struct Dav1dMCDSPContext: Dispatch through indirect fn ptrs inside of bitdepth-dependent fns

[kkysen]

This may help reduce the instruction count and thus not be worse in performance compared to the purely fn ptr version, but we'll see exactly how much (if at all).

[kkysen]

@thedataking, could you re-benchmark this one, too, on the AVX512 machine? Thanks!

[kkysen]

It doesn't compile on arm yet, but this should be enough to benchmark it on x86_64 to see if it's worth it at all.

[thedataking]

It looks like this change would regress performance.

main on amy:

 Performance counter stats for 'target/release/dav1d_main -i /home/perl/.phoronix-test-suite/installed-tests/pts/dav1d-1.13.0/summer_nature_1080p.ivf -o /dev/null' (10 runs):

        431,493.49 msec task-clock                       #  137.049 CPUs utilized               ( +-  9.58% )
         2,686,890      context-switches                 #   11.358 K/sec                       ( +-  9.56% )
           241,000      cpu-migrations                   #    1.019 K/sec                       ( +-  9.51% )
           229,047      page-faults                      #  968.242 /sec                        ( +-  9.53% )
 2,087,648,036,957      cycles                           #    8.825 GHz                         ( +-  9.57% )  (33.24%)
     4,227,888,146      stalled-cycles-frontend          #    0.37% frontend cycles idle        ( +-  9.51% )  (33.29%)
    95,775,550,882      stalled-cycles-backend           #    8.36% backend cycles idle         ( +-  9.58% )  (33.35%)
 3,280,933,862,233      instructions                     #    2.86  insn per cycle
                                                  #    0.02  stalled cycles per insn     ( +-  9.58% )  (33.46%)
   365,372,275,615      branches                         #    1.545 G/sec                       ( +-  9.58% )  (33.58%)
    10,750,889,079      branch-misses                    #    5.35% of all branches             ( +-  9.58% )  (33.66%)
 1,173,408,463,591      L1-dcache-loads                  #    4.960 G/sec                       ( +-  9.58% )  (33.67%)
    54,907,344,945      L1-dcache-load-misses            #    8.51% of all L1-dcache accesses   ( +-  9.57% )  (33.61%)
   <not supported>      LLC-loads
   <not supported>      LLC-load-misses
   182,350,420,327      L1-icache-loads                  #  770.844 M/sec                       ( +-  9.58% )  (33.51%)
     2,130,409,531      L1-icache-load-misses            #    2.12% of all L1-icache accesses   ( +-  9.56% )  (33.47%)
    11,047,434,742      dTLB-loads                       #   46.700 M/sec                       ( +-  9.54% )  (33.43%)
       699,776,941      dTLB-load-misses                 #   11.58% of all dTLB cache accesses  ( +-  9.55% )  (33.37%)
       586,365,696      iTLB-loads                       #    2.479 M/sec                       ( +-  9.02% )  (33.36%)
        34,390,790      iTLB-load-misses                 #    9.99% of all iTLB cache accesses  ( +-  9.43% )  (33.32%)
    13,044,351,246      L1-dcache-prefetches             #   55.142 M/sec                       ( +-  9.58% )  (33.27%)
   <not supported>      L1-dcache-prefetch-misses

            3.1485 +- 0.0103 seconds time elapsed  ( +-  0.33% )

kkysen/devirtualize-mc-indirect on amy:

 Performance counter stats for 'target/release/dav1d_mc_indirect -i /home/perl/.phoronix-test-suite/installed-tests/pts/dav1d-1.13.0/summer_nature_1080p.ivf -o /dev/null' (10 runs):

        462,105.07 msec task-clock                       #  135.885 CPUs utilized               ( +-  9.59% )
         2,530,994      context-switches                 #    9.988 K/sec                       ( +-  9.50% )
           268,008      cpu-migrations                   #    1.058 K/sec                       ( +-  9.84% )
           228,258      page-faults                      #  900.738 /sec                        ( +-  9.60% )
 2,235,047,462,245      cycles                           #    8.820 GHz                         ( +-  9.59% )  (33.23%)
     4,595,687,183      stalled-cycles-frontend          #    0.37% frontend cycles idle        ( +-  9.57% )  (33.31%)
   108,933,740,731      stalled-cycles-backend           #    8.89% backend cycles idle         ( +-  9.61% )  (33.31%)
 3,408,763,954,940      instructions                     #    2.78  insn per cycle
                                                  #    0.02  stalled cycles per insn     ( +-  9.57% )  (33.41%)
   379,986,315,022      branches                         #    1.499 G/sec                       ( +-  9.57% )  (33.51%)
    10,818,380,238      branch-misses                    #    5.18% of all branches             ( +-  9.57% )  (33.59%)
 1,223,546,578,914      L1-dcache-loads                  #    4.828 G/sec                       ( +-  9.57% )  (33.53%)
    64,870,772,256      L1-dcache-load-misses            #    9.64% of all L1-dcache accesses   ( +-  9.56% )  (33.52%)
   <not supported>      LLC-loads
   <not supported>      LLC-load-misses
   186,083,422,221      L1-icache-loads                  #  734.311 M/sec                       ( +-  9.57% )  (33.49%)
     2,394,506,643      L1-icache-load-misses            #    2.34% of all L1-icache accesses   ( +-  9.56% )  (33.51%)
    13,531,234,110      dTLB-loads                       #   53.396 M/sec                       ( +-  9.62% )  (33.47%)
       721,226,217      dTLB-load-misses                 #    9.73% of all dTLB cache accesses  ( +-  9.61% )  (33.45%)
       564,822,964      iTLB-loads                       #    2.229 M/sec                       ( +-  9.62% )  (33.45%)
        38,534,660      iTLB-load-misses                 #   12.54% of all iTLB cache accesses  ( +-  9.51% )  (33.38%)
    13,718,626,090      L1-dcache-prefetches             #   54.136 M/sec                       ( +-  9.58% )  (33.27%)
   <not supported>      L1-dcache-prefetch-misses

            3.4007 +- 0.0247 seconds time elapsed  ( +-  0.73% )

main on donna:

 Performance counter stats for 'target/release/dav1d_main -i /tmp/summer_nature_1080p.ivf -o /dev/null' (10 runs):

         72,656.29 msec task-clock                #   15.901 CPUs utilized            ( +-  0.38% )
           722,725      context-switches          #    9.837 K/sec                    ( +-  0.41% )
            19,175      cpu-migrations            #  260.997 /sec                     ( +-  1.08% )
            45,929      page-faults               #  625.155 /sec                     ( +-  0.35% )
   279,644,470,112      cycles                    #    3.806 GHz                      ( +-  0.70% )  (33.14%)
     1,885,631,074      stalled-cycles-frontend   #    0.66% frontend cycles idle     ( +-  0.18% )  (33.04%)
    25,263,784,676      stalled-cycles-backend    #    8.90% backend cycles idle      ( +-  1.54% )  (32.88%)
   361,993,846,839      instructions              #    1.28  insn per cycle
                                                  #    0.07  stalled cycles per insn  ( +-  0.07% )  (33.15%)
    38,210,726,676      branches                  #  520.099 M/sec                    ( +-  0.05% )  (33.22%)
     1,213,950,217      branch-misses             #    3.18% of all branches          ( +-  0.16% )  (33.44%)
   124,785,807,301      L1-dcache-loads           #    1.699 G/sec                    ( +-  0.04% )  (33.68%)
     5,584,477,464      L1-dcache-load-misses     #    4.47% of all L1-dcache accesses  ( +-  0.28% )  (33.83%)
   <not supported>      LLC-loads
   <not supported>      LLC-load-misses
    28,724,499,259      L1-icache-loads           #  390.979 M/sec                    ( +-  0.05% )  (33.76%)
       448,893,743      L1-icache-load-misses     #    1.56% of all L1-icache accesses  ( +-  0.24% )  (33.87%)
     1,771,108,036      dTLB-loads                #   24.107 M/sec                    ( +-  0.13% )  (33.71%)
       131,660,378      dTLB-load-misses          #    7.44% of all dTLB cache accesses  ( +-  0.16% )  (33.66%)
        67,381,885      iTLB-loads                #  917.157 K/sec                    ( +-  0.35% )  (33.61%)
         5,307,938      iTLB-load-misses          #    7.92% of all iTLB cache accesses  ( +-  1.06% )  (33.44%)
     1,830,368,661      L1-dcache-prefetches      #   24.914 M/sec                    ( +-  0.34% )  (33.24%)
   <not supported>      L1-dcache-prefetch-misses

            4.5693 +- 0.0201 seconds time elapsed  ( +-  0.44% )

kkysen/devirtualize-mc-indirect on donna:

 Performance counter stats for 'target/release/dav1d_mc_indirect -i /tmp/summer_nature_1080p.ivf -o /dev/null' (10 runs):

         77,115.91 msec task-clock                #   16.603 CPUs utilized            ( +-  0.34% )
           700,553      context-switches          #    9.306 K/sec                    ( +-  0.25% )
            17,129      cpu-migrations            #  227.538 /sec                     ( +-  1.36% )
            44,295      page-faults               #  588.406 /sec                     ( +-  0.42% )
   298,743,542,507      cycles                    #    3.968 GHz                      ( +-  0.47% )  (33.36%)
     1,899,534,123      stalled-cycles-frontend   #    0.66% frontend cycles idle     ( +-  0.20% )  (33.25%)
    28,351,565,762      stalled-cycles-backend    #    9.79% backend cycles idle      ( +-  0.93% )  (33.54%)
   373,279,062,170      instructions              #    1.29  insn per cycle
                                                  #    0.07  stalled cycles per insn  ( +-  0.06% )  (33.58%)
    39,490,583,291      branches                  #  524.585 M/sec                    ( +-  0.07% )  (33.56%)
     1,208,988,030      branch-misses             #    3.05% of all branches          ( +-  0.11% )  (33.65%)
   129,535,941,286      L1-dcache-loads           #    1.721 G/sec                    ( +-  0.08% )  (33.91%)
     6,374,806,377      L1-dcache-load-misses     #    4.91% of all L1-dcache accesses  ( +-  0.14% )  (33.49%)
   <not supported>      LLC-loads
   <not supported>      LLC-load-misses
    29,380,057,051      L1-icache-loads           #  390.279 M/sec                    ( +-  0.14% )  (33.67%)
       362,368,616      L1-icache-load-misses     #    1.24% of all L1-icache accesses  ( +-  0.33% )  (33.57%)
     2,061,205,631      dTLB-loads                #   27.381 M/sec                    ( +-  0.12% )  (33.34%)
       131,258,869      dTLB-load-misses          #    6.36% of all dTLB cache accesses  ( +-  0.22% )  (33.12%)
        49,283,601      iTLB-loads                #  654.674 K/sec                    ( +-  0.71% )  (33.33%)
         5,110,826      iTLB-load-misses          #   10.42% of all iTLB cache accesses  ( +-  1.12% )  (33.17%)
     1,912,415,593      L1-dcache-prefetches      #   25.404 M/sec                    ( +-  0.35% )  (33.28%)
   <not supported>      L1-dcache-prefetch-misses

            4.6448 +- 0.0160 seconds time elapsed  ( +-  0.34% )

[kkysen]

Okay, it looks like this way isn't worth it either. I'll start working on a type-erased fn ptr version of #322.