Google Engineer Breaks Down Steps for VP8 Optimization
Road map for VP8 codec improvements laid bare.
Google software engineer John Koleszar addressed the open source community regarding the VP8 codec and the steps needed to further optimize to codec within the WebM project.
Koleszar notes Scott LaVarnway's work in creating an x86 version of the quantizer and moved on to request a SIMD version of the ARNR temporal filtering code from the community. Koleszar also asked for newer extensions for the assembly code, as it currently only takes advantage of the SSE2 instruction set.
The last improvement Koleszar called for in the VP8 encoder was for someone to explore alternative motion search strategies, eventually hoping to decouple motion search entirely, leaving the motion field calculations to the graphics processors.
For the decoder, Koleszar highlights the work of Jeff Muizelaar, Johan Koenig, and Tim Terriberry. While he doesn't specifically ask for help on any one item as he did with the encoder, he does highlight some of the ongoing work. Terriberry is working had on the bool decoder, which is called multiple times per each bit in the input stream. Currently, the code uses a simple clamp on the innermost loops for checking and performs less frequent copies into a circular buffer. Terriberry's patch uses a more complex clamp and removes the circular buffer.
Meanwhile Muizelaar's work has combined IDCT and summation with the predicted block into a single function. Doing this reduces memory transfers and therefore reduces cache pollution. Koenig is implementing Muizellaar's work into ARM processors.
Speaking of embedded processors, Koleszar ended his post with a description of the work being done on not-desktop platforms. Fritz Koenig is working to optimize the VP8 codec for the Atom platform, quite a task considering the x86 assembly code for the codec was written for an out-of-order processor.
The Atom, of course, is in-order, so Koleszar and company are debating scheduling the code for Atom and then checking to see what performance issues arise on x86. Regardless, Koleszar notes that a lot of work lies ahead.
Finally, he spends some time on intrinsics and whether or he and his fellow programmers should use them when trying optimize the codec for multiple processors and platforms.
"If you have experience in dealing with a lot of assembly code across several similar-but-kinda-different platforms, these maintainability issues might be familiar to you. I hope you'll share your thoughts and experiences on the codec-devel mailing list," Koleszar said.
dgdgWorking with a lot of assembly code reminds me of another project.
They use Orc, it seems to help them a lot:
The company is collaborating with Google and Intel to use Kubernetes as an engine for Fuel
Customers can take a free test drive of SLES for HPC on the Azure Cloud
San Francisco-based chip company announces their first fully open source chip platform.
The whole distro gets rebuilt on glibc 2.3
Ubuntu Vendor tries to solve app packaging and distribution problem across distributions.
Founder of ownCloud launches the Nextcloud project.
Will The Machine change the way future programmers think about memory?
The new Torus distributed storage system is available under an open source license on GitHub
Juries decides Google’s use of Java APIs Was Fair Use