Implement difference clip rects with window rectangles

Implement difference clip rects with window rectangles

Plumbs the pipeline for window rectangles and uses them for a very
basic implementation of difference clip rects. This puts a common
Blink pattern on fast path, but we will still eventually need to make
more comprehensive use of window rectangles during clipping.

GTX 960 perf result:
                              gpu                     glinst4                 glinst16
desk_jsfiddlebigcar.skp       0.254 -> 0.177 [70%]    0.279 -> 0.197 [71%]    0.577 -> 0.196 [34%]
keymobi_sfgate_com_.skp       0.697 -> 0.513 [74%]    0.766 -> 0.451 [59%]    0.769 -> 0.597 [78%]
keymobi_blogger.skp           0.406 -> 0.314 [77%]    0.436 -> 0.292 [67%]    0.696 -> 0.319 [46%]
desk_pokemonwiki.skp          0.121 -> 0.098 [81%]     0.13 -> 0.105 [81%]    0.216 -> 0.097 [45%]
desk_wikipedia.skp            0.121 -> 0.098 [81%]     0.13 -> 0.104 [80%]    0.199 -> 0.104 [52%]
keymobi_androidpolice_co...   0.443 -> 0.382 [86%]    0.447 -> 0.398 [89%]    0.444 -> 0.396 [89%]
keymobi_booking_com_sear...   1 .15 ->  1.03 [90%]     1.17 ->  1.06 [91%]     1.17 ->  1.05 [90%]
keymobi_theverge_com.skp      0.417 -> 0.396 [95%]    0.426 -> 0.405 [95%]    0.429 ->   0.4 [93%]

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2251573002

Committed: https://skia.googlesource.com/skia/+/28341fad8438b294ed7311edbc68d8cb34a990ab

[csmartdalton, 2016-08-15 23:46:46]

Description was changed from

==========
Implement difference clip rects with window rectangles

Plumbs the pipeline for window rectangles and uses them for a very
basic implementation of difference clip rects. This puts a common
Blink pattern on fast path, but we will still eventually need to make
more comprehensive use of window rectangles during clipping.

GTX 960 perf result:
                              gpu                     glinst4                
glinst16
desk_jsfiddlebigcar.skp       0.254 -> 0.177 [70%]    0.279 -> 0.197 [71%]   
0.577 -> 0.196 [34%]
desk_pokemonwiki.skp          0.121 -> 0.098 [81%]     0.13 -> 0.105 [81%]   
0.216 -> 0.097 [45%]
desk_wikipedia.skp            0.121 -> 0.098 [81%]     0.13 -> 0.104 [80%]   
0.199 -> 0.104 [52%]
keymobi_blogger.skp           0.406 -> 0.314 [77%]    0.436 -> 0.292 [67%]   
0.696 -> 0.319 [46%]
keymobi_androidpolice_co...   0.443 -> 0.382 [86%]    0.447 -> 0.398 [89%]   
0.444 -> 0.396 [89%]
keymobi_sfgate_com_.skp       0.697 -> 0.513 [74%]    0.766 -> 0.451 [59%]   
0.769 -> 0.597 [78%]
keymobi_theverge_com.skp      0.417 -> 0.396 [95%]    0.426 -> 0.405 [95%]   
0.429 ->   0.4 [93%]
keymobi_booking_com_sear...   1 .15 ->  1.03 [90%]     1.17 ->  1.06 [91%]    
1.17 ->  1.05 [90%]

BUG=skia:
==========

to

==========
Implement difference clip rects with window rectangles

Plumbs the pipeline for window rectangles and uses them for a very
basic implementation of difference clip rects. This puts a common
Blink pattern on fast path, but we will still eventually need to make
more comprehensive use of window rectangles during clipping.

GTX 960 perf result:
                              gpu                     glinst4                
glinst16
desk_jsfiddlebigcar.skp       0.254 -> 0.177 [70%]    0.279 -> 0.197 [71%]   
0.577 -> 0.196 [34%]
desk_pokemonwiki.skp          0.121 -> 0.098 [81%]     0.13 -> 0.105 [81%]   
0.216 -> 0.097 [45%]
desk_wikipedia.skp            0.121 -> 0.098 [81%]     0.13 -> 0.104 [80%]   
0.199 -> 0.104 [52%]
keymobi_blogger.skp           0.406 -> 0.314 [77%]    0.436 -> 0.292 [67%]   
0.696 -> 0.319 [46%]
keymobi_androidpolice_co...   0.443 -> 0.382 [86%]    0.447 -> 0.398 [89%]   
0.444 -> 0.396 [89%]
keymobi_sfgate_com_.skp       0.697 -> 0.513 [74%]    0.766 -> 0.451 [59%]   
0.769 -> 0.597 [78%]
keymobi_theverge_com.skp      0.417 -> 0.396 [95%]    0.426 -> 0.405 [95%]   
0.429 ->   0.4 [93%]
keymobi_booking_com_sear...   1 .15 ->  1.03 [90%]     1.17 ->  1.06 [91%]    
1.17 ->  1.05 [90%]

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2251573002
==========

[csmartdalton, 2016-08-15 23:50:09]

csmartdalton@google.com changed reviewers:
+ bsalomon@google.com

[csmartdalton, 2016-08-15 23:50:10]

We don't have any bots that can test this functionality. Will it be released
soon in a public driver? Or is there any reason we can't just put the Siggraph
driver on the bots?

[csmartdalton, 2016-08-16 02:19:34]

Description was changed from

==========
Implement difference clip rects with window rectangles

Plumbs the pipeline for window rectangles and uses them for a very
basic implementation of difference clip rects. This puts a common
Blink pattern on fast path, but we will still eventually need to make
more comprehensive use of window rectangles during clipping.

GTX 960 perf result:
                              gpu                     glinst4                
glinst16
desk_jsfiddlebigcar.skp       0.254 -> 0.177 [70%]    0.279 -> 0.197 [71%]   
0.577 -> 0.196 [34%]
desk_pokemonwiki.skp          0.121 -> 0.098 [81%]     0.13 -> 0.105 [81%]   
0.216 -> 0.097 [45%]
desk_wikipedia.skp            0.121 -> 0.098 [81%]     0.13 -> 0.104 [80%]   
0.199 -> 0.104 [52%]
keymobi_blogger.skp           0.406 -> 0.314 [77%]    0.436 -> 0.292 [67%]   
0.696 -> 0.319 [46%]
keymobi_androidpolice_co...   0.443 -> 0.382 [86%]    0.447 -> 0.398 [89%]   
0.444 -> 0.396 [89%]
keymobi_sfgate_com_.skp       0.697 -> 0.513 [74%]    0.766 -> 0.451 [59%]   
0.769 -> 0.597 [78%]
keymobi_theverge_com.skp      0.417 -> 0.396 [95%]    0.426 -> 0.405 [95%]   
0.429 ->   0.4 [93%]
keymobi_booking_com_sear...   1 .15 ->  1.03 [90%]     1.17 ->  1.06 [91%]    
1.17 ->  1.05 [90%]

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2251573002
==========

to

==========
Implement difference clip rects with window rectangles

Plumbs the pipeline for window rectangles and uses them for a very
basic implementation of difference clip rects. This puts a common
Blink pattern on fast path, but we will still eventually need to make
more comprehensive use of window rectangles during clipping.

GTX 960 perf result:
                              gpu                     glinst4                
glinst16
desk_jsfiddlebigcar.skp       0.254 -> 0.177 [70%]    0.279 -> 0.197 [71%]   
0.577 -> 0.196 [34%]
keymobi_sfgate_com_.skp       0.697 -> 0.513 [74%]    0.766 -> 0.451 [59%]   
0.769 -> 0.597 [78%]
keymobi_blogger.skp           0.406 -> 0.314 [77%]    0.436 -> 0.292 [67%]   
0.696 -> 0.319 [46%]
desk_pokemonwiki.skp          0.121 -> 0.098 [81%]     0.13 -> 0.105 [81%]   
0.216 -> 0.097 [45%]
desk_wikipedia.skp            0.121 -> 0.098 [81%]     0.13 -> 0.104 [80%]   
0.199 -> 0.104 [52%]
keymobi_androidpolice_co...   0.443 -> 0.382 [86%]    0.447 -> 0.398 [89%]   
0.444 -> 0.396 [89%]
keymobi_booking_com_sear...   1 .15 ->  1.03 [90%]     1.17 ->  1.06 [91%]    
1.17 ->  1.05 [90%]
keymobi_theverge_com.skp      0.417 -> 0.396 [95%]    0.426 -> 0.405 [95%]   
0.429 ->   0.4 [93%]

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2251573002
==========

[csmartdalton, 2016-08-16 05:20:18]

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles.h
File src/gpu/GrWindowRectangles.h (right):

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles....
src/gpu/GrWindowRectangles.h:33: SkIRect& addWindow(const GrCaps& caps);
I'm thinking I'll replace this GrCaps param with just an int that says your max
window rectangles.

[dogben, 2016-08-16 12:34:51]

On 2016/08/15 at 23:50:10, csmartdalton wrote:
> We don't have any bots that can test this functionality. Will it be released
soon in a public driver? Or is there any reason we can't just put the Siggraph
driver on the bots?

We can put a different driver on the bot. As of now, we have a single Windows
GTX960 bot. (I assume the reason you're testing with GTX960 is that the driver
works only for Maxwell or later.) We will have some additional Linux bots
hopefully in a couple weeks. If we change the driver, we will have no way to
test the standard driver.

[csmartdalton, 2016-08-16 13:34:16]

On 2016/08/16 12:34:51, Ben Wagner wrote:
> (I assume the reason you're testing with GTX960 is that the driver
> works only for Maxwell or later.)

Oh. Right. This feature actually ought to work on just about any NVIDIA GPU. I
was just requiring Maxwell out of habit. Does that make it easier?

[csmartdalton, 2016-08-16 18:38:47]

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrClipStackClip.cpp
File src/gpu/GrClipStackClip.cpp (right):

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrClipStackClip.cpp...
src/gpu/GrClipStackClip.cpp:294: element->getRect().round(&window);
Needs to be offset to screen space from clip space.

[bsalomon, 2016-08-16 19:03:55]

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles.h
File src/gpu/GrWindowRectangles.h (right):

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles....
src/gpu/GrWindowRectangles.h:39: constexpr static int kNumLocalWindows = 1;
What in practice is the limit for the number of window rects?

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles....
src/gpu/GrWindowRectangles.h:52: class GrWindowRectangles::Rec {
Inherit from GrNonAtomicRef?

[csmartdalton, 2016-08-16 19:49:46]

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles.h
File src/gpu/GrWindowRectangles.h (right):

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles....
src/gpu/GrWindowRectangles.h:39: constexpr static int kNumLocalWindows = 1;
On 2016/08/16 19:03:55, bsalomon wrote:
> What in practice is the limit for the number of window rects?

I believe it's 8 on NVIDIA. GrGLCaps limits us to 32 for now so we don't
accidentally track huge amounts.

But the way the pipeline works, this object gets copied several times. So if you
have, say, 8 windows, I think it's better to just take new refs on the Rec than
it would be to memcpy 8 rectangles multiple times.

2 local windows may be better than one...

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles....
src/gpu/GrWindowRectangles.h:52: class GrWindowRectangles::Rec {
On 2016/08/16 19:03:55, bsalomon wrote:
> Inherit from GrNonAtomicRef?

GrWindowRectangles does need to peek at the ref count. Or at least, check if it
only has one ref. Would it be ok to add either refCnt() or hasSingleRef() to
GrNonAtomicRef?

[bsalomon, 2016-08-16 20:05:52]

On 2016/08/16 19:49:46, csmartdalton wrote:
> https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles.h
> File src/gpu/GrWindowRectangles.h (right):
> 
>
https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles....
> src/gpu/GrWindowRectangles.h:39: constexpr static int kNumLocalWindows = 1;
> On 2016/08/16 19:03:55, bsalomon wrote:
> > What in practice is the limit for the number of window rects?
> 
> I believe it's 8 on NVIDIA. GrGLCaps limits us to 32 for now so we don't
> accidentally track huge amounts.
> 
> But the way the pipeline works, this object gets copied several times. So if
you
> have, say, 8 windows, I think it's better to just take new refs on the Rec
than
> it would be to memcpy 8 rectangles multiple times.
> 
> 2 local windows may be better than one...
> 
>
https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles....
> src/gpu/GrWindowRectangles.h:52: class GrWindowRectangles::Rec {
> On 2016/08/16 19:03:55, bsalomon wrote:
> > Inherit from GrNonAtomicRef?
> 
> GrWindowRectangles does need to peek at the ref count. Or at least, check if
it
> only has one ref. Would it be ok to add either refCnt() or hasSingleRef() to
> GrNonAtomicRef?

Sure, our other ref counting classes have "bool unique() const".

[csmartdalton, 2016-08-16 22:00:12]

Patchset #2 (id:20001) has been deleted

[csmartdalton, 2016-08-16 22:19:45]

Patchset #3 (id:60001) has been deleted

[csmartdalton, 2016-08-16 23:38:34]

Much happier with it now.

I'll have to redo the benchmark now that the window rect is actually aligned
correctly...

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrClipStackClip.cpp
File src/gpu/GrClipStackClip.cpp (right):

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrClipStackClip.cpp...
src/gpu/GrClipStackClip.cpp:294: element->getRect().round(&window);
On 2016/08/16 18:38:47, csmartdalton wrote:
> Needs to be offset to screen space from clip space.

Done.

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles.h
File src/gpu/GrWindowRectangles.h (right):

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles....
src/gpu/GrWindowRectangles.h:33: SkIRect& addWindow(const GrCaps& caps);
On 2016/08/16 05:20:17, csmartdalton wrote:
> I'm thinking I'll replace this GrCaps param with just an int that says your
max
> window rectangles.

Done. Just used a compile-time constant max as discussed offline.

https://codereview.chromium.org/2251573002/diff/1/src/gpu/GrWindowRectangles....
src/gpu/GrWindowRectangles.h:52: class GrWindowRectangles::Rec {
On 2016/08/16 19:49:45, csmartdalton wrote:
> On 2016/08/16 19:03:55, bsalomon wrote:
> > Inherit from GrNonAtomicRef?
> 
> GrWindowRectangles does need to peek at the ref count. Or at least, check if
it
> only has one ref. Would it be ok to add either refCnt() or hasSingleRef() to
> GrNonAtomicRef?

Done.

[Mark Kilgard, 2016-08-17 04:24:24]

On 2016/08/16 13:34:16, csmartdalton wrote:
> On 2016/08/16 12:34:51, Ben Wagner wrote:
> > (I assume the reason you're testing with GTX960 is that the driver
> > works only for Maxwell or later.)
> 
> Oh. Right. This feature actually ought to work on just about any NVIDIA GPU. I
> was just requiring Maxwell out of habit. Does that make it easier?

Yes, EXT_window_rectangles support goes back to Fermi (GeForce 4xx series) and
all subsequent GPU generations.  So not just Maxwell and up; Fermi and Kepler
have the support.

Public fully QA'ed drivers are now available:

 
http://us.download.nvidia.com/Windows/372.54/372.54-desktop-win10-64bit-inter...

I verified the 372.54 has EXT_window_rectangles support.

This (technically beta) driver should have the support though I didn't try it:

 
http://us.download.nvidia.com/XFree86/Linux-x86_64/370.23/NVIDIA-Linux-x86_64...

[bsalomon, 2016-08-17 13:28:18]

lgtm

[csmartdalton, 2016-08-17 14:45:55]

Just a bit of  bikeshedding

[bsalomon, 2016-08-17 14:49:33]

On 2016/08/17 14:45:55, csmartdalton wrote:
> Just a bit of  bikeshedding

still lgtm

[csmartdalton, 2016-08-17 16:41:33]

The CQ bit was checked by csmartdalton@google.com

[csmartdalton, 2016-08-17 16:41:33]

The patchset sent to the CQ was uploaded after l-g-t-m from bsalomon@google.com
Link to the patchset: https://codereview.chromium.org/2251573002/#ps200001
(title: "return type")

[commit-bot: I haz the power, 2016-08-17 16:41:44]

CQ is trying da patch. Follow status at
 
https://chromium-cq-status.appspot.com/v2/patch-status/codereview.chromium.or...

[commit-bot: I haz the power, 2016-08-17 17:00:27]

Description was changed from

==========
Implement difference clip rects with window rectangles

Plumbs the pipeline for window rectangles and uses them for a very
basic implementation of difference clip rects. This puts a common
Blink pattern on fast path, but we will still eventually need to make
more comprehensive use of window rectangles during clipping.

GTX 960 perf result:
                              gpu                     glinst4                
glinst16
desk_jsfiddlebigcar.skp       0.254 -> 0.177 [70%]    0.279 -> 0.197 [71%]   
0.577 -> 0.196 [34%]
keymobi_sfgate_com_.skp       0.697 -> 0.513 [74%]    0.766 -> 0.451 [59%]   
0.769 -> 0.597 [78%]
keymobi_blogger.skp           0.406 -> 0.314 [77%]    0.436 -> 0.292 [67%]   
0.696 -> 0.319 [46%]
desk_pokemonwiki.skp          0.121 -> 0.098 [81%]     0.13 -> 0.105 [81%]   
0.216 -> 0.097 [45%]
desk_wikipedia.skp            0.121 -> 0.098 [81%]     0.13 -> 0.104 [80%]   
0.199 -> 0.104 [52%]
keymobi_androidpolice_co...   0.443 -> 0.382 [86%]    0.447 -> 0.398 [89%]   
0.444 -> 0.396 [89%]
keymobi_booking_com_sear...   1 .15 ->  1.03 [90%]     1.17 ->  1.06 [91%]    
1.17 ->  1.05 [90%]
keymobi_theverge_com.skp      0.417 -> 0.396 [95%]    0.426 -> 0.405 [95%]   
0.429 ->   0.4 [93%]

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2251573002
==========

to

==========
Implement difference clip rects with window rectangles

Plumbs the pipeline for window rectangles and uses them for a very
basic implementation of difference clip rects. This puts a common
Blink pattern on fast path, but we will still eventually need to make
more comprehensive use of window rectangles during clipping.

GTX 960 perf result:
                              gpu                     glinst4                
glinst16
desk_jsfiddlebigcar.skp       0.254 -> 0.177 [70%]    0.279 -> 0.197 [71%]   
0.577 -> 0.196 [34%]
keymobi_sfgate_com_.skp       0.697 -> 0.513 [74%]    0.766 -> 0.451 [59%]   
0.769 -> 0.597 [78%]
keymobi_blogger.skp           0.406 -> 0.314 [77%]    0.436 -> 0.292 [67%]   
0.696 -> 0.319 [46%]
desk_pokemonwiki.skp          0.121 -> 0.098 [81%]     0.13 -> 0.105 [81%]   
0.216 -> 0.097 [45%]
desk_wikipedia.skp            0.121 -> 0.098 [81%]     0.13 -> 0.104 [80%]   
0.199 -> 0.104 [52%]
keymobi_androidpolice_co...   0.443 -> 0.382 [86%]    0.447 -> 0.398 [89%]   
0.444 -> 0.396 [89%]
keymobi_booking_com_sear...   1 .15 ->  1.03 [90%]     1.17 ->  1.06 [91%]    
1.17 ->  1.05 [90%]
keymobi_theverge_com.skp      0.417 -> 0.396 [95%]    0.426 -> 0.405 [95%]   
0.429 ->   0.4 [93%]

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2251573002

Committed:
https://skia.googlesource.com/skia/+/28341fad8438b294ed7311edbc68d8cb34a990ab
==========

[commit-bot: I haz the power, 2016-08-17 17:00:30]

Committed patchset #9 (id:200001) as
https://skia.googlesource.com/skia/+/28341fad8438b294ed7311edbc68d8cb34a990ab

[Mark Kilgard, 2016-08-17 17:59:16]

markkilgard@gmail.com changed reviewers:
+ markkilgard@gmail.com

[Mark Kilgard, 2016-08-17 17:59:17]

LGTM

mild concern about capping fMaxWindowRectangles to avoid future performance
pitfalls if large number of window rectangles are not "free"

https://codereview.chromium.org/2251573002/diff/160001/src/gpu/gl/GrGLCaps.cpp
File src/gpu/gl/GrGLCaps.cpp (right):

https://codereview.chromium.org/2251573002/diff/160001/src/gpu/gl/GrGLCaps.cp...
src/gpu/gl/GrGLCaps.cpp:465: GR_GL_GetIntegerv(gli, GR_GL_MAX_WINDOW_RECTANGLES,
&fMaxWindowRectangles);
I actually think protecting yourself from a large number of window rectangles is
a good idea.  We've had code get burned when something like the number of
texture units goes up a bunch (Skia is an example of this...).

On NVIDIA GPUs, 8 window rectangles are "free" (meaning as fast as not having
them when they do no discarding), but if an implementation supported many more,
it's likely to have some kind of cost to cycle more than a fixed number.

OpenVG has a minimum limit of like 32 or something.  It's unlikely that many
window rectangles can all be "free".

I recommend you cap at 8.  If an implementation has a higher limit, you'd be
well-served to evaluate the performance of that higher limit rather than just
accept it.  Chrome updates fast enough that moving to a higher limit on future
hardware will be fast.

[csmartdalton, 2016-08-17 18:23:55]

On 2016/08/17 17:59:17, Mark Kilgard wrote:
> LGTM
> 
> mild concern about capping fMaxWindowRectangles to avoid future performance
> pitfalls if large number of window rectangles are not "free"
> 
> https://codereview.chromium.org/2251573002/diff/160001/src/gpu/gl/GrGLCaps.cpp
> File src/gpu/gl/GrGLCaps.cpp (right):
> 
>
https://codereview.chromium.org/2251573002/diff/160001/src/gpu/gl/GrGLCaps.cp...
> src/gpu/gl/GrGLCaps.cpp:465: GR_GL_GetIntegerv(gli,
GR_GL_MAX_WINDOW_RECTANGLES,
> &fMaxWindowRectangles);
> I actually think protecting yourself from a large number of window rectangles
is
> a good idea.  We've had code get burned when something like the number of
> texture units goes up a bunch (Skia is an example of this...).
> 
> On NVIDIA GPUs, 8 window rectangles are "free" (meaning as fast as not having
> them when they do no discarding), but if an implementation supported many
more,
> it's likely to have some kind of cost to cycle more than a fixed number.
> 
> OpenVG has a minimum limit of like 32 or something.  It's unlikely that many
> window rectangles can all be "free".
> 
> I recommend you cap at 8.  If an implementation has a higher limit, you'd be
> well-served to evaluate the performance of that higher limit rather than just
> accept it.  Chrome updates fast enough that moving to a higher limit on future
> hardware will be fast.

Mark, see the comment in GrCaps.cpp.

[csmartdalton, 2016-08-17 18:24:06]

https://codereview.chromium.org/2251573002/diff/200001/src/gpu/GrCaps.cpp
File src/gpu/GrCaps.cpp (right):

https://codereview.chromium.org/2251573002/diff/200001/src/gpu/GrCaps.cpp#new...
src/gpu/GrCaps.cpp:144: fMaxWindowRectangles = GrWindowRectangles::kMaxWindows;
We still cap. The cap just got moved here instead. (To the platform-independent
part. i.e. this will cap vulkan too)

kMaxWindows is 16. But I agree your suggestion of 8 makes more sense. Lets drop
to 8 and always increase later if needed.