mirror of
https://github.com/libretro/RetroArch.git
synced 2026-11-21 12:29:43 -10:00
The six XMB background effects (ribbon, simple ribbon, simple snow, snow, bokeh, snowflake) were baked into every video driver as fixed shader slots, VIDEO_SHADER_MENU to VIDEO_SHADER_MENU_6. Each driver carried its own copies of the sources, built them whether or not a menu used them, and had a case per slot in its draw code. Now the menu owns them. XMB keeps one table of the effects (menu/drivers/xmb_effects.h) with each effect's sources for every backend - GLSL (legacy and core), SPIR-V, HLSL SM3 and SM4, Metal and GX2 - and how it draws (a mesh or a quad, its blend, whether it takes the ribbon data, whether it flips Y). XMB installs them through gfx_display_effect_install() and draws one by handing the draw its effect. The drivers have a single generic path: an effect_install callback that makes a program from the backend's source, and a draw that goes by the effect's flags. VIDEO_SHADER_MENU..MENU_6 are gone, and so is the shader backend's init_menu_shaders. Installs are kept in a small lock-free table in gfx_display. The programs are made on the drawing thread when the first effect is drawn, all at once, so nothing is compiled while an effect is on screen. When the video driver goes away the programs are dropped and made again under the next one. What this removes or speeds up: - Menus that draw no effect (Ozone, MaterialUI, RGUI, XMB with the effect off) build no effect programs at all. - GL: effect draws use their own program with every uniform and attribute location looked up once at compile. Before, every effect draw looked up "time", "Ribbon" and "OutputSize" by name through the shader backend, with prefix retries, each frame. The quads draw from the full-screen mesh's buffer, as the ribbon draws from its own: no vertices are uploaded per draw, and nothing is read from client memory, which a core profile does not allow. - GL core: effects no longer switch to and from the shader chain's program through the chain path; the chain program is put back only after an effect was drawn. - D3D9 (HLSL and Cg): the constant registers/parameters are found once when an effect is compiled, not on every draw. - Vulkan: display.pipelines shrinks from 9 to 3 (plastic stays at its own slot); the effect pipelines are made from the effect's own SPIR-V and remade with the passes. - D3D12: effect bytecode is compiled once; only the PSOs are remade when the render target format changes. - The seven Cg effect sources (gl_shaders/pipeline_*.cg.h) are gone. Unchanged: the plastic effect and the stock shaders keep their ids; fonts and other menus use them, and moving them buys nothing. Behaviour notes: - GL (gl2) effects need GLSL; a Cg-only GL build no longer draws them. - GL core draws effects only when built with slang (they come from the SPIR-V); without it the background is plain. - D3D9 HLSL no longer falls back to stock shading when an effect fails to compile; the effect is skipped instead. Tests: - samples/gfx/menu_effects (new): every effect carries a source for every backend, its flags hold together, and each source names the inputs and uniforms the drivers set for those flags. Each GLSL program is made through shader_glsl.c as gl2 makes it - GL legacy and core profile, GLES 3 and 2, on Mesa's EGL device - its locations checked and drawn once into a framebuffer. The SPIR-V goes through spirv-val. Cross-built with mingw-w64, the HLSL is compiled under Wine as D3D9 (SM3), D3D10/11 (SM4) and D3D12 (SM5) compile it; what Wine's compiler does not yet support (SM1 branches and loops, atan) is reported as skipped. Run in samples-gfx and samples-gfx-d3d. - samples/gfx/gfx_display_mesh covers install-once, re-making after the driver is lost, and no draw without an install. - samples/gfx/d3d9_failed_lock's ribbon test follows the new ids.