terminate 调用后抛出 'thrust::system::system_error' 实例：what(): after cudaGetLastError: 遇到非法内存访问
terminate called after throwing an instance of 'thrust::system::system_error' what(): after cudaGetLastError: an illegal memory access was encountered
ID: cuda/thrust-vector-alloc-failure
版本兼容性
| 版本 | 状态 | 引入 | 弃用 | 备注 |
|---|---|---|---|---|
| CUDA 11.8 | active | — | — | — |
| CUDA 12.1 | active | — | — | — |
| Thrust 1.17.2 | active | — | — | — |
| Thrust 2.1.0 | active | — | — | — |
根因分析
Thrust 算法（如 thrust::sort、thrust::reduce）内部启动 CUDA 内核；当设备向量越界访问或主机指针被传递给设备端操作时，会发生非法内存访问，导致内核崩溃，Thrust 将其报告为 system_error。
English
Thrust algorithms (e.g., thrust::sort, thrust::reduce) internally launch CUDA kernels; an illegal memory access occurs when a device vector is accessed out of bounds or a host pointer is passed to a device-side operation, causing a kernel crash that Thrust reports as a system_error.
官方文档
https://thrust.github.io/doc/group__cuda.html解决方案
-
Use cuda-memcheck or compute-sanitizer to pinpoint the exact illegal memory access location. Example: run the program with compute-sanitizer --tool memcheck ./my_program, then fix the out-of-bounds index.
-
Ensure all device vectors are properly sized and that host pointers are not accidentally passed to Thrust algorithms. Use thrust::device_vector for device data and thrust::host_vector for host data. Example: change thrust::sort(host_ptr, host_ptr+N) to thrust::sort(d_vec.begin(), d_vec.end()).
无效尝试
常见但无效的做法:
-
90% 失败
Resetting the device (cudaDeviceReset) does not fix the root cause; the illegal access will recur on the next Thrust call.
-
75% 失败
Increasing the device vector size arbitrarily may mask the out-of-bounds access but does not guarantee correctness; the real bug is in the algorithm logic.