def get_kernel()

in generate_kernels.py [0:0]

• 49 lines of code
• 15 McCabe index (conditional complexity)

def get_kernel(kernel):

    major, minor = 5, 0
    arch = "sm_%d%d" % (major, minor)
    libprefix = "PERL5LIB=%s" % maxas_dir
    maxas_i = [libprefix, os.path.join(maxas_dir, "maxas.pl") + " -i -w"]
    maxas_p = [libprefix, os.path.join(maxas_dir, "maxas.pl") + " -p"]

    sass_name   = kernel[0]
    kernel_spec = kernel_specs[sass_name]
    kernel_name = kernel_spec["basename"]
    args_spec   = str(kernel[1:])

    for opt in kernel[1:]:
        maxas_i.append("-D%s 1" % opt)
        maxas_p.append("-D%s 1" % opt)
        kernel_name += "_" + opt

    maxas_i.insert(2, "-k " + kernel_name)

    sass_name += ".sass"
    cubin_name = kernel_name + ".cubin"
    cubin_dir  = _get_cache_dir([arch, 'cubin'])

    ptx_version = "4.2" if major < 6 else "5.0"
    ptx_file   = get_ptx_file(kernel_spec, args_spec, kernel_name, arch, ptx_version)
    sass_file  = os.path.join(sass_dir, sass_name)
    cubin_file = os.path.join(cubin_dir, cubin_name)

    if not os.path.exists(sass_file):
        raise RuntimeError("Missing: %s for kernel: %s" % (sass_name, kernel_name))

    ptx_mtime   = os.path.getmtime(ptx_file)
    cubin_mtime = os.path.getmtime(cubin_file) if os.path.exists(cubin_file) else 0

    build_cubin = False
    if ptx_mtime > cubin_mtime:
        build_cubin = True

    includes = extract_includes(sass_name)
    for include, include_mtime in includes:
        if include_mtime > cubin_mtime:
            build_cubin = True
            break

    if build_cubin:
        # build the cubin and run maxas in the same command
        # we don't want the chance of a generated cubin not processed by maxas (in case user hits ^C in between these steps)
        run_command([ "ptxas -v -arch", arch, "-o", cubin_file, ptx_file, ";" ] + maxas_i + [sass_file, cubin_file])
        cubin_mtime = time.time()

    # output preprocessed and disassembled versions in debug mode
    if debug:
        pre_dir  = _get_cache_dir([arch, 'pre'])
        dump_dir = _get_cache_dir([arch, 'dump'])

        pre_file   = os.path.join(pre_dir,  kernel_name + "_pre.sass")
        dump_file  = os.path.join(dump_dir, kernel_name + "_dump.sass")
        pre_mtime  = os.path.getmtime(pre_file)  if os.path.exists(pre_file)  else 0
        dump_mtime = os.path.getmtime(dump_file) if os.path.exists(dump_file) else 0

        for include, include_mtime in includes:
            if include_mtime > pre_mtime:
                run_command(maxas_p + [sass_file, pre_file])
                break

        # if cubin_mtime > dump_mtime:
        #     run_command(["nvdisasm -c", cubin_file, ">", dump_file])

    return kernel_name, cubin_file