Brian | b87c1ab | 2008-01-24 09:15:31 -0700 | [diff] [blame] | 1 | <HTML> |
| 2 | |
| 3 | <TITLE>Cell Driver</TITLE> |
| 4 | |
| 5 | <link rel="stylesheet" type="text/css" href="mesa.css"></head> |
| 6 | |
| 7 | <BODY> |
| 8 | |
Brian Paul | 2c0ce92 | 2009-01-08 16:15:31 -0700 | [diff] [blame] | 9 | <H1>Mesa/Gallium Cell Driver</H1> |
Brian | b87c1ab | 2008-01-24 09:15:31 -0700 | [diff] [blame] | 10 | |
| 11 | <p> |
| 12 | The Mesa |
| 13 | <a href="http://en.wikipedia.org/wiki/Cell_%28microprocessor%29" target="_parent">Cell</a> |
| 14 | driver is part of the |
| 15 | <a href="http://www.tungstengraphics.com/wiki/index.php/Gallium3D" target="_parent">Gallium3D</a> |
| 16 | architecture. |
| 17 | </p> |
| 18 | |
| 19 | <p> |
| 20 | <a href="http://www.tungstengraphics.com/" target="_parent">Tungsten Graphics</a> |
| 21 | is leading the project. |
| 22 | Two phases are planned. |
| 23 | First, to implement the framework for parallel rasterization using the Cell |
| 24 | SPEs, including texture mapping. |
| 25 | Second, to implement a full-featured OpenGL driver with support for GLSL, etc. |
Brian Paul | 2c0ce92 | 2009-01-08 16:15:31 -0700 | [diff] [blame] | 26 | The second phase is now underway. |
Brian | b87c1ab | 2008-01-24 09:15:31 -0700 | [diff] [blame] | 27 | </p> |
| 28 | |
| 29 | |
| 30 | <H2>Source Code</H2> |
| 31 | |
| 32 | <p> |
Brian Paul | 9c101c4 | 2009-02-10 16:56:58 -0700 | [diff] [blame] | 33 | The latest Cell driver source code is on the master branch of the Mesa |
| 34 | git repository. |
Brian | 99a1260 | 2008-02-28 09:48:26 -0700 | [diff] [blame] | 35 | </p> |
Brian | 99a1260 | 2008-02-28 09:48:26 -0700 | [diff] [blame] | 36 | <p> |
Brian | b87c1ab | 2008-01-24 09:15:31 -0700 | [diff] [blame] | 37 | To build the driver you'll need the IBM Cell SDK (version 2.1 or 3.0). |
| 38 | To use the driver you'll need a Cell system, such as a PS3 running Linux, |
| 39 | or the Cell Simulator (untested, though). |
| 40 | </p> |
| 41 | |
| 42 | <p> |
Brian Paul | 2c0ce92 | 2009-01-08 16:15:31 -0700 | [diff] [blame] | 43 | If using Cell SDK 2.1, see the configs/linux-cell file for some |
| 44 | special changes. |
Brian | cabcee9 | 2008-02-27 18:01:37 -0700 | [diff] [blame] | 45 | </p> |
| 46 | |
| 47 | <p> |
Brian | b87c1ab | 2008-01-24 09:15:31 -0700 | [diff] [blame] | 48 | To compile the code, run <code>make linux-cell</code>. |
Brian Paul | 36c1e75 | 2009-02-10 19:36:52 -0700 | [diff] [blame^] | 49 | Or to build in debug mode, run <code>make linux-cell-debug</code>. |
Brian | b87c1ab | 2008-01-24 09:15:31 -0700 | [diff] [blame] | 50 | </p> |
| 51 | |
| 52 | <p> |
Brian Paul | 36c1e75 | 2009-02-10 19:36:52 -0700 | [diff] [blame^] | 53 | To use the library, make sure your current directory is the top of the |
| 54 | Mesa tree, then set <code>LD_LIBRARY_PATH</code> like this: |
| 55 | <pre> |
| 56 | export LD_LIBRARY_PATH=$PWD/lib/gallium:$PWD/lib/ |
| 57 | </pre> |
Brian | b87c1ab | 2008-01-24 09:15:31 -0700 | [diff] [blame] | 58 | |
| 59 | <p> |
Brian Paul | 36c1e75 | 2009-02-10 19:36:52 -0700 | [diff] [blame^] | 60 | Verify that the Cell driver is being used by running |
| 61 | <code>progs/xdemos/glxinfo</code> and looking for: |
Brian | b87c1ab | 2008-01-24 09:15:31 -0700 | [diff] [blame] | 62 | <pre> |
Brian Paul | 2c0ce92 | 2009-01-08 16:15:31 -0700 | [diff] [blame] | 63 | OpenGL renderer string: Gallium 0.2, Cell on Xlib |
Brian | b87c1ab | 2008-01-24 09:15:31 -0700 | [diff] [blame] | 64 | </pre> |
| 65 | |
| 66 | |
Brian | cef71f5 | 2008-02-27 18:13:21 -0700 | [diff] [blame] | 67 | <H2>Driver Implementation Summary</H2> |
| 68 | |
| 69 | <p> |
| 70 | Rasterization is parallelized across the SPUs in a tiled-based manner. |
| 71 | Batches of transformed triangles are sent to the SPUs (actually, pulled by from |
| 72 | main memory by the SPUs). |
| 73 | Each SPU loops over a set of 32x32-pixel screen tiles, rendering the triangles |
| 74 | into each tile. |
| 75 | Because of the limited SPU memory, framebuffer tiles are paged in/out of |
| 76 | SPU local store as needed. |
| 77 | Similarly, textures are tiled and brought into local store as needed. |
| 78 | </p> |
| 79 | |
Brian | cef71f5 | 2008-02-27 18:13:21 -0700 | [diff] [blame] | 80 | |
Brian | b87c1ab | 2008-01-24 09:15:31 -0700 | [diff] [blame] | 81 | <H2>Status</H2> |
| 82 | |
| 83 | <p> |
Brian Paul | 2c0ce92 | 2009-01-08 16:15:31 -0700 | [diff] [blame] | 84 | As of October 2008, the driver runs quite a few OpenGL demos. |
| 85 | Features that work include: |
Brian | b87c1ab | 2008-01-24 09:15:31 -0700 | [diff] [blame] | 86 | </p> |
Brian Paul | 2c0ce92 | 2009-01-08 16:15:31 -0700 | [diff] [blame] | 87 | <ul> |
| 88 | <li>Point/line/triangle rendering, glDrawPixels |
| 89 | <li>2D, NPOT and cube texture maps with nearest/linear/mipmap filtering |
| 90 | <li>Dynamic SPU code generation for fragment shaders, but not complete |
| 91 | <li>Dynamic SPU code generation for fragment ops (blend, Z-test, etc), but not complete |
| 92 | <li>Dynamic PPU/PPC code generation for vertex shaders, but not complete |
| 93 | </ul> |
| 94 | <p> |
| 95 | Performance has recently improved with the addition of PPC code generation |
| 96 | for vertex shaders, but the code quality isn't too great yet. |
| 97 | </p> |
| 98 | <p> |
| 99 | Another bottleneck is SwapBuffers. It may be the limiting factor for |
| 100 | many simple GL tests. |
| 101 | </p> |
| 102 | |
| 103 | |
| 104 | |
| 105 | <H2>Debug Options</H2> |
| 106 | |
| 107 | <p> |
| 108 | The CELL_DEBUG env var can be set to a comma-separated list of one or |
| 109 | more of the following debug options: |
| 110 | </p> |
| 111 | <ul> |
| 112 | <li><b>checker</b> - use a different background clear color for each SPU. |
| 113 | This lets you see which SPU is rendering which screen tiles. |
| 114 | <li><b>sync</b> - wait/synchronize after each DMA transfer |
| 115 | <li><b>asm</b> - print generated SPU assembly code to stdout |
| 116 | <li><b>fragops</b> - emit fragment ops debug messages |
| 117 | <li><b>fragopfallback</b> - don't use codegen for fragment ops |
| 118 | <li><b>cmd</b> - print SPU commands as their received |
| 119 | <li><b>cache</b> - print texture cache statistics when program exits |
| 120 | </ul> |
| 121 | <p> |
| 122 | Note that some of these options may only work for linux-cell-debug builds. |
| 123 | </p> |
| 124 | |
| 125 | <p> |
| 126 | If the GALLIUM_NOPPC env var is set, PPC code generation will not be used |
| 127 | and vertex shaders will be run with the TGSI interpreter. |
| 128 | </p> |
| 129 | <p> |
| 130 | If the GALLIUM_NOCELL env var is set, the softpipe driver will be used |
| 131 | intead of the Cell driver. |
| 132 | This is useful for comparison/validation. |
| 133 | </p> |
| 134 | |
Brian | b87c1ab | 2008-01-24 09:15:31 -0700 | [diff] [blame] | 135 | |
| 136 | |
| 137 | <H2>Contributing</H2> |
| 138 | |
| 139 | <p> |
| 140 | If you're interested in contributing to the effort, familiarize yourself |
| 141 | with the code, join the <a href="lists.html">mesa3d-dev mailing list</a>, |
| 142 | and describe what you'd like to do. |
| 143 | </p> |
| 144 | |
| 145 | |
| 146 | </BODY> |
| 147 | </HTML> |