Shaders and Coordinate Spaces
The Major Coordinate Spaces
When writing complex shaders, it is often necessary to do a lot of coordinate system conversion. In order to get this right, it is important to be aware of all the different coordinate spaces that panda uses. You must know what “space” the coordinate is in. Here is a list of the major coordinate spaces:
- Model Space
If a coordinate is in model space, then it is relative to the center of the model currently being rendered. The vertex arrays are in model space, therefore, if you access the vertex position using vtx_position, you have a coordinate in model space. Model space is z-up right-handed.
- World Space
If a coordinate is in world space, then it is relative to the scene’s origin. World space is z-up right-handed.
- View Space
If a coordinate is in view space, then it is relative to the camera. View space is z-up right-handed.
- API View Space
This coordinate space is identical to view space, except that the axes may be flipped to match the natural orientation of the rendering API. In the case of OpenGL, API view space is y-up right-handed. In the case of DirectX, API view space is y-up left-handed.
- Clip Space
Panda’s clip space is a coordinate system in which (X/W, Y/W) maps to a screen pixel, and (Z/W) maps to a depth-buffer value. All values in this space range over [-1,1].
- API Clip Space
This coordinate space is identical to clip space, except that the axes may be flipped to match the natural orientation of the rendering API, and the numeric ranges may be rescaled to match the needs of the rendering API. In the case of OpenGL, the (Z/W) values range from [-1, 1]. In the case of DirectX, the (Z/W) values range from [0,1].
In OpenGL, “clip space” and “API clip space” are equivalent.
Supplying Translation Matrices to a Shader
You can use a shader parameter named “trans_x_to_y” to automatically obtain a matrix that converts any coordinate system to any other. The words x and y can be “model,” “world,” “view,” “apiview,” “clip,” or “apiclip.” Using this notation, you can build up almost any transform matrix that you might need. Here is a short list of popular matrices that can be recreated using this syntax. Of course, this isn’t even close to exhaustive: there are seven keywords, so there are 7x7 possible matrices, of which 7 are the identity matrix.
Desired Matrix |
Source |
Syntax |
GLSL input |
|---|---|---|---|
The Modelview Matrix |
gsg.getInternalTransform() |
trans_model_to_apiview |
p3d_ModelViewMatrix |
The Projection Matrix |
gsg.getProjectionMat() |
trans_apiview_to_apiclip |
p3d_ProjectionMatrix |
the DirectX world matrix |
model.getNetTransform() |
trans_model_to_world |
p3d_ModelMatrix |
the DirectX view matrix |
scene.getCsWorldTransform() |
trans_world_to_apiview |
p3d_ViewMatrix |
scene.getCameraTransform() |
trans_view_to_world |
||
scene.getWorldTransform() |
trans_world_to_view |
||
gsg.getExternalTransform() |
trans_model_to_view |
||
gsg.getCsTransform() |
trans_view_to_apiview |
||
gsg.getInvCsTransform() |
trans_apiview_to_view |
A note about GLSL inputs
The p3d_ModelViewMatrix and p3d_ProjectionMatrix by default transform to and
from “apiview” space, in order to match the behavior of the equivalent
gl_-prefixed inputs from earlier GLSL versions. Panda3D traditionally uses a
right-handed Y-up coordinate space for all OpenGL operations because some OpenGL
fixed-function features rely on this space in order to produce the correct
results.
However, if you develop a largely shader-based application and/or don’t really use features like fixed-function sphere mapping, you may choose to disable this conversion to Y-up space. This will define “apiview” space to be equivalent to “view” space, which simplifies many things, and will reduce overhead due to unnecessary coordinate space conversion, especially as “apiclip” and “clip” are already equivalent in OpenGL as well.
To do this, place gl-coordinate-system default in your Config.prc file.
Recommendation: Don’t use API View Space or API Clip Space
The coordinate systems “API View Space” and “API Clip Space” are not very useful. The fact that their behavior changes from one rendering API to the next makes them extremely hard to work with. Of course, you have to use the composed modelview/projection matrix to transform your vertices, and in doing so, you are implicitly using these spaces. But aside from that, it is strongly recommended that you not use these spaces for anything else.
Model_of_x, View_of_x, Clip_of_x
When you use the word “model” in a trans directive, you implicitly mean “the
model currently being rendered.” But you can make any nodepath accessible to the
shader subsystem using NodePath.set_shader_input():
NodePath myhouse = window->load_model(framework.get_models(), "myhouse");
window->get_render().set_shader_input("myhouse", myhouse);
Then, in the shader, you can convert coordinates to or from the model-space of this particular nodepath:
uniform float4x4 trans_world_to_model_of_myhouse
or, use the syntactic shorthand:
uniform float4x4 trans_world_to_myhouse
Likewise, you can create a camera and pass it into the shader subsystem. This is particularly useful when doing shadow mapping:
render.set_shader_input("shadowcam", shadowcam);
Now you can transform vertices into the clip-space of the given camera using this notation:
uniform float4x4 trans_model_to_clip_of_shadowcam
If you transform your model’s vertices from model space into the clip space of a shadow camera, the resulting (X/W,Y/W) values can be used as texture coordinates to projectively texture the shadow map onto the scene (after rescaling them), and the (Z/W) value can be compared to the value stored in the depth map (again, after rescaling it).
Panda does support the notation “trans_x_to_apiclip_of_y”, but again, our recommendation is not to use it.
You can transform a vertex to the view space of an alternate camera, using “view of x.” In fact, this is exactly identical to “model of x,” but it’s probably good form to use “view of x” when x is a camera.