Apply unsupported blending modes in the shader #6070

unknownbrackets · 2014-05-11T16:59:22Z

This handles unsupported fixed color combinations, alpha doubling, etc.

Where possible, tries to avoid it - it means using a blit (without the framebuffer fetch extension), which can slow things down a lot with tons of drawcalls.

This fixes some effects in games which use doubling blend modes, but I haven't tested all the cases (especially since I don't have all the games.) It also does absdiff blending and mismatched fixed color blending. It will not help GLES2, and requires GLES3+ or OpenGL 3.0+.

I'm worried this might need a setting. Before I made AlphaToColorDoubling() || CanDoubleSrcBlendMode() bail, it made Popolocrois severely slower. However, it does allow more accuracy.

-[Unknown]

unknownbrackets · 2014-05-11T17:04:26Z

Potentially we could also try not blitting as described in #3001 (comment), for certain GPUs.

-[Unknown]

hrydgard · 2014-05-11T17:11:24Z

We could blit on ES2 on Tegra using NV_framebuffer_blit - but of course, on Tegra we don't need it as we can use gl_LastFragColor. Anyway, we could fake blitting the color buffer fairly easily on plain ES2 too by simply rendering a quad with a pass through shader, although a little more expensive. That's for future pull requests though.

I'll merge this once it has received a little bit of testing.

unknownbrackets · 2014-05-11T17:20:30Z

Right, well, we could replace all the unsupported blits with "BlitFramebuffer_", I guess.

-[Unknown]

hrydgard · 2014-05-11T17:35:19Z

The ones that blit (or care about) only the color buffer, yeah. Can't do depth that way.

unknownbrackets · 2014-05-11T17:49:25Z

What about a GL_DEPTH24_STENCIL8 texture? Can't those be bound too?

-[Unknown]

hrydgard · 2014-05-11T18:02:36Z

Not as textures, In unextended ES 2.0, there are no depth textures (although you can write to depth in the shader). There are extensions though so it can work on some chips (maybe quite a lot).

hrydgard · 2014-05-11T18:56:00Z

GPU/GLES/FragmentShaderGenerator.cpp

+		return true;
+
+	case GE_SRCBLEND_FIXA:
+		if (funcB == GE_DSTBLEND_FIXB) {


Here you could also filter out the cases where one of the fix colors is 0xFFFFFF or 0x0, as those are already handled correctly by regular blending by replacing them with GL_ONE/GL_ZERO.

Just did it, yeah.

-[Unknown]

solarmystic · 2014-05-11T21:51:27Z

Do permit me to do the usual and report back later. Hopefully the hit isn't too bad on my system.

unknownbrackets · 2014-05-11T22:00:01Z

Well, if it gets used a lot the hit will be bad, but these blending modes are only used in certain games and usually not many times per frame. Here are some of them:

http://report.ppsspp.org/logs/kind/353
http://report.ppsspp.org/logs/kind/83 (probably mostly when things are crossfading...)

Hmm, I wonder how this impacts God Eater 2.... demo seems not majorly impacted.

-[Unknown]

Otherwise, the frag shader might be generated accounting for different vertexFullAlpha values, or otherwise won't match up with the id, since we change flags between the two Apply calls.

solarmystic · 2014-05-11T23:20:29Z

@unknownbrackets @hrydgard
Performance test results for your perusal. I opted not to publish one for functionality since all games in the test suite remained graphically unaffected as compared to the current master. (Persona 2, which was supposed to be fixed by this pull request had already been resolved by #6072)

General Observations:-

Gods Eater Burst undergoes a minor, but noticeable performance regression.
The rest of the games in the list are unchanged/or close to it in terms of their performance.

Conclusion:-

Seems like the pull request is good to go. I was expecting a severe performance hit in GEB (since it's a part of the lists http://report.ppsspp.org/logs/kind/353 and http://report.ppsspp.org/logs/kind/83), but it's relatively minor.

unknownbrackets · 2014-05-12T00:13:15Z

Well, Gods Eater Burst actually loses 4% for me with this (947% -> 910%), but only in areas and conditions where the blend is triggered. It does a blend of 0xfefefe/0x808080, which this currently triggers for (I could probably allow a variance of 0xfefefe though...

Anyway, I'm able now to reproduce (without forcing it) a box in the top left when blending, but I can't figure out why. The framebuffer and fbotex are the right size. It's not strongly noticeable, but i don't get what's causing it. You can reproduce as follows:

Load the demo.
Get to the "den", which is the first area where you can move. There's a reception desk and people sitting/standing around.
Walk upstairs (this triggers the blending.)
Note a slight line in the top left quadrant of the screen.

This occurs for 1x as well as 2x/3x/etc., so it's not something to do with sizing. The buffer is already selected and estimated by the time it hits ApplyDrawState, so it shouldn't be sizing either... I don't get it.

-[Unknown]

dbz400 · 2014-05-12T00:43:13Z

Persona 2 and Super Robot War Portable A both look good however DBZ Tag got black background in-game

dbz400 · 2014-05-12T00:46:00Z

GPU/GLES/FragmentShaderGenerator.cpp

+				if (funcA != GE_SRCBLEND_DSTCOLOR || funcB != GE_DSTBLEND_SRCCOLOR) {
+					WARN_LOG_REPORT(G3D, "Using MIN blend equation with odd factors: %d, %d", funcA, funcB);
+				}
+				WRITE(p, "  v.rgb = min(v.rgb, destColor.rgb);\n", srcFactor, dstFactor);


I think MIN/MAX/ABSDIFF don't need the srcFactor and dstFactor (though no %s and being used in equation)

unknownbrackets · 2014-05-12T00:56:27Z

Dragon Ball Z uses a CLUT on an FBO, which is not supported without the other changes. That's why it's black.

-[Unknown]

dbz400 · 2014-05-12T00:57:33Z

Arr yes . I almost forget it has to been used with framebuffer-clut branch.

dbz400 · 2014-05-12T01:00:42Z

This is another game which use ABSDIFF blending but not showing propertly .Not too sure if it is related to overlap polygon issue and also happen in clut branch. (btw , how to check ?)

dbz400 · 2014-05-12T01:47:18Z

texelFetch is not supported on ES 2.0 .Wondering would it be good to fall back to use texture2D though need more work to cope with the screen resolution .

unknownbrackets · 2014-05-12T02:01:19Z

What is the actual texture it's using absdiff with? Do we have an issue already for it (I don't know the name so can't search)?

-[Unknown]

daniel229 · 2014-05-12T02:05:40Z

Naruto Shippuden: Narutimate Accel 3 ABSDIFF blending works fine.

master build

test build

dbz400 · 2014-05-12T02:07:31Z

@unknownbrackets , yep , there is already an issue tracking it though it is closed #1982

I'll check the actual texture when i back home .

On our current implemenation for ABSDIFF blending , we are passing GL_MAX to approximate it .

-softGPU

-GLES (GL_MAX approximate)

-this commit

daniel229 · 2014-05-12T04:14:29Z

FF zero type ABSDIFF blending,don't know if right.

test build

master build

softGPU

daniel229 · 2014-05-12T04:40:07Z

God eater 2 has a bug in this area.

test build

master build

daniel229 · 2014-05-12T04:50:44Z

demo can reproduce the problem in this mission,go to this area and walk around.

unknownbrackets · 2014-05-12T04:58:34Z

So, that looks similar (but much worse) to the problem in God Eater 1 and Gods Eater Burst. I don't know what's causing it, but it's rendering at the wrong size somehow. There shouldn't be anything to trigger that here? I tried scaling the params to texelFetch, but that just moves the window up and down.. maybe I did it wrong.

Maybe I should not trust abs()?
http://stackoverflow.com/questions/13335507/glsl-abs-broken

Does it help to change this:
WRITE(p, " v.rgb = abs(v.rgb - destColor.rgb);\n");

To:
WRITE(p, " v.r = max(v.r - destColor.r, destColor.r - v.r);\n");
WRITE(p, " v.g = max(v.g - destColor.g, destColor.g - v.g);\n");
WRITE(p, " v.b = max(v.b - destColor.b, destColor.b - v.b);\n");

Also, are any of these games reporting odd blending factors?

-[Unknown]

This handles unsupported fixed color combinations, alpha doubling, etc. Where possible, tries to avoid it - it means using a blit (without the framebuffer fetch extension), which can slow things down a lot with tons of drawcalls.

Was all lies.

Just in case there are problems with decimating it or etc.

Seems from reports that they are indeed ignored, just wanted to verify.

unknownbrackets · 2014-05-27T08:55:01Z

Aha, I think I found it. Scissor affects glBlitFramebuffer()?

"When values are written to the draw buffers, blit operations bypass most of the fragment pipeline. The only fragment operations which affect a blit are the pixel ownership test, the scissor test, and sRGB conversion"

Well, so it does. Mystery solved, fix upcoming.

-[Unknown]

Because it applies.

hrydgard · 2014-05-27T09:05:01Z

Ah, nice!!

What does the performance impact of this branch look like now?

unknownbrackets · 2014-05-27T09:07:21Z

@daniel229 how is God Eater 2? @raven02 how is Dai2Ji Super Robot Taisen Z Hakaihen now?

@hrydgard I don't know, I'm too tired now to test it, before it was not really that bad but it depends on how many blits it does... it will report "Lots of blits needed for obscure blending: %d per frame, blend %d/%d/%d" on a problem. Could use wider testing, mostly of these:
http://report.ppsspp.org/logs/kind/83
http://report.ppsspp.org/logs/kind/353

I have not even tried Tactics Ogre yet...

-[Unknown]

hrydgard · 2014-05-27T09:09:24Z

Alright, I'll wait for more testing before deciding whether we need a setting, and merging.

dbz400 · 2014-05-27T09:52:38Z

Both look good now .

daniel229 · 2014-05-27T11:36:26Z

@unknownbrackets I think you mean revert this #6067 and test,because that is hiding issue.Revert it,and God Eater 2 is still fine with this branch.

daniel229 · 2014-05-27T12:10:37Z

Macross Triangle Frontier report this,but I don't where is the problem.
W[G3D]: GLES\StateMapping.cpp:184 Lots of blits needed for obscure blending: 26 per frame, blend 6/3/0

Beats,
FF-Type 0,
Higurashi Daybreak Portable,
Naruto Shippuden: Narutimate Accel 3,
Tenchu 4
are still fine,the others I don't know when they would suing this feature.

unknownbrackets · 2014-05-27T14:36:52Z

Well, it would mainly be during effects, but yeah, probably hard to find...

That's a GE_SRCBLEND_DOUBLESRCALPHA / GE_DSTBLEND_INVSRCALPHA / GE_BLENDMODE_MUL_AND_ADD blend. What that means is textures/vertices that don't use full alpha, so we can't just flip it to color doubling.

As long as performance was okay, it might be alright. 26 is high but not super bad. Will be worse on slower devices, though...

Yeah, I had to revert that change to test it.

-[Unknown]

hrydgard · 2014-05-27T15:45:26Z

Let's merge and see how it goes. If there's severe slowdown this merits a setting but it seems for most games it works fine so we'll default to on.

Apply unsupported blending modes in the shader

brujo5 · 2014-05-30T06:08:03Z

#6168

…n github or send a pull request.

hrydgard reviewed May 11, 2014
View reviewed changes

This was referenced May 11, 2014

Dark Overworlds in Persona 2 Innocent Sin #3379

Closed

Optimize double alpha blending for 1/0 textures #6072

Merged

dbz400 reviewed May 12, 2014
View reviewed changes

dbz400 mentioned this pull request May 12, 2014

GLES: Implement in-shader alpha blending modes #5917

Closed

dbz400 mentioned this pull request May 12, 2014

Not handling the 2x dest ones .Fixes #3379 #5844

Closed

unknownbrackets added 9 commits May 27, 2014 01:45

Apply unsupported blending modes in the shader.

4b978ed

This handles unsupported fixed color combinations, alpha doubling, etc. Where possible, tries to avoid it - it means using a blit (without the framebuffer fetch extension), which can slow things down a lot with tons of drawcalls.

Double using multiply not add, should be fine.

a673277

Skip a few more fixa/fixb combos.

41ca303

Log when blitting a ton per frame, it can get slow.

b02fe1d

Remove outdated comment.

f80dee4

Was all lies.

Oops, typo.

b8a2410

Set a bit to account for fixa/fixb/etc. logic.

0eae30d

Unbind the fbo texture when it's not used.

8f9a2b9

Just in case there are problems with decimating it or etc.

Ignore the funcs for min/max/absdiff.

fce7d8c

Seems from reports that they are indeed ignored, just wanted to verify.

unknownbrackets added 2 commits May 27, 2014 02:00

Disable the scissor test around blits.

3e24786

Because it applies.

Make sure to rebind a texture after changing it.

ba2cf4f

hrydgard added a commit that referenced this pull request May 27, 2014

Merge pull request #6070 from unknownbrackets/gpu-blend

8d84c91

Apply unsupported blending modes in the shader

hrydgard merged commit 8d84c91 into hrydgard:master May 27, 2014

unknownbrackets deleted the gpu-blend branch May 28, 2014 06:15

daniel229 mentioned this pull request May 30, 2014

God Eater Burst Issues: 9.8-890 version to latest versions #6185

Closed

unknownbrackets mentioned this pull request Jun 1, 2014

Alpha problem ULJS00379 Dai2Ji Super Robot Taisen Z Hakaihen #1982

Closed

daniel229 mentioned this pull request Jun 4, 2014

Gensou suikuden 1&2 various issues #6226

Closed

daniel229 mentioned this pull request Jun 17, 2014

Support alpha and color test masks on desktop/gles3 #6341

Merged

solarmystic referenced this pull request Jul 22, 2014

Update What's New in the README. If I forgot something big, comment o…

a979480

…n github or send a pull request.

lightfirst mentioned this pull request Oct 8, 2015

Feature Request: Add absdiff blend effect jpcsp/jpcsp#398

Open

Apply unsupported blending modes in the shader #6070

Apply unsupported blending modes in the shader #6070

Conversation

unknownbrackets commented May 11, 2014

unknownbrackets commented May 11, 2014

hrydgard commented May 11, 2014

unknownbrackets commented May 11, 2014

hrydgard commented May 11, 2014

unknownbrackets commented May 11, 2014

hrydgard commented May 11, 2014

hrydgard May 11, 2014

Choose a reason for hiding this comment

unknownbrackets May 11, 2014

Choose a reason for hiding this comment

solarmystic commented May 11, 2014

unknownbrackets commented May 11, 2014

solarmystic commented May 11, 2014

unknownbrackets commented May 12, 2014

dbz400 commented May 12, 2014

dbz400 May 12, 2014

Choose a reason for hiding this comment

unknownbrackets commented May 12, 2014

dbz400 commented May 12, 2014

dbz400 commented May 12, 2014

dbz400 commented May 12, 2014

unknownbrackets commented May 12, 2014

daniel229 commented May 12, 2014

dbz400 commented May 12, 2014

daniel229 commented May 12, 2014

daniel229 commented May 12, 2014

daniel229 commented May 12, 2014

unknownbrackets commented May 12, 2014

unknownbrackets commented May 27, 2014

hrydgard commented May 27, 2014

unknownbrackets commented May 27, 2014

hrydgard commented May 27, 2014

dbz400 commented May 27, 2014

daniel229 commented May 27, 2014

daniel229 commented May 27, 2014

unknownbrackets commented May 27, 2014

hrydgard commented May 27, 2014

brujo5 commented May 30, 2014