Profiling with Tracy

Tracy is a profiler that puts together in a single view:

• Both CPU and GPU profiling.
• Both sampling and instrumentation.
• Both specifics of our own process, and whole-system profiling a la “systrace”.

Since Tracy relies on instrumentation, it requires IREE binaries to be built with a special flag to enable it.

There are two components to Tracy. They communicate over a TCP socket.

• The “client” is the program being profiled.
• The “server” is the Tracy profiler UI.

The Tracy manual

The primary source of Tracy documentation, including for build instructions, is a PDF manual that's part of each numbered release. You can download the latest or preview it.

Install dependencies

Linux

Install dependencies (Debian-based distributions):

sudo apt install libcapstone-dev libtbb-dev libglfw3-dev libfreetype6-dev libgtk-3-dev

If you only build the command line tool you can install:

sudo apt install libcapstone-dev libtbb-dev libzstd-dev

Mac

brew install capstone glfw freetype

Building IREE with Tracy instrumentation (the “client”)

IREE needs to be build with Tracy instrumentation enabled. This enables both the collection of data, and its streaming to the Tracy server over a socket.

This is only supported in the CMake build system of IREE, not in Bazel.

In the initial CMake configuration command:

• Set IREE_ENABLE_RUNTIME_TRACING to ON.
• Use the RelWithDebInfo build type.

For example:

$ cmake \
  -DIREE_ENABLE_RUNTIME_TRACING=ON \
  -DCMAKE_BUILD_TYPE=RelWithDebInfo \
  ... # other cmake arguments as usual

For tracing the compiler, additionally set IREE_ENABLE_COMPILER_TRACING to ON. Compiler tracing is less stable, particularly on Linux with MLIR threading enabled (https://github.com/google/iree/issues/6404).

Building the Tracy Command Line Capture (the alternative “server”)

Optionally build the command capture tool with -DIREE_BUILD_TRACY=ON

$ cmake \
  -DIREE_BUILD_TRACY=ON \
  -DIREE_ENABLE_RUNTIME_TRACING=ON \
  -DCMAKE_BUILD_TYPE=RelWithDebInfo \
  ... # other cmake arguments as usual

The built command line tool will be found at ./build/tracy/iree-tracy-capture

Building the Tracy UI (the “server”)

This is explained in “Building the server” section 2.3 of the manual for Windows and Linux. Here we give some more detailed instructions for some systems.

This is built as part of the IREE build above if you have installed the dependencies. The tool would be located at ./build/tracy/iree-tracy-profiler

If you would like to build the “server” for your desktop to connect remotely follow along.

The IREE repository contains its own clone of the Tracy repository in third_party/tracy, so there is no need to make a separate clone of it. You can use one if you want, but be aware that the Tracy client/server protocol gets updated sometimes. Building both sides from the same iree/third_party/tracy lowers the risk of running into a protocol version mismatch.

Build on Mac and Linux

Build (from your iree/ clone root directory):

make -C third_party/tracy/profiler/build/unix -j12 release

Running the profiled program

Tracing doesn't work properly on VMs (see “Problematic Platforms / Virtual Machines” section 2.1.6.4 of the manual). To get sampling, you should run the profiled program on bare metal.

Run your IREE workload as you normally would: now that it's been built with Tracy instrumentation enabled, it should do all the right things automatically.

The only change that you are likely to need in your command line is to set the TRACY_NO_EXIT=1 environment variable. This ensures that your program does not exit until the Tracy server (the UI) has connected to it and finished uploading the profile data.

Typically, TRACY_NO_EXIT=1 is needed when profiling iree-benchmark-module. It wouldn't be needed when profiling a real user-facing application.

Example:

TRACY_NO_EXIT=1 /data/local/tmp/iree-benchmark-module \
  --driver=dylib \
  --module_file=/data/local/tmp/android_module.fbvm \
  --entry_function=serving_default \
  --function_input=1x384xi32 \
  --function_input=1x384xi32 \
  --function_input=1x384xi32

Running the Tracy Capture CLI, connecting and saving profiles

While the program that you want to profile is still running (thanks to TRACY_NO_EXIT=1), start the Tracy capture tool in another terminal / ttyl. From the IREE root directory:

./build/tracy/iree-tracy-capture -o myprofile.tracy
Connecting to 127.0.0.1:8086...

It should connect to the IREE client and save the output to myprofile.tracy that can be visualized by the client below. You can start the capture tool first to make sure you don't miss any capture events.

Running the Tracy profiler UI, connecting and visualizing

While the program that you want to profile is still running (possibly thanks to TRACY_NO_EXIT=1), start the Tracy profiler UI which we had built above. From the IREE root directory:

If you built it part of IREE:

./build/tracy/iree-tracy-profiler

If you built the standalone “server” on MAC or Linux:

./third_party/tracy/profiler/build/unix/Tracy-release

It should show a dialog offering to connect to a client i.e. a profiled program:

If connecting doesn't work:

• If the profiled program is on a separate machine, make sure you've correctly set up port forwarding.
  • On Android, the adb forward many need to be run again.
• Make sure that the profiled program is still running. Do you need TRACY_NO_EXIT=1?
• Kill the profiled program and restart it.

You should then start seeing a profile. The initial view should look like this:

Before going further, take a second to check that your recorded profile data has all the data that it should have. Permissions issues, as discussed above, could cause it to lack “sampling” or “CPU data” information. For example, here is what he initial view looks like when one forgot to run the profiled program as root on Desktop Linux (where running as root is required, as explained above):

Notice how the latter screenshot is lacking the following elements:

• No ‘CPU data’ header on the left side, with the list of all CPU cores. The ‘CPU usage’ graph is something else.
• No ‘ghost’ icon next to the ‘Main thread’ header.

When running into any of the above issues, refer to the above Permissions section. Look for any interesting stderr message (in the profiled program's terminal).

Click the ‘Statistics’ button at the top. It will open a window like this:

See how the above screenshot has two radio buttons at the top: ‘Instrumentation’ and ‘Sampling’. At this point, if you don't see the ‘Sampling’ radio button, you need to resolve that first, as discussed above about possible permissions issues.

These ‘Instrumentation’ and ‘Sampling’ statistics correspond the two kinds of data that Tracy collects about your program. In the Tracy main view, they correspond, respectively, to ‘instrumentation’ and ‘ghost’ zones. Refer to the Tracy PDF manual for a general introduction to these concepts. For each thread, the ghost icon toggles the view between these two kinds of zones.

Back to the main view, look for the part of the timeline that is of interest to you. Your area of interest might not be on the Main thread. In fact, it might be on a thread that's not visible in the initial view at all. Indeed, the initial view tends to be zoomed-in a lot. Either use the mouse directly to zoom out and navigate, to look for the ‘Frame’ control at the top of the Tracy window. Use the ‘next frame’ arrow button until more interesting threads appear. Typically, IREE generated code tends to run on a thread named worker[0] (the screenshot here shows old names that were like cpu0).

Once you have identified the thread of interest, use its ghost icon to toggle between instrumentation and ghost zones, and zoom until you have found the zone of interest.

Here is what you should get when clicking on a ghost zone:

The percentages column to the left of the disassembly shows where time is being spent. This is unique to the sampling data (ghost zones) and has no equivalent in the instrumentation data (instrumentation zones). Here is what we get clicking on the corresponding instrumentation zone:

This still has a ‘Source’ button but that only shows the last C++ caller that had explicit Tracy information, so here we see a file under iree/hal whereas the Ghost zone saw into the IREE compiled module that that calls into, with the source view pointing to the .mlir file.

Permissions issues

The profiled application (i.e. the Tracy client) needs to have appropriate permissions so perform the special I/O required to collect the profile information. This is OS-specific.

Desktop Linux

On desktop Linux, the profiled application must be run as root, e.g. with sudo. Otherwise, profile data will lack important components.

Android

On Android it is not necessary to run as root and in fact, Android graphical applications never run as root, so it's advisable to run all programs as non-root for consistency.

The Android device must be prepared as follows to enable Tracy profiling.

• The device must be rooted.
  • That means that in adb shell, the command su must succeed.
  • That does NOT mean doing adb root. The effect of adb root is to have the adbd daemon itself run as root, which causes adb shell to give you a root shell by default. If you are in that case, consider doing adb unroot to restart the adbd server as non-root. Not mandatory, but again, running anything as root on Android is a deviation from normal user conditions.
• Execute the following commands in a root shell on the device (i.e. adb shell, then su, then the following commands). These are from the manual, but hard to find there, and copy-pasting from PDF introduces unwanted whitespace. These settings normally persist until the next reboot of the device.
  • setenforce 0
  • mount -o remount,hidepid=0 /proc
  • echo 0 > /proc/sys/kernel/perf_event_paranoid

Port forwarding

The Tracy client and server communicate by default over port 8086. When they run on different machines, e.g. with embedded/Android profiling or remote profiling, port forwarding must be set up.

Between a computer and a local Android device connected to it by USB

Run this command. You might need to run it again more a little frequently than you reboot the device. When experiencing connection issues, try that first.

adb forward tcp:8086 tcp:8086

Between two computers over the network

Setup you ssh tunnel Make sure you can telnet to your local port and you can hit the remote server port 8086. Once that is done supply the port address for the caputre tool to connect to.

Configuring Tracy instrumentation

Set IREE's IREE_TRACING_MODE value (defined in iree/base/tracing.h) to adjust which tracing features, such as allocation tracking and callstacks, are enabled.