Dynamically allocate arrays using `MAX_SCREENWIDTH`/`HEIGHT` #45

MrAlaux · 2023-06-20T02:50:42Z

Alright, let's get some automated builds and test performance.

It's not without its bugs: flat rendering is distorted and weapon sprites can be drawn cut off when changing resolutions.

If `cachedheight` is a dynamic array, that `memset()` call is problematic. We'll just keep it fixed-size for now.

Well, that was quick...

liPillON · 2023-06-20T13:17:01Z

Hi, finally had the time to try the build resulting from this pr, here are the result.
For comparison, I've included a bunch of data from previous tests.

It seems that this PR has brought Nugget's performances in line with Woof 11.2
The only exceptions being:

demonastery (better than nugg-master, worst than woof-11.2)
alien vendetta (better than both nugg-master and woof-11.2)
pln.wad (same as nugg-master, worst than woof-11.2)

Hopefully when the optimizations from Woof will be merged in Nugget, the performance boost will compensate the costs of making 800p and 1600p available.

I'm still curious about testing the current PR with MAX_HIRES 2

pwad	woof 11.2	woof master	nugg master	nugg pr45
pln	63	106	33	34
eviternity	315	338	290	312
alien vendetta	575	609	550	596
saturnine chapel	260	268	245	258
grove	249	258	238	250
demonastery	471	479	415	431

MrAlaux · 2023-06-20T19:04:56Z

Hmm... I take it these tests were performed by warping to some maps and standing still, right? Any chance you could try some timedemos for each build?

liPillON · 2023-06-20T19:34:44Z

With the exception of pln.wad, all tests were timedemos (three runs each, average fps).
This is the same methodology I've used for the results posted in #44

win11 x64
intel i7 1260p
16gb ddr4 dual channel
iris xe graphics
1920x1200 display
-noautoload -nosound
clean-from-scratch cfg (only changes: vsync off, widescreen auto, demo progeessbar on)
between each run I waited for cpu/gpu cool off, no thermal throttling should have happened

MrAlaux · 2023-06-20T21:24:49Z

I see, thanks.

@ceski-1 thoughts?

ceski-1 · 2023-06-20T21:36:00Z

Looks like it's time to review visplane_t again.

liPillON · 2023-06-21T08:36:15Z

for the record, the optimizations I've mentioned were introduced in Woof with these PRs:
fabiangreffrath#1108
fabiangreffrath#1110
fabiangreffrath#1111

fabiangreffrath · 2023-06-23T11:38:59Z

src/r_bsp.c

@@ -124,7 +124,7 @@ static void R_ClipWallSegment(int first, int last, boolean solid)

 void R_ClearClipSegs (void)
 {
-  memset(solidcol, 0, MAX_SCREENWIDTH);
+  memset(solidcol, 0, SCREENWIDTH << hires); // [Nugget] Dynamic arrays


You should be able to get away with viewwidth whenever you use SCREENWIDTH << hires in this code change.

src/r_draw.c

src/r_main.c

src/z_zone.h

liPillON · 2023-06-23T13:04:31Z

BTW I re-run all timedemos because while testing the latest PR I noticed some inconsistencies..
Here are the results: I believe they are more trustworthy since this time I used the same methodology for all tests (the previous table was simple an accumulation of old results).

All timedemos were ran 5 times (letting the cpu cool down in between runs).
Best and Worst results were excluded.
I then calculated the average of the remaining 3 results.

For PLN.WAD I simply warped into the level and stood still in one corner of the map, looking at the opposite corner.

Noteworthy considerations:

nugg 1.14 performing better than woof 11.2
sensible performance gain in woof master (artifacts run 798: various optimizations)
sensible performance hit in nugget master (run69: supporting rendering resolutions up to 1600p)
even stronger performance hit with this PR (pr45, run67: dynamic array allocations)

In particular, the last point confirms what @MrAlaux was anticipating here #44 (comment)

Sorry for the misleading previous results...

TIMEDEMOS	woof 11.2	woof run798	nugg 1.14	nugg run69	nugg pr45
alien vendetta MAP20	605	643	637	612	582
demonastery	458	508	498	460	448
eviternity MAP26	337	358	347	324	312
grove	256	266	264	253	248
planisfere 2	363	379	370	343	331
saturnine chapel	272	279	275	263	258
sunlust MAP30	275	285	278	258	256

VISPLANE TEST MAP	woof 11.2	woof run798	nugg 1.14	nugg run69	nugg pr45
PLN.WAD	60	100	59	35	35

DEMOS:
https://dsdarchive.com/files/demos/av/28588/av20-1357.zip
https://dsdarchive.com/files/demos/demonastery/48438/demonastery-1326.zip
https://dsdarchive.com/files/demos/eviternity/44536/evit26-658.zip
https://dsdarchive.com/files/demos/grove/12807/grov-724.zip
https://dsdarchive.com/files/demos/planisf2/48969/planisf2os910.zip
https://dsdarchive.com/files/demos/satchap/20510/satchap-706.zip
https://dsdarchive.com/files/demos/sunlust/57817/sl30m2321.zip

MrAlaux · 2023-06-23T18:06:29Z

Thanks for this, @liPillON.

It seems I was right about this PR's effects, then. If these changes are not improving performance in their current state, I don't think I'll be able to figure out how to make them do so. I'll try tweaking the allocation process as @fabiangreffrath suggested, but as it stands, that as far as I can go.

Nugget 11.4 performing better than Woof 11.2 was a complete surprise, I wonder how that happened.

The performance hit of allowing higher resolutions is certainly unwanted, but I guess we can live with it considering the unexpected performance gain over Woof, and also the fact that further down the road, Woof's latest optimizations will be merged. I might make a branch merging those optimizations early so you can test it.

liPillON · 2023-06-23T18:47:09Z

Sure! When the time comes, feel free to @ me in the discussion for issue 44

While we're here: what about benchmarking a MAX_HIRES 2 branch?

MrAlaux · 2023-06-25T09:32:59Z

I changed some things based on most of Fabian's suggestions, but to no avail it seems; the only change that seems to have increased performance was making negonearray static again, hinting at the fact that making arrays dynamic causes a performance loss. I don't think these last changes are worth benchmarking.

I'm pretty sure that I'm the one doing something wrong here, and I'm certain I won't be able to figure out what exactly. As it stands, I can't do anything else with this branch, but I guess I'll leave the PR open for a while in case anyone wants to chime in.

MrAlaux added 6 commits June 19, 2023 06:59

Initial implementation of dynamic arrays

0d8b49a

It's not without its bugs: flat rendering is distorted and weapon sprites can be drawn cut off when changing resolutions.

Misc changes

ee83f46

Temp. fix for flat distortion

b93e19e

If `cachedheight` is a dynamic array, that `memset()` call is problematic. We'll just keep it fixed-size for now.

Proper fix for flat distortion

25345f0

Well, that was quick...

Temp. fix for weapon sprite drawing

c490695

Clean up

b744f78

MrAlaux changed the title ~~Dynamically allocate screen arrays~~ Dynamically allocate most rendering-related arrays Jun 20, 2023

MrAlaux added 2 commits June 20, 2023 00:17

Attempt to fix workflow runs

82fa2dd

Fix crash, make negonearray dynamic

bf8de84

liPillON mentioned this pull request Jun 20, 2023

400p performance regression? #44

Closed

MrAlaux changed the title ~~Dynamically allocate most rendering-related arrays~~ Dynamically allocate arrays using MAX_SCREENWIDTH/HEIGHT Jun 22, 2023

fabiangreffrath reviewed Jun 23, 2023

View reviewed changes

src/r_draw.c Outdated Show resolved Hide resolved

fabiangreffrath reviewed Jun 23, 2023

View reviewed changes

src/r_draw.c Outdated Show resolved Hide resolved

fabiangreffrath reviewed Jun 23, 2023

View reviewed changes

src/r_main.c Show resolved Hide resolved

fabiangreffrath reviewed Jun 23, 2023

View reviewed changes

src/z_zone.h Outdated Show resolved Hide resolved

MrAlaux added 2 commits June 25, 2023 05:59

Misc. changes

9dc3cae

Make negonearray static again

2717c07

MrAlaux closed this Jul 10, 2023

liPillON mentioned this pull request Nov 28, 2023

any resolution (i.e. any multipe of 200p) rendering fabiangreffrath/woof#1292

Closed

MrAlaux deleted the screen_alloc branch February 23, 2024 08:26

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Dynamically allocate arrays using `MAX_SCREENWIDTH`/`HEIGHT` #45

Dynamically allocate arrays using `MAX_SCREENWIDTH`/`HEIGHT` #45

MrAlaux commented Jun 20, 2023

liPillON commented Jun 20, 2023 •

edited

MrAlaux commented Jun 20, 2023

liPillON commented Jun 20, 2023 •

edited

MrAlaux commented Jun 20, 2023

ceski-1 commented Jun 20, 2023

liPillON commented Jun 21, 2023 •

edited

fabiangreffrath Jun 23, 2023

liPillON commented Jun 23, 2023 •

edited

MrAlaux commented Jun 23, 2023 •

edited

liPillON commented Jun 23, 2023 •

edited

MrAlaux commented Jun 25, 2023

Dynamically allocate arrays using MAX_SCREENWIDTH/HEIGHT #45

Dynamically allocate arrays using MAX_SCREENWIDTH/HEIGHT #45

Conversation

MrAlaux commented Jun 20, 2023

liPillON commented Jun 20, 2023 • edited

MrAlaux commented Jun 20, 2023

liPillON commented Jun 20, 2023 • edited

MrAlaux commented Jun 20, 2023

ceski-1 commented Jun 20, 2023

liPillON commented Jun 21, 2023 • edited

fabiangreffrath Jun 23, 2023

Choose a reason for hiding this comment

liPillON commented Jun 23, 2023 • edited

MrAlaux commented Jun 23, 2023 • edited

liPillON commented Jun 23, 2023 • edited

MrAlaux commented Jun 25, 2023

Dynamically allocate arrays using `MAX_SCREENWIDTH`/`HEIGHT` #45

Dynamically allocate arrays using `MAX_SCREENWIDTH`/`HEIGHT` #45

liPillON commented Jun 20, 2023 •

edited

liPillON commented Jun 20, 2023 •

edited

liPillON commented Jun 21, 2023 •

edited

liPillON commented Jun 23, 2023 •

edited

MrAlaux commented Jun 23, 2023 •

edited

liPillON commented Jun 23, 2023 •

edited