gpu_control_flow_offload_pass.h

Go to the documentation of this file.

#pragma once
 
#include <string>
 
#include "sdfg/data_flow/library_node.h"
#include "sdfg/passes/pass.h"
#include "sdfg/transformations/offloading/cuda_offload_transform.h"
#include "sdfg/transformations/offloading/rocm_offload_transform.h"
 
namespace sdfg {
namespace passes {
 
struct CUDAControlFlowOffloadTraits {
    using OffloadTransform = cuda::CUDAOffloadTransform;
 
    static std::string name() {
        return "CUDAControlFlowOffloadPass";
    }
 
    static bool is_gpu_implementation(const data_flow::ImplementationType& implementation_type);
};
 
struct ROCMControlFlowOffloadTraits {
    using OffloadTransform = rocm::ROCMOffloadTransform;
 
    static std::string name() {
        return "ROCMControlFlowOffloadPass";
    }
 
    static bool is_gpu_implementation(const data_flow::ImplementationType& implementation_type);
};

template<typename Traits>
class GPUControlFlowOffloadPass : public Pass {
private:
    bool allow_dynamic_sizes_;
 
public:
    explicit GPUControlFlowOffloadPass(bool allow_dynamic_sizes = false);
    ~GPUControlFlowOffloadPass() override = default;
 
    bool run_pass(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) override;
 
    std::string name() override {
        return Traits::name();
    }
};
 
using CUDAControlFlowOffloadPass = GPUControlFlowOffloadPass<CUDAControlFlowOffloadTraits>;
using ROCMControlFlowOffloadPass = GPUControlFlowOffloadPass<ROCMControlFlowOffloadTraits>;
 
} // namespace passes
} // namespace sdfg