6 replies to this topic

[aeroz]

I'm learning how to use GLSL shaders at the moment. I'm implementing a Gaussian blur shader which uses 2 passes, one for horizontal and one for vertical blur.
My question is, do I need to create a texture target for each pass? This way I would render the scene to the first texture, then I would render the first texture to the second texture using the horizontal blur shader, and finally render the second texture to the screen using the vertical shader. I'm not sure if this is the right way.
The second possibility I have is to render the scene to the same texture two times once with each shader (and add a * 0.5 to the end of each shader). But with this solution, all the scene geometry has to be drawn twice (and not just 2D textures).

Is the double texture approach better? Is there another one?
Thank you!

Edited by aeroz, 02 December 2012 - 01:02 AM.

[Yours3!f]

I'm learning how to use GLSL shaders at the moment. I'm implementing a Gaussian blur shader which uses 2 passes, one for horizontal and one for vertical blur.
My question is, do I need to create a texture target for each pass? This way I would render the scene to the first texture, then I would render the first texture to the second texture using the horizontal blur shader, and finally render the second texture to the screen using the vertical shader. I'm not sure if this is the right way.
The second possibility I have is to render the scene to the same texture two times once with each shader (and add a * 0.5 to the end of each shader). But with this solution, all the scene geometry has to be drawn twice (and not just 2D textures).

Is the double texture approach better? Is there another one?
Thank you!

This is how you do it:
-render the scene to a fbo #1 (ie to texture #1 attached to it).
-blur texture #1 using gaussian blur horizontally (or vertically) outputting to fbo #2 (ie to texture #2 attached to it)
-blur texture #2 outputting to fbo #3 (ie texture #1 that is attached to it)

You can see that I used 3 fbos and only 2 textures. This is because an fbo is just a concept, it only determines WHAT TO the gpu will output the pixels. The storage (which has megabytes of magnitude) is the texture attached to it.
But this "ping pong" rule applies to many other effects like depth of field etc. Where you need the same type of output texture, you can simply cleverly reuse the existing ones, thus saving gpu memory (which you can use later for storing additional assets).

Edited by Yours3!f, 02 December 2012 - 02:38 AM.

[SaTANO]

I am not sure if I understand correctly your rendering cycle but if you have scene already rendered in texture and you want to use separable blur like gaussian you dont need one texture for each pass.
when you are rendering you only need destination target (texture or renderbuffer). When you want to do 2 pass blur you need texture A as source and texture B as destination. After first pass you can simply switch them and use texture B as source and texture A as destination.
So this mean than you only need 2 textures (surce and destination) but you must done this in final step because you wont be able to acces original, non blured, scene texture

[aeroz]

Thank you for the replies!
Your solutions make sense.
I'm actually implementing a kind of simple glow shader. I want to perform a blur with a specific color and then draw the original scene on top of it.
What can I do if I want to have a background image behind the glow? Do you use alpha channels in the FBO's textures?
I'm planning to do the following:
- Draw background to screen
- Draw part of the scene that should have the glow to Tex#1 with alpha channel.
- Blur Tex#1. (using the 2 pass method you explained)
- Draw Tex#1 to screen. (Background will still be visible because it has an alpha channel)
- Draw full scene to screen.

I'm not really sure if this is a good solution. Any comments are appreciated.

[Yours3!f]

Thank you for the replies!
Your solutions make sense.
I'm actually implementing a kind of simple glow shader. I want to perform a blur with a specific color and then draw the original scene on top of it.
What can I do if I want to have a background image behind the glow? Do you use alpha channels in the FBO's textures?
I'm planning to do the following:
- Draw background to screen
- Draw part of the scene that should have the glow to Tex#1 with alpha channel.
- Blur Tex#1. (using the 2 pass method you explained)
- Draw Tex#1 to screen. (Background will still be visible because it has an alpha channel)
- Draw full scene to screen.

I'm not really sure if this is a good solution. Any comments are appreciated.

I think there are many options, among those one would be masking with the alpha channel:
draw scene with glow parameter in alpha channel (ie alpha = 0 if no glow, alpha = 1 if glow)
blur the scene, and multiply each sample with the original (middle) sample's alpha.
so if the given pixel should glow ==> 1 * (sample #1 + sample #2 + ... + sample #n) = some color
if the given pixel should not glow ==> 0 * (sample #1 + sample #2 + ... + sample #n) = 0
output = original + blurred

or using stencil masking:
draw scene with setting the stencil bit where it needs to glow.
do the blurring with stencil testing enabled.

Edited by Yours3!f, 02 December 2012 - 05:23 PM.

[aeroz]

FBO, shaders, and those modern computer graphics concepts are very new to me.
But they are very interesting!
Thank you for the tips!

[larspensjo]

There are some tricks you can do if you want to blur. If you set the GL_TEXTURE_MAG_FILTER to GL_LINEAR, and then sample in the middle between 4 pixels you will get an automatic average of these 4, almost for free. Though I haven't studied the algorithm for Gaussian blur so I am not sure how much it helps? But I have a vague memory of someone doing a more efficient blurring with the combination of these two.

The disadvantage is that the resulting points are 0.5 pixels shifted.