AVX Optimized App Development

[Miep]

For completion: first one boinc running, second one boinc suspended

[Claggy]

Here's the J34 results for my Atom N450 @1.66Mhz (5 runs with Boinc and 2 apps running, 5 runs with Boinc shut down)

Claggy

[Claggy]

Here's the J34 run on my C2D T8100 @2.1GHz (5 runs with Boinc and one v7 r246 task running, one AP r409 task running and one collatz_mini 2.05 Cuda task running, then 5 runs of Boinc and apps shut down)

Claggy

[Josef W. Segur]

Here's the latest version, including a fix for the AVX 8x4 transpose plus folding for AVX. The folding is as simple as I could make it, but there's no guarantee it's all correct even so. If it works I'm sure it can be improved.

I dropped the extra set of Transpose tests, there's enough data to thoroughly confuse me already. I may reactivate them later if I have some useful thoughts.
                                                                                                    Joe
Edit: Attachment deleted, newer version in later post.

[Jason G]

All legacy functions... err.. functioning.

[arkayn]

Code: [Select]

=========================================================
Ftst_v7_J34 started.

Optimal function choices:
--------------------------------------------------------
                            name   timing   error
--------------------------------------------------------
                v_BaseLineSmooth (no other)

              v_GetPowerSpectrum 0.000480 0.00000  test
             v_vGetPowerSpectrum 0.000301 0.00000  test
            v_vGetPowerSpectrum2 0.000327 0.00000  test
     v_vGetPowerSpectrumUnrolled 0.000314 0.00000  test
    v_vGetPowerSpectrumUnrolled2 0.000294 0.00000  test
    v_vGetPowerSpectrumUnrolled2 0.000294 0.00000  choice

                     v_ChirpData 0.019478 0.00000  test
                   fpu_ChirpData 0.025356 0.00000  test
               fpu_opt_ChirpData 0.015757 0.00000  test
             v_vChirpData_x86_64 0.079464 0.00000  test
               sse1_ChirpData_ak 0.011689 0.00000  test
               sse2_ChirpData_ak 0.011893 0.00000  test
              sse2_ChirpData_ak8 0.008098 0.00000  test
               sse3_ChirpData_ak 0.011029 0.00000  test
              sse2_ChirpData_ak8 0.008098 0.00000  choice

                     v_Transpose 0.041660 0.00000  test
                    v_Transpose2 0.025839 0.00000  test
                    v_Transpose4 0.012987 0.00000  test
                    v_Transpose8 0.020351 0.00000  test
                  v_pfTranspose2 0.025092 0.00000  test
                  v_pfTranspose4 0.012726 0.00000  test
                  v_pfTranspose8 0.019991 0.00000  test
                   v_vTranspose4 0.012808 0.00000  test
                 v_vTranspose4np 0.013273 0.00000  test
                v_vTranspose4ntw 0.008225 0.00000  test
              v_vTranspose4x8ntw 0.008911 0.00000  test
             v_vTranspose4x16ntw 0.007548 0.00000  test
            v_vpfTranspose8x4ntw 0.008831 0.00000  test
             v_vTranspose4x16ntw 0.007548 0.00000  choice

                 FPU opt folding 0.003467 0.00000  test
                  AK SSE folding 0.001317 0.00000  test
                  BH SSE folding 0.001285 0.00000  test
                  BH SSE folding 0.001285 0.00000  choice

                   Test duration     6.44 seconds

Ftst_v7 completed successfully.

[Miep]

1st boinc running 2nd snoozed

[Claggy]

@arkayn your posted stderr.txt says Ftst_v7_J34 and not J37

Claggy

[Claggy]

Here'a run on my C2D E8500 @4.14GHz with J37 (5 times with Boinc and apps running and 5 times with Boinc shut down)

Claggy

[Fredericx51]

Did some reading about AVX and checked its output with this Test-file.

Whithout BOINC running: sterr.txt :

=========================================================
Ftst_v7_J34 started.

Optimal function choices:
--------------------------------------------------------
                            name   timing   error
--------------------------------------------------------
                v_BaseLineSmooth (no other)

              v_GetPowerSpectrum 0.000105 0.00000  test
             v_vGetPowerSpectrum 0.000052 0.00000  test
            v_vGetPowerSpectrum2 0.000063 0.00000  test
     v_vGetPowerSpectrumUnrolled 0.000049 0.00000  test
    v_vGetPowerSpectrumUnrolled2 0.000066 0.00000  test
           v_avxGetPowerSpectrum 0.000043 0.00000  test
           v_avxGetPowerSpectrum 0.000043 0.00000  choice

                     v_ChirpData 0.005899 0.00000  test
                   fpu_ChirpData 0.010711 0.00000  test
               fpu_opt_ChirpData 0.005305 0.00000  test
             v_vChirpData_x86_64 0.051195 0.00000  test
               sse1_ChirpData_ak 0.006250 0.00000  test
               sse2_ChirpData_ak 0.005789 0.00000  test
              sse2_ChirpData_ak8 0.003679 0.00000  test
               sse3_ChirpData_ak 0.005621 0.00000  test
                 avx_ChirpData_a 0.001884 0.00000  test
                 avx_ChirpData_b 0.002139 0.00000  test
                 avx_ChirpData_a 0.001884 0.00000  choice

                     v_Transpose 0.002753 0.00000  test
                    v_Transpose2 0.002947 0.00000  test
                    v_Transpose4 0.001516 0.00000  test
                    v_Transpose8 0.002775 0.00000  test
                  v_pfTranspose2 0.001659 0.00000  test
                  v_pfTranspose4 0.001586 0.00000  test
                  v_pfTranspose8 0.002802 0.00000  test
                   v_vTranspose4 0.000915 0.00000  test
                 v_vTranspose4np 0.001169 0.00000  test
                v_vTranspose4ntw 0.007690 0.00000  test
              v_vTranspose4x8ntw 0.003222 0.00000  test
             v_vTranspose4x16ntw 0.000900 0.00000  test
            v_vpfTranspose8x4ntw 0.007704 0.00000  test
            v_avxTranspose4x8ntw 0.003195 0.00000  test
           v_avxTranspose4x16ntw 0.000817 0.00000  test
            v_avxTranspose8x4ntw 0.007712 0.00000  test
          v_avxTranspose8x8ntw_a 0.002666 0.00000  test
          v_avxTranspose8x8ntw_b 0.003011 0.00000  test
           v_avxTranspose4x16ntw 0.000817 0.00000  choice

                 FPU opt folding 0.002047 0.00000  test
                  AK SSE folding 0.000464 0.00000  test
                  BH SSE folding 0.000451 0.00000  test
                  JS AVX folding 0.000405 0.00000  test
                  JS AVX folding 0.000405 0.00000  choice

                   Test duration     2.90 seconds

Ftst_v7 completed successfully.

With BOINC (6.10.60 X64)(i7-2600 + 2x HD5870) 8x MB (CPU)+ 4 ATi  MB rev.177 or AP rev.524.

=========================================================
Ftst_v7_J34 started.

Optimal function choices:
--------------------------------------------------------
                            name   timing   error
--------------------------------------------------------
                v_BaseLineSmooth (no other)

              v_GetPowerSpectrum 0.000105 0.00000  test
             v_vGetPowerSpectrum 0.000052 0.00000  test
            v_vGetPowerSpectrum2 0.000063 0.00000  test
     v_vGetPowerSpectrumUnrolled 0.000049 0.00000  test
    v_vGetPowerSpectrumUnrolled2 0.000066 0.00000  test
           v_avxGetPowerSpectrum 0.000043 0.00000  test
           v_avxGetPowerSpectrum 0.000043 0.00000  choice

                     v_ChirpData 0.005899 0.00000  test
                   fpu_ChirpData 0.010711 0.00000  test
               fpu_opt_ChirpData 0.005305 0.00000  test
             v_vChirpData_x86_64 0.051195 0.00000  test
               sse1_ChirpData_ak 0.006250 0.00000  test
               sse2_ChirpData_ak 0.005789 0.00000  test
              sse2_ChirpData_ak8 0.003679 0.00000  test
               sse3_ChirpData_ak 0.005621 0.00000  test
                 avx_ChirpData_a 0.001884 0.00000  test
                 avx_ChirpData_b 0.002139 0.00000  test
                 avx_ChirpData_a 0.001884 0.00000  choice

                     v_Transpose 0.002753 0.00000  test
                    v_Transpose2 0.002947 0.00000  test
                    v_Transpose4 0.001516 0.00000  test
                    v_Transpose8 0.002775 0.00000  test
                  v_pfTranspose2 0.001659 0.00000  test
                  v_pfTranspose4 0.001586 0.00000  test
                  v_pfTranspose8 0.002802 0.00000  test
                   v_vTranspose4 0.000915 0.00000  test
                 v_vTranspose4np 0.001169 0.00000  test
                v_vTranspose4ntw 0.007690 0.00000  test
              v_vTranspose4x8ntw 0.003222 0.00000  test
             v_vTranspose4x16ntw 0.000900 0.00000  test
            v_vpfTranspose8x4ntw 0.007704 0.00000  test
            v_avxTranspose4x8ntw 0.003195 0.00000  test
           v_avxTranspose4x16ntw 0.000817 0.00000  test
            v_avxTranspose8x4ntw 0.007712 0.00000  test
          v_avxTranspose8x8ntw_a 0.002666 0.00000  test
          v_avxTranspose8x8ntw_b 0.003011 0.00000  test
           v_avxTranspose4x16ntw 0.000817 0.00000  choice

                 FPU opt folding 0.002047 0.00000  test
                  AK SSE folding 0.000464 0.00000  test
                  BH SSE folding 0.000451 0.00000  test
                  JS AVX folding 0.000405 0.00000  test
                  JS AVX folding 0.000405 0.00000  choice

                   Test duration     2.90 seconds

Ftst_v7 completed successfully.
=========================================================
Ftst_v7_J34 started.

Optimal function choices:
--------------------------------------------------------
                            name   timing   error
--------------------------------------------------------
                v_BaseLineSmooth (no other)

              v_GetPowerSpectrum 0.000234 0.00000  test
             v_vGetPowerSpectrum 0.000105 0.00000  test
            v_vGetPowerSpectrum2 0.000100 0.00000  test
     v_vGetPowerSpectrumUnrolled 0.000082 0.00000  test
    v_vGetPowerSpectrumUnrolled2 0.000098 0.00000  test
           v_avxGetPowerSpectrum 0.000061 0.00000  test
           v_avxGetPowerSpectrum 0.000061 0.00000  choice

                     v_ChirpData 0.011899 0.00000  test
                   fpu_ChirpData 0.019045 0.00000  test
               fpu_opt_ChirpData 0.012640 0.00000  test
             v_vChirpData_x86_64 0.063979 0.00000  test
               sse1_ChirpData_ak 0.010132 0.00000  test
               sse2_ChirpData_ak 0.009260 0.00000  test
              sse2_ChirpData_ak8 0.006961 0.00000  test
               sse3_ChirpData_ak 0.008636 0.00000  test
                 avx_ChirpData_a 0.003490 0.00000  test
                 avx_ChirpData_b 0.003833 0.00000  test
                 avx_ChirpData_a 0.003490 0.00000  choice

                     v_Transpose 0.007700 0.00000  test
                    v_Transpose2 0.004792 0.00000  test
                    v_Transpose4 0.008537 0.00000  test
                    v_Transpose8 0.014129 0.00000  test
                  v_pfTranspose2 0.015112 0.00000  test
                  v_pfTranspose4 0.011302 0.00000  test
                  v_pfTranspose8 0.012998 0.00000  test
                   v_vTranspose4 0.002625 0.00000  test
                 v_vTranspose4np 0.005798 0.00000  test
                v_vTranspose4ntw 0.008330 0.00000  test
              v_vTranspose4x8ntw 0.004689 0.00000  test
             v_vTranspose4x16ntw 0.002755 0.00000  test
            v_vpfTranspose8x4ntw 0.008488 0.00000  test
            v_avxTranspose4x8ntw 0.003759 0.00000  test
           v_avxTranspose4x16ntw 0.002249 0.00000  test
            v_avxTranspose8x4ntw 0.008294 0.00000  test
          v_avxTranspose8x8ntw_a 0.003551 0.00000  test
          v_avxTranspose8x8ntw_b 0.004706 0.00000  test
           v_avxTranspose4x16ntw 0.002249 0.00000  choice

                 FPU opt folding 0.003407 0.00000  test
                  AK SSE folding 0.000878 0.00000  test
                  BH SSE folding 0.000816 0.00000  test
                  JS AVX folding 0.000656 0.00000  test
                  JS AVX folding 0.000656 0.00000  choice

                   Test duration     5.03 seconds

Ftst_v7 completed successfully.

[arkayn]

@arkayn your posted stderr.txt says Ftst_v7_J34 and not J37

Claggy

So does yours and everyone else, I am guessing that Joe forgot to change that number.

[Josef W. Segur]

@arkayn your posted stderr.txt says Ftst_v7_J34 and not J37

Claggy

So does yours and everyone else, I am guessing that Joe forgot to change that number.

'Tis true. Luckily having only one set of transposes serves well to distinguish J37 from J34.

The test outputs are appreciated, I'd hate to break something and not know it because of the limited systems I have.

Did some reading about AVX and checked its output with this Test-file.
...

Thanks very much, it seems I've reached the initial goal of having working avx variants for all the tested categories. I'll do some tweaking to see if I can improve performance.
                                                                                            Joe

[Raistmer]

Pulse finding is another area which is more limited by processing speed than memory access, though, and it also accounts for a sizable fraction of run time.

On GPU pulse finding looks memory-bound...
Maybe GaussFit ? It's compute intensive search.

[Josef W. Segur]

Here's another test, added a third variant AVX chirp which might be marginally faster and a second variant AVX folding set, ditto. AFAIK no changes to the non-AVX code other than making sure this one identifies itself correctly 
                                                                                                 Joe
Edit: Attachment deleted, newer version in later post.

[Jason G]

... AFAIK no changes to the non-AVX code other than making sure this one identifies itself correctly ...

  Checked anyway, All good including:

Ftst_v7_J39 started.