Re: [PATCH] SVE popcount support

Поиск
Список
Период
Сортировка
От Nathan Bossart
Тема Re: [PATCH] SVE popcount support
Дата
Msg-id Z-R1OP2s3mYs_DIP@nathan
обсуждение исходный текст
Ответ на Re: [PATCH] SVE popcount support  (Nathan Bossart <nathandbossart@gmail.com>)
Ответы Re: [PATCH] SVE popcount support
Список pgsql-hackers
I've attached a new set of patches in which I've tried to address John's
feedback.  I ran some new benchmarks with these patches.  "M3" is an Apple
M3 (my laptop), "G3" is an r7g.4xlarge, and "G4" is an r8g.4xlarge.  "no
SVE" means the patches are applied but the function pointer points to the
Neon implementation.  "SVE" and "patched" mean all the patches are applied
with no changes.

 8 byte words | M3 HEAD | M3 patched | G3 HEAD | G3 no SVE | G3 SVE  | G4 HEAD | G4 no SVE | G4 SVE
--------------+---------+------------+---------+-----------+---------+---------+-----------+---------
            1 |     3.6 |     3.0    |     3.1 |     2.9   |     3.1 |     2.5 |     2.2   |     1.8
            2 |     6.4 |     4.4    |     3.1 |     3.0   |     3.1 |     2.5 |     2.5   |     2.0
            3 |     7.3 |     6.9    |     3.5 |     3.5   |     3.1 |     3.3 |     3.2   |     2.0
            4 |     8.0 |     3.8    |     4.0 |     2.7   |     4.7 |     3.6 |     2.2   |     2.7
            5 |     9.4 |     5.5    |     4.6 |     2.8   |     4.6 |     3.9 |     2.5   |     2.7
            6 |     7.9 |     5.0    |     5.1 |     3.5   |     4.7 |     4.3 |     3.1   |     3.4
            7 |    10.2 |     7.4    |     5.9 |     4.0   |     4.7 |     4.7 |     3.6   |     3.4
            8 |    12.0 |     5.4    |     6.5 |     4.0   |     5.9 |     5.0 |     3.2   |     2.5
            9 |    11.7 |     6.5    |     7.2 |     4.3   |     5.9 |     5.4 |     3.6   |     2.5
           10 |    12.5 |     5.4    |     8.0 |     4.8   |     5.9 |     6.2 |     3.9   |     3.1
           11 |    14.0 |     8.6    |     8.5 |     5.5   |     5.9 |     6.1 |     5.0   |     3.1
           12 |    13.1 |     5.7    |     9.1 |     5.1   |     7.4 |     6.4 |     3.9   |     3.6
           13 |    12.1 |     6.8    |     9.8 |     5.4   |     7.3 |     6.8 |     4.3   |     3.6
           14 |    16.4 |     7.8    |    10.4 |     5.9   |     7.4 |     7.2 |     4.7   |     4.4
           15 |    17.4 |     8.0    |    11.1 |     6.6   |     7.4 |     7.5 |     5.7   |     4.4
           16 |    15.5 |     5.7    |    11.8 |     5.7   |     4.7 |     7.9 |     5.0   |     3.5
           32 |    26.0 |    16.2    |    22.7 |    10.3   |     6.2 |    16.8 |     8.4   |     5.2
           64 |    38.5 |    20.3    |    42.7 |    20.1   |     9.3 |    31.8 |    15.4   |     8.8
          128 |    75.1 |    35.7    |    86.1 |    35.0   |    15.4 |    80.2 |    28.6   |    16.3
          256 |   117.7 |    51.8    |   179.6 |    68.2   |    27.8 |   154.0 |    55.7   |    30.9
          512 |   198.5 |    93.1    |   329.3 |   134.4   |    52.4 |   246.5 |   110.2   |    59.4
         1024 |   355.0 |   159.2    |   673.6 |   265.8   |   101.7 |   487.0 |   219.0   |   114.7
         2048 |   669.5 |   288.8    |  1294.7 |   529.7   |   200.3 |   969.3 |   438.7   |   228.5
         4096 |  1308.0 |   552.8    |  2784.3 |  1063.0   |   397.4 |  1934.5 |   874.4   |   455.9

IMHO these are acceptable results, at least for the use-cases I see in the
tree.  We might be able to minimize the difference between the Neon and SVE
implementations on the low end with some additional code, but I'm really
not sure if it's worth the effort.

Barring feedback or objections, I'm planning to commit these on Friday.

-- 
nathan

Вложения

В списке pgsql-hackers по дате отправления: