Auto-Vectorizing LLVMpipe For Performance Gains

Here's a hint that may allow for some notable performance gains out of the Gallium3D LLVMpipe driver for multi-core systems.

Recently I wrote about the Gallium3D LLVMpipe Driver On LLVM 3.1. There wasn't really any major performance changes to note for this Gallium3D "software rasterizer" driver that uses the CPU and not the GPU as a means of fallback support, but this Sunday I received a message from Hal Finkel. This developer from the Argonne National Laboratory wrote the new basic-block auto-vectorization pass for LLVM 3.1. Below is the message he sent over to me regarding this work and LLVMpipe.

Michael,

I looked through your benchmark results on "Gallium3D LLVMpipe Driver On LLVM 3.1", and, as author of LLVM's new basic-block autovectorization pass, I wonder if turning on autovectorization would improve the 3.1 results.

If you're curious about the autovectorization pass, you can see my presentation at the developer's conference earlier this year: http://llvm.org/devmtg/2012-04-12/

It seems like llvmpipe uses its own pass manager, which is probably not a bad thing, but it is either intentionally minimalistic, or is pretty out of date (or both). There are many other optimization passes available (see PassManagerBuilder::populateModulePassManager in lib/Transforms/IPO/PassManagerBuilder.cpp to get an idea of what is available). It looks like the current pass manager used by llvmpipe is in create_pass_manager in src/gallium/auxiliary/gallivm/lp_bld_init.c and it adds only a few optimization passes.

To add autovectorization to llvmpipe, I think adding something like this after the call to LLVMAddGVNPass(gallivm->passmgr); will work:

#if HAVE_LLVM >= 0x301
LLVMAddBBVectorizePass(gallivm->passmgr);
/* BBVectorize depends on other passes for cleanup... */
LLVMAddInstructionCombiningPass(gallivm->passmgr);
LLVMAddEarlyCSEPass(gallivm->passmgr);
#endif

I've never really used the C API, and I'm not really sure how to pass tuning parameters through (and some tuning might be helpful or necessary here -- integration between the optimization passes and the backends is not as good yet as it should be), but this would be a start.

Thanks in advance,
Hal

P.S. Please feel free to forward or repost this message elsewhere if you feel that would be helpful or more appropriate.

I'm still out of the country on this nearly month-long business trip, but hopefully a few Phoronix readers would be interested in poking around with making these code changes as it could yield some interesting performance gains for this promising GPU fallback driver. LLVMpipe is now used as the fallback for Ubuntu and Fedora when no hardware GPU driver is available.