parse_kernel_info.py

# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import paddle


class KernelInfo:
    def __init__(self, op_type):
        self.op_type = op_type
        self.supported_dtypes = set()

    def parse_phi_dtypes(self, registered_info_list, device="GPU"):
        assert isinstance(registered_info_list, list)
        assert device in ["CPU", "GPU"]

        # registered_info_list is in format as follows:
        # ['(GPU, Undefined(AnyLayout), float32)', '(GPU, Undefined(AnyLayout), float64)']
        for kernel_str in registered_info_list:
            kernel_strs = (
                kernel_str.replace("(", "").replace(")", "").split(",")
            )
            # for GPU, the kernel type can be GPUDNN.
            if device in kernel_strs[0]:
                self.supported_dtypes.add(kernel_strs[-1].replace(" ", ""))

        # if len(self.supported_dtypes) == 0:
        #    print("-- [WARNING] No dtypes for op_type={}, device={}. Registered info: {}".format(self.op_type, device, registered_info_list))

    def parse_fluid_dtypes(self, registered_info_list, device="gpu"):
        assert isinstance(registered_info_list, list)
        assert device in ["cpu", "gpu"]

        # registered_info_list is in format as follows:
        # ['{data_type[::paddle::platform::bfloat16]; data_layout[Undefined(AnyLayout)]; place[Place(gpu:0)]; library_type[PLAIN]}', ...}']
        for kernel_str in registered_info_list:
            kernel_strs = kernel_str.split(";")
            if "place" in kernel_strs[2] and device in kernel_strs[2]:
                assert "data_type" in kernel_strs[0]
                dtype_str = kernel_strs[0].replace("{data_type[", "")
                dtype_str = dtype_str.replace("::paddle::platform::", "")
                dtype_str = dtype_str.replace("]", "")
                self.supported_dtypes.add(dtype_str)


class KernelRegistryStatistics:
    def __init__(self):
        self.num_ops_for_dtypes = {
            "all": 0,
            "float32": 0,
            "float16": 0,
            "bfloat16": 0,
        }

    def update(self, supported_dtypes):
        for dtype in supported_dtypes:
            if dtype in self.num_ops_for_dtypes.keys():
                self.num_ops_for_dtypes[dtype] += 1
            elif dtype == "float":
                self.num_ops_for_dtypes["float32"] += 1
        self.num_ops_for_dtypes["all"] += 1

    def __str__(self):
        res = "{ "
        num_floats = int(self.num_ops_for_dtypes["float32"])
        for dtype, num in self.num_ops_for_dtypes.items():
            res += f"{dtype}: {num:4d}"
            if dtype in ["float16", "bfloat16"]:
                if num_floats != 0:
                    percent = float(self.num_ops_for_dtypes[dtype]) / float(
                        num_floats
                    )
                    res += f"({percent * 100:.2f}%)"
                else:
                    res += f"({0:.2f}%)"
            res += " "
        res += "}"
        return res


def parse_paddle_kernels(lib="phi", kernel_type="function", print_detail=False):
    assert lib in ["fluid", "phi"]

    if lib == "phi":
        assert kernel_type in ["function", "structure", "all"]
        # phi kernel type can be: function, structure, all
        kernel_infos = paddle.base.core._get_registered_phi_kernels(kernel_type)
    else:
        # fluid, phi, all
        assert kernel_type in ["fluid", "phi", "all"]
        kernel_infos = paddle.base.core._get_all_register_op_kernels(
            kernel_type
        )

    max_op_type_lengths = 0
    stats = KernelRegistryStatistics()

    kernel_info_dict = {}
    for key, value in kernel_infos.items():
        info = KernelInfo(key)
        if lib == "phi":
            info.parse_phi_dtypes(value, device="GPU")
        else:
            info.parse_fluid_dtypes(value, device="gpu")
        kernel_info_dict[key] = info
        if len(info.op_type) > max_op_type_lengths:
            max_op_type_lengths = len(info.op_type)
        stats.update(info.supported_dtypes)

    if print_detail:
        print(
            "==================== lib={}, kernel_type={} ====================".format(
                lib, kernel_type
            )
        )
        print(
            "{} : {}".format(
                "op_type".ljust(max_op_type_lengths + 4),
                "supported_dtypes for GPU",
            )
        )
        for key, value in sorted(kernel_info_dict.items()):
            print(
                "{} : {}".format(
                    value.op_type.ljust(max_op_type_lengths + 4),
                    value.supported_dtypes,
                )
            )
        print("")
    return stats


def main(lib):
    assert lib in ["fluid", "phi"]

    print_detail = False
    if lib == "phi":
        phi_function_kernels_stats = parse_paddle_kernels(
            lib, "function", print_detail=False
        )
        phi_structure_kernels_stats = parse_paddle_kernels(
            lib, "structure", print_detail=False
        )
        phi_all_kernels_stats = parse_paddle_kernels(
            lib, "all", print_detail=print_detail
        )
        print(
            "==================================   phi kernels summary   =================================="
        )
        print(f"phi function  kernels : {phi_function_kernels_stats}")
        print(f"phi structure kernels : {phi_structure_kernels_stats}")
        print(f"phi all       kernels : {phi_all_kernels_stats}")
        print("")
    else:
        fluid_ops_stats = parse_paddle_kernels(lib, "fluid", print_detail=False)
        phi_ops_stats = parse_paddle_kernels(lib, "phi", print_detail=False)
        all_ops_stats = parse_paddle_kernels(
            lib, "all", print_detail=print_detail
        )
        print(
            "================================== fluid operators summary =================================="
        )
        print(f"fluid operators : {fluid_ops_stats}")
        print(f"phi   operators : {phi_ops_stats}")
        print(f"all   operators : {all_ops_stats}")
        print("")


main(lib="fluid")
main(lib="phi")