IVGCVSW-7550 GpuFsa Op: Add ElementWiseBinary Operator ADD
* Adding support for Gpu Add operator
* Added tests for layer support, end to end and optimization
Signed-off-by: Tracy Narine <tracy.narine@arm.com>
Change-Id: Ie9328d269c5c0ff60a7e10133b728ac9265033af
diff --git a/src/backends/gpuFsa/layers/CMakeLists.txt b/src/backends/gpuFsa/layers/CMakeLists.txt
index c174c51..bba795e 100644
--- a/src/backends/gpuFsa/layers/CMakeLists.txt
+++ b/src/backends/gpuFsa/layers/CMakeLists.txt
@@ -8,6 +8,8 @@
GpuFsaConvolution2d.hpp
GpuFsaDepthwiseConvolution2d.cpp
GpuFsaDepthwiseConvolution2d.hpp
+ GpuFsaElementwiseBinaryAdd.cpp
+ GpuFsaElementwiseBinaryAdd.hpp
)
add_library(armnnGpuFsaBackendLayerValidators OBJECT ${armnnGpuFsaBackendLayerValidators_sources})
diff --git a/src/backends/gpuFsa/layers/GpuFsaDepthwiseConvolution2d.cpp b/src/backends/gpuFsa/layers/GpuFsaDepthwiseConvolution2d.cpp
index 01a36f2..dd55d4d 100644
--- a/src/backends/gpuFsa/layers/GpuFsaDepthwiseConvolution2d.cpp
+++ b/src/backends/gpuFsa/layers/GpuFsaDepthwiseConvolution2d.cpp
@@ -18,7 +18,6 @@
#include <arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h>
#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadContext.h>
-#include <src/dynamic_fusion/sketch/gpu/GpuWorkloadContextImpl.h>
#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuDepthwiseConv2d.h>
#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h>
diff --git a/src/backends/gpuFsa/layers/GpuFsaElementwiseBinaryAdd.cpp b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinaryAdd.cpp
new file mode 100644
index 0000000..bc8b370
--- /dev/null
+++ b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinaryAdd.cpp
@@ -0,0 +1,86 @@
+//
+// Copyright © 2024 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "GpuFsaElementwiseBinaryAdd.hpp"
+
+#include <armnn/Types.hpp>
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <arm_compute/core/CL/CLKernelLibrary.h>
+
+#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuAdd.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 GpuFsaElementwiseBinaryAddValidate(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 GpuAdd::validate_op(sketch, inputInfo0, inputInfo1);
+}
+
+void GpuFsaElementwiseBinaryAddCreateOp(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 = 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
new file mode 100644
index 0000000..8221f0e
--- /dev/null
+++ b/src/backends/gpuFsa/layers/GpuFsaElementwiseBinaryAdd.hpp
@@ -0,0 +1,26 @@
+//
+// Copyright © 2024 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+#pragma once
+
+#include <armnn/Descriptors.hpp>
+#include <armnn/Tensor.hpp>
+
+#include <arm_compute/core/Error.h>
+#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h>
+#include <gpuFsa/GpuFsaBackend.hpp>
+
+namespace armnn
+{
+
+ using namespace arm_compute::experimental::dynamic_fusion;
+
+ 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