validate_op(sketch, inputInfo0, inputInfo1);
+        }
+        case BinaryOperation::Sub:
+        {
+            return GpuSub::validate_op(sketch, inputInfo0, inputInfo1);
+        }
+        default:
+            return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR,
+                                       std::string("Elementwise Binary operation not supported in GpuFsa: ")
+                                       + GetBinaryOperationAsCString(descriptor.m_Operation));
+    }
+}
+
+void GpuFsaElementwiseBinaryCreateOp(GpuFsaPreCompiledBlob* blob,
+                                     const TensorInfo& input0,
+                                     const TensorInfo& input1,
+                                     const ElementwiseBinaryDescriptor& descriptor)
+{
+    GpuWorkloadSketch* sketch           = blob->sketch.get();
+    GpuWorkloadContext* workloadContext = blob->workloadContext.get();
+    std::vector<arm_compute::ITensorInfo*> inputTensorInfos  = {};
+    std::vector<arm_compute::ITensorInfo*> outputTensorInfos = {};
+
+    arm_compute::TensorInfo aclInput0Info = BuildArmComputeTensorInfo(input0, input0.GetNumDimensions());
+    arm_compute::TensorInfo aclInput1Info = BuildArmComputeTensorInfo(input1, input1.GetNumDimensions());
+
+    aclInput0Info.set_are_values_constant(input0.IsConstant());
+    aclInput1Info.set_are_values_constant(input1.IsConstant());
+
+    inputTensorInfos.emplace_back(workloadContext->create_tensor_info(aclInput0Info));
+    inputTensorInfos.emplace_back(workloadContext->create_tensor_info(aclInput1Info));
+
+    // Validate operator, check status and update reasonIfUnsupported
+    // Validate operator, check status and update reasonIfUnsupported
+    arm_compute::Status aclStatus{};
+    switch (descriptor.m_Operation)
+    {
+        case BinaryOperation::Add:
+        {
+            aclStatus = GpuAdd::validate_op(*sketch, inputTensorInfos[0], inputTensorInfos[1]);
+            break;
+        }
+        case BinaryOperation::Mul:
+        {
+            aclStatus = GpuMul::validate_op(*sketch, inputTensorInfos[0], inputTensorInfos[1]);
+            break;
+        }
+        case BinaryOperation::Sub:
+        {
+            aclStatus = GpuSub::validate_op(*sketch, inputTensorInfos[0], inputTensorInfos[1]);
+            break;
+        }
+        default:
+            throw InvalidArgumentException(std::string("Elementwise Binary operation not supported in GpuFsa: ")
+                                           + GetBinaryOperationAsCString(descriptor.m_Operation));
+    }
+
+    const bool supported = aclStatus.error_code() == arm_compute::ErrorCode::OK;
+    if (!supported)
+    {
+        throw BackendCapabilityException("\"GpuFsa\" backend failed during elementwise binary add validation");
+    }
+
+    arm_compute::ITensorInfo* elementwiseBinaryOutputInfo{};
+    switch (descriptor.m_Operation)
+    {
+        case BinaryOperation::Add:
+        {
+            elementwiseBinaryOutputInfo = GpuAdd::create_op(*sketch, inputTensorInfos[0], inputTensorInfos[1]);
+            break;
+        }
+        case BinaryOperation::Mul:
+        {
+            elementwiseBinaryOutputInfo = GpuMul::create_op(*sketch, inputTensorInfos[0], inputTensorInfos[1]);
+            break;
+        }
+        case BinaryOperation::Sub:
+        {
+            elementwiseBinaryOutputInfo = GpuSub::create_op(*sketch, inputTensorInfos[0], inputTensorInfos[1]);
+            break;
+        }
+        default:
+            throw InvalidArgumentException(std::string("Elementwise Binary operation not supported in GpuFsa: ")
+                                           + GetBinaryOperationAsCString(descriptor.m_Operation));
+    }
+
+    // Temporary fix until fusing attempt is make for GpuFsa backend and Output layer workload is created.
+    outputTensorInfos.emplace_back(workloadContext->create_tensor_info());
+    GpuOutput::create_op(*sketch, elementwiseBinaryOutputInfo, outputTensorInfos[0]);
+
+    // Store the TensorInfos within the blob as unique_ptrs to be used later
+    blob->inputTensorInfos  = std::make_unique<std::vector<arm_compute::ITensorInfo*>>(inputTensorInfos);
+    blob->outputTensorInfos = std::make_unique<std::vector<arm_compute::ITensorInfo*>>(outputTensorInfos);
+}
+
+} // namespace armnn
\ No newline at end of file
diff --git a/src/backends/gpuFsa/layers/GpuFsaElementwiseBinary.hpp b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinary.hpp
new file mode 100644
index 0000000000..11583f1dc7
--- /dev/null
+++ b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinary.hpp
@@ -0,0 +1,22 @@
+//
+// Copyright © 2024 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+#pragma once
+
+#include <armnn/Descriptors.hpp>
+
+#include <gpuFsa/GpuFsaBackend.hpp>
+
+namespace armnn
+{
+arm_compute::Status GpuFsaElementwiseBinaryValidate(const TensorInfo& input0,
+                                                    const TensorInfo& input1,
+                                                    const ElementwiseBinaryDescriptor& descriptor);
+
+void GpuFsaElementwiseBinaryCreateOp(GpuFsaPreCompiledBlob* blob,
+                                     const TensorInfo& input0,
+                                     const TensorInfo& input1,
+                                     const ElementwiseBinaryDescriptor& descriptor);
+
+} // namespace armnn
\ No newline at end of file
diff --git a/src/backends/gpuFsa/layers/GpuFsaElementwiseBinaryAdd.cpp b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinaryAdd.cpp
deleted file mode 100644
index d6404dd67e..0000000000
--- a/src/backends/gpuFsa/layers/GpuFsaElementwiseBinaryAdd.cpp
+++ /dev/null
@@ -1,79 +0,0 @@
-//
-// Copyright © 2024 Arm Ltd and Contributors. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#include "GpuFsaElementwiseBinaryAdd.hpp"
-
-#include <aclCommon/ArmComputeTensorUtils.hpp>
-
-#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadContext.h>
-#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h>
-#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuAdd.h>
-#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h>
-
-using namespace arm_compute::experimental::dynamic_fusion;
-using namespace armnn::armcomputetensorutils;
-
-namespace armnn
-{
-
-arm_compute::Status GpuFsaElementwiseBinaryAddValidate(const TensorInfo& input0,
-                                                       const TensorInfo& input1)
-{
-    // Create a new workload sketch, for validation purposes
-    auto compileCtx         = arm_compute::CLKernelLibrary::get().get_compile_context();
-    auto workloadContext    = GpuWorkloadContext(&compileCtx);
-    GpuWorkloadSketch sketch{ &workloadContext };
-
-    arm_compute::TensorInfo aclInput0Info = BuildArmComputeTensorInfo(input0, input0.GetNumDimensions());
-    arm_compute::TensorInfo aclInput1Info = BuildArmComputeTensorInfo(input1, input1.GetNumDimensions());
-
-    aclInput0Info.set_are_values_constant(input0.IsConstant());
-    aclInput1Info.set_are_values_constant(input1.IsConstant());
-
-    arm_compute::ITensorInfo*  inputInfo0 = workloadContext.create_tensor_info(aclInput0Info);
-    arm_compute::ITensorInfo*  inputInfo1 = workloadContext.create_tensor_info(aclInput1Info);
-
-    return GpuAdd::validate_op(sketch, inputInfo0, inputInfo1);
-}
-
-void GpuFsaElementwiseBinaryAddCreateOp(GpuFsaPreCompiledBlob* blob,
-                                        const TensorInfo& input0,
-                                        const TensorInfo& input1)
-{
-    GpuWorkloadSketch* sketch           = blob->sketch.get();
-    GpuWorkloadContext* workloadContext = blob->workloadContext.get();
-    std::vector<arm_compute::ITensorInfo*> inputTensorInfos  = {};
-    std::vector<arm_compute::ITensorInfo*> outputTensorInfos = {};
-
-    arm_compute::TensorInfo aclInput0Info = BuildArmComputeTensorInfo(input0, input0.GetNumDimensions());
-    arm_compute::TensorInfo aclInput1Info = BuildArmComputeTensorInfo(input1, input1.GetNumDimensions());
-
-    aclInput0Info.set_are_values_constant(input0.IsConstant());
-    aclInput1Info.set_are_values_constant(input1.IsConstant());
-
-    inputTensorInfos.emplace_back(workloadContext->create_tensor_info(aclInput0Info));
-    inputTensorInfos.emplace_back(workloadContext->create_tensor_info(aclInput1Info));
-
-    // Validate operator, check status and update reasonIfUnsupported
-    arm_compute::Status aclStatus = GpuAdd::validate_op(*sketch, inputTensorInfos[0], inputTensorInfos[1]);
-    const bool supported = aclStatus.error_code() == arm_compute::ErrorCode::OK;
-    if (!supported)
-    {
-        throw BackendCapabilityException("\"GpuFsa\" backend failed during elementwise binary add validation");
-    }
-
-    arm_compute::ITensorInfo* addOutputInfo =
-            GpuAdd::create_op(*sketch, inputTensorInfos[0], inputTensorInfos[1]);
-
-    // Temporary fix until fusing attempt is make for GpuFsa backend and Output layer workload is created.
-    outputTensorInfos.emplace_back(workloadContext->create_tensor_info());
-    GpuOutput::create_op(*sketch, addOutputInfo, outputTensorInfos[0]);
-
-    // Store the TensorInfos within the blob as unique_ptrs to be used later
-    blob->inputTensorInfos  = std::make_unique<std::vector<arm_compute::ITensorInfo*>>(inputTensorInfos);
-    blob->outputTensorInfos = std::make_unique<std::vector<arm_compute::ITensorInfo*>>(outputTensorInfos);
-}
-
-} // namespace armnn
\ No newline at end of file
diff --git a/src/backends/gpuFsa/layers/GpuFsaElementwiseBinaryAdd.hpp b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinaryAdd.hpp
deleted file mode 100644
index 1392d01257..0000000000
--- a/src/backends/gpuFsa/layers/GpuFsaElementwiseBinaryAdd.hpp
+++ /dev/null
@@ -1,20 +0,0 @@
-//
-// Copyright © 2024 Arm Ltd and Contributors. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-#pragma once
-
-#include <armnn/Descriptors.hpp>
-
-#include <gpuFsa/GpuFsaBackend.hpp>
-
-namespace armnn
-{
-arm_compute::Status GpuFsaElementwiseBinaryAddValidate(const TensorInfo& input0,
-                                                       const TensorInfo& input1);
-
-void GpuFsaElementwiseBinaryAddCreateOp(GpuFsaPreCompiledBlob* blob,
-                                        const TensorInfo& input0,
-                                        const TensorInfo& input1);
-
-} // namespace armnn
\ No newline at end of file
diff --git a/src/backends/gpuFsa/layers/GpuFsaElementwiseBinarySub.cpp b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinarySub.cpp
deleted file mode 100644
index 5e0f478686..0000000000
--- a/src/backends/gpuFsa/layers/GpuFsaElementwiseBinarySub.cpp
+++ /dev/null
@@ -1,79 +0,0 @@
-//
-// Copyright © 2024 Arm Ltd and Contributors. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#include "GpuFsaElementwiseBinarySub.hpp"
-
-#include <aclCommon/ArmComputeTensorUtils.hpp>
-
-#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadContext.h>
-#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h>
-#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSub.h>
-#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h>
-
-using namespace arm_compute::experimental::dynamic_fusion;
-using namespace armnn::armcomputetensorutils;
-
-namespace armnn
-{
-
-arm_compute::Status GpuFsaElementwiseBinarySubValidate(const TensorInfo& input0,
-                                                       const TensorInfo& input1)
-{
-    // Create a new workload sketch, for validation purposes
-    auto compileCtx         = arm_compute::CLKernelLibrary::get().get_compile_context();
-    auto workloadContext    = GpuWorkloadContext(&compileCtx);
-    GpuWorkloadSketch sketch{ &workloadContext };
-
-    arm_compute::TensorInfo aclInput0Info = BuildArmComputeTensorInfo(input0, input0.GetNumDimensions());
-    arm_compute::TensorInfo aclInput1Info = BuildArmComputeTensorInfo(input1, input1.GetNumDimensions());
-
-    aclInput0Info.set_are_values_constant(input0.IsConstant());
-    aclInput1Info.set_are_values_constant(input1.IsConstant());
-
-    arm_compute::ITensorInfo*  inputInfo0 = workloadContext.create_tensor_info(aclInput0Info);
-    arm_compute::ITensorInfo*  inputInfo1 = workloadContext.create_tensor_info(aclInput1Info);
-
-    return GpuSub::validate_op(sketch, inputInfo0, inputInfo1);
-}
-
-void GpuFsaElementwiseBinarySubCreateOp(GpuFsaPreCompiledBlob* blob,
-                                        const TensorInfo& input0,
-                                        const TensorInfo& input1)
-{
-    GpuWorkloadSketch* sketch           = blob->sketch.get();
-    GpuWorkloadContext* workloadContext = blob->workloadContext.get();
-    std::vector<arm_compute::ITensorInfo*> inputTensorInfos  = {};
-    std::vector<arm_compute::ITensorInfo*> outputTensorInfos = {};
-
-    arm_compute::TensorInfo aclInput0Info = BuildArmComputeTensorInfo(input0, input0.GetNumDimensions());
-    arm_compute::TensorInfo aclInput1Info = BuildArmComputeTensorInfo(input1, input1.GetNumDimensions());
-
-    aclInput0Info.set_are_values_constant(input0.IsConstant());
-    aclInput1Info.set_are_values_constant(input1.IsConstant());
-
-    inputTensorInfos.emplace_back(workloadContext->create_tensor_info(aclInput0Info));
-    inputTensorInfos.emplace_back(workloadContext->create_tensor_info(aclInput1Info));
-
-    // Validate operator, check status and update reasonIfUnsupported
-    arm_compute::Status aclStatus = GpuSub::validate_op(*sketch, inputTensorInfos[0], inputTensorInfos[1]);
-    const bool supported = aclStatus.error_code() == arm_compute::ErrorCode::OK;
-    if (!supported)
-    {
-        throw BackendCapabilityException("\"GpuFsa\" backend failed during elementwise binary subtract validation");
-    }
-
-    arm_compute::ITensorInfo* addOutputInfo =
-            GpuSub::create_op(*sketch, inputTensorInfos[0], inputTensorInfos[1]);
-
-    // Temporary fix until fusing attempt is make for GpuFsa backend and Output layer workload is created.
-    outputTensorInfos.emplace_back(workloadContext->create_tensor_info());
-    GpuOutput::create_op(*sketch, addOutputInfo, outputTensorInfos[0]);
-
-    // Store the TensorInfos within the blob as unique_ptrs to be used later
-    blob->inputTensorInfos  = std::make_unique<std::vector<arm_compute::ITensorInfo*>>(inputTensorInfos);
-    blob->outputTensorInfos = std::make_unique<std::vector<arm_compute::ITensorInfo*>>(outputTensorInfos);
-}
-
-}
\ No newline at end of file
diff --git a/src/backends/gpuFsa/layers/GpuFsaElementwiseBinarySub.hpp b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinarySub.hpp
deleted file mode 100644
index 4d58f313b6..0000000000
--- a/src/backends/gpuFsa/layers/GpuFsaElementwiseBinarySub.hpp
+++ /dev/null
@@ -1,20 +0,0 @@
-//
-// Copyright © 2024 Arm Ltd and Contributors. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-#pragma once
-
-#include <armnn/Descriptors.hpp>
-
-#include <gpuFsa/GpuFsaBackend.hpp>
-
-namespace armnn
-{
-arm_compute::Status GpuFsaElementwiseBinarySubValidate(const TensorInfo& input0,
-                                                       const TensorInfo& input1);
-
-void GpuFsaElementwiseBinarySubCreateOp(GpuFsaPreCompiledBlob* blob,
-                                        const TensorInfo& input0,
-                                        const TensorInfo& input1);
-
-} // namespace armnn
\ No newline at end of file