Program Listing for File cuda.cuh

↰ Return to documentation for file (include/flamegpu/detail/cuda.cuh)

#ifndef INCLUDE_FLAMEGPU_DETAIL_CUDA_CUH_
#define INCLUDE_FLAMEGPU_DETAIL_CUDA_CUH_

#include <cuda_runtime.h>
#include <cuda.h>
#include <limits>
#include <cstdint>
#include "flamegpu/exception/FLAMEGPUException.h"

namespace flamegpu {
namespace detail {

namespace cuda {

inline cudaError_t cudaFree(void* devPtr) {
    cudaError_t status = cudaSuccess;
    // Check the pointer attribtues to detect if it is a valid ptr for the current context.
    // @todo - version which checks the device ordinal is a match for the active context too, potenitally flip-flopping the device.
    cudaPointerAttributes attributes = {};
    status = cudaPointerGetAttributes(&attributes, devPtr);
    // valid device pointers have a type of cudaMemoryTypeDevice (2), or we could check the device is non negative (and matching the current device index?), or the devicePointer will be non null.
    if (status == cudaSuccess && attributes.type == cudaMemoryTypeDevice) {
        status = ::cudaFree(devPtr);
        // Forward any status on
        return status;
    }
    // If the pointer attribtues were not correct, return cudaSuccess to avoid bad error checking.
    return cudaSuccess;
}

inline cudaError_t cudaFreeHost(void* devPtr) {
    cudaError_t status = cudaSuccess;
    // Check the pointer attribtues to detect if it is a valid ptr for the current context.
    // @todo - version which checks the device ordinal is a match for the active context too, potenitally flip-flopping the device.
    cudaPointerAttributes attributes = {};
    status = cudaPointerGetAttributes(&attributes, devPtr);
    // valid pointers allocated using cudaMallocHost have a type of cudaMemoryTypeHost
    if (status == cudaSuccess && attributes.type == cudaMemoryTypeHost) {
        status = ::cudaFreeHost(devPtr);
        // Forward on any cuda errors returned.
        return status;
    }
    // If the pointer attribtues were not correct, return cudaSuccess to avoid bad error checking.
    return cudaSuccess;
}

inline bool cuDevicePrimaryContextIsActive(int ordinal) {
    // Throw an exception if a negative device ordinal is passed
    if (ordinal < 0) {
        THROW exception::InvalidCUDAdevice("CUDA Device ordinals must be non-negative integers, in detail::cuda::cuDevicePrimaryContextIsActive()");
    }

    int deviceCount = 0;
    CUresult cuErr = CUDA_SUCCESS;
    // Get the device count, possible errors are all about bad context / state  deinitialisation, so eat those silently.
    cuErr = cuDeviceGetCount(&deviceCount);
    if (cuErr == CUDA_SUCCESS) {
        // If the device count is 0, throw.
        if (deviceCount == 0) {
            THROW exception::InvalidCUDAdevice("Error no CUDA devices found!, in detail::cuda::cuDevicePrimaryContextIsActive()");
        }
        // If the ordinal is invalid, throw
        if (ordinal >= deviceCount) {
            THROW exception::InvalidCUDAdevice("Requested CUDA device %d is not valid, only %d CUDA devices available!, in detail::cuda::cuDevicePrimaryContextIsActive()", ordinal, deviceCount);
        }
        // Get the CUdevice handle, silently dismissing any cuErrors as they are falsey
        CUdevice deviceHandle;
        cuErr = cuDeviceGet(&deviceHandle, ordinal);
        if (cuErr == CUDA_SUCCESS) {
            // Get the status of the primary context, again silently treating any cuda driver api errors returned as false-y values as they are effectively what we are checking for with this method.
            unsigned int primaryCtxflags = 0;
            int primaryCtxIsActive = false;
            cuErr = cuDevicePrimaryCtxGetState(deviceHandle, &primaryCtxflags, &primaryCtxIsActive);
            if (cuErr == CUDA_SUCCESS) {
                return primaryCtxIsActive;
            }
        }
    }
    // If we could not return the active state, return false.
    return false;
}

#if __CUDACC_VER_MAJOR__ >= 12
inline std::uint64_t cuGetCurrentContextUniqueID() {
    static_assert(sizeof(unsigned long long int) == sizeof(std::uint64_t));  // NOLINT
    CUresult cuErr = CUDA_SUCCESS;
    // Get the handle to the current context
    CUcontext ctx = NULL;
    cuErr = cuCtxGetCurrent(&ctx);
    if (cuErr == CUDA_SUCCESS) {
        // Getand return the unique id
        unsigned long long int ctxid = std::numeric_limits<std::uint64_t>::max();  // NOLINT
        cuErr = cuCtxGetId(ctx, &ctxid);
        if (cuErr == CUDA_SUCCESS) {
            return static_cast<std::uint64_t>(ctxid);
        }
    }
    return std::numeric_limits<std::uint64_t>::max();
}
#endif  // __CUDACC_VER_MAJOR__ >= 12

}  // namespace cuda
}  // namespace detail
}  // namespace flamegpu

#endif  // INCLUDE_FLAMEGPU_DETAIL_CUDA_CUH_