colossalai

1
import platform
2

3
from ..cuda_extension import _CudaExtension
4
from ..utils import append_nvcc_threads
5

6

7
class CpuAdamX86Extension(_CudaExtension):
8
    def __init__(self):
9
        super().__init__(name="cpu_adam_x86")
10

11
    def is_hardware_available(self) -> bool:
12
        return platform.machine() == "x86_64" and super().is_hardware_available()
13

14
    def assert_hardware_compatible(self) -> None:
15
        arch = platform.machine()
16
        assert (
17
            arch == "x86_64"
18
        ), f"[extension] The {self.name} kernel requires the CPU architecture to be x86_64 but got {arch}"
19
        super().assert_hardware_compatible()
20

21
    # necessary 4 functions
22
    def sources_files(self):
23
        ret = [
24
            self.csrc_abs_path("cuda/cpu_adam.cpp"),
25
        ]
26
        return ret
27

28
    def include_dirs(self):
29
        return [self.csrc_abs_path("includes"), self.get_cuda_home_include()]
30

31
    def cxx_flags(self):
32
        extra_cxx_flags = [
33
            "-std=c++14",
34
            "-std=c++17",
35
            "-lcudart",
36
            "-lcublas",
37
            "-g",
38
            "-Wno-reorder",
39
            "-fopenmp",
40
            "-march=native",
41
        ]
42
        return ["-O3"] + self.version_dependent_macros + extra_cxx_flags
43

44
    def nvcc_flags(self):
45
        extra_cuda_flags = [
46
            "-std=c++14",
47
            "-std=c++17",
48
            "-U__CUDA_NO_HALF_OPERATORS__",
49
            "-U__CUDA_NO_HALF_CONVERSIONS__",
50
            "-U__CUDA_NO_HALF2_OPERATORS__",
51
            "-DTHRUST_IGNORE_CUB_VERSION_CHECK",
52
        ]
53
        ret = ["-O3", "--use_fast_math"] + self.version_dependent_macros + extra_cuda_flags
54
        return append_nvcc_threads(ret)
55