This demo several approaches for computing barycentric coordinates in the pixel shader.
Geometry is rendered using non-indexed draws. Vertex shader explicitly loads indices and vertices from index and vertex buffers and writes out barycentric coordinates. This approach is similar to using a geometry shader that outputs per-vertex barycentrics.
This approach results in geometry throughput ~2x slower than regular indexed rendering.
Geometry shader is used to output new triangles with explicit per-vertex barycentric coordinates. This approach does not require custom vertex fetching (unlike mode 1).
Performance is slightly worse than mode 1 on AMD Fury X, but better on NVIDIA 1080. In general, we are still looking at ~2x slower rendering in geometry-bound scenes.
Primitive indices and vertices are loaded in the pixel shader based on primitive ID. Positions are transformed into world space and resulting triangle is intersected with the eye ray to calculate barycentrics.
Despite doing quite a lot of work per pixel, this mode is much faster than modes 1 and 2 in geometry-heavy scenes.
On NVIDIA 1080 this runs ~25% slower than baseline "speed-of-light" shader that simply outpts texture coordinates in a geometry-bound scene. Interestingly, even with no geometry visible on screen (camera facing away, but still shading all vertices) performance is ~13% slower than "speed-of-light". It appears that simply using gl_PrimitiveID incurs an overhead.
AMD Fury X is ~10% slower than "speed-of-light" on average and ~7% slower in pure geometry-bound case, again suggesting an overhead from using gl_PrimitiveID.
This mode uses VK_NV_geometry_shader_passthrough extension. Fast / passthrough geometry shader is used to output world positions of triangles to the pixel shader, which then performs a ray-triangle intersection similar to mode 3.
Performance is slightly better than mode 3, averaging ~15% slowdown compared to "speed-of-light". With no geometry in view, performance matches the baseline (no primitive ID overhead, unlike mode 3).
This mode uses VK_AMD_shader_explicit_vertex_parameter extension. This approach is described in GPUOpen blog post.
Vertex shader writes gl_VertexIndex into 2 separate outputs. Pixel shader accesses those parameters through flat and __explicitInterpAMD interpolators to establish the order of native barycentrics available through gl_BaryCoordSmoothAMD.
Performance matches the "speed-of-light". There is no measurable overhead from accessing barycentrics with this method.
Geometry-heavy scene used for testing is San Miguel 2.0 from http://casual-effects.com/data.
In more balanced scenes, performance delta between different methods can be much less dramatic. In any case, Mode 4 seems to be the most preferable one on NVIDIA and Mode 5 is obviously hard to compete with on AMD.
Mode 3 may be the best cross-platform mechanism at this point, though it would be interesting to implement a way to avoid gl_PrimitiveID overhead.