fix_missing_stdint

[wu-kan]

I encountered a compile error when compiling cutlass@3.8.0. It seems to be a missing header . See also #2148.

...
[ 52%] Generating nvrtc/cute/atom/mma_traits_sm90.hpp
cd /tmp/mizuno-ai/spack-stage/spack-stage-cutlass-3.8.0-taomxjkncofl27l4ed6y4azkxfgb2yll/spack-build-taomxjk/test/unit/nvrtc && /mnt/sda/2022-0526/public/wuk/v0.23.0.12.3.0.12.20241120/spack/opt/spack/linux-debian11-zen2/gcc-13.2.0/cmake-3.30.5-a4w5ckchyi3ih7t4slujnfrhr4vsc64s/bin/cmake -DFILE_IN="/tmp/mizuno-ai/spack-stage/spack-stage-cutlass-3.8.0-taomxjkncofl27l4ed6y4azkxfgb2yll/spack-src/include/cute/atom/mma_traits_sm90.hpp" -DFILE_OUT="/tmp/mizuno-ai/spack-stage/spack-stage-cutlass-3.8.0-taomxjkncofl27l4ed6y4azkxfgb2yll/spack-build-taomxjk/test/unit/nvrtc/nvrtc/cute/atom/mma_traits_sm90.hpp" -DVARIABLE_NAME="cute_atom_mma_traits_sm90_hpp" -P /tmp/mizuno-ai/spack-stage/spack-stage-cutlass-3.8.0-taomxjkncofl27l4ed6y4azkxfgb2yll/spack-src/bin2hex.cmake
/mnt/sda/2022-0526/public/wuk/v0.23.0.12.3.0.12.20241120/spack/opt/spack/linux-debian11-zen2/gcc-13.2.0/gmake-4.4.1-3ncebplrwxrbfxmcxvansf4cxl52gl6f/bin/make  -f tools/library/CMakeFiles/cutlass_library_gemm_sm80_f16_s16816gemm_u8_f16.dir/build.make tools/library/CMakeFiles/cutlass_library_gemm_sm80_f16_s16816gemm_u8_f16.dir/depend
/tmp/mizuno-ai/spack-stage/spack-stage-cutlass-3.8.0-taomxjkncofl27l4ed6y4azkxfgb2yll/spack-src/include/cute/arch/cluster_sm90.hpp(153): error: identifier "uint32_t" is undefined
             uint32_t block_rank_in_cluster()
...

[d-k-b]

Directly including cstdint is going to cause compilation issues with nvrtc on some other systems, unfortunately. Notice the other included wrapped with checks for __CUDACC_RTC__. CUTE and CUTLASS workaround this by including the CUDA std library on NVRTC. See

cutlass/include/cute/util/type_traits.hpp

Line 37 in 62750a2

#include <cuda/std/cstdint>

. You should be able to include that header to fix the issue.

[wu-kan]

Directly including cstdint is going to cause compilation issues with nvrtc on some other systems, unfortunately. Notice the other included wrapped with checks for __CUDACC_RTC__. CUTE and CUTLASS workaround this by including the CUDA std library on NVRTC. See

cutlass/include/cute/util/type_traits.hpp

Line 37 in 62750a2

#include <cuda/std/cstdint>

. You should be able to include that header to fix the issue.

Done

[d-k-b]

Thanks, I am running some additional tests internally.