(beta) Building a Simple CPU Performance Profiler with FX

Author: James Reed

In this tutorial, we are going to use FX to do the following:

1. Capture PyTorch Python code in a way that we can inspect and gather statistics about the structure and execution of the code
2. Build out a small class that will serve as a simple performance “profiler”, collecting runtime statistics about each part of the model from actual runs.

For this tutorial, we are going to use the torchvision ResNet18 model for demonstration purposes.

import torch
import torch.fx
import torchvision.models as models

rn18 = models.resnet18()
rn18.eval()

Now that we have our model, we want to inspect deeper into its performance. That is, for the following invocation, which parts of the model are taking the longest?

input = torch.randn(5, 3, 224, 224)
output = rn18(input)

A common way of answering that question is to go through the program source, add code that collects timestamps at various points in the program, and compare the difference between those timestamps to see how long the regions between the timestamps take.

That technique is certainly applicable to PyTorch code, however it would be nicer if we didn’t have to copy over model code and edit it, especially code we haven’t written (like this torchvision model). Instead, we are going to use FX to automate this “instrumentation” process without needing to modify any source.

First, let’s get some imports out of the way (we will be using all of these later in the code).

import statistics, tabulate, time
from typing import Any, Dict, List
from torch.fx import Interpreter

Note

tabulate is an external library that is not a dependency of PyTorch. We will be using it to more easily visualize performance data. Please make sure you’ve installed it from your favorite Python package source.

Capturing the Model with Symbolic Tracing

Next, we are going to use FX’s symbolic tracing mechanism to capture the definition of our model in a data structure we can manipulate and examine.

traced_rn18 = torch.fx.symbolic_trace(rn18)
print(traced_rn18.graph)

Out:

graph():
    %x : torch.Tensor [#users=1] = placeholder[target=x]
    %conv1 : [#users=1] = call_module[target=conv1](args = (%x,), kwargs = {})
    %bn1 : [#users=1] = call_module[target=bn1](args = (%conv1,), kwargs = {})
    %relu : [#users=1] = call_module[target=relu](args = (%bn1,), kwargs = {})
    %maxpool : [#users=2] = call_module[target=maxpool](args = (%relu,), kwargs = {})
    %layer1_0_conv1 : [#users=1] = call_module[target=layer1.0.conv1](args = (%maxpool,), kwargs = {})
    %layer1_0_bn1 : [#users=1] = call_module[target=layer1.0.bn1](args = (%layer1_0_conv1,), kwargs = {})
    %layer1_0_relu : [#users=1] = call_module[target=layer1.0.relu](args = (%layer1_0_bn1,), kwargs = {})
    %layer1_0_conv2 : [#users=1] = call_module[target=layer1.0.conv2](args = (%layer1_0_relu,), kwargs = {})
    %layer1_0_bn2 : [#users=1] = call_module[target=layer1.0.bn2](args = (%layer1_0_conv2,), kwargs = {})
    %add : [#users=1] = call_function[target=operator.add](args = (%layer1_0_bn2, %maxpool), kwargs = {})
    %layer1_0_relu_1 : [#users=2] = call_module[target=layer1.0.relu](args = (%add,), kwargs = {})
    %layer1_1_conv1 : [#users=1] = call_module[target=layer1.1.conv1](args = (%layer1_0_relu_1,), kwargs = {})
    %layer1_1_bn1 : [#users=1] = call_module[target=layer1.1.bn1](args = (%layer1_1_conv1,), kwargs = {})
    %layer1_1_relu : [#users=1] = call_module[target=layer1.1.relu](args = (%layer1_1_bn1,), kwargs = {})
    %layer1_1_conv2 : [#users=1] = call_module[target=layer1.1.conv2](args = (%layer1_1_relu,), kwargs = {})
    %layer1_1_bn2 : [#users=1] = call_module[target=layer1.1.bn2](args = (%layer1_1_conv2,), kwargs = {})
    %add_1 : [#users=1] = call_function[target=operator.add](args = (%layer1_1_bn2, %layer1_0_relu_1), kwargs = {})
    %layer1_1_relu_1 : [#users=2] = call_module[target=layer1.1.relu](args = (%add_1,), kwargs = {})
    %layer2_0_conv1 : [#users=1] = call_module[target=layer2.0.conv1](args = (%layer1_1_relu_1,), kwargs = {})
    %layer2_0_bn1 : [#users=1] = call_module[target=layer2.0.bn1](args = (%layer2_0_conv1,), kwargs = {})
    %layer2_0_relu : [#users=1] = call_module[target=layer2.0.relu](args = (%layer2_0_bn1,), kwargs = {})
    %layer2_0_conv2 : [#users=1] = call_module[target=layer2.0.conv2](args = (%layer2_0_relu,), kwargs = {})
    %layer2_0_bn2 : [#users=1] = call_module[target=layer2.0.bn2](args = (%layer2_0_conv2,), kwargs = {})
    %layer2_0_downsample_0 : [#users=1] = call_module[target=layer2.0.downsample.0](args = (%layer1_1_relu_1,), kwargs = {})
    %layer2_0_downsample_1 : [#users=1] = call_module[target=layer2.0.downsample.1](args = (%layer2_0_downsample_0,), kwargs = {})
    %add_2 : [#users=1] = call_function[target=operator.add](args = (%layer2_0_bn2, %layer2_0_downsample_1), kwargs = {})
    %layer2_0_relu_1 : [#users=2] = call_module[target=layer2.0.relu](args = (%add_2,), kwargs = {})
    %layer2_1_conv1 : [#users=1] = call_module[target=layer2.1.conv1](args = (%layer2_0_relu_1,), kwargs = {})
    %layer2_1_bn1 : [#users=1] = call_module[target=layer2.1.bn1](args = (%layer2_1_conv1,), kwargs = {})
    %layer2_1_relu : [#users=1] = call_module[target=layer2.1.relu](args = (%layer2_1_bn1,), kwargs = {})
    %layer2_1_conv2 : [#users=1] = call_module[target=layer2.1.conv2](args = (%layer2_1_relu,), kwargs = {})
    %layer2_1_bn2 : [#users=1] = call_module[target=layer2.1.bn2](args = (%layer2_1_conv2,), kwargs = {})
    %add_3 : [#users=1] = call_function[target=operator.add](args = (%layer2_1_bn2, %layer2_0_relu_1), kwargs = {})
    %layer2_1_relu_1 : [#users=2] = call_module[target=layer2.1.relu](args = (%add_3,), kwargs = {})
    %layer3_0_conv1 : [#users=1] = call_module[target=layer3.0.conv1](args = (%layer2_1_relu_1,), kwargs = {})
    %layer3_0_bn1 : [#users=1] = call_module[target=layer3.0.bn1](args = (%layer3_0_conv1,), kwargs = {})
    %layer3_0_relu : [#users=1] = call_module[target=layer3.0.relu](args = (%layer3_0_bn1,), kwargs = {})
    %layer3_0_conv2 : [#users=1] = call_module[target=layer3.0.conv2](args = (%layer3_0_relu,), kwargs = {})
    %layer3_0_bn2 : [#users=1] = call_module[target=layer3.0.bn2](args = (%layer3_0_conv2,), kwargs = {})
    %layer3_0_downsample_0 : [#users=1] = call_module[target=layer3.0.downsample.0](args = (%layer2_1_relu_1,), kwargs = {})
    %layer3_0_downsample_1 : [#users=1] = call_module[target=layer3.0.downsample.1](args = (%layer3_0_downsample_0,), kwargs = {})
    %add_4 : [#users=1] = call_function[target=operator.add](args = (%layer3_0_bn2, %layer3_0_downsample_1), kwargs = {})
    %layer3_0_relu_1 : [#users=2] = call_module[target=layer3.0.relu](args = (%add_4,), kwargs = {})
    %layer3_1_conv1 : [#users=1] = call_module[target=layer3.1.conv1](args = (%layer3_0_relu_1,), kwargs = {})
    %layer3_1_bn1 : [#users=1] = call_module[target=layer3.1.bn1](args = (%layer3_1_conv1,), kwargs = {})
    %layer3_1_relu : [#users=1] = call_module[target=layer3.1.relu](args = (%layer3_1_bn1,), kwargs = {})
    %layer3_1_conv2 : [#users=1] = call_module[target=layer3.1.conv2](args = (%layer3_1_relu,), kwargs = {})
    %layer3_1_bn2 : [#users=1] = call_module[target=layer3.1.bn2](args = (%layer3_1_conv2,), kwargs = {})
    %add_5 : [#users=1] = call_function[target=operator.add](args = (%layer3_1_bn2, %layer3_0_relu_1), kwargs = {})
    %layer3_1_relu_1 : [#users=2] = call_module[target=layer3.1.relu](args = (%add_5,), kwargs = {})
    %layer4_0_conv1 : [#users=1] = call_module[target=layer4.0.conv1](args = (%layer3_1_relu_1,), kwargs = {})
    %layer4_0_bn1 : [#users=1] = call_module[target=layer4.0.bn1](args = (%layer4_0_conv1,), kwargs = {})
    %layer4_0_relu : [#users=1] = call_module[target=layer4.0.relu](args = (%layer4_0_bn1,), kwargs = {})
    %layer4_0_conv2 : [#users=1] = call_module[target=layer4.0.conv2](args = (%layer4_0_relu,), kwargs = {})
    %layer4_0_bn2 : [#users=1] = call_module[target=layer4.0.bn2](args = (%layer4_0_conv2,), kwargs = {})
    %layer4_0_downsample_0 : [#users=1] = call_module[target=layer4.0.downsample.0](args = (%layer3_1_relu_1,), kwargs = {})
    %layer4_0_downsample_1 : [#users=1] = call_module[target=layer4.0.downsample.1](args = (%layer4_0_downsample_0,), kwargs = {})
    %add_6 : [#users=1] = call_function[target=operator.add](args = (%layer4_0_bn2, %layer4_0_downsample_1), kwargs = {})
    %layer4_0_relu_1 : [#users=2] = call_module[target=layer4.0.relu](args = (%add_6,), kwargs = {})
    %layer4_1_conv1 : [#users=1] = call_module[target=layer4.1.conv1](args = (%layer4_0_relu_1,), kwargs = {})
    %layer4_1_bn1 : [#users=1] = call_module[target=layer4.1.bn1](args = (%layer4_1_conv1,), kwargs = {})
    %layer4_1_relu : [#users=1] = call_module[target=layer4.1.relu](args = (%layer4_1_bn1,), kwargs = {})
    %layer4_1_conv2 : [#users=1] = call_module[target=layer4.1.conv2](args = (%layer4_1_relu,), kwargs = {})
    %layer4_1_bn2 : [#users=1] = call_module[target=layer4.1.bn2](args = (%layer4_1_conv2,), kwargs = {})
    %add_7 : [#users=1] = call_function[target=operator.add](args = (%layer4_1_bn2, %layer4_0_relu_1), kwargs = {})
    %layer4_1_relu_1 : [#users=1] = call_module[target=layer4.1.relu](args = (%add_7,), kwargs = {})
    %avgpool : [#users=1] = call_module[target=avgpool](args = (%layer4_1_relu_1,), kwargs = {})
    %flatten : [#users=1] = call_function[target=torch.flatten](args = (%avgpool, 1), kwargs = {})
    %fc : [#users=1] = call_module[target=fc](args = (%flatten,), kwargs = {})
    return fc

This gives us a Graph representation of the ResNet18 model. A Graph consists of a series of Nodes connected to each other. Each Node represents a call-site in the Python code (whether to a function, a module, or a method) and the edges (represented as args and kwargs on each node) represent the values passed between these call-sites. More information about the Graph representation and the rest of FX’s APIs ca be found at the FX documentation https://pytorch.org/docs/master/fx.html.

Creating a Profiling Interpreter

Next, we are going to create a class that inherits from torch.fx.Interpreter. Though the GraphModule that symbolic_trace produces compiles Python code that is run when you call a GraphModule, an alternative way to run a GraphModule is by executing each Node in the Graph one by one. That is the functionality that Interpreter provides: It interprets the graph node- by-node.

By inheriting from Interpreter, we can override various functionality and install the profiling behavior we want. The goal is to have an object to which we can pass a model, invoke the model 1 or more times, then get statistics about how long the model and each part of the model took during those runs.

Let’s define our ProfilingInterpreter class:

class ProfilingInterpreter(Interpreter):
    def __init__(self, mod : torch.nn.Module):
        # Rather than have the user symbolically trace their model,
        # we're going to do it in the constructor. As a result, the
        # user can pass in any ``Module`` without having to worry about
        # symbolic tracing APIs
        gm = torch.fx.symbolic_trace(mod)
        super().__init__(gm)

        # We are going to store away two things here:
        #
        # 1. A list of total runtimes for ``mod``. In other words, we are
        #    storing away the time ``mod(...)`` took each time this
        #    interpreter is called.
        self.total_runtime_sec : List[float] = []
        # 2. A map from ``Node`` to a list of times (in seconds) that
        #    node took to run. This can be seen as similar to (1) but
        #    for specific sub-parts of the model.
        self.runtimes_sec : Dict[torch.fx.Node, List[float]] = {}

    ######################################################################
    # Next, let's override our first method: ``run()``. ``Interpreter``'s ``run``
    # method is the top-level entrypoint for execution of the model. We will
    # want to intercept this so that we can record the total runtime of the
    # model.

    def run(self, *args) -> Any:
        # Record the time we started running the model
        t_start = time.time()
        # Run the model by delegating back into Interpreter.run()
        return_val = super().run(*args)
        # Record the time we finished running the model
        t_end = time.time()
        # Store the total elapsed time this model execution took in the
        # ProfilingInterpreter
        self.total_runtime_sec.append(t_end - t_start)
        return return_val

    ######################################################################
    # Now, let's override ``run_node``. ``Interpreter`` calls ``run_node`` each
    # time it executes a single node. We will intercept this so that we
    # can measure and record the time taken for each individual call in
    # the model.

    def run_node(self, n : torch.fx.Node) -> Any:
        # Record the time we started running the op
        t_start = time.time()
        # Run the op by delegating back into Interpreter.run_node()
        return_val = super().run_node(n)
        # Record the time we finished running the op
        t_end = time.time()
        # If we don't have an entry for this node in our runtimes_sec
        # data structure, add one with an empty list value.
        self.runtimes_sec.setdefault(n, [])
        # Record the total elapsed time for this single invocation
        # in the runtimes_sec data structure
        self.runtimes_sec[n].append(t_end - t_start)
        return return_val

    ######################################################################
    # Finally, we are going to define a method (one which doesn't override
    # any ``Interpreter`` method) that provides us a nice, organized view of
    # the data we have collected.

    def summary(self, should_sort : bool = False) -> str:
        # Build up a list of summary information for each node
        node_summaries : List[List[Any]] = []
        # Calculate the mean runtime for the whole network. Because the
        # network may have been called multiple times during profiling,
        # we need to summarize the runtimes. We choose to use the
        # arithmetic mean for this.
        mean_total_runtime = statistics.mean(self.total_runtime_sec)

        # For each node, record summary statistics
        for node, runtimes in self.runtimes_sec.items():
            # Similarly, compute the mean runtime for ``node``
            mean_runtime = statistics.mean(runtimes)
            # For easier understanding, we also compute the percentage
            # time each node took with respect to the whole network.
            pct_total = mean_runtime / mean_total_runtime * 100
            # Record the node's type, name of the node, mean runtime, and
            # percent runtim
            node_summaries.append(
                [node.op, str(node), mean_runtime, pct_total])

        # One of the most important questions to answer when doing performance
        # profiling is "Which op(s) took the longest?". We can make this easy
        # to see by providing sorting functionality in our summary view
        if should_sort:
            node_summaries.sort(key=lambda s: s[2], reverse=True)

        # Use the ``tabulate`` library to create a well-formatted table
        # presenting our summary information
        headers : List[str] = [
            'Op type', 'Op', 'Average runtime (s)', 'Pct total runtime'
        ]
        return tabulate.tabulate(node_summaries, headers=headers)

Note

We use Python’s time.time function to pull wall clock timestamps and compare them. This is not the most accurate way to measure performance, and will only give us a first- order approximation. We use this simple technique only for the purpose of demonstration in this tutorial.

Investigating the Performance of ResNet18

We can now use ProfilingInterpreter to inspect the performance characteristics of our ResNet18 model;

interp = ProfilingInterpreter(rn18)
interp.run(input)
print(interp.summary(True))

Out:

Op type        Op                       Average runtime (s)    Pct total runtime
-------------  ---------------------  ---------------------  -------------------
call_module    conv1                            0.00458789             7.45099
call_module    maxpool                          0.00450206             7.3116
call_module    layer4_0_conv2                   0.0036912              5.99471
call_module    layer1_0_conv1                   0.00354171             5.75193
call_module    layer1_1_conv1                   0.00320005             5.19707
call_module    layer4_1_conv2                   0.00315881             5.13008
call_module    layer4_1_conv1                   0.00312328             5.07239
call_module    layer1_1_conv2                   0.00302052             4.9055
call_module    layer1_0_conv2                   0.00269413             4.37542
call_module    layer3_1_conv2                   0.00267029             4.3367
call_module    layer2_1_conv2                   0.00263286             4.27591
call_module    layer3_1_conv1                   0.00259113             4.20815
call_module    layer3_0_conv2                   0.00254703             4.13651
call_module    layer4_0_conv1                   0.0023818              3.86818
call_module    layer2_1_conv1                   0.00235248             3.82055
call_module    layer2_0_conv2                   0.00232291             3.77254
call_module    layer2_0_conv1                   0.00197101             3.20103
call_module    layer3_0_conv1                   0.0017333              2.81498
call_module    layer2_0_downsample_0            0.0013783              2.23843
call_function  add_1                            0.000653505            1.06133
call_function  add                              0.000454903            0.738788
call_module    layer3_0_downsample_0            0.000435829            0.707811
call_module    bn1                              0.000406265            0.659798
call_module    layer4_0_downsample_0            0.00036335             0.590101
call_function  add_3                            0.00028348             0.460387
call_module    layer1_0_bn1                     0.000208855            0.339192
call_module    fc                               0.000207663            0.337256
call_module    relu                             0.000199556            0.324091
call_module    layer1_1_bn1                     0.000158787            0.257879
call_module    layer1_1_bn2                     0.000146627            0.238131
call_module    layer1_0_bn2                     0.000142097            0.230774
call_module    avgpool                          0.000127554            0.207155
call_module    layer2_0_downsample_1            0.000122547            0.199023
call_module    layer2_0_bn1                     0.00012207             0.198249
call_module    layer3_1_bn2                     0.000110149            0.178889
call_module    layer2_1_bn2                     0.000107765            0.175017
call_module    layer4_0_bn2                     0.000107288            0.174242
call_module    layer3_1_bn1                     0.00010705             0.173855
call_module    layer2_1_bn1                     0.000106335            0.172694
call_module    layer2_0_bn2                     0.000103474            0.168047
call_module    layer4_1_bn2                     0.000102997            0.167273
call_module    layer1_0_relu                    0.000100851            0.163788
call_module    layer4_0_bn1                     9.89437e-05            0.16069
call_module    layer3_0_bn2                     9.53674e-05            0.154882
call_module    layer4_1_bn1                     9.53674e-05            0.154882
call_module    layer3_0_downsample_1            9.48906e-05            0.154108
call_module    layer3_0_bn1                     9.32217e-05            0.151397
call_module    layer1_1_relu_1                  9.27448e-05            0.150623
call_module    layer1_1_relu                    8.86917e-05            0.14404
call_module    layer4_0_downsample_1            8.74996e-05            0.142104
call_module    layer1_0_relu_1                  8.46386e-05            0.137458
call_function  add_2                            8.39233e-05            0.136296
call_function  add_5                            7.12872e-05            0.115774
call_module    layer3_0_relu                    6.96182e-05            0.113064
call_module    layer3_1_relu                    6.96182e-05            0.113064
call_module    layer4_0_relu                    6.86646e-05            0.111515
call_module    layer4_1_relu                    6.77109e-05            0.109966
call_module    layer2_0_relu                    6.74725e-05            0.109579
call_function  add_6                            6.65188e-05            0.10803
call_module    layer2_0_relu_1                  6.62804e-05            0.107643
call_function  add_7                            6.38962e-05            0.103771
call_module    layer2_1_relu_1                  6.34193e-05            0.102997
call_module    layer2_1_relu                    6.31809e-05            0.102609
call_function  add_4                            6.22272e-05            0.101061
call_module    layer3_0_relu_1                  5.72205e-05            0.0929292
call_module    layer3_1_relu_1                  5.62668e-05            0.0913804
call_module    layer4_0_relu_1                  5.62668e-05            0.0913804
call_module    layer4_1_relu_1                  5.57899e-05            0.090606
call_function  flatten                          3.45707e-05            0.0561448
output         output                           1.5974e-05             0.0259427
placeholder    x                                1.38283e-05            0.0224579

There are two things we should call out here:

• MaxPool2d takes up the most time. This is a known issue: https://github.com/pytorch/pytorch/issues/51393
• BatchNorm2d also takes up significant time. We can continue this line of thinking and optimize this in the Conv-BN Fusion with FX tutorial TODO: link

Conclusion

As we can see, using FX we can easily capture PyTorch programs (even ones we don’t have the source code for!) in a machine-interpretable format and use that for analysis, such as the performance analysis we’ve done here. FX opens up an exiciting world of possibilities for working with PyTorch programs.

Finally, since FX is still in beta, we would be happy to hear any feedback you have about using it. Please feel free to use the PyTorch Forums (https://discuss.pytorch.org/) and the issue tracker (https://github.com/pytorch/pytorch/issues) to provide any feedback you might have.