See the results in http://jsperf.com/dommatrix-perf/7 for reasons. Also, this should pack things more tightly, at the very least with bhackett's work on optimizing typed array memory usage in bug 1058340.

The JS perf results don't have a test case that match our implementation. I'll experiment, but I can't imagine why Float arrays would necessarily be faster.

As it turns out, normal JS Arrays seem fastest by quite a margin once the test is fixed to only create the temporary arrays once and reuse them: http://jsperf.com/dommatrix-perf/9 I also added a simple JSMatrix, which is about as fast as the float array version, but a lot slower than the JS Array version.

Looks like Float64Arrays are actually faster than Float32Arrays. Here is a PR that changes both the GFX and Flash Matrix class. https://github.com/mozilla/shumway/pull/1706 Till, what do you think about creating matrix pools so we can reuse them.