Optimize pg_popcount() with AVX-512 instructions.
authorNathan Bossart <[email protected]>
Sun, 7 Apr 2024 02:56:23 +0000 (21:56 -0500)
committerNathan Bossart <[email protected]>
Sun, 7 Apr 2024 02:56:23 +0000 (21:56 -0500)
Presently, pg_popcount() processes data in 32-bit or 64-bit chunks
when possible.  Newer hardware that supports AVX-512 instructions
can use 512-bit chunks, which provides a nice speedup, especially
for larger buffers.  This commit introduces the infrastructure
required to detect compiler and CPU support for the required
AVX-512 intrinsic functions, and it adds a new pg_popcount()
implementation that uses these functions.  If CPU support for this
optimized implementation is detected at runtime, a function pointer
is updated so that it is used by subsequent calls to pg_popcount().

Most of the existing in-tree calls to pg_popcount() should benefit
from these instructions, and calls with smaller buffers should at
least not regress compared to v16.  The new infrastructure
introduced by this commit can also be used to optimize
visibilitymap_count(), but that is left for a follow-up commit.

Co-authored-by: Paul Amonson, Ants Aasma
Reviewed-by: Matthias van de Meent, Tom Lane, Noah Misch, Akash Shankaran, Alvaro Herrera, Andres Freund, David Rowley
Discussion: https://postgr.es/m/BL1PR11MB5304097DF7EA81D04C33F3D1DCA6A%40BL1PR11MB5304.namprd11.prod.outlook.com

15 files changed:
config/c-compiler.m4
configure
configure.ac
meson.build
src/Makefile.global.in
src/include/pg_config.h.in
src/include/port/pg_bitutils.h
src/makefiles/meson.build
src/port/Makefile
src/port/meson.build
src/port/pg_bitutils.c
src/port/pg_popcount_avx512.c [new file with mode: 0644]
src/port/pg_popcount_avx512_choose.c [new file with mode: 0644]
src/test/regress/expected/bit.out
src/test/regress/sql/bit.sql

index 3268a780bb059a043cdac6f453d76b0afbcf1e43..10f8c7bd0a9fbb12a2e7c0060560fe385dd9d01b 100644 (file)
@@ -694,3 +694,61 @@ if test x"$Ac_cachevar" = x"yes"; then
 fi
 undefine([Ac_cachevar])dnl
 ])# PGAC_LOONGARCH_CRC32C_INTRINSICS
+
+# PGAC_XSAVE_INTRINSICS
+# ---------------------
+# Check if the compiler supports the XSAVE instructions using the _xgetbv
+# intrinsic function.
+#
+# An optional compiler flag can be passed as argument (e.g., -mxsave).  If the
+# intrinsic is supported, sets pgac_xsave_intrinsics and CFLAGS_XSAVE.
+AC_DEFUN([PGAC_XSAVE_INTRINSICS],
+[define([Ac_cachevar], [AS_TR_SH([pgac_cv_xsave_intrinsics_$1])])dnl
+AC_CACHE_CHECK([for _xgetbv with CFLAGS=$1], [Ac_cachevar],
+[pgac_save_CFLAGS=$CFLAGS
+CFLAGS="$pgac_save_CFLAGS $1"
+AC_LINK_IFELSE([AC_LANG_PROGRAM([#include <immintrin.h>],
+  [return _xgetbv(0) & 0xe0;])],
+  [Ac_cachevar=yes],
+  [Ac_cachevar=no])
+CFLAGS="$pgac_save_CFLAGS"])
+if test x"$Ac_cachevar" = x"yes"; then
+  CFLAGS_XSAVE="$1"
+  pgac_xsave_intrinsics=yes
+fi
+undefine([Ac_cachevar])dnl
+])# PGAC_XSAVE_INTRINSICS
+
+# PGAC_AVX512_POPCNT_INTRINSICS
+# -----------------------------
+# Check if the compiler supports the AVX-512 popcount instructions using the
+# _mm512_setzero_si512, _mm512_maskz_loadu_epi8, _mm512_popcnt_epi64,
+# _mm512_add_epi64, and _mm512_reduce_add_epi64 intrinsic functions.
+#
+# Optional compiler flags can be passed as argument (e.g., -mavx512vpopcntdq
+# -mavx512bw).  If the intrinsics are supported, sets
+# pgac_avx512_popcnt_intrinsics and CFLAGS_POPCNT.
+AC_DEFUN([PGAC_AVX512_POPCNT_INTRINSICS],
+[define([Ac_cachevar], [AS_TR_SH([pgac_cv_avx512_popcnt_intrinsics_$1])])dnl
+AC_CACHE_CHECK([for _mm512_popcnt_epi64 with CFLAGS=$1], [Ac_cachevar],
+[pgac_save_CFLAGS=$CFLAGS
+CFLAGS="$pgac_save_CFLAGS $1"
+AC_LINK_IFELSE([AC_LANG_PROGRAM([#include <immintrin.h>],
+  [const char buf@<:@sizeof(__m512i)@:>@;
+   PG_INT64_TYPE popcnt = 0;
+   __m512i accum = _mm512_setzero_si512();
+   const __m512i val = _mm512_maskz_loadu_epi8((__mmask64) 0xf0f0f0f0f0f0f0f0, (const __m512i *) buf);
+   const __m512i cnt = _mm512_popcnt_epi64(val);
+   accum = _mm512_add_epi64(accum, cnt);
+   popcnt = _mm512_reduce_add_epi64(accum);
+   /* return computed value, to prevent the above being optimized away */
+   return popcnt == 0;])],
+  [Ac_cachevar=yes],
+  [Ac_cachevar=no])
+CFLAGS="$pgac_save_CFLAGS"])
+if test x"$Ac_cachevar" = x"yes"; then
+  CFLAGS_POPCNT="$1"
+  pgac_avx512_popcnt_intrinsics=yes
+fi
+undefine([Ac_cachevar])dnl
+])# PGAC_AVX512_POPCNT_INTRINSICS
index 36feeafbb23a6ead538c446803ade19dac431493..cfbd2a096f4d29f516413a5b15d2a52edd68ff3d 100755 (executable)
--- a/configure
+++ b/configure
@@ -647,6 +647,9 @@ MSGFMT_FLAGS
 MSGFMT
 PG_CRC32C_OBJS
 CFLAGS_CRC
+PG_POPCNT_OBJS
+CFLAGS_POPCNT
+CFLAGS_XSAVE
 LIBOBJS
 OPENSSL
 ZSTD
@@ -17404,6 +17407,40 @@ $as_echo "#define HAVE__GET_CPUID 1" >>confdefs.h
 
 fi
 
+{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for __get_cpuid_count" >&5
+$as_echo_n "checking for __get_cpuid_count... " >&6; }
+if ${pgac_cv__get_cpuid_count+:} false; then :
+  $as_echo_n "(cached) " >&6
+else
+  cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+/* end confdefs.h.  */
+#include <cpuid.h>
+int
+main ()
+{
+unsigned int exx[4] = {0, 0, 0, 0};
+  __get_cpuid_count(7, 0, &exx[0], &exx[1], &exx[2], &exx[3]);
+
+  ;
+  return 0;
+}
+_ACEOF
+if ac_fn_c_try_link "$LINENO"; then :
+  pgac_cv__get_cpuid_count="yes"
+else
+  pgac_cv__get_cpuid_count="no"
+fi
+rm -f core conftest.err conftest.$ac_objext \
+    conftest$ac_exeext conftest.$ac_ext
+fi
+{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $pgac_cv__get_cpuid_count" >&5
+$as_echo "$pgac_cv__get_cpuid_count" >&6; }
+if test x"$pgac_cv__get_cpuid_count" = x"yes"; then
+
+$as_echo "#define HAVE__GET_CPUID_COUNT 1" >>confdefs.h
+
+fi
+
 { $as_echo "$as_me:${as_lineno-$LINENO}: checking for __cpuid" >&5
 $as_echo_n "checking for __cpuid... " >&6; }
 if ${pgac_cv__cpuid+:} false; then :
@@ -17438,6 +17475,221 @@ $as_echo "#define HAVE__CPUID 1" >>confdefs.h
 
 fi
 
+{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for __cpuidex" >&5
+$as_echo_n "checking for __cpuidex... " >&6; }
+if ${pgac_cv__cpuidex+:} false; then :
+  $as_echo_n "(cached) " >&6
+else
+  cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+/* end confdefs.h.  */
+#include <intrin.h>
+int
+main ()
+{
+unsigned int exx[4] = {0, 0, 0, 0};
+  __get_cpuidex(exx[0], 7, 0);
+
+  ;
+  return 0;
+}
+_ACEOF
+if ac_fn_c_try_link "$LINENO"; then :
+  pgac_cv__cpuidex="yes"
+else
+  pgac_cv__cpuidex="no"
+fi
+rm -f core conftest.err conftest.$ac_objext \
+    conftest$ac_exeext conftest.$ac_ext
+fi
+{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $pgac_cv__cpuidex" >&5
+$as_echo "$pgac_cv__cpuidex" >&6; }
+if test x"$pgac_cv__cpuidex" = x"yes"; then
+
+$as_echo "#define HAVE__CPUIDEX 1" >>confdefs.h
+
+fi
+
+# Check for XSAVE intrinsics
+#
+CFLAGS_XSAVE=""
+{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for _xgetbv with CFLAGS=" >&5
+$as_echo_n "checking for _xgetbv with CFLAGS=... " >&6; }
+if ${pgac_cv_xsave_intrinsics_+:} false; then :
+  $as_echo_n "(cached) " >&6
+else
+  pgac_save_CFLAGS=$CFLAGS
+CFLAGS="$pgac_save_CFLAGS "
+cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+/* end confdefs.h.  */
+#include <immintrin.h>
+int
+main ()
+{
+return _xgetbv(0) & 0xe0;
+  ;
+  return 0;
+}
+_ACEOF
+if ac_fn_c_try_link "$LINENO"; then :
+  pgac_cv_xsave_intrinsics_=yes
+else
+  pgac_cv_xsave_intrinsics_=no
+fi
+rm -f core conftest.err conftest.$ac_objext \
+    conftest$ac_exeext conftest.$ac_ext
+CFLAGS="$pgac_save_CFLAGS"
+fi
+{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $pgac_cv_xsave_intrinsics_" >&5
+$as_echo "$pgac_cv_xsave_intrinsics_" >&6; }
+if test x"$pgac_cv_xsave_intrinsics_" = x"yes"; then
+  CFLAGS_XSAVE=""
+  pgac_xsave_intrinsics=yes
+fi
+
+if test x"$pgac_xsave_intrinsics" != x"yes"; then
+  { $as_echo "$as_me:${as_lineno-$LINENO}: checking for _xgetbv with CFLAGS=-mxsave" >&5
+$as_echo_n "checking for _xgetbv with CFLAGS=-mxsave... " >&6; }
+if ${pgac_cv_xsave_intrinsics__mxsave+:} false; then :
+  $as_echo_n "(cached) " >&6
+else
+  pgac_save_CFLAGS=$CFLAGS
+CFLAGS="$pgac_save_CFLAGS -mxsave"
+cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+/* end confdefs.h.  */
+#include <immintrin.h>
+int
+main ()
+{
+return _xgetbv(0) & 0xe0;
+  ;
+  return 0;
+}
+_ACEOF
+if ac_fn_c_try_link "$LINENO"; then :
+  pgac_cv_xsave_intrinsics__mxsave=yes
+else
+  pgac_cv_xsave_intrinsics__mxsave=no
+fi
+rm -f core conftest.err conftest.$ac_objext \
+    conftest$ac_exeext conftest.$ac_ext
+CFLAGS="$pgac_save_CFLAGS"
+fi
+{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $pgac_cv_xsave_intrinsics__mxsave" >&5
+$as_echo "$pgac_cv_xsave_intrinsics__mxsave" >&6; }
+if test x"$pgac_cv_xsave_intrinsics__mxsave" = x"yes"; then
+  CFLAGS_XSAVE="-mxsave"
+  pgac_xsave_intrinsics=yes
+fi
+
+fi
+if test x"$pgac_xsave_intrinsics" = x"yes"; then
+
+$as_echo "#define HAVE_XSAVE_INTRINSICS 1" >>confdefs.h
+
+fi
+
+
+# Check for AVX-512 popcount intrinsics
+#
+CFLAGS_POPCNT=""
+PG_POPCNT_OBJS=""
+if test x"$host_cpu" = x"x86_64"; then
+  { $as_echo "$as_me:${as_lineno-$LINENO}: checking for _mm512_popcnt_epi64 with CFLAGS=" >&5
+$as_echo_n "checking for _mm512_popcnt_epi64 with CFLAGS=... " >&6; }
+if ${pgac_cv_avx512_popcnt_intrinsics_+:} false; then :
+  $as_echo_n "(cached) " >&6
+else
+  pgac_save_CFLAGS=$CFLAGS
+CFLAGS="$pgac_save_CFLAGS "
+cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+/* end confdefs.h.  */
+#include <immintrin.h>
+int
+main ()
+{
+const char buf[sizeof(__m512i)];
+   PG_INT64_TYPE popcnt = 0;
+   __m512i accum = _mm512_setzero_si512();
+   const __m512i val = _mm512_maskz_loadu_epi8((__mmask64) 0xf0f0f0f0f0f0f0f0, (const __m512i *) buf);
+   const __m512i cnt = _mm512_popcnt_epi64(val);
+   accum = _mm512_add_epi64(accum, cnt);
+   popcnt = _mm512_reduce_add_epi64(accum);
+   /* return computed value, to prevent the above being optimized away */
+   return popcnt == 0;
+  ;
+  return 0;
+}
+_ACEOF
+if ac_fn_c_try_link "$LINENO"; then :
+  pgac_cv_avx512_popcnt_intrinsics_=yes
+else
+  pgac_cv_avx512_popcnt_intrinsics_=no
+fi
+rm -f core conftest.err conftest.$ac_objext \
+    conftest$ac_exeext conftest.$ac_ext
+CFLAGS="$pgac_save_CFLAGS"
+fi
+{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $pgac_cv_avx512_popcnt_intrinsics_" >&5
+$as_echo "$pgac_cv_avx512_popcnt_intrinsics_" >&6; }
+if test x"$pgac_cv_avx512_popcnt_intrinsics_" = x"yes"; then
+  CFLAGS_POPCNT=""
+  pgac_avx512_popcnt_intrinsics=yes
+fi
+
+  if test x"$pgac_avx512_popcnt_intrinsics" != x"yes"; then
+    { $as_echo "$as_me:${as_lineno-$LINENO}: checking for _mm512_popcnt_epi64 with CFLAGS=-mavx512vpopcntdq -mavx512bw" >&5
+$as_echo_n "checking for _mm512_popcnt_epi64 with CFLAGS=-mavx512vpopcntdq -mavx512bw... " >&6; }
+if ${pgac_cv_avx512_popcnt_intrinsics__mavx512vpopcntdq__mavx512bw+:} false; then :
+  $as_echo_n "(cached) " >&6
+else
+  pgac_save_CFLAGS=$CFLAGS
+CFLAGS="$pgac_save_CFLAGS -mavx512vpopcntdq -mavx512bw"
+cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+/* end confdefs.h.  */
+#include <immintrin.h>
+int
+main ()
+{
+const char buf[sizeof(__m512i)];
+   PG_INT64_TYPE popcnt = 0;
+   __m512i accum = _mm512_setzero_si512();
+   const __m512i val = _mm512_maskz_loadu_epi8((__mmask64) 0xf0f0f0f0f0f0f0f0, (const __m512i *) buf);
+   const __m512i cnt = _mm512_popcnt_epi64(val);
+   accum = _mm512_add_epi64(accum, cnt);
+   popcnt = _mm512_reduce_add_epi64(accum);
+   /* return computed value, to prevent the above being optimized away */
+   return popcnt == 0;
+  ;
+  return 0;
+}
+_ACEOF
+if ac_fn_c_try_link "$LINENO"; then :
+  pgac_cv_avx512_popcnt_intrinsics__mavx512vpopcntdq__mavx512bw=yes
+else
+  pgac_cv_avx512_popcnt_intrinsics__mavx512vpopcntdq__mavx512bw=no
+fi
+rm -f core conftest.err conftest.$ac_objext \
+    conftest$ac_exeext conftest.$ac_ext
+CFLAGS="$pgac_save_CFLAGS"
+fi
+{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $pgac_cv_avx512_popcnt_intrinsics__mavx512vpopcntdq__mavx512bw" >&5
+$as_echo "$pgac_cv_avx512_popcnt_intrinsics__mavx512vpopcntdq__mavx512bw" >&6; }
+if test x"$pgac_cv_avx512_popcnt_intrinsics__mavx512vpopcntdq__mavx512bw" = x"yes"; then
+  CFLAGS_POPCNT="-mavx512vpopcntdq -mavx512bw"
+  pgac_avx512_popcnt_intrinsics=yes
+fi
+
+  fi
+  if test x"$pgac_avx512_popcnt_intrinsics" = x"yes"; then
+    PG_POPCNT_OBJS="pg_popcount_avx512.o pg_popcount_avx512_choose.o"
+
+$as_echo "#define USE_AVX512_POPCNT_WITH_RUNTIME_CHECK 1" >>confdefs.h
+
+  fi
+fi
+
+
+
 # Check for Intel SSE 4.2 intrinsics to do CRC calculations.
 #
 # First check if the _mm_crc32_u8 and _mm_crc32_u64 intrinsics can be used
index 57f734879e19fe8756c3fc7e4fd463dbf5451607..67e738d92b1cca54ce51f6ea56b926d210f4f020 100644 (file)
@@ -2052,6 +2052,17 @@ if test x"$pgac_cv__get_cpuid" = x"yes"; then
   AC_DEFINE(HAVE__GET_CPUID, 1, [Define to 1 if you have __get_cpuid.])
 fi
 
+AC_CACHE_CHECK([for __get_cpuid_count], [pgac_cv__get_cpuid_count],
+[AC_LINK_IFELSE([AC_LANG_PROGRAM([#include <cpuid.h>],
+  [[unsigned int exx[4] = {0, 0, 0, 0};
+  __get_cpuid_count(7, 0, &exx[0], &exx[1], &exx[2], &exx[3]);
+  ]])],
+  [pgac_cv__get_cpuid_count="yes"],
+  [pgac_cv__get_cpuid_count="no"])])
+if test x"$pgac_cv__get_cpuid_count" = x"yes"; then
+  AC_DEFINE(HAVE__GET_CPUID_COUNT, 1, [Define to 1 if you have __get_cpuid_count.])
+fi
+
 AC_CACHE_CHECK([for __cpuid], [pgac_cv__cpuid],
 [AC_LINK_IFELSE([AC_LANG_PROGRAM([#include <intrin.h>],
   [[unsigned int exx[4] = {0, 0, 0, 0};
@@ -2063,6 +2074,46 @@ if test x"$pgac_cv__cpuid" = x"yes"; then
   AC_DEFINE(HAVE__CPUID, 1, [Define to 1 if you have __cpuid.])
 fi
 
+AC_CACHE_CHECK([for __cpuidex], [pgac_cv__cpuidex],
+[AC_LINK_IFELSE([AC_LANG_PROGRAM([#include <intrin.h>],
+  [[unsigned int exx[4] = {0, 0, 0, 0};
+  __get_cpuidex(exx[0], 7, 0);
+  ]])],
+  [pgac_cv__cpuidex="yes"],
+  [pgac_cv__cpuidex="no"])])
+if test x"$pgac_cv__cpuidex" = x"yes"; then
+  AC_DEFINE(HAVE__CPUIDEX, 1, [Define to 1 if you have __cpuidex.])
+fi
+
+# Check for XSAVE intrinsics
+#
+CFLAGS_XSAVE=""
+PGAC_XSAVE_INTRINSICS([])
+if test x"$pgac_xsave_intrinsics" != x"yes"; then
+  PGAC_XSAVE_INTRINSICS([-mxsave])
+fi
+if test x"$pgac_xsave_intrinsics" = x"yes"; then
+  AC_DEFINE(HAVE_XSAVE_INTRINSICS, 1, [Define to 1 if you have XSAVE intrinsics.])
+fi
+AC_SUBST(CFLAGS_XSAVE)
+
+# Check for AVX-512 popcount intrinsics
+#
+CFLAGS_POPCNT=""
+PG_POPCNT_OBJS=""
+if test x"$host_cpu" = x"x86_64"; then
+  PGAC_AVX512_POPCNT_INTRINSICS([])
+  if test x"$pgac_avx512_popcnt_intrinsics" != x"yes"; then
+    PGAC_AVX512_POPCNT_INTRINSICS([-mavx512vpopcntdq -mavx512bw])
+  fi
+  if test x"$pgac_avx512_popcnt_intrinsics" = x"yes"; then
+    PG_POPCNT_OBJS="pg_popcount_avx512.o pg_popcount_avx512_choose.o"
+    AC_DEFINE(USE_AVX512_POPCNT_WITH_RUNTIME_CHECK, 1, [Define to 1 to use AVX-512 popcount instructions with a runtime check.])
+  fi
+fi
+AC_SUBST(CFLAGS_POPCNT)
+AC_SUBST(PG_POPCNT_OBJS)
+
 # Check for Intel SSE 4.2 intrinsics to do CRC calculations.
 #
 # First check if the _mm_crc32_u8 and _mm_crc32_u64 intrinsics can be used
index 87437960bc321d53c719286faea77f9b49798ed6..5acf083ce3c7a9f05db64c1c08f0afa743da4a10 100644 (file)
@@ -1783,6 +1783,30 @@ elif cc.links('''
 endif
 
 
+# Check for __get_cpuid_count() and __cpuidex() in a similar fashion.
+if cc.links('''
+    #include <cpuid.h>
+    int main(int arg, char **argv)
+    {
+        unsigned int exx[4] = {0, 0, 0, 0};
+        __get_cpuid_count(7, 0, &exx[0], &exx[1], &exx[2], &exx[3]);
+    }
+    ''', name: '__get_cpuid_count',
+    args: test_c_args)
+  cdata.set('HAVE__GET_CPUID_COUNT', 1)
+elif cc.links('''
+    #include <intrin.h>
+    int main(int arg, char **argv)
+    {
+        unsigned int exx[4] = {0, 0, 0, 0};
+        __cpuidex(exx, 7, 0);
+    }
+    ''', name: '__cpuidex',
+    args: test_c_args)
+  cdata.set('HAVE__CPUIDEX', 1)
+endif
+
+
 # Defend against clang being used on x86-32 without SSE2 enabled.  As current
 # versions of clang do not understand -fexcess-precision=standard, the use of
 # x87 floating point operations leads to problems like isinf possibly returning
@@ -1996,6 +2020,69 @@ int main(void)
 endif
 
 
+###############################################################
+# Check for the availability of XSAVE intrinsics.
+###############################################################
+
+cflags_xsave = []
+if host_cpu == 'x86' or host_cpu == 'x86_64'
+
+  prog = '''
+#include <immintrin.h>
+
+int main(void)
+{
+    return _xgetbv(0) & 0xe0;
+}
+'''
+
+  if cc.links(prog, name: 'XSAVE intrinsics without -mxsave',
+        args: test_c_args)
+    cdata.set('HAVE_XSAVE_INTRINSICS', 1)
+  elif cc.links(prog, name: 'XSAVE intrinsics with -mxsave',
+        args: test_c_args + ['-mxsave'])
+    cdata.set('HAVE_XSAVE_INTRINSICS', 1)
+    cflags_xsave += '-mxsave'
+  endif
+
+endif
+
+
+###############################################################
+# Check for the availability of AVX-512 popcount intrinsics.
+###############################################################
+
+cflags_popcnt = []
+if host_cpu == 'x86_64'
+
+  prog = '''
+#include <immintrin.h>
+
+int main(void)
+{
+    const char buf[sizeof(__m512i)];
+    INT64 popcnt = 0;
+    __m512i accum = _mm512_setzero_si512();
+    const __m512i val = _mm512_maskz_loadu_epi8((__mmask64) 0xf0f0f0f0f0f0f0f0, (const __m512i *) buf);
+    const __m512i cnt = _mm512_popcnt_epi64(val);
+    accum = _mm512_add_epi64(accum, cnt);
+    popcnt = _mm512_reduce_add_epi64(accum);
+    /* return computed value, to prevent the above being optimized away */
+    return popcnt == 0;
+}
+'''
+
+  if cc.links(prog, name: 'AVX-512 popcount without -mavx512vpopcntdq -mavx512bw',
+        args: test_c_args + ['-DINT64=@0@'.format(cdata.get('PG_INT64_TYPE'))])
+    cdata.set('USE_AVX512_POPCNT_WITH_RUNTIME_CHECK', 1)
+  elif cc.links(prog, name: 'AVX-512 popcount with -mavx512vpopcntdq -mavx512bw',
+        args: test_c_args + ['-DINT64=@0@'.format(cdata.get('PG_INT64_TYPE'))] + ['-mavx512vpopcntdq'] + ['-mavx512bw'])
+    cdata.set('USE_AVX512_POPCNT_WITH_RUNTIME_CHECK', 1)
+    cflags_popcnt += ['-mavx512vpopcntdq'] + ['-mavx512bw']
+  endif
+
+endif
+
 
 ###############################################################
 # Select CRC-32C implementation.
index 8b3f8c24e08e8a7527f3735686b515403efac521..36d880d22515705dcfb1ff17f6053f3b144f3b5d 100644 (file)
@@ -262,7 +262,9 @@ CFLAGS_SL_MODULE = @CFLAGS_SL_MODULE@
 CXXFLAGS_SL_MODULE = @CXXFLAGS_SL_MODULE@
 CFLAGS_UNROLL_LOOPS = @CFLAGS_UNROLL_LOOPS@
 CFLAGS_VECTORIZE = @CFLAGS_VECTORIZE@
+CFLAGS_POPCNT = @CFLAGS_POPCNT@
 CFLAGS_CRC = @CFLAGS_CRC@
+CFLAGS_XSAVE = @CFLAGS_XSAVE@
 PERMIT_DECLARATION_AFTER_STATEMENT = @PERMIT_DECLARATION_AFTER_STATEMENT@
 CXXFLAGS = @CXXFLAGS@
 
@@ -758,6 +760,9 @@ LIBOBJS = @LIBOBJS@
 # files needed for the chosen CRC-32C implementation
 PG_CRC32C_OBJS = @PG_CRC32C_OBJS@
 
+# files needed for the chosen popcount implementation
+PG_POPCNT_OBJS = @PG_POPCNT_OBJS@
+
 LIBS := -lpgcommon -lpgport $(LIBS)
 
 # to make ws2_32.lib the last library
index 591e1ca3df66486e6968c362c22567459d428709..f8d3e3b6b843cf77e0a956ca0f7b51c6f2650eb1 100644 (file)
 /* Define to 1 if the assembler supports X86_64's POPCNTQ instruction. */
 #undef HAVE_X86_64_POPCNTQ
 
+/* Define to 1 if you have XSAVE intrinsics. */
+#undef HAVE_XSAVE_INTRINSICS
+
 /* Define to 1 if the system has the type `_Bool'. */
 #undef HAVE__BOOL
 
 /* Define to 1 if you have __cpuid. */
 #undef HAVE__CPUID
 
+/* Define to 1 if you have __cpuidex. */
+#undef HAVE__CPUIDEX
+
 /* Define to 1 if you have __get_cpuid. */
 #undef HAVE__GET_CPUID
 
+/* Define to 1 if you have __get_cpuid_count. */
+#undef HAVE__GET_CPUID_COUNT
+
 /* Define to 1 if your compiler understands _Static_assert. */
 #undef HAVE__STATIC_ASSERT
 
 /* Define to 1 to build with assertion checks. (--enable-cassert) */
 #undef USE_ASSERT_CHECKING
 
+/* Define to 1 to use AVX-512 popcount instructions with a runtime check. */
+#undef USE_AVX512_POPCNT_WITH_RUNTIME_CHECK
+
 /* Define to 1 to build with Bonjour support. (--with-bonjour) */
 #undef USE_BONJOUR
 
index de480da71e1483208dd2636fe410d4549a1fdf9d..b453f84d8f154f0a602f3ccebea7012c7dc0e0e7 100644 (file)
@@ -304,6 +304,17 @@ extern PGDLLIMPORT int (*pg_popcount32) (uint32 word);
 extern PGDLLIMPORT int (*pg_popcount64) (uint64 word);
 extern PGDLLIMPORT uint64 (*pg_popcount_optimized) (const char *buf, int bytes);
 
+/*
+ * We can also try to use the AVX-512 popcount instruction on some systems.
+ * The implementation of that is located in its own file because it may
+ * require special compiler flags that we don't want to apply to any other
+ * files.
+ */
+#ifdef USE_AVX512_POPCNT_WITH_RUNTIME_CHECK
+extern bool pg_popcount_avx512_available(void);
+extern uint64 pg_popcount_avx512(const char *buf, int bytes);
+#endif
+
 #else
 /* Use a portable implementation -- no need for a function pointer. */
 extern int pg_popcount32(uint32 word);
index b0f4178b3d9eb45bc42d456e4388172e6b9d99de..5618050b30630149166c8824f1233147a472ba94 100644 (file)
@@ -100,8 +100,10 @@ pgxs_kv = {
     ' '.join(cflags_no_decl_after_statement),
 
   'CFLAGS_CRC': ' '.join(cflags_crc),
+  'CFLAGS_POPCNT': ' '.join(cflags_popcnt),
   'CFLAGS_UNROLL_LOOPS': ' '.join(unroll_loops_cflags),
   'CFLAGS_VECTORIZE': ' '.join(vectorize_cflags),
+  'CFLAGS_XSAVE': ' '.join(cflags_xsave),
 
   'LDFLAGS': var_ldflags,
   'LDFLAGS_EX': var_ldflags_ex,
@@ -177,7 +179,7 @@ pgxs_empty = [
   'WANTED_LANGUAGES',
 
   # Not needed because we don't build the server / PLs with the generated makefile
-  'LIBOBJS', 'PG_CRC32C_OBJS', 'TAS',
+  'LIBOBJS', 'PG_CRC32C_OBJS', 'PG_POPCNT_OBJS', 'TAS',
   'DTRACEFLAGS', # only server has dtrace probes
 
   'perl_archlibexp', 'perl_embed_ccflags', 'perl_embed_ldflags', 'perl_includespec', 'perl_privlibexp',
index dcc8737e6823da297f83f34484f0d8e686ebdbe1..db7c02117b02d1d5f07eaea33e204884c56ef685 100644 (file)
@@ -38,6 +38,7 @@ LIBS += $(PTHREAD_LIBS)
 OBJS = \
    $(LIBOBJS) \
    $(PG_CRC32C_OBJS) \
+   $(PG_POPCNT_OBJS) \
    bsearch_arg.o \
    chklocale.o \
    inet_net_ntop.o \
@@ -92,6 +93,16 @@ pg_crc32c_armv8.o: CFLAGS+=$(CFLAGS_CRC)
 pg_crc32c_armv8_shlib.o: CFLAGS+=$(CFLAGS_CRC)
 pg_crc32c_armv8_srv.o: CFLAGS+=$(CFLAGS_CRC)
 
+# all versions of pg_popcount_avx512_choose.o need CFLAGS_XSAVE
+pg_popcount_avx512_choose.o: CFLAGS+=$(CFLAGS_XSAVE)
+pg_popcount_avx512_choose_shlib.o: CFLAGS+=$(CFLAGS_XSAVE)
+pg_popcount_avx512_choose_srv.o: CFLAGS+=$(CFLAGS_XSAVE)
+
+# all versions of pg_popcount_avx512.o need CFLAGS_POPCNT
+pg_popcount_avx512.o: CFLAGS+=$(CFLAGS_POPCNT)
+pg_popcount_avx512_shlib.o: CFLAGS+=$(CFLAGS_POPCNT)
+pg_popcount_avx512_srv.o: CFLAGS+=$(CFLAGS_POPCNT)
+
 #
 # Shared library versions of object files
 #
index 92b593e6ef3dccdbe181749c2b44d0e49ab7e999..fd9ee199d1b0be4708e1da579cfcbabb6b7abe97 100644 (file)
@@ -84,6 +84,8 @@ replace_funcs_pos = [
   ['pg_crc32c_sse42', 'USE_SSE42_CRC32C_WITH_RUNTIME_CHECK', 'crc'],
   ['pg_crc32c_sse42_choose', 'USE_SSE42_CRC32C_WITH_RUNTIME_CHECK'],
   ['pg_crc32c_sb8', 'USE_SSE42_CRC32C_WITH_RUNTIME_CHECK'],
+  ['pg_popcount_avx512', 'USE_AVX512_POPCNT_WITH_RUNTIME_CHECK', 'popcnt'],
+  ['pg_popcount_avx512_choose', 'USE_AVX512_POPCNT_WITH_RUNTIME_CHECK', 'xsave'],
 
   # arm / aarch64
   ['pg_crc32c_armv8', 'USE_ARMV8_CRC32C'],
@@ -98,8 +100,8 @@ replace_funcs_pos = [
   ['pg_crc32c_sb8', 'USE_SLICING_BY_8_CRC32C'],
 ]
 
-pgport_cflags = {'crc': cflags_crc}
-pgport_sources_cflags = {'crc': []}
+pgport_cflags = {'crc': cflags_crc, 'popcnt': cflags_popcnt, 'xsave': cflags_xsave}
+pgport_sources_cflags = {'crc': [], 'popcnt': [], 'xsave': []}
 
 foreach f : replace_funcs_neg
   func = f.get(0)
index 6271acea600301266af57b394f364883ea283822..411be90f734b12fdbf984792115068c19e8e1fdd 100644 (file)
@@ -163,6 +163,11 @@ choose_popcount_functions(void)
        pg_popcount64 = pg_popcount64_slow;
        pg_popcount_optimized = pg_popcount_slow;
    }
+
+#ifdef USE_AVX512_POPCNT_WITH_RUNTIME_CHECK
+   if (pg_popcount_avx512_available())
+       pg_popcount_optimized = pg_popcount_avx512;
+#endif
 }
 
 static int
diff --git a/src/port/pg_popcount_avx512.c b/src/port/pg_popcount_avx512.c
new file mode 100644 (file)
index 0000000..9088176
--- /dev/null
@@ -0,0 +1,81 @@
+/*-------------------------------------------------------------------------
+ *
+ * pg_popcount_avx512.c
+ *   Holds the AVX-512 pg_popcount() implementation.
+ *
+ * Copyright (c) 2024, PostgreSQL Global Development Group
+ *
+ * IDENTIFICATION
+ *   src/port/pg_popcount_avx512.c
+ *
+ *-------------------------------------------------------------------------
+ */
+#include "c.h"
+
+#include <immintrin.h>
+
+#include "port/pg_bitutils.h"
+
+/*
+ * It's probably unlikely that TRY_POPCNT_FAST won't be set if we are able to
+ * use AVX-512 intrinsics, but we check it anyway to be sure.  We piggy-back on
+ * the function pointers that are only used when TRY_POPCNT_FAST is set.
+ */
+#ifdef TRY_POPCNT_FAST
+
+/*
+ * pg_popcount_avx512
+ *     Returns the number of 1-bits in buf
+ */
+uint64
+pg_popcount_avx512(const char *buf, int bytes)
+{
+   __m512i     val,
+               cnt;
+   __m512i     accum = _mm512_setzero_si512();
+   const char *final;
+   int         tail_idx;
+   __mmask64   mask = ~UINT64CONST(0);
+
+   /*
+    * Align buffer down to avoid double load overhead from unaligned access.
+    * Calculate a mask to ignore preceding bytes.  Find start offset of final
+    * iteration and ensure it is not empty.
+    */
+   mask <<= ((uintptr_t) buf) % sizeof(__m512i);
+   tail_idx = (((uintptr_t) buf + bytes - 1) % sizeof(__m512i)) + 1;
+   final = (const char *) TYPEALIGN_DOWN(sizeof(__m512i), buf + bytes - 1);
+   buf = (const char *) TYPEALIGN_DOWN(sizeof(__m512i), buf);
+
+   /*
+    * Iterate through all but the final iteration.  Starting from the second
+    * iteration, the mask is ignored.
+    */
+   if (buf < final)
+   {
+       val = _mm512_maskz_loadu_epi8(mask, (const __m512i *) buf);
+       cnt = _mm512_popcnt_epi64(val);
+       accum = _mm512_add_epi64(accum, cnt);
+
+       buf += sizeof(__m512i);
+       mask = ~UINT64CONST(0);
+
+       for (; buf < final; buf += sizeof(__m512i))
+       {
+           val = _mm512_load_si512((const __m512i *) buf);
+           cnt = _mm512_popcnt_epi64(val);
+           accum = _mm512_add_epi64(accum, cnt);
+       }
+   }
+
+   /* Final iteration needs to ignore bytes that are not within the length */
+   mask &= (~UINT64CONST(0) >> (sizeof(__m512i) - tail_idx));
+
+   val = _mm512_maskz_loadu_epi8(mask, (const __m512i *) buf);
+   cnt = _mm512_popcnt_epi64(val);
+   accum = _mm512_add_epi64(accum, cnt);
+
+   return _mm512_reduce_add_epi64(accum);
+}
+
+#endif                         /* TRY_POPCNT_FAST */
diff --git a/src/port/pg_popcount_avx512_choose.c b/src/port/pg_popcount_avx512_choose.c
new file mode 100644 (file)
index 0000000..ae3fa3d
--- /dev/null
@@ -0,0 +1,88 @@
+/*-------------------------------------------------------------------------
+ *
+ * pg_popcount_avx512_choose.c
+ *    Test whether we can use the AVX-512 pg_popcount() implementation.
+ *
+ * Copyright (c) 2024, PostgreSQL Global Development Group
+ *
+ * IDENTIFICATION
+ *    src/port/pg_popcount_avx512_choose.c
+ *
+ *-------------------------------------------------------------------------
+ */
+#include "c.h"
+
+#if defined(HAVE__GET_CPUID) || defined(HAVE__GET_CPUID_COUNT)
+#include <cpuid.h>
+#endif
+
+#ifdef HAVE_XSAVE_INTRINSICS
+#include <immintrin.h>
+#endif
+
+#if defined(HAVE__CPUID) || defined(HAVE__CPUIDEX)
+#include <intrin.h>
+#endif
+
+#include "port/pg_bitutils.h"
+
+/*
+ * It's probably unlikely that TRY_POPCNT_FAST won't be set if we are able to
+ * use AVX-512 intrinsics, but we check it anyway to be sure.  We piggy-back on
+ * the function pointers that are only used when TRY_POPCNT_FAST is set.
+ */
+#ifdef TRY_POPCNT_FAST
+
+/*
+ * Returns true if the CPU supports the instructions required for the AVX-512
+ * pg_popcount() implementation.
+ */
+bool
+pg_popcount_avx512_available(void)
+{
+   unsigned int exx[4] = {0, 0, 0, 0};
+
+   /* Does CPUID say there's support for AVX-512 popcount instructions? */
+#if defined(HAVE__GET_CPUID_COUNT)
+   __get_cpuid_count(7, 0, &exx[0], &exx[1], &exx[2], &exx[3]);
+#elif defined(HAVE__CPUIDEX)
+   __cpuidex(exx, 7, 0);
+#else
+#error cpuid instruction not available
+#endif
+   if ((exx[2] & (1 << 14)) == 0)  /* avx512-vpopcntdq */
+       return false;
+
+   /* Does CPUID say there's support for AVX-512 byte and word instructions? */
+   memset(exx, 0, sizeof(exx));
+#if defined(HAVE__GET_CPUID_COUNT)
+   __get_cpuid_count(7, 0, &exx[0], &exx[1], &exx[2], &exx[3]);
+#elif defined(HAVE__CPUIDEX)
+   __cpuidex(exx, 7, 0);
+#else
+#error cpuid instruction not available
+#endif
+   if ((exx[1] & (1 << 30)) == 0)  /* avx512-bw */
+       return false;
+
+   /* Does CPUID say there's support for XSAVE instructions? */
+   memset(exx, 0, sizeof(exx));
+#if defined(HAVE__GET_CPUID)
+   __get_cpuid(1, &exx[0], &exx[1], &exx[2], &exx[3]);
+#elif defined(HAVE__CPUID)
+   __cpuid(exx, 1);
+#else
+#error cpuid instruction not available
+#endif
+   if ((exx[2] & (1 << 26)) == 0)  /* xsave */
+       return false;
+
+   /* Does XGETBV say the ZMM registers are enabled? */
+#ifdef HAVE_XSAVE_INTRINSICS
+   return (_xgetbv(0) & 0xe0) != 0;
+#else
+   return false;
+#endif
+}
+
+#endif                         /* TRY_POPCNT_FAST */
index e17cbf42ca2f1996f88f7a915bc7296e330cf63d..6a436288bb15f867bf20a0e869d7c298a235edb7 100644 (file)
@@ -740,6 +740,30 @@ SELECT bit_count(B'1111111111'::bit(10));
         10
 (1 row)
 
+SELECT bit_count(repeat('0', 100)::bit(100));
+ bit_count 
+-----------
+         0
+(1 row)
+
+SELECT bit_count(repeat('1', 100)::bit(100));
+ bit_count 
+-----------
+       100
+(1 row)
+
+SELECT bit_count(repeat('01', 500)::bit(1000));
+ bit_count 
+-----------
+       500
+(1 row)
+
+SELECT bit_count(repeat('10101', 200)::bit(1000));
+ bit_count 
+-----------
+       600
+(1 row)
+
 -- This table is intentionally left around to exercise pg_dump/pg_upgrade
 CREATE TABLE bit_defaults(
   b1 bit(4) DEFAULT '1001',
index 34230b99fba2eeb367f7ecd13523d04b1dabfae9..8ba6facd032f7003906ecb5a322d6eca4c988b83 100644 (file)
@@ -223,6 +223,10 @@ SELECT overlay(B'0101011100' placing '001' from 20);
 -- bit_count
 SELECT bit_count(B'0101011100'::bit(10));
 SELECT bit_count(B'1111111111'::bit(10));
+SELECT bit_count(repeat('0', 100)::bit(100));
+SELECT bit_count(repeat('1', 100)::bit(100));
+SELECT bit_count(repeat('01', 500)::bit(1000));
+SELECT bit_count(repeat('10101', 200)::bit(1000));
 
 -- This table is intentionally left around to exercise pg_dump/pg_upgrade
 CREATE TABLE bit_defaults(