PROOF-OF-CONCEPT Replace x86_64 asm by CryptOpt output #1254

1. 
   real-or-random commented at 3:28 pm on March 30, 2023: contributor

   @dderjoel

   It managed to convince autotools to compile your asm. It seems slower than our C code, at least on my machine:

   12th Gen Intel(R) Core(TM) i7-1260P, TurboBoost disabled

    0./bench_internal field
    1
    2Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)
    3
    4this PR (e2684293b1b72a1ab5974a2864549cea2788cf95), --with-asm=x86_64 (which is the default)
    5field_sqr                     ,     0.0296    ,     0.0297    ,     0.0297 
    6field_mul                     ,     0.0339    ,     0.0341    ,     0.0344 
    7
    8master (464a9115b4eda46b464d22829ece4f51985944bf), --with-asm=x86_64 (which is the default)
    9field_sqr                     ,     0.0270    ,     0.0271    ,     0.0272 
   10field_mul                     ,     0.0359    ,     0.0359    ,     0.0360 
   11
   12master, --with-asm=no (464a9115b4eda46b464d22829ece4f51985944bf), gcc 12.2.1 -02:
   13Benchmark                     ,    Min(us)    ,    Avg(us)    ,    Max(us)
   14field_sqr                     ,     0.0236    ,     0.0238    ,     0.0240 
   15field_mul                     ,     0.0283    ,     0.0284    ,     0.0286
   

   By the way, we should really get rid of our current ASM in the short term, at least for the field. GCC beats it significantly.

2. Replace x86_64 field asm by CryptOpt output

   Co-authored-by: Tim Ruffing <crypto@timruffing.de>

   e2684293b1
3. 
   dderjoel commented at 4:44 am on March 31, 2023: none

   1. What are the magic incantations to convince Autotools?
   2. I’ve added the fiat-dettman-c code the benchmark is attached (but probably not too representative, as I cant compare against the asm, as I don’t have the magic words for auto tools)
4. 
   dderjoel commented at 6:36 am on March 31, 2023: none

   original C

   $ ./bench_ecmult Benchmark , Min(us) , Avg(us) , Max(us)

   ecmult_gen , 18.4 , 18.6 , 18.8
   ecmult_const , 36.3 , 36.7 , 39.0
   ecmult_1p , 28.9 , 29.3 , 29.8
   ecmult_0p_g , 20.2 , 20.6 , 23.7
   ecmult_1p_g , 16.9 , 17.0 , 17.2
   ecmult_multi_0p_g , 20.0 , 20.3 , 20.7

5. 
   dderjoel commented at 6:38 am on March 31, 2023: none

   cryptopt asm

   $ ./bench_ecmult
   Benchmark , Min(us) , Avg(us) , Max(us)

   ecmult_gen , 17.7 , 18.1 , 19.2
   ecmult_const , 34.1 , 35.1 , 37.9
   ecmult_1p , 27.2 , 28.1 , 29.4
   ecmult_0p_g , 18.7 , 19.5 , 21.4
   ecmult_1p_g , 16.2 , 16.9 , 18.6
   ecmult_multi_0p_g , 18.8 , 19.4 , 20.2

6. 
   dderjoel commented at 7:09 am on March 31, 2023: none

   fiat-dettman c ./bench_ecmult Benchmark , Min(us) , Avg(us) , Max(us)

   ecmult_gen , 19.5 , 20.1 , 21.5
   ecmult_const , 38.9 , 40.4 , 41.4
   ecmult_1p , 30.3 , 31.2 , 33.5
   ecmult_0p_g , 20.8 , 21.3 , 24.0
   ecmult_1p_g , 17.5 , 18.1 , 18.7
   ecmult_multi_0p_g , 22.1 , 23.2 , 25.3

7. fixup! Replace x86_64 field asm by CryptOpt output 7e00a589eb
8. 
   real-or-random commented at 5:39 am on April 6, 2023: contributor

   Tracked in #1261.
9. real-or-random closed this on Apr 6, 2023

   ---

Contributors
real-or-random dderjoel