[BUG]: Static assertion failure in atomic_ref for double scan on sm_120 (<=16 vs <16 size check mismatch)

[zkhatami]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Compile-time Error

Component

CUB

Describe the bug

This test fails when targeting sm_120 architecture:

#include <execution>
#include <numeric>
#include <vector>

int32_t
main()
{
  std::vector<double> in(1000);
  std::vector<double> out(1000);
  auto orr = std::inclusive_scan(std::execution::par, in.begin(), in.end(), out.begin());
  return 0;
}

it fails with:

cuda/std/__atomic/functions/host.h", line 85: error: static assertion failed with "atomic_ref<T> where sizeof(T) > 8 is not supported on this system."
    _LIBCUDACXX_INT128_WARN(_Tp)
    ^
cuda/std/__atomic/functions/host.h", line 85: note: the final comparison was 16 < 16
    _LIBCUDACXX_INT128_WARN(_Tp)
    ^

On sm_120, we use Policy1000, which includes the WarpspeedPolicy structure. For type double, scan_use_warpspeed evaluates to true, leading to the instantiation of atomic_ref<tile_state_t<double>>.
The function storeTileAggregate in cub/detail/warpspeed/look_ahead.cuh contains:

  if constexpr (sizeof(tile_state_t<AccumT>) <= 16

Meanwhile, cuda/std/__atomic/functions/host.h enforces:

static_assert(sizeof(_Tp) < 16, "atomic_ref<T> where sizeof(T) > 8 is not supported on this system.");

Because storeTileAggregate allows sizeof(tile_state_t<AccumT>) == 16, it takes the atomic_ref path.
However, atomic_ref later asserts against sizeof(T) < 16, causing the build to fail.
This results in an inconsistent condition between these two checks (<= 16 vs < 16), leading to a static assertion failure when building for sm_120.

How to Reproduce

nvc++ -stdpar --c++17  test_scan.cpp

Expected behavior

Compilation should succeed, or the size conditions should be made consistent to prevent this mismatch.

Reproduction link

No response

Operating System

No response

nvidia-smi output

No response

NVCC version

No response

[github-actions]

Hi @zkhatami!

Thanks for submitting this issue - the CCCL team has been notified and we'll get back to you as soon as we can!
In the mean time, feel free to add any relevant information to this issue.

[bernhardmgruber]

Thx for the bug report. I think

  if constexpr (sizeof(tile_state_t<AccumT>) <= 16

is correct for the use of __nv_atomic_store etc. but in case we don't have _CUDACC_DEVICE_ATOMIC_BUILTINS__, we probably must change the condition to < 16.

@wmaxey why does cuda::atomic_ref not support 16-byte wide atomics on Blackwell?