CUDAUtils.h

Go to the documentation of this file.

// ----------------------------------------------------------------------------
// -                        CloudViewer: www.cloudViewer.org                  -
// ----------------------------------------------------------------------------
// Copyright (c) 2018-2024 www.cloudViewer.org
// SPDX-License-Identifier: MIT
// ----------------------------------------------------------------------------
 
 
#pragma once
 
#include <Logging.h>
 
#include "cloudViewer/core/Device.h"
 
#ifdef BUILD_CUDA_MODULE
 
#include <Optional.h>
#include <cuda.h>
#include <cuda_runtime.h>
 
#include <memory>
#include <vector>
 
#define CLOUDVIEWER_FORCE_INLINE __forceinline__
#define CLOUDVIEWER_HOST_DEVICE __host__ __device__
#define CLOUDVIEWER_DEVICE __device__
#define CLOUDVIEWER_ASSERT_HOST_DEVICE_LAMBDA(type)                       \
    static_assert(__nv_is_extended_host_device_lambda_closure_type(type), \
                  #type " must be a __host__ __device__ lambda")
#define CLOUDVIEWER_CUDA_CHECK(err) \
    cloudViewer::core::__CLOUDVIEWER_CUDA_CHECK(err, __FILE__, __LINE__)
#define CLOUDVIEWER_GET_LAST_CUDA_ERROR(message) \
    __CLOUDVIEWER_GET_LAST_CUDA_ERROR(message, __FILE__, __LINE__)
#define CUDA_CALL(cuda_function, ...) cuda_function(__VA_ARGS__);
 
#else  // #ifdef BUILD_CUDA_MODULE
 
#define CLOUDVIEWER_FORCE_INLINE inline
#define CLOUDVIEWER_HOST_DEVICE
#define CLOUDVIEWER_DEVICE
#define CLOUDVIEWER_ASSERT_HOST_DEVICE_LAMBDA(type)
#define CLOUDVIEWER_CUDA_CHECK(err)
#define CLOUDVIEWER_GET_LAST_CUDA_ERROR(message)
#define CUDA_CALL(cuda_function, ...) \
    cloudViewer::utility::LogError(   \
            "Not built with CUDA, cannot call " #cuda_function);
 
#endif  // #ifdef BUILD_CUDA_MODULE
 
namespace cloudViewer {
namespace core {
 
#ifdef BUILD_CUDA_MODULE
 
class CUDAScopedDevice {
public:
    explicit CUDAScopedDevice(int device_id);
 
    explicit CUDAScopedDevice(const Device& device);
 
    ~CUDAScopedDevice();
 
    CUDAScopedDevice(const CUDAScopedDevice&) = delete;
    CUDAScopedDevice& operator=(const CUDAScopedDevice&) = delete;
 
private:
    int prev_device_id_;
};
 
class CUDAScopedStream {
private:
    struct CreateNewStreamTag {
        CreateNewStreamTag(const CreateNewStreamTag&) = delete;
        CreateNewStreamTag& operator=(const CreateNewStreamTag&) = delete;
        CreateNewStreamTag(CreateNewStreamTag&&) = delete;
        CreateNewStreamTag& operator=(CreateNewStreamTag&&) = delete;
    };
 
public:
    constexpr static CreateNewStreamTag CreateNewStream = {};
 
    explicit CUDAScopedStream(const CreateNewStreamTag&);
 
    explicit CUDAScopedStream(cudaStream_t stream);
 
    ~CUDAScopedStream();
 
    CUDAScopedStream(const CUDAScopedStream&) = delete;
    CUDAScopedStream& operator=(const CUDAScopedStream&) = delete;
 
private:
    cudaStream_t prev_stream_;
    cudaStream_t new_stream_;
    bool owns_new_stream_ = false;
};
 
class CUDAState {
public:
    static CUDAState& GetInstance();
 
    CUDAState(const CUDAState&) = delete;
    CUDAState& operator=(const CUDAState&) = delete;
 
    bool IsP2PEnabled(int src_id, int tar_id) const;
 
    bool IsP2PEnabled(const Device& src, const Device& tar) const;
 
    void ForceDisableP2PForTesting();
 
private:
    CUDAState();
 
    std::vector<std::vector<bool>> p2p_enabled_;
};
 
int GetCUDACurrentWarpSize();
 
int GetCUDACurrentDeviceTextureAlignment();
 
size_t GetCUDACurrentTotalMemSize();
 
#else
 
class CUDAScopedDevice {
public:
    explicit CUDAScopedDevice(int device_id) {}
    explicit CUDAScopedDevice(const Device& device) {}
    ~CUDAScopedDevice() {}
    CUDAScopedDevice(const CUDAScopedDevice&) = delete;
    CUDAScopedDevice& operator=(const CUDAScopedDevice&) = delete;
};
 
#endif
 
namespace cuda {
 
int DeviceCount();
 
bool IsAvailable();
 
void ReleaseCache();
 
void Synchronize();
 
void Synchronize(const Device& device);
 
void AssertCUDADeviceAvailable(int device_id);
 
void AssertCUDADeviceAvailable(const Device& device);
 
bool SupportsMemoryPools(const Device& device);
 
#ifdef BUILD_CUDA_MODULE
 
int GetDevice();
cudaStream_t GetStream();
cudaStream_t GetDefaultStream();
 
#endif
 
}  // namespace cuda
}  // namespace core
}  // namespace cloudViewer
 
// Exposed as implementation detail of macros at the end of the file.
#ifdef BUILD_CUDA_MODULE
 
namespace cloudViewer {
namespace core {
 
void __CLOUDVIEWER_CUDA_CHECK(cudaError_t err,
                              const char* file,
                              const int line);
 
void __CLOUDVIEWER_GET_LAST_CUDA_ERROR(const char* message,
                                       const char* file,
                                       const int line);
 
}  // namespace core
}  // namespace cloudViewer
 
#endif