Integrate new preset parser and rewrite/refactor preset rendering.

[kblaschke]

First of all, big sorry for creating such a huge PR.

Sadly, just replacing the preset and expression parser with the newly written one wasn't exactly easy, as the old code was highly integrated in basically all major effect classes. It worked on a few assumptions no longer compatible with the new expression compiler, so I had to reorganize code. This lead to the decision to reorganize the whole Milkdrop preset renderer into an overall better structure, removing the "ping pong" way of how the preset parts were rendered previously and replacing it with a more straightforward approach, effectively containing the rendering of a preset into a single function call.

Having to restructure classes, I've also went fully down into the rabbit hole and checked every rendering step for Milkdrop compatibility, fixed issues and added missing code. Also added framebuffer usage to the rendering process for various reasons listed below.

Changes

Here's a list of changes, improvements and other notable things contained in this PR, in no particular order:

• Removed all existing preset file parser and expression-related code from libprojectM and replaced it with a new, simpler file parser class and the newly written projectM-eval expression compiler library.
• Renamed target "MilkdropPresetFactory" to "MilkdropPreset".
• Moved all Milkdrop-specific rendering code to "MilkdropPreset".
• Moved and refactored code from "Renderer" and "ShaderEngine" classes into other places, e,g, new classes. This includes blur textures, comp/warp shader code and some other math for generating the composite mesh.
• Added lazy texture loading to TextureManager, so it doesn't load all images into VRAM on startup.
• Added unloading logic for unused textures, very close to the code used in Milkdrop to save additional VRAM.
• Moved warp mesh texture coordinate calculation code to a vertex shader, relieving the CPU of some work.
• Properly implemented the motion vector grid effect, was only a simple grid of static points before. Now it properly reverse-propagates the warp mesh "motion" of the last frame and draws lines as it should. In contrast to Milkdrop, the actual motion vector calculation is completely done on the GPU inside the vertex shader, using the U/V coordinate color attachment of the previous frame.
• Properly implemented the "Darken Center" effect.
• Consolidated code for classic 1.x filter effects like "Darken", "Invert" and "Solarize".
• Made preset rendering per-frame aware of external parameter changes like screen resolution and warp mesh size, only reallocating the textures and meshes which actually have changed since the last frame. This makes the process of resizing the window way faster.
• Blur textures are now rendered in their own (smaller) framebuffer instead of directly drawing into the main image texture.

Work left for later

Will fix or add the following features and bugs later, so we can get this huge changeset merged ASAP:

• Fix "upside-down" rendering issues and, if possible, pass the correct U/V coordinates to the HLSL shaders to prevent math errors.
• Implement random texture selection.
• Implement simple smooth preset blending using shaders.
• Audio-related issues, especially the spectrum values, which are off the charts, causing a multitude of rendering issues in many presets.
• Proper framebuffer/render-to-texture support for rendering a frame.
• User-configurable blending shaders and options.

[kblaschke]

Emscripten version builds again, a crash was fixed and cleanups are done so far. I'd say the branch is in a good enough state to be merged to master, even if there's some work left to get it into a releasable state.

Please test extensively, review the changes and give your opinions.

[JohannesKauffmann]

First of all, big sorry for creating such a huge PR.

Forgive me if this may come across as bikeshedding, but some CMake parts of bf344b6b49adfcd3b5488759edd9622bfe9f96cd (the PROJECTM_STATIC_DEFINE parts) could be extracted to make this PR a bit smaller :)

[kblaschke]

Replaced the color masking with attaching/detaching the U/V texture for use with motion vectors, as GLES only supports per-buffer color masking with v3.2.

[kblaschke]

First of all, big sorry for creating such a huge PR.

Forgive me if this may come across as bikeshedding, but some CMake parts of bf344b6 (the PROJECTM_STATIC_DEFINE parts) could be extracted to make this PR a bit smaller :)

What exactly are you proposing there? Can't remember that I've done anything with the linking stuff, besides additing the evaluation library. Also unsure how this would make the PR substantially smaller, as the overwhelming part was rewriting code in the Renderer and MilkdropPreset[Factory] targets.

If we can make the static linking part easier, that's definitely a good thing to do. I'd do that separately though, as this PR already changes enough things, or better said, too many to even properly review. Could you create an issue please, detailing the proposed changes? This way I won't forget about that!

[kblaschke]

I'd personally say this changeset is as good as it gets without adding even more stuff to it, so I'd like anyone to pick up the branch, build and test it for stability. I've not seen any serious issues besides a few remaining rendering problems noted in the description, but these will be addressed later. Overall, it's a huge improvement over the current master, visually and performance-wise.

[kblaschke]

FYI, last push was a rebase on latest master to prepare for merging.

[JohannesKauffmann]

What exactly are you proposing there?

The commit A few Windows build fixes from this PR changed both CMakeLists.txt such that PROJECTM_STATIC_DEFINE is defined for a wider range/scope of targets. My proposal was to move these changes specifically to a separate commit/PR, as these CMakeLists.txt changes didn't seem related to the new preset parser/rendering rewrite, but would apply more globally. It'd make this PR a bit smaller (not much smaller admittedly).

Not sure it is worth changing now, however. PR looks good either way.

[kblaschke]

Ah, I see. Yes, I had to do this because otherwise the class definitions in the PCM class headers were different between the compile units, which the MSVC compiler doesn't like (different declspec attributes), causing linker errors. The MilkdropPreset target uses the PCM class to retrieve the frame audio data, and I didn't also want to refactor this class structure as it needs more attention later due to the broken/high spectrum and beat detection values. Could've done this in an earlier PR, but it's really just a small fix for a typo and adding missing defines to other targets. Proper cleanup will be done in issue #708.

[kblaschke]

Merging now, as the current branch state is stable enough for most uses. Emscripten might still have some issues left on top of the other things on the to-do list, so @Blaquewithaq please create an issue if there are any problems running on Emscripten, I'll tackle these in a later fix.