Render to texture, etc.
Author
Afternoon. I was wondering if someone could answer a few questions I have, to make absolutely sure that I'm solving some particular problems I have at the moment in the right way. Here goes:
With render to texture, is there any direct OpenGL support for it, apart from glCopyTexImage2D? The method I always saw used for render to texture was to set up a suitable viewport and projection, render to the frame buffer, then use glCopyTexImage2D to read the frame buffer into the texture. Is this still used? Or is there a better/faster/more widely used method?
When creating/implementing post-process effects (motion blur, HDR, etc.), is it still essentially a render to texture operation (in that you render the entire scene to a large texture, then process that texture in n passes to get the desired after-effect)? I understand that regular render to texture (as described above) would probably kill the frame rate. Should I use frame buffer objects for post-process effects? Is there a specific method for handling post-process effects?
In vertex and fragment programs, varying and uniform variables are used. I take this to mean that varying variables can change between the vertex and fragment program (ie. the vertex program can change a texture coordinate), and uniform variables are fixed, and cannot change inside either the vertex program or fragment program. Is this correct? Does a vertex or fragment program ever change the data that is sent to it?
I know these seem like silly questions. Most of them I'd be able to find out by testing, but both of my goddamn computers are broken, so I can't test anything out until they're fixed, and I need to know about these questions in the meantime. Thanks in advance for any replies.
PS. There are no hardware constraints; I'm looking for the best and fastest methods available.
Well, let's see if I can be of any help.
First off, yes, post processing effects always need a previous render step. The easiest way for doing it is just by doing what you say: "change the viewport size to some valid texture size, render, then use glCopyTexImage2D or something similar".
The problem with this is that it completely trashes whatever previous rendering operations you were doing in the framebuffer, so you need to adapt your engine code to make the actual rendering pass the last one; that is, you first perform all the post process framebuffers, and once you got them, proceed with the final render.
This can be quite messy, so, [un]fortunately, there're some GL extensions you can use to make things "easier", like the "render to texture" (can't remember the exact name)...but then, you need to be sure the extension is supported by the current graphics card :( But if it is, then you just can "bind" a texture to the GL, so all rendering operations can be sent to it, keeping the standard framebuffer "safe" until you "unbind" it to continue your normal rendering.
Take a look at the OpenGL site and check the extensions; there should be some one called "GL_EXT_framebuffer_object" or similar. Or you can visit www.realtech-vr.com, they have made a tiny little free program called "OpenGL Extensions Viewer" that can help you find wich card supports wich extension, and can look the extension specificacions for you in the OpenGL site :P
And regarding shaders and GLSL, yes: you're right. Uniforms are some kind of "vertex shader local variables" than can't be modified during the vertex program's execution, but can be changed programatically through the GL API before the actual rendering takes place.
The "varyings" are some kind of "comunication mechanism" the vertex shader can use to pass information to the fragment shader. Basically the vertex shader can pass any float value to the fragment shader, but that value will be interpolated (linearly I guess) acrosss all fragments; so if the vertex shader sets a varying to 1.0 as the result for the first vertex of a line, and a 0.0 as the result for the second vertex, then the fragments belonging to the linle will be receiving a linear progression of values going smoothly from 1.0 to 0.0.
Regards,
First off, yes, post processing effects always need a previous render step. The easiest way for doing it is just by doing what you say: "change the viewport size to some valid texture size, render, then use glCopyTexImage2D or something similar".
The problem with this is that it completely trashes whatever previous rendering operations you were doing in the framebuffer, so you need to adapt your engine code to make the actual rendering pass the last one; that is, you first perform all the post process framebuffers, and once you got them, proceed with the final render.
This can be quite messy, so, [un]fortunately, there're some GL extensions you can use to make things "easier", like the "render to texture" (can't remember the exact name)...but then, you need to be sure the extension is supported by the current graphics card :( But if it is, then you just can "bind" a texture to the GL, so all rendering operations can be sent to it, keeping the standard framebuffer "safe" until you "unbind" it to continue your normal rendering.
Take a look at the OpenGL site and check the extensions; there should be some one called "GL_EXT_framebuffer_object" or similar. Or you can visit www.realtech-vr.com, they have made a tiny little free program called "OpenGL Extensions Viewer" that can help you find wich card supports wich extension, and can look the extension specificacions for you in the OpenGL site :P
And regarding shaders and GLSL, yes: you're right. Uniforms are some kind of "vertex shader local variables" than can't be modified during the vertex program's execution, but can be changed programatically through the GL API before the actual rendering takes place.
The "varyings" are some kind of "comunication mechanism" the vertex shader can use to pass information to the fragment shader. Basically the vertex shader can pass any float value to the fragment shader, but that value will be interpolated (linearly I guess) acrosss all fragments; so if the vertex shader sets a varying to 1.0 as the result for the first vertex of a line, and a 0.0 as the result for the second vertex, then the fragments belonging to the linle will be receiving a linear progression of values going smoothly from 1.0 to 0.0.
Regards,
Pixel buffer objects are what you should try and use for RTT. You can set them as render target, or buffer with single calls:
Pretty straight forward to use, a total pain to set up.
On vertex / fragment programs:
'varying' variables are ones which the GPU will interpolate for you across the triangle, as it does with color, normal, and UVs. 'uniform' variables are not interpolated.
bool PBuffer::BindAsRenderTarget(){ m_holdDC = wglGetCurrentDC(); m_holdGLRC = wglGetCurrentContext(); if (!wglMakeCurrent(m_hDC, m_hGLRC)) return(false); m_bIsCurrent = true; return(true);}void PBuffer::BindAsTexture(){ if (!m_bIsTexture) return; wglBindTexImageARB(m_hPBuffer, WGL_FRONT_LEFT_ARB); m_bIsBound = true;}Pretty straight forward to use, a total pain to set up.
On vertex / fragment programs:
'varying' variables are ones which the GPU will interpolate for you across the triangle, as it does with color, normal, and UVs. 'uniform' variables are not interpolated.
Mmmm...I tried pbuffers too, and I was really disappointed by the poor performance they showed.
Basically you need to make a reder context change for each pbuffer you want to render to, and that could be a real pain as it involves lot of GL state.
Besides, if not specified, pbuffers do NOT share texture objects, and if you make the GL call to make it share texture objects among pbuffers, then the render context change is even slower.
I would rather use the GL_EXT_framebuffer_object or whatever the extension name is, as it doesn't imply a context switch. So it *should* be faster. The only problem is that it is a relative "new" extension, so not all cards may support it.
Basically you need to make a reder context change for each pbuffer you want to render to, and that could be a real pain as it involves lot of GL state.
Besides, if not specified, pbuffers do NOT share texture objects, and if you make the GL call to make it share texture objects among pbuffers, then the render context change is even slower.
I would rather use the GL_EXT_framebuffer_object or whatever the extension name is, as it doesn't imply a context switch. So it *should* be faster. The only problem is that it is a relative "new" extension, so not all cards may support it.
firstly, pbuffers != Pixel buffer objects, PBOs are a different extension which enable the functions which read back from the framebuffer todo so async.
The modern way to do RTT is via frame buffer objects (FBO) as derodo pointed out. They should be supported on anything from the 9500 up from ATI and from the GFFX series from NV. The speed and features might vary however.
However, while they might be only for newer hardware, they are MUCH nicer to work with and considerably more flexible, if only because you are getting a TRUE RRT, not a proxy system like you do with pbuffers.
That said, for all their faults pbuffers do still have their useage as they are a completely new context and as such are completely self contained and there are times when you need this functionality.
As for the shading question;
As pointed out varying change across the face of a polygon, uniforms are more like 'per object' variables (well, per polygon, but there is a fair amount of overhead when it comes to changing them so per object or per batch of objects is a better useage pattern).
As for changing the data sent to them, shaders do change it however vertex shaders can only read from uniform and attribute variables and write out to varying variables. On the other hand fragment programs can only write out to the framebuffer and read from varying and uniforms (and ofcourse textures). Neither can write to uniforms, think of them as run-time constants.
The modern way to do RTT is via frame buffer objects (FBO) as derodo pointed out. They should be supported on anything from the 9500 up from ATI and from the GFFX series from NV. The speed and features might vary however.
However, while they might be only for newer hardware, they are MUCH nicer to work with and considerably more flexible, if only because you are getting a TRUE RRT, not a proxy system like you do with pbuffers.
That said, for all their faults pbuffers do still have their useage as they are a completely new context and as such are completely self contained and there are times when you need this functionality.
As for the shading question;
As pointed out varying change across the face of a polygon, uniforms are more like 'per object' variables (well, per polygon, but there is a fair amount of overhead when it comes to changing them so per object or per batch of objects is a better useage pattern).
As for changing the data sent to them, shaders do change it however vertex shaders can only read from uniform and attribute variables and write out to varying variables. On the other hand fragment programs can only write out to the framebuffer and read from varying and uniforms (and ofcourse textures). Neither can write to uniforms, think of them as run-time constants.
from my testing FBO is only maximum 10% faster (in benchmark) than using the backbuffer and updating a texture with glCopyTexSubImage2d(..)
so dont expect speed increases from using them
the FBO whilst not much quicker does offer other benifits eg, youre not limited to the framebuffers constraints
so dont expect speed increases from using them
the FBO whilst not much quicker does offer other benifits eg, youre not limited to the framebuffers constraints
FBO is the immediate equivalent of setting a texture surface as render target in DirectX.
However, render-to-surface textures may perform poorly if you use them in an arbitrary alignment, whereas they will perform well if you just use it as a screen-aligned texture. This is because they're usually not tiled on the card, but instead stored with regular scanline alignment.
Meanwhile, CopyTexSubImage textures (the equivalent of using StretchRect blits in D3D9, I think) will swizzle the texture data to perform equally well in all alignments, so if you don't just use the RTT as a screen overlay or multi-pass, the blit might actually be better.
However, render-to-surface textures may perform poorly if you use them in an arbitrary alignment, whereas they will perform well if you just use it as a screen-aligned texture. This is because they're usually not tiled on the card, but instead stored with regular scanline alignment.
Meanwhile, CopyTexSubImage textures (the equivalent of using StretchRect blits in D3D9, I think) will swizzle the texture data to perform equally well in all alignments, so if you don't just use the RTT as a screen overlay or multi-pass, the blit might actually be better.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
