ANGLE Performance Tests

angle_perftests is a standalone microbenchmark testing suite that contains tests for the OpenGL API. angle_trace_tests is a suite to run captures traces for correctness and performance. Because the traces contain confidential data, they are not publicly available. For more details on ANGLE's tracer please see the docs.

The tests currently run on the Chromium ANGLE infrastructure and report results to the Chromium perf dashboard. Please refer to thepublic dashboard docs for help

Running the Tests

You can follow the usual instructions to check out and build ANGLE. Build the angle_perftests or angle_trace_tests targets. Note that all test scores are higher-is-better. You should also ensure is_debug=false in your build. Running with angle_assert_always_on or debug validation enabled is not recommended.

Variance can be a problem when benchmarking. We have a test harness to run a tests repeatedly to find a lower variance measurement. See src/tests/run_perf_tests.py.

To use the script first build angle_perftests or angle_trace_tests, set your working directory your build directory, and invoke the run_perf_tests.py script. Use --test-suite to specify your test suite, and --filter to specify a test filter.

Choosing the Test to Run

You can choose individual tests to run with --gtest_filter=*TestName*. To select a particular ANGLE back-end, add the name of the back-end to the test filter. For example: DrawCallPerfBenchmark.Run/gl or DrawCallPerfBenchmark.Run/d3d11. Many tests have sub-tests that run slightly different code paths. You might need to experiment to find the right sub-test and its name.

Null/No-op Configurations

ANGLE implements a no-op driver for OpenGL, D3D11 and Vulkan. To run on these configurations use the gl_null, d3d11_null or vulkan_null test configurations. These null drivers will not do any GPU work. They will skip the driver entirely. These null configs are useful for diagnosing performance overhead in ANGLE code.

Command-line Arguments

Each test runs N trials and prints metrics for each trial. Trials are limited by time (default), step/frame limits can also be set. Note that at the beginning performance might be affected by hitting new code paths, cold caches etc (see warmup below) but longer runs on some devices trigger thermal throttling affecting performance (known: phones, desktop perf CI bots).

Several command-line arguments control how the tests run:

  • --run-to-key-frame: If the trace specifies a key frame, run to that frame and stop. Traces without a KeyFrames entry in their JSON will default to frame 1. This is primarily to save cycles on our bots that do screenshot quality comparison.
  • --enable-trace: Write a JSON event log that can be loaded in Chrome.
  • --trace-file file: Name of the JSON event log for --enable-trace.
  • --steps-per-trial x: Fixed number of steps to run for each test trial.
  • --max-steps-performed x: Upper maximum on total number of steps for the entire test run. For a quick smoke test, you can specify 1.
  • --render-test-output-dir=dir: Directory to store test artifacts (including screenshots but unlike --screenshot-dir, dir here is always a local directory regardless of platform and --save-screenshots isn't implied).
  • --verbose: Print extra timing information.
  • --trial-time x or --max-trial-time x: Run each test trial under this max time. Defaults to 10 seconds.
  • --fixed-test-time x: Run the tests until this much time has elapsed.
  • --warmup: Run a warmup phase before the test. Defaults to off.
  • --fixed-test-time-with-warmup x: Start with a warmup, then run the tests until this much time has elapsed.
  • --trials: Number of times to repeat testing. Defaults to 3.
  • --no-finish: Don't call glFinish after each test trial.
  • --validation: Enable serialization validation in the trace tests. Normally used with SwiftShader and retracing.
  • --perf-counters: Additional performance counters to include in the result output. Separate multiple entries with colons: ‘:’.

The command line arguments implementations are located in ANGLEPerfTestArgs.cpp.

Test Breakdown

Microbenchmarks

  • DrawCallPerfBenchmark: Runs a tight loop around DrawArarys calls.
    • validation_only: Skips all rendering.
    • render_to_texture: Render to a user Framebuffer instead of the default FBO.
    • vbo_change: Applies a Vertex Array change between each draw.
    • tex_change: Applies a Texture change between each draw.
  • UniformsBenchmark: Tests performance of updating various uniforms counts followed by a DrawArrays call.
    • vec4: Tests vec4 Uniforms.
    • matrix: Tests using Matrix uniforms instead of vec4.
    • multiprogram: Tests switching Programs between updates and draws.
    • repeating: Skip the update of uniforms before each draw call.
  • DrawElementsPerfBenchmark: Similar to DrawCallPerfBenchmark but for indexed DrawElements calls.
  • BindingsBenchmark: Tests Buffer binding performance. Does no draw call operations.
    • 100_objects_allocated_every_iteration: Tests repeated glBindBuffer with new buffers allocated each iteration.
    • 100_objects_allocated_at_initialization: Tests repeated glBindBuffer the same objects each iteration.
  • TexSubImageBenchmark: Tests glTexSubImage update performance.
  • BufferSubDataBenchmark: Tests glBufferSubData update performance.
  • TextureSamplingBenchmark: Tests Texture sampling performance.
  • TextureBenchmark: Tests Texture state change performance.
  • LinkProgramBenchmark: Tests performance of glLinkProgram.
  • glmark2: Runs the glmark2 benchmark.

Many other tests can be found that have documentation in their classes.

Trace Tests

Trace tests take command line arguments that pick the run configuration:

  • --use-gl=native: Runs the tests against the default system GLES implementation instad of your local ANGLE.
  • --use-angle=backend: Picks an ANGLE back-end. e.g. vulkan, d3d11, d3d9, gl, gles, metal, or swiftshader. Vulkan is the default.
  • --offscreen: Run with an offscreen surface instead of swapping every frame.
  • --vsync: Run with vsync enabled, and measure CPU and GPU work insead of wall clock time.
  • --minimize-gpu-work: Modify API calls so that GPU work is reduced to minimum.
  • --screenshot-dir dir: Directory to store test screenshots. Implies --save-screenshots. On Android this directory is on device, not local (see also --render-test-output-dir). Only implemented in TracePerfTest.
  • --save-screenshots: Save screenshots. Only implemented in TracePerfTest.
  • --screenshot-frame <frame>: Which frame to capture a screenshot of. Defaults to first frame (1). Using -1 will capture every frame rendered, including those after Reset for multiple loops. Only implemented in TracePerfTest.
  • --include-inactive-resources : Include all resources captured at trace-time during replay. Only resources which are active during trace execution are replayed by default.

For example, for an endless run with no warmup on swiftshader, run:

angle_trace_tests --gtest_filter=TraceTest.trex_200 --use-angle=swiftshader --trial-time 1000000

Understanding the Metrics

  • cpu_time: Amount of CPU time consumed by an iteration of the test. This is backed by GetProcessTimes on Windows, getrusage on Linux/Android, and zx_object_get_info on Fuchsia.
    • This value may sometimes be larger than wall_time. That is because we are summing up the time on all CPU threads for the test.
  • wall_time: Wall time taken to run a single iteration, calculated by dividing the total wall clock time by the number of test iterations.
    • For trace tests, each rendered frame is an iteration.
  • gpu_time: Estimated GPU elapsed time per test iteration. We compute the estimate using GLES timestamp queries at the beginning and ending of each test loop.
    • For trace tests, this metric is only enabled in vsync mode.