Transparency #589

psieg · 2014-04-21T02:01:46Z

We should have done this earlier....
Here comes Order Independent Transparency using Weighted Average
Because we use instancing etc it was not possible to directly use the glow implementation, but most is stolen from there.
With two render passes and one additional render target, you get transparency for any Voxel in a VoxelCluster and any Particle
Please try this ASAP and be blown by the awesomeness 😄

xchrdw · 2014-04-21T07:08:20Z

looks good. I tried to make these yellow things in the gamescenario transparent but it didn't work. What did I do wrong?

psieg · 2014-04-21T09:38:56Z

nothing, seems to work for me. note you edited the gamescenario (the big battle) and not the frozen one.

mrzzzrm · 2014-04-21T09:53:34Z

Is that a freaking transparent cockpit on the mox? wooooha!

psieg · 2014-04-21T10:02:15Z

data/shader/postprocessing/combine.frag

+void main() {	
+	vec4 opaque = texture(color, v_uv) + texture(bloom, v_uv);
+	vec4 accumulated = texture(transparencyAcc, v_uv);
+	accumulated.a = texture(transparencyCnt, v_uv).z; // Why is Acc.a always 1?!


I could use some input with this:
Even though I switched the textures to RGBA the alpha value always returns 1. It's no problem because I have space in the count buffer, but why?

mrzzzrm · 2014-04-21T10:16:27Z

I don't know the limitations of weighted average, but would it be possible to make a whole planet or maybe at least the yellow enemies (-> frozengamescenario) appear transparent? Currently the planet turns almost white and the enemies appear transparent only from some angles.

psieg · 2014-04-21T10:23:08Z

you should be able to...

can you be more specific about the enemies?

psieg · 2014-04-21T10:24:53Z

because all voxels are rendred those on the outer edge appear more transparent from some angles, because there are only 2 transparent layers rather than the >4 everywhere else, is that what you mean?

mrzzzrm · 2014-04-21T10:28:54Z

That's probably it. We should use multiple layers of transparent voxels then, I guess.
Also performance drops extremly here when flying throught the transparent planet. Is that inevitable?

psieg · 2014-04-21T10:42:44Z

The performance impact from the trancparency is constant (well, the number of transparent pixels might have some impact), the perfomance drop must result from splitdetection.
Also, please explain what you mean by multiple layers or transparent voxels

mrzzzrm · 2014-04-21T10:45:51Z

I flew through the planet via freecam, so no splitdetection. Are there ifs only triggered on transparency in the shaders? Sometimes they are evil. :D
I mean, I don't think the transparent Wall (when viewed from the front, just turns white), or let alone the planet, look good anymore. So we should be carefull not to overdo it when using transparency.

psieg · 2014-04-21T10:52:57Z

data/shader/voxelparticle/voxelparticle.frag

+    }
+    if (transparentPass == 0 && f_color.a < 0.9999) {
+        discard;
+    }


there are ifs, I don't see how we can get around these though...
And indeed the game spends 90% in nvogl.dll when in the planet
Weighted average is one of the faster algorithms, so yes, it's not perfect, especially with many layers of transparent pixels. But if you don't start making whole ships or objects transparent, I think it does the job quite well

I don't see why there is a performance difference with transparent voxels though. The scene is rendered twice, any fragment passes either trough the first if in the first run or the second if in the second run. This means any fragment passes the part below the ifs once per frame.
I can't explain it via Z-Buffer either because the z test is applied after the fragment shader has run. Can blending be the problem? With many transparent fragments on one pixel, the GPU has to blend a lot

psieg · 2014-04-21T13:18:34Z

GTX460 video load in the transparency2_perf branch (without moving the camera, shader doesn't discard anything):
single rendering: 54%
render, clear, render: 84% 81%
render, clear, render without writing z-buffer: 84% 88%
render, clear, render without backface culling: 84% 87%
render, clear, render without both: 99% 98%
single render without z-buffer: 80%
single render without backface culling: 56%
single render without both: 99%
=> the if in the shader isn't the problem
=> transparency (a) renders twice and (b) renders the second run without writing to z and backface -> way more fragments
=> don't make too many layers transparent....
The weird thing is the z buffer test should happen after the fragment shader, what is it that produces the load then? the fact that the already-shaded fragment is not written/blended to the buffer?

mrzzzrm · 2014-04-22T13:44:30Z

Need some less euphoria? Crashes on my bp-pc

#0  0x0000000000000000 in ?? ()
#1  0x00007ffff67ca999 in glow::Buffer::clearData (this=0x160de40, internalformat=33334, format=6403, type=5125, 
    data=0x7ffff665f388 <glowutils::WeightedAverageAlgorithm::draw(std::function<void (glow::Program*)> const&, glowutils::Camera*, int, int)::initialDepthComplexity>)
    at /home/moritz/workspace/voxellancer/lib/glow/source/glow/source/Buffer.cpp:292
#2  0x00007ffff63e7c55 in glowutils::WeightedAverageAlgorithm::draw(std::function<void (glow::Program*)> const&, glowutils::Camera*, int, int) (this=0x16082a0, 
    drawFunction=..., camera=0x1607e30, width=1280, height=800) at /home/moritz/workspace/voxellancer/lib/glow/source/glowutils/source/WeightedAverageAlgorithm.cpp:57
#3  0x0000000000777037 in GamePlayScene::drawGame (this=0x1214590, camera=..., destinationViewport=...)
    at /home/moritz/workspace/voxellancer/src/gamestate/gameplay/gameplayscene.cpp:122
#4  0x0000000000776907 in GamePlayScene::draw (this=0x1214590, camera=..., target=0x7ffff6b408c0 <glow::FrameBufferObject::s_defaultFBO>, destinationViewport=..., side=None)
    at /home/moritz/workspace/voxellancer/src/gamestate/gameplay/gameplayscene.cpp:62
#5  0x0000000000762af8 in MonoView::draw (this=0x1217d80, scene=..., cameraHead=...) at /home/moritz/workspace/voxellancer/src/display/monoview.cpp:58
#6  0x0000000000762fcd in Viewer::draw (this=0x1219110, scene=..., cameraHead=...) at /home/moritz/workspace/voxellancer/src/display/viewer.cpp:48
#7  0x000000000076a98f in Game::draw (this=0x11b0b40) at /home/moritz/workspace/voxellancer/src/gamestate/game.cpp:56
#8  0x00000000007ad4f5 in mainloop () at /home/moritz/workspace/voxellancer/src/main.cpp:155
#9  0x00000000007adefa in main (argc=1, argv=0x7fffffffdf48) at /home/moritz/workspace/voxellancer/src/main.cpp:251

mrzzzrm · 2014-04-22T13:45:14Z

    at /home/moritz/workspace/voxellancer/lib/glow/source/glow/source/Buffer.cpp:292
292         glClearNamedBufferDataEXT(m_id, internalformat, format, type, data);

Probably because glClearNamedBufferDataEXT isnt set by glewInit....

mrzzzrm · 2014-05-08T20:36:27Z

src/gamestate/gameplay/gameplayscene.cpp

+    glBlendFunc(GL_ONE, GL_ONE);
+    CheckGLError();
+    drawGame(camera, true);
+    //drawGameAlpha(camera);


remove may 🐝 ?

mrzzzrm · 2014-05-08T21:01:51Z

Not quite done yet, I need to think about splitting the particleengine a bit, but so far it looks pretty good.
Could you place some explaining words into the combine shader how the weighted-average-algorithm works?

psieg · 2014-05-10T13:00:58Z

data/shader/postprocessing/combine.frag

+    vec4 accumulated = texture(transparencyAcc, v_uv);
+    // HACK: actually accumulated.a should be the accumulation of the a values
+    // It seems to always be 1, while using the count-buffer's z for the same purpose works
+    accumulated.a = texture(transparencyCnt, v_uv).z;


open to ideas... I didn't try cnt.a, don't the textures have alpha? The framebuffer creates them with
texture->image2D(0, GL_RGBA, m_resolution.x, m_resolution.y, 0, GL_RGBA, GL_FLOAT, nullptr);

Didn't look but you have enabled GL_BLEND? Otherwise I wouldn't be surprised if OpenGL optimizes values out that aren't explicitely set (as is the case for the alpha value of the color)

…rency2

… again

mrzzzrm · 2014-05-18T16:37:39Z

data/shader/voxelparticle/voxelparticle.frag

+    vec3 rgbColor = voxelFragmentColor(f_color.rgb, f_emissiveness, f_normal, f_modelposition);
+    fragColor = vec4(rgbColor * f_color.a, f_color.a);
+    emissiveness = voxelFragmentEmissiveness(f_color.xyz, f_emissiveness);
+    if(transparentPass) {


how about
normalz = (1-transparentPass) * voxelFragmenNormalZ(f_normal);?

ifs are evil, I have been told, as we need voxelFragmenNormalZ() in most cases (=most of the pixels won't be transparent), this should be better

so long as this operation is properly defined with bool, I'm on. Will try

you will probably need to cast to float explicitely

mrzzzrm · 2014-05-18T16:44:12Z

src/display/rendering/buffernames.h

@@ -23,6 +24,7 @@ namespace {
        "NormalZ",
        "Emissisiveness",
        "BlurTmp",
-        "Bloom",


Could be considered premature optimization... and its a bit confusing as well. And what is the pupose to name a buffer "Emissiveness" if it is used for something else as well?

Well, the use still is only Emissiveness == Bloom. Since they are the same buffer now it could be named bloom. Otherwise it's not an optimization but was needed to get two more buffers (limit is 8 on most GPUs)

mrzzzrm · 2014-05-18T16:49:01Z

When I run the PirateScenario and fly through the enginetrail of the normandy, I suffer slowdowns. Can you reproduce this?

psieg · 2014-05-18T17:58:15Z

my load goes from 33% without significant transparent voxels up to 80% when the entire screen is filled with transparent voxels.
The reason is the same as with the planet, many non-culling non-z-testing fragment shader passes I think. Maybe we can ask daniel about it but I wouldn't know how to avoid this

mrzzzrm · 2014-05-18T18:04:45Z

Well, it's definitely to much of a drop for that few particles there, that definitely needs to be fixed somehow. Or we have to forbid transparent particles

psieg · 2014-05-18T18:08:31Z

note the performance critical part is screen coverage (pixels), not particle count. with a bad perspective, you probably can mess up any scenario by overlapping n transparent surfaces and filling the entire screen with it. I agree though, we'll have to find some solution. Maybe a performance setting (so HD4000's don't have it but GTX780 players may enjoy it)

mrzzzrm · 2014-05-18T18:23:29Z

I have problems to believe this. Even my HD5830 has a pixelfillrate of 200Mio/Frame. That's each pixel 100 times in HD. And I am not even running the game in HD! http://techreport.com/review/18521/amd-radeon-hd-5830-graphics-card/5

psieg · 2014-05-18T18:52:09Z

I know too few about GPUs to argue.
Try disabling z testing in the current master and looking at the planet:
glDepthMask(false); in gameplayscene:56
with depth mask I'm at 49%, without >90%, so the issue isn't directly linked to the transparency (as in blending, ifs, etc)

psieg and others added 10 commits April 20, 2014 19:45

Add Transparency buffer

6a6f044

Add TransparencyCount buffer, repurpose Emissiveness for Bloom

42c31dc

code checkpoint

f1955b4

Switch color format to include alpha

dada5fd

fix shader

1b89943

Fix shader even more, the rest is algorithm error

977f4e5

Should have done that earlier

d5e8ca6

Transparency! In every VoxelCluster

df7330e

And because we like it, here is particle transparency

ed4536f

try to make something transparent

1ce8cfc

Make even more things transparent

ea77924

psieg reviewed Apr 21, 2014
View reviewed changes

psieg added 4 commits April 21, 2014 12:04

shader formatting

5c2043f

Add missing cockpitvoxels

3f0f122

more shader formatting

56e4239

buffer clear order

f9d38ae

make things transparent for mrzzzrm

b860371

psieg reviewed Apr 21, 2014
View reviewed changes

mrzzzrm reviewed May 8, 2014
View reviewed changes

psieg added 5 commits May 9, 2014 08:42

Lots of curious bugs

3671959

Some shader documentation

f62c540

Add bloom last

510b04a

remove drawGameAlpha development leftovers

cc48cd4

switch gameplayscene to glow::State

f4e319e

psieg reviewed May 10, 2014
View reviewed changes

psieg and others added 3 commits May 10, 2014 15:23

fix crucial voxel color not rendered

46a677c

Merge branch 'master' of github.com:voxelinc/voxellancer into transpa…

2b3b506

…rency2

made objects that aren't rendered beautifully when transparent opaque…

b134b9f

… again

mrzzzrm reviewed May 18, 2014
View reviewed changes

some formating

4a1ab19

mrzzzrm reviewed May 18, 2014
View reviewed changes

Avoid if in fragment shader

7bacd1d

mrzzzrm mentioned this pull request May 22, 2014

Transparency merge #618

Open

Merge branch 'master' into transparency2

97b6d57

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Transparency #589

Transparency #589

psieg commented Apr 21, 2014

xchrdw commented Apr 21, 2014

psieg commented Apr 21, 2014

mrzzzrm commented Apr 21, 2014

psieg Apr 21, 2014

mrzzzrm commented Apr 21, 2014

psieg commented Apr 21, 2014

psieg commented Apr 21, 2014

mrzzzrm commented Apr 21, 2014

psieg commented Apr 21, 2014

mrzzzrm commented Apr 21, 2014

psieg Apr 21, 2014

psieg Apr 21, 2014

psieg commented Apr 21, 2014

mrzzzrm commented Apr 22, 2014

mrzzzrm commented Apr 22, 2014

mrzzzrm May 8, 2014

mrzzzrm commented May 8, 2014

psieg May 10, 2014

mrzzzrm May 18, 2014

mrzzzrm May 18, 2014

mrzzzrm May 18, 2014

psieg May 18, 2014

mrzzzrm May 18, 2014

mrzzzrm May 18, 2014

psieg May 18, 2014

mrzzzrm commented May 18, 2014

psieg commented May 18, 2014

mrzzzrm commented May 18, 2014

psieg commented May 18, 2014

mrzzzrm commented May 18, 2014

psieg commented May 18, 2014

Transparency #589

Are you sure you want to change the base?

Transparency #589

Conversation

psieg commented Apr 21, 2014

xchrdw commented Apr 21, 2014

psieg commented Apr 21, 2014

mrzzzrm commented Apr 21, 2014

Choose a reason for hiding this comment

mrzzzrm commented Apr 21, 2014

psieg commented Apr 21, 2014

psieg commented Apr 21, 2014

mrzzzrm commented Apr 21, 2014

psieg commented Apr 21, 2014

mrzzzrm commented Apr 21, 2014

Choose a reason for hiding this comment

Choose a reason for hiding this comment

psieg commented Apr 21, 2014

mrzzzrm commented Apr 22, 2014

mrzzzrm commented Apr 22, 2014

Choose a reason for hiding this comment

mrzzzrm commented May 8, 2014

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mrzzzrm commented May 18, 2014

psieg commented May 18, 2014

mrzzzrm commented May 18, 2014

psieg commented May 18, 2014

mrzzzrm commented May 18, 2014

psieg commented May 18, 2014