Enable dispatch towards AVX512

.gitignore

@@ -1 +1,3 @@
 .osc
+*.tar
+xh

Makefile

@@ -0,0 +1,8 @@
+test: xh linux-6.6.1.tar linux-6.6.2.tar linux-6.6.3.tar linux-6.6.4.tar linux-6.6.5.tar
+	cat linux-6.*.tar{,,,,,,,} | time ./xh
+
+%.tar:
+	wget -O- https://cdn.kernel.org/pub/linux/kernel/v6.x/$@.xz | xz -cd >$@
+
+xh: xh.cpp
+	g++ -o xh xh.cpp -lxxhash -Wall -O2 -g

cpu-1135g7

@@ -0,0 +1,111 @@
+
+=== test ===
+
+standard openSUSE xxhash packages of today
+
+$ make
+cat linux-6.*.tar{,,,,,,,} | time ./xh
+5df8e93736b5a2bddd0b324a455b61fc
+3.24user 6.69system 0:13.69elapsed 72%CPU (0avgtext+0avgdata 3824maxresident)k
+0inputs+0outputs (0major+178minor)pagefaults 0swaps
+14:04 f3:../xx/xxhash $ make
+cat linux-6.*.tar{,,,,,,,} | time ./xh
+5df8e93736b5a2bddd0b324a455b61fc
+3.29user 6.05system 0:13.04elapsed 71%CPU (0avgtext+0avgdata 3720maxresident)k
+0inputs+0outputs (0major+178minor)pagefaults 0swaps
+14:05 f3:../xx/xxhash $ make
+cat linux-6.*.tar{,,,,,,,} | time ./xh
+5df8e93736b5a2bddd0b324a455b61fc
+3.20user 6.29system 0:13.19elapsed 72%CPU (0avgtext+0avgdata 3664maxresident)k
+0inputs+0outputs (0major+177minor)pagefaults 0swaps
+14:05 f3:../xx/xxhash $ make
+cat linux-6.*.tar{,,,,,,,} | time ./xh
+5df8e93736b5a2bddd0b324a455b61fc
+3.38user 5.91system 0:12.89elapsed 72%CPU (0avgtext+0avgdata 3712maxresident)k
+0inputs+0outputs (0major+178minor)pagefaults 0swaps
+
+Mean/Spread: 13.20 ± 0.31 sec (2.3% error)
+
+
+=== switch packages ===
+
+# rpm -U ...
+
+
+=== verify AVX512 is now in use ===
+
+Watch this space closely.
+
+$ gdb xh
+Reading symbols from xh...
+(gdb) b main
+Breakpoint 1 at 0x4010b0: file xh.cpp, line 7.
+(gdb) r <xh.cpp
+Starting program: /home/jengelh/obs_nosave/zu/home/jengelh/xx/xxhash/xh <xh.cpp
+[Thread debugging using libthread_db enabled]
+Using host libthread_db library "/lib64/libthread_db.so.1".
+
+Breakpoint 1, main () at xh.cpp:7
+7               auto state = XXH3_createState();
+(gdb) b XXH_setDispatch
+Breakpoint 2 at 0x7ffff7f9a6f4: file /usr/src/debug/xxHash-0.8.2/xxh_x86dispatch.c, line 706.
+(gdb) r
+Restart?-yes
+Starting program: /home/jengelh/obs_nosave/zu/home/jengelh/xx/xxhash/xh <xh.cpp
+[Thread debugging using libthread_db enabled]
+Using host libthread_db library "/lib64/libthread_db.so.1".
+
+Breakpoint 2, XXH_setDispatch () at /usr/src/debug/xxHash-0.8.2/xxh_x86dispatch.c:706
+706     {
+(gdb) bt
+#0  XXH_setDispatch () at /usr/src/debug/xxHash-0.8.2/xxh_x86dispatch.c:706
+#1  0x00007ffff7fca53e in call_init (env=0x7fffffffd998, argv=0x7fffffffd988, argc=1, l=<optimized out>) at dl-init.c:74
+#2  call_init (l=<optimized out>, argc=1, argv=0x7fffffffd988, env=0x7fffffffd998) at dl-init.c:26
+#3  0x00007ffff7fca63c in _dl_init (main_map=0x7ffff7fbe000, argc=1, argv=0x7fffffffd988, env=0x7fffffffd998) at dl-init.c:121
+#4  0x00007ffff7fe2ba0 in _dl_start_user () from /lib64/ld-linux-x86-64.so.2
+#5  0x0000000000000001 in ?? ()
+#6  0x00007fffffffde8c in ?? ()
+#7  0x0000000000000000 in ?? ()
+(gdb) n
+707         int vecID = XXH_featureTest();
+(gdb) 
+719         XXH_g_dispatch = XXH_kDispatch[vecID];
+(gdb) p vecID
+$1 = 3  [= XXH_AVX512!]
+
+
+=== test ===
+
+$ make test
+cat linux-6.*.tar{,,,,,,,} | time ./xh
+5df8e93736b5a2bddd0b324a455b61fc
+3.37user 6.16system 0:13.23elapsed 72%CPU (0avgtext+0avgdata 3708maxresident)k
+0inputs+0outputs (0major+177minor)pagefaults 0swaps
+14:07 f3:../xx/xxhash $ make test
+cat linux-6.*.tar{,,,,,,,} | time ./xh
+5df8e93736b5a2bddd0b324a455b61fc
+3.23user 6.27system 0:13.27elapsed 71%CPU (0avgtext+0avgdata 3664maxresident)k
+0inputs+0outputs (0major+176minor)pagefaults 0swaps
+14:07 f3:../xx/xxhash $ make test
+cat linux-6.*.tar{,,,,,,,} | time ./xh
+5df8e93736b5a2bddd0b324a455b61fc
+3.26user 6.77system 0:13.75elapsed 73%CPU (0avgtext+0avgdata 3712maxresident)k
+0inputs+0outputs (0major+178minor)pagefaults 0swaps
+14:07 f3:../xx/xxhash $ make test
+cat linux-6.*.tar{,,,,,,,} | time ./xh
+5df8e93736b5a2bddd0b324a455b61fc
+3.19user 6.07system 0:12.96elapsed 71%CPU (0avgtext+0avgdata 3712maxresident)k
+0inputs+0outputs (0major+178minor)pagefaults 0swaps
+
+Mean/Spread: 13.30 ± 0.34s (2.5% error)
+
+== Lessons learned ==
+
+Prerequisite for good measurement:
+ * consistent(!) error rate  (4 individual measurements is a bit on the low side, but ... anyway)
+
+Result:
+ * The observed 13.2s mean is within the margin of measurement error of the 13.3 mean and viceversa
+
+Conclusion:
+ * DISPATCH=1 made no impact on this machine

xh.cpp

@@ -0,0 +1,18 @@
+#include <xxhash.h>
+#include <unistd.h>
+#include <cstdio>
+
+int main()
+{
+	auto state = XXH3_createState();
+	XXH3_128bits_reset(state);
+	char buffer[1048576];
+	ssize_t count;
+	while ((count = read(STDIN_FILENO, buffer, sizeof(buffer))) > 0)
+		XXH3_128bits_update(state, buffer, count);
+	XXH128_canonical_t canon{};
+	XXH128_canonicalFromHash(&canon, XXH3_128bits_digest(state));
+	for (auto u : canon.digest)
+		printf("%02x", u);
+	printf("\n");
+}

xxhash.spec

@@ -60,9 +60,10 @@ Headers and other development files for xxHash.
 
 %build
 # ALLOW_AVX just means "we guarantee we policed our %%optflags".
-export CFLAGS="%{optflags} -DXXH_X86DISPATCH_ALLOW_AVX=1"
+export CFLAGS="%{optflags} -DXXH_X86DISPATCH_ALLOW_AVX=1 -Og"
 export CXXFLAGS="$CFLAGS"
 export LDFLAGS="%{?build_ldflags}"
+export DISPATCH=1
 # DISPATCH=1 if you want AVX2/AVX512. But it does not seem to perform any
 # better than the lowest-denomimation code on at least the 1135G7 and 5950X
 # CPUs, and for both LP64 as well as ILP32 — it seems to be all within margin
@@ -70,16 +71,18 @@ export LDFLAGS="%{?build_ldflags}"
 %make_build prefix=%{_prefix} libdir=%{_libdir}
 
 %install
-export CFLAGS="%{optflags} -DXXH_X86DISPATCH_ALLOW_AVX=1"
+export CFLAGS="%{optflags} -DXXH_X86DISPATCH_ALLOW_AVX=1 -Og"
 export CXXFLAGS="%{optflags}"
 export LDFLAGS="%{?build_ldflags}"
+export DISPATCH=1
 %make_install prefix=%{_prefix} libdir=%{_libdir}
 rm -rf %{buildroot}%{_libdir}/libxxhash.a
 
 %check
-export CFLAGS="%{optflags}"
+export CFLAGS="%{optflags} -DXXH_X86DISPATCH_ALLOW_AVX=1"
 export CXXFLAGS="%{optflags}"
 export LDFLAGS="%{?build_ldflags}"
+export DISPATCH=1
 # not safe for parallel execution as it removes xxhash.o and recreates it with different flags
 # the list is taken from test-all with non-working/irrelevant ones (such as ones that change the toolchain) removed
 %make_build -j1 test test-unicode listL120 trailingWhitespace test-xxh-nnn-sums
@@ -102,8 +105,7 @@ export LDFLAGS="%{?build_ldflags}"
 %{_libdir}/libxxhash.so.*
 
 %files devel
-%{_includedir}/xxhash.h
-%{_includedir}/xxh3.h
+%{_includedir}/*
 %{_libdir}/pkgconfig/libxxhash.pc
 %{_libdir}/libxxhash.so