Merge pull request #411 from MegEngine/add-tools

feat(imperative): add more tools for megengine

imperative/python/megengine/tools/README.md

@@ -1,8 +1,156 @@
 # MegEngine Tools
 
-This directory contains executable python files.
-Use these files in the following way (replace `xxx` to specific file name, like `network_visualize`):
+MegEngine 相关的工具汇总。使用方法如下(可将 `xxx` 替换成任一脚本文件，如 `network_visualize`):
 
-```
+```bash
 python -m megengine.tools.xxx
-```
+```
+
+工具列表:
+
+### accuracy_shake_var_tree
+
+将精度抖动分析结果构造成树结构，方便锁定引起抖动的根节点，以及查找依赖关系。
+
+输入: compare_binary_iodump 的输出存入到的一个文件
+
+输出: 第一个出现结果不一致的输出结点
+
+执行命令: accuracy_shake_var_tree 中定义了一些函数组件，可按需集成到实际代码中。下面有一个测试代码:
+
+```python
+import megengine.tools.accuracy_shake_var_tree as st
+
+r = st.parse('diff.txt')
+for key, value in r.items():
+    n = st.varNode.get_varNode(key)
+    n.show_src_info()
+    print("reference nodes:")
+    for i in n.get_reference_list():
+        print(i.id)
+```
+
+### benchmark_op
+
+逐个运行 functional op(并不是所有的 functional op)，对比 MegEngine 与 PyTorch 的性能，通过量化结果来指导如何进行下一步的优化。
+
+输入: 无
+
+输出: 打印一个列表，对比在小输入和大输入的情况下 MegEngine 和 Pytorch 执行一些 functional op 的速度对比
+
+执行命令: `python3 -m megengine.tools.benchmark_op`
+
+### compare_binary_iodump
+
+分析同一模型在不同平台下给定相同输入之后的输出是否完全一致。
+
+输入: 两个目录(假设分别为 expect/ 和 actual/)，分别存有不同平台下运行的 tensor 结果
+
+输出: 打印所有的输出 tensor 信息，如果某个 tensor 在两个平台上的值不一致，那么会打印出第一个不一致的值
+
+执行命令: `python3 -m megengine.tools.compare_binary_iodump expect/ actual/`
+
+### draw_graph
+
+用来查看静态图的 op 序列，有助于理解 MegEngine 的静态图在动态图的基础上做了哪些优化。
+
+输入: `megengine.core.tensor.megbrain_graph.Graph._to_json` 得出的静态图描述文件，为 json 格式
+
+输出: 一个 dot 文件，可通过 dot 命令绘制出图片
+
+执行命令:
+
+```bash
+python3 -m megengine.tools.draw_graph -i dump.json -o dump.dot
+dot -Tpng dump.dot -o dump.png
+```
+
+### dump_with_testcase_mge
+
+将待测数据提前注入模型文件，并在本地运行得到期望结果，可与实际运行的结果进行比对以检查是否出错。
+
+输入: 一个 MegEngine 模型文件，可选一些 npy 文件作为模型输入(也可以随机生成输入，如下面的命令示例)
+
+输出: 一个带输入的 MegEngine 模型文件
+
+执行命令: `python3 -m megengine.tools.dump_with_testcase_mge model.mge -d "#rand(0,255,14,2)"`
+
+### graph_info_analyze
+
+将图和内存信息的 json 文件的文件夹 logs 转换为 TensorBoard 的输入文件夹 logs_p。以便 TensorBoard 对图结构以及内存信息进行可视化。
+
+输入: 图和内存信息的 json 文件的文件夹
+
+输出: TensorBoard 的输入文件夹
+
+执行命令: `python3 -m megengine.tools.graph_info_analyze -i logs -o logs_p`
+
+### load_network_and_run
+
+python 版本的 load_and_run。
+
+输入: MegEngine 的模型文件，可选一些 npy 文件作为模型输入
+
+输出: 模型执行并打印一些测速信息
+
+执行命令: `python3 -m megengine.tools.load_network_and_run model.mge --iter 10`
+
+### network_visualize
+
+1. 分析给定的 MegEngine 模型中参数量信息，包括 shape、dtype、mean、std 以及 size 占比等。
+2. 分析给定的 MegEngine 模型中算子 FLOPs 计算量以及占比，还有算子的 inputs\outputs shape、感受野、stride 等。
+
+输入: MegEngine 的模型文件
+
+输出: 模型中的参数量信息或计算量信息
+
+执行命令:
+
+```bash
+# 分析参数量
+python3 -m megengine.tools.network_visualize model.mge --cal_params --logging_to_stdout
+
+# 分析计算量
+python3 -m megengine.tools.network_visualize model.mge --cal_flops --logging_to_stdout
+```
+
+### profile_analyze
+
+对于 load_and_run --profile 运行模型生成的 profile.json 文件或者 trace 模式下开启 profiling 功能并通过 trace.get_profile() 得到的 json 文件进行分析，得到静态图中算子的时间和显存占比等信息，以表格形式呈现。
+
+输入: load_and_run 生成的 profile 文件
+
+输出: 一个按照参数在输入文件中筛选得出的数据表格
+
+执行命令:
+
+```bash
+# 生成供分析的 json 文件
+python3 -m megengine.tools.load_network_and_run model.mge --warm-up --iter 10 --profile profile.json
+
+#分析耗时前 3 的单个算子
+python3 -m megengine.tools.profile_analyze profile.json -t 3
+
+#筛选用时超过 10us 的 conv 按 flops 排序
+python3 -m megengine.tools.profile_analyze profile.json -t 3 --order-by +flops --min-time 1e-5 --type ConvolutionForward
+```
+
+### profiler
+
+对给定的训练程序，记录训练过程并以通用格式存储，可在浏览器上可视化。
+
+输入: 需要一个 MegEngine 的训练程序(称之为 train.py，其中包含一个典型的 MegEngine 训练过程)
+
+输出: 一些记录 profile 过程的 json 文件，默认在 profile 子目录下，可用 https://ui.perfetto.dev/ 进行加载并且可视化
+
+执行命令: `python3 -m megengine.tools.profiler train.py`
+
+### svg_viewer
+
+查看 MegEngine 生成的显存占用图，可以帮助用户了解显存使用情况.
+
+输入: 显存占用的 svg 图片
+
+输出: 网页展示的可视化
+
+执行命令: `python3 -m megengine.tools.svg_viewer`

imperative/python/megengine/tools/benchmark_op.py

@@ -0,0 +1,151 @@
+# -*- coding: utf-8 -*-
+# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+#
+# Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+import time
+import numpy as np
+import megengine as mge
+import megengine.module as MM
+import megengine.functional as MF
+import torch
+import torch.nn as nn
+import torch.nn.functional as TF
+ 
+from tabulate import tabulate
+ 
+module_cache = {
+    "conv2d": (MM.Conv2d(32, 32, 3, 1, 0), nn.Conv2d(32, 32, 3, 1, 0).cuda()),
+    "dw_conv2d": (MM.Conv2d(32, 32, 3, 1, 0, groups=32), nn.Conv2d(32, 32, 3, 1, 0, groups=32).cuda()),
+    "conv3d": (MM.Conv3d(32, 32, 3, 1, 0), nn.Conv3d(32, 32, 3, 1, 0).cuda()),
+    "ConvTranspose2d": (MM.ConvTranspose2d(32, 32, 3, 1, 0), nn.ConvTranspose2d(32, 32, 3, 1, 0).cuda()),
+    "BatchNorm2d": (MM.BatchNorm2d(64), nn.BatchNorm2d(64).cuda()),
+    "Linear": (MM.Linear(1000, 1000), nn.Linear(1000, 1000).cuda()),
+}
+ 
+test_cases = [
+    # (mge op, torch op, small inps, large inps, unpack_inps, rep)
+    ("adaptive_avg_pool2d", lambda x: MF.adaptive_avg_pool2d(x, (7,7)), lambda x: TF.adaptive_avg_pool2d(x, (7,7)), [(2, 32, 16, 16)], [(64, 512, 16, 16)], True, 1000),
+    ("adaptive_max_pool2d", lambda x: MF.adaptive_max_pool2d(x, (7,7)), lambda x: TF.adaptive_max_pool2d(x, (7,7)), [(2, 32, 16, 16)], [(64, 512, 16, 16)], True, 1000),
+    ("argsort", MF.argsort, torch.argsort, [(1000,)], [(1000, 1000),], True, 1000),
+    ("avg_pool2d", lambda x: MF.avg_pool2d(x, 2), lambda x: TF.avg_pool2d(x, 2), [(2, 32, 16, 16)], [(64, 512, 16, 16)], True, 1000),
+    ("broadcast", lambda x: MF.broadcast_to(x, (5,) + x.shape), lambda x: torch.broadcast_to(x, (5,)+x.shape), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("batchedmatmul", MF.matmul, torch.matmul, [(8, 64, 32), (8, 32, 64)], [(8, 2048, 512), (8, 512, 2048)], True, 1000),
+    ("batchnrom2d", lambda x: module_cache["BatchNorm2d"][0](x), lambda x: module_cache["BatchNorm2d"][1](x), [(2, 64, 16, 16)], [(64, 64, 128, 128)], True, 1000),
+    ("concat", MF.concat, torch.cat, [(20, 100), (50, 100), (30, 100)], [(64, 512, 16, 16), (64, 512, 16, 16), (64, 512, 16, 16)], False, 1000),
+    ("conv2d", lambda x: module_cache["conv2d"][0](x), lambda x: module_cache["conv2d"][1](x), [(2, 32, 16, 16)], [(32, 32, 128, 128)], True, 1000),
+    ("conv3d", lambda x: module_cache["conv3d"][0](x), lambda x: module_cache["conv3d"][1](x), [(2, 32, 8, 8, 8)], [(32, 32, 16, 16, 16)], True, 1000),
+    ("convTranspose2d", lambda x: module_cache["ConvTranspose2d"][0](x), lambda x: module_cache["ConvTranspose2d"][1](x), [(2, 32, 16, 16)], [(32, 32, 128, 128)], True, 1000),
+    ("dropout", lambda x: MF.dropout(x, 0.5), TF.dropout, [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("dw_conv2d", lambda x: module_cache["dw_conv2d"][0](x), lambda x: module_cache["dw_conv2d"][1](x), [(2, 32, 16, 16)], [(32, 32, 128, 128)], True, 1000),
+    ("elemwise.unary", MF.log, torch.log, [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("elemwise.binary", MF.add, torch.add, [(100,100), (100,100)], [(64, 512, 16, 16), (64, 512, 16, 16)], True, 1000),
+    ("expand_dims", lambda x: MF.expand_dims(x, 0), lambda x: torch.unsqueeze(x, 0), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("gelu", MF.gelu, TF.gelu, [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("hswish", MF.hswish, TF.hardswish, [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("hsigmoid", MF.hsigmoid, TF.hardsigmoid, [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("isinf", MF.isinf, torch.isinf, [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("indeixngMultiAxisVec", lambda x: x[[1,3,5], [1,3,5], [1,3,5], [1,3,5]], lambda x: x[[1,3,5], [1,3,5], [1,3,5], [1,3,5]], [(10,10,10,10)], [(64, 512, 16, 16)], True, 1000),
+    ("logsigmoid", MF.logsigmoid, TF.logsigmoid, [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("leaky_relu", lambda x: MF.leaky_relu(x, 0.5), lambda x: TF.leaky_relu(x, 0.5), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("linear", lambda x: module_cache["Linear"][0](x), lambda x: module_cache["Linear"][1](x), [(10, 1000)], [(64, 128, 1000)], True, 1000),
+    ("matinv", MF.matinv, torch.inverse, [(10,10)], [(30, 30)], True, 1000),
+    ("matmul", MF.matmul, torch.matmul, [(64,32), (32, 64)], [(2048, 1024), (1024, 2048)], True, 1000),
+    ("max_pool2d", lambda x: MF.max_pool2d(x, 2), lambda x: TF.max_pool2d(x, 2), [(2, 32, 16, 16)], [(64, 512, 16, 16)], True, 1000),
+    ("normal", lambda x: mge.random.normal(0,1, x.shape), lambda x: torch.randn(x.shape, device="cuda"), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("prelu", MF.prelu, TF.prelu, [(100,100), (1,)], [(64, 512, 16, 16), (1,)], True, 1000),
+    ("reduce.max", lambda x: MF.max(x, 0), lambda x: torch.max(x, 0), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("reduce.mean", lambda x: MF.mean(x, 0), lambda x: torch.mean(x, 0), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("reduce.mean", lambda x: MF.mean(x, 0), lambda x: torch.mean(x, 0), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("relu", MF.relu, TF.relu, [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("relu6", MF.relu6, TF.relu6, [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("repeat", lambda x: MF.repeat(x, 5), lambda x: torch.repeat_interleave(x, 5), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("silu", MF.silu, TF.silu, [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("split", lambda x: MF.split(x, 5), lambda x: torch.split(x, 5), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("sigmoid", MF.sigmoid, TF.sigmoid, [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("softmax", lambda x: MF.softmax(x, axis=1), lambda x: TF.softmax(x, dim=1), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("softplus", MF.softplus, TF.softplus, [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("squeeze", lambda x: MF.squeeze(x, 0), lambda x: torch.squeeze(x, 0), [(1, 100,100)], [(1, 64, 512, 16, 16)], True, 1000),
+    ("stack", MF.stack, torch.stack, [(100,100), (100,100)], [(64, 512, 16, 16), (64, 512, 16, 16)], False, 10000),
+    ("subtensor", lambda x: x[0:20, 10:60], lambda x: x[0:20, 10:60], [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("topk", lambda x: MF.topk(x, 10), lambda x: torch.topk(x, 10), [(100,100)], [(1000, 1000)], True, 1000),
+    ("tile", lambda x: MF.tile(x, (2,)*len(x.shape)), lambda x: torch.tile(x, (2,)*len(x.shape)), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("transpose", lambda x: MF.transpose(x, list(range(len(x.shape)))[::-1]), lambda x: torch.permute(x, list(range(len(x.shape)))[::-1]), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("where", lambda x: MF.where(x > 0.5, x, x), lambda x: torch.where(x > 0.5, x, x), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+    ("uniform", lambda x: mge.random.uniform(0,1, x.shape), lambda x: torch.rand(x.shape, device="cuda"), [(100,100)], [(64, 512, 16, 16)], True, 1000),
+]
+ 
+ 
+def perf_func(func, inps, reps, unpack_inps, is_mge):
+    if is_mge:
+        mge._full_sync()
+        tik = time.time()
+        for _ in range(reps):
+            if unpack_inps:
+                out = func(*inps)
+            else:
+                out = func(inps)
+        mge._full_sync()
+    else:
+        torch.cuda.synchronize()
+        with torch.no_grad():
+            tik = time.time()
+            for _ in range(reps):
+                if unpack_inps:
+                    out = func(*inps)
+                else:
+                    out = func(inps)
+        torch.cuda.synchronize()
+    return time.time() - tik
+ 
+ 
+def get_avg_time(func, inps, reps, unpack_inps, is_mge):
+    # warm up
+    for _ in range(2):
+        t = perf_func(func, inps, reps, unpack_inps, is_mge)
+ 
+    times = []
+    for _ in range(5):
+        t = perf_func(func, inps, reps, unpack_inps, is_mge)
+        times.append(t)
+    return np.mean(times)
+ 
+ 
+def get_perf_results(mge_func, torch_func, shapes, unpack_inps, reps):
+    inps = [
+        np.random.randn(*shape) for shape in shapes
+    ]
+ 
+    inps_mge = [mge.tensor(inp, dtype="float32") for inp in inps]
+    avg_time_mge = get_avg_time(mge_func, inps_mge, reps, unpack_inps, True)
+ 
+    inps_torch = [torch.Tensor(inp).type(torch.float).cuda() for inp in inps]
+    avg_time_torch = get_avg_time(torch_func, inps_torch, reps, unpack_inps, False)
+ 
+    return avg_time_mge, avg_time_torch
+ 
+ 
+if __name__ == "__main__":
+    header = ["opr_name", "time(mge/pytorch; small input)", "time(mge/pytorch; large input)"]
+    table = []
+    for case in test_cases:
+        assert len(case) == 7
+        name, mge_func, torch_func, small_shapes, large_shapes, unpack_inps, reps = case
+        data = []
+        data.append(name)
+        print("========== op: {}".format(name))
+ 
+        avg_time_mge, avg_time_torch = get_perf_results(mge_func, torch_func, small_shapes, unpack_inps, reps)
+        print("mge time: {}".format(avg_time_mge))
+        print("torch time: {}".format(avg_time_torch))
+        data.append("{:.2f}".format(avg_time_mge / avg_time_torch))
+ 
+        avg_time_mge, avg_time_torch = get_perf_results(mge_func, torch_func, large_shapes, unpack_inps, reps)
+        print("mge time: {}".format(avg_time_mge))
+        print("torch time: {}".format(avg_time_torch))
+        data.append("{:.2f}".format(avg_time_mge / avg_time_torch))
+        table.append(data)
+    print(tabulate(table, header, tablefmt="github"))

imperative/python/megengine/tools/dump_with_testcase_mge.py

@@ -0,0 +1,535 @@
+# -*- coding: utf-8 -*-
+# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+#
+# Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+import argparse
+import os
+import re
+import struct
+
+import cv2
+import numpy as np
+
+import megengine as mge
+import megengine.core._imperative_rt as rt
+import megengine.core.tensor.megbrain_graph as G
+from megengine import tensor
+from megengine.core.ops import builtin
+from megengine.utils import comp_graph_tools as cgtools
+
+logger = mge.get_logger(__name__)
+
+
+def auto_reformat_image(args, path, data, dst_shape):
+    """reformat image to target shape
+
+    :param data: image data as numpy array
+    :param dst_shape: target shape
+    """
+    dim3_format = False  # required input format does not contain batch
+    hwc_format = False  # required input format is NHWC
+
+    if not dst_shape:  # input tensor shape is not predefined
+        if len(data.shape) == 2:
+            chl = 1
+            h = data.shape[0]
+            w = data.shape[1]
+        else:
+            assert len(data.shape) == 3, "Input image must be of dimension 2 or 3"
+            h, w, chl = data.shape
+        dst_shape = (1, chl, h, w)
+
+    if len(dst_shape) == 3:
+        dst_shape = (1,) + dst_shape
+        dim3_format = True
+
+    assert len(dst_shape) == 4, "bad dst_shape: {}".format(dst_shape)
+    chl = dst_shape[1]
+    if chl in [1, 3]:
+        n, c, h, w = dst_shape
+        dst_shape = (n, h, w, c)
+    else:
+        chl = dst_shape[3]
+        assert chl in [1, 3], "can not infer input format from shape: {}".format(
+            dst_shape
+        )
+        hwc_format = True
+
+    # dst_shape has now been normalized to NHWC format
+
+    if args.resize_input:
+        h, w = dst_shape[1:3]
+        data = cv2.resize(data, (w, h))
+        logger.info("input {} resized to {}".format(path, data.shape))
+
+    if chl == 1:
+        data = cv2.cvtColor(data, cv2.COLOR_BGR2GRAY)
+        data = data[:, :, np.newaxis]
+
+    assert data.ndim == 3
+    data = data[np.newaxis]
+    # data normalized to NHWC format
+
+    if not hwc_format:
+        data = np.transpose(data, (0, 3, 1, 2))
+
+    if dim3_format:
+        data = np.squeeze(data, 0)
+
+    return data
+
+
+def read_input_data(args, dst_shape, dtype, path, repeat):
+    def check_shape_equal(dst_shape, data_shape):
+        if len(dst_shape):
+            assert len(data_shape) == len(
+                dst_shape
+            ), "input/data shapes mismatch: {} vs {}".format(dst_shape, data_shape)
+
+            if data_shape[1:] != dst_shape[1:]:
+                logger.warning(
+                    "dst_shape is {}; data_shape is {}".format(dst_shape, data_shape)
+                )
+
+    if path.startswith("#"):
+        assert not args.resize_input
+        assert not args.input_transform
+        spec = path
+        m = re.match(r"^#rand\(([-0-9.]*)\s*,\s*([-0-9.]*)\s*(,[^\)]+)?\)$", spec)
+        assert m, "bad spec {}".format(spec)
+
+        rng_min = float(m.group(1))
+        rng_max = float(m.group(2))
+        if m.group(3):
+            shape_str = m.group(3)
+            try:
+                shape = shape_str[1:].split(",")
+                if shape[-1].strip() == "...":
+                    shape = shape[:-1]
+                    shape.extend(list(dst_shape[len(shape) :]))
+                data_shape = tuple(map(int, shape))
+            except ValueError as e:
+                raise ValueError("bad spec {}: {}".format(spec, e.args))
+        else:
+            data_shape = dst_shape
+
+        check_shape_equal(dst_shape, data_shape)
+        return np.random.uniform(rng_min, rng_max, data_shape).astype(dtype)
+
+    # try to load image
+    data = cv2.imread(path, cv2.IMREAD_COLOR)
+    if data is None:
+        assert not args.resize_input
+        data = np.load(path)
+        assert isinstance(data, np.ndarray)
+    else:
+        # load image succeeds, so we expect input format is image format
+        data = auto_reformat_image(args, path, data, dst_shape)
+
+    data = np.repeat(data, repeat, axis=0)
+    if repeat > 1:
+        logger.info(
+            "repeat input for {} times, data shape is {}".format(repeat, data.shape)
+        )
+
+    check_shape_equal(dst_shape, data.shape)
+
+    if args.input_transform:
+        data = eval(args.input_transform, {"data": data, "np": np})
+
+    return data
+
+
+def gen_one_testcase(args, inputs, spec):
+    paths = spec.split(";")
+    if len(paths) != len(inputs):
+        if len(paths) == 1 and paths[0].startswith("#"):
+            paths = ["{}:{}".format(name, paths[0]) for name in inputs.keys()]
+    assert len(paths) == len(inputs), "required inputs: {}; data paths: {}".format(
+        inputs.keys(), paths
+    )
+    if len(paths) == 1 and ":" not in paths[0]:
+        paths[0] = next(iter(inputs.keys())) + ":" + paths[0]
+
+    ret = {}
+    for path in paths:
+        var, path = path.split(":")
+        if args.repeat:
+            repeat = args.repeat
+        else:
+            repeat = 1
+        ret[var] = read_input_data(
+            args, inputs[var].shape, inputs[var].dtype, path, repeat
+        )
+    return ret
+
+
+def make_feeds(args):
+    ret = G.load_graph(args.input)
+    cg_rt, outputs = ret.graph, ret.output_vars_list
+    inputs = cgtools.get_dep_vars(outputs, "Host2DeviceCopy")
+
+    inputs = {i.name: i for i in inputs}
+    if not args.no_assert:
+
+        replace_varmap = {}
+        inp_map = {}
+        # replace var use InputNode
+        for name, var in inputs.items():
+            inp = G.InputNode(
+                device="xpux", dtype=var.dtype, shape=var.shape, graph=cg_rt
+            )
+            replace_varmap[var] = inp.outputs[0]
+            inp_map[name] = inp
+
+        new = cgtools.replace_vars(outputs, replace_varmap)
+        if isinstance(new, rt.VarNode):
+            new = list(new)
+
+        output_nodes = [G.OutputNode(var) for var in new]
+        func = cg_rt.compile([node.outputs[0] for node in output_nodes])
+
+        def make_dev_tensor(value, dtype=None, device=None):
+            return tensor(value, dtype=dtype, device=device)._dev_tensor()
+
+        def calculate(*args, **kwargs):
+            output_val = []
+            # set inputs value
+            for name, var in inputs.items():
+                val = kwargs.pop(name, None)
+                assert val is not None, "miss input name{}".format(name)
+                dev_tensor = make_dev_tensor(val, dtype=var.dtype, device="xpux")
+                inp_map[name].set_value(dev_tensor)
+
+            func.execute()
+
+            for res in output_nodes:
+                output_val.append(res.get_value().numpy())
+            return output_val
+
+        def expect_name(var):
+            return "{}:expect".format(var.name)
+
+    testcases = []
+
+    np.set_printoptions(precision=2, threshold=4, suppress=True)
+
+    data_list = []
+    for item in args.data:
+        if item.startswith("@"):
+            with open(item[1:], "r") as f:
+                data_list.extend([line.rstrip() for line in f if line.rstrip() != ""])
+        else:
+            data_list.append(item)
+
+    for inp_spec in data_list:
+        cur_testcase = gen_one_testcase(args, inputs, inp_spec)
+        assert len(cur_testcase) == len(
+            inputs
+        ), "required inputs: {}; given data: {}".format(
+            inputs.keys(), cur_testcase.keys()
+        )
+
+        if not args.no_assert:
+            outputs_get = calculate(**cur_testcase)
+            for var, val in zip(outputs, outputs_get):
+                cur_testcase[expect_name(var)] = val
+                logger.info(
+                    "generate test groundtruth: var={} shape={} range=({}, {})"
+                    " mean={} var={}".format(
+                        var, val.shape, val.min(), val.max(), np.mean(val), np.var(val)
+                    )
+                )
+        testcases.append(cur_testcase)
+        logger.info(
+            "add testcase: \n {}".format(
+                "\n ".join(
+                    "{}: shape={} dtype={} range=({:.2f},{:.2f}) "
+                    "mean={:.2f} sd={:.2f}".format(
+                        k, v.shape, v.dtype, v.min(), v.max(), np.mean(v), np.std(v)
+                    )
+                    for k, v in sorted(cur_testcase.items())
+                )
+            )
+        )
+
+    if not args.no_assert:
+
+        def expect_shp(var):
+            ret = var.shape
+            if ret:
+                return ret
+            return testcases[0][expect_name(var)].shape
+
+        def assert_equal(expect, real, **kwargs):
+            op = builtin.AssertEqual(**kwargs)
+            (res,) = G.apply_normal_varnode(op, expect, real)
+            return res
+
+        verbose = not args.silent
+
+        outputs_new = []
+        for i in outputs:
+            device = rt.CompNode("xpux")
+            dtype = i.dtype
+            name = expect_name(i)
+            shape = expect_shp(i)
+            # make expect output as one input of model.
+            expect_get = rt.make_h2d(cg_rt, device, dtype, shape, name)
+            # insert assert opr to check expect and real.
+            outputs_new.append(
+                assert_equal(
+                    expect_get,
+                    i,
+                    verbose=verbose,
+                    maxerr=args.maxerr,
+                )
+            )
+            inputs[expect_name(i)] = expect_get
+        outputs = outputs_new
+
+    return {"outputs": outputs, "testcases": testcases}
+
+
+def optimize_for_inference(args, outputs):
+    args_list = [
+        "enable_io16xc32",
+        "enable_ioc16",
+        "enable_hwcd4",
+        "enable_nchw4",
+        "enable_nchw88",
+        "enable_nchw44",
+        "enable_nchw44_dot",
+        "enable_nchw32",
+        "enable_chwn4",
+        "enable_fuse_conv_bias_nonlinearity",
+        "enable_fuse_conv_bias_with_z",
+        "enable_fuse_preprocess",
+    ]
+    kwargs = {}
+    for k in args_list:
+        if getattr(args, k):
+            kwargs[k] = True
+
+    if args.optimize_for_inference:
+        outputs = G.optimize_for_inference(outputs, **kwargs)
+
+    return outputs
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Pack computing graph, input values and expected output "
+        "values into one file for checking correctness. README.md gives more "
+        "details on the usage",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument("input", help="MegEngine dumped model file")
+    parser.add_argument("-o", "--output", help="output file", required=True)
+    parser.add_argument(
+        "-d",
+        "--data",
+        default=[],
+        action="append",
+        required=True,
+        help="Given input test data when input file is a network, "
+        "and current network output would be used as groundtruth. "
+        "The format is var0:file0;var1:file1... to specify data files for "
+        "input vars. It can also be #rand(min,max,shape...) for generating "
+        "random input data, for example, #rand(0,255), "
+        "#rand(0,255,1,3,224,224) or #rand(0, 255, 1, ...) where `...` means "
+        "the remaining part of the original shape. "
+        "If the shape is not specified, the shape of "
+        "corresponding input tensors in the network will be used. "
+        "If there is only one input var, its name can be omitted. "
+        "Each data file can either be an image which can be loaded by opencv, "
+        "or a pickled numpy.ndarray. "
+        "This option can be given multiple times to add multiple testcases. "
+        " *NOTE* "
+        "If you start the data with the letter @, the rest should be a "
+        "filename, and each line in the file should be a single datum in "
+        "the format described above. ",
+    )
+    parser.add_argument(
+        "--repeat",
+        type=int,
+        default=1,
+        help="Specify how many times the input image is repeated. "
+        "Useful when running benchmark for batch size other than one. "
+        "Have no effect on randomly generated input data.",
+    )
+    parser.add_argument(
+        "--silent",
+        action="store_true",
+        help="set verbose to False in asserti_equal opr",
+    )
+    parser.add_argument(
+        "--optimize-for-inference",
+        action="store_true",
+        help="enable optimization for inference",
+    )
+    parser.add_argument(
+        "--no-assert",
+        action="store_true",
+        help="do not insert assert_equal opr to check result; "
+        "this option is useful for benchmarking",
+    )
+    parser.add_argument(
+        "--maxerr",
+        type=float,
+        default=1e-4,
+        help="max error for assert_equal check during runtime",
+    )
+    parser.add_argument(
+        "--resize-input",
+        action="store_true",
+        help="resize input image to fit input var shape",
+    )
+    parser.add_argument(
+        "--input-transform",
+        help="a python expression to transform the input data. "
+        "Example: data / np.std(data)",
+    )
+    parser.add_argument(
+        "--discard-var-name",
+        action="store_true",
+        help="discard variable and param names in the " "generated output",
+    )
+    parser.add_argument(
+        "--output-strip-info", action="store_true", help="output code strip information"
+    )
+    parser.add_argument(
+        "--enable-io16xc32",
+        action="store_true",
+        help="transform the mode to float16 io float32 compute",
+    )
+    parser.add_argument(
+        "--enable-ioc16",
+        action="store_true",
+        help="transform the dtype of the model to float16 io " "and compute",
+    )
+    parser.add_argument(
+        "--enable-fuse-conv-bias-nonlinearity",
+        action="store_true",
+        help="fuse convolution bias and nonlinearity opr to a "
+        "conv_bias opr and compute",
+    )
+    parser.add_argument(
+        "--enable-hwcd4",
+        action="store_true",
+        help="transform the model format from NCHW to NHWCD4 "
+        "for inference; you may need to disable CUDA and set "
+        "MGB_USE_MEGDNN_DBG=2",
+    )
+    parser.add_argument(
+        "--enable-nchw4",
+        action="store_true",
+        help="transform the model format from NCHW to NCHW4 " "for inference",
+    )
+    parser.add_argument(
+        "--enable-nchw88",
+        action="store_true",
+        help="transform the model format from NCHW to NCHW88 " "for inference",
+    )
+    parser.add_argument(
+        "--enable-nchw44",
+        action="store_true",
+        help="transform the model format from NCHW to NCHW44 " "for inference",
+    )
+    parser.add_argument(
+        "--enable-nchw44-dot",
+        action="store_true",
+        help="transform the model format from NCHW to NCHW44_DOT "
+        "for optimizing armv8.2 dot in inference",
+    )
+    parser.add_argument(
+        "--enable-nchw32",
+        action="store_true",
+        help="transform the model format from NCHW4 to NCHW32 "
+        "for inference on nvidia TensoCore",
+    )
+    parser.add_argument(
+        "--enable-chwn4",
+        action="store_true",
+        help="transform the model format to CHWN4 "
+        "for inference, mainly used for nvidia tensorcore",
+    )
+    parser.add_argument(
+        "--enable-fuse-conv-bias-with-z",
+        action="store_true",
+        help="fuse conv_bias with z input for inference on "
+        "nvidia GPU (this optimization pass will result in mismatch "
+        "of the precision of output of training and inference)",
+    )
+    parser.add_argument(
+        "--enable-fuse-preprocess",
+        action="store_true",
+        help="fuse astype\pad_channel\dimshuffle and etc opr "
+        "from h2d opr",
+    )
+    args = parser.parse_args()
+
+    feeds = make_feeds(args)
+
+    assert isinstance(feeds, dict) and feeds["testcases"], "testcases can not be empty"
+
+    output_mgbvars = feeds["outputs"]
+    output_mgbvars = optimize_for_inference(args, output_mgbvars)
+
+    inputs = cgtools.get_dep_vars(output_mgbvars, "Host2DeviceCopy")
+    inputs = sorted((i.name, i.dtype) for i in inputs)
+
+    if args.discard_var_name:
+        sereg_kwargs = dict(keep_var_name=0, keep_param_name=False)
+    else:
+        sereg_kwargs = dict(keep_var_name=2, keep_param_name=True)
+
+    strip_info_file = args.output + ".json" if args.output_strip_info else None
+
+    with open(args.output, "wb") as fout:
+        fout.write(b"mgbtest0")
+        fout.write(struct.pack("I", len(feeds["testcases"])))
+        dump_content, stat = G.dump_graph(
+            output_mgbvars,
+            append_json=True,
+            strip_info_file=strip_info_file,
+            **sereg_kwargs,
+        )
+        fout.write(dump_content)
+
+        logger.info(
+            "graph dump sizes: tot_size={:.3f}KiB overhead={:.3f}KiB".format(
+                stat.tot_bytes / 1024, (stat.tot_bytes - stat.tensor_value_bytes) / 1024
+            )
+        )
+
+    def make_dev_tensor(value, dtype=None, device=None):
+        return tensor(value, dtype=dtype, device=device)._dev_tensor()
+
+    for testcase in feeds["testcases"]:
+        assert isinstance(testcase, dict)
+        cg = G.Graph()
+        output_mgbvars = []
+        for name, dtype in inputs:
+            output_mgbvars.append(
+                cg.make_const(
+                    make_dev_tensor(testcase.pop(name), dtype=dtype, device="cpux")
+                )
+            )
+        assert not testcase, "extra inputs provided in testcase: {}".format(
+            testcase.keys()
+        )
+        with open(args.output, "ab") as fout:
+            dump_content, _ = G.dump_graph(
+                output_mgbvars, strip_info_file=strip_info_file, append_json=True
+            )
+            fout.write(dump_content)
+
+
+if __name__ == "__main__":
+    main()
+