cuda_check.hpp File Reference

PULSATRIX_CUDA_CHECK – converts a CUDA runtime failure into a thrown C++ exception at the DeviceBackend boundary, per gpu_backend_programming/context_gpu_cuda_kernel_mechanics.md's Error Checking pattern. Only compiled when PULSATRIX_ENABLE_CUDA is set. More...

#include <cuda_runtime.h>
#include <stdexcept>
#include <string>

Include dependency graph for cuda_check.hpp:

Go to the source code of this file.

Macros
#define	PULSATRIX_CUDA_CHECK(call)
	Checks a CUDA runtime call's cudaError_t result; throws std::runtime_error with the CUDA error string and source location if it did not succeed.

Detailed Description

PULSATRIX_CUDA_CHECK – converts a CUDA runtime failure into a thrown C++ exception at the DeviceBackend boundary, per gpu_backend_programming/context_gpu_cuda_kernel_mechanics.md's Error Checking pattern. Only compiled when PULSATRIX_ENABLE_CUDA is set.

Macro Definition Documentation

PULSATRIX_CUDA_CHECK

#define PULSATRIX_CUDA_CHECK

(

call

)

Value:

    do {                                                                                          \
        cudaError_t pulsatrix_cuda_check_result = (call);                                               \
        if (pulsatrix_cuda_check_result != cudaSuccess) {                                                \
            throw std::runtime_error(std::string("CUDA error: ") +                                  \
                                      cudaGetErrorString(pulsatrix_cuda_check_result) + " at " __FILE__ + \
                                      ":" + std::to_string(__LINE__));                                 \
        }                                                                                                \
    } while (0)

Checks a CUDA runtime call's cudaError_t result; throws std::runtime_error with the CUDA error string and source location if it did not succeed.

Note

Kernel launches are asynchronous – this macro catches launch-configuration errors immediately via cudaGetLastError(), but in-kernel runtime errors only surface at the next synchronizing call. Route every allocate/copy/kernel-launch/sync call in CUDABackend through this macro, including a cudaGetLastError() check right after every kernel launch.