cuda.hpp

/*
 * SPDX-FileCopyrightText: Copyright (c) 2022-2025, NVIDIA CORPORATION.
 * SPDX-License-Identifier: Apache-2.0
 */
#pragma once
 
#include <any>
#include <functional>
 
#include <cuda.h>
#include <kvikio/shim/utils.hpp>
#include <stdexcept>
 
namespace kvikio {
 
namespace detail {
class AnyCallable {
 private:
  std::any _callable;
 
 public:
  template <typename Callable>
  void set(Callable&& c)
  {
    _callable = std::function(c);
  }
 
  void reset() { _callable.reset(); }
 
  template <typename... Args>
  CUresult operator()(Args... args)
  {
    using T = std::function<CUresult(Args...)>;
    if (!_callable.has_value()) {
      throw std::runtime_error("No callable has been assigned to the wrapper yet.");
    }
    return std::any_cast<T&>(_callable)(args...);
  }
 
  operator bool() const { return _callable.has_value(); }
};
 
}  // namespace detail
 
class cudaAPI {
 public:
  int driver_version{0};
 
  decltype(cuInit)* Init{nullptr};
  decltype(cuMemHostAlloc)* MemHostAlloc{nullptr};
  decltype(cuMemFreeHost)* MemFreeHost{nullptr};
  decltype(cuMemHostRegister)* MemHostRegister{nullptr};
  decltype(cuMemHostUnregister)* MemHostUnregister{nullptr};
  decltype(cuMemcpyHtoDAsync)* MemcpyHtoDAsync{nullptr};
  decltype(cuMemcpyDtoHAsync)* MemcpyDtoHAsync{nullptr};
 
  detail::AnyCallable MemcpyBatchAsync{};
 
  decltype(cuPointerGetAttribute)* PointerGetAttribute{nullptr};
  decltype(cuPointerGetAttributes)* PointerGetAttributes{nullptr};
  decltype(cuCtxPushCurrent)* CtxPushCurrent{nullptr};
  decltype(cuCtxPopCurrent)* CtxPopCurrent{nullptr};
  decltype(cuCtxGetCurrent)* CtxGetCurrent{nullptr};
  decltype(cuCtxGetDevice)* CtxGetDevice{nullptr};
  decltype(cuMemGetAddressRange)* MemGetAddressRange{nullptr};
  decltype(cuGetErrorName)* GetErrorName{nullptr};
  decltype(cuGetErrorString)* GetErrorString{nullptr};
  decltype(cuDeviceGet)* DeviceGet{nullptr};
  decltype(cuDeviceGetCount)* DeviceGetCount{nullptr};
  decltype(cuDeviceGetAttribute)* DeviceGetAttribute{nullptr};
  decltype(cuDevicePrimaryCtxRetain)* DevicePrimaryCtxRetain{nullptr};
  decltype(cuDevicePrimaryCtxRelease)* DevicePrimaryCtxRelease{nullptr};
  decltype(cuStreamSynchronize)* StreamSynchronize{nullptr};
  decltype(cuStreamCreate)* StreamCreate{nullptr};
  decltype(cuStreamDestroy)* StreamDestroy{nullptr};
  decltype(cuDriverGetVersion)* DriverGetVersion{nullptr};
 
 private:
  cudaAPI();
 
 public:
  cudaAPI(cudaAPI const&)        = delete;
  void operator=(cudaAPI const&) = delete;
 
  KVIKIO_EXPORT static cudaAPI& instance();
};
 
bool is_cuda_available();
 
}  // namespace kvikio