Setting the size of each descriptor pool to one
was a hack introduced to make the code work on android.
Changing the code to now allocate all descriptor sets at once
should fix the previous crash.
Not tested yet.
Before this patch, sometimes rendering a texture
on an android device would result in random things
appearing in the canvas, apart from the stuff
that has been drawn properly.
With these changes this does not seem to appear anymore.
Assigning to and accessing extra shader variables can increase VGPR / wave-variant register usage in shaders, which is often a limiting factor for performance in non-trivial shaders. Accessing the uniforms directly should bypass that.
Since these objects basically get used every frame
we shouldn't store them in a hashmap with potentially
worse performance.
Instead storing them in member variables should be better.
It's not sure if this extension is supported everywhere
This commit uses normal descriptor pools, sets and writes,
which are all availabel in base vulkan.
Until now we only considered the builtin uniform variables.
This commit attempts to make the code in the vulkan Shader
implementation more general, as to allow for custom uniforms.
This is not possible yet, but should be easier to implement now.
Until now we've used a command buffer for every swap chain image,
however it makes more sense to just have a command buffer for every
frame in flight, since we won't be ever recording more than that.
This makes it more in line with the data transfer command buffers,
which are also just created per frame in flight.
This should make more sense.
There is no real effect yet when rendering, however debugging the
application with renderdoc reveals that the mipmap generation is
already correct. There's probably something weird going on with the
sampling. I'll have to look into that next.
niki
slime
nikeinikei