Hah this took a while, as there was no texture scaffolding in place at
all. Thus all this had to be added and tested for the first time:
* 2D textures
* 2D texture arrays
* Texture transformation uniforms
* Texture transformation UBOs
* Instanced texture offset
This also means that MeshVisualizer can be used to visualize arbitrary
(single-channel) integer textures now, not just render meshes with
object ID textures. Yay for feature parity!
Forgot that gl_VertexID includes the base offset in the multidraw
scenarios, so we need to take all vertices into account, not just the
largest view. The wraparound would cause nasty output differences among
drivers.
Mainly to have feature parity with Flat and Phong -- otherwise switching
to draw a wireframe on an instanced mesh would be too annoying. Also, if
we have multidraw there already, why not instancing as well.
Originally the uniform wasn't present with the assumption that users
could easily adjust color map offset to achieve the same effect. That
was however unnecessarily annoying and error-prone in cases where it's
essential to have the same object IDs from multiple draws have a
matching color, and it was complicating multidraw workflows as the color
map offset was not a part of per-draw data, but rather material data.
Do it always when Flag::TextureArrays is set, not inside handling of
some particular texture. Because that way it won't work when other
textures are added/tested.
Before the object ID was enabled and tested always, which may lead to
some error being undetectable. Plus this makes the test more flexible
for further additions.
In cases when specular highlights are not desired, results in 30%
speedup (on Intel) and ~25% speedup on AMD, compared to setting the
specular color to transparent black.
Testing was easy thanks to already having a ground truth image for this
case.
After several failed attempts to make UBO performance not suck on Intel
Mesa and Windows drivers, I ended up hiding the dynamic aspect under a
flag. That way it's still possible to get the proper perf in UBO
workflows that don't do light culling, and for workflows where light
culling matters the 2x slowdown might be still better than looping
through several extra lights that don't contribute anything.
"Luckily", thanks to the DRAW_COUNT=1 and MATERIAL_COUNT=1 optimizations
not everything blows up, so i don't need to skip absolutely everything,
unfortunately Phong lights are affected by this insane crapfest as well
so basically nothing from Phong UBO support is tested there. FFS.
So it's all having the same workflow. This one results in even more
saved UBO slots per-draw than in the case of Flat, and the slowdown on
Intel is as bad as expected.
While it's one additional indirection (that has an extra cost on Intel
GPUs apparently, like with Phong and MeshVisualizer and
DistanceFieldVector already), with the assumption that draws usually
share the material info it allows to cram more draws into the 16/64k UBO
limit as the per-draw data are now one vec4 smaller.
For the indirection overhead I can imagine adding a new flag which makes
material mapping implicit (materialId == drawId). That seems to put the
benchmark numbers back to the original speed. Same could be done for
other shaders.
These deliberately share the same binding (because there's very little
space), but the shader wasn't guarding that. Discovered completely by
accident when adding tests for "multidraw with all the things" -- Mesa
gives just a warning, but ANGLE straight out fails the shader
compilation, so better have an assert there.
Besides expanding the tested platform set and updating thresholds where
needed, it makes more sense to list what is tested than what is not,
because when I forget to update the list it looks like I tested while I
did not.
Interestingly, shaders that have indirect material references are about
2x slower on Intel. Not the Flat or Vector, which contain the full
material in the DrawUniform. Will probably need extra
Intel-specific optimizations (like avoiding the indirection if
MATERIAL_COUNT=1).
Took me a while (several years?) to figure out a way to benchmark this
without basically duplicating the testing effort and without new
variants being too hard to add.
Vladimír Vondruš