https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116582
--- Comment #2 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
it is mysterious. I was looking into why in some cases the gather is a win in
micro-benchmark and loss in real benchmark. Indeed distribution of indices
makes difference.
If I make indices random then the performance effect is neutral:
jh@shroud:/tmp> ~/trunk-install-znver5/bin/g++ -DT=float -march=native
gather.c -Ofast -mtune-ctrl=^use_gather_4parts -mtune-ctrl=^use_gather_8parts
-mtune-ctrl=^use_gather -mtune=native -fdump-tree-all-details ; objdump -d
a.out | grep gather ; perf stat ./a.out
Performance counter stats for './a.out':
454.77 msec task-clock:u # 0.999 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
663 page-faults:u # 1.458 K/sec
1,854,500,227 cycles:u # 4.078 GHz
4,788,337 stalled-cycles-frontend:u # 0.26% frontend
cycles idle
651,597,070 instructions:u # 0.35 insn per
cycle
# 0.01 stalled cycles per
insn
58,222,408 branches:u # 128.027 M/sec
60,269 branch-misses:u # 0.10% of all
branches
0.455155383 seconds time elapsed
0.455154000 seconds user
0.000000000 seconds sys
jh@shroud:/tmp> ~/trunk-install-znver5/bin/g++ -DT=float -march=native
gather.c -Ofast -mtune-ctrl=use_gather_4parts -mtune-ctrl=use_gather_8parts
-mtune-ctrl=use_gather -mtune=native -fdump-tree-all-details ; objdump -d a.out
| grep gather ; perf stat ./a.out
401212: 62 f2 7d 4a 92 04 8d vgatherdps
0x404080(,%zmm1,4),%zmm0{%k2}
Performance counter stats for './a.out':
448.84 msec task-clock:u # 0.999 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
663 page-faults:u # 1.477 K/sec
1,834,437,666 cycles:u # 4.087 GHz
4,522,424 stalled-cycles-frontend:u # 0.25% frontend
cycles idle
160,137,040 instructions:u # 0.09 insn per
cycle
# 0.03 stalled cycles per
insn
27,502,394 branches:u # 61.274 M/sec
60,328 branch-misses:u # 0.22% of all
branches
0.449240415 seconds time elapsed
0.449224000 seconds user
0.000000000 seconds sys
If I make stride 8 then it is a win:
#include <stdlib.h>
#define M 1024*1024
int indices[M];
T a[M], b[M];
__attribute__ ((noipa))
void
test ()
{
for (int i = 0; i < 1024* 16; i++)
a[i] += b[indices[i]];
}
int
main()
{
for (int i = 0 ; i < M; i++)
indices[i] = (i * 8)%M;
for (int i = 0 ; i < 10000; i++)
test ();
return 0;
}
jh@shroud:/tmp> ~/trunk-install-znver5/bin/g++ -DT=float -march=native
gather.c -Ofast -mtune-ctrl=^use_gather_4parts -mtune-ctrl=^use_gather_8parts
-mtune-ctrl=^use_gather -mtune=native -fdump-tree-all-details ; objdump -d
a.out | grep gather ; perf stat ./a.out
Performance counter stats for './a.out':
5,827.78 msec task-clock:u # 1.000 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
222 page-faults:u # 38.093 /sec
23,975,482,386 cycles:u # 4.114 GHz
784,362,546 stalled-cycles-frontend:u # 3.27% frontend
cycles idle
576,680,806 instructions:u # 0.02 insn per
cycle
# 1.36 stalled cycles per
insn
41,523,290 branches:u # 7.125 M/sec
53,461 branch-misses:u # 0.13% of all
branches
5.828522527 seconds time elapsed
5.828224000 seconds user
0.000000000 seconds sys
jh@shroud:/tmp> ~/trunk-install-znver5/bin/g++ -DT=float -march=native
gather.c -Ofast -mtune-ctrl=use_gather_4parts -mtune-ctrl=use_gather_8parts
-mtune-ctrl=use_gather -mtune=native -fdump-tree-all-details ; objdump -d a.out
| grep gather ; perf stat ./a.out
401252: 62 f2 7d 4a 92 04 8d vgatherdps
0x404080(,%zmm1,4),%zmm0{%k2}
Performance counter stats for './a.out':
825.60 msec task-clock:u # 0.999 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
220 page-faults:u # 266.472 /sec
3,398,434,808 cycles:u # 4.116 GHz
76,631,825 stalled-cycles-frontend:u # 2.25% frontend
cycles idle
85,196,162 instructions:u # 0.03 insn per
cycle
# 0.90 stalled cycles per
insn
10,790,219 branches:u # 13.069 M/sec
458,624 branch-misses:u # 4.25% of all
branches
0.826054290 seconds time elapsed
0.826079000 seconds user
0.000000000 seconds sys
if I make it fully sequential:
#include <stdlib.h>
#define M 1024*1024
int indices[M];
T a[M], b[M];
__attribute__ ((noipa))
void
test ()
{
for (int i = 0; i < 1024* 16; i++)
a[i] += b[indices[i]];
}
int
main()
{
for (int i = 0 ; i < M; i++)
indices[i] = i;
for (int i = 0 ; i < 10000; i++)
test ();
return 0;
}
then it is loss:
jh@shroud:/tmp> ~/trunk-install-znver5/bin/g++ -DT=float -march=native
gather.c -Ofast -mtune-ctrl=^use_gather_4parts -mtune-ctrl=^use_gather_8parts
-mtune-ctrl=^use_gather -mtune=native -fdump-tree-all-details ; objdump -d
a.out | grep gather ; perf stat ./a.out
Performance counter stats for './a.out':
36.87 msec task-clock:u # 0.990 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
174 page-faults:u # 4.720 K/sec
145,666,993 cycles:u # 3.951 GHz
764,475 stalled-cycles-frontend:u # 0.52% frontend
cycles idle
576,521,394 instructions:u # 3.96 insn per
cycle
# 0.00 stalled cycles per
insn
41,508,228 branches:u # 1.126 G/sec
23,851 branch-misses:u # 0.06% of all
branches
0.037254683 seconds time elapsed
0.037398000 seconds user
0.000000000 seconds sys
jh@shroud:/tmp> ~/trunk-install-znver5/bin/g++ -DT=float -march=native
gather.c -Ofast -mtune-ctrl=use_gather_4parts -mtune-ctrl=use_gather_8parts
-mtune-ctrl=use_gather -mtune=native -fdump-tree-all-details ; objdump -d a.out
| grep gather ; perf stat ./a.out
401252: 62 f2 7d 4a 92 04 8d vgatherdps
0x404080(,%zmm1,4),%zmm0{%k2}
Performance counter stats for './a.out':
59.06 msec task-clock:u # 0.994 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
172 page-faults:u # 2.912 K/sec
236,520,114 cycles:u # 4.005 GHz
879,734 stalled-cycles-frontend:u # 0.37% frontend
cycles idle
85,061,573 instructions:u # 0.36 insn per
cycle
# 0.01 stalled cycles per
insn
10,788,320 branches:u # 182.674 M/sec
24,354 branch-misses:u # 0.23% of all
branches
0.059442906 seconds time elapsed
0.059527000 seconds user
0.000000000 seconds sys
if I make it inverse sequential then it is also loss:
jh@shroud:/tmp> ~/trunk-install-znver5/bin/g++ -DT=float -march=native
gather.c -Ofast -mtune-ctrl=^use_gather_4parts -mtune-ctrl=^use_gather_8parts
-mtune-ctrl=^use_gather -mtune=native -fdump-tree-all-details ; objdump -d
a.out | grep gather ; perf stat ./a.out
Performance counter stats for './a.out':
36.84 msec task-clock:u # 0.985 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
158 page-faults:u # 4.289 K/sec
146,386,127 cycles:u # 3.974 GHz
778,318 stalled-cycles-frontend:u # 0.53% frontend
cycles idle
576,521,379 instructions:u # 3.94 insn per
cycle
# 0.00 stalled cycles per
insn
41,508,213 branches:u # 1.127 G/sec
23,784 branch-misses:u # 0.06% of all
branches
0.037391272 seconds time elapsed
0.037422000 seconds user
0.000000000 seconds sys
jh@shroud:/tmp> ~/trunk-install-znver5/bin/g++ -DT=float -march=native
gather.c -Ofast -mtune-ctrl=use_gather_4parts -mtune-ctrl=use_gather_8parts
-mtune-ctrl=use_gather -mtune=native -fdump-tree-all-details ; objdump -d a.out
| grep gather ; perf stat ./a.out
401252: 62 f2 7d 4a 92 04 8d vgatherdps
0x404080(,%zmm1,4),%zmm0{%k2}
Performance counter stats for './a.out':
59.14 msec task-clock:u # 0.993 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
160 page-faults:u # 2.705 K/sec
236,706,166 cycles:u # 4.002 GHz
921,999 stalled-cycles-frontend:u # 0.39% frontend
cycles idle
85,061,576 instructions:u # 0.36 insn per
cycle
# 0.01 stalled cycles per
insn
10,788,314 branches:u # 182.419 M/sec
24,412 branch-misses:u # 0.23% of all
branches
0.059586646 seconds time elapsed
0.059648000 seconds user
0.000000000 seconds sys
Stride 2 is neutral, stride 4 is win, stride 8 is win, stride 16 is win.
I checked that this is reproducible
On zen5 gather is a 5-10% loss on parest. I suppose CPU's prefetching logic is
behaving a lot different when gathers are used.
