Optimizing Marvell Graphics With iwMMXt

Thanks in large part to iwMMXt fixes/improvements found in GCC 4.8, the ARM Marvell graphics performance will be much-improved, which will benefit the OLPC XO-1.75.

Matt Turner, the student developer who previously worked on a 3Dlabs KMS driver, began working on optimizing the (Marvell) ARM graphics performance last summer to benefit the One Laptop Per Child XO-1.75 notebook.

As outlined today on Matt's blog, "I started working with Chris and Jon Nettleton on improving the graphics performance of the ARM-based XO 1.75 laptop. The graphics drivers were in a state of flux, and in a number of cases the Sugar interface felt noticeably slower than on the VIA-powered XO 1.5. We wanted to know why it was slower and how to quantitatively measure graphics performance of real-world applications."

What Matt Turner ended up focusing upon for the XO-1.75 graphics optimizations were allowing Pixman's MMX-optimized compositing functions to work on the OLPC ARM hardware. Marvell's Sheeva ARM SoC as used in the XO-1.75 is based upon Intel's XScale ARM architecture that Marvell acquired.

Thanks to the Intel lineage, the ARM SoC features the iwMMXt SIMD instruction set, which is effectively compatible with Intel x86 MMX. In one hit, x86/MMX and ARM/iwMMXt can be targeted. With Pixman already having the MMX optimizations, Matt focused on making them work for the OLPC ARM target with iwMMXt.

Unfortunately, the GCC support for iwMMXt had regressed with it not being tested since the Intel XScale days. As a result, Turner had to fix the GCC support, which he patched, Marvell also provided GCC improvements for iwMMXt scheduling and support. Therefore with GCC 4.8 will be improved iwMMXt code generation support. Thanks to this, the graphics performance goes up to benefit the OLPC XO-1.75. Some positive benchmark results are shared in Matt's aforementioned blog post.

There's also an unfortunate problem in this situation, illustrating another problem with American universities. As said on Matt's blog about his GCC patch to fix iwMMXt support, "I submitted a small patch to fix the problem, but my school (NC State) took eight months to acknowledge that they don't own my work, and by that time Marvell had contributed a five-patch series to significantly improve iwMMXt scheduling and support." So it took over a half-year for North Carolina State University to let Matt submit a patch to the GPLv3-licensed GNU Compiler Collection. (Had this not happened, his patch could have already landed with GCC 4.7.)

With having around an OLPC XO-1.75, it looks like it's time to run some new benchmarks.