aosp12/external/mesa3d/.gitlab-ci/tracie

Tracie - Mesa Traces Continuous Integration System

Home of the Mesa trace testing effort.

Traces definition file

The trace definition file contains information about the traces to run along with their expected image checksums on each device, and optionally from where to download them. An example:

traces-db:
  download-url: https://minio-packet.freedesktop.org/mesa-tracie-public/

traces:
  - path: glmark2/jellyfish.rdc
    expectations:
      - device: gl-intel-0x3185
        checksum: 58359ea4caf6ad44c6b65526881bbd17
      - device: gl-vmware-llvmpipe
        checksum: d82267c25a0decdad7b563c56bb81106
  - path: supertuxkart/supertuxkart-antediluvian-abyss.rdc
    expectations:
      - device: gl-intel-0x3185
        checksum: ff827f7eb069afd87cc305a422cba939

The traces-db entry can be absent, in which case it is assumed that the traces can be found in the CWD/traces-db directory.

Traces that don't have an expectation for the current device are skipped during trace replay.

Adding a new trace to the list involves commiting the trace to the git repo and adding an entry to the traces list. The reference checksums can be calculated with the image_checksum.py script. Alternatively, an arbitrary checksum can be used, and during replay (see below) the scripts will report the mismatch and expected checksum.

Trace-db download urls

The trace-db:download-url property contains an HTTPS url from which traces can be downloaded, by appending traces:path properties to it.

Enabling trace testing on a new device

To enable trace testing on a new device:

1. Create a new job in .gitlab-ci.yml. The job will need to be tagged to run on runners with the appropriate hardware.

   1. If you mean to test GL traces, use the .traces-test-gl template jobs as a base, and make sure you set a unique value for the DEVICE_NAME variable and the name of the Mesa driver as DRIVER_NAME:

   my-hardware-gl-traces:
     extends: .traces-test-gl
     variables:
       DEVICE_NAME: "gl-myhardware"
       DRIVER_NAME: "mydriver"
   

   2. If you mean to test Vulkan traces, use the .traces-test-vk template jobs as a base, set the VK_DRIVER variable, and make sure you set a unique value for the DEVICE_NAME variable:

   my-hardware-vk-traces:
     extends: .traces-test-vk
     variables:
       VK_DRIVER: "radeon"
       DEVICE_NAME: "vk-myhardware"
       DRIVER_NAME: "radv"
   

2. Update the .gitlab-ci/traces-$DRIVER_NAME.yml file with expectations for the new device. Ensure that the device name used in the expectations matches the one set in the job. For more information, and tips about how to calculate the checksums, see the section describing the trace definition files.

Trace files

Tracie supports renderdoc (.rdc), apitrace (.trace) and gfxreconstruct (.gfxr) files. Trace files need to have the correct extension so that tracie can detect them properly.

The trace files that are contained in public traces-db repositories must be legally redistributable. This is typically true for FOSS games and applications. Traces for proprietary games and application are typically not redistributable, unless specific redistribution rights have been granted by the publisher.

Trace files in a given repository are expected to be immutable once committed for the first time, so any changes need to be accompanied by a change in the file name (eg. by appending a _v2 suffix to the file).

Replaying traces

Mesa traces CI uses a set of scripts to replay traces and check the output against reference checksums.

The high level script tracie.py accepts a traces definition file and the name of the device to be tested:

tracie.py --file .gitlab-ci/traces-llvmpipe.yml --device-name gl-vmware-llvmpipe

tracie.py copies the produced artifacts to the $CI_PROJECT_DIR/result directory. By default, created images from traces are only stored in case of a checksum mismatch. The TRACIE_STORE_IMAGES CI/environment variable can be set to 1 to force storing images, e.g., to get a complete set of reference images.

At a lower level the dump_trace_images.py script is called, which replays a trace, dumping a set of images in the process. By default only the image corresponding to the last frame of the trace is dumped, but this can be changed with the --calls parameter. The dumped images are stored in a subdirectory test/<device-name> next to the trace file itself, with names of the form tracefilename-callnum.png. The full log of any commands used while dumping the images is also saved in a file in the 'test/' subdirectory, named after the trace name with '.log' appended.

Examples:

python3 dump_traces_images.py --device-name=gl-vmware-llvmpipe mytrace.trace
python3 dump_traces_images.py --device-name=gl-vmware-llvmpipe --calls=2075,3300 mytrace.trace

Running the replay scripts locally

It's often useful, especially during development, to be able to run the scripts locally.

Depending on the target 3D API, the scripts require a recent version of apitrace being in the path, and also the renderdoc python module being available, for GL traces.

To ensure python3 can find the renderdoc python module you need to set PYTHONPATH to point to the location of renderdoc.so (binary python modules) and LD_LIBRARY_PATH to point to the location of librenderdoc.so. In the renderdoc build tree, both of these are in renderdoc/<builddir>/lib. Note that renderdoc doesn't install the renderdoc.so python module.

In the case of Vulkan traces, the scripts need a recent version of gfxrecon-replay being in the path, and also the VK_LAYER_LUNARG_screenshot Vulkan layer from LunarG's VulkanTools.

To ensure that this layer can be found when running the trace you need to set VK_LAYER_PATH to point to the location of VkLayer_screenshot.json and LD_LIBRARY_PATH to point to the location of libVkLayer_screenshot.so.

In the case of DXGI traces, the scripts require Wine, a recent version of DXVK installed in the default WINEPREFIX, and a recent binary version of apitrace for Windows which should be reachable through Windows' PATH environment variable.