ch4/shm: fix performance degradation on Sapphire Rapids with Intel Compiler

configure.ac

@@ -906,11 +906,14 @@ for option in $enable_fast ; do
  ;;
  alwaysinline) # No op in MPICH. See mpl/configure.ac
  ;;
- sse2) # No op in MPICH. See mpl/configure.ac
+ sse2)
+ enable_fast_sse2_instr=yes
  ;;
- avx) # No op in MPICH. See mpl/configure.ac
+ avx)
+ enable_fast_avx_instr=yes
  ;;
- avx512f) # No op in MPICH. See mpl/configure.ac
+ avx512f)
+ enable_fast_avx512f_instr=yes
  ;;
  all|yes)
  enable_fast_ndebug=yes
@@ -955,6 +958,81 @@ if test -z "$enable_fast_no_strict_alignment" ; then
  AC_DEFINE(NEEDS_STRICT_ALIGNMENT,1,[Define if strict alignment memory access is required])
 fi
 
+if test "$enable_fast_sse2_instr" = "yes" ; then
+ AC_CACHE_CHECK([whether -msse2 is supported], pac_cv_found_sse2,
+ [PAC_C_CHECK_COMPILER_OPTION([-msse2],pac_cv_found_sse2=yes,pac_cv_found_sse2=no)],
+ pac_cv_found_sse2=no,pac_cv_found_sse2=yes)
+ PAC_PUSH_FLAG([CFLAGS])
+ PAC_APPEND_FLAG([-msse2],[CFLAGS])
+ AC_CACHE_CHECK([whether SSE2 is supported by the CPU], pac_cv_found_sse2_runnable,[
+ AC_RUN_IFELSE([AC_LANG_SOURCE([[
+ #include <emmintrin.h>
+
+ int main() {
+ __m128i a = _mm_set1_epi32(1);
+ __asm__ volatile("" : : "x" (a) : "memory");
+ return 0;
+ }
+ ]])], pac_cv_found_sse2_runnable="yes",
+ pac_cv_found_sse2_runnable="no",
+ pac_cv_found_sse2_runnable="unknown")
+ ])
+ PAC_POP_FLAG([CFLAGS])
+ if test "$pac_cv_found_sse2" = "yes" && test "$pac_cv_found_sse2_runnable" = "yes"; then
+ PAC_APPEND_FLAG([-msse2],[CFLAGS])
+ fi
+fi
+
+if test "$enable_fast_avx_instr" = "yes" ; then
+ AC_CACHE_CHECK([whether -mavx is supported], pac_cv_found_avx,
+ [PAC_C_CHECK_COMPILER_OPTION([-mavx],pac_cv_found_avx=yes,pac_cv_found_avx=no)],
+ pac_cv_found_avx=no,pac_cv_found_avx=yes)
+ PAC_PUSH_FLAG([CFLAGS])
+ PAC_APPEND_FLAG([-mavx],[CFLAGS])
+ AC_CACHE_CHECK([whether AVX is supported by the CPU], pac_cv_found_avx_runnable,[
+ AC_RUN_IFELSE([AC_LANG_SOURCE([[
+ #include <immintrin.h>
+
+ int main() {
+ __m256i a = _mm256_set1_epi32(1);
+ __asm__ volatile("" : : "x" (a) : "memory");
+ return 0;
+ }
+ ]])], pac_cv_found_avx_runnable="yes",
+ pac_cv_found_avx_runnable="no",
+ pac_cv_found_avx_runnable="unknown")
+ ])
+ PAC_POP_FLAG([CFLAGS])
+ if test "$pac_cv_found_avx" = "yes" && test "$pac_cv_found_avx_runnable" = "yes"; then
+ PAC_APPEND_FLAG([-mavx],[CFLAGS])
+ fi
+fi
+
+if test "$enable_fast_avx512f_instr" = "yes" ; then
+ AC_CACHE_CHECK([whether -mavx512f is supported], pac_cv_found_avx512f,
+ [PAC_C_CHECK_COMPILER_OPTION([-mavx512f],pac_cv_found_avx512f=yes,pac_cv_found_avx512f=no)],
+ pac_cv_found_avx512f=no,pac_cv_found_avx512f=yes)
+ PAC_PUSH_FLAG([CFLAGS])
+ PAC_APPEND_FLAG([-mavx512f],[CFLAGS])
+ AC_CACHE_CHECK([whether AVX512F is supported by the CPU], pac_cv_found_avx512f_runnable,[
+ AC_RUN_IFELSE([AC_LANG_SOURCE([[
+ #include <immintrin.h>
+
+ int main() {
+ __m512i a = _mm512_set1_epi32(1);
+ __asm__ volatile("" : : "x" (a) : "memory");
+ return 0;
+ }
+ ]])], pac_cv_found_avx512f_runnable="yes",
+ pac_cv_found_avx512f_runnable="no",
+ pac_cv_found_avx512f_runnable="unknown")
+ ])
+ PAC_POP_FLAG([CFLAGS])
+ if test "$pac_cv_found_avx512f" = "yes" && test "$pac_cv_found_avx512f_runnable" = "yes"; then
+ PAC_APPEND_FLAG([-mavx512f],[CFLAGS])
+ fi
+fi
+
 # error-checking
 # Change default into the specific value of the default
 if test "$enable_error_checking" = "yes" ; then

src/mpid/ch4/shm/posix/posix_init.c

@@ -40,7 +40,7 @@
  - name : MPIR_CVAR_CH4_SHM_POSIX_TOPO_ENABLE
  category : CH4
  type : boolean
- default : false
+ default : true
  class : none
  verbosity : MPI_T_VERBOSITY_USER_BASIC
  scope : MPI_T_SCOPE_ALL_EQ