Room for more render performance

[MrBrixican]

There is room for much more performance in respect to degenerate and standard cases where Graphics.Submit must be called frequently:

• Drawing different textures in alternating pattern
• Drawing many models/meshes

The main detractor from performance in those scenarios is redundant calls to OpenGL methods.

• FosterDraw_OpenGL calls a lot of methods to set values that may already be set correctly (glUseProgram, glBindFramebuffer, glTexParameteri, glEnable, etc).
• Unfortunately OpenGL drivers don't make a strong effort (at least on my machine) to prevent redundant calls from having nontrivial overhead.

Luckily, most redundant calls are pretty simple to catch, as we can just keep track of what we set/used previously and ignore the call if the same parameters will be used. This method was used in the original Foster to great effect.

I have personally tested this and received a 4x speedup in the Froggymark sample (using built in Batcher) when drawing alternating textures (causing a draw call each). I (very roughly) implemented the following adjustments:

• glUseProgram calls in FosterDraw_OpenGL and FosterShaderSetUniform_OpenGL are ignored if the shader has already been set
• glTexParameteri calls are ignored if the texture has already been set to use those parameters
• ignore calls relating to the following global state if they are using previous values:
  • frame buffer
  • blend state
  • depth state
  • cull state
  • viewport
  • scissor
• removed glBindVertexArray(0) from end and avoid redundantly calling fgl.glBindVertexArray(mesh->id)

There are a few more possible tweaks, but they will require more involved solutions:

• avoid glActiveTexture, glBindTexture, and possibly glUniform1iv calls if textures/slots haven't changed
• avoid setting uniforms if their value is unchanged from last set (this may not be worth it, needs testing)

[NoelFB]

This all sounds good to me. I think at some point I had them check for some these more often but removed it because I assumed drivers would just not do redundant work. But yeah, I agree it's better to just check since, as you've pointed out, not all are nice about it haha.

If you've already made changes to the rendering in the C library I'd be happy to merge that in. I can look into the texture/uniform binding as well.

[NoelFB]

I've pushed changes that handle most of these, with the exception of uniforms/texture binding. I'll look into those more later; might make a new PR for it. But the majority of cases should now be resolved!

[MrBrixican]

Looks excellent! Great work! Way cleaner than the slapdash solution I had, haha.

[NoelFB]

Alright I've got improved a lot of glActiveTexture/glBindTexture/updating Sampler states now. The only thing that could potentially also be improved is caching Shader Uniform values but I think it's not worth the trouble, or should maybe happen from the C# side.

https://github.com/NoelFB/Foster/commit/36088f6be8db594d439d7e5f049ac748662cdfac

Will close this!