Skip to content

CUDALibrarySamples: cuFFT callback samples hang (legacy) or fail with illegal address (LTO) through the shim #596

Description

@kevmo314

Found by running NVIDIA/CUDALibrarySamples (3437729) through the shim (test/run_cuda_library_samples.sh), client CUDA 13.1 / server RTX 4090 driver 590.48. The same binaries pass when run natively on the server host.

  • r2c_c2r_legacy_callback_example never finishes (killed after 180 s).
  • r2c_c2r_lto_callback_example and r2c_c2r_lto_nvrtc_callback_example report Example failed in CUDA API on line 69 with error 700 (cudaErrorIllegalAddress) at the first execution after cufftXtSetJITCallback.

cuFFT callbacks are device function pointers / LTO-linked device code resolved by cuFFT at plan time, so the kernel ends up dereferencing a client-side address or an unlinked symbol.

Units (listed in test/cuda-library-samples/known_failures.txt until fixed): cuFFT/lto_callback_window_1d/bin/r2c_c2r_legacy_callback_example, cuFFT/lto_callback_window_1d/bin/r2c_c2r_lto_callback_example, cuFFT/lto_callback_window_1d/bin/r2c_c2r_lto_nvrtc_callback_example

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions