- Increase the Auto-LoD distance on tracks with a low scene complexity (fix#5065)
- Increase the Auto-LoD distance when the old formula gives a too low base distance (small objects). This helps with some of the worst popping issues for comparatively little performance cost.
- Make the LoD distance auto-compute code clearer by moving the settings-related multiplier after the squaring step
- Also fix a couple of unrelated warnings about comparison between signed and unsigned integers
With the existing engine and tracks, when the framerate is rather GPU-limited, the performance cost of increasing LoD distance is minor. It becomes more significant at higher framerates when the CPU can be more of a limiting factor, but it also noticeably enhances the player experience by removing the distracting 'popping' when an object visibly changes shape or appears because of LoD.
- Lowest is changed from 0.75 to 0.8
- Low is reduced to 0.8, just like Lowest. The difference is now only the display of optional geometry objects, but it's a big difference.
- Medium is changed from 1.25 to 1.3
- High is changed from 1.3 to 1.8
- Very high is changed from 1.8 to 2.4
- Ultra is changed from 3.0 to 3.2
- Ensure that optional geometry objects are properly removed. This provides a massive performance boost, at the cost of visuals. Benchmark on graphics 2 with lowest geometry level: bugged has 202 steady FPS, 230 mostly stable FPS, fixed has 427 steady FPS, 495 mostly stable FPS.
- Restrict the removal of objects based on geometry detail to the lowest geometry level. This ignore the different levels of geometry removal that track authors can set for objects, and which was mostly unused: all objects marked for removal at any level are either all removed or all kept.
The assert could be triggered when a new sign-in request was sent before the first one was resolved, usually when using the enter key to request the sign in. Since it can happen in normal operations to have the player profile status different from signing out or signed out, an assert is incorrect.
Also fix server-only compilation for the previous commit.
* Format the standard CPU and GPU profiler data as CSV, to simplify its analysis
* Include the internal identifier name of the track in the report name, or 'menu' if there is no active track. This uses the current track at the time of report saving. It is possible to keep the profiler on through multiple tracks and menus, but it's on people enabling the artist debug mode and the profiler to use it sensibly
* Stop the profiler when saving data, if it's not already stopped
* Compute, when profiling is stopped, for all integer FPS values up to 1000, three key metrics:
- The number of frames too slow to reach this FPS value
- The proportion of total time spent in frames slower than this FPS value
- The proportion of total time spent in frames beyond the maximum duration associated with this FPS value
* Compute three key indicator values:
- Steady FPS (highest FPS with no more than 1% time spent in slower frames and no more than 0.1% time spent beyond the maximum duration)
- Mostly stable FPS (highest FPS with no more than 12% time spent in slower frames and no more than 2% time spent beyond the maximum duration)
- Typical FPS (highest FPS with no more than 50% time spent in slower frames and no more than 10% time spent beyond the maximum duration)
* Save an additional 'perf-report' containing this new data
Traditional performance measurements, such as average FPS and 1% lows (which are typically based on the 1% slowest frames, not on the 1% slowest gameplay time), are flawed because they unduly reward outputting more fast frames. Increasing the number of fast frames without speeding up the slow frames is useless when it comes to making the game feel smooth, which is especially critical in a racing game.
Missing features:
- Displaying a summary of the performance test instead of the normal end screen
- Confirm to the user that the report has been successfully saved
- And more advanced options.
- Customize the race result screen to offer three options : leave to main menu, save test data, and return to video settings
- Ensure all three options do as they say
- Ensure that the benchmark mode and the profiler are always turned off when leaving through the end screen
* Use explicit activation and desactivation functions for the profiler to improve code clarity and robustness
* In the pause dialog, only display the 'back to race' and 'exit' buttons in benchmark mode
* Relabel the buttons in the pause menu in benchmark mode
* Use 'Performance test' in user-facing texts to ensure better translations
* Fix parts of the UI disappearing when leaving the pause menu in benchmark mode
Additional known issue introduced in the previous benchmark commit:
* With big text sizes, the performance test button overflows in the graphics settings.
* Add a button in the graphics options to start a benchmark
* Added logic to play a custom replay for benchmarking
* Added logic to automatically start and end the profiling as the race starts and ends, when in benchmark mode
* Disable the profiler drawings in benchmark mode, they take additional resources and are useless
* Keep the race going when the pause menu is opened, as it would otherwise distort the results
* Added logic to clean up the profiler data each time profiling is switched back from off to on. This avoids multiple profiling sessions piling up in one report, for example when benchmarking tw
* Added a sped up Black Forest replay for use during the benchmark
Missing features and known issues:
* The end screen should display a benchmark result summary instead of a normal end screen
* Entering and leaving the pause menu in benchmark mode makes some parts of the race UI disappear
* The end screen should send the player back to the graphics settings, not the main menu
* Pausing and quitting can leave the profiler enabled when it should not be
* The options in the pause menu should be customized in benchmark mode
* The replay used for benchmarking should be configurable in a data file
* Automatically testing multiple graphics settings and providing an overall summary
* And more advanced options.
* Quick cull invisible LOD nodes
* Reduce dynamic cast
* Fix for Vulkan
Maintainer's comment: Helps only in select tracks and more in already fast frames than in the more problematic slow frames, but an improvement nonetheless.
