/cuda-cooperative-group-bug-report

Primary LanguageCuda

Race condition reported in cg::reduce()

This is a minimum reproducible test case that causes the compute-sanitizer to report a race condition when using the cooperative groups cg::reduce() method.

Test Environment

  • GPU: Quadro RTX 5000
  • CUDA Compiler: nvcc V12.2.140
  • Host Compiler: GCC 10 (via devtoolset-10)
  • Operating System: RHEL7

Steps to reproduce

# Build it.
mkdir build && cd build
cmake .. && make

# Run the executable under the racecheck tool.
compute-sanitizer --tool=racecheck ./demo

# Optional: run with --racecheck-report all for more output.
compute-sanitizer --tool=racecheck --racecheck-report all ./demo

Expected output

========= COMPUTE-SANITIZER
========= Error: Race reported between Write access at 0xa60 in void Kernel<(int)256>(const int *, int, int *)
=========     and Read access at 0xcc0 in void Kernel<(int)256>(const int *, int, int *) [28 hazards]
=========
========= Error: Race reported between Write access at 0xd60 in void Kernel<(int)256>(const int *, int, int *)
=========     and Read access at 0xf30 in void Kernel<(int)256>(const int *, int, int *) [4 hazards]
=========
Host min = 10; Device min = 10
Host and device calculation match - we're good!
========= RACECHECK SUMMARY: 2 hazards displayed (2 errors, 0 warnings)