More tweaking, a hidden and relevant comment moved to a less invisible place.

In principle the machinery should turn task groups on when there are too many
processors for FFT parallelization. In practice, I don't think it does that,
otherwise it would stop on those cases that have no task group implementation
(e.g. metaGGA). I don't yet know what is really happening and whether the old
task groups are still present in the code, reachable and working.

Modules/command_line_options.f90

@@ -20,8 +20,12 @@ MODULE command_line_options
   !
   ! ... Number of arguments in command line
   INTEGER :: nargs = 0
-  ! ... QE arguments read from command line
-  INTEGER :: nimage_= 1, npool_= 0, ndiag_ = 0, nband_= 1, ntg_= 1, nyfft_ = 1, nmany_ = 1
+  ! ... QE arguments read from command line, default 1
+  INTEGER :: nimage_= 1, nband_= 1, nyfft_ = 1, nmany_ = 1
+  ! ... As above, default 0, in order to distinguish the "not set" and the
+  ! ... "set to 1" cases - useful for automatically choosing those values
+  INTEGER :: npool_= 0, ndiag_ = 0, ntg_= 0
+  ! ... Undocumented options
   LOGICAL :: pencil_decomposition_ = .false., rmm_with_paro_ = .false. 
   ! ... Indicate if using library init
   LOGICAL :: library_init = .FALSE.
@@ -101,16 +105,7 @@ CONTAINS
               ENDIF
               READ ( arg, *, ERR = 15, END = 15) npool_
               narg = narg + 1
-! FIXME: following comment should be moved to a more visible place
-! special case : task group parallelization and nyfft parallelization, both 
-!                introduced to improve scaling coexist and are in part interchangeable
-!                if TG is available it's faster that NYFFT becouse it communicates larger
-!                data chuncks less times. But sometimes it is not available as for instance
-!                when metagga is used or realus or for conjugate gradient. nyfft can be used.
-!-ntg and -nyfft are both alloved flags set the same value for both ntg and nyfft. 
-!                These variables are kept separated to help understanding which operation belong
-!                to TG or to NYFFT. This can enable to make them different if the need arises.
-!
+
            CASE ( '-nt', '-ntg', '-ntask_groups', '-nyfft')   
               IF (read_string) THEN
                  CALL my_getarg ( input_command_line, narg, arg )

Modules/mp_global.f90

@@ -52,6 +52,17 @@ CONTAINS
     !                      convenient to call it in serial execution as well
     !  IMPORTANT NOTICE 2: most parallelization levels are initialized here 
     !                      but at least some will be moved to a later stage
+    !  SPECIAL CASE: command-line options "-ntg" and "-nyfft", introduced to
+    !                improve scaling, coexist and are in part interchangeable.
+    !                If task groups are available, -ntg is faster than -nyfft
+    !                because it communicates larger data chuncks less frequently
+    !                Sometimes task groups are not available as for instance
+    !                when metagga is used or realus or for conjugate gradient.
+    !                For those cases, -nyfft can be used instead.
+    !                You may specify one or another: the same value will be set
+    !                for both ntg and nyfft. These variables are kept separated
+    !                to help understanding which operation belong to task groups
+    !                or to nyfft, allowing to differenciate them if need arises.
     !
     USE command_line_options, ONLY : get_command_line, &
         nimage_, npool_, nband_, ntg_, nyfft_
@@ -88,6 +99,8 @@ CONTAINS
     ! npool_ is 0 if not specified in command line
     IF ( npool_== 0 ) npool_ = 1
     CALL mp_start_pools ( npool_, intra_image_comm )
+    ! ntg_ is 0 if not specified in command line
+    IF ( ntg_== 0 ) ntg_ = 1
 #if defined (__CUDA_OPTIMIZED)
     CALL mp_start_bands ( 1 , ntg_, nyfft_, intra_pool_comm )
 #else

PW/src/setup.f90

@@ -696,6 +696,7 @@ SUBROUTINE setup_para ( nr3, nkstot, nbnd )
   !
   LOGICAL, EXTERNAL  :: check_gpu_support
   LOGICAL, SAVE :: first = .TRUE.
+  INTEGER :: maxtask
   !
   ! do not execute twice: unpredictable results may follow
   !
@@ -705,13 +706,13 @@ SUBROUTINE setup_para ( nr3, nkstot, nbnd )
   ! k-point parallelization first
   !
   IF ( npool_== 0 ) THEN
+     npool_ = 1
      !
-     ! check if too many mpi processes for this fft dimension,
-     ! use k-point parallelization if available
-     !
-     if ( nproc_image <= nr3/2 ) then
-        npool_ = 1
-     else
+     if ( nproc_image > nr3/2 .and. nkstot > 1 ) then
+        !
+        ! if too many mpi processes for this fft dimension,
+        ! use k-point parallelization if available
+        !
         do npool_ = 2, nkstot
            ! npool should be a divisor of the number of k-points
            if ( mod(nkstot, npool_) /= 0 ) cycle
@@ -735,17 +736,23 @@ SUBROUTINE setup_para ( nr3, nkstot, nbnd )
   use_gpu = check_gpu_support( )
   !
   ! Set "task_groups" if still too many processors for PW parallelization
+  !
+  IF ( ntask_groups == 0 ) THEN
+     ntask_groups = 1
+     if ( nproc_bgrp > nr3 ) THEN
+        maxtask = min (nbnd, 16)
+        do ntask_groups = 2, maxtask
+           if ( mod(nproc_bgrp,ntask_groups) == 0 .and. &
+                    nproc_bgrp/ntask_groups  <= nr3 .or.&
+                    nproc_bgrp/ntask_groups  <= nr3/4 .or.&
+                ntask_groups == maxtask ) exit
+        end do
+      end if
+  END IF
+  !
   ! Note that "task_groups" require to set pencil_decomposition to .true.
   !
-  IF ( ( ntask_groups /= 1 ) ) pencil_decomposition_ = .true. 
-  IF ( ( ntask_groups == 1 ) .AND. ( nproc_bgrp > nr3 ) ) THEN
-     pencil_decomposition_ = .true. 
-     do ntask_groups = 2, min(nbnd,16)
-        if ( mod(nproc_bgrp,ntask_groups) == 0 .and. &
-                 nproc_bgrp/ntask_groups  <= nr3 ) exit 
-        if ( ntask_groups == min(nbnd,16) ) exit
-     end do
-  END IF
+  IF ( ntask_groups /= 1 ) pencil_decomposition_ = .true. 
   !
   ! printout - same as in environment.f90
   !