From help@cs.rice.edu Thu Jun 1 10:44:43 1995 Return-Path: Received: from terka.cs.rice.edu by cs.rice.edu (KAA01654); Thu, 1 Jun 1995 10:44:42 -0500 Received: from localhost.cs.rice.edu by terka.cs.rice.edu (AA12423); Thu, 1 Jun 95 10:44:41 CDT Message-Id: <9506011544.AA12423@terka.cs.rice.edu> To: hpff-distribute@cs.rice.edu Cc: mmzn@terka.cs.rice.edu Subject: test for new hpff mail list - pls ignore Date: Thu, 01 Jun 95 10:44:40 -0500 From: mmzn@terka.cs.rice.edu piffle From help@cs.rice.edu Fri Jul 21 08:30:16 1995 Return-Path: Received: from squid.icase.edu by cs.rice.edu (IAA03009); Fri, 21 Jul 1995 08:30:15 -0500 Received: by squid.icase.edu (8.6.11/lanleaf8.6.4) id JAA20090; Fri, 21 Jul 1995 09:30:16 -0400 Message-Id: <199507211330.JAA20090@squid.icase.edu> Date: Fri, 21 Jul 1995 09:30:16 -0400 From: Piyush Mehrotra To: hpff-distribute@cs.rice.edu Subject: Proposal for mapping derived types Mapping of derived type components: Proposal by: Barbara Chapman, Piyush Mehrotra, Guy Robinson and Hans Zima Remove the first constraints under H303-311 and H312-318. That is, a distributee or an alignee may be a subobject designator. Add the constraint: Constraint: If a component of a derived type is mapped (either through a distribute or an align statement) then no other component below or above it in the derived type tree may be mapped. {the above constraint needs to use more precise language about "below" and "above" in the "derived type tree"} Note that a subcomponent of a derived type can be declared DYNAMIC and can be redistributed or realigned as long as the above constraint is not violated. Example: ! MULTI is an array of MESHes each of which has components of type GRID TYPE GRID .... REAL S(100) !HPF$ DISTRIBUTE S(BLOCK) END TYPE GRID TYPE MESH INTEGER SIZE REAL RV(100,100) INTEGER KV(100,100) ! Both RV and KV can be distributed !HPF$ DISTRIBUTE (BLOCK, BLOCK):: RV, KV TYPE(GRID) SG(100) !Illegal: SG cannot be distributed since a subobject (S) is already distributed !HPF$ DISTRIBUTE SG(BLOCK) END TYPE MESH TYPE(GRID) MULTI(20) !Illegal distribution !HPF$ DISTRIBUTE MULTI(BLOCK) Here, the derived type tree is as follows: MULTI | | MESH | | GRID Since a subobject of GRID is distributed and subobjects of MESH of TYPE GRID cannot be distributed. Other subobjects of MESH can be distributed. However, MULTI cannot be distributed since its subobjects are distributed. Here is another example: allocation of blocks in a multiblock code; !HPF$ PROCESSORS P( number_of_processors() ) TYPE SUBGRID INTEGER size INTEGER lo, hi ! target subset of processors REAL, POINTER bl(:) !HPF$ DYNAMIC bl END TYPE SUBGRID TYPE (SUBGRID), ALLOCATABLE:: grid(:) READ (*,*) gridnos ALLOCATE GRID(gridnos) DO i = 1, gridnos READ(*,*) isize grid(i)%size = isize END DO ! Computes processor subsets for each subgrid setting the lo, hi values CALL procs ( grid, number_of_processors()) ! Allocate each subgrid and distribute to computed subset DO i = 1, gridnos ALLOCATE(grid(i)%bl(isize)) !HPF$ REDISTRIBUTE grid(i)%bl(BLOCK) ONTO P( grid(i)%lo: grid(i)%hi) END DO .... From help@cs.rice.edu Fri Jul 21 08:30:36 1995 Return-Path: Received: from squid.icase.edu by cs.rice.edu (IAA03033); Fri, 21 Jul 1995 08:30:36 -0500 Received: by squid.icase.edu (8.6.11/lanleaf8.6.4) id JAA20094; Fri, 21 Jul 1995 09:30:36 -0400 Message-Id: <199507211330.JAA20094@squid.icase.edu> Date: Fri, 21 Jul 1995 09:30:36 -0400 From: Piyush Mehrotra To: hpff-distribute@cs.rice.edu Subject: Proposal for Generalized Blocks Generalized Block Distributions: Proposal by: Barbara Chapman, Piyush Mehrotra, Guy Robinson and Hans Zima TO H309 add: dist-format is BLOCK ... or CYCLIC .... or * or GEN_BLOCK (int-array) Constraint: int-array must be a one-dimensional integer array of size equal to the extent of the corresponding dimension of the target processor array. Constraint: The sum of the values of int-array must match the size of the dimension of the array being distributed using this distribution. This distribution maps contiguous blocks of the array elements onto processor with the sizes of the blocks being given by the integer array . Thus, if S(i) = lb + sum of k = 1 to i int-array(i), the lb is the lower bound of the array being distributed, then P(1) owns the array elements [ lb : S(1)-1] and P(i) owns array elements [ S(i-1)+1, S(i)]] Example: PARAMETER (START = /4,46,46,4/) !HPF$ PROCESSORS P(4) REAL A(100), B(200), new(4) !HPF$ DISTRIBUTE A( GEN_BLOCK( START) ) ONTO P !HPF$ DYNAMIC B ... new = ... !HPF$ REDISTRIBUTE ( B(GEN_BLOCK (new) ) Array elements A(1:4) are on P(1) A(5:50) are on P(2) A(51:96) are on P(3) A(97:100) are on P(4) General block distributions have been implemented in the Vienna Fortran Compilation System. From help@cs.rice.edu Fri Jul 21 08:31:20 1995 Return-Path: Received: from squid.icase.edu by cs.rice.edu (IAA03044); Fri, 21 Jul 1995 08:31:19 -0500 Received: by squid.icase.edu (8.6.11/lanleaf8.6.4) id JAA20098; Fri, 21 Jul 1995 09:31:20 -0400 Message-Id: <199507211331.JAA20098@squid.icase.edu> Date: Fri, 21 Jul 1995 09:31:20 -0400 From: Piyush Mehrotra To: hpff-distribute@cs.rice.edu Subject: Proposal for irregular mappings Irregular mapping: Proposal by: Barbara Chapman, Piyush Mehrotra, Guy Robinson and Hans Zima TO H309 add dist-format is BLOCK ... or CYCLIC .... or * or INDIRECT (int-array) Constraint: int-array must be a one-dimensional integer array of size equal to the extent of the corresponding dimension of the array being distributed. Here, the int-array acts as a mapping array and provides the processor number onto which the array element is mapped. Thus, array element i is mapped to processor index int-array(i). Example: PARAMETER (map = /1,2,4,3,3,2,1,4, ..../) !HPF$ PROCESSORS P(4) REAL A(100) INTEGER map(100) !HPF$ DISTRIBUTE A( INDIRECT(map) ) ONTO P Indirect distributions using map arrays have been implemented in the Vienna Fortran Compilation System. Mapping of derived type components: Remove the first constraints under H303-311 and H312-318. That is, a distributee or an alignee may be a subobject designator. Add the constraint: Constraint: If a component of a derived type is mapped (either through a distribute or an align statement) then no other component below or above it in the derived type tree may be mapped. {the above constraint needs to use more precise language about "below" and "above" in the "derived type tree"} Note that a subcomponent of a derived type can be declared DYNAMIC and can be redistributed or realigned as long as the above constraint is not violated. Example: ! MULTI is an array of MESHes each of which has components of type GRID TYPE GRID .... REAL S(100) !HPF$ DISTRIBUTE S(BLOCK) END TYPE GRID TYPE MESH INTEGER SIZE REAL RV(100,100) INTEGER KV(100,100) ! Both RV and KV can be distributed !HPF$ DISTRIBUTE (BLOCK, BLOCK):: RV, KV TYPE(GRID) SG(100) !Illegal: SG cannot be distributed since a subobject (S) is already distributed !HPF$ DISTRIBUTE SG(BLOCK) END TYPE MESH TYPE(GRID) MULTI(20) !Illegal distribution !HPF$ DISTRIBUTE MULTI(BLOCK) Here, the derived type tree is as follows: MULTI | | MESH | | GRID Since a subobject of GRID is distributed and subobjects of MESH of TYPE GRID cannot be distributed. Other subobjects of MESH can be distributed. However, MULTI cannot be distributed since its subobjects are distributed. Here is another example: allocation of blocks in a multiblock code; !HPF$ PROCESSORS P( number_of_processors() ) TYPE SUBGRID INTEGER size INTEGER lo, hi ! target subset of processors REAL, POINTER bl(:) !HPF$ DYNAMIC bl END TYPE SUBGRID TYPE (SUBGRID), ALLOCATABLE:: grid(:) READ (*,*) gridnos ALLOCATE GRID(gridnos) DO i = 1, gridnos READ(*,*) isize grid(i)%size = isize END DO ! Computes processor subsets for each subgrid setting the lo, hi values CALL procs ( grid, number_of_processors()) ! Allocate each subgrid and distribute to computed subset DO i = 1, gridnos ALLOCATE(grid(i)%bl(isize)) !HPF$ REDISTRIBUTE grid(i)%bl(BLOCK) ONTO P( grid(i)%lo: grid(i)%hi) END DO .... From help@cs.rice.edu Fri Jul 21 08:32:03 1995 Return-Path: Received: from squid.icase.edu by cs.rice.edu (IAA03053); Fri, 21 Jul 1995 08:32:02 -0500 Received: by squid.icase.edu (8.6.11/lanleaf8.6.4) id JAA20104; Fri, 21 Jul 1995 09:32:03 -0400 Message-Id: <199507211332.JAA20104@squid.icase.edu> Date: Fri, 21 Jul 1995 09:32:03 -0400 From: Piyush Mehrotra To: hpff-distribute@cs.rice.edu Subject: Proposal for mapping to processor subsets Mapping to Processor Subsets: Proposal by: Barbara Chapman, Piyush Mehrotra, Guy Robinson and Hans Zima Replace HPF syntax rule H311 by the following: dist target is processor-name [( subscript-triplet )] or * processor-name [( subscript-triplet )] or * Constraint: The subscript triplet must specify a subset of the named processor array. This allows a processor subset to be specified wherever a processor could be named, e.g., in distribute statements. Note that this does not allow processor subsets to be named. Example: !HPF$ PROCESSORS P(10) REAL A(100) !HPF$ DISTRIBUTE A(BLOCK) ONTO P(2:5) This distributes A onto processors P(2) to P(5). This facility has been implemented in the Vienna Fortran Compilation System. Extension: SUBGROUP statement for allowing subgroups to be named, i.e., restricted equivalencing of processors arrays SUBGROUP OF (subscript-triple) This allows to be a named subgroup of . From help@cs.rice.edu Fri Jul 21 16:54:53 1995 Return-Path: Received: from phoenix.ocf.llnl.gov by cs.rice.edu (QAA27310); Fri, 21 Jul 1995 16:54:52 -0500 Received: by phoenix.ocf.llnl.gov (5.65/LLNL-1.18) id AA08809; Fri, 21 Jul 1995 14:54:34 -0700 Date: Fri, 21 Jul 1995 14:54:34 -0700 From: zosel@phoenix.ocf.llnl.gov (Mary E Zosel) Message-Id: <9507212154.AA08809@phoenix.ocf.llnl.gov> To: hpff-distribute@cs.rice.edu Subject: test message - ignore test distribution. -mz-