[BUG]: BLOCK_THREADS won't compile with NVC++ due to non-constexpr assertion macro

[dkolsen-pgi]

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

PR #2773 change the definitions of constexpr values RegBoundScaling::BLOCK_THREADS and MemBoundScaling::BLOCK_THREADS in cub/util_arch.cuh to use the constexpr function ::cuda::ceil_div. This breaks NVC++ stdpar, because ::cuda::ceil_div is not constexpr in NVC++. ceil_div contains two _CCCL_ASSERT statements. With NVC++ in stdpar mode, _CCCL_ASSERT is defined as

#  define _CCCL_ASSERT(expression, message) \
    NV_IF_ELSE_TARGET(                      \
      NV_IS_DEVICE, (_CCCL_ASSERT_DEVICE(expression, message);), (_CCCL_ASSERT_HOST(expression, message);))

That works fine at runtime. But NV_IF_ELSE_TARGET is not constexpr. It can't be evaluated by the front end during constexpr evaluation because the front end can't know what target the code will eventually run on. This prevents ceil_dev from being a constexpr function and leads to compilation errors such as:

"/home/dolsen/work/pgi/dev/nv/Linux_x86_64/mine/compilers/include-stdpar/cub/util_arch.cuh", line 136: error: expression must have a constant value
      ::cuda::std::min(Nominal4ByteBlockThreads,
      ^
"/home/dolsen/work/pgi/dev/nv/Linux_x86_64/mine/compilers/include-stdpar/cuda/__cmath/ceil_div.h", line 63: note: cannot call non-constexpr function "__builtin_i
s_device_code" (declared implicitly)
    _CCCL_ASSERT(__a >= _Tp(0), "cuda::ceil_div: a must be non negative");
    ^

How to Reproduce

The NVHPC stdpar automated tests have many hundreds of failures due to this. Any test that uses a parallel algorithm that uses CUB will run into this.

Expected behavior

nvc++ -stdpar should work.

Reproduction link

No response

Operating System

No response

nvidia-smi output

No response

NVCC version

No response

[dkolsen-pgi]

I think the best fix is to make _CCCL_ASSERT be constexpr friendly when compiling with NVC++.

Change:

#if _CCCL_CUDA_COMPILER(NVHPC) // NVHPC needs to use NV_IF_TARGET instead of __CUDA_ARCH__
#  define _CCCL_VERIFY(expression, message) \
    NV_IF_ELSE_TARGET(                      \
      NV_IS_DEVICE, (_CCCL_ASSERT_IMPL_DEVICE(expression, message);), (_CCCL_ASSERT_IMPL_HOST(expression, message);))
#  define _CCCL_ASSERT(expression, message) \
    NV_IF_ELSE_TARGET(                      \
      NV_IS_DEVICE, (_CCCL_ASSERT_DEVICE(expression, message);), (_CCCL_ASSERT_HOST(expression, message);))

to

#if _CCCL_CUDA_COMPILER(NVHPC) // NVHPC can't have different behavior for host and device.
                               // The host version of the assert will also work in device code.
#  define _CCCL_VERIFY(expression, message) _CCCL_ASSERT_IMPL_HOST(expression, message)
#  if defined(CCCL_ENABLE_HOST_ASSERTIONS) || defined(CCCL_ENABLE_DEVICE_ASSERTIONS)
#    define _CCCL_ASSERT(expression, message) _CCCL_ASSERT_HOST(expression, message)
#  else
#    define _CCCL_ASSERT(expression, message) ((void) 0)
#  endif

It is simply not possible to have a constexpr-friendly assertion with NVC++ that is checked in host code but not device code or vice versa. The assertion needs to be checked everywhere or nowhere.

[miscco]

I am wondering whether it would be possible to wrap the condition inside an NV_IF_TARGET(NV_IS_HOST, true, condition)