GCC 4.6 Compiler Performance With AVX On Sandy Bridge

Like the HMMer results, with GraphicsMagick, any Intel Core 2 / Core i7 architecture/tuning optimizations really hurt the performance of this graphics / image processing application derived from ImageMagick.

The GraphicsMagick performance drop with the flags was not limited to a sub-set of the results, but all of the GraphicsMagick tests. With the image resizing operation, we also see that the tested GCC 4.6 snapshot has regressed rather negatively compared to GCC 4.5.2, which ran the best of any compiler tested.

With Gcrypt, the flags once again hurt the performance compared to the vanilla builds.

For those that have been asking about support for the Advanced Vector Extensions and its performance under Linux, these are the current findings, which are similar to the results from other test profiles we also tested. In some workloads, we have found when tapping the AVX capabilities of the Sandy Bridge processor the performance has improved a lot, including when using GCC's AVX floating-point math library, but in some, this actually led to a performance regression.

While the AVX instruction set has been known for years, it was just last month that the support began appearing in Intel's CPUs via the Sandy Bridge family and it will not be until later in the year that we find support for this ISA on the AMD side. Surely, in the coming months once AVX-supportive processors are more widespread, we will find more software being optimized under Linux to take advantage of the Advanced Vector Extensions support at the application level. For now though at least the Linux kernel and GCC compiler support for AVX is ready for the early-adopters. More tests shall come a few months down the road.