Improved profiled

Version:1.2.4

src/command_line.py

@@ -46,6 +46,7 @@ options['z'] = [ 'openmp'       , 'Automatic openMP tasks (may not work)', 0 ]
 options['l'] = [ 'align'        , 'Align arrays using compiler directives', 1 ]
 options['s'] = [ 'substitute'   , 'Substitute values for loop max values', 1 ]
 options['r'] = [ 'no_directives', 'Ignore compiler directives !DEC$ and !DIR$', 0 ]
+options['n'] = [ 'inline'       , 'all|providers|builders : Force inlining of providers or builders', 1 ]
 
 class CommandLine(object):
 
@@ -64,6 +65,16 @@ class CommandLine(object):
     return self._defined
   defined = property(fget=defined)
 
+  def inline(self):
+    if '_inline' not in self.__dict__:
+      self._inline = ""
+      for o,a in self.opts:
+        if o in [ "-n", '--'+options['n'][0] ]:
+          self._inline = a
+          break
+    return self._inline
+  inline = property(fget=inline)
+
   def substituted(self):
     if '_substituted' not in self.__dict__:
       self._substituted = {}

src/makefile.py

@@ -89,7 +89,7 @@ def run():
 
     print >>file, "OBJ1 = $(patsubst %%, %s%%,$(notdir $(OBJ))) %sirp_touches.irp.o"%(irpdir,irpdir),
     if command_line.do_profile:
-      print >>file, " %sirp_profile.irp.o"%(irpdir), " %sirp_rdtsc.o"%(irpdir),
+      print >>file, " %sirp_profile.irp.o"%(irpdir), " irp_rdtsc.o",
     if command_line.do_openmp:
       print >>file, " %sirp_locks.irp.o"%(irpdir),
     else:

src/preprocessed_text.py

@@ -34,6 +34,8 @@ from util import *
 re_endif = re.compile("end\s+if")
 re_elseif = re.compile("else\s+if")
 re_enddo  = re.compile("end\s+do")
+re_endtype= re.compile("end\s+type")
+re_endmodule = re.compile("end\s+module")
 re_endselect  = re.compile("end\s+select")
 
 # Local variables
@@ -100,6 +102,8 @@ def get_type (i, filename, line, is_doc):
   # Replacements
   lower_line = re_elseif.sub("elseif",lower_line)
   lower_line = re_enddo.sub("enddo",lower_line)
+  lower_line = re_endtype.sub("endtype",lower_line)
+  lower_line = re_endmodule.sub("endmodule",lower_line)
   lower_line = re_endif.sub("endif",lower_line)
   lower_line = re_endselect.sub("endselect",lower_line)
   for c in """()'"[]""":

src/profile.py

@@ -29,8 +29,7 @@ def build_rdtsc():
   file.write(rdtsc)
   file.close()
   def t():
-    p = subprocess.Popen(["gcc","-O2",filename,"-c","-o","IRPF90_temp/irp_rdtsc.o"])
+    p = subprocess.Popen(["gcc","-O2",filename,"-c","-o","irp_rdtsc.o"])
-
     p.communicate()
     os.remove(filename)
 
@@ -39,14 +38,52 @@ def build_rdtsc():
 def build_module():
   data = """
 module irp_timer
- double precision :: irp_profile(2,%(n)d) 
+ double precision :: irp_profile(3,%(n)d) 
+ integer          :: irp_order(%(n)d) 
  character*(64)   :: irp_profile_label(%(n)d)
+ double precision :: irp_rdtsc_shift
+
+ contains
+
+   subroutine profile_sort ()
+    implicit none
+    character*(64)         :: xtmp
+    integer                :: i, i0, j, jmax
+
+    do i=1,size(irp_profile_label)
+     irp_order(i)=i
+    enddo
+    do i=1,size(irp_profile_label)
+     xtmp = irp_profile_label(i)
+     i0 = irp_order(i)
+     j = i-1
+     do j=i-1,1,-1
+      if ( irp_profile_label(j) > xtmp ) then
+       irp_profile_label(j+1) = irp_profile_label(j)
+       irp_order(j+1) = irp_order(j)
+      else
+       exit
+      endif
+     enddo
+     irp_profile_label(j+1) = xtmp
+     irp_order(j+1) = i0
+    enddo
+   end subroutine profile_sort
+
 end module
 
 subroutine irp_init_timer
  use irp_timer
  implicit none
- irp_profile = 0.
+ integer :: i
+ double precision :: irp_rdtsc, t0
+ irp_profile = 0.d0
+ irp_rdtsc_shift = 0.d0
+ do i=1,1000
+   t0 = irp_rdtsc()
+   irp_rdtsc_shift = irp_rdtsc_shift + (irp_rdtsc()-t0)
+ enddo
+ irp_rdtsc_shift = 1.d-3*irp_rdtsc_shift
 %(text)s
 end
 
@@ -54,33 +91,55 @@ subroutine irp_set_timer(i,value)
  use irp_timer
  implicit none
  integer, intent(in) :: i
- double precision, intent(in) :: value
+ double precision, intent(inout) :: value
+ value = value - irp_rdtsc_shift
  irp_profile(1,i) = irp_profile(1,i) + value
- irp_profile(2,i) = irp_profile(2,i) + 1.d0
+ irp_profile(2,i) = irp_profile(2,i) + value*value
+ irp_profile(3,i) = irp_profile(3,i) + 1.d0
 end
 
 subroutine irp_print_timer()
  use irp_timer
  implicit none
- integer :: i
+ integer :: i, ii
- print '(A24,A8,4(X,A14))', 'Calls', 'Tot Cycles', 'Avge Cycles', &
+ double precision :: error, sigma2, average, average2, frequency, t0
-                          'Tot Secs(1GHz)', 'Avge Secs(1GHz)'
+ double precision :: irp_rdtsc
+ t0 = irp_rdtsc()
+ call sleep(1)
+ frequency = (irp_rdtsc()-t0-irp_rdtsc_shift)
+
+ call profile_sort()
+ print '(A24,A8,A17,A20,A13,A20)', '', 'N.Calls', 'Tot Cycles', 'Avg Cycles', &
+                          'Tot Secs', 'Avg Secs'
    print '(A)', '----------------------------------------------'// &
                 '----------------------------------------------'
- do i=1,%(n)d
+ do ii=1,%(n)d
-  if (irp_profile(2,i) > 0.) then
+  i = irp_order(ii)
-   print '(A24,F8.0,2(X,F14.0),2(X,F14.8))', &
+  if (irp_profile(3,i) > 0.) then
-     irp_profile_label(i), irp_profile(2,i), &
+   error = 0.d0
-     irp_profile(1,i), irp_profile(1,i)/irp_profile(2,i), &
+   average  = irp_profile(1,i)/irp_profile(3,i)
-     irp_profile(1,i)*1.d-9, 1.d-9*irp_profile(1,i)/irp_profile(2,i)
+   if (irp_profile(3,i) > 1.d0) then
+     average2 = irp_profile(2,i)/irp_profile(3,i)
+     sigma2   = (average2 - average*average)
+     error = sqrt(sigma2/(irp_profile(3,i)+1.d0))
+   endif
+   print '(A24 , F8.0 , X,F12.0 , X,F12.0,A3,F8.0, X,F12.8, X,F8.5,A3,F8.5 )', &
+     irp_profile_label(ii), &
+     irp_profile(3,i), &
+     irp_profile(1,i), &
+     average, '+/-', error, &
+     irp_profile(1,i)/frequency, &
+     average/frequency, '+/-', error/frequency
   endif
  enddo
+ print *, 'Frequency     :', frequency*1.d-9, ' GHz'
+ print *, 'rdtsc latency :', irp_rdtsc_shift, ' cycles'
 end
   """
   label = {}
   for i in variables:
     vi = variables[i]
-    label[vi.label] = vi.name
+    label[vi.label] = vi.same_as
   text = []
   lmax = 0
   for l in label:

src/regexps.py

@@ -46,6 +46,7 @@ re_decl    = re.compile( "".join( [ r"^\ *",
  r"|intrinsic *(::)?",
  r"|external *(::)?",
  r"|equivalence *(::)?",
+ r"|type",
  r")[^=(]"
 ] ) )
 

src/variable.py

@@ -507,7 +507,10 @@ class Variable(object):
           " endif" ]
         return result
 
-      result = [ "subroutine provide_%s"%(name) ] 
+      result = []
+      if command_line.directives and command_line.inline in ["all","providers"]:
+        result += [ "!DEC$ ATTRIBUTES FORCEINLINE :: provide_%s"%(name) ]
+      result += [ "subroutine provide_%s"%(name) ] 
       result += build_use( [same_as]+self.to_provide )
       result.append("  implicit none")
       length = len("provide_%s"%(name))
@@ -570,7 +573,10 @@ class Variable(object):
         text = map(lambda x: x[1], text)
         for line in filter(lambda x: type(x) not in [ Begin_doc, End_doc, Doc], text):
           if type(line) == Begin_provider:
-            result = [ "subroutine bld_%s"%(name) ]
+            result = []
+            if command_line.directives and command_line.inline in ["all","builders"]:
+              result += [ "!DEC$ ATTRIBUTES FORCEINLINE :: bld_%s"%(same_as) ]
+            result += [ "subroutine bld_%s"%(name) ]
             result += build_use([name]+self.needs)
           elif type(line) == Cont_provider:
             pass

src/version.py

@@ -1 +1 @@
-version = "1.2.3"
+version = "1.2.4"