|
|
|
|
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Am 17.01.2019 um 15:05 schrieb William F Pokorny:
>> - Performance has degraded a bit, but I'm willing to accept this for
>> the sake of extended functionality and easier maintenance.
>
> Hmm, I'm surprised some by this. Are your test character strings really
> short? In the existing text shape code all the characters ended up more
> or less as one huge glyph as you know. As the string to the text shape
> got large, performance slowed substantially. Your implementing
> internally as a union of prisms should address this shortcoming well -
> providing a lot of room for other code to slow without overall impact.
> Wonder... Is the prism object a lot slower than the text object was?
> (I've never compared)
My strings are roundabout one or two dozen characters.
My first guess was that it was caused by 3rd order vs. 2nd order
splines, but that doesn't seem to make much of a difference.
Another potential cause for slowdown could be rooted in prism being able
to support conic sweeps.
My current guess however is that there's some "poor man's bounding"
hidden in the TrueType code that the prism doesn't have.
What I can say is that the same text strings composed into a single
prism gave abysmal performance.
What I also can say is that my performance comparisons were only cursory.
>> - POV-Ray v3.7 (haven't checked whether this has been fixed already)
>
> It is fixed in the current 3.8 master. You picked up - or made
> independently - a change quite like one Jérôme made in hgpovray.
Yeah, found text to that effect in the changelog, so it must be true ;)
>> - (*) For normal TrueType fonts, that is. Based on the FreeType
>> documentation, I totally expect the new implementation to get this
>> wrong, and/or to mess up the "holes" (e.g. in characters like `o`),
>> with /some/ font file formats or even individual fonts. Please keep
>> your eyes peeled, I'm keen to get my hands on any such font specimens.
>
> I have somewhere a crude test set up / scene collection I created when I
> did my text shape hacks. I'll put it on my list to run those.
>
> Plus, IIRC, I saw font sites offering fonts in multiple formats. Might
> be we can turn up problems with holes / overlaps or whatever by running
> such format sets and comparing resultant images. I now have a pretty
> good semi-automated testing framework for that sort of scene to sceen
> image testing so shouldn't take too much to get something going -
> provided I'm right about the availability of fonts in multiple formats.
If you would like to pick up that glove some time in the near future,
I'd very much appreciate it.
Right now, I have found one font where the surface normal orientation is
broken for at least one character, namely the closing square bracket
("]") in the "Arial Black" font. I blame FreeType, with the mechanism
being that the font defines that glyph by scaled reference to the
opening square bracket glyph ("["), mirroring it horizontally (and thus
messing up control point order).
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
in news:5c40b549$1@news.povray.org clipka wrote:
> support conic sweeps
interesting feature for fonts ;)
ingo
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Am 17.01.2019 um 18:41 schrieb ingo:
> in news:5c40b549$1@news.povray.org clipka wrote:
>
>> support conic sweeps
>
> interesting feature for fonts ;)
Not currently supported; would need extra work to make sure all
characters converge towards the same apex point.
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Am 17.01.2019 um 15:05 schrieb William F Pokorny:
>> - Performance has degraded a bit, but I'm willing to accept this for
>> the sake of extended functionality and easier maintenance.
>>
>
> Hmm, I'm surprised some by this. Are your test character strings really
> short? In the existing text shape code all the characters ended up more
> or less as one huge glyph as you know. As the string to the text shape
> got large, performance slowed substantially.
Actually, wading through the old code for unrelated reasons, I just
noticed that this isn't true: The old `text` primitive has actually been
a CSG union all along, with one child per character.
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
On 2/3/19 4:04 AM, clipka wrote:
> Am 17.01.2019 um 15:05 schrieb William F Pokorny:
>
>>> - Performance has degraded a bit, but I'm willing to accept this for
>>> the sake of extended functionality and easier maintenance.
>>>
>>
>> Hmm, I'm surprised some by this. Are your test character strings
>> really short? In the existing text shape code all the characters ended
>> up more or less as one huge glyph as you know. As the string to the
>> text shape got large, performance slowed substantially.
>
> Actually, wading through the old code for unrelated reasons, I just
> noticed that this isn't true: The old `text` primitive has actually been
> a CSG union all along, with one child per character.
Hmm. Not my recollection or experience. I was focused on inside tests if
those were perhaps done differently than intersections. The glyph loop
range testing I added helped regular intersection performance too, but
less.
Anyway. I'll keep what you saw in mind. Possible the code work I did was
pointless, and the performance gains seen false, for reasons of
self-foolery.
I'm maintaining my text branch with the thought the new text object
might also be too slow for what I want. Means, if motivated, I can again
do performance comparisons to the 'old' text object as well as your new
one off master, but, I probably won't so long as the new is fast enough.
Bill P.
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Le 19-02-03 à 04:04, clipka a écrit :
> Am 17.01.2019 um 15:05 schrieb William F Pokorny:
>
>>> - Performance has degraded a bit, but I'm willing to accept this for
>>> the sake of extended functionality and easier maintenance.
>>>
>>
>> Hmm, I'm surprised some by this. Are your test character strings
>> really short? In the existing text shape code all the characters ended
>> up more or less as one huge glyph as you know. As the string to the
>> text shape got large, performance slowed substantially.
>
> Actually, wading through the old code for unrelated reasons, I just
> noticed that this isn't true: The old `text` primitive has actually been
> a CSG union all along, with one child per character.
That's coherent with the object count : Adding a long text object really
increase the number of objects in a scene.
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
clipka <ano### [at] anonymousorg> wrote:
> - Previously, the text was aligned so that x=0 coincided with the
> /actual/ left side of the first character (i.e. its outline). The new
> implementation instead has x=0 coincide with the first character's
> /nominal/ left side (i.e. its "character box"), leaving a small gap
> typically matching half the distance between characters. Since I expect
> this to give aesthetically more pleasing results with multiple lines, I
> hesitate to invest time and effort to reproduce the old behaviour.
*must go through old scenes to see if/where I make this ASSumption.*
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
On 2/3/19 10:35 AM, William F Pokorny wrote:
> On 2/3/19 4:04 AM, clipka wrote:
>> Am 17.01.2019 um 15:05 schrieb William F Pokorny:
>>
>>>> - Performance has degraded a bit, but I'm willing to accept this for
>>>> the sake of extended functionality and easier maintenance.
>>>>
>>>
>>> Hmm, I'm surprised some by this. Are your test character strings
>>> really short? In the existing text shape code all the characters
>>> ended up more or less as one huge glyph as you know. As the string to
>>> the text shape got large, performance slowed substantially.
>>
>> Actually, wading through the old code for unrelated reasons, I just
>> noticed that this isn't true: The old `text` primitive has actually
>> been a CSG union all along, with one child per character.
>
> Hmm. Not my recollection or experience. I was focused on inside tests if
> those were perhaps done differently than intersections. The glyph loop
> range testing I added helped regular intersection performance too, but
> less.
>
...
>
Trying to avoid sliding too much sideways into work I did almost two
years ago given I've already got more going than I'll ever finish. Can't
completely help it I guess. Over the past days, kept asking myself how
could I see such large performance improvements if the text object was
already a union (Christoph and Alain being almost certainly right).
- Remembered two years ago I was mostly going after performance using
hardware counter analysis. I was going after hot spots.
- Remembered a decade ago with the objectAsIso experiments how I looked
hard at converting other than the simplest csg to a mesh because the
inside test performance got so slow with larger/complex csg.
- Remembered Lanuhum's Blender hair scene and thinking it shouldn't
really be that slow, even with super slow sphere_sweeps.
- Remembered thinking on trying it years ago the +bm2 mode should
provide more performance than it does.
- Remembered learning early last year while finding the cause of a
particular speckling bug, that POV-Ray does inside tests when it creates
original rays.
- Remembered thinking fixes to that particular speckling bug and issues
like: https://github.com/POV-Ray/povray/issues/139 will likely require
multiple inside tests over one to account for floating point noise no
matter any improved accuracy.
---
These last two thoughts led me to doing a new performance test with our
ttf1.pov sample scene. One where all I changed was whether or not AA
used. My thinking is the performance improvement of my text branch
should be more or less the same with and without AA. It should be the
same, unless, the text object is a union and the real performance
problem is the csg inside test mechanism doing something like just
trundling through ALL the shapes in a csg.
Results were not similar. This leads me to a new suspicion / unproven
theory we are sitting on a csg inside test performance issue. One
perhaps affecting things generally. My code changes to the text object
only treated another problem.
I've spent zero time in the csg code and I'm working/playing elsewhere
for the foreseeable future. I'm put digging here on my, maybe, someday
list - for what little that's worth! My new theory is perhaps wrong too
- but there is something going on inside test wise which is not very
efficient.
Aside: Even looking at single shapes of a type, the inside test
performance is often awful. This is why that isosurface peeling paint
skin test of Thomas's superellipsoid using the hard/soft object patch
was so extremely slow. IIRC, a box replacement was more than 20x faster.
Bill P.
--------------------------- Data for those interested.
/usr/bin/time povray ttf1.pov +am2 +a0.1 +r4 +wt1 \
-j -cc -fn -d -p +w2000 +h2000
p380
---------
12.39user 0.04system 0:12.71elapsed With +a0.1
12.21user 0.03system 0:12.51elapsed
12.27user 0.03system 0:12.56elapsed
-----
36.87
4.40user 0.03system 0:04.70elapsed With -a0.1
4.36user 0.05system 0:04.65elapsed
4.40user 0.02system 0:04.66elapsed
----
13.16
p380 + my text shape branch
---------
12.08user 0.04system 0:12.40elapsed With +a0.1
12.08user 0.04system 0:12.36elapsed
12.18user 0.03system 0:12.45elapsed
-----
36.34 -1.44%
4.39user 0.01system 0:04.64elapsed With -a0.1
4.32user 0.04system 0:04.59elapsed
4.34user 0.03system 0:04.64elapsed
----
13.05 -0.84% (??? Hmm)
----------- Repeating a few previous perf tests for my text branch.
soft_object.pov p380 + hard/soft object only.
----
345.96user 0.08system 5:46.50elapsed As 'text union'
64.91user 0.04system 1:05.52elapsed Individual chars at top (-81.24%)
soft_object.pov p380 + hard/soft object + my text shape branch.
----
161.22user 0.05system 2:41.78elapsed As 'text union' -53.40%
69.42user 0.02system 1:09.99elapsed Individual chars at top (+6.95%)
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Am 07.02.2019 um 17:31 schrieb William F Pokorny:
> Trying to avoid sliding too much sideways into work I did almost two
> years ago given I've already got more going than I'll ever finish.
Welcome to the world of POV-Ray development ;)
> - Remembered a decade ago with the objectAsIso experiments how I looked
> hard at converting other than the simplest csg to a mesh because the
> inside test performance got so slow with larger/complex csg.
>
> - Remembered Lanuhum's Blender hair scene and thinking it shouldn't
> really be that slow, even with super slow sphere_sweeps.
Let's play a game of "shapes that could benefit from internal bounding
but don't have any" ;)
> - Remembered learning early last year while finding the cause of a
> particular speckling bug, that POV-Ray does inside tests when it creates
> original rays.
Just a spontaneous thought here: Would it be worth it to cache the
result of the camera ray origin inside test, and only re-do it if the
origin changes (due to special camera or focal blur)?
> Results were not similar. This leads me to a new suspicion / unproven
> theory we are sitting on a csg inside test performance issue. One
> perhaps affecting things generally. My code changes to the text object
> only treated another problem.
I'm not so sure it's a CSG thing you're seeing. By default, POV-Ray
"flattens" CSG unions wherever possible, promoting all children to
top-level objects - presumably this is also the case with implicit
unions like the text primitive. So unless you took some measure to
specifically prevent this, you're not actually testing a CSG object.
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
On 2/7/19 4:34 PM, clipka wrote:
> Am 07.02.2019 um 17:31 schrieb William F Pokorny:
>
...
>
> Let's play a game of "shapes that could benefit from internal bounding
> but don't have any" ;)
On doing the solver work I'm - a while now - very aware of the internal
bounding issues with the sphere_sweep(1). However, if memory serves,
what Lanuhum had was hundreds of thousands of short sphere_sweeps/hairs.
Why would this be so much slower than one, or several, equally screen
filling ones, if bounding is working well(2). A possible answer is the
inside test on setting up original rays is for some reason doing the
test on all the hairs.
(1) - I've made a changes to the inside test of the sphere_sweep as
currently updated for my solver changes. Specifically a change to break
out of the test loop on the first segment where the point tests inside
true. Today the test runs through all the segments no matter. It's not a
change necessary for the solver changes - just something I noticed.
(2) - His textures looked to be simple.
>
...
>
> Just a spontaneous thought here: Would it be worth it to cache the
> result of the camera ray origin inside test, and only re-do it if the
> origin changes (due to special camera or focal blur)?
>
Maybe, yes. Globally turning the inside testing off when it's not
necessary perhaps another idea, but that would add some bookkeeping to
the parser.
>
...
>
> I'm not so sure it's a CSG thing you're seeing. By default, POV-Ray
> "flattens" CSG unions wherever possible, promoting all children to
> top-level objects - presumably this is also the case with implicit
> unions like the text primitive. So unless you took some measure to
> specifically prevent this, you're not actually testing a CSG object.
Agree. It's a theory until proven.
Will say, a reason I chose the shipped ttf1.pov for the AA and not-AA
testing is there are just two top level text objects in that scene. If
any text objects will hierarchically flatten into the prime/top level
seems like those would.
Do you know offhand what sorts of measures prevent flattening? Merge
pops to mind as suppose we'd need to remember the grouping there. The
two ttf1 text objects are textured and transformed - things almost
always the case with the text /csg - but that's it.
Another test which could be done pretty quickly, I guess, would be to
hack the code to eliminate the inside test on ray initialization for
ttf1.pov testing as it really isn't needed in that scene. I might try
that later today as I'm now curious. But, I'm off to do some other stuff
at the moment.
Bill P.
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
|
|