Task parallelization

src/command_line.py

@@ -37,6 +37,8 @@ options['h'] = [ 'help'         , 'Print this help', 0 ]
 options['i'] = [ 'init'         , 'Initialize current directory', 0 ]
 options['D'] = [ 'define'       , 'Define variable', 1 ]
 options['p'] = [ 'preprocess'   , 'Preprocess file', 1 ]
+options['o'] = [ 'openmp'       , 'Auto-parallelization', 0 ]
+options['m'] = [ 'memory'       , 'Debug memory', 0 ]
 
 class CommandLine(object):
 

src/irp_stack.py

@@ -30,6 +30,7 @@ from command_line import command_line
 
 do_assert = command_line.do_assert
 do_debug = command_line.do_debug
+do_openmp = command_line.do_openmp
 
 import irpf90_t
 
@@ -52,8 +53,9 @@ subroutine irp_enter(irp_where)
  integer       :: ithread
  integer       :: nthread
  character*(*) :: irp_where
- ithread = 0
- nthread = 1
+$OMP_DECL
+ ithread = $OMP_GET_THREAD_NUM
+ nthread = $OMP_GET_NUM_THREADS
 $1
 $2
 end subroutine
@@ -63,12 +65,26 @@ subroutine irp_leave (irp_where)
   character*(*) :: irp_where
   integer       :: ithread
   double precision :: cpu
-  ithread = 0
+$OMP_DECL
+  ithread = $OMP_GET_THREAD_NUM
 $3
 $4
 end subroutine
 """
 
+  # $OMP_DECL
+  if do_openmp:
+    txt = txt.replace("$OMP_DECL","""
+  integer :: omp_get_num_threads
+  integer :: omp_get_thread_num
+""")
+    txt = txt.replace("$OMP_GET_NUM_THREADS","omp_get_num_threads()")
+    txt = txt.replace("$OMP_GET_THREAD_NUM","omp_get_thread_num()")
+  else:
+    txt = txt.replace("$OMP_DECL","")
+    txt = txt.replace("$OMP_GET_NUM_THREADS","1")
+    txt = txt.replace("$OMP_GET_THREAD_NUM","0")
+
   # $1
   if do_assert or do_debug:
     txt = txt.replace("$1","""
@@ -82,6 +98,11 @@ end subroutine
 !$OMP END CRITICAL
  stack_index(ithread+1) = stack_index(ithread+1)+1
  irp_stack(stack_index(ithread+1),ithread+1) = irp_where""")
+    if command_line.do_memory:
+      txt+="""
+  print *, 'Allocating irp_stack(',STACKMAX,','nthread,')'
+  print *, 'Allocating irp_cpu(',STACKMAX,','nthread,')'
+  print *, 'Allocating stack_index(',nthread,')'"""
   else:
     txt = txt.replace("$1","")
 

src/module.py

@@ -28,6 +28,7 @@
 from irpf90_t import *
 from variable import *
 from variables import variables
+from command_line import command_line
 import preprocessed_text
 from util import *
 
@@ -169,11 +170,16 @@ class Fmodule(object):
       result    = map(lambda x: x[1], result)
       result    = map(lambda x: x.text, result)
       if self.is_main:
-        result = [ \
-        "program irp_program",
-        " call %s"%(self.prog_name),
-        "end program",
-        ] + result
+        temp  = [ "program irp_program" ]
+        if command_line.do_openmp:
+          temp += [ "!$OMP PARALLEL" ]
+          temp += [ "!$OMP MASTER" ]
+        temp += [ " call %s"%(self.prog_name) ]
+        if command_line.do_openmp:
+          temp += [ "!$OMP END MASTER" ]
+          temp += [ "!$OMP END PARALLEL" ]
+        temp += [ "end program" ]
+        result = temp + result
       self._residual_text = result
     return self._residual_text
   residual_text = property(residual_text)

src/variable.py

@@ -352,6 +352,13 @@ class Variable(object):
       result = [ "!","! >>> FREE %s"%(self.name),
         "  %s_is_built = .False."%(self.same_as) ] 
       if self.dim != []:
+        if command_line.do_memory:
+          result += [ \
+          "  if (allocated(%s)) then"%(name),
+          "    deallocate (%s)"%(name),
+          "    print *, 'Deallocating %s'"%(name),
+          "  endif" ]
+        else:
           result += [ \
           "  if (allocated(%s)) then"%(name),
           "    deallocate (%s)"%(name),
@@ -398,6 +405,9 @@ class Variable(object):
         def do_allocate():
           result = "    allocate(%s(%s),stat=irp_err)"
           result = result%(name,','.join(self.dim))
+          if command_line.do_memory:
+            tmp = "\n   print *, 'Allocating %s(%s)'"
+            result += tmp%(name,','.join(self.dim))
           return result
 
         result = [ " if (allocated (%s) ) then"%(name) ]
@@ -405,6 +415,9 @@ class Variable(object):
         result += [\
           "  if (.not.irp_dimensions_OK) then",
           "   deallocate(%s)"%(name) ]
+        if command_line.do_memory:
+          result += [\
+          "   print *, 'Deallocating %s'"%(name) ]
         result.append(check_dimensions())
         result.append(do_allocate())
         result += [\
@@ -438,10 +451,14 @@ class Variable(object):
         result.append("  call irp_enter(irp_here)")
       result += call_provides(self.to_provide)
       result += flatten( map(build_alloc,[self.same_as]+self.others) )
-      result += [\
-      "  call bld_%s"%(same_as),
-      "  %s_is_built = .True."%(same_as),
-      "" ]
+      if command_line.do_openmp:
+        result += [ "!$OMP CRITICAL (%s_critical)"%(same_as) ]
+        result += [ " if (.not.%s_is_built) then"%(same_as) ]
+      result += [ "  call bld_%s"%(same_as) ]
+      result += [ "  %s_is_built = .True."%(same_as), "" ]
+      if command_line.do_openmp:
+        result += [ " endif" ]
+        result += [ "!$OMP END CRITICAL (%s_critical)"%(same_as) ]
       if command_line.do_assert or command_line.do_debug:
         result.append("  call irp_leave(irp_here)")
       result.append("end subroutine provide_%s"%(name) )

src/variables.py

@@ -27,6 +27,7 @@
 
 from variable import *
 from irpf90_t import *
+from command_line import command_line
 from util import *
 
 ######################################################################
@@ -65,11 +66,27 @@ def call_provides(vars,opt=False):
     all_children = flatten( map(lambda x: variables[x].children, vars ))
     vars = filter(lambda x: x not in all_children,vars)
   def fun(x):
-    return [ \
+    if command_line.do_openmp:
+      result  = [ "!$OMP TASK" ]
+    else:
+      result = []
+    result += [ \
     "  if (.not.%s_is_built) then"%(x),
     "    call provide_%s"%(x),
     "  endif" ]
-  return flatten ( map (fun, vars) )
+    if command_line.do_openmp:
+      result += [ "!$OMP END TASK" ]
+    return result
+
+  result = flatten ( map (fun, vars) )
+  if command_line.do_openmp and result != []:
+   result.reverse()
+   result.remove("!$OMP TASK")
+   result.remove("!$OMP END TASK")
+   result.reverse()
+   if "!$OMP TASK" in result:
+     result += [ "!$OMP TASKWAIT" ]
+  return result
 
 ######################################################################
 if __name__ == '__main__':