Jump to content

  • Log In with Google      Sign In   
  • Create Account


Deferred Depth Reconstruction


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
3 replies to this topic

#1 EnlightenedOne   Members   -  Reputation: 142

Like
0Likes
Like

Posted 19 April 2014 - 04:25 PM

Hi All,

 

I am porting/migrating my pipeline from forward to deferred rendering because I want precomputed atmospheric scattering. If I do not learn how to achieve the technique I will never die a happy man (and vice versa smile.png). The goal is:

 

 

The paper with it is interesting reading as is prolands source code example, however prolands demo does not actually use a GBuffer so I need to blend their high level behaviour from this DX port to achieve the effect in GLSL 330.

 

I am binding my render buffer as a texture (24bit 8 bit stencil), I can read and draw it to a quad on screen, I know that the none linear depth values it writes are valid and my shader/model binding process is tried and true (right handed). My current goal is to establish all the fiddly unpleasant transforms required to get from GBuffer post processing exposed and to better understand the most practical way to handle transforms around depth. My first stop was the realisation that if I can get the post process surfacePos below I will be home free as all the lighting is similar it is just the sources used to lookup values that have changed.

 

Here is the code I am attempting to port, the transforms through various coordinate space I have a loose grasp of, but the part I do not get is how the SV_POSITION translates in GLSL. Changing the gl_FragDepth to try and mimic screen space changes ends badly.

 

GBUFFER GEOMETRY___________________________________________________

 

struct VS_OUT {

float4 posH : SV_POSITION;

float3 posW : POSITION;

float3 tangent : TANGENT0;

float3 bitangent : TANGENT1;

float2 texC : TEXCOORD0;

};

 

...

 

Vertex shader snippet of interest to position:

 

output.posH = mul(float4(posWorld, 1.0f), g_viewProj);

output.posH.z = output.posH.z * output.posH.w * g_invFarPlane;

 

POST PROCESSING_____________________________________________________

 

Vertex Shader

static const float EPSILON_ATMOSPHERE = 0.002f;

static const float EPSILON_INSCATTER = 0.004f;

 

Texture2D g_depth;

Texture2D g_color;

Texture2D g_normal;

Texture2D g_texIrradiance;

Texture3D g_texInscatter;

float3 g_cameraPos;

float3 g_sunVector;

float4x4 g_cameraWorld;

float4 g_frustumFar[4];

float4 g_frustumNear[4];

 

struct VS_IN {

float3 posL : POSITION;

float2 texC : TEXCOORD0;

uint index : TEXCOORD1;

};

 

struct VS_OUT {

float4 posH : SV_POSITION;

float2 texC : TEXCOORD0;

float3 nearToFar : TEXCOORD2;

float3 cameraToNear : TEXCOORD3;

};

 

VS_OUT VS(VS_IN input) {

VS_OUT output;

output.posH = float4(input.posL,1.0f);

output.texC = input.texC;

float3 frustumFarWorld = mul(float4(g_frustumFar[input.index].xyz, 1.0f), g_cameraWorld).xyz;

float3 frustumNearWorld = mul(float4(g_frustumNear[input.index].xyz, 1.0f), g_cameraWorld).xyz;

output.cameraToNear = frustumNearWorld - g_cameraPos;

output.nearToFar = frustumFarWorld - frustumNearWorld;

return output;

}

 

Pixel Shader:

float4 PS_PLANET_DEFERRED(VS_OUT input) : SV_TARGET0 {

// reconstructing world space postion by interpolation

float depthVal = g_depth.SampleLevel( PointSamplerClamp, input.texC, 0 ).r;

float3 surfacePos = g_cameraPos + input.cameraToNear + depthVal * input.nearToFar;

 

// obtaining the view direction vector

float3 viewDir = normalize(input.nearToFar);

 

  • Can anyone confirm how the PosH value is impacting the encoding of the depth buffer and what the openGL equivalent would be?
  • Can anyone tell me the real values of those vec4[4] fustrums and where they can be derived from? I have no problem adding an index to my screen quad to link to it or building a clipping frustrum near or far plane. The problem I have is that it is glossed over and I am worried my deductive reasoning will be slower than the my goal. I really want my GBuffer ready to start assembling this post processing effect before the Easter holiday is done.

 

I believe g_cameraWorld is the world rotation of the camera.

 

I have gone through all of the following resources to try and understand how people are handling this process:

 

So far I have had no success reconstructing raw depth by blending these snippets input on the problem or testing them in relative isolation, I have a feeling that everyone is tampering with depth output in the geometry buffer but its not very clear in many of the snippets I have found exactly what parameters they are using to do this and why. I am going to try and focus on filling in the gaps from the model above because the resources above suggest it is still an efficient mechanism for solving reconstructing the desirable spaces in most post processes.

 

Does anyone have a tutorial where this reconstruction process is applied as a holistic piece of functioning code? I would love to see the implementation for these frustum shapes on the near and far plane. I just need to see a proper GBuffer pipeline using depth to reconstruct position and linear depth so I can reverse engineer and inspect its properties to understand the bugs in my own code and move on.

 

I really cannot wait to play with that effect. If I get my shaders to reconstruct from depth I will post them up and describe the parts that have as of yet confounded me.

 

I welcome any input.

 

Many Thanks,

Enlightened One



Sponsor:

#2 EnlightenedOne   Members   -  Reputation: 142

Like
0Likes
Like

Posted 20 April 2014 - 07:30 AM

I solved a bug in one of the references where they multiply 2 * zNear * zFar when they only need 2 * zNear in part of the equation, this got me seeing linearised space from a none linear depth in the real depth buffer. I reread the details of the precomputed atmospheric scattering paper and concluded that the SVPosition and Position allow the real position and clip space depth? to be set independently. Although I might have really screwed up in my interpretation of their duality.

 

Ok so I am attempting to understand how to force linear depth in the hardware Depth buffer to bring myself closer to the DX PCAS paper, the goal is linear depth in the real depth buffer with the world space position and by extension the view direction cheaply solved in the pixel shader from this depth.

 

Right now I am studying Yours3!f's description to progress:

http://www.gamedev.net/topic/624649-position-reconstruction-from-depth-in-opengl/

I am concerned he is using a separate texture from the real depth buffer... if he is using something to map to the SV_Poistion equivalent he is not hinting at it.

 

This article is interesting but just too abstract for me to apply:

http://mynameismjp.wordpress.com/2010/09/05/position-from-depth-3/

Off of it is a sample for "attack the depth buffer", it is a great sample but their pixel shader code has:

float2 DepthCS      : DEPTH;
float4 PositionCS   : SV_POSITION;
Again it seems not understanding the GL version of clip space is hurting me.

 

To better qualify my FBO I have 24/8 depth/stencil buffer, normal 32 and colour 32, I believe that should be enough.



#3 EnlightenedOne   Members   -  Reputation: 142

Like
1Likes
Like

Posted 20 April 2014 - 08:46 AM

My goodness it worked!

 

VERTEX

#version 330
 
in vec3 inPos;
in vec4 inCol;
//model being tested has no normals yet
 
uniform mat4 inverseProjectionMatrix;
 
uniform mat4 wvpMatrix;
uniform mat4 modelViewMatrix;
 
out vec3 pass_colour;
out vec4 worldPosition;
out float depth;
 
const float far_plane_distance = 15000.0f;
 
void main(void) {
gl_Position = wvpMatrix * vec4(inPos, 1);
 
worldPosition = gl_Position;
 
vec4 viewSpacePos = modelViewMatrix * vec4(inPos, 1);
 
depth = viewSpacePos.z / - far_plane_distance;
 
pass_colour = inCol.xyz;
}
 
 
FRAGMENT
 
#version 330
 
in vec3 pass_colour;
in vec4 worldPosition;
in float depth;
 
layout (location = 0) out vec4 ColourOut;   
layout (location = 1) out vec4 NormalOut;
layout (location = 2) out vec4 PosOut;     
 
void main(void) {
ColourOut = vec4(pass_colour, 1);
NormalOut = vec4(0.5,0.5,0.5,1); //dummy values please ignore
 
PosOut.xyz = worldPosition.xyz; //intending to use this to compare to the reconstructed positions as soon as I have either working smile.png
PosOut.w = 1;
gl_FragDepth = depth;
}

 

Then just sampling the depth:

    d = texture2D(texture_depth, pass_texCoord).x;
  DepthOut = vec4(d, d, d, 1.0);
 
My old linearisation from none linear depth buffer approach when using this is redundant as the depth texture is linear. Following on from his approach I expect I will get the desired view space positions necessary to get around.
 
I dug out this:
 
It suggests what I am doing is a recipe for performance disaster, however "gl_Position.z = viewSpacePos.z / - far_plane_distance;" simply causes depth testing to fail.
 
Interestingly but obviously I get less close banding from the none linear depth calculation, reconstructing linear depth without storing it must be the way forward.
 
I am about to find out if I stick with post GBuffer linear depth reconstruction can I still reconstruct the surface position using Yours3!f's method? more challenging will be to get the viewDir from this data.

Edited by EnlightenedOne, 20 April 2014 - 09:25 AM.


#4 EnlightenedOne   Members   -  Reputation: 142

Like
0Likes
Like

Posted 20 April 2014 - 10:54 AM

Looks like I can get to both view and world space if I plug my derived linear depth into:

http://mynameismjp.wordpress.com/2009/05/05/reconstructing-position-from-depth-continued/

 

Based on the attack the depth buffer pixel shader:

    float3 viewRay = float3(Input.PositionVS.xy / Input.PositionVS.z, 1.0f);
    float3 positionVS = LinearizedDepth(depthTextureForPizel) * viewRay;

Where positionVS is the world view project multiplied vertex as per standard transform, Yours3!f's method includes additional transforms that may be required for OpenGL but I haven't tested that yet.

 

The real work will be getting to world space without matrix multiplication per pixel that is my current task.






Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS