3.8 KiB

Raw Blame History

Benchmark demo

Overview

The benchmark demo tests the performance in various cases. For example rectangle, border, shadow, text, image blending, image transformation, blending modes, etc. All tests are repeated with 50% opacity.

The size and position of the objects during testing are set with a pseudo random number to make the benchmark repeatable.

On to top of the screen the title of the current test step, and the result of the previous step is displayed.

Run the benchmark

In lv_conf.h or equivalent places set LV_USE_DEMO_BENCHMARK 1
After lv_init() and initializing the drivers call lv_demo_benchmark()
If you only want to run a specific scene for any purpose (e.g. debug, performance optimization etc.), you can call lv_demo_benchmark_run_scene() instead of lv_demo_benchmark()and pass the scene number.
If you enabled trace output by setting macro LV_USE_LOG to 1 and trace level LV_LOG_LEVEL to LV_LOG_LEVEL_USER or higher, benchmark results are printed out in csv format.

Interpret the result

The FPS is measured like this:

load the next step
in the display driver's monitor_cb accumulate the time-to-render and the number of cycles
measure for 1 second
calculate FPS = time_sum / render_cnt

Note that it can result in very high FPS results for simple cases. E.g. if some simple rectangles are drawn in 5 ms, the benchmark will tell it's 200 FPS. So it ignores LV_DISP_REFR_PERIOD which tells LVGL how often it should refresh the screen. In other words, the benchmark shows the FPS from the pure rendering time.

By default, only the changed areas are refreshed. It means if only a few pixels are changed in 1 ms the benchmark will show 1000 FPS. To measure the performance with full screen refresh uncomment lv_obj_invalidate(lv_scr_act()) in monitor_cb() in lv_demo_benchmark.c.

If you are doing performance analysis for 2D image processing optimization, LCD latency (flushing data to LCD) introduced by disp_flush() might dilute the performance results of the LVGL drawing process, hence make it harder to see your optimization results (gain or loss). To avoid such problem, please:

Temporarily remove the code for flushing data to LCD inside disp_flush(). For example:

static void disp_flush(lv_disp_drv_t * disp_drv, const lv_area_t * area, lv_color_t * color_p)
{
#if 0 //!< remove LCD latency
    GLCD_DrawBitmap(area->x1,               //!< x
                    area->y1,               //!< y
                    area->x2 - area->x1 + 1,    //!< width
                    area->y2 - area->y1 + 1,    //!< height
                    (const uint8_t *)color_p);
#endif
    /*IMPORTANT!!!
     *Inform the graphics library that you are ready with the flushing*/
    lv_disp_flush_ready(disp_drv);
}

Use trace output to get the benchmark results by:
- Setting macro LV_USE_LOG to 1
- Setting trace level LV_LOG_LEVEL to LV_LOG_LEVEL_USER or higher.

Result summary

In the end, a table is created to display measured FPS values.

On top of the summary screen, the "Weighted FPS" value is shown. In this, the result of the more common cases are taken into account with a higher weight.

"Opa. speed" shows the speed of the measurements with opacity compared to full opacity. E.g. "Opa. speed = 90%" means that rendering with opacity is 10% slower.

In the first section of the table, "Slow but common cases", those cases are displayed which are considered common but were slower than 20 FPS.

Below this in the "All cases section" all the results are shown. The < 10 FPS results are shown with red, the >= 10 but < 20 FPS values are displayed with orange.

3.8 KiB Raw Blame History