Thales contribution

Context

In the context of implementing three operators for the Quantization of Vision Transformers (ShiftMax, ShiftGELU and I-LayerNorm). Here is the first Thales contribution with the implementation of the forward of ShiftMax and ShiftGELU. In aidge_backend_cuda, we implement the kernel cuda for the forward of SHiftMax and ShiftGELU. We didn't make any modifications on the existing file (except cuda.hpp where we include our file).

Add files

• src/operator/ShiftGELUImpl.cpp and include/aidge/backend/cuda/operator/ShiftGELUImpl.hpp : Implementation of the forward for ShiftGELU
• src/operator/ShiftGELUImpl_CUDA_kernels.cu and include/aidge/backend/cuda/operator/ShiftGELUImpl_CUDA_kernels.hpp : Implementation of the kernel cuda for ShiftGELU
• src/operator/ShiftMaxImpl.cpp and include/aidge/backend/cuda/operator/ShiftMaxImpl.hpp : Implementation of the forward for ShiftMax
• src/operator/ShiftMaxImpl_CUDA_kernels.cu and include/aidge/backend/cuda/operator/ShiftMaxImpl_CUDA_kernels.hpp : Implementation of the kernel cuda for ShiftMax
• unit_tests/Test_ShiftGELUImpl.cpp : Unitary test for showing how work the operator ShiftGELU
• unit_tests/Test_ShiftMaxImpl.cpp : Unitary test for showing how work the operator ShiftMax

TODO

The functions that launch the Kernel (ShiftGELULaunchKernel and ShiftMaxLaunchKernel) are not yet templated. We didn't fix the error yet.

include/aidge/backend/cuda/operator/ShiftGELUImpl.hpp

+/********************************************************************************
+ * Copyright (c) 2024 Thales
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Eclipse Public License 2.0 which is available at
+ * http://www.eclipse.org/legal/epl-2.0.
+ *
+ * SPDX-License-Identifier: EPL-2.0
+ * Author: Lucas RAKOTOARIVONY, Thales Research & Technology France
+ * Date: 25.06.2024
+ *
+ ********************************************************************************/
+
+#ifndef AIDGE_BACKEND_CUDA_OPERATOR_SHIFTGELUIMPL_H_
+#define AIDGE_BACKEND_CUDA_OPERATOR_SHIFTGELUIMPL_H_
+
+#include <array>
+#include <memory>
+#include <tuple>
+#include <vector>
+
+#include <cudnn.h>
+
+#include "aidge/backend/OperatorImpl.hpp"
+#include "aidge/operator/ShiftGELU.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+
+#include "aidge/backend/cuda/utils/CudaUtils.hpp"
+
+namespace Aidge {
+class ShiftGELUImpl_cuda : public OperatorImpl {
+private:
+    std::shared_ptr<Tensor> mInputFallback;
+public:
+    ShiftGELUImpl_cuda(const ShiftGELU_Op &op) : OperatorImpl(op, "cuda") {}
+
+    static std::unique_ptr<ShiftGELUImpl_cuda> create(const ShiftGELU_Op &op) {
+        return std::make_unique<ShiftGELUImpl_cuda>(op);
+    }
+
+public:
+    void forward();
+    //~ShiftGELUImpl_cuda();
+
+private:
+    template <class T> void forward_(const Tensor& input);
+    
+};
+
+namespace {
+// add cuda backend to ShiftGELU_Op implementation registry
+static Registrar<ShiftGELU_Op> registrarShiftGELUImpl_cuda("cuda", Aidge::ShiftGELUImpl_cuda::create);
+}  // namespace
+}  // namespace Aidge
+
+#endif /* AIDGE_BACKEND_CUDA_OPERATOR_SHIFTGELUIMPL_H_ */
\ No newline at end of file