IPP 6, 7 and 8 performance comparison

You can do the experiments yourself. IPP comes with a performance measurement utility, usually "ps * .exe" in the ipp \ tools \ perfsys directory. It's hard to say how it was during IPP 6.x, but it should be the same. The "ps * .exe" executables allow you to measure the specific performance of the IPP function in terms of hours per element (the lower the better, of course) for various CPU optimizations. Key features for these performances. tests "-?", "-e" show all functions in the test, "-T" only enables certain CPU optimizations, "-r" saves the output to a csv file.

Let's say you want to measure the ippsIIR64f_32s_Sfs function for AVX, SSE41 and SSE3 processors. You need to run ps_ipps.exe (which is a 1D domain performance test) three times:

ps_ipps.exe -fippsIIR64f_32s_Sfs -B -R -TAVX    (you'll get csv file with AVX optimization results)
ps_ipps.exe -fippsIIR64f_32s_Sfs -B -R -TSSE41  (SSE4.1 perf. data will be appended to csv)
ps_ipps.exe -fippsIIR64f_32s_Sfs -B -R -TSSE3"   (SSE3 performance data will be appended).

      

        
        
        
      

    

Then grep csv file for required function / argument combination like

find "ippsIIR64f,32s,Sfs,32768,6,numBq_DF1" ps_ipps.csv

      

        
        
        
      

    

For example, I get

ippsIIR64f,32s,Sfs,32768,6,numBq_DF1,-,-,0,nLps=2048,1.30,cpMac,512,-
ippsIIR64f,32s,Sfs,32768,6,numBq_DF1,-,-,0,nLps=8,1.56,cpMac,613,-
ippsIIR64f,32s,Sfs,32768,6,numBq_DF1,-,-,0,nLps=4,5.61,cpMac,2.21e+003,-

      

        
        
        
      

    

This means 5.1 hours for SSE3, 1.56 hours for SSE4.1 and 1.30 hours for AVX. You must support the highest command set that you want to measure. For IPP 7 and 8, you can download versions of Intel products (Composer or Parallel Studio) from Intel website to run tests.