Trace and Replay Tools
This document describes the vktrace and vkreplay tools which are used for tracing and replaying Vulkan API calls.
Tracing and Trace Files
The trace program is named
vktrace. It is used to record an application's Vulkan API calls to a trace file. The call information is stored in the trace file in a compact binary format. The trace files normally have a
.vktrace suffix. The application can be either a local or remote application.
Options for the
vktrace command are:
|Command line arguments to pass to the application to be traced||none|
|Name of the generated trace file||vktrace_out.vktrace|
|Name of the application to trace||if not provided, server mode tracing is enabled|
|Print trace messages to console||on|
|Frame numbers of which to take screen shots. String arg is one of:
comma separated list of frames
<start>- <count>- <interval>
|Alternate working directory||the application's directory|
|Trace persistently mapped buffers||true|
|Start/stop trim by hotkey or frame range. String arg is one of:
|Verbosity mode - "quiet", "errors", "warnings", or "full"||errors|
In local tracing mode, both the
vktrace and application executables reside on the same system.
An example command to trace the sample
cube application in local mode follows.
$ vktrace -p ./cube -o cubetrace.vktrace
PS> vktrace -p cube.exe -o cubetrace.vktrace
Note: Subsequent command examples in this document are written using Linux shell commands. These example commands can be translated and used as Windows Powershell commands.
Trace packets are written to the file
cubetrace.vktrace in the local directory. Output messages from the replay operation are written to
vktrace runs with the same
-o option value will overwrite the trace file, preventing the generation of multiple, large trace files. Be sure to specify a unique output trace file name for each
vktrace invocation if you do not desire this behaviour.
The tools also support tracing Vulkan applications in client/server mode, where the trace server resides on a local or a remote system.
In client/server mode, the
vktrace server is started without specifying the
--Program option. Invoked in this fashion,
vktrace enters a waiting mode for trace packets.
$ mkdir examples/traces $ cd examples/traces $ vktrace -o cubetrace_s.vktrace No program (-p) parameter found: Running vktrace as server.
The trace file will be written to
cubetrace_s.vktrace. If additional programs are traced with this trace server, subsequent trace files will be named
>.vktrace, with the trace server incrementing
N for each time the application is run.
The tracer is implemented as a Vulkan layer. When tracing in server mode, the local or remote client must enable the
Vktrace layer. The
Vktrace layer must be the first layer identified in the
$ export VK_INSTANCE_LAYERS=VK_LAYER_LUNARG_vktrace
Local client/server mode is particularly useful for Vulkan applications with multiple runtime options or complicated startup.
In local client/server mode, the tracer layer must be enabled before the local client application to be traced is launched.
$ cd examples/build $ export VK_INSTANCE_LAYERS=VK_LAYER_LUNARG_vktrace $ ./cube
The generated trace file is found at
cube application is used to demonstrate tracing in client/server mode.
cube is a very simple application, and would not be traced in this mode in practice.
Remote client/server mode is useful if the client is a mobile device or running in a lightweight environment that might not have the disk or other capabilities to store large trace files.
In remote client/server mode, the remote client must additionally identify the IP address of the trace server system.
$ cd examples/build $ export VKTRACE_LIB_IPADDR=<ip address of trace server system> $ export VK_INSTANCE_LAYERS=Vktrace $ ./cube
The generated trace file is located at
examples/traces/cubetrace_s.vktrace on the remote trace server system.
The vkreplay command is used to replay a Vulkan application trace.
Important: Trace files generated with earlier versions of the vktrace tools may not be replayable with later versions of the tools. The trace file format has evolved over time - vkreplay attempts to identify incompatible versions during replay and can often successfully replay them, but it does not handle all such cases.
Important: Trace files generated on one GPU may or may not be replayable on other GPUs, as well as trace files generated on different OSes and different driver versions. vkreplay attempts to replay such trace files, translating API calls as needed for the new platform, but it does not handle all such cases.
vkreplay command-line options are:
|Name of trace file to open and replay||required|
|Number of times to replay the trace file||1|
|The start frame number of the loop range||0|
|The end frame number of the loop range||the last frame in the tracefile|
|Enable compatibility mode - modify api calls as needed when replaying trace file created on different platform than replay platform. For example: Convert trace file memory indices to replay device memory indices.||true|
|Comma-separated list of frame numbers of which to take screen shots||no screenshots|
|Color Space format of screenshot files. Formats are UNORM, SNORM, USCALED, SSCALED, UINT, SINT, SRGB||Format of swapchain image|
|Verbosity mode - "quiet", "errors", "warnings", or "full"||errors|
To replay the cube application trace captured in the example above:
$ cd examples/build $ vkreplay -o cubetrace.vktrace -l 5
If the trace is rather short, the replay may finish quickly. Specify the
--NumLoops option to replay the trace
NumLoops option value times.
Output messages from the replay operation are written to
Replayer Interaction with Layers
The Vulkan validation layers may be enabled for trace replay. Replaying a trace with layers activated provides many benefits. Developers can take advantage of new validation capabilities as they are developed with older and existing trace files.
To activate specific layers on a trace replay, set the
VK_INSTANCE_LAYERS environment variable to a colon-separated list of layer names before replaying the trace. Refer to the Vulkan Validation and Debugging Layers guide for additional information on layers and how to configure layer output options.
The vktraceviewer tool allows interactive creation and viewing of Vulkan trace files. In the future, it will include support for interactively playing back trace files. This is alpha software.
vktraceviewer has a dependency on Qt5 libraries; they need to be installed on the system.
Persistently Mapped Buffers and vktrace
If a Vulkan program uses persistently mapped buffers (PMB) that are allocated via vkMapMemory, vktrace can track changes to PMB and automatically copy modified PMB pages to the trace file, rather than requiring that the Vulkan program call vkFlushMappedMemoryRanges to specify what PMB buffers should be copied. On Windows, the trace layer detects changes to PMB pages by setting the PAGE_GUARD flag for mapped memory pages and installing an exception handler for PAGE_GUARD that keeps track of which pages have been modified. On Linux, the trace layer detects changes to PMB pages by examining /proc/self/pagemap.
Tracking of changes to PMB using the above techniques is enabled by default. If you wish to disable PMB tracking, it can be disabled by with the
--PMB false option to the vktrace command. Disabling PMB tracking can result in some mapped memory changes not being detected by the trace layer, a larger trace file, and/or slower trace/replay.
Trace Tools Enviroment Variables
Several environment variables can be set to change the behavior of vktrace/vkrepay:
VKTRACE_PMB_ENABLE enables tracking of PMB if its value is 1. Other values disable PMB tracking. If this environment variable is not set, PMB tracking is enabled. When creating a trace using client/server mode, set this variable to 0 when starting the client if you wish to disable PMB tracking.
VKTRACE_PAGEGUARD_ENABLE_READ_PMB enables read PMB support if set to a non-NULL value. If PMB data changes comes from the GPU side, PMB tracking does not usually capture those changes. This environment variable is used to enable capture of such GPU initiated PMB data changes. It is supported only on Windows.
VKTRACE_PAGEGUARD_ENABLE_READ_POST_PROCESS, when set to a non-null value, enables post processing when read PMB support is enabled. When VKTRACE_PAGEGUARD_ENABLE_READ_PMB is set, PMB processing will sometimes miss writes following reads if writes occur on the same page as a read. Set this environment variable to enable post processing to fix missed pmb writes. It is supported only on Windows.
To record a trace on Android, enable port forwarding from the device to the host:
adb reverse localabstract:vktrace tcp:34201
Start up vktrace on the host in server mode:
vktrace -v full -o cube.vktrace
Run your application with the following layer enabled:
The trace will be recorded on the host.
To replay a trace, push the trace to your device
adb push cube.vktrace /sdcard/cube.vktrace
Grant vkreplay the ability to read it
adb shell pm grant com.example.vkreplay android.permission.READ_EXTERNAL_STORAGE adb shell pm grant com.example.vkreplay android.permission.WRITE_EXTERNAL_STORAGE
And start the native activity
adb shell am start -a android.intent.action.MAIN -c android-intent.category.LAUNCH -n com.example.vkreplay/android.app.NativeActivity --es args "-v\ full\ -t\ /sdcard/cube.vktrace"
Linux Display Server Support
To build vkreplay to use a different display server implementation than XCB, the CMake variable VKREPLAY_WSI_SELECTION can be set. Currently, the available options are XCB and WAYLAND.
Example for buiding for Wayland: (see 'Linux Build')
cmake -H. -Bdbuild -DCMAKE_BUILD_TYPE=Debug -DVKREPLAY_WSI_SELECTION=WAYLAND
Note: In the future, vkreplay will choose a display server based on a command-line argument.