AuroraMiddleware/glibc
1
0
Fork 0
mirror of https://sourceware.org/git/glibc.git synced 2024-09-19 16:10:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

x86: Disable non-temporal memset on Skylake Server

Browse Source 
The original commit enabling non-temporal memset on Skylake Server had
erroneous benchmarks (actually done on ICX).

Further benchmarks indicate non-temporal stores may in fact by a
regression on Skylake Server.

This commit may be over-cautious in some cases, but should avoid any
regressions for 2.40.

Tested using qemu on all x86_64 cpu arch supported by both qemu +
GLIBC.

Reviewed-by: DJ Delorie <dj@redhat.com>
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
This commit is contained in:
Noah Goldstein 2024-07-15 16:19:17 +08:00
parent 2dcc908538
commit 5bcf6265f2
5 changed files with 25 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
sysdeps/x86/cpu-features.c
View File

@ -870,11 +870,18 @@ init_cpu_features (struct cpu_features *cpu_features)
/* Newer Bigcore microarch (larger non-temporal store
threshold). */
case INTEL_BIGCORE_SKYLAKE:
case INTEL_BIGCORE_KABYLAKE:
case INTEL_BIGCORE_COMETLAKE:
case INTEL_BIGCORE_SKYLAKE_AVX512:
case INTEL_BIGCORE_CANNONLAKE:
/* Benchmarks indicate non-temporal memset is not
necessarily profitable on SKX (and in some cases much
worse). This is likely unique to SKX due its it unique
mesh interconnect (not present on ICX or BWD). Disable
non-temporal on all Skylake servers. */
cpu_features->preferred[index_arch_Avoid_Non_Temporal_Memset]
|= bit_arch_Avoid_Non_Temporal_Memset;
case INTEL_BIGCORE_COMETLAKE:
case INTEL_BIGCORE_SKYLAKE:
case INTEL_BIGCORE_KABYLAKE:
case INTEL_BIGCORE_ICELAKE:
case INTEL_BIGCORE_TIGERLAKE:
case INTEL_BIGCORE_ROCKETLAKE:

5
sysdeps/x86/cpu-tunables.c
View File

@ -243,6 +243,11 @@ TUNABLE_CALLBACK (set_hwcaps) (tunable_val_t *valp)
(n, cpu_features, MathVec_Prefer_No_AVX512, AVX512F, 24);
}
break;
case 25:
{
CHECK_GLIBC_IFUNC_PREFERRED_BOTH (n, cpu_features,
Avoid_Non_Temporal_Memset, 25);
}
case 26:
{
CHECK_GLIBC_IFUNC_PREFERRED_NEED_BOTH

13
sysdeps/x86/dl-cacheinfo.h
View File

@ -991,13 +991,14 @@ dl_init_cacheinfo (struct cpu_features *cpu_features)
/* Non-temporal stores are more performant on Intel and AMD hardware above
non_temporal_threshold. Enable this for both Intel and AMD hardware. */
unsigned long int memset_non_temporal_threshold = SIZE_MAX;
if (cpu_features->basic.kind == arch_kind_intel
|| cpu_features->basic.kind == arch_kind_amd)
memset_non_temporal_threshold = non_temporal_threshold;
if (!CPU_FEATURES_ARCH_P (cpu_features, Avoid_Non_Temporal_Memset)
&& (cpu_features->basic.kind == arch_kind_intel
|| cpu_features->basic.kind == arch_kind_amd))
memset_non_temporal_threshold = non_temporal_threshold;
/* For AMD CPUs that support ERMS (Zen3+), REP MOVSB is in a lot of
cases slower than the vectorized path (and for some alignments,
it is really slow, check BZ #30994). */
/* For AMD CPUs that support ERMS (Zen3+), REP MOVSB is in a lot of
cases slower than the vectorized path (and for some alignments,
it is really slow, check BZ #30994). */
if (cpu_features->basic.kind == arch_kind_amd)
rep_movsb_threshold = non_temporal_threshold;

1
sysdeps/x86/include/cpu-features-preferred_feature_index_1.def
View File

@ -33,3 +33,4 @@ BIT (Prefer_No_AVX512)
BIT (MathVec_Prefer_No_AVX512)
BIT (Prefer_FSRM)
BIT (Avoid_Short_Distance_REP_MOVSB)
BIT (Avoid_Non_Temporal_Memset)

4
sysdeps/x86/tst-hwcap-tunables.c
View File

@ -60,7 +60,7 @@ static const struct test_t
/* Disable everything. */
"-Prefer_ERMS,-Prefer_FSRM,-AVX,-AVX2,-AVX512F,-AVX512VL,"
"-SSE4_1,-SSE4_2,-SSSE3,-Fast_Unaligned_Load,-ERMS,"
"-AVX_Fast_Unaligned_Load",
"-AVX_Fast_Unaligned_Load,-Avoid_Non_Temporal_Memset",
test_1,
array_length (test_1)
},
@ -68,7 +68,7 @@ static const struct test_t
/* Same as before, but with some empty suboptions. */
",-,-Prefer_ERMS,-Prefer_FSRM,-AVX,-AVX2,-AVX512F,-AVX512VL,"
"-SSE4_1,-SSE4_2,-SSSE3,-Fast_Unaligned_Load,,-,"
"-ERMS,-AVX_Fast_Unaligned_Load,-,",
"-ERMS,-AVX_Fast_Unaligned_Load,-Avoid_Non_Temporal_Memset,-,",
test_1,
array_length (test_1)
}