• FEATURED

View more

View more

View more

### Image of the Day Submit

IOTD | Top Screenshots

### The latest, straight to your Inbox.

Subscribe to GameDev.net Direct to receive the latest updates and exclusive content.

# GPU voxelization - Conservative rasterization issue.

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.

2 replies to this topic

### #1elurahu  Members

Posted 27 December 2012 - 01:59 PM

While implementing the technique outlined in Crassins / Greens work found in OpenGL insights I've run into some issues.

To get proper voxel coverage I need to dilate the triangles before they are sent off to rasterization. As written in the article they are using an older technique found in GPU Gems 2 on conservative rasterization.

My issue is that the my geometry shader projects the triangles using orthographic projections where the conservative rasterization technique only works when using perspective. As far as I can tell.

### #2Bacterius  Members

Posted 27 December 2012 - 04:08 PM

I don't see why you shouldn't be able to use conservative rasterization with an orthographic projection - the only difference is that you accept all pixels which are ever-so-slightly covered by a triangle (conservative estimate) instead of discarding those with less than 50% coverage. Though obviously the math will be different. In fact, it will probably be simpler since the projection is not as complicated, for instance I would start by getting the triangle's screen-space dimensions. Since this is an orthographic projection, this is trivial and you can just use the post-transformation (x, y) coordinates of each vertex.

At this point you have all the information you need to apply the corrective algorithms (42.2) in the GPU Gems article, obtain the new optimal vertex coordinates, and invert the transformation to get your corrected vertices back into world space (if you need this for rendering - the rasterization step doesn't care and just wants clip space coordinates).

Most of the math at the end of this article concerns how to correct depth as well, since this is important in perspective rendering. However, that's not the case for orthographic projections, as there is no depth to worry about then, so it seems to me that you can ignore all of that.

At least that's my take on it - I've never implemented any of this, just going on mathematical principles and logic here.

“If I understand the standard right it is legal and safe to do this but the resulting value could be anything.”

### #3elurahu  Members

Posted 27 December 2012 - 04:44 PM

I agree that having orthographic projects makes alot of stuff easiere as I can just extend the clipspace vertices directly. The problem lies in HOW much to extend. The article outlines that you'll need to extend along the worst-case semidiagonal. The author uses the fact that a line in clipspace can be represented as a plane through the camera (0,0,0) in projected space and then dilutes along the plane normal - Then to find the new vertex positions they do a plane - plane intersection using a cross product.

I just don't think this works for orthographic projects as all points on a direction vector from the origin doesn't project to a single point.

For reference here my geometry code for the impl.

// Triangle bounding box.
vec4 f4AABB;

// Compute v0.
vec4 f4CSV0 = m4ViewProj * RF_WORLD * gl_in[0].gl_Position;
f4AABB.xy = f4CSV0.xy;
f4AABB.zw = f4CSV0.xy;

// Compute v1.
vec4 f4CSV1 = m4ViewProj * RF_WORLD * gl_in[1].gl_Position;
f4AABB.xy = min(f4AABB.xy, f4CSV1.xy);
f4AABB.zw = max(f4AABB.zw, f4CSV1.xy);

// Compute v2.
vec4 f4CSV2 = m4ViewProj * RF_WORLD * gl_in[2].gl_Position;
f4AABB.xy = min(f4AABB.xy, f4CSV2.xy);
f4AABB.zw = max(f4AABB.zw, f4CSV2.xy);

// Extend and set AABB.
f4AABB.xy -= vec2(fHalfPixel);
f4AABB.zw += vec2(fHalfPixel);
OUT.f4AABB = f4AABB;

// Compute dialated edges.
vec3 f3Plane[3];
f3Plane[0] = cross(f4CSV0.xyw - f4CSV2.xyw, f4CSV2.xyw);
f3Plane[1] = cross(f4CSV1.xyw - f4CSV0.xyw, f4CSV0.xyw);
f3Plane[2] = cross(f4CSV2.xyw - f4CSV1.xyw, f4CSV1.xyw);
f3Plane[0].z -= dot(vec2(fHalfPixel), abs(f3Plane[0].xy));
f3Plane[1].z -= dot(vec2(fHalfPixel), abs(f3Plane[1].xy));
f3Plane[2].z -= dot(vec2(fHalfPixel), abs(f3Plane[2].xy));

// Compute plane intersections.
f4CSV0.xyw = cross(f3Plane[0], f3Plane[1]);
f4CSV1.xyw = cross(f3Plane[1], f3Plane[2]);
f4CSV2.xyw = cross(f3Plane[2], f3Plane[0]);

// Emit vertex data.
OUT.f3CSPosition = m3Rotation * (f4CSV0.xyz / f4CSV0.w);
OUT.f3Color = f3Color;
gl_Position = f4CSV0;
EmitVertex();

OUT.f3CSPosition = m3Rotation * (f4CSV1.xyz / f4CSV1.w);
OUT.f3Color = f3Color;
gl_Position = f4CSV1;
EmitVertex();

OUT.f3CSPosition = m3Rotation * (f4CSV2.xyz / f4CSV2.w);
OUT.f3Color = f3Color;
gl_Position = f4CSV2;
EmitVertex();

EndPrimitive();

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.