WebRtcXR/Plugin~/WebRTCPlugin/GraphicsDevice/Cuda/CudaContext.cpp

#include "pch.h"

#include "CudaContext.h"

#if SUPPORT_VULKAN
#include "GraphicsDevice/Vulkan/VulkanUtility.h"
#endif

#if SUPPORT_D3D11
using namespace Microsoft::WRL;
#endif

namespace unity
{
namespace webrtc
{
    static void* s_hModule = nullptr;
    static bool FindModule()
    {
        if (s_hModule)
            return true;

#if UNITY_WIN
        // dll delay load
        HMODULE module = LoadLibrary(TEXT("nvcuda.dll"));
        if (!module)
        {
            RTC_LOG(LS_INFO) << "nvcuda.dll is not found.";
            return false;
        }
        s_hModule = module;
#elif UNITY_LINUX
        s_hModule = dlopen("libcuda.so.1", RTLD_LAZY | RTLD_GLOBAL);
        if (!s_hModule)
            return false;

        // Close handle immediately because going to call `dlopen` again
        // in the implib module when cuda api called on Linux.
        dlclose(s_hModule);
        s_hModule = nullptr;
#endif
        return true;
    }

    static CUresult CheckDriverVersion()
    {
        int driverVersion = 0;
        CUresult result = cuDriverGetVersion(&driverVersion);
        if (result != CUDA_SUCCESS)
        {
            return result;
        }

        if (kRequiredDriverVersion > driverVersion)
        {
            RTC_LOG(LS_ERROR) << "CUDA driver version is not higher than the required version. " << driverVersion;
            return CUDA_ERROR_NO_DEVICE;
        }
        return CUDA_SUCCESS;
    }

    CudaContext::CudaContext()
        : m_context(nullptr)
    {
    }

    CUresult CudaContext::FindCudaDevice(const uint8_t* uuid, CUdevice* cuDevice)
    {
        bool found = FindModule();
        if (!found)
            return CUDA_ERROR_NO_DEVICE;

        CUdevice _cuDevice = 0;
        CUresult result = CUDA_SUCCESS;
        int numDevices = 0;
        result = cuDeviceGetCount(&numDevices);
        if (result != CUDA_SUCCESS)
        {
            return result;
        }
        CUuuid id = {};

        // Loop over the available devices and identify the CUdevice  corresponding to the physical device in use by
        // this Vulkan instance. This is required because there is no other way to match GPUs across API boundaries.
        for (int i = 0; i < numDevices; i++)
        {
            result = cuDeviceGet(&_cuDevice, i);
            if (result != CUDA_SUCCESS)
            {
                return result;
            }
            result = cuDeviceGetUuid(&id, _cuDevice);
            if (result != CUDA_SUCCESS)
            {
                return result;
            }

            if (!std::memcmp(static_cast<const void*>(&id), static_cast<const void*>(uuid), sizeof(CUuuid)))
            {
                if (cuDevice != nullptr)
                    *cuDevice = _cuDevice;
                return CUDA_SUCCESS;
            }
        }
        return CUDA_ERROR_NO_DEVICE;
    }

    CUresult CudaContext::Init(const VkInstance instance, VkPhysicalDevice physicalDevice)
    {
        // dll check
        bool found = FindModule();
        if (!found)
        {
            return CUDA_ERROR_NOT_FOUND;
        }

        CUresult result = CheckDriverVersion();
        if (result != CUDA_SUCCESS)
        {
            return result;
        }

        CUdevice cuDevice = 0;
        result = cuInit(0);
        if (result != CUDA_SUCCESS)
        {
            return result;
        }

        int numDevices = 0;
        result = cuDeviceGetCount(&numDevices);
        if (result != CUDA_SUCCESS)
        {
            return result;
        }

        CUuuid id = {};
        std::array<uint8_t, VK_UUID_SIZE> deviceUUID;
        if (!VulkanUtility::GetPhysicalDeviceUUIDInto(instance, physicalDevice, &deviceUUID))
        {
            return CUDA_ERROR_INVALID_DEVICE;
        }

        // Loop over the available devices and identify the CUdevice  corresponding to the physical device in use by
        // this Vulkan instance. This is required because there is no other way to match GPUs across API boundaries.
        bool foundDevice = false;
        for (int i = 0; i < numDevices; i++)
        {
            cuDeviceGet(&cuDevice, i);

            cuDeviceGetUuid(&id, cuDevice);

            if (!std::memcmp(
                    static_cast<const void*>(&id), static_cast<const void*>(deviceUUID.data()), sizeof(CUuuid)))
            {
                foundDevice = true;
                break;
            }
        }

        if (!foundDevice)
        {
            return CUDA_ERROR_NO_DEVICE;
        }

        result = cuCtxCreate(&m_context, 0, cuDevice);
        return result;
    }
    //---------------------------------------------------------------------------------------------------------------------

#if defined(SUPPORT_D3D11)
    CUresult CudaContext::Init(ID3D11Device* device)
    {
        bool found = FindModule();
        if (!found)
        {
            return CUDA_ERROR_NOT_FOUND;
        }

        CUresult result = CheckDriverVersion();
        if (result != CUDA_SUCCESS)
        {
            return result;
        }

        result = cuInit(0);
        if (result != CUDA_SUCCESS)
        {
            return result;
        }
        int numDevices = 0;
        result = cuDeviceGetCount(&numDevices);
        if (result != CUDA_SUCCESS)
        {
            return result;
        }

        ComPtr<IDXGIDevice> pDxgiDevice = nullptr;
        if (device->QueryInterface(IID_PPV_ARGS(&pDxgiDevice)) != S_OK)
        {
            return CUDA_ERROR_NO_DEVICE;
        }
        ComPtr<IDXGIAdapter> pDxgiAdapter = nullptr;
        if (pDxgiDevice->GetAdapter(&pDxgiAdapter) != S_OK)
        {
            return CUDA_ERROR_NO_DEVICE;
        }
        CUdevice dev;
        if (cuD3D11GetDevice(&dev, pDxgiAdapter.Get()) != CUDA_SUCCESS)
        {
            return CUDA_ERROR_NO_DEVICE;
        }

        result = cuCtxCreate(&m_context, 0, dev);
        return result;
    }
#endif
    //---------------------------------------------------------------------------------------------------------------------

#if defined(SUPPORT_D3D12)
    CUresult CudaContext::Init(ID3D12Device* device)
    {

        bool found = FindModule();
        if (!found)
        {
            return CUDA_ERROR_NOT_FOUND;
        }

        CUresult result = CheckDriverVersion();
        if (result != CUDA_SUCCESS)
        {
            return result;
        }

        result = cuInit(0);
        if (result != CUDA_SUCCESS)
        {
            return result;
        }

        int numDevices = 0;
        result = cuDeviceGetCount(&numDevices);
        if (result != CUDA_SUCCESS)
        {
            return result;
        }

        LUID luid = device->GetAdapterLuid();

        CUdevice cuDevice = 0;
        bool deviceFound = false;

        for (int32_t deviceIndex = 0; deviceIndex < numDevices; deviceIndex++)
        {
            result = cuDeviceGet(&cuDevice, deviceIndex);
            if (result != CUDA_SUCCESS)
            {
                return result;
            }
            char luid_[8];
            unsigned int nodeMask;
            result = cuDeviceGetLuid(luid_, &nodeMask, cuDevice);
            if (result != CUDA_SUCCESS)
            {
                return result;
            }
            if (memcmp(&luid.LowPart, luid_, sizeof(luid.LowPart)) == 0 &&
                memcmp(&luid.HighPart, luid_ + sizeof(luid.LowPart), sizeof(luid.HighPart)) == 0)
            {
                deviceFound = true;
                break;
            }
        }

        if (!deviceFound)
            return CUDA_ERROR_NO_DEVICE;
        return cuCtxCreate(&m_context, 0, cuDevice);
    }
#endif
//---------------------------------------------------------------------------------------------------------------------

// todo(kazuki):: not supported on windows
#if defined(SUPPORT_OPENGL_UNIFIED) && defined(UNITY_LINUX)
    CUresult CudaContext::InitGL()
    {
        // dll check
        bool found = FindModule();
        if (!found)
        {
            return CUDA_ERROR_NOT_FOUND;
        }

        CUresult result = CheckDriverVersion();
        if (result != CUDA_SUCCESS)
        {
            return result;
        }

        result = cuInit(0);
        if (result != CUDA_SUCCESS)
        {
            return result;
        }

        int numDevices;
        result = cuDeviceGetCount(&numDevices);
        if (CUDA_SUCCESS != result)
        {
            return result;
        }
        if (numDevices == 0)
        {
            return CUDA_ERROR_NO_DEVICE;
        }

        // TODO:: check GPU capability
        int cuDevId = 0;
        CUdevice cuDevice = 0;
        result = cuDeviceGet(&cuDevice, cuDevId);
        if (CUDA_SUCCESS != result)
        {
            return result;
        }
        result = cuCtxCreate(&m_context, 0, cuDevice);
        if (CUDA_SUCCESS != result)
        {
            return result;
        }
        return CUDA_SUCCESS;
    }
#endif
    //---------------------------------------------------------------------------------------------------------------------
    CUcontext CudaContext::GetContext() const
    {
        RTC_DCHECK(m_context);

        CUcontext current;
        if (cuCtxGetCurrent(&current) != CUDA_SUCCESS)
        {
            throw;
        }
        if (m_context == current)
        {
            return m_context;
        }
        if (cuCtxSetCurrent(m_context) != CUDA_SUCCESS)
        {
            throw;
        }
        return m_context;
    }

    void CudaContext::Shutdown()
    {
        if (m_context)
        {
            cuCtxDestroy(m_context);
            m_context = nullptr;
        }
        if (s_hModule)
        {
#if UNITY_WIN
            FreeLibrary((HMODULE)s_hModule);
#elif UNITY_LINUX
            dlclose(s_hModule);
#endif
            s_hModule = nullptr;
        }
    }

} // end namespace webrtc
} // end namespace unity