nesdev.com
https://forums.nesdev.com/

visual timing tests for hblank / sprite 0 hit
https://forums.nesdev.com/viewtopic.php?f=3&t=16308
Page 1 of 1

Author:  rainwarrior [ Sun Aug 06, 2017 1:08 am ]
Post subject:  visual timing tests for hblank / sprite 0 hit

When discussing trying to do a palette swap in hblank in this thread, I had a thought that a visual test could help calibrate timings for emulators.

I wrote this very quick/simple visual test as a very rough start. This generates a sprite 0 hit, waits a configurable number of cycles, then writes $2001 to set the greyscale flag, giving a visual indication of where that occurs. This is intended to give a rough estimate of where emulators think the horizontal blank is, and maybe try to shed some light on the question "how much can I trust an emulator with hblank timing?

In this initial rough test, I determined where the first and last "off-screen" delay settings appear:
Code:
My NES/Famicom: $13-$29
FCEUX New PPU:  $10-$28
Nintendulator:  $12-$28
nestopia:       $13-$28

Note: on my NES and Famicom the change is clearly visible in the areas to the left and right of the picture, i.e. I can see about 8 pixels to the left and 5 pixels to the right of the rendered picture on my TV, so at delays adjacent to the ones stated I can still see the visual change, but they're off the edge of the picture an emulator would display.

So, I mean this is hopeful looking at least. In general emulators seem to agree where the end of hblank is relative to a sprite 0 hit, at least at the CPU cycle granularity. I think Nestopia and Nintendulator are very close to each other, like maybe 1 pixel different at either end, and both seem to show the change about 2 pixels later than my NES. (The CPU cycle counts above exaggerate the difference slightly, but I'm not sure how accurately I can measure pixels visually on my TV at the moment.)

FCEUX's New PPU has an 8-pixel granularity on its rendering, so I think that mostly accounts for the difference in its low delay value (i.e. the appearance of the write is delayed up to 8 pixels anyway, 2.6 cycles).

My NES and Famicom appear to be consistent with each other.

In the debuggers, however, the FCEUX's "pixel" value seems to be about 25 higher than Nintendulator's "tick" value, and I don't know why. I'm not sure if this is just a different numbering convention or what... will have to investigate.


So that's just one quick test trying to see if there's much variability between emulators. It could easily be adapted to investigate whether the timing consequences of other effects diverge, scroll changes, palette changes, etc. (i.e. things which may have their specific timings emulated differently) the source is included if you want to modify it for different tests.

Attachments:
hblank_test.zip [5.12 KiB]
Downloaded 58 times

Author:  Sour [ Sun Aug 06, 2017 5:14 am ]
Post subject:  Re: visual timing tests for hblank / sprite 0 hit

Thanks for making this!

Mesen displays no glitch between $14 and $29. $13 & $2A have a pixel or 2 flashing.
This test is affected by the logic Mesen/Bizhawk implement related to the grayscale bit, though.
According to Alyosha and other people's findings, the effect of the grayscale bit is delayed by 2 pixels (link)

So changing that delay to 1 pixel, for example, will make Mesen display the glitch between $13 & $28 (but results in a less accurate result for demo_ntsc, which Nestopia/Nintendulator both fail, iirc).

Other emulators:
puNES: $12-$28
MyNes (the emulator): $12-$28
Nintaco: $13-$28
Bizhawk: $14-$29

So I guess no emulator gets its exactly the same as your NES, but most of them are pretty close.
Assuming the other emus don't implement the 2-cycle delay (and assuming the NES actually has this delay) and adjusting for it:
NES: $13-$29
Bizhawk/Mesen: $14-$29
puNES/MyNes/Nintendulator: $14-$2A
Nintaco/Nestopia: $15-$2A
FCEUX: $12-$2A

Author:  Zepper [ Sun Aug 06, 2017 6:35 am ]
Post subject:  Re: visual timing tests for hblank / sprite 0 hit

*ahem*

RockNES: $12-$28.

Author:  Alyosha_TAS [ Sun Aug 06, 2017 8:45 am ]
Post subject:  Re: visual timing tests for hblank / sprite 0 hit

Oh cool another test using the black / white flag. It would be good to nail down the exact behaviour here. I really don't see any way the nmi_sync test can work without some kind of delay involved with this flag, whether in latching the write or in applying the correction somewhere later in the rendering pipeline.

Hopefully this can shed some light on it.

Author:  rainwarrior [ Sun Aug 06, 2017 12:56 pm ]
Post subject:  Re: visual timing tests for hblank / sprite 0 hit

Well, really the test should be measured in pixels not CPU cycles. The point of my using CPU cycles was to estimate how much range software has for getting this correct. A rig for testing how many cycles you can push your code back or forth until it will break.

If you're trying to tune an emulator, you'll want to measure exactly how many pixels and compare to an NES. This was proving difficult to measure on my TV though for the moment, the signal quality made it hard to tell. Probably there's a better choice of background image to compare against I could make here to help...

Also, maybe I should allow moving the sprite as well. You could get pixel resolution instead of cycle resolution...

Author:  Alyosha_TAS [ Mon Aug 07, 2017 3:04 pm ]
Post subject:  Re: visual timing tests for hblank / sprite 0 hit

rainwarrior wrote:
Well, really the test should be measured in pixels not CPU cycles. The point of my using CPU cycles was to estimate how much range software has for getting this correct. A rig for testing how many cycles you can push your code back or forth until it will break.

If you're trying to tune an emulator, you'll want to measure exactly how many pixels and compare to an NES. This was proving difficult to measure on my TV though for the moment, the signal quality made it hard to tell. Probably there's a better choice of background image to compare against I could make here to help...

Also, maybe I should allow moving the sprite as well. You could get pixel resolution instead of cycle resolution...


Also it would be cool to be able to check the timing relative to other similar PPU signals, like BG enable/disable. For example if on one scanline you enable greyscale and exactly one scanline later you disable BG, then it would be obvious if there was any relative delay between the two. Just an idea.

Author:  tokumaru [ Mon Aug 07, 2017 4:31 pm ]
Post subject:  Re: visual timing tests for hblank / sprite 0 hit

Alyosha_TAS wrote:
For example if on one scanline you enable greyscale and exactly one scanline later you disable BG

Problem is that there's no way to do something "exactly one scanline later", since a scanline doesn't have a whole number of CPU cycles (NTSC = 341 / 3 = 113.666... PAL = 341 / 3.2 = 106,5625). So I guess that the best you can do in NTSC is "exactly 3 scanlines later", and "exactly 16 scanlines later" in PAL.

Author:  Zepper [ Mon Aug 07, 2017 4:58 pm ]
Post subject:  Re: visual timing tests for hblank / sprite 0 hit

tokumaru wrote:
Alyosha_TAS wrote:
For example if on one scanline you enable greyscale and exactly one scanline later you disable BG

Problem is that there's no way to do something "exactly one scanline later", since a scanline doesn't have a whole number of CPU cycles (NTSC = 341 / 3 = 113.666... PAL = 341 / 3.2 = 106,5625). So I guess that the best you can do in NTSC is "exactly 3 scanlines later", and "exactly 16 scanlines later" in PAL.

...and blargg never did that on his tests?

Author:  tokumaru [ Mon Aug 07, 2017 5:14 pm ]
Post subject:  Re: visual timing tests for hblank / sprite 0 hit

Zepper wrote:
...and blargg never did that on his tests?

Even on blargg's demos the effects shift 1 or 2 pixels horizontally from scanline to scanline. It's mathematically impossible to trigger something on the exact same pixel on consecutive scanlines, because CPU would have to wait a fractional number of cycles, which is just not possible.

Just look at blargg's full palette demo... the boundaries between colors are slightly jittery, and only align vertically every 3 scanlines, because that's how long it takes to realign the CPU and the PPU. Not only that, but the number of CPU cycles in a frame is also not a whole number, so the jittery boundaries change position from frame to frame. Brargg got as close as possible to a straight vertical line, it's mathematically impossible to do any better.

Page 1 of 1 All times are UTC - 7 hours
Powered by phpBB® Forum Software © phpBB Group
http://www.phpbb.com/