| 1 | #!/usr/bin/env python |
|---|
| 2 | |
|---|
| 3 | from mapping import * |
|---|
| 4 | |
|---|
| 5 | ################################################################################### |
|---|
| 6 | # file : classif.py |
|---|
| 7 | # date : november 2014 |
|---|
| 8 | # author : Alain Greiner |
|---|
| 9 | ################################################################################### |
|---|
| 10 | # This file describes the mapping of the multi-threaded "classif" |
|---|
| 11 | # application on a multi-clusters, multi-processors architecture. |
|---|
| 12 | # The mapping of threads on processors is the following: |
|---|
| 13 | # - the "main" on cluster[0][0] |
|---|
| 14 | # - one "load" thread per cluster containing processors, |
|---|
| 15 | # - one "store" thread per cluster containing processors, |
|---|
| 16 | # - (nprocs-2) "analyse" thread per cluster containing processors. |
|---|
| 17 | # The mapping of virtual segments is the following: |
|---|
| 18 | # - There is one shared data vseg in cluster[0][0] |
|---|
| 19 | # - The code vsegs are replicated on all clusters containing processors. |
|---|
| 20 | # - There is one heap vseg per cluster containing processors. |
|---|
| 21 | # - The stacks vsegs are distibuted on all clusters containing processors. |
|---|
| 22 | # This mapping uses 5 platform parameters, (obtained from the "mapping" argument) |
|---|
| 23 | # - x_size : number of clusters in a row |
|---|
| 24 | # - y_size : number of clusters in a column |
|---|
| 25 | # - x_width : number of bits for x field |
|---|
| 26 | # - y_width : number of bits for y field |
|---|
| 27 | # - nprocs : number of processors per cluster |
|---|
| 28 | # |
|---|
| 29 | # WARNING: The target architecture cannot contain less |
|---|
| 30 | # than 3 processors per cluster. |
|---|
| 31 | ################################################################################## |
|---|
| 32 | |
|---|
| 33 | ###################### |
|---|
| 34 | def extend( mapping ): |
|---|
| 35 | |
|---|
| 36 | x_size = mapping.x_size |
|---|
| 37 | y_size = mapping.y_size |
|---|
| 38 | nprocs = mapping.nprocs |
|---|
| 39 | x_width = mapping.x_width |
|---|
| 40 | y_width = mapping.y_width |
|---|
| 41 | |
|---|
| 42 | assert (nprocs >= 3) and (nprocs <= 8) |
|---|
| 43 | |
|---|
| 44 | # define vsegs base & size |
|---|
| 45 | code_base = 0x10000000 |
|---|
| 46 | code_size = 0x00010000 # 64 Kbytes (per cluster) |
|---|
| 47 | |
|---|
| 48 | data_base = 0x20000000 |
|---|
| 49 | data_size = 0x00010000 # 64 Kbytes (non replicated) |
|---|
| 50 | |
|---|
| 51 | heap_base = 0x30000000 |
|---|
| 52 | heap_size = 0x00200000 # 2M bytes (per cluster) |
|---|
| 53 | |
|---|
| 54 | stack_base = 0x40000000 |
|---|
| 55 | stack_size = 0x00010000 # 64 Kbytes (per thread) |
|---|
| 56 | |
|---|
| 57 | # create vspace |
|---|
| 58 | vspace = mapping.addVspace( name = 'classif', startname = 'classif_data', active = False ) |
|---|
| 59 | |
|---|
| 60 | # data vseg : shared / cluster[0][0] |
|---|
| 61 | mapping.addVseg( vspace, 'classif_data', data_base , data_size, |
|---|
| 62 | 'C_WU', vtype = 'ELF', x = 0, y = 0, pseg = 'RAM', |
|---|
| 63 | binpath = 'bin/classif/appli.elf', |
|---|
| 64 | local = False ) |
|---|
| 65 | |
|---|
| 66 | # heap vsegs : shared (one per cluster) |
|---|
| 67 | for x in xrange (x_size): |
|---|
| 68 | for y in xrange (y_size): |
|---|
| 69 | cluster_id = (x * y_size) + y |
|---|
| 70 | if ( mapping.clusters[cluster_id].procs ): |
|---|
| 71 | size = heap_size |
|---|
| 72 | base = heap_base + (cluster_id * size) |
|---|
| 73 | |
|---|
| 74 | mapping.addVseg( vspace, 'classif_heap_%d_%d' %(x,y), base , size, |
|---|
| 75 | 'C_WU', vtype = 'HEAP', x = x, y = y, pseg = 'RAM', |
|---|
| 76 | local = False, big = True ) |
|---|
| 77 | |
|---|
| 78 | # code vsegs : local (one copy per cluster) |
|---|
| 79 | for x in xrange (x_size): |
|---|
| 80 | for y in xrange (y_size): |
|---|
| 81 | cluster_id = (x * y_size) + y |
|---|
| 82 | if ( mapping.clusters[cluster_id].procs ): |
|---|
| 83 | |
|---|
| 84 | mapping.addVseg( vspace, 'classif_code_%d_%d' %(x,y), |
|---|
| 85 | code_base , code_size, |
|---|
| 86 | 'CXWU', vtype = 'ELF', x = x, y = y, pseg = 'RAM', |
|---|
| 87 | binpath = 'bin/classif/appli.elf', |
|---|
| 88 | local = True ) |
|---|
| 89 | |
|---|
| 90 | # stacks vsegs: local (one stack per processor => nprocs stacks per cluster) |
|---|
| 91 | # ... plus main_stack in cluster[0][0] |
|---|
| 92 | mapping.addVseg( vspace, 'main_stack', |
|---|
| 93 | stack_base, stack_size, 'C_WU', vtype = 'BUFFER', |
|---|
| 94 | x = 0 , y = 0 , pseg = 'RAM', |
|---|
| 95 | local = True ) |
|---|
| 96 | |
|---|
| 97 | for x in xrange (x_size): |
|---|
| 98 | for y in xrange (y_size): |
|---|
| 99 | cluster_id = (x * y_size) + y |
|---|
| 100 | if ( mapping.clusters[cluster_id].procs ): |
|---|
| 101 | for p in xrange( nprocs ): |
|---|
| 102 | proc_id = (((x * y_size) + y) * nprocs) + p |
|---|
| 103 | base = stack_base + (proc_id * stack_size) + stack_size |
|---|
| 104 | |
|---|
| 105 | mapping.addVseg( vspace, 'classif_stack_%d_%d_%d' % (x,y,p), |
|---|
| 106 | base, stack_size, 'C_WU', vtype = 'BUFFER', |
|---|
| 107 | x = x , y = y , pseg = 'RAM', |
|---|
| 108 | local = True ) |
|---|
| 109 | |
|---|
| 110 | # distributed threads / one thread per processor |
|---|
| 111 | # ... plus main on P[0][0][0] |
|---|
| 112 | mapping.addThread( vspace, 'main', True, 0, 0, 1, |
|---|
| 113 | 'main_stack', |
|---|
| 114 | 'classif_heap_0_0', |
|---|
| 115 | 0 ) # index in start_vector |
|---|
| 116 | |
|---|
| 117 | for x in xrange (x_size): |
|---|
| 118 | for y in xrange (y_size): |
|---|
| 119 | cluster_id = (x * y_size) + y |
|---|
| 120 | if ( mapping.clusters[cluster_id].procs ): |
|---|
| 121 | for p in xrange( nprocs ): |
|---|
| 122 | if ( p== 0 ): # thread load |
|---|
| 123 | start_index = 3 |
|---|
| 124 | thread_name = 'load_%d_%d_%d' %(x,y,p) |
|---|
| 125 | elif ( p== 1 ): # thread store |
|---|
| 126 | start_index = 2 |
|---|
| 127 | thread_name = 'stor_%d_%d_%d' %(x,y,p) |
|---|
| 128 | else : # thread analyse |
|---|
| 129 | start_index = 1 |
|---|
| 130 | thread_name = 'anal_%d_%d_%d' % (x,y,p) |
|---|
| 131 | |
|---|
| 132 | mapping.addThread( vspace, thread_name, False , x, y, p, |
|---|
| 133 | 'classif_stack_%d_%d_%d' % (x,y,p), |
|---|
| 134 | 'classif_heap_%d_%d' % (x,y), |
|---|
| 135 | start_index ) # index in start_vector |
|---|
| 136 | |
|---|
| 137 | # extend mapping name |
|---|
| 138 | mapping.name += '_classif' |
|---|
| 139 | |
|---|
| 140 | return vspace # useful for test |
|---|
| 141 | |
|---|
| 142 | ################################ test ############################################ |
|---|
| 143 | |
|---|
| 144 | if __name__ == '__main__': |
|---|
| 145 | |
|---|
| 146 | vspace = extend( Mapping( 'test', 2, 2, 4 ) ) |
|---|
| 147 | print vspace.xml() |
|---|
| 148 | |
|---|
| 149 | |
|---|
| 150 | # Local Variables: |
|---|
| 151 | # tab-width: 4; |
|---|
| 152 | # c-basic-offset: 4; |
|---|
| 153 | # c-file-offsets:((innamespace . 0)(inline-open . 0)); |
|---|
| 154 | # indent-tabs-mode: nil; |
|---|
| 155 | # End: |
|---|
| 156 | # |
|---|
| 157 | # vim: filetype=python:expandtab:shiftwidth=4:tabstop=4:softtabstop=4 |
|---|
| 158 | |
|---|