SHMEM_DOUBLE_MAX_TO_ALL(3)

📦
version 5.0.10 · 4 versions
, in ubuntu stonking

🎀 NAME 🎀

okay bestie, hydrate, sit down, because this page is a whole squad of functions and they all share one job. This is SHMEM_DOUBLE_MAX_TO_ALL(3) straight out of Open MPI, and the fam is:

and what do they do? "Performs a maximum function reduction across a set of processing elements (PEs)." 💅 iconic behavior, no notes.

✨ SYNOPSIS ✨

C or C++:

#include <mpp/shmem.h>

void shmem_double_max_to_all(double *target, const double *source,
  int nreduce, int PE_start, int logPE_stride, int PE_size,
  double *pWrk, long *pSync)

void shmem_float_max_to_all(float *target, const float *source,
  int nreduce, int PE_start, int logPE_stride, int PE_size,
  float *pWrk, long *pSync)

void shmem_int_max_to_all(int *target, const int *source,
  int nreduce, int PE_start, int logPE_stride, int PE_size,
  int *pWrk, long *pSync)

void shmem_long_max_to_all(long *target, const long *source,
  int nreduce, int PE_start, int logPE_stride, int PE_size,
  long *pWrk, long *pSync)

void shmem_longdouble_max_to_all(long double *target,
  const long double *source, int nreduce, int PE_start,
  int logPE_stride, int PE_size, long double *pWrk, long *pSync)

void shmem_longlong_max_to_all(long long *target,
  const long long *source, int nreduce,  int PE_start,
  int logPE_stride, int PE_size, long long *pWrk, long *pSync)

 void shmem_short_max_to_all(short *target, const short *source,
  int nreduce, int PE_start, int logPE_stride, int PE_size,
  short *pWrk, long *pSync)

Fortran:

INCLUDE "mpp/shmem.fh"

INTEGER pSync(SHMEM_REDUCE_SYNC_SIZE)

INTEGER nreduce, PE_start, logPE_stride, PE_size

CALL SHMEM_INT4_MAX_TO_ALL(target, source, nreduce,
& PE_start, logPE_stride, PE_size, pWrk, pSync)

CALL SHMEM_INT8_MAX_TO_ALL(target, source, nreduce,
& PE_start, logPE_stride, PE_size, pWrk, pSync)

CALL SHMEM_REAL4_MAX_TO_ALL(target, source, nreduce,
& PE_start, logPE_stride, PE_size, pWrk, pSync)

CALL SHMEM_REAL8_MAX_TO_ALL(target, source, nreduce,
& PE_start, logPE_stride, PE_size, pWrk, pSync)

CALL SHMEM_REAL16_MAX_TO_ALL(target, source, nreduce,
& PE_start, logPE_stride, PE_size, pWrk, pSync)

🌸 DESCRIPTION 🌸

So the shared memory (SHMEM) reduction routines compute one or more reductions across symmetric arrays on multiple virtual PEs. A reduction performs an associative binary operation across a set of values. Want the full roster of other SHMEM reduction routines? Go peek at intro_shmem(3). 📚

Like literally every SHMEM collective routine ever, each of these assumes that only PEs in the active set call the routine. If a PE that is not in the active set calls a SHMEM collective routine…

⚠️ undefined behavior results. we do not do that here, bestie. 🚫

The nreduce argument decides how many separate reductions you're doing. The source array on all PEs in the active set hands over one element per reduction. Then the results land in the target array on all PEs in the active set. And the active set itself? Defined by the PE_start, logPE_stride, PE_size triplet. 🎯

Oh and here's a cute little fact: the source and target arrays may be the same array, but they may not be overlapping arrays. same, yes. overlapping, absolutely not. 🙅‍♀️

The arguments are as follows:

  • target — a symmetric array of length nreduce elements that receives the results of the reduction operations. Its data type varies with whichever version of the reduction routine you called. Calling from C? Go check the SYNOPSIS section for the data type tea. ☕

When calling from Fortran, the target data types are as follows:

  • shmem_comp8_max_to_all — Complex, with an element size equal to two 8-byte real values.
  • shmem_int4_max_to_all — Integer, with an element size of 4 bytes.
  • shmem_int8_max_to_all — Integer, with an element size of 8 bytes.
  • shmem_real4_max_to_all — Real, with an element size of 4 bytes.
  • shmem_real16_max_to_all — Real, with an element size of 16 bytes.
  • source — a symmetric array of length nreduce elements containing one element for each separate reduction operation. It must have the same data type as target. matching outfits or it's not happening. 👯
  • nreduce — the number of elements in the target and source arrays. It must be of type integer. In Fortran? It must be a default integer value.
  • PE_start — the lowest virtual PE number of the active set of PEs. Must be of type integer. In Fortran, it must be a default integer value.
  • logPE_stride — the log (base 2) of the stride between consecutive virtual PE numbers in the active set. Must be of type integer. In Fortran, it must be a default integer value. (yes babe, log base 2. math is the vibe.) 🧮
  • PE_size — the number of PEs in the active set. Must be of type integer. In Fortran, it must be a default integer value.
  • pWrk — a symmetric work array. It must have the same data type as target. In C/C++, this contains max(nreduce/2 + 1, _SHMEM_REDUCE_MIN_WRKDATA_SIZE) elements. In Fortran, this contains max(nreduce/2 + 1, SHMEM_REDUCE_MIN_WRKDATA_SIZE) elements.
  • pSync — a symmetric work array. In C/C++, pSync is of type long and of size _SHMEM_REDUCE_SYNC_SIZE. In Fortran, pSync is of type integer and of size SHMEM_REDUCE_SYNC_SIZE. If you're in Fortran it must be a default integer value. Every single element of this array must be initialized with the value _SHMEM_SYNC_VALUE (in C/C++) or SHMEM_SYNC_VALUE (in Fortran) before any of the PEs in the active set enter the reduction routine. pre-game ritual, non-negotiable. 💖

The values of the arguments nreduce, PE_start, logPE_stride, and PE_size must be equal on all PEs in the active set. The same target and source arrays, and the same pWrk and pSync work arrays, must be passed to all PEs in the active set. everybody on the same page, no freestyling. 🤝

Before any PE calls a reduction routine, you must make sure these conditions exist (synchronization via a barrier or some other method is often needed to make this happen, fr fr):

  • The pWrk and pSync arrays on all PEs in the active set are not still in use from a prior call to a collective SHMEM routine.
  • The target array on all PEs in the active set is ready to accept the results of the reduction.

Upon return from a reduction routine, these are true for the local PE:

  • The target array is updated.
  • The values in the pSync array are restored to the original values. ✨

💅 NOTES 💅

The terms collective, symmetric, and cache aligned are all defined in intro_shmem(3) — go read that if you need the lore. 📖

All SHMEM reduction routines reset the values in pSync before they return, so a particular pSync buffer only needs to be initialized the first time it's used. one and done, we love a low-maintenance array.

You must ensure that the pSync array is not being updated on any PE in the active set while any of the PEs participate in processing of a SHMEM reduction routine. Be careful of the following situations:

  • If the pSync array is initialized at run time, some type of synchronization is needed to ensure that all PEs in the working set have initialized pSync before any of them enter a SHMEM routine called with the pSync synchronization array.
  • A pSync or pWrk array can be reused in a subsequent reduction routine call only if none of the PEs in the active set are still processing a prior reduction routine call that used the same pSync or pWrk arrays.
😭 In general, this can be assured only by doing some type of synchronization.

BUT. here's the slay: in the special case of reduction routines being called with the same active set, you can allocate two pSync and pWrk arrays and alternate between them on successive calls. double the arrays, double the (safe) reuse. 👑

🦋 EXAMPLES 🦋

Example 1: This Fortran example statically initializes the pSync array and finds the maximum value of real variable FOO across all even PEs.

INCLUDE "mpp/shmem.fh"

INTEGER PSYNC(SHMEM_REDUCE_SYNC_SIZE)
DATA PSYNC /SHMEM_REDUCE_SYNC_SIZE*SHMEM_SYNC_VALUE/
PARAMETER (NR=1)
REAL FOO, FOOMAX, PWRK(MAX(NR/2+1,SHMEM_REDUCE_MIN_WRKDATA_SIZE))
COMMON /COM/ FOO, FOOMAX, PWRK
INTRINSIC MY_PE

IF ( MOD(MY_PE(),2) .EQ. 0) THEN
  CALL SHMEM_REAL8_MAX_TO_ALL(FOOMAX, FOO, NR, 0, 1, N$PES/2,
  & PWRK, PSYNC)
  PRINT *, 'Result on PE ', MY_PE(), ' is ', FOOMAX
ENDIF

Example 2: Consider the following C/C++ call:

shmem_int_max_to_all( target, source, 3, 0, 0, 8, pwrk, psync );

The preceding call is more efficient than, but semantically equivalent to, the combination of the following calls:

shmem_int_max_to_all(&(target[0]), &(source[0]), 1, 0, 0, 8,
  pwrk1, psync1);
shmem_int_max_to_all(&(target[1]), &(source[1]), 1, 0, 0, 8,
  pwrk2, psync2);
shmem_int_max_to_all(&(target[2]), &(source[2]), 1, 0, 0, 8,
  pwrk1, psync1);

Note that two sets of pWrk and pSync arrays are used alternately, because no synchronization is done between calls. that's the alternating trick from the notes, in the wild. 🌈

See also:

2003-2026, The Open MPI Community

and that's the whole page, bestie 🫶 this one came from Open MPI, dated Sep 10, 2026, filed under SHMEM_DOUBLE_MAX_TO_ALL(3) — go forth, take the max, and never forget to initialize your pSync. 💖✨🎀

📦 where this page comes from