Kernel 12 doesn't work with CUDA Toolkit <12

[siboehm]

Looks like there's an issue with building kernel 12 when using an old CUDA Toolkit. Have to investigate

[ardianumam]

I wanna share an error I encountered that might be similar to this issue. I tried to build the project with CUDA 11.4 and got the below error. My GPU is A5000, and based on here, the compute compatibility is 86, so I modified the corresponding cmake script to set(CUDA_COMPUTE_CAPABILITY 86). Let me know if anyone finds a way to solve the issue.

Anw, thanks much for this repo!

Scanning dependencies of target sgemm
[ 71%] Building CUDA object CMakeFiles/sgemm.dir/sgemm.cu.o
[ 85%] Building CUDA object CMakeFiles/sgemm.dir/src/runner.cu.o
In file included from /usr/local/cuda-11.4/bin/../targets/x86_64-linux/include/cuda/barrier:10,
                 from /disk2/aumam/dev/cuda/cuda-course/07_Faster_Matmul/SGEMM_CUDA/src/kernels/12_kernel_double_buffering.cuh:9,
                 from /disk2/aumam/dev/cuda/cuda-course/07_Faster_Matmul/SGEMM_CUDA/src/kernels.cuh:5,
                 from /disk2/aumam/dev/cuda/cuda-course/07_Faster_Matmul/SGEMM_CUDA/src/runner.cu:1:
/usr/local/cuda-11.4/bin/../targets/x86_64-linux/include/cuda/std/barrier:11:4: error: #error "CUDA synchronization primitives are only supported for sm_70 and up."
   11 | #  error "CUDA synchronization primitives are only supported for sm_70 and up."
      |    ^~~~~
In file included from /usr/local/cuda-11.4/bin/../targets/x86_64-linux/include/cuda/std/atomic:49,
                 from /usr/local/cuda-11.4/bin/../targets/x86_64-linux/include/cuda/std/barrier:17,
                 from /usr/local/cuda-11.4/bin/../targets/x86_64-linux/include/cuda/barrier:10,
                 from /disk2/aumam/dev/cuda/cuda-course/07_Faster_Matmul/SGEMM_CUDA/src/kernels/12_kernel_double_buffering.cuh:9,
                 from /disk2/aumam/dev/cuda/cuda-course/07_Faster_Matmul/SGEMM_CUDA/src/kernels.cuh:5,
                 from /disk2/aumam/dev/cuda/cuda-course/07_Faster_Matmul/SGEMM_CUDA/src/runner.cu:1:
/usr/local/cuda-11.4/bin/../targets/x86_64-linux/include/cuda/std/detail/__atomic:11:4: error: #error "CUDA atomics are only supported for sm_60 and up on *nix and sm_70 and up on Windows."
   11 | #  error "CUDA atomics are only supported for sm_60 and up on *nix and sm_70 and up on Windows."
      |    ^~~~~
make[2]: *** [CMakeFiles/sgemm.dir/build.make:76: CMakeFiles/sgemm.dir/src/runner.cu.o] Error 1
make[1]: *** [CMakeFiles/Makefile2:134: CMakeFiles/sgemm.dir/all] Error 2
make: *** [Makefile:84: all] Error 2