Benchmark: converting array of structures to structure of arrays

[LemonParty]

After this topic https://forum.lazarus.freepascal.org/index.php/topic,74158.0.html it become interesting to me how fast the pointed convertion can be done.

Here is a simple benchmark work in both x86 64-bit and AArch64. Should work at Linux too. There may be a problems with old compiler versions when compile under AArch64.

Results Intel Core Ultra 7 258V:

Code: Pascal  [Select][+][-]

1. 4096 ELEMENTS BY 100 RESULTS
2. Naive    : 4792
3. Unrolled : 3412
4. SIMD     : 2810
5. 1048576 ELEMENTS BY 4 RESULTS
6. Naive    : 1477308
7. Unrolled : 1093105
8. SIMD     : 760640
9.  
10. 4096 ELEMENTS BY 100 RESULTS
11. Naive    : 5152
12. Unrolled : 3441
13. SIMD     : 2701
14. 1048576 ELEMENTS BY 4 RESULTS
15. Naive    : 1460518
16. Unrolled : 1080429
17. SIMD     : 766542
18.  
19. 4096 ELEMENTS BY 100 RESULTS
20. Naive    : 5429
21. Unrolled : 5035
22. SIMD     : 2695
23. 1048576 ELEMENTS BY 4 RESULTS
24. Naive    : 1467502
25. Unrolled : 1067923
26. SIMD     : 791595
27.  
28. 4096 ELEMENTS BY 100 RESULTS
29. Naive    : 5038
30. Unrolled : 3374
31. SIMD     : 3051
32. 1048576 ELEMENTS BY 4 RESULTS
33. Naive    : 1464983
34. Unrolled : 1054564
35. SIMD     : 775087
36.  
37. 4096 ELEMENTS BY 100 RESULTS
38. Naive    : 4840
39. Unrolled : 3608
40. SIMD     : 2955
41. 1048576 ELEMENTS BY 4 RESULTS
42. Naive    : 1483384
43. Unrolled : 1108301
44. SIMD     : 806910

Results Raspberry Pi 5:

Code: Pascal  [Select][+][-]

1. 4096 ELEMENTS BY 100 RESULTS
2. Naive    : 47603
3. Unrolled : 46531
4. SIMD     : 5484
5. 1048576 ELEMENTS BY 4 RESULTS
6. Naive    : 8293058
7. Unrolled : 7676892
8. SIMD     : 1382494
9.  
10. Naive    : 47603
11. Unrolled : 46531
12. SIMD     : 5484
13. 1048576 ELEMENTS BY 4 RESULTS
14. Naive    : 8293058
15. Unrolled : 7676892
16. SIMD     : 1382494
17.  
18. 4096 ELEMENTS BY 100 RESULTS
19. Naive    : 29455
20. Unrolled : 29864
21. SIMD     : 3426
22. 1048576 ELEMENTS BY 4 RESULTS
23. Naive    : 8010053
24. Unrolled : 7741117
25. SIMD     : 1422446
26.  
27. 4096 ELEMENTS BY 100 RESULTS
28. Naive    : 29617
29. Unrolled : 29096
30. SIMD     : 3426
31. 1048576 ELEMENTS BY 4 RESULTS
32. Naive    : 8031165
33. Unrolled : 7690447
34. SIMD     : 1343017
35.  
36. 4096 ELEMENTS BY 100 RESULTS
37. Naive    : 43800
38. Unrolled : 43603
39. SIMD     : 5153
40. 1048576 ELEMENTS BY 4 RESULTS
41. Naive    : 8238321
42. Unrolled : 7689008
43. SIMD     : 1389328

Notes:
1. Unrolling on x86 work really well (almost +50% speed), when on ARM it only gives +7%;
2. SIMD version on ARM not uses actually SIMD, it just utilize a lot of registers. Even in this case a speed difference is almost 9.0x on small chunks and 6.0x on big chunks. That difference is impressive (I hope I don't make a mistake in "SIMD" code);
3. The SIMD solution on x86 is twice faster than Naive approach.