x86: Add new cpu-flag `Prefer_Non_Temporal`
Noah Goldstein
goldstein.w.n@gmail.com
Tue Aug 13 15:42:08 GMT 2024
On Tue, Aug 13, 2024 at 11:33 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>
> On Tue, Aug 13, 2024 at 8:24 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
> >
> > On Tue, Aug 13, 2024 at 11:12 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> > >
> > > On Tue, Aug 13, 2024 at 8:05 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
> > > >
> > > > On Tue, Aug 13, 2024 at 10:45 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> > > > >
> > > > > On Tue, Aug 13, 2024 at 6:55 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
> > > > > >
> > > > > > The goal of this flag is to allow targets which don't prefer/have ERMS
> > > > > > to still access the non-temporal memset implementation.
> > > > > > ---
> > > > > > sysdeps/x86/cpu-tunables.c | 4 ++
> > > > > > sysdeps/x86/dl-cacheinfo.h | 46 ++++++++++++++-----
> > > > > > ...cpu-features-preferred_feature_index_1.def | 1 +
> > > > > > sysdeps/x86/tst-hwcap-tunables.c | 6 ++-
> > > > > > sysdeps/x86_64/multiarch/ifunc-memset.h | 19 ++++++--
> > > > > > 5 files changed, 58 insertions(+), 18 deletions(-)
> > > > > >
> > > > > > diff --git a/sysdeps/x86/cpu-tunables.c b/sysdeps/x86/cpu-tunables.c
> > > > > > index ccc6b64dc2..b01b703517 100644
> > > > > > --- a/sysdeps/x86/cpu-tunables.c
> > > > > > +++ b/sysdeps/x86/cpu-tunables.c
> > > > > > @@ -255,6 +255,10 @@ TUNABLE_CALLBACK (set_hwcaps) (tunable_val_t *valp)
> > > > > > (n, cpu_features, Prefer_PMINUB_for_stringop, SSE2, 26);
> > > > > > }
> > > > > > break;
> > > > > > + case 32:
> > > > > > + CHECK_GLIBC_IFUNC_PREFERRED_BOTH (n, cpu_features,
> > > > > > + Prefer_Non_Temporal_Without_ERMS, 32);
> > > > > > + break;
> > > > > > }
> > > > > > }
> > > > > > }
> > > > > > diff --git a/sysdeps/x86/dl-cacheinfo.h b/sysdeps/x86/dl-cacheinfo.h
> > > > > > index a1c03b8903..600bd66a93 100644
> > > > > > --- a/sysdeps/x86/dl-cacheinfo.h
> > > > > > +++ b/sysdeps/x86/dl-cacheinfo.h
> > > > > > @@ -988,14 +988,6 @@ dl_init_cacheinfo (struct cpu_features *cpu_features)
> > > > > > if (CPU_FEATURE_USABLE_P (cpu_features, FSRM))
> > > > > > rep_movsb_threshold = 2112;
> > > > > >
> > > > > > - /* 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_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). */
> > > > > > @@ -1017,6 +1009,15 @@ dl_init_cacheinfo (struct cpu_features *cpu_features)
> > > > > > if (tunable_size != 0)
> > > > > > shared = tunable_size;
> > > > > >
> > > > > > + /* 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_ARCH_P (cpu_features, Avoid_Non_Temporal_Memset)
> > > > > > + && (CPU_FEATURES_ARCH_P (cpu_features, Prefer_Non_Temporal_Without_ERMS)
> > > > > > + || cpu_features->basic.kind == arch_kind_intel
> > > > > > + || cpu_features->basic.kind == arch_kind_amd))
> > > > >
> > > > > Please add Prefer_Non_Temporal instead and enable it for Intel and AMD.
> > > > >
> > > > > > + memset_non_temporal_threshold = non_temporal_threshold;
> > > > > > +
> > > > > > tunable_size = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
> > > > > > if (tunable_size > minimum_non_temporal_threshold
> > > > > > && tunable_size <= maximum_non_temporal_threshold)
> > > > > > @@ -1042,14 +1043,37 @@ dl_init_cacheinfo (struct cpu_features *cpu_features)
> > > > > > slightly better than ERMS. */
> > > > > > rep_stosb_threshold = SIZE_MAX;
> > > > > >
> > > > > > + /*
> > > > > > + For memset, the non-temporal implementation is only accessed through the
> > > > > > + stosb code. ie:
> > > > > > + ```
> > > > > > + if (size >= rep_stosb_thresh)
> > > > > > + {
> > > > > > + if (size >= non_temporal_thresh)
> > > > > > + {
> > > > > > + do_non_temporal ();
> > > > > > + }
> > > > > > + do_stosb ();
> > > > > > + }
> > > > > > + do_normal_vec_loop ();
> > > > > > + ```
> > > > > > + So if we prefer non-temporal, set `rep_stosb_thresh = non_temporal_thresh`
> > > > > > + to enable the implementation. If `rep_stosb_thresh = non_temporal_thresh`,
> > > > > > + `rep stosb` will never be used.
> > > > > > + */
> > > > > > + TUNABLE_SET_WITH_BOUNDS (x86_memset_non_temporal_threshold,
> > > > > > + memset_non_temporal_threshold,
> > > > > > + minimum_non_temporal_threshold, SIZE_MAX);
> > > > > > + if (CPU_FEATURES_ARCH_P (cpu_features, Prefer_Non_Temporal_Without_ERMS))
> > > > >
> > > > > Please add another bit, Avoid_STOSB, and check it instead.
> > > > >
> > > > > > + rep_stosb_threshold
> > > > > > + = TUNABLE_GET (x86_memset_non_temporal_threshold, long int, NULL);
> > > > >
> > > > > Please move and reuse the existing TUNABLE_GET
> > > > > (x86_memset_non_temporal_threshold, long int, NULL)
> > > > > call.
Would prefer to leave as is. We need the final values to be equal
and it seems less bug prone to use the final value as opposed to
having a seperate bounds check here.
> > > > >
> > > > > > +
> > > > > > +
> > > > > > TUNABLE_SET_WITH_BOUNDS (x86_data_cache_size, data, 0, SIZE_MAX);
> > > > > > TUNABLE_SET_WITH_BOUNDS (x86_shared_cache_size, shared, 0, SIZE_MAX);
> > > > > > TUNABLE_SET_WITH_BOUNDS (x86_non_temporal_threshold, non_temporal_threshold,
> > > > > > minimum_non_temporal_threshold,
> > > > > > maximum_non_temporal_threshold);
> > > > > > - TUNABLE_SET_WITH_BOUNDS (x86_memset_non_temporal_threshold,
> > > > > > - memset_non_temporal_threshold,
> > > > > > - minimum_non_temporal_threshold, SIZE_MAX);
> > > > >
> > > > > Please leave it here.
> > > > >
> > > > > > TUNABLE_SET_WITH_BOUNDS (x86_rep_movsb_threshold, rep_movsb_threshold,
> > > > > > minimum_rep_movsb_threshold, SIZE_MAX);
> > > > > > TUNABLE_SET_WITH_BOUNDS (x86_rep_stosb_threshold, rep_stosb_threshold, 1,
> > > > > > diff --git a/sysdeps/x86/include/cpu-features-preferred_feature_index_1.def b/sysdeps/x86/include/cpu-features-preferred_feature_index_1.def
> > > > > > index 61bbbc2e89..60a3a382ff 100644
> > > > > > --- a/sysdeps/x86/include/cpu-features-preferred_feature_index_1.def
> > > > > > +++ b/sysdeps/x86/include/cpu-features-preferred_feature_index_1.def
> > > > > > @@ -34,3 +34,4 @@ BIT (MathVec_Prefer_No_AVX512)
> > > > > > BIT (Prefer_FSRM)
> > > > > > BIT (Avoid_Short_Distance_REP_MOVSB)
> > > > > > BIT (Avoid_Non_Temporal_Memset)
> > > > > > +BIT (Prefer_Non_Temporal_Without_ERMS)
> > > > > > \ No newline at end of file
> > > > >
> > > > > Please add a newline.
> > > > >
> > > > > > diff --git a/sysdeps/x86/tst-hwcap-tunables.c b/sysdeps/x86/tst-hwcap-tunables.c
> > > > > > index 94307283d7..4495d795e4 100644
> > > > > > --- a/sysdeps/x86/tst-hwcap-tunables.c
> > > > > > +++ b/sysdeps/x86/tst-hwcap-tunables.c
> > > > > > @@ -60,7 +60,8 @@ 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,-Avoid_Non_Temporal_Memset",
> > > > > > + "-AVX_Fast_Unaligned_Load,-Avoid_Non_Temporal_Memset,"
> > > > > > + "-Prefer_Non_Temporal_Without_ERMS",
> > > > > > test_1,
> > > > > > array_length (test_1)
> > > > > > },
> > > > > > @@ -68,7 +69,8 @@ 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,-Avoid_Non_Temporal_Memset,-,",
> > > > > > + "-ERMS,-AVX_Fast_Unaligned_Load,-Avoid_Non_Temporal_Memset,"
> > > > > > + "-Prefer_Non_Temporal_Without_ERMS,-,",
> > > > > > test_1,
> > > > > > array_length (test_1)
> > > > > > }
> > > > > > diff --git a/sysdeps/x86_64/multiarch/ifunc-memset.h b/sysdeps/x86_64/multiarch/ifunc-memset.h
> > > > > > index 7a637ef7ca..3ca64bb853 100644
> > > > > > --- a/sysdeps/x86_64/multiarch/ifunc-memset.h
> > > > > > +++ b/sysdeps/x86_64/multiarch/ifunc-memset.h
> > > > > > @@ -61,7 +61,9 @@ IFUNC_SELECTOR (void)
> > > > > > && X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, AVX512BW)
> > > > > > && X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, BMI2))
> > > > > > {
> > > > > > - if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))
> > > > > > + if (CPU_FEATURE_USABLE_P (cpu_features, ERMS)
> > > > > > + || CPU_FEATURES_ARCH_P (cpu_features,
> > > > > > + Prefer_Non_Temporal_Without_ERMS))
> > > > >
> > > > > Please check only Prefer_Non_Temporal instead.
> > > > >
> > > > > > return OPTIMIZE (avx512_unaligned_erms);
> > > > > >
> > > > > > return OPTIMIZE (avx512_unaligned);
> > > > > > @@ -76,7 +78,9 @@ IFUNC_SELECTOR (void)
> > > > > > && X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, AVX512BW)
> > > > > > && X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, BMI2))
> > > > > > {
> > > > > > - if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))
> > > > > > + if (CPU_FEATURE_USABLE_P (cpu_features, ERMS)
> > > > > > + || CPU_FEATURES_ARCH_P (cpu_features,
> > > > > > + Prefer_Non_Temporal_Without_ERMS))
> > > > > > return OPTIMIZE (evex_unaligned_erms);
> > > > >
> > > > > Please check only Prefer_Non_Temporal instead.
> > > > >
> > > > > >
> > > > > > return OPTIMIZE (evex_unaligned);
> > > > > > @@ -84,7 +88,9 @@ IFUNC_SELECTOR (void)
> > > > > >
> > > > > > if (CPU_FEATURE_USABLE_P (cpu_features, RTM))
> > > > > > {
> > > > > > - if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))
> > > > > > + if (CPU_FEATURE_USABLE_P (cpu_features, ERMS)
> > > > > > + || CPU_FEATURES_ARCH_P (cpu_features,
> > > > > > + Prefer_Non_Temporal_Without_ERMS))
> > > > > > return OPTIMIZE (avx2_unaligned_erms_rtm);
> > > > >
> > > > > Please check only Prefer_Non_Temporal instead.
> > > > >
> > > > > >
> > > > > > return OPTIMIZE (avx2_unaligned_rtm);
> > > > > > @@ -93,14 +99,17 @@ IFUNC_SELECTOR (void)
> > > > > > if (X86_ISA_CPU_FEATURES_ARCH_P (cpu_features,
> > > > > > Prefer_No_VZEROUPPER, !))
> > > > > > {
> > > > > > - if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))
> > > > > > + if (CPU_FEATURE_USABLE_P (cpu_features, ERMS)
> > > > > > + || CPU_FEATURES_ARCH_P (cpu_features,
> > > > > > + Prefer_Non_Temporal_Without_ERMS))
> > > > > > return OPTIMIZE (avx2_unaligned_erms);
> > > > >
> > > > > Please check only Prefer_Non_Temporal instead.
> > > > >
> > > > > >
> > > > > > return OPTIMIZE (avx2_unaligned);
> > > > > > }
> > > > > > }
> > > > > >
> > > > > > - if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))
> > > > > > + if (CPU_FEATURE_USABLE_P (cpu_features, ERMS)
> > > > > > + || CPU_FEATURES_ARCH_P (cpu_features, Prefer_Non_Temporal_Without_ERMS))
> > > > > > return OPTIMIZE (sse2_unaligned_erms);
> > > > >
> > > > > Please check only Prefer_Non_Temporal instead.
> > > > >
> > > >
> > > > Hmm? Why remove the ERMS check here? Shouldn't we want to select
> > > > the `_erms` implemention if we want either non-temporal OR erms?
> > >
> > > You said NT store must go through ERMS. You can do
> > >
> > > if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))
> > > Set Prefer_Non_Temporal
> > >
> > > for Intel and AMD.
> >
> > What about machines that want ERMS but not NT?
>
> We can check CPUID to turn off Prefer_Non_Temporal.
This seems overly complicated. We have a code path that goes
to ERMS or NT. The natural check for that seems to be `ERMS || NT`.
>
> > >
> > > >
> > > > Everything else sounds fine though.
> > > >
> > > > > >
> > > > > > return OPTIMIZE (sse2_unaligned);
> > > > > > --
> > > > > > 2.34.1
> > > > > >
> > > > >
> > > > >
> > > > > --
> > > > > H.J.
> > >
> > >
> > >
> > > --
> > > H.J.
>
>
>
> --
> H.J.
More information about the Libc-alpha
mailing list