README_ALGORITHM.md
1SkQP Render Test Algorithm
2==========================
3
4The following is a description of the render test validation algorithm that
5will be used by the version of SkQP that will be released for Android Q-release.
6
7There is a global macro constant: `SK_SKQP_GLOBAL_ERROR_TOLERANCE`, which
8reflects the `gn` variable `skia_skqp_global_error_tolerance`. This is usually
9set to 8.
10
11First, look for a file named `skqp/rendertests.txt` in the
12`platform_tools/android/apps/skqp/src/main/assets` directory. The format of
13this file is: each line contains one render test name, followed by a comma,
14followed by an integer. The integer is the `passing_threshold` for that test.
15
16For each test, we have a `max_image` and a `min_image`. These are PNG-encoded
17images stored in SkQP's APK's asset directory (in the paths `gmkb/${TEST}/min.png`
18and `gmkb/${TEST}/max.png`).
19
20The test input is a rendered image. This will be produced by running one of
21the render tests against the either the `vk` (Vulkan) or `gles` (OpenGL ES)
22Skia backend.
23
24Here is psuedocode for the error calculation:
25
26 function calculate_pixel_error(pixel_value, pixel_max, pixel_min):
27 pixel_error = 0
28
29 for color_channel in { red, green, blue, alpha }:
30 value = get_color(pixel_value, color_channel)
31 v_max = get_color(pixel_max, color_channel)
32 v_min = get_color(pixel_min, color_channel)
33
34 if value > v_max:
35 channel_error = value - v_max
36 elif value < v_min:
37 channel_error = v_min - value
38 else:
39 channel_error = 0
40 pixel_error = max(pixel_error, channel_error)
41
42 return max(0, pixel_error - SK_SKQP_GLOBAL_ERROR_TOLERANCE);
43
44 function get_error(rendered_image, max_image, min_image):
45 assert(dimensions(rendered_image) == dimensions(max_image))
46 assert(dimensions(rendered_image) == dimensions(min_image))
47
48 max_error = 0
49 bad_pixels = 0
50 total_error = 0
51
52 error_image = allocate_bitmap(dimensions(rendered_image))
53
54 for xy in list_all_pixel_coordinates(rendered_image):
55 pixel_error = calculate_pixel_error(rendered_image(xy),
56 max_image(xy),
57 min_image(xy))
58 if pixel_error > 0:
59 for neighboring_xy in find_neighbors(xy):
60 if not inside(neighboring_xy, dimensions(rendered_image)):
61 continue
62 pixel_error = min(pixel_error,
63 calculate_pixel_error(rendered_image(xy),
64 max_image(neighboring_xy),
65 min_image(neighboring_xy)))
66
67 if pixel_error > 0:
68 max_error = max(max_error, pixel_error)
69 bad_pixels += 1
70 total_error += pixel_error
71
72 error_image(xy) = linear_interpolation(black, red, pixel_error)
73 else:
74 error_image(xy) = white
75
76 return ((total_error, max_error, bad_pixels), error_image)
77
78For each render test, there is a threshold value for `total_error`, :
79`passing_threshold`.
80
81If `passing_threshold >= 0 && total_error > passing_threshold`, then the test
82is a failure and is included in the report. if `passing_threshold == -1`, then
83the test always passes, but we do execute the test to verify that the driver
84does not crash.
85
86We generate a report with the following information for each test:
87
88 backend_name,render_test_name,max_error,bad_pixels,total_error
89
90in CSV format in the file `out.csv`. A HTML report of just the failing tests
91is written to the file `report.html`. This version includes four images for
92each test: `rendered_image`, `max_image`, `min_image`, and `error_image`, as
93well as the three metrics: `max_error`, `bad_pixels`, and `total_error`.
94
95
96
97
README_GENERATING_MODELS.md
1How SkQP Generates Render Test Models
2=====================================
3
4We will, at regular intervals, generate new models from the [master branch of
5Skia][1]. Here is how that process works:
6
70. Choose a commit to make the branch from
8
9 COMMIT=origin/master
10
11 Or use the script to find the best one:
12
13 cd SKIA_SOURCE_DIRECTORY
14 git fetch origin
15 COMMIT=$(python tools/skqp/find_commit_with_best_gold_results.py \
16 origin/master ^origin/skqp/dev)
17
181. Get the positively triaged results from Gold and generate models:
19
20 cd SKIA_SOURCE_DIRECTORY
21 git fetch origin
22 git checkout "$COMMIT"
23 python tools/skqp/cut_release.py HEAD~10 HEAD
24
25 This will create the following files:
26
27 platform_tools/android/apps/skqp/src/main/assets/files.checksum
28 platform_tools/android/apps/skqp/src/main/assets/skqp/rendertests.txt
29 platform_tools/android/apps/skqp/src/main/assets/skqp/unittests.txt
30
31 These three files can be commited to Skia to create a new commit. Make
32 `origin/skqp/dev` a parent of this commit (without merging it in), and
33 push this new commit to `origin/skqp/dev`, using this script:
34
35 sh tools/skqp/branch_skqp_dev.sh
36
37 Review and submit the change:
38
39 git push origin HEAD:refs/for/skqp/dev
40 bin/sysopen https://review.skia.org/$(bin/gerrit-number HEAD)
41
42 (Optional) Make a SkQP APK.
43
44 tools/skqp/docker_build_universal_apk.sh
45
46 (Optional) Test the SkQP APK:
47
48 tools/skqp/test_apk.sh (LOCATION)/skqp-universal-debug.apk
49
50 (Once changes land) Upload the SkQP APK.
51
52 tools/skqp/upload_apk HEAD (LOCATION)/skqp-universal-debug.apk
53
54
55`tools/skqp/cut_release.py`
56---------------------------
57
58This tool will call `make_skqp_model` to generate the `m{ax,in}.png` files for
59each render test.
60
61Then it calls `jitter_gms` to see which render tests pass the jitter test.
62`jitter_gms` respects the `bad_gms.txt` file by ignoring the render tests
63enumerated in that file. Tests which pass the jitter test are enumerated in
64the file `good.txt`, those that fail in the `bad.txt` file.
65
66Next, the `skqp/rendertests.txt` file is created. This file lists the render
67tests that will be executed by SkQP. These are the union of the tests
68enumerated in the `good.txt` and `bad.txt` files. If the render test is found
69in the `good.txt` file and the model exists, its per-test threshold is set
70to 0 (a later CL can manually change this, if needed). Otherwise, the
71threshold is set to -1; this indicated that the rendertest will be executed (to
72verify that the driver will not crash), but the output will not be compared
73against the model. Unnecessary models will be removed.
74
75Next, all of the files that represent the models are uploaded to cloud storage.
76A single checksum hash is kept in the `files.checksum` file. This is enough
77to re-download those files later, but we don't have to fill the git repository
78with a lot of binary data.
79
80Finally, a list of the current gpu unit tests is created and stored in
81`skqp/unittests.txt`.
82
83[1]: https://skia.googlesource.com/skia/+log/master "Skia Master Branch"
84[2]: https://gold.skia.org/search "Skia Gold Search"
85
