Changeset 720 for soft/giet_vm/applications/router/router.py

Timestamp:

Oct 11, 2015, 6:23:18 PM (9 years ago)

Author:

alain

Message:

Adapt the router application to the POSIX API.

File:

• soft/giet_vm/applications/router/router.py (modified) (3 diffs)

soft/giet_vm/applications/router/router.py

@@ -10 +10 @@
 #  This file describes the mapping of the multi-threaded "router" 
 #  application on a multi-clusters, multi-processors architecture.
-#  This application contains N+2 parallel tasks communicating through MWMR channels:
+#  This application contains N+2 parallel threads communicating through MWMR channels:
 #  The mapping of virtual segments on the clusters is the following:
+#    - the data vseg is mapped on cluster[0,0].
 #    - The code vsegs are replicated on all clusters.
-#    - the data_0 vseg is mapped on cluster[0,0].
-#    - the data_1 vseg is mapped on cluster[x_size-1,y_size-1].
-#    - the stacks vsegs are distibuted on all clusters.
-#  The mapping of tasks on processors is the following:
-#    - one "producer" task  => on proc[0,0,0]
-#    - one "consume"  task  => on proc[x_size-1,y_size-1,nprocs-1]
-#    - N   "router"   tasks => on all others processors
+#    - the stack vsegs are distibuted on all clusters.
+#    - the heap vsegs are distributed on all clusters
+#  The mapping of threads on processors is the following:
+#    - the "main" thread      => on proc[0,0,0]
+#    - the "producer" thread  => on proc[0,0,0]
+#    - the "consume"  thread  => on proc[0,0,1]
+#    - N   "compute"  threads => on all others processors
 #  This mapping uses 5 platform parameters, (obtained from the "mapping" argument)
 #    - x_size    : number of clusters in a row
@@ -39 +40 @@
     code_size    = 0x00010000     # 64 Kbytes (replicated in each cluster)
     
-    data_0_base  = 0x20000000
-    data_0_size  = 0x00010000     # 64 Kbytes (non replicated)
+    data_base    = 0x20000000
+    data_size    = 0x00010000     # 64 Kbytes (non replicated)
 
-    data_1_base  = 0x30000000
-    data_1_size  = 0x00010000     # 64 Kbytes (non replicated)
+    stack_base   = 0x30000000 
+    stack_size   = 0x00010000     # 64 Kbytes (per thread)
 
-    stack_base   = 0x40000000 
-    stack_size   = 0x00200000     # 2 Mbytes (per cluster)
+    heap_base    = 0x40000000
+    heap_size    = 0x00200000     # 2 Mbytes (per cluster)
 
     # create vspace
-    vspace = mapping.addVspace( name = 'router', startname = 'router_data_0' )
+    vspace = mapping.addVspace( name = 'router', 
+                                startname = 'router_data',
+                                active = False )
     
-    # data_0 vseg : shared / in cluster [0,0]
-    mapping.addVseg( vspace, 'router_data_0', data_0_base , data_0_size, 
+    # data vseg : shared / in cluster [0,0]
+    mapping.addVseg( vspace, 'router_data', data_base , data_size, 
                      'C_WU', vtype = 'ELF', x = 0, y = 0, pseg = 'RAM', 
-                     binpath = 'bin/router/appli.elf',
-                     local = False )
-
-    # data_1 vseg : shared / in cluster[x_size-1,y_size-1]
-    mapping.addVseg( vspace, 'router_data_1', data_1_base , data_1_size, 
-                     'C_WU', vtype = 'ELF', x = x_size - 1, y = y_size - 1, pseg = 'RAM', 
                      binpath = 'bin/router/appli.elf',
                      local = False )
@@ -71 +68 @@
                              local = True )
 
-    # stacks vsegs: local (one stack per processor => nprocs stacks per cluster)            
+    # heap vsegs : shared (one per cluster) 
     for x in xrange (x_size):
         for y in xrange (y_size):
-            for p in xrange( nprocs ):
-                proc_id = (((x * y_size) + y) * nprocs) + p
-                size    = stack_size / nprocs
-                base    = stack_base + (proc_id * size)
-                mapping.addVseg( vspace, 'router_stack_%d_%d_%d' % (x,y,p), base, size,
-                                 'C_WU', vtype = 'BUFFER', x = x , y = y , pseg = 'RAM',
-                                 local = True, big = True )
+            cluster_id = (x * y_size) + y
+            if ( mapping.clusters[cluster_id].procs ):
+                size  = heap_size
+                base  = heap_base + (cluster_id * size)
 
-    # distributed tasks / one task per processor
+                mapping.addVseg( vspace, 'router_heap_%d_%d' %(x,y), base , size, 
+                                 'C_WU', vtype = 'HEAP', x = x, y = y, pseg = 'RAM', 
+                                 local = False, big = True )
+
+    # stacks vsegs: local (one stack per thread => nprocs stacks per cluster)
+    # ... plus main_stack in cluster[0][0]
+    mapping.addVseg( vspace, 'main_stack',
+                     stack_base, stack_size, 'C_WU', vtype = 'BUFFER', 
+                     x = 0 , y = 0 , pseg = 'RAM',
+                     local = True )
+
     for x in xrange (x_size):
         for y in xrange (y_size):
-            for p in xrange( nprocs ):
-                trdid = (((x * y_size) + y) * nprocs) + p
-                if   (x==0) and (y==0) and (p== 0):                      # task producer
-                    task_index = 2
-                    task_name  = 'producer'            
-                elif (x==x_size-1) and (y==y_size-1) and (p==nprocs-1):  # task consumer 
-                    task_index = 1  
-                    task_name  = 'consumer'
-                else :                                                   # task router
-                    task_index = 0
-                    task_name  = 'router_%d_%d_%d' % (x,y,p)
-                mapping.addTask( vspace, task_name, trdid, x, y, p,
-                                 'router_stack_%d_%d_%d' % (x,y,p), '' , task_index )
+            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
+                    base    = stack_base + (proc_id * stack_size) + stack_size 
+
+                    mapping.addVseg( vspace, 'router_stack_%d_%d_%d' % (x,y,p), 
+                                     base, stack_size, 'C_WU', vtype = 'BUFFER', 
+                                     x = x , y = y , pseg = 'RAM',
+                                     local = True )
+
+    # distributed threads / one thread per processor
+    # ... plus main on P[0][0][0]
+    mapping.addThread( vspace, 'main', True, 0, 0, 1,
+                       'main_stack',
+                       'router_heap_0_0',
+                       0 )                      # index in start_vector
+
+    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):  # thread producer
+                        start_index = 3
+                        thread_name  = 'producer_0_0_0'            
+                    elif (x==0) and (y==0) and (p==1):  # thread consumer 
+                        start_index = 2  
+                        thread_name  = 'consumer_0_0_1'
+                    else :                              # thread compute
+                        start_index = 1
+                        thread_name  = 'compute_%d_%d_%d' % (x,y,p)
+
+                    mapping.addThread( vspace, thread_name, False , x, y, p,
+                                      'router_stack_%d_%d_%d' % (x,y,p),
+                                      'router_heap_%d_%d' %(x,y),
+                                      start_index )
 
     # extend mapping name