The Good & Bad For LLVMpipe With Intel's GLSL2 Compiler

So what good is there from this new OpenGL Shading Language compiler for Mesa? In some cases, it seems to be more efficient and for LLVMpipe results in a lower CPU usage. Below is the CPU usage charted by the Phoronix Test Suite when OpenArena was run at 1600 x 900.

The CPU load on the Intel Core i7 CPU was measurably less with OpenArena (and Tremulous) when using the new shader compiler. The average CPU usage was 74% usage when running with the current mainline Mesa code versus 68% when merging in the new GLSL compiler.

When monitoring World of Padman, however, there was only a 1% difference with this very fast mobile CPU in the Lenovo ThinkPad W510 when using the new shader compiler.

While we were not expecting the Mesa GLSL compiler or Intel's new solution to cause any major changes for the in-game frame-rate performance of the tests, when running this code that's expected to be merged into the mainline Mesa code-base later this month, the frame-rate dropped when using Gallium3D's LLVMpipe in some tests (Tremulous and World of Padman) and it broke the Urban Terror game. The only good news we have found thus far from pairing the new shader compiler with the LLVM-using driver is that the CPU usage is lower in some areas, but due to bottlenecks elsewhere in the driver or Gallium3D, this is not resulting in any performance gains at this point.