.TL libgraphics: Design and Implementation .DA .AU Rodrigo G. López rgl@antares-labs.eu .SH Introduction .LP .I Libgraphics is a 3D computer graphics library that provides a way to set up a scene, fill it up with a bunch of models (with their own meshes and materials), lights and cameras, and start taking pictures at the user request. It implements a fully concurrent retained mode software renderer, with support for vertex and fragment/pixel shaders written in C (not GPU ones, at least for now), a z-buffer, front- and back-face culling, textures and skyboxes, directional and punctual lights, tangent-space normal mapping, ??? .SH The renderer .LP The .I renderer is the core of the library. It follows a .B "retained mode" model, which means that the user won't get a picture until the entire scene has been rendered. Thanks to this we can also clear and swap the framebuffers without their intervention, they only need to concern themselves with shooting and “developing” a camera. .LP It's implemented as a tree of concurrent processes connected by .CW Channel s—as seen in .B "Figure 1" —, spawned with a call to .CW initgraphics , each representing a stage of the pipeline: .IP The .B renderer process, the root of the tree, waits on a .CW channel for a .CW Renderjob sent by another user process, specifying a scene, a camera and a shader table. It walks the scene and sends each .CW Entity individually to the entityproc. .IP The .B entityproc receives an entity and splits its geometry equitatively among the tilers, sending a batch for each of them to process. .IP Next, each .B tiler gets to work on their subset of the geometry (potentially in parallel)—see .B "Figure 2" . They walk the list of primitives, then for each of them apply the .B "vertex shader" to its vertices (which expects clip space coordinates in return), perform frustum culling and clipping, back-face culling, and then project them into the viewport (screen space). Following this step, they build a bounding box, used to allocate each primitive into a rasterization bucket, or .B tile , managed by one of the rasterizers; this is illustrated in .B "Figure 3" . If it spans multiple tiles, it will be copied and sent to each of them. .IP Finally, the .B rasterizers receive the primitive in screen space, slice it to fit their tile, and apply a rasterization routine based on its type (only .I points , .I lines and .I triangles are supported). For each of the pixels, a .B "depth test" is performed, discarding fragments that are further away. Then a .B "fragment shader" is applied and the result written to the framebuffer after blending. .PS .ps 7 circlerad = 0.3 moveht = 0.1 arrowhead = 9 box "Renderjob" arrow R: circle "renderer" arrow E: circle "entityproc" move Tiler: [ down T0: circle "tiler 1" move T1: circle "tiler 2" move Td: circle "…" move Tn: circle "tiler n" ] move Raster: [ down R0: circle "rasterizer 1" move R1: circle "rasterizer 2" move Rd: circle "…" move Rn: circle "rasterizer n" ] arrow from E to Tiler.T0 chop arrow from E to Tiler.T1 chop arrow from E to Tiler.Td chop arrow from E to Tiler.Tn chop arrow from Tiler.T0 to Raster.R0 chop arrow from Tiler.T0 to Raster.R1 chop arrow from Tiler.T0 to Raster.Rd chop arrow from Tiler.T0 to Raster.Rn chop arrow from Tiler.T1 to Raster.R0 chop arrow from Tiler.T1 to Raster.R1 chop arrow from Tiler.T1 to Raster.Rd chop arrow from Tiler.T1 to Raster.Rn chop .ps 10 .PE .B "Figure 1" : The rendering graph for a .B 2n processor machine. .SH Tile-based rendering .PP .PS .ps 7 Tiles: [ boxht = 0.2 boxwid = 1.25 down T0: box dashed "tile 1" T1: box dashed "tile 2" Td: box dashed "…" Tn: box dashed "tile n" ] box ht last [].ht+0.1 wid last [].wid+0.1 at last [] "Screen" rjust with .se at last [].nw - (0.1,0) Raster: [ moveht = 0.1 down R0: circle "rasterizer 1" move R1: circle "rasterizer 2" move Rd: circle "…" move Rn: circle "rasterizer n" ] with .w at Tiles.e + (0.5,0) line from Tiles.T0.e to Raster.R0.w line from Tiles.T1.e to Raster.R1.w line from Tiles.Td.e to Raster.Rd.w line from Tiles.Tn.e to Raster.Rn.w .ps 10 .PE .B "Figure 2" : Per tile rasterizers. .PS .ps 7 Tiles: [ boxht = 0.2 boxwid = 1.25 down T0: box dashed "1" T1: box dashed "2" Td: box dashed "…" Tn: box dashed "n" ] line from last [].w + (0.1,-0.05) to last [].n - (-0.1,0.25) line to last [].se - (0.3,-0.1) line to 1st line box ht last [].ht+0.1 wid last [].wid+0.1 at last [] "Screen" rjust with .se at last [].nw - (0.1,0) Raster: [ moveht = 0.1 down R0: circle "rasterizer 1" move R1: circle "rasterizer 2" move Rd: circle "…" move Rn: circle "rasterizer n" ] with .w at Tiles.e + (0.5,0) arrow from Tiles.T1.e to Raster.R1.w arrow from Tiles.Td.e to Raster.Rd.w arrow from Tiles.Tn.e to Raster.Rn.w .ps 10 .PE .B "Figure 3" : Raster task scheduling. .SH The scene .PP .PS .ps 7 boxwid = 0.5 boxht = 0.2 linewid = 0.1 lineht = 0.2 box "Scene" down; line from last box.s; right; line box "Entity" down; line from last box.s; right; line box "Model" down; line from last box.s; right; line box "Mesh" down; line from last box.s; right; line box "Primitive" down line from 2nd last line.s; line; right; line box "Material" .ps 10 .PE .SH Frames of reference .PP Frames are right-handed throughout every stage. .PS .ps 7 RFrame: [ pi = 3.1415926535 circle fill rad 0.01 at (0,0) "p" at last circle.c - (0.1,0) xa = -5*pi/180 arrow from (0,0) to (cos(xa),sin(xa)) "bx" at last arrow.end + (0.1,0) arrow from (0,0) to (0,1) "by" at last arrow.end - (0.1,0) za = -150*pi/180 arrow from (0,0) to (cos(za)+0.1,sin(za)+0.1) "bz" at last arrow.end - (0.1,0) ] .ps 10 .PE .B "Figure 4" : Example right-handed rframe. .SH Viewports .PP .PS .ps 7 View: [ boxwid = 3 boxht = 2 box with .nw at (-1,1) "framebuffer" at last box.s + (0,0.2) circle fill rad 0.01 at (-1,1) "p" at last circle.c - (0.1,0) arrow from (-1,1) to (-1,1) + (1,0) "bx" at last arrow.end + (0,0.1) arrow from (-1,1) to (-1,1) - (0,1) "by" at last arrow.end - (0.1,0) ] .ps 10 .PE .B "Figure 5" : Illustration of a 3:2 viewport.