| 1 | #!/usr/bin/env python |
|---|
| 2 | |
|---|
| 3 | from mapping import * |
|---|
| 4 | |
|---|
| 5 | ##################################################################################### |
|---|
| 6 | # file : convol.py |
|---|
| 7 | # date : may 2014 |
|---|
| 8 | # author : Alain Greiner |
|---|
| 9 | ##################################################################################### |
|---|
| 10 | # This file describes the mapping of the multi-threaded "convol" |
|---|
| 11 | # application on a multi-clusters, multi-processors architecture. |
|---|
| 12 | # This include both the mapping of virtual segments on the clusters, |
|---|
| 13 | # and the mapping of threads on processors. |
|---|
| 14 | # There is one thread per processor. |
|---|
| 15 | # The mapping of virtual segments is the following: |
|---|
| 16 | # - There is one shared data vseg in cluster[0][0] |
|---|
| 17 | # - The code vsegs are replicated on all clusters containing processors. |
|---|
| 18 | # - There is one heap vseg per cluster containing processors. |
|---|
| 19 | # - The stacks vsegs are distibuted on all clusters containing processors. |
|---|
| 20 | # This mapping uses 5 platform parameters, (obtained from the "mapping" argument) |
|---|
| 21 | # - x_size : number of clusters in a row |
|---|
| 22 | # - y_size : number of clusters in a column |
|---|
| 23 | # - x_width : number of bits coding x coordinate |
|---|
| 24 | # - y_width : number of bits coding y coordinate |
|---|
| 25 | # - nprocs : number of processors per cluster |
|---|
| 26 | #################################################################################### |
|---|
| 27 | |
|---|
| 28 | ###################### |
|---|
| 29 | def extend( mapping ): |
|---|
| 30 | |
|---|
| 31 | x_size = mapping.x_size |
|---|
| 32 | y_size = mapping.y_size |
|---|
| 33 | nprocs = mapping.nprocs |
|---|
| 34 | x_width = mapping.x_width |
|---|
| 35 | y_width = mapping.y_width |
|---|
| 36 | |
|---|
| 37 | # define vsegs base & size |
|---|
| 38 | code_base = 0x10000000 |
|---|
| 39 | code_size = 0x00010000 # 64 Kbytes (replicated in each cluster) |
|---|
| 40 | |
|---|
| 41 | data_base = 0x20000000 |
|---|
| 42 | data_size = 0x00020000 # 128 Kbytes (non replicated) |
|---|
| 43 | |
|---|
| 44 | heap_base = 0x40000000 |
|---|
| 45 | heap_size = 0x01000000 # 16 Mbytes (per cluster) |
|---|
| 46 | |
|---|
| 47 | stack_base = 0x50000000 |
|---|
| 48 | stack_size = 0x00200000 # 2 Mbytes (per cluster) |
|---|
| 49 | |
|---|
| 50 | # create Vspace |
|---|
| 51 | vspace = mapping.addVspace( name = 'convol', startname = 'conv_data', active = False ) |
|---|
| 52 | |
|---|
| 53 | # data vseg in cluster[0,0] : non local |
|---|
| 54 | mapping.addVseg( vspace, 'conv_data', data_base , data_size, |
|---|
| 55 | 'C_WU', vtype = 'ELF', x = 0, y = 0, pseg = 'RAM', |
|---|
| 56 | binpath = 'bin/convol/appli.elf', |
|---|
| 57 | local = False ) |
|---|
| 58 | |
|---|
| 59 | # code vsegs : local (one copy in each cluster) |
|---|
| 60 | for x in xrange (x_size): |
|---|
| 61 | for y in xrange (y_size): |
|---|
| 62 | cluster_id = (x * y_size) + y |
|---|
| 63 | if ( mapping.clusters[cluster_id].procs ): |
|---|
| 64 | size = code_size |
|---|
| 65 | base = code_base |
|---|
| 66 | |
|---|
| 67 | mapping.addVseg( vspace, 'conv_code_%d_%d' % (x,y), base, size, |
|---|
| 68 | 'CXWU', vtype = 'ELF', x = x , y = y , pseg = 'RAM', |
|---|
| 69 | binpath = 'bin/convol/appli.elf', |
|---|
| 70 | local = True ) |
|---|
| 71 | |
|---|
| 72 | # stack vsegs : local (one stack per processor) |
|---|
| 73 | for x in xrange (x_size): |
|---|
| 74 | for y in xrange (y_size): |
|---|
| 75 | cluster_id = (x * y_size) + y |
|---|
| 76 | if ( mapping.clusters[cluster_id].procs ): |
|---|
| 77 | for p in xrange( nprocs ): |
|---|
| 78 | proc_id = (((x * y_size) + y) * nprocs) + p |
|---|
| 79 | size = (stack_size / nprocs) & 0xFFFFF000 |
|---|
| 80 | base = stack_base + (proc_id * size) |
|---|
| 81 | |
|---|
| 82 | mapping.addVseg( vspace, 'conv_stack_%d_%d_%d' % (x,y,p), |
|---|
| 83 | base, size, 'C_WU', vtype = 'BUFFER', |
|---|
| 84 | x = x , y = y , pseg = 'RAM', |
|---|
| 85 | local = True, big = True ) |
|---|
| 86 | |
|---|
| 87 | # heap vsegs : distributed but non local (any heap can be accessed by any thread) |
|---|
| 88 | for x in xrange (x_size): |
|---|
| 89 | for y in xrange (y_size): |
|---|
| 90 | cluster_id = (x * y_size) + y |
|---|
| 91 | if ( mapping.clusters[cluster_id].procs ): |
|---|
| 92 | size = heap_size |
|---|
| 93 | base = heap_base + (cluster_id * size) |
|---|
| 94 | |
|---|
| 95 | mapping.addVseg( vspace, 'conv_heap_%d_%d' % (x,y), base, size, |
|---|
| 96 | 'C_WU', vtype = 'HEAP', x = x , y = y , pseg = 'RAM', |
|---|
| 97 | local = False, big = True ) |
|---|
| 98 | |
|---|
| 99 | # distributed threads : one thread per processor |
|---|
| 100 | for x in xrange (x_size): |
|---|
| 101 | for y in xrange (y_size): |
|---|
| 102 | cluster_id = (x * y_size) + y |
|---|
| 103 | if ( mapping.clusters[cluster_id].procs ): |
|---|
| 104 | for p in xrange( nprocs ): |
|---|
| 105 | if (x == 0) and (y == 0) and (p == 0) : # main thread |
|---|
| 106 | startid = 1 |
|---|
| 107 | is_main = True |
|---|
| 108 | else : # trsp thread |
|---|
| 109 | startid = 0 |
|---|
| 110 | is_main = False |
|---|
| 111 | |
|---|
| 112 | mapping.addThread( vspace, |
|---|
| 113 | 'conv_%d_%d_%d' % (x,y,p), |
|---|
| 114 | is_main, |
|---|
| 115 | x, y, p, |
|---|
| 116 | 'conv_stack_%d_%d_%d' % (x,y,p), |
|---|
| 117 | 'conv_heap_%d_%d' % (x,y), |
|---|
| 118 | startid ) |
|---|
| 119 | |
|---|
| 120 | # extend mapping name |
|---|
| 121 | mapping.name += '_convol' |
|---|
| 122 | |
|---|
| 123 | return vspace # useful for test |
|---|
| 124 | |
|---|
| 125 | ################################ test ################################################ |
|---|
| 126 | |
|---|
| 127 | if __name__ == '__main__': |
|---|
| 128 | |
|---|
| 129 | vspace = extend( Mapping( 'test', 2, 2, 4 ) ) |
|---|
| 130 | print vspace.xml() |
|---|
| 131 | |
|---|
| 132 | |
|---|
| 133 | # Local Variables: |
|---|
| 134 | # tab-width: 4; |
|---|
| 135 | # c-basic-offset: 4; |
|---|
| 136 | # c-file-offsets:((innamespace . 0)(inline-open . 0)); |
|---|
| 137 | # indent-tabs-mode: nil; |
|---|
| 138 | # End: |
|---|
| 139 | # |
|---|
| 140 | # vim: filetype=python:expandtab:shiftwidth=4:tabstop=4:softtabstop=4 |
|---|
| 141 | |
|---|
