Nvcc fatal : Value 'sm_86' is not defined for option 'gpu-architecture'

comaniac · November 9, 2021, 5:59pm

We do support -arch in NVCC utility:

apache/tvm/blob/main/python/tvm/contrib/nvcc.py#L77


    if nd.cuda(0).exist:
        # auto detect the compute arch argument
        arch = "sm_" + "".join(nd.cuda(0).compute_version.split("."))
    else:
        raise ValueError("arch(sm_xy) is not passed, and we cannot detect it from env")


file_target = path_target if path_target else temp_target
cmd = ["nvcc"]
cmd += ["--%s" % target, "-O3"]
if isinstance(arch, list):
    cmd += arch
else:
    cmd += ["-arch", arch]


if options:
    if isinstance(options, str):
        cmd += [options]
    elif isinstance(options, list):
        cmd += options
    else:
        raise ValueError("options must be str or list of str")

But currently we use a weird way to pass this argument:

github.com

apache/tvm/blob/main/python/tvm/autotvm/measure/measure_methods.py#L789




@tvm._ffi.register_func
def tvm_callback_cuda_compile(code):
    """use nvcc to generate ptx code for better optimization"""
    curr_cuda_target_arch = AutotvmGlobalScope.current.cuda_target_arch
    # e.g., target arch could be [
    #   "-gencode", "arch=compute_52,code=sm_52",
    #   "-gencode", "arch=compute_70,code=sm_70"
    # ]
    target = "fatbin" if isinstance(curr_cuda_target_arch, list) else "ptx"
    ptx = nvcc.compile_cuda(code, target=target, arch=AutotvmGlobalScope.current.cuda_target_arch)
    return ptx




def set_cuda_target_arch(arch):
    """set target architecture of nvcc compiler


    Parameters
    ----------
    arch: str or list
        The argument of nvcc -arch. (e.g. "sm_51", "sm_62")

As a result, target= "cuda -arch=xx" is not effective. Instead, you need to set the global variable as follows:

from tvm.autotvm.measure.measure_methods import set_cuda_target_arch
set_cuda_target_arch('sm_80')