Your message dated Sun, 01 Dec 2024 17:56:55 -0600
with message-id <15868802.tv2OnDr8pf@riemann>
and subject line Re: gmp: please add NEON-optimized variant for armhf
has caused the Debian Bug report #918314,
regarding gmp: please add NEON-optimized variant for armhf
to be marked as done.
This means that you claim that the problem has been dealt with.
If this is not the case it is now your responsibility to reopen the
Bug report if necessary, and/or fix the problem forthwith.
(NB: If you are a system administrator and have no idea what this
message is talking about, this may indicate a serious mail system
misconfiguration somewhere. Please contact ow...@bugs.debian.org
immediately.)
--
918314: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=918314
Debian Bug Tracking System
Contact ow...@bugs.debian.org with problems
--- Begin Message ---
Source: gmp
Version: 2:6.1.2+dfsg-1
Severity: wishlist
Tags: patch
Dear Maintainer,
As --enable-fat works only on i386^1, on armhf neon-optimized code is
never used, rendering gmp much slower than it can be.
As a workaround, I suggest to compile and install separate
neon-optimized library in $(libdir)/neon/vfp and rely on ld.so for
runtime-detection.
Debdiff attached, passed limited testing ([1] cross-compiled [implies
nocheck] `pbuilder --host-arch=armhf --arch=i386`, installed on
rpi3b+/raspbian-stretch, benchmarked; [2] native recompilation on
rpi3b+/raspbian [passed regression tests], then installed and benchmarked);
No idea how well this would work on hurd, kfreebsd, etc.
I hardcoded armv7 for neon code: debian's armhf requires minimum armv7,
so this should be acceptable; I have not tested, but this should not
break "fake armhf" from raspbian (rpi1 is armv6 without neon, so new
neon-optimized variant would not be picked; newer rpi are armv7+ and
have neon).
Potential pitfalls: it enables previously untested (at least, by debian)
code on affected platforms, some bugs can lurk there.
About expected user-visible effect, on Raspberry Pi 3B+ (BCM2837B0,
4-core Cortex-A53 @1.4GHz), `gnutls-cli --benchmark-tls-kx`:
Before:
Testing key exchanges (RSA/DH bits: 3072, EC bits: 256)
(TLS1.2)-(DHE-RSA-3072)-(AES-128-CBC)-(SHA1) 5.50 transactions/sec
(avg. handshake time: 181.71 ms, sample variance: 0.43)
(TLS1.2)-(ECDHE-RSA-SECP256R1)-(AES-128-CBC)-(SHA1) 12.08 transactions/sec
(avg. handshake time: 82.77 ms, sample variance: 0.18)
(TLS1.2)-(ECDHE-RSA-X25519)-(AES-128-CBC)-(SHA1) 12.32 transactions/sec
(avg. handshake time: 81.03 ms, sample variance: 0.03)
(TLS1.2)-(ECDHE-ECDSA-SECP256R1)-(AES-128-CBC)-(SHA1) 77.88
transactions/sec
(avg. handshake time: 12.73 ms, sample variance: 0.20)
(TLS1.2)-(ECDHE-ECDSA-X25519)-(AES-128-CBC)-(SHA1) 88.98 transactions/sec
(avg. handshake time: 11.13 ms, sample variance: 0.11)
(TLS1.2)-(RSA)-(AES-128-CBC)-(SHA1) 13.15 transactions/sec
(avg. handshake time: 75.86 ms, sample variance: 0.12)
After:
Testing key exchanges (RSA/DH bits: 3072, EC bits: 256)
(TLS1.2)-(DHE-RSA-3072)-(AES-128-CBC)-(SHA1) 8.40 transactions/sec
(avg. handshake time: 118.98 ms, sample variance: 0.27)
(TLS1.2)-(ECDHE-RSA-SECP256R1)-(AES-128-CBC)-(SHA1) 18.42 transactions/sec
(avg. handshake time: 54.23 ms, sample variance: 0.18)
(TLS1.2)-(ECDHE-RSA-X25519)-(AES-128-CBC)-(SHA1) 18.88 transactions/sec
(avg. handshake time: 52.82 ms, sample variance: 0.15)
(TLS1.2)-(ECDHE-ECDSA-SECP256R1)-(AES-128-CBC)-(SHA1) 93.89
transactions/sec
(avg. handshake time: 10.54 ms, sample variance: 0.25)
(TLS1.2)-(ECDHE-ECDSA-X25519)-(AES-128-CBC)-(SHA1) 106.83 transactions/sec
(avg. handshake time: 9.25 ms, sample variance: 0.19)
(TLS1.2)-(RSA)-(AES-128-CBC)-(SHA1) 20.46 transactions/sec
(avg. handshake time: 48.78 ms, sample variance: 0.18)
That is, 20% to 50% speedup.
^1 --enable-fat works on amd64 too - but debian disables it; maybe, it's
time to reconsider? some related bugs was fixed upstream since last
attempt (which resulted in #671866); that said, on my cpu amd64+fat is
slower than current debian-packaged "fat-free" code, so I'm not very
much interested.
-- System Information:
Debian Release: 9.6
APT prefers stable-updates
APT policy: (500, 'stable-updates'), (500, 'stable-debug'), (500,
'proposed-updates-debug'), (500, 'proposed-updates'), (500, 'stable')
Architecture: i386 (x86_64)
Foreign Architectures: amd64
Kernel: Linux 4.9.0-6-amd64 (SMP w/2 CPU cores)
Locale: LANG=ru_RU.KOI8-R, LC_CTYPE=ru_RU.KOI8-R (charmap=KOI8-R),
LANGUAGE=ru_RU.KOI8-R (charmap=KOI8-R)
Shell: /bin/sh linked to /bin/dash
Init: systemd (via /run/systemd/system)
diff -Nru gmp-6.1.2+dfsg/debian/rules gmp-6.1.2+dfsg/debian/rules
--- gmp-6.1.2+dfsg/debian/rules 2016-12-21 08:38:23.000000000 +0300
+++ gmp-6.1.2+dfsg/debian/rules 2019-01-02 22:52:33.000000000 +0300
@@ -68,9 +68,20 @@
confflags_ma = $(confflags) $(confflags_build)
--libdir=/usr/lib/$(DEB_HOST_MULTIARCH)
+FLAVORS = main
+LIBDIR_main =
+
CC = $(DEB_HOST_GNU_TYPE)-gcc
CXX = $(DEB_HOST_GNU_TYPE)-g++
+ifneq (,$(filter armhf, $(DEB_HOST_ARCH)))
+FLAVORS += neon
+
+LIBDIR_neon = neon/vfp
+neon_host_type = $(patsubst
arm-%,armcortexa7neon-unknown-%,$(DEB_HOST_GNU_TYPE))
+confflags_neon = --host=$(neon_host_type) --target=$(neon_host_type)
--libdir=/usr/lib/$(DEB_HOST_MULTIARCH)/$(LIBDIR_neon)
+CFLAGS_neon = -march=armv7-a -mfpu=neon
+endif
get-orig-source: gmp-$(ORIG_SRC_VERSION).tar.xz
tar --xz -xf $<
@@ -88,25 +99,34 @@
gmp-$(ORIG_SRC_VERSION).tar.xz:
wget https://gmplib.org/download/gmp/$@
-configure: configure-stamp
-configure-stamp:
- mkdir -p build
- cd build && ../configure $(confflags_ma) \
- AR=$(AR) CC="$(CC)" CFLAGS="$(CFLAGS)" \
- CXX="$(CXX)" CXXFLAGS="$(CXXFLAGS)"
+$(patsubst %,configure-%,$(FLAVORS)): configure-%: configure-stamp-%
+$(patsubst %,configure-stamp-%,$(FLAVORS)): configure-stamp-%:
+ mkdir -p build-$*
+ cd build-$* && ../configure $(confflags_ma) \
+ $(confflags_$*) \
+ AR=$(AR) CC="$(CC)" CFLAGS="$(CFLAGS) $(CFLAGS_$*)" \
+ CXX="$(CXX)" CXXFLAGS="$(CXXFLAGS) $(CFLAGS_$*)"
touch $@
-build: build-stamp
-build-stamp: configure
+$(patsubst %,build-%,$(FLAVORS)): build-%: build-stamp-%
+$(patsubst %,build-stamp-%,$(FLAVORS)): build-stamp-%: configure-%
dh_testdir
- $(MAKE) $(JOBSFLAG) -C build
- $(MAKE_CHECK) -C build
+ $(MAKE) $(JOBSFLAG) -C build-$*
+ $(MAKE_CHECK) -C build-$*
+ touch $@
+
+build: $(patsubst %,build-%,$(FLAVORS))
+build-stamp: $(patsubst %,build-stamp-%,$(FLAVORS))
touch $@
clean:
dh_testdir
dh_testroot
rm -rf build build-stamp
+ rm -rf $(patsubst %,build-%,$(FLAVORS))
+ rm -rf $(patsubst %,build-stamp-%,$(FLAVORS))
+ rm -rf configure-stamp
+ rm -rf $(patsubst %,configure-stamp-%,$(FLAVORS))
dh_clean
install-prep:
@@ -115,13 +135,17 @@
dh_prep
dh_installdirs
+$(patsubst %,install-%,$(FLAVORS)): install-%: build-stamp-%
+ $(MAKE) DESTDIR=`pwd`/debian/tmp
includeexecdir=/usr/include/$(DEB_HOST_MULTIARCH) -C build-$* install-exec
+ dh_install -plibgmp10 usr/lib/*/$(LIBDIR_$*)/libgmp.so.*
+
install: build-stamp install-prep
rm -rf debian/tmp
+ $(MAKE) -f debian/rules $(patsubst %,install-%,$(FLAVORS))
# Install places gmp.h in 'includeexecdir' which is non-standard and
cannot be set at compile time,
# so override it at install.
- $(MAKE) DESTDIR=`pwd`/debian/tmp
includeexecdir=/usr/include/$(DEB_HOST_MULTIARCH) -C build install
+ $(MAKE) DESTDIR=`pwd`/debian/tmp
includeexecdir=/usr/include/$(DEB_HOST_MULTIARCH) -C build-main install
- dh_install -plibgmp10 usr/lib/*/libgmp.so.*
dh_install -plibgmpxx4ldbl usr/lib/*/libgmpxx.so.*
dh_install -plibgmp-dev usr/lib/*/lib*.so
--- End Message ---
--- Begin Message ---
On Sat, 5 Jan 2019 01:45:40 +0300 "Yuriy M. Kaminskiy"
<yumkam+deb...@gmail.com> wrote:
> Source: gmp
> Version: 2:6.1.2+dfsg-1
> Severity: wishlist
> Tags: patch
>
> Dear Maintainer,
>
> As --enable-fat works only on i386^1, on armhf neon-optimized code is
> never used, rendering gmp much slower than it can be.
>
> As a workaround, I suggest to compile and install separate
> neon-optimized library in $(libdir)/neon/vfp and rely on ld.so for
> runtime-detection.
>
> Debdiff attached, passed limited testing ([1] cross-compiled [implies
> nocheck] `pbuilder --host-arch=armhf --arch=i386`, installed on
> rpi3b+/raspbian-stretch, benchmarked; [2] native recompilation on
> rpi3b+/raspbian [passed regression tests], then installed and benchmarked);
>
> No idea how well this would work on hurd, kfreebsd, etc.
>
> I hardcoded armv7 for neon code: debian's armhf requires minimum armv7,
> so this should be acceptable; I have not tested, but this should not
> break "fake armhf" from raspbian (rpi1 is armv6 without neon, so new
> neon-optimized variant would not be picked; newer rpi are armv7+ and
> have neon).
Please work with upstream on incorporating this change.
Best,
-Steve
signature.asc
Description: This is a digitally signed message part.
--- End Message ---
--
debian-science-maintainers mailing list
debian-science-maintainers@alioth-lists.debian.net
https://alioth-lists.debian.net/cgi-bin/mailman/listinfo/debian-science-maintainers
