Implement shader pipelines #1587
Conversation
Also make sure to display the full shader name rather than just the main shader.
Previously, the GLSL shader building was done in the following manner: 1. When the renderer is started, some of the shader source was initialised/processed. This included the main shader text, #insert's, and material system post-processing. 2. When `buildAll()` was called, all permutations with empty deform shaders were built. The deform shaders themselves are loaded either at the start (for the default, empty one), or when the map itself is being loaded. These permutations included adding macros to shaders in the form of `#ifndef #define #endif`, and attaching the default deform shader. When possible, shader program binaries were saved/loaded to/from disk. 3. UI shader and non-0 deform shaders were built on-demand. There were multiple issues with that system: 1. Shader caches were unreliable. This is likely because the checksum was calculated *before* macros and such were added to the shader text. 2. Every shader permutation was always compiled and linked every time. 3. Shader programs with non-0 deformVertexes were never cached. This commit changes the underlying system to use an `std::vector` of shader program descriptors, and another one for shader descriptors. The initialisation/building of shaders is now more generic for each type, and as a result each unique shader is only compiled once. The permutations are still linked each time (unless downloaded from cache), but this change lays the foundation to easily add shader pipelines, which would solve this problem. Shader cache binaries are now identified by each shader compiled into the corresponding shader program, which allows caching non-0 deformVertexes too. This also fixes the incorrect #line numbers emitted by #insert due to macros and such being added after #insert is processed. Also adds more information into how many shader/shader programs were built and how much time it took (there's currently a small discrepancy, because the output is shown at the end of `BuildAll()`, but is gathered every time `BuildPermutation()` or `InitShader()` is called).
This would be much easier if we drop support for hardware straight out of Pleistocene. |
|
It looks like Mesa supports |
It doesn't make sense for GL1 cards, since they don't support the programmable pipeline anyway. |
Requires #1586
This further improves the shader building by using GL_ARB_separate_shader_objects when it is available. This extension allows linking different shader stages separately, thereby decreasing the total amount of linking required.
As an example, on map
station15with all the shader stuff except liquid mapping and static reflections enabled, it went from 433 permutations/~866 compiled shaders/433 linked programs to 433 permutations/287 compiled shaders/286 linked programs, and cut the total amount of time it took to build shaders almost in half.I'd had to add some ugly workarounds to make it work with both shader pipelines and regular shader programs due to how our GL uniforms work. Because of that this is currently draft, as I'm thinking of moving all these uniforms to a UBO when shader pipelines are available.