Performance boost of JSBench using doubleToAtom_sse2

Haghighat, Mohammad R mohammad.r.haghighat at intel.com
Tue Jan 22 12:39:41 PST 2008


By changing AvmCore::doubleToAtom to call AvmCore::doubleToAtom_sse2
when the sse2 field is set (i.e., when sse2 is available), we observe a
good range of performance gains on the benchmarks in JSBench on Tamarin
Central. As a result of this optimization, seven out of the eleven
benchmarks gain performance by 2.5% up to +50%. MolDyn is the benchmark
that gets the 52% performance gain. One test (Series) shows 2.25
performance degradation that we're investigating.

               Time (sec)        Time (sec)              
Test           Baseline        DoubleToAtom_SSE2       %Perf Improvement
Moldyn         478.17               314.00                  52.28
LUFact         123.60               114.93                   7.54
FFT            403.70               380.10                   6.21
Crypt           54.97                52.20                   5.30
HeapSort        41.60                39.83                   4.44
sparseMatMult   52.47                50.90                   3.08
RayTracer      295.33               288.17                   2.49
Euler          349.67               346.93                   0.79
Search         481.47               479.27                   0.46
SOR             72.57                73.20                  -0.87
Series         173.63               177.63                  -2.25

We have not yet observed gains by converting a similar helper function:
integer_d to integer_d_sse2.

- moh
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.mozilla.org/pipermail/tamarin-devel/attachments/20080122/b4f91cc1/attachment.html 


More information about the Tamarin-devel mailing list