Hi!
I have read that deferred shading is quite simple rendering technique. Although I have tried to get it work for a week already. My demo program is written in OpenGL.
I know how to use Frame Buffers, Multiple Render Targets, Shaders and I know some theory behind deferred shading (have read many articles, slides, googled a lot, downloaded samples).
Questions about G-Buffers.
1) Do I need to render depth or position to the one of the texture? Is there any difference?
2) Is it correct to render depth/positions and normals in camera space or in world space? At the moment position and normal colors are moving with camera.
My english isn't very good, sorry for that :)
Any help will be appreciated!
Karmux
Questions about deferred shading
Hello,
1.) You can render depth, positions or both. Although currently it is common to render depth only (it saves space, fillrate, etc... e.g. it is faster at same quality).
If you render positions, you have to use 32-bits on all three axis, e.g. GL_RGB32F texture at least.
If you render depth only, you have to use just 32-bits on depth (or even just 24-bits) with GL_R32F (it is most common today), and also I think (not sure) that in GL 3 and higher you can store and use directly depth buffer from frame buffer (it can be either 24 or 32 bits), which will save you one GL_R32F texture in memory and some more fillrate.
Although you have to reconstruct positions from depth of course if you store depth only. Which is fairly simple, but on the other hand imagine an example ... you've got lots and lots of fillrate but not enough computational power - one can store positions then and even have faster application in final ... so of course using depth or positions depends! Although in most cases depth-approach beats positions in performance.
2.) You can work in any space you wish (as long as you have everything in correct = same space), although it is most common to use view (or one can call it camera) space.
1.) You can render depth, positions or both. Although currently it is common to render depth only (it saves space, fillrate, etc... e.g. it is faster at same quality).
If you render positions, you have to use 32-bits on all three axis, e.g. GL_RGB32F texture at least.
If you render depth only, you have to use just 32-bits on depth (or even just 24-bits) with GL_R32F (it is most common today), and also I think (not sure) that in GL 3 and higher you can store and use directly depth buffer from frame buffer (it can be either 24 or 32 bits), which will save you one GL_R32F texture in memory and some more fillrate.
Although you have to reconstruct positions from depth of course if you store depth only. Which is fairly simple, but on the other hand imagine an example ... you've got lots and lots of fillrate but not enough computational power - one can store positions then and even have faster application in final ... so of course using depth or positions depends! Although in most cases depth-approach beats positions in performance.
2.) You can work in any space you wish (as long as you have everything in correct = same space), although it is most common to use view (or one can call it camera) space.
I personally prefer using the depth rather than a position map. There are issues with the accuracy for some things, but that can be worked around. Plus you get the depth map for free anyway.
Your main issue is that a position map will kill your bandwidth. Even on PC you will be severely bottle necking yourself at higher resolutions. Also there's no way you could fit a position map on the 360 without pain.
I personally do everything in world space, even if it does require the odd matrix inversion and such. I just find it easier to visualise the problem.
Your main issue is that a position map will kill your bandwidth. Even on PC you will be severely bottle necking yourself at higher resolutions. Also there's no way you could fit a position map on the 360 without pain.
I personally do everything in world space, even if it does require the odd matrix inversion and such. I just find it easier to visualise the problem.
Author
Thanks for your informative replies! Some things are much clearer.
Now I'm trying to write gbuffer shaders according to your posts.
How do I get vertex position, depth and normal vector in world space in GLSL?
Now I'm trying to write gbuffer shaders according to your posts.
How do I get vertex position, depth and normal vector in world space in GLSL?
Surely you're doing that anyway in your normal lighting calculations???
Transform the vertex into world space then pass through its position to the pixel shader and there you have it. Depth is the distance between camera position and pixel position obviously :)
Transform the vertex into world space then pass through its position to the pixel shader and there you have it. Depth is the distance between camera position and pixel position obviously :)
Author
Quote:Original post by Darg
Transform the vertex into world space then pass through its position to the pixel shader and there you have it.
Following screen (not mine) is correct?
http://home.comcast.net/~agentsnoop/correctwpos.png
-------- My rendering cycle:
1) Clear scene
2) Bind FBO with MRT
3) Bind shader
4)
glGetFloatv(GL_MODELVIEW_MATRIX, mvm);
glGetFloatv(GL_PROJECTION_MATRIX, pm);
gbuffer_shader->set("mvm", mvm);
gbuffer_shader->set("pm", pm);
5) Move camera
6) Draw scene
7) Unbind shader
8) Render FBO to the fullscreen quad
-------- Vertex shader:
varying vec3 vertex_position;
varying vec3 normal_vector;
uniform mat4 mvm;
uniform mat4 pm;
void main() {
gl_Position = mvm * pm * gl_Vertex;
vertex_position = normalize(vec3(mvm * gl_Vertex));
normal_vector = normalize(vec3(gl_NormalMatrix * gl_Normal));
gl_TexCoord[0] = gl_TextureMatrix[0] * gl_MultiTexCoord0;
gl_TexCoord[1] = gl_TextureMatrix[1] * gl_MultiTexCoord1;
}
I don't see nothing on screen. What I'm doing wrong? :(
[Edited by - Karmux on December 10, 2010 1:41:27 PM]
Quote:Original post by KarmuxQuote:Original post by Darg
Transform the vertex into world space then pass through its position to the pixel shader and there you have it.
Following screen (not mine) is correct?
http://home.comcast.net/~agentsnoop/correctwpos.png
-------- My rendering cycle:
1) Clear scene
2) Bind FBO with MRT
3) Bind shader
4)
glGetFloatv(GL_MODELVIEW_MATRIX, mvm);
glGetFloatv(GL_PROJECTION_MATRIX, pm);
gbuffer_shader->set("mvm", mvm);
gbuffer_shader->set("pm", pm);
5) Move camera
6) Draw scene
7) Unbind shader
8) Render FBO to the fullscreen quad
-------- Vertex shader:
varying vec3 vertex_position;
varying vec3 normal_vector;
uniform mat4 mvm;
uniform mat4 pm;
void main() {
gl_Position = mvm * pm * gl_Vertex;
vertex_position = normalize(vec3(mvm * gl_Vertex));
normal_vector = normalize(vec3(gl_NormalMatrix * gl_Normal));
gl_TexCoord[0] = gl_TextureMatrix[0] * gl_MultiTexCoord0;
gl_TexCoord[1] = gl_TextureMatrix[1] * gl_MultiTexCoord1;
}
I don't see nothing on screen. What I'm doing wrong? :(
Fragment shader?
Author
I had a little progress. Now I see scene but all shapes are in position (0,0,0). Translations and rotations in draw_scene() doesn't give any effect.
What I have done so far:
* put 5) before 4) in pipeline
* changed 3rd variable in glUniformMatrix4fv to false
* updated shaders
-------- Vertex Shader:
-------- Fragment Shader:
Screen to illustrate things:
http://www.upload.ee/image/977435/positions2.png
This one rectangle should be a box made of 6 rotated rectangles.
There is something wrong with projection matrix.
[Edited by - Karmux on December 11, 2010 5:25:32 AM]
What I have done so far:
* put 5) before 4) in pipeline
* changed 3rd variable in glUniformMatrix4fv to false
* updated shaders
-------- Vertex Shader:
varying vec3 vertex_position;varying vec3 normal_vector;uniform mat4 mm;uniform mat4 pm;void main() { gl_Position = pm * mm * gl_Vertex; vertex_position = vec3(mm * gl_Vertex); normal_vector = vec3(gl_NormalMatrix * gl_Normal); gl_TexCoord[0] = gl_TextureMatrix[0] * gl_MultiTexCoord0; gl_TexCoord[1] = gl_TextureMatrix[1] * gl_MultiTexCoord1;}-------- Fragment Shader:
varying vec3 vertex_position;varying vec3 normal_vector;uniform sampler2D texture0, texture1;vec4 texel, texel0, texel1;void main() { texel0 = texture2D(texture0, gl_TexCoord[0].st); texel1 = texture2D(texture1, gl_TexCoord[1].st); texel = texel0 * texel1; gl_FragData[0] = texel; gl_FragData[1] = vec4(vertex_position, 1); gl_FragData[2] = vec4(normal_vector, 1);}Screen to illustrate things:
http://www.upload.ee/image/977435/positions2.png
This one rectangle should be a box made of 6 rotated rectangles.
There is something wrong with projection matrix.
[Edited by - Karmux on December 11, 2010 5:25:32 AM]
You don't seem to ever be taking the world matrix of the individual objects into account.
The correct steps to render a model are as follow:
1: Obtain the objects world matrix from its position, rotation and scale.
2: Pass this world matrix to the vertex shader each time you draw an object.
3: Multiply the incoming vertex position by the world matrix. This gives you the world position of the vertex which is what you want to output to the GBuffer either as position in three 32-bit channels or as depth in one 32-bit channel, I would recommend depth.
4: Multiply this world position by the View and Projection matrices to obtain the view space position of the vertex, this is what is passed through to the pixel shader.
You appear to be completely ignoring the individual objects world matrices.
The correct steps to render a model are as follow:
1: Obtain the objects world matrix from its position, rotation and scale.
2: Pass this world matrix to the vertex shader each time you draw an object.
3: Multiply the incoming vertex position by the world matrix. This gives you the world position of the vertex which is what you want to output to the GBuffer either as position in three 32-bit channels or as depth in one 32-bit channel, I would recommend depth.
4: Multiply this world position by the View and Projection matrices to obtain the view space position of the vertex, this is what is passed through to the pixel shader.
You appear to be completely ignoring the individual objects world matrices.
Author
I decided to do deferred rendering in view space.
Depth texture looks fine but when I reconstruct positions then entire screen is green-yellow-red-black.
If I multiply positions by color texture then only my model is green-yellow-red-black and empty space is black and colors are moving with camera.
Any suggestions are welcome :)
Depth texture looks fine but when I reconstruct positions then entire screen is green-yellow-red-black.
If I multiply positions by color texture then only my model is green-yellow-red-black and empty space is black and colors are moving with camera.
// Vertex gl_Position = gl_ModelViewProjectionMatrix * gl_Vertex; unpro = gl_ModelViewMatrix * gl_Vertex;// Fragment float linearize_depth(float z) { float n = 0.001; float f = 10.0; return (2.0 * n) / (f + n - z * (f - n)); } ... vec3 color = texture2D(color_buffer, gl_TexCoord[0].xy).rgb; float depth = linearize_depth(texture2D(depth_buffer, gl_TexCoord[0].xy).r); vec3 normal = texture2D(normal_buffer, gl_TexCoord[0].xy).rgb; vec4 position = unpro * vec4(gl_FragCoord.xy, depth, 1.0); position /= position.w; ... gl_FragColor = position;Any suggestions are welcome :)
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
