AMD Bulldozer Review, AVX Performance using SONAR benchmark

Its pretty cool that Cakewalks AVX optimization work was featured in this review of AMD’s bulldozer from Tom’s Hardware. For those unfamiliar with Tom’s Hardware, the site is the holy grail of hardware reviews and benchmarks. Their depth of knowledge and coverage of the state of the art in computer hardware is unparalleled.

A few months ago I was contacted by Chris Angelini, the Editor in chief at Tom’s Hardware. He had come across a white paper that I co authored with Intel, featuring the AVX optimization’s in SONAR X1, and was very interested in knowing more about our experiences with AVX in relation to Bulldozer vs Intel’s SandyBrige.
Continue reading “AMD Bulldozer Review, AVX Performance using SONAR benchmark”

Utilizing Intel® AVX with Cakewalk SONAR X1

This is a whitepaper (Utilizing Intel® AVX with Cakewalk SONAR X1) which I co-authored with Intel engineer Rajshree Chabukswar, highlighting the advantages of optimizing for the Intel AVX chipset, with a focus on digital audio processing in a modern DAW like SONAR X1.

We’re excited with our synergetic relationship with Intel, which allows us to take  advantage of their bleeding edge technology in ways that directly beneft our users, allowing them to squeeze the most power out of their systems. While the paper is technical and requires an understanding of some low level programming, it also offers insight into the nuts and bolts of whats involved in optimizations for Intel CPU architectures in a modern DAW.

The paper features a real world case study of SONAR X1 code that was optimized in to take advantage of the benefits of the 256 bit AVX instruction set. If you have an Intel CPU from the Sandy Bridge processor family, it supports AVX and SONAR X1 will take advantage of it.
(While AVX is an Intel instruction set, it has also been adopted by AMD will be available in their upcoming Bulldozer processors. )

Code which is optimized for AVX  vectorization capabilities can work with 256-bit vectors, allowing working on 8 32-bit floating point values per iteration. In other words, this is twice the data throughput of earlier SSE instruction set! While this doesn’t necessarily translate to twice as fast, it is a huge step up in performance in many cases as the white paper illustrates.

The first step in any optimization task is what is referred to as “hotspot analysis”. In this phase you identify the bottlenecks in the code or that would benefit most from AVX optimization. We did analysis running through stress test projects and workflows that showed some classic hotspots. Once these were identified, the code was AVX optimized using the new AVX intrinsics available in Visual Studio 2010.

Click below to read the paper or download the PDF from Intel’s site:
Utilizing Intel® AVX with Cakewalk SONAR X1

[ Additional credits to Keith Albright and Bob Currie from Cakewalk, for hotspot analysis, development, and troubleshooting ]