Universal benchmark with 17 tests for CPU, graphics, memory and disk
Tests can be toggled and configured
Generates derailed reports in HTML
Remove MGB and FSPEED from System Panel, move them from IMG to ISO
- Add `bbench` to IMG and System Panel
- Universal benchmark with 17 tests for CPU, graphics, memory and disk
- Tests can be toggled and configured
- Generates derailed reports in HTML
- Remove `MGB` and `FSPEED` from System Panel, move them from IMG to ISO
<img alt="image.png" src="attachments/724df0b9-ec6d-4145-a701-cf86bb9a5591">
Functionality duplication.
This PR adds bbench to IMG image.
Should we move existing benchmarks from IMG to ISO and replace them with bbench in System Panel?
Blocker:
Closing the window mid-run leaks the temp file. AbortRequested() handles the window-close button (id==1) with an immediate ExitProcess(), bypassing Disk_Cleanup().
Fix: call Disk_Cleanup() before ExitProcess() (or from an exit handler).
Minor:
Seq Read depends on Seq Write.
With "Seq Write" unticked but "Seq Read" ticked, t_disk_read reads a missing bbtst.tmp → score 0 (no crash). The Random tests self-create the file via EnsureRandFile; the sequential ones don't. Give Seq Read the same guard.
LANG_RUS/LANG_ENG define is dead. Tupfile.lua passes /D=$(C_LANG) but no source references it - no localization despite the build wiring. Please, consider adding localization to all system languages.
Unbounded string buffers. disk_store slots (24 B), base[48], pf[80], disk_menu[260] are filled via strcpy/strcat from disk names with no length check. Safe for the short KolibriOS mount names, but still a risk of long name overflows.
Direct LFB write bypasses window clipping. vram_fill_gs writes the fill pattern to absolute screen coords for ~1 s; anything overlapping BirdBench's rectangle (another always-on-top window, a popup) gets painted over.
Cosmetic, but maybe worth fixing.
**Infrastructure:**
- **Functionality duplication.**
This PR adds `bbench` to IMG image.
Should we move existing benchmarks from IMG to ISO and replace them with `bbench` in System Panel?
**Blocker:**
- **Closing the window mid-run leaks the temp file.**
`AbortRequested()` handles the window-close button (`id==1`) with an immediate `ExitProcess()`, bypassing `Disk_Cleanup()`.
Fix: call `Disk_Cleanup()` before `ExitProcess()` (or from an exit handler).
**Minor:**
- **Seq Read depends on Seq Write.**
With "Seq Write" unticked but "Seq Read" ticked, `t_disk_read` reads a missing `bbtst.tmp` → score 0 (no crash). The Random tests self-create the file via `EnsureRandFile`; the sequential ones don't. Give Seq Read the same guard.
- **`LANG_RUS`/`LANG_ENG` define is dead.**
`Tupfile.lua` passes `/D=$(C_LANG)` but no source references it - no localization despite the build wiring. Please, consider adding localization to all system languages.
- **Unbounded string buffers.**
`disk_store` slots (24 B), `base[48]`, `pf[80]`, `disk_menu[260]` are filled via `strcpy`/`strcat` from disk names with no length check. Safe for the short KolibriOS mount names, but still a risk of long name overflows.
- **Direct LFB write bypasses window clipping.**
`vram_fill_gs` writes the fill pattern to absolute screen coords for ~1 s; anything overlapping BirdBench's rectangle (another always-on-top window, a popup) gets painted over.
Cosmetic, but maybe worth fixing.
Functionality duplication.
This PR adds bbench to IMG image.
Should we move existing benchmarks from IMG to ISO and replace them with bbench in System Panel?
You decide. Not in the scope of current push.
I would leave them.
Direct LFB write bypasses window clipping.
vram_fill_gs writes the fill pattern to absolute screen coords for ~1 s; anything overlapping BirdBench's rectangle (another always-on-top window, a popup) gets painted over.
Cosmetic, but maybe worth fixing.
This is fine.
> Functionality duplication.
> This PR adds bbench to IMG image.
> Should we move existing benchmarks from IMG to ISO and replace them with bbench in System Panel?
You decide. Not in the scope of current push.
I would leave them.
> Direct LFB write bypasses window clipping.
> vram_fill_gs writes the fill pattern to absolute screen coords for ~1 s; anything overlapping BirdBench's rectangle (another always-on-top window, a popup) gets painted over.
> Cosmetic, but maybe worth fixing.
This is fine.
Leency
requested review from Burer 2026-07-12 10:26:44 +00:00
Burer
changed title from bbench - Bird Benchmark for KolibriOS. Benches CPU and Memory, Graphics and Disk to WIP: apps/bbench - CPU, graphics, memory and disk benchmark for KolibriOS2026-07-13 10:13:36 +00:00
Burer
self-assigned this 2026-07-13 10:13:42 +00:00
Burer
marked the pull request as ready for review 2026-07-19 14:00:25 +00:00
A test that cannot run on this machine returned 0, which the section geomean
clamped to 1 and treated as the worst possible result: VRAM Write on a
banked-VGA mode, or a driver without folder ops, pulled a six-test section from
1000 down to about 300. Those now return BB_SKIP, carry no score, stay out of
the mean and appear as "skipped" in the report.
The REF_ values are guesses, and calibrating meant transcribing every metric by
hand, so nobody would. The report now ends with a ready-to-paste block holding
this machine's #define REF_ lines and its cpudb_add line. Scores only mean
something within one set of REF_ values, so BB_REVISION stamps the report and
cpudb entries from another revision are dropped instead of silently compared.
Also: bound the report buffer before writing to it, check the work-buffer
allocations, trim the VRAM row to whole dwords so the fill does not drift by a
byte per row at 24/16 bpp, and fold CompareRow's copy of the bar drawing into
Hbar.
The block is only useful while the scale is being set, and the REF_ names it
printed cost ~400 bytes of literals in every build. C-- has no macros with
arguments, so those could not be hidden behind the ifdef without one around
each of the seventeen registrations - drop the names and label the values with
the test name instead, which lands in the same order as the defines.
BB_REVISION stays unconditional: it is the version of the tests and their REF_
values, and a report that does not carry it cannot be compared with anything.
The CPU compare rows used name(22)+score(8)+mhz(11) against the test rows'
17+8+16. The totals matched, so the bars agreed, but the score and the third
column sat five characters apart. Three shared defines instead, wide enough
for the longest name in cpudb and the longest unit string.
BirdBench covers what mgb and fspeed measure, so keeping all three on a 1.44 MB
floppy buys nothing. They move to kolibrios/utils on the ISO and lose their
panel entries; the panel gains one Benchmark entry pointing at bbench, which
stays in the image. Icon 121 is the one mgb used - 48 already belongs to GrafX2.
All three syspanel.ini end up smaller than before, so none of them crosses into
a third 512-byte cluster. ru_RU is CP866 and es_ES is LF-only; both were edited
byte-wise to keep that.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
bad_Dr3dd0x
Burer
bbenchto IMG and System PanelMGBandFSPEEDfrom System Panel, move them from IMG to ISOInfrastructure:
This PR adds
bbenchto IMG image.Should we move existing benchmarks from IMG to ISO and replace them with
bbenchin System Panel?Blocker:
AbortRequested()handles the window-close button (id==1) with an immediateExitProcess(), bypassingDisk_Cleanup().Fix: call
Disk_Cleanup()beforeExitProcess()(or from an exit handler).Minor:
With "Seq Write" unticked but "Seq Read" ticked,
t_disk_readreads a missingbbtst.tmp→ score 0 (no crash). The Random tests self-create the file viaEnsureRandFile; the sequential ones don't. Give Seq Read the same guard.LANG_RUS/LANG_ENGdefine is dead.Tupfile.luapasses/D=$(C_LANG)but no source references it - no localization despite the build wiring. Please, consider adding localization to all system languages.disk_storeslots (24 B),base[48],pf[80],disk_menu[260]are filled viastrcpy/strcatfrom disk names with no length check. Safe for the short KolibriOS mount names, but still a risk of long name overflows.vram_fill_gswrites the fill pattern to absolute screen coords for ~1 s; anything overlapping BirdBench's rectangle (another always-on-top window, a popup) gets painted over.Cosmetic, but maybe worth fixing.
You decide. Not in the scope of current push.
I would leave them.
This is fine.
@Burer other are fixed
bbench - Bird Benchmark for KolibriOS. Benches CPU and Memory, Graphics and Diskto WIP: apps/bbench - CPU, graphics, memory and disk benchmark for KolibriOS13cc3fdd94to1175b4c906