cpu: x64: fix perf issue for f32 conv

[xuxinzen]

Partially fixes MFDNN-14037, MFDNN-14054

This patch intends to fix the performance issue in f32 conv.
brgconv_f32_perf.xlsx

[densamoilov]

If the regression is related to introducing safe prefetching for prefetch offsets greater than INT_MAX, then the main contributor to the degradation could be register spilling.

Have you tried to move this spill to the higher level, for example to ld_loop_body? Or even higher.

oneDNN/src/cpu/x64/brgemm/jit_brgemm_kernel.cpp

Line 2606 in 48872b1

auto ld_loop_body = [&](dim_t vpad, bool last_bdb) {

[xuxinzen]

Yes, I tried to move the spill to higher level but the perf gap were still there.

Spill in ldb_loop():
export  OMP_NUM_THREADS=56 ;export KMP_AFFINITY=granularity=fine,compact; export OMP_PROC_BIND close ; numactl -m 0 -N 0 ./tests/benchdnn/benchdnn -v5 --conv --dir=FWD_I --mode=p mb21_ic256oc128_ih40oh40kh3sh1dh0ph1_iw40ow40kw3sw1dw0pw1
create: --mode=P --conv --dir=FWD_I mb21ic256ih40oc128oh40kh3ph1
oneDNN implementation: brg_conv_fwd:avx512_core
run: --mode=P --conv --dir=FWD_I mb21ic256ih40oc128oh40kh3ph1
Output template: perf,%engine%,%impl%,%name%,%prb%,%Gops%,%+ctime%,%-time%,%-Gflops%,%0time%,%0Gflops%
perf,cpu,brg_conv_fwd:avx512_core,,--mode=P --conv --dir=FWD_I mb21ic256ih40oc128oh40kh3ph1,19.163,100.7,4.60156,4164.45,4.69376,4082.65
============================================================
= Implementation statistics (--summary=no-impl to disable) =
============================================================
| brg_conv_fwd:avx512_core : 1 (100%)                      |
============================================================
tests:1 passed:1 skipped:0 mistrusted:0 unimplemented:0 invalid_arguments:0 failed:0 listed:0
total perf: min(ms):4.60156 avg(ms):4.69376

Disable safe prefetching when offset <= INT_MAX:
export  OMP_NUM_THREADS=56 ;export KMP_AFFINITY=granularity=fine,compact; export OMP_PROC_BIND close ; numactl -m 0 -N 0 ./tests/benchdnn/benchdnn -v5 --conv --dir=FWD_I --mode=p mb21_ic256oc128_ih40oh40kh3sh1dh0ph1_iw40ow40kw3sw1dw0pw1
create: --mode=P --conv --dir=FWD_I mb21ic256ih40oc128oh40kh3ph1
oneDNN implementation: brg_conv_fwd:avx512_core
run: --mode=P --conv --dir=FWD_I mb21ic256ih40oc128oh40kh3ph1
Output template: perf,%engine%,%impl%,%name%,%prb%,%Gops%,%+ctime%,%-time%,%-Gflops%,%0time%,%0Gflops%
perf,cpu,brg_conv_fwd:avx512_core,,--mode=P --conv --dir=FWD_I mb21ic256ih40oc128oh40kh3ph1,19.163,3.57324,4.14893,4618.78,4.25633,4502.23
============================================================
= Implementation statistics (--summary=no-impl to disable) =
============================================================
| brg_conv_fwd:avx512_core : 1 (100%)                      |
============================================================
tests:1 passed:1 skipped:0 mistrusted:0 unimplemented:0 invalid_arguments:0 failed:0 listed:0
total perf: min(ms):4.14893 avg(ms):4.25633
total: 3.13s; create_pd: 0.00s (0%); create_prim: 0.00s (0%); fill: 0.06s (2%); execute: 0.01s (0%);

[densamoilov]

I see. Btw, if you collected the performance data before the fix then it could affect the data because we skipped some prefetch blocks.

[xuxinzen]

make test