mindspore2022/tests/st/dump/test_data_dump.py

# Copyright 2020-2021 Huawei Technologies Co., Ltd
#
# 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 os
import sys
import tempfile
import time
import shutil
import glob
import csv
from importlib import import_module
from pathlib import Path
import numpy as np
import pytest
import mindspore.context as context
import mindspore.nn as nn
from mindspore import Tensor
from mindspore.ops import operations as P
from mindspore.nn import Cell
from mindspore.nn import Dense
from mindspore.nn import SoftmaxCrossEntropyWithLogits
from mindspore.nn import Momentum
from mindspore.nn import TrainOneStepCell
from mindspore.nn import WithLossCell
from dump_test_utils import generate_dump_json, generate_dump_json_with_overflow, \
    generate_statistic_dump_json, check_dump_structure, find_nth_pos
from tests.security_utils import security_off_wrap


class Net(nn.Cell):
    def __init__(self):
        super(Net, self).__init__()
        self.add = P.Add()

    def construct(self, x_, y_):
        return self.add(x_, y_)


x = np.array([[1, 2, 3], [4, 5, 6]]).astype(np.float32)
y = np.array([[7, 8, 9], [10, 11, 12]]).astype(np.float32)


def run_async_dump(test_name):
    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
    with tempfile.TemporaryDirectory(dir='/tmp') as tmp_dir:
        dump_path = os.path.join(tmp_dir, 'async_dump')
        dump_config_path = os.path.join(tmp_dir, 'async_dump.json')
        generate_dump_json(dump_path, dump_config_path, test_name)
        os.environ['MINDSPORE_DUMP_CONFIG'] = dump_config_path
        dump_file_path = os.path.join(dump_path, 'rank_0', 'Net', '0', '0')
        if os.path.isdir(dump_path):
            shutil.rmtree(dump_path)
        add = Net()
        add(Tensor(x), Tensor(y))
        for _ in range(3):
            if not os.path.exists(dump_file_path):
                time.sleep(2)
        check_dump_structure(dump_path, dump_config_path, 1, 1, 1)
        assert len(os.listdir(dump_file_path)) == 1
        del os.environ['MINDSPORE_DUMP_CONFIG']


@pytest.mark.level1
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
@security_off_wrap
def test_async_dump():
    """
    Feature: async dump on Ascend
    Description: test async dump with default file_format value
    Expectation: dump data are generated as protobuf file format (suffix with timestamp)
    """
    run_async_dump("test_async_dump")


@pytest.mark.skip(reason="wait for run package updates in Dec 01")
@pytest.mark.level1
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
@security_off_wrap
def test_async_dump_file_format():
    """
    Feature: async dump on Ascend in npy format
    Description: test async dump with file_format is configured as npy
    Expectation: dump data are generated as npy file format
    """
    run_async_dump("test_async_dump_file_format")


def run_e2e_dump():
    if sys.platform != 'linux':
        return
    with tempfile.TemporaryDirectory(dir='/tmp') as tmp_dir:
        dump_path = os.path.join(tmp_dir, 'e2e_dump')
        dump_config_path = os.path.join(tmp_dir, 'e2e_dump.json')
        generate_dump_json(dump_path, dump_config_path, 'test_e2e_dump')
        os.environ['MINDSPORE_DUMP_CONFIG'] = dump_config_path
        dump_file_path = os.path.join(dump_path, 'rank_0', 'Net', '0', '0')
        if os.path.isdir(dump_path):
            shutil.rmtree(dump_path)
        add = Net()
        add(Tensor(x), Tensor(y))
        if context.get_context("device_target") == "Ascend":
            assert len(os.listdir(dump_file_path)) == 5
            output_name = "Add.Add-op*.0.0.*.output.0.DefaultFormat.npy"
        elif context.get_context("device_target") == "CPU":
            assert len(os.listdir(dump_file_path)) == 5
            output_name = "Add.Add-op*.0.0.*.output.0.DefaultFormat.npy"
        else:
            assert len(os.listdir(dump_file_path)) == 3
            output_name = "Add.Add-op*.0.0.*.output.0.DefaultFormat.npy"
        output_path = glob.glob(os.path.join(dump_file_path, output_name))[0]
        real_path = os.path.realpath(output_path)
        output = np.load(real_path)
        expect = np.array([[8, 10, 12], [14, 16, 18]], np.float32)
        assert output.dtype == expect.dtype
        assert np.array_equal(output, expect)
        for _ in range(3):
            if not os.path.exists(dump_file_path):
                time.sleep(2)
        check_dump_structure(dump_path, dump_config_path, 1, 1, 1)
        del os.environ['MINDSPORE_DUMP_CONFIG']


@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
@security_off_wrap
def test_e2e_dump():
    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
    run_e2e_dump()


@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
@security_off_wrap
def test_e2e_dump_with_hccl_env():
    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
    os.environ["RANK_TABLE_FILE"] = "invalid_file.json"
    os.environ["RANK_ID"] = "4"
    run_e2e_dump()
    del os.environ['RANK_TABLE_FILE']
    del os.environ['RANK_ID']


@pytest.mark.level0
@pytest.mark.platform_x86_cpu
@pytest.mark.env_onecard
@security_off_wrap
def test_cpu_e2e_dump():
    context.set_context(mode=context.GRAPH_MODE, device_target="CPU")
    run_e2e_dump()


@pytest.mark.level0
@pytest.mark.platform_x86_cpu
@pytest.mark.env_onecard
@security_off_wrap
def test_cpu_e2e_dump_with_hccl_set():
    context.set_context(mode=context.GRAPH_MODE, device_target="CPU")
    os.environ["RANK_TABLE_FILE"] = "invalid_file.json"
    os.environ["RANK_ID"] = "4"
    run_e2e_dump()
    del os.environ['RANK_TABLE_FILE']
    del os.environ['RANK_ID']


@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
@security_off_wrap
def test_gpu_e2e_dump():
    context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
    run_e2e_dump()


@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
@security_off_wrap
def test_gpu_e2e_dump_with_hccl_set():
    context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
    os.environ["RANK_TABLE_FILE"] = "invalid_file.json"
    os.environ["RANK_ID"] = "4"
    run_e2e_dump()
    del os.environ['RANK_TABLE_FILE']
    del os.environ['RANK_ID']


class ReluReduceMeanDenseRelu(Cell):
    def __init__(self, kernel, bias, in_channel, num_class):
        super().__init__()
        self.relu = P.ReLU()
        self.mean = P.ReduceMean(keep_dims=False)
        self.dense = Dense(in_channel, num_class, kernel, bias)

    def construct(self, x_):
        x_ = self.relu(x_)
        x_ = self.mean(x_, (2, 3))
        x_ = self.dense(x_)
        x_ = self.relu(x_)
        return x_


@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
@security_off_wrap
def test_async_dump_net_multi_layer_mode1():
    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
    with tempfile.TemporaryDirectory(dir='/tmp') as tmp_dir:
        dump_path = os.path.join(tmp_dir, 'async_dump_net_multi_layer_mode1')
        json_file_path = os.path.join(tmp_dir, "test_async_dump_net_multi_layer_mode1.json")
        generate_dump_json(dump_path, json_file_path, 'test_async_dump_net_multi_layer_mode1')
        os.environ['MINDSPORE_DUMP_CONFIG'] = json_file_path
        weight = Tensor(np.ones((1000, 2048)).astype(np.float32))
        bias = Tensor(np.ones((1000,)).astype(np.float32))
        net = ReluReduceMeanDenseRelu(weight, bias, 2048, 1000)
        criterion = SoftmaxCrossEntropyWithLogits(sparse=False)
        optimizer = Momentum(learning_rate=0.1, momentum=0.1,
                             params=filter(lambda x: x.requires_grad, net.get_parameters()))
        net_with_criterion = WithLossCell(net, criterion)
        train_network = TrainOneStepCell(net_with_criterion, optimizer)
        train_network.set_train()
        inputs = Tensor(np.random.randn(32, 2048, 7, 7).astype(np.float32))
        label = Tensor(np.zeros(shape=(32, 1000)).astype(np.float32))
        net_dict = train_network(inputs, label)
        dump_file_path = os.path.join(dump_path, 'rank_0', 'test', '0', '0')
        dump_file_name = list(Path(dump_file_path).rglob("*SoftmaxCrossEntropyWithLogits*"))[0]
        dump_file_full_path = os.path.join(dump_file_path, dump_file_name)
        npy_path = os.path.join(dump_path, "npy_files")
        if os.path.exists(npy_path):
            shutil.rmtree(npy_path)
        os.mkdir(npy_path)
        tool_path_search_list = list(Path('/usr/local/Ascend').rglob('msaccucmp.py*'))
        if tool_path_search_list:
            converter = import_module("mindspore.offline_debug.convert_async")
            converter.AsyncDumpConverter([dump_file_full_path], npy_path).convert_files()
            npy_result_file = list(Path(npy_path).rglob("*output.0.*.npy"))[0]
            dump_result = np.load(os.path.join(npy_path, npy_result_file))
            for index, value in enumerate(net_dict):
                assert value.asnumpy() == dump_result[index]
        else:
            print('Failed to find hisi convert tools: msaccucmp.py or msaccucmp.pyc.')
        del os.environ['MINDSPORE_DUMP_CONFIG']


@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
@security_off_wrap
def test_dump_with_diagnostic_path():
    """
    Test e2e dump when path is not set (set to empty) in dump json file and MS_DIAGNOSTIC_DATA_PATH is set.
    Data is expected to be dumped into MS_DIAGNOSTIC_DATA_PATH/debug_dump.
    """
    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
    with tempfile.TemporaryDirectory(dir='/tmp') as tmp_dir:
        dump_config_path = os.path.join(tmp_dir, 'e2e_dump.json')
        generate_dump_json('', dump_config_path, 'test_e2e_dump')
        os.environ['MINDSPORE_DUMP_CONFIG'] = dump_config_path
        diagnose_path = os.path.join(tmp_dir, 'e2e_dump')
        os.environ['MS_DIAGNOSTIC_DATA_PATH'] = diagnose_path
        dump_file_path = os.path.join(diagnose_path, 'debug_dump', 'rank_0', 'Net', '0', '0')
        if os.path.isdir(diagnose_path):
            shutil.rmtree(diagnose_path)
        add = Net()
        add(Tensor(x), Tensor(y))
        assert len(os.listdir(dump_file_path)) == 5
        del os.environ['MINDSPORE_DUMP_CONFIG']
        del os.environ['MS_DIAGNOSTIC_DATA_PATH']


def run_e2e_dump_execution_graph():
    """Run e2e dump and check execution order."""
    if sys.platform != 'linux':
        return
    with tempfile.TemporaryDirectory(dir='/tmp') as tmp_dir:
        dump_path = os.path.join(tmp_dir, 'e2e_dump_exe_graph')
        dump_config_path = os.path.join(tmp_dir, 'e2e_dump.json')
        generate_dump_json(dump_path, dump_config_path, 'test_e2e_dump')
        os.environ['MINDSPORE_DUMP_CONFIG'] = dump_config_path
        if os.path.isdir(dump_path):
            shutil.rmtree(dump_path)
        add = Net()
        add(Tensor(x), Tensor(y))
        exe_graph_path = os.path.join(dump_path, 'rank_0', 'execution_order')
        assert len(os.listdir(exe_graph_path)) == 2
        del os.environ['MINDSPORE_DUMP_CONFIG']


@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
@security_off_wrap
def test_dump_with_execution_graph():
    """Test dump with execution graph on GPU."""
    context.set_context(mode=context.GRAPH_MODE, device_target='GPU')
    run_e2e_dump_execution_graph()


def run_overflow_dump():
    """Run async dump and generate overflow"""
    if sys.platform != 'linux':
        return
    overflow_x = np.array([60000, 60000]).astype(np.float16)
    with tempfile.TemporaryDirectory(dir='/tmp') as tmp_dir:
        dump_path = os.path.join(tmp_dir, 'overflow_dump')
        dump_config_path = os.path.join(tmp_dir, 'overflow_dump.json')
        generate_dump_json_with_overflow(dump_path, dump_config_path, 'test_async_dump', 3)
        os.environ['MINDSPORE_DUMP_CONFIG'] = dump_config_path
        if os.path.isdir(dump_path):
            shutil.rmtree(dump_path)
        add = Net()
        add(Tensor(overflow_x), Tensor(overflow_x))
        exe_graph_path = os.path.join(dump_path, 'rank_0', 'Net', '0', '0')
        for _ in range(5):
            if not os.path.exists(exe_graph_path):
                time.sleep(2)
        check_dump_structure(dump_path, dump_config_path, 1, 1, 1)
        # check if overflow dump generate exact two files, and the naming format
        assert len(os.listdir(exe_graph_path)) == 2
        output_path = glob.glob(os.path.join(exe_graph_path, "Add.Default_Add-op0.*.*.*"))[0]
        overflow_path = glob.glob(os.path.join(exe_graph_path, "Opdebug.Node_OpDebug.*.*.*"))[0]
        assert output_path
        assert overflow_path
        # check if generated files have matching task and stream id
        output_file_name = os.path.split(output_path)
        overflow_file_name = os.path.split(overflow_path)
        output_second_dot_pos = find_nth_pos(output_file_name[1], ".", 2)
        output_third_dot_pos = find_nth_pos(output_file_name[1], ".", 3)
        output_fourth_dot_pos = find_nth_pos(output_file_name[1], ".", 4)
        output_task_id = output_file_name[1][output_second_dot_pos+1:output_third_dot_pos]
        output_stream_id = output_file_name[1][output_third_dot_pos+1:output_fourth_dot_pos]

        overflow_second_dot_pos = find_nth_pos(overflow_file_name[1], ".", 2)
        overflow_third_dot_pos = find_nth_pos(overflow_file_name[1], ".", 3)
        overflow_fourth_dot_pos = find_nth_pos(overflow_file_name[1], ".", 4)
        overflow_task_id = overflow_file_name[1][overflow_second_dot_pos+1:overflow_third_dot_pos]
        overflow_stream_id = overflow_file_name[1][overflow_third_dot_pos+1:overflow_fourth_dot_pos]
        assert output_task_id == overflow_task_id
        assert output_stream_id == overflow_stream_id
        # check if overflow dump file contains same task and stream id as file name
        with open(overflow_path, 'rb') as f:
            f.seek(321, 0)
            raw_data = f.read()
            task_id_infile = int.from_bytes(raw_data[24:25], 'little')
            stream_id_infile = int.from_bytes(raw_data[16:17], 'little')
            assert output_task_id == str(task_id_infile)
            assert output_stream_id == str(stream_id_infile)
        del os.environ['MINDSPORE_DUMP_CONFIG']


def run_not_overflow_dump():
    """Run async dump and not generate overflow"""
    if sys.platform != 'linux':
        return
    overflow_x = np.array([60000, 60000]).astype(np.float16)
    overflow_y = np.array([2, 2]).astype(np.float16)
    with tempfile.TemporaryDirectory(dir='/tmp') as tmp_dir:
        dump_path = os.path.join(tmp_dir, 'overflow_dump')
        dump_config_path = os.path.join(tmp_dir, 'overflow_dump.json')
        generate_dump_json_with_overflow(dump_path, dump_config_path, 'test_async_dump', 3)
        os.environ['MINDSPORE_DUMP_CONFIG'] = dump_config_path
        if os.path.isdir(dump_path):
            shutil.rmtree(dump_path)
        add = Net()
        add(Tensor(overflow_x), Tensor(overflow_y))
        exe_graph_path = os.path.join(dump_path, 'rank_0', 'Net', '0', '0')
        # check no overflow is happening, and path should not be generated
        assert not os.path.exists(exe_graph_path)
        del os.environ['MINDSPORE_DUMP_CONFIG']

@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
@security_off_wrap
def test_ascend_overflow_dump():
    """
    Feature: Overflow Dump
    Description: Test overflow dump
    Expectation: Overflow is occurred, and overflow dump file is in correct format
    """
    context.set_context(mode=context.GRAPH_MODE, device_target='Ascend')
    run_overflow_dump()


@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
@security_off_wrap
def test_ascend_not_overflow_dump():
    """
    Feature: Overflow Dump
    Description: Test overflow dump
    Expectation: Overflow is not occurred, and overflow dump file is not generated
    """
    context.set_context(mode=context.GRAPH_MODE, device_target='Ascend')
    run_not_overflow_dump()

def check_statistic_dump(dump_file_path):
    output_name = "statistic.csv"
    output_path = glob.glob(os.path.join(dump_file_path, output_name))[0]
    real_path = os.path.realpath(output_path)
    with open(real_path) as f:
        reader = csv.DictReader(f)
        input1 = next(reader)
        assert input1['IO'] == 'input'
        assert input1['Min Value'] == '1'
        assert input1['Max Value'] == '6'
        input2 = next(reader)
        assert input2['IO'] == 'input'
        assert input2['Min Value'] == '7'
        assert input2['Max Value'] == '12'
        output = next(reader)
        assert output['IO'] == 'output'
        assert output['Min Value'] == '8'
        assert output['Max Value'] == '18'

def check_data_dump(dump_file_path):
    output_name = "Add.Add-op*.0.0.*.output.0.DefaultFormat.npy"
    output_path = glob.glob(os.path.join(dump_file_path, output_name))[0]
    real_path = os.path.realpath(output_path)
    output = np.load(real_path)
    expect = np.array([[8, 10, 12], [14, 16, 18]], np.float32)
    assert np.array_equal(output, expect)

def run_gpu_e2e_dump(saved_data):
    """Run gpu e2e dump"""
    if sys.platform != 'linux':
        return
    pwd = os.getcwd()
    with tempfile.TemporaryDirectory(dir=pwd) as tmp_dir:
        dump_path = os.path.join(tmp_dir, 'gpu_e2e_dump')
        dump_config_path = os.path.join(tmp_dir, 'gpu_e2e_dump.json')
        generate_statistic_dump_json(dump_path, dump_config_path, 'test_gpu_e2e_dump', saved_data)
        os.environ['MINDSPORE_DUMP_CONFIG'] = dump_config_path
        dump_file_path = os.path.join(dump_path, 'rank_0', 'Net', '0', '0')
        if os.path.isdir(dump_path):
            shutil.rmtree(dump_path)
        add = Net()
        add(Tensor(x), Tensor(y))
        for _ in range(3):
            if not os.path.exists(dump_file_path):
                time.sleep(2)
        check_dump_structure(dump_path, dump_config_path, 1, 1, 1)
        if saved_data in ('statistic', 'full'):
            check_statistic_dump(dump_file_path)
        if saved_data in ('tensor', 'full'):
            check_data_dump(dump_file_path)
        if saved_data == 'statistic':
            # assert only file is statistic.csv, tensor data is not saved
            assert len(os.listdir(dump_file_path)) == 1
        elif saved_data == 'tensor':
            # assert only tensor data is saved, not statistics
            stat_path = os.path.join(dump_file_path, 'statistic.csv')
            assert not os.path.isfile(stat_path)
        del os.environ['MINDSPORE_DUMP_CONFIG']

@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
@security_off_wrap
def test_gpu_e2e_statistic_dump():
    """
    Feature: GPU Statistics Dump
    Description: Test GPU statistics dump
    Expectation: Statistics are stored in statistic.csv files
    """
    context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
    run_gpu_e2e_dump('statistic')

@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
@security_off_wrap
def test_gpu_e2e_tensor_dump():
    """
    Feature: GPU Tensor Dump
    Description: Test GPU tensor dump
    Expectation: Tensor data are stored in npy files
    """
    context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
    run_gpu_e2e_dump('tensor')

@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
@security_off_wrap
def test_gpu_e2e_full_dump():
    """
    Feature: GPU Full Dump
    Description: Test GPU full dump
    Expectation: Tensor are stored in npy files and their statistics stored in statistic.csv
    """
    context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
    run_gpu_e2e_dump('full')