IVGCVSW-7568 Implement Sub ElementwiseBinary operator GpuFsa

* Added support for Gpu Sub operator
* Added unit tests

Signed-off-by: John Mcloughlin <john.mcloughlin@arm.com>
Change-Id: I1efaa485772a3716e3781566843bd50bd9bab811

11 files changed, 228 insertions, 0 deletions

diff --git a/src/backends/cl/test/ClEndToEndTests.cpp b/src/backends/cl/test/ClEndToEndTests.cpp
index 8b2d6a514d..878054f7ba 100644
--- a/src/backends/cl/test/ClEndToEndTests.cpp
+++ b/src/backends/cl/test/ClEndToEndTests.cpp
@@ -90,6 +90,15 @@ TEST_CASE("ClSquaredDifferenceEndToEndTestUint8")
     ElementwiseBinarySimpleEndToEnd<armnn::DataType::QAsymmU8>(clDefaultBackends, BinaryOperation::SqDiff);
 }
 
+TEST_CASE("ClSubtractionEndToEndFloat32Simple3DTest")
+{
+    ElementwiseBinarySimple3DEndToEnd<armnn::DataType::Float32>(clDefaultBackends, BinaryOperation::Sub);
+}
+TEST_CASE("ClSubtractionEndToEndFloat16Simple3DTest")
+{
+    ElementwiseBinarySimple3DEndToEnd<armnn::DataType::Float16>(clDefaultBackends, BinaryOperation::Sub);
+}
+
 // Batch Mat Mul
 TEST_CASE("ClBatchMatMulEndToEndFloat32Test")
 {
diff --git a/src/backends/gpuFsa/GpuFsaBackend.cpp b/src/backends/gpuFsa/GpuFsaBackend.cpp
index 7951b17d9d..8b62aec9e6 100644
--- a/src/backends/gpuFsa/GpuFsaBackend.cpp
+++ b/src/backends/gpuFsa/GpuFsaBackend.cpp
@@ -23,6 +23,7 @@
 #include "layers/GpuFsaConvolution2d.hpp"
 #include "layers/GpuFsaDepthwiseConvolution2d.hpp"
 #include "layers/GpuFsaElementwiseBinaryAdd.hpp"
+#include "layers/GpuFsaElementwiseBinarySub.hpp"
 
 namespace armnn
 {
@@ -305,6 +306,13 @@ OptimizationViews GpuFsaBackend::OptimizeSubgraphView(const SubgraphView& subgra
 
                     GpuFsaElementwiseBinaryAddCreateOp(preCompiledBlobPtr, input0, input1);
                 }
+                else if (desc->m_Operation == BinaryOperation::Sub)
+                {
+                    auto input0 = base.GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo();
+                    auto input1 = base.GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo();
+
+                    GpuFsaElementwiseBinarySubCreateOp(preCompiledBlobPtr, input0, input1);
+                }
                 break;
             }
             default:
diff --git a/src/backends/gpuFsa/GpuFsaLayerSupport.cpp b/src/backends/gpuFsa/GpuFsaLayerSupport.cpp
index e98275c166..2e5c7d5a53 100644
--- a/src/backends/gpuFsa/GpuFsaLayerSupport.cpp
+++ b/src/backends/gpuFsa/GpuFsaLayerSupport.cpp
@@ -13,6 +13,7 @@
 #include "layers/GpuFsaConvolution2d.hpp"
 #include "layers/GpuFsaDepthwiseConvolution2d.hpp"
 #include "layers/GpuFsaElementwiseBinaryAdd.hpp"
+#include "layers/GpuFsaElementwiseBinarySub.hpp"
 #endif
 
 #include <vector>
@@ -144,6 +145,13 @@ bool GpuFsaLayerSupport::IsLayerSupported(const LayerType& type,
                                             infos[0],
                                             infos[1]);
             }
+            else if (desc->m_Operation == BinaryOperation::Sub)
+            {
+                FORWARD_LAYER_VALIDATE_FUNC(GpuFsaElementwiseBinarySubValidate,
+                                            reasonIfUnsupported,
+                                            infos[0],
+                                            infos[1]);
+            }
             else
             {
                 throw InvalidArgumentException("Invalid ElementwiseBinary BinaryOperation operation.");
diff --git a/src/backends/gpuFsa/layers/CMakeLists.txt b/src/backends/gpuFsa/layers/CMakeLists.txt
index bba795eedb..182a32c121 100644
--- a/src/backends/gpuFsa/layers/CMakeLists.txt
+++ b/src/backends/gpuFsa/layers/CMakeLists.txt
@@ -10,6 +10,8 @@ list(APPEND armnnGpuFsaBackendLayerValidators_sources
         GpuFsaDepthwiseConvolution2d.hpp
         GpuFsaElementwiseBinaryAdd.cpp
         GpuFsaElementwiseBinaryAdd.hpp
+        GpuFsaElementwiseBinarySub.cpp
+        GpuFsaElementwiseBinarySub.hpp
     )
 
 add_library(armnnGpuFsaBackendLayerValidators OBJECT ${armnnGpuFsaBackendLayerValidators_sources})
diff --git a/src/backends/gpuFsa/layers/GpuFsaElementwiseBinarySub.cpp b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinarySub.cpp
new file mode 100644
index 0000000000..4e7eb77190
--- /dev/null
+++ b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinarySub.cpp
@@ -0,0 +1,82 @@
+//
+// 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/operators/GpuSub.h>
+#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h>
+#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadContext.h>
+#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h>
+
+using namespace arm_compute::experimental::dynamic_fusion;
+
+namespace armnn
+{
+
+arm_compute::Status GpuFsaElementwiseBinarySubValidate(const TensorInfo& input0,
+                                                       const TensorInfo& input1)
+{
+    using namespace armcomputetensorutils;
+
+    // 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)
+{
+    using namespace armcomputetensorutils;
+
+    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
new file mode 100644
index 0000000000..59d8189f1f
--- /dev/null
+++ b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinarySub.hpp
@@ -0,0 +1,23 @@
+//
+// 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
+{
+
+    using namespace arm_compute::experimental::dynamic_fusion;
+
+    arm_compute::Status GpuFsaElementwiseBinarySubValidate(const TensorInfo& input0,
+                                                           const TensorInfo& input1);
+
+    void GpuFsaElementwiseBinarySubCreateOp(GpuFsaPreCompiledBlob* blob,
+                                            const TensorInfo& input0,
+                                            const TensorInfo& input1);
+}
\ No newline at end of file
diff --git a/src/backends/gpuFsa/test/GpuFsaEndToEndTests.cpp b/src/backends/gpuFsa/test/GpuFsaEndToEndTests.cpp
index 9972b24637..26c7cb8d63 100644
--- a/src/backends/gpuFsa/test/GpuFsaEndToEndTests.cpp
+++ b/src/backends/gpuFsa/test/GpuFsaEndToEndTests.cpp
@@ -45,4 +45,15 @@ TEST_CASE("GpuFsaElementwiseBinaryAddTestFloat16")
     ElementwiseBinarySimple3DEndToEnd<armnn::DataType::Float16>(gpuFsaDefaultBackends, BinaryOperation::Add);
 }
 
+// ElementwiseBinary Sub
+TEST_CASE("GpuFsaElementwiseBinarySubTestFloat32")
+{
+    ElementwiseBinarySimple3DEndToEnd<armnn::DataType::Float32>(gpuFsaDefaultBackends, BinaryOperation::Sub);
+}
+
+TEST_CASE("GpuFsaElementwiseBinarySubTestFloat16")
+{
+    ElementwiseBinarySimple3DEndToEnd<armnn::DataType::Float16>(gpuFsaDefaultBackends, BinaryOperation::Sub);
+}
+
 }
diff --git a/src/backends/gpuFsa/test/GpuFsaLayerSupportTests.cpp b/src/backends/gpuFsa/test/GpuFsaLayerSupportTests.cpp
index 49ddadea59..9d4b3b9367 100644
--- a/src/backends/gpuFsa/test/GpuFsaLayerSupportTests.cpp
+++ b/src/backends/gpuFsa/test/GpuFsaLayerSupportTests.cpp
@@ -81,4 +81,24 @@ TEST_CASE("IsLayerSupportedGpuFsaElementWiseBinaryAdd")
     CHECK(supported);
 }
 
+TEST_CASE("IsLayerSupportedGpuFsaElementWiseBinarySub")
+{
+    TensorInfo input0Info({ 2, 2 }, DataType::Float32);
+    TensorInfo input1Info({ 2, 2 }, DataType::Float32);
+    TensorInfo outputInfo({ 2, 2 }, DataType::Float32);
+
+    ElementwiseBinaryDescriptor desc;
+    desc.m_Operation = BinaryOperation::Sub;
+
+    GpuFsaLayerSupport supportChecker;
+    std::string reasonIfNotSupported;
+    auto supported = supportChecker.IsLayerSupported(LayerType::ElementwiseBinary,
+                                                     {input0Info, input1Info, outputInfo},
+                                                     desc,
+                                                     EmptyOptional(),
+                                                     EmptyOptional(),
+                                                     reasonIfNotSupported);
+    CHECK(supported);
+}
+
 }
\ No newline at end of file
diff --git a/src/backends/gpuFsa/test/GpuFsaOptimizedNetworkTests.cpp b/src/backends/gpuFsa/test/GpuFsaOptimizedNetworkTests.cpp
index 673a52a5a0..4b27f3bff0 100644
--- a/src/backends/gpuFsa/test/GpuFsaOptimizedNetworkTests.cpp
+++ b/src/backends/gpuFsa/test/GpuFsaOptimizedNetworkTests.cpp
@@ -182,4 +182,52 @@ TEST_CASE("ElementwiseBinaryAddSupportedOptimizedNetwork")
                         &IsLayerOfType<OutputLayer>));
 }
 
+TEST_CASE("ElementwiseBinarySubSupportedOptimizedNetwork")
+{
+    using namespace armnn;
+
+    const float qScale = 1.0f;
+    const int32_t qOffset = 0;
+
+    const TensorShape& input1Shape  = { 2, 2, 2 };
+    const TensorShape& input2Shape  = { 2, 2, 2 };
+    const TensorShape& outputShape  = { 2, 2, 2 };
+
+    TensorInfo input1TensorInfo(input1Shape, DataType::Float32, qScale, qOffset, true);
+    TensorInfo input2TensorInfo(input2Shape, DataType::Float32, qScale, qOffset, true);
+    TensorInfo outputTensorInfo(outputShape, DataType::Float32, qScale, qOffset);
+
+    IRuntime::CreationOptions options;
+    IRuntimePtr runtime(IRuntime::Create(options));
+    INetworkPtr network(INetwork::Create());
+
+    IConnectableLayer* input1 = network->AddInputLayer(0, "input0");
+    IConnectableLayer* input2 = network->AddInputLayer(1, "input1");
+
+    ElementwiseBinaryDescriptor desc;
+    desc.m_Operation = BinaryOperation::Sub;
+
+    IConnectableLayer* elementwiseBinaryLayer = network->AddElementwiseBinaryLayer(desc, "elementwiseBinary");
+    IConnectableLayer* output = network->AddOutputLayer(2, "output");
+
+    Connect(input1, elementwiseBinaryLayer, input1TensorInfo, 0, 0);
+    Connect(input2, elementwiseBinaryLayer, input2TensorInfo, 0, 1);
+    Connect(elementwiseBinaryLayer, output, outputTensorInfo, 0, 0);
+
+    std::vector<BackendId> backends = { "GpuFsa" };
+
+    OptimizerOptionsOpaque optimizedOptions;
+    IOptimizedNetworkPtr optNet = Optimize(*network, backends, runtime->GetDeviceSpec(), optimizedOptions);
+    CHECK(optNet);
+
+    Graph& graph = GetGraphForTesting(optNet.get());
+
+    // Check graph layer sequence to ensure that the network has been replaced with a PreCompiledLayer
+    CHECK(CheckSequence(graph.cbegin(), graph.cend(),
+                        &IsLayerOfType<InputLayer>,
+                        &IsLayerOfType<InputLayer>,
+                        &IsLayerOfType<PreCompiledLayer>,
+                        &IsLayerOfType<OutputLayer>));
+}
+
 }
\ No newline at end of file
diff --git a/src/backends/neon/test/NeonEndToEndTests.cpp b/src/backends/neon/test/NeonEndToEndTests.cpp
index 58c55e5d72..37f6d3845b 100644
--- a/src/backends/neon/test/NeonEndToEndTests.cpp
+++ b/src/backends/neon/test/NeonEndToEndTests.cpp
@@ -177,6 +177,15 @@ TEST_CASE("NeonSquaredDifferenceEndToEndTestUint8")
     ElementwiseBinarySimpleEndToEnd<armnn::DataType::QAsymmU8>(neonDefaultBackends, BinaryOperation::SqDiff);
 }
 
+TEST_CASE("NeonSubtractionEndToEndFloat32Simple3DTest")
+{
+    ElementwiseBinarySimple3DEndToEnd<armnn::DataType::Float32>(neonDefaultBackends, BinaryOperation::Sub);
+}
+TEST_CASE("NeonSubtractionEndToEndFloat16Simple3DTest")
+{
+    ElementwiseBinarySimple3DEndToEnd<armnn::DataType::Float16>(neonDefaultBackends, BinaryOperation::Sub);
+}
+
 // Batch Mat Mul
 TEST_CASE("NeonBatchMatMulEndToEndFloat32Test")
 {
diff --git a/src/backends/reference/test/RefEndToEndTests.cpp b/src/backends/reference/test/RefEndToEndTests.cpp
index 185deb685d..c09304e4d1 100644
--- a/src/backends/reference/test/RefEndToEndTests.cpp
+++ b/src/backends/reference/test/RefEndToEndTests.cpp
@@ -1819,6 +1819,14 @@ TEST_CASE("RefSubEndToEndTestUint8")
 {
     ElementwiseBinarySimpleEndToEnd<armnn::DataType::QAsymmU8>(defaultBackends, BinaryOperation::Sub);
 }
+TEST_CASE("RefSubEndToEndTestFloat32Simple3D")
+{
+    ElementwiseBinarySimple3DEndToEnd<armnn::DataType::Float32>(defaultBackends, BinaryOperation::Sub);
+}
+TEST_CASE("RefSubEndToEndTestFloat16Simple3D")
+{
+    ElementwiseBinarySimple3DEndToEnd<armnn::DataType::Float16>(defaultBackends, BinaryOperation::Sub);
+}
 TEST_CASE("RefMaximumEndToEndTestFloat32")
 {
     ElementwiseBinarySimpleEndToEnd<armnn::DataType::Float32>(defaultBackends, BinaryOperation::Maximum);