This extension requires multiViewport and geometryShader
features to be useful.
Contributors
Daniel Koch, NVIDIA
Jeff Bolz, NVIDIA
Description
This extension provides a new per-viewport swizzle that can modify the
position of primitives sent to each viewport. New viewport swizzle state
is added for each viewport, and a new position vector is computed for
each vertex by selecting from and optionally negating any of the four
components of the original position vector.
This new viewport swizzle is useful for a number of algorithms,
including single-pass cube map rendering (broadcasting a primitive to
multiple faces and reorienting the vertex position for each face) and
voxel rasterization. The per-viewport component remapping and negation
provided by the swizzle allows application code to re-orient
three-dimensional geometry with a view along any of the X, Y, or
Z axes. If a perspective projection and depth buffering is required,
1/W buffering should be used, as described in the single-pass cube map
rendering example in the “Issues” section below.
1) Where does viewport swizzling occur in the pipeline?
RESOLVED: Despite being associated with the viewport, viewport
swizzling must happen prior to the viewport transform. In particular, it
needs to be performed before clipping and perspective division.
The viewport mask expansion (VK_NV_viewport_array2) and the viewport
swizzle could potentially be performed before or after transform
feedback, but feeding back several viewports worth of primitives with
different swizzles does not seem particularly useful. This specification
applies the viewport mask and swizzle after transform feedback, and
makes primitive queries only count each primitive once.
2) Any interesting examples of how this extension,
VK_NV_viewport_array2, and VK_NV_geometry_shader_passthrough can be
used together in practice?
RESOLVED: One interesting use case for this extension is for
single-pass rendering to a cube map. In this example, the application
would attach a cube map texture to a layered FBO where the six cube
faces are treated as layers. Vertices are sent through the vertex shader
without applying a projection matrix, where the gl_Position output is
(x,y,z,1) and the center of the cube map is at (0,0,0). With unextended
Vulkan, one could have a conventional instanced geometry shader that
looks something like the following:
layout(invocations = 6) in; // separate invocation per face
layout(triangles) in;
layout(triangle_strip) out;
layout(max_vertices = 3) out;
in Inputs {
vec2 texcoord;
vec3 normal;
vec4 baseColor;
} v[];
out Outputs {
vec2 texcoord;
vec3 normal;
vec4 baseColor;
};
void main()
{
int face = gl_InvocationID; // which face am I?
// Project gl_Position for each vertex onto the cube map face.
vec4 positions[3];
for (int i = 0; i < 3; i++) {
positions[i] = rotate(gl_in[i].gl_Position, face);
}
// If the primitive does not project onto this face, we are done.
if (shouldCull(positions)) {
return;
}
// Otherwise, emit a copy of the input primitive to the
// appropriate face (using gl_Layer).
for (int i = 0; i < 3; i++) {
gl_Layer = face;
gl_Position = positions[i];
texcoord = v[i].texcoord;
normal = v[i].normal;
baseColor = v[i].baseColor;
EmitVertex();
}
}
With passthrough geometry shaders, this can be done using a much simpler
shader:
layout(triangles) in;
layout(passthrough) in Inputs {
vec2 texcoord;
vec3 normal;
vec4 baseColor;
}
layout(passthrough) in gl_PerVertex {
vec4 gl_Position;
} gl_in[];
layout(viewport_relative) out int gl_Layer;
void main()
{
// Figure out which faces the primitive projects onto and
// generate a corresponding viewport mask.
uint mask = 0;
for (int i = 0; i < 6; i++) {
if (!shouldCull(face)) {
mask |= 1U << i;
}
}
gl_ViewportMask = mask;
gl_Layer = 0;
}
The application code is set up so that each of the six cube faces has a
separate viewport (numbered 0 to 5). Each face also has a separate
swizzle, programmed via the PipelineViewportSwizzleStateCreateInfoNV
pipeline state. The viewport swizzle feature performs the coordinate
transformation handled by the rotate() function in the original
shader. The viewport_relative layout qualifier says that the viewport
number (0 to 5) is added to the base gl_Layer value of 0 to determine
which layer (cube face) the primitive should be sent to.
Note that the use of the passed through input normal in this example
suggests that the fragment shader in this example would perform an
operation like per-fragment lighting. The viewport swizzle would
transform the position to be face-relative, but normal would remain in
the original coordinate system. It seems likely that the fragment shader
in either version of the example would want to perform lighting in the
original coordinate system. It would likely do this by reconstructing
the position of the fragment in the original coordinate system using
gl_FragCoord, a constant or uniform holding the size of the cube face,
and the input gl_ViewportIndex (or gl_Layer), which identifies the
cube face. Since the value of normal is in the original coordinate
system, it would not need to be modified as part of this coordinate
transformation.
Note that while the rotate() operation in the regular geometry shader
above could include an arbitrary post-rotation projection matrix, the
viewport swizzle does not support arbitrary math. To get proper
projection, 1/W buffering should be used. To do this:
Program the viewport swizzles to move the pre-projection W eye
coordinate (typically 1.0) into the Z coordinate of the swizzle
output and the eye coordinate component used for depth into the W
coordinate. For example, the viewport corresponding to the +Z face
might use a swizzle of (+X, -Y, +W, +Z). The Z normalized device
coordinate computed after swizzling would then be z'/w' =
1/Zeye.
On NVIDIA implementations supporting floating-point depth buffers
with values outside [0,1], prevent unwanted near plane clipping by
enabling depthClampEnable. Ensure that the depth clamp does not
mess up depth testing by programming the depth range to very large
values, such as minDepthBounds=-z, maxDepthBounds=+z, where z =
It should be possible to use IEEE infinity encodings also
(0xFF800000 for -INF, 0x7F800000 for +INF). Even when
near/far clipping is disabled, primitives extending behind the eye
will still be clipped because one or more vertices will have a
negative W coordinate and fail X/Y clipping tests.
On other implementations, scale X, Y, and Z eye coordinates so that
vertices on the near plane have a post-swizzle W coordinate of 1.0.
For example, if the near plane is at Zeye = 1/256, scale X, Y, and
Z by 256.
Adjust depth testing to reflect the fact that 1/W values are large
near the eye and small away from the eye. Clear the depth buffer to
zero (infinitely far away) and use a depth test of
COMPARE_OP_GREATER instead of
COMPARE_OP_LESS.