--------------------------
Tela/Matlab benchmark data
--------------------------
Benchmark-1 Benchmark-2
Tela-1.12
SGI/R4400 115 s 16
IBM Power-2 86 7.6
Matlab-4.2a
SGI/R4400 892 24
IBM Power-2 624 13
Tela/Matlab speed ratio
SGI/R4400 7.75 1.5
IBM Power-2 7.25 1.7
IBM/SGI speed ratio
Tela 1.3 2.1
Matlab 1.4 1.8
Both benchmarks are "real-world" examples provided by Ari.Viljanen@fmi.fi.
Benchmark-1 is scalar code and benchmark-2 operates mainly on complex vectors.
Notice that Tela/Matlab speed ratio is rather independent on architecture.
On platform comparison, IBM is relatively faster on more vectorized
computations, which is not surprising.
-------------------------
Experiences in using Tela
-------------------------
Case 1, provided by Ari.Viljanen@fmi.fi
=======================================
There were 1300 lines of badly vectorizable Matlab code, some of it quite old.
It was translated using m2t to tela. It took about 2 workdays to get
the program running in Tela. Most of the time went to correcting calls
to zeros function, which was translated incorrectly by m2t. This and
some other bugs were afterwards fixed in m2t.
CPU timings (seconds). The four machines are all Irises, simppu has
R3000 processor, sumppu has R4000, and merta and nuotta have R4400
processors. The program does both computation and binary and ASCII I/O.
Part of the computation uses complex numbers.
merta nuotta simppu sumppu
MatLab 26.7 27.6 85.2 39.4
Tela 10.7 10.3 39.9 15.7
Slight tuning of the Tela-code yielded still 20 percent improvement, but
the above comparison is more fair one.
Case 2, provided by Ari.Viljanen@fmi.fi
=======================================
CPU timings (seconds). Nuotta is SGI R4400 (150/75 MHz), simppu is
SGI R4600PC (without L2 cache), sumppu is R4000 (100/50 MHz).
nuotta simppu sumppu
MatLab 123 212 240
Tela 59 91 86