source: soft/giet_vm/applications/fft/fft.py

#!/usr/bin/env python

from mapping import *

######################################################################################
#   file   : fft.py  (for the fft application)
#   date   : mars 2016
#   author : Alain Greiner
#######################################################################################
#  This file describes the mapping of the multi-threaded "fft" 
#  application on a multi-clusters, multi-processors architecture.
#  There is one thread per processor.
#  The mapping of virtual segments is the following:
#    - There is one shared data vseg in cluster[0][0]
#    - The code vsegs are replicated on all clusters containing processors.
#    - There is one heap vseg per cluster containing processors.
#    - The stacks vsegs are distributed on all clusters containing processors.
#  This mapping uses 5 platform parameters, (obtained from the "mapping" argument)
#  - x_size    : number of clusters in a row
#  - y_size    : number of clusters in a column
#  - x_width   : number of bits coding x coordinate
#  - y_width   : number of bits coding y coordinate
#  - nprocs : number of processors per cluster
####################################################################################

######################
def extend( mapping ):

    x_size    = mapping.x_size
    y_size    = mapping.y_size
    nprocs    = mapping.nprocs
    x_width   = mapping.x_width
    y_width   = mapping.y_width

    # define vsegs base & size
    code_base  = 0x10000000
    code_size  = 0x00010000     # 64 Kbytes (replicated in each cluster)
    
    data_base  = 0x20000000
    data_size  = 0x00010000     # 64 Kbytes (non replicated)

    heap_base  = 0x40000000     
    heap_size  = 0x02000000     # 16 Mbytes (per cluster)

    stack_base = 0x30000000     
    stack_size = 0x00040000     # 256 Kbytes (per thread) 

    # create Vspace
    vspace = mapping.addVspace( name = 'fft', startname = 'fft_data' , active = True )
    
    # data vseg in cluster[0,0] : non local
    mapping.addVseg( vspace, 'fft_data', data_base , data_size, 
                     'C_WU', vtype = 'ELF', x = 0 , y = 0 , pseg = 'RAM', 
                     binpath = 'bin/fft/appli.elf',
                     local = False )

    # code vsegs : local (one copy in each cluster)
    for x in xrange (x_size):
        for y in xrange (y_size):
            cluster_id = (x * y_size) + y
            if ( mapping.clusters[cluster_id].procs ):

                mapping.addVseg( vspace, 'fft_code_%d_%d' % (x,y), code_base, code_size,
                                 'CXWU', vtype = 'ELF', x = x , y = y , pseg = 'RAM', 
                                 binpath = 'bin/fft/appli.elf', 
                                 local = True )

    # stack vsegs : local (one stack per processor)
    for x in xrange (x_size):
        for y in xrange (y_size):
            cluster_id = (x * y_size) + y
            if ( mapping.clusters[cluster_id].procs ):
                for p in xrange( nprocs ):
                    proc_id = (((x * y_size) + y) * nprocs) + p
                    size    = stack_size - 4096
                    base    = stack_base + (proc_id * stack_size)

                    mapping.addVseg( vspace, 'fft_stack_%d_%d_%d' % (x,y,p), base, size,
                                 'C_WU', vtype = 'BUFFER', x = x , y = y , pseg = 'RAM',
                                 local = True )
            
    # heap vsegs : distributed but non local (all heap vsegs can be accessed by all tasks)
    for x in xrange (x_size):
        for y in xrange (y_size):
            cluster_id = (x * y_size) + y
            if ( mapping.clusters[cluster_id].procs ):
                size       = heap_size
                base       = heap_base + (cluster_id * size)

                mapping.addVseg( vspace, 'fft_heap_%d_%d' % (x,y), base, size,
                                 'C_WU', vtype = 'HEAP', x = x , y = y , pseg = 'RAM',
                                 local = False, big = True )

    # distribute one thread per processor / Main on P[0,0,0]
    for x in xrange (x_size):
        for y in xrange (y_size):
            cluster_id = (x * y_size) + y
            if ( mapping.clusters[cluster_id].procs ):
                for p in xrange( nprocs ):
                    if x == 0 and y == 0 and p == 0 :    # main
                        startid = 1
                        is_main = True
                    else :                               # slaves
                        startid = 0
                        is_main = False
                    
                    mapping.addThread( vspace, 
                                       'fft_%d_%d_%d' % (x,y,p),
                                       is_main,
                                       x, y, p,
                                       'fft_stack_%d_%d_%d' % (x,y,p),
                                       'fft_heap_%d_%d' % (x,y),
                                       startid )

    # extend mapping name
    mapping.name += '_fft'

    return vspace  # useful for test
            
################################ test ######################################################

if __name__ == '__main__':

    vspace = fft( Mapping( 'test', 2, 2, 4 ) )
    print vspace.xml()


# Local Variables:
# tab-width: 4;
# c-basic-offset: 4;
# c-file-offsets:((innamespace . 0)(inline-open . 0));
# indent-tabs-mode: nil;
# End:
#
# vim: filetype=python:expandtab:shiftwidth=4:tabstop=4:softtabstop=4