Changeset 823 for soft/giet_vm/applications/rosenfeld/scripts

Timestamp:

Jun 14, 2016, 5:23:56 PM (9 years ago)

Author:

meunier

Message:

• Improved scripts for simulations and graphes
• Continued to clean up the lib nrc2 (from nrio2x.x to nrmem1.c)
• Added a version (Fast - Parmerge - No stats)

Location:

soft/giet_vm/applications/rosenfeld/scripts

Files:

• create_graph.py (modified) (6 diffs)
• run_simus.py (modified) (12 diffs)

soft/giet_vm/applications/rosenfeld/scripts/create_graph.py

@@ -16 +16 @@
 import filecmp
 import random
+import math
 
 from stack import Stack
@@ -23 +24 @@
 
 use_rand_images = True
-with_features = True
+with_features = False
 
-threads = [1, 2, 4]
-nb_step = 0
+threads = [1, 2, 4, 8, 16, 32, 64]
 use_dsk = True
-img_size = 1024
+img_size = 2048
 granularity = 1
 rand_seed = 7
@@ -42 +42 @@
 # Each of these configuration must have been run with both features activated and deactivated
 configs = [
-        {'SLOW':'1', 'FAST':'0', 'PARMERGE':'0', 'ARSP':'0'},
-        {'SLOW':'0', 'FAST':'1', 'PARMERGE':'0', 'ARSP':'0'},
         #{'SLOW':'1', 'FAST':'0', 'PARMERGE':'0', 'ARSP':'0'},
         #{'SLOW':'0', 'FAST':'1', 'PARMERGE':'0', 'ARSP':'0'},
         #{'SLOW':'0', 'FAST':'1', 'PARMERGE':'1', 'ARSP':'0'},
-        #{'SLOW':'0', 'FAST':'1', 'PARMERGE':'1', 'ARSP':'1'},
+        {'SLOW':'0', 'FAST':'1', 'PARMERGE':'1', 'ARSP':'1'},
 ]
 
@@ -55 +53 @@
     code = compile(f.read(), pyconf_file, 'exec')
     exec(code)
+
+nb_step = 0
 
 if use_dsk:
@@ -125 +125 @@
         plotter.add_x(X)
         assert(nb_step == 4)
-        YPAR1 = [float(exec_time[fconfig][thread]["NF"][density][0]) for density in range(0, 101)] # Parallel Labeling
-        YPAR2 = [float(exec_time[fconfig][thread]["NF"][density][1]) for density in range(0, 101)] # Merging (Parallel or Pyramidal)
-        YPTC2 = [float(exec_time[fconfig][thread]["NF"][density][2]) for density in range(0, 101)] # Parallel Transitive Closure
-        YPAR3 = [float(exec_time[fconfig][thread]["NF"][density][3]) for density in range(0, 101)] # Parallel Relabeling
+        YPAR1 = [float(int(exec_time[fconfig][thread]["NF"][density][0]) / math.pow(img_size, 2)) for density in range(0, 101)] # Parallel Labeling
+        YPAR2 = [float(int(exec_time[fconfig][thread]["NF"][density][1]) / math.pow(img_size, 2)) for density in range(0, 101)] # Merging (Parallel or Pyramidal)
+        YPTC2 = [float(int(exec_time[fconfig][thread]["NF"][density][2]) / math.pow(img_size, 2)) for density in range(0, 101)] # Parallel Transitive Closure
+        YPAR3 = [float(int(exec_time[fconfig][thread]["NF"][density][3]) / math.pow(img_size, 2)) for density in range(0, 101)] # Parallel Relabeling
 
         plotter.add_y(YPAR1)
@@ -136 +136 @@
 
         if with_features:
-            YPAR1f = [float(exec_time[fconfig][thread]["FT"][density][0]) for density in range(0, 101)] # Parallel Labeling
-            YPAR2f = [float(exec_time[fconfig][thread]["FT"][density][1]) for density in range(0, 101)] # Merging (Parallel or Pyramidal)
+            YPAR1f = [float(int(exec_time[fconfig][thread]["FT"][density][0]) / math.pow(img_size, 2)) for density in range(0, 101)] # Parallel Labeling
+            YPAR2f = [float(int(exec_time[fconfig][thread]["FT"][density][1]) / math.pow(img_size, 2)) for density in range(0, 101)] # Merging (Parallel or Pyramidal)
             deltaYPAR1 = [max(0, x - y) for x, y in zip (YPAR1f, YPAR1)]
             deltaYPAR2 = [max(0, x - y) for x, y in zip (YPAR2f, YPAR2)]

soft/giet_vm/applications/rosenfeld/scripts/run_simus.py

@@ -5 +5 @@
 # python 2 as it will execute on computation servers
 
+# TODO:
+#       Can we do something about assert in perf eval?
 
 from __future__ import print_function
@@ -37 +39 @@
 
 # Parameters
-num_runs = 10
+# - with eval_perf, num_internal_runs should be used, as this allows to mitigate the cost of the extra
+#     run required to have the correct "ne" value (number of labels), and only the times from last application run are taken
+# - With check_results, num_app_runs should be used, so as to have a number of checks equals to the number of runs,
+#     because only one check per application run is performed
+num_app_runs = 1        # Number of times the application is launched per configuration
+num_internal_runs = 20  # Number of times the image is processed inside the application
 check_results = False
 eval_perf = True
 use_valgrind = False
 use_rand_images = True
-threads = [1, 2, 4]
+threads = [1, 2, 4, 8, 16, 32, 64]
+#threads = [1, 2]
 use_dsk = True
 # Using dsk will store generated random images, otherwise they are re-generated at each run to save disk space
@@ -48 +56 @@
 # Configurations
 configs = [
-        {'SLOW':'1', 'FAST':'0', 'FEATURES':'0', 'PARMERGE':'0', 'ARSP':'0'},
-        {'SLOW':'0', 'FAST':'1', 'FEATURES':'0', 'PARMERGE':'0', 'ARSP':'0'},
-        {'SLOW':'1', 'FAST':'0', 'FEATURES':'1', 'PARMERGE':'0', 'ARSP':'0'},
-        {'SLOW':'0', 'FAST':'1', 'FEATURES':'1', 'PARMERGE':'0', 'ARSP':'0'},
+        #{'SLOW':'1', 'FAST':'0', 'FEATURES':'0', 'PARMERGE':'0', 'ARSP':'0'},
+        #{'SLOW':'0', 'FAST':'1', 'FEATURES':'0', 'PARMERGE':'0', 'ARSP':'0'},
+        #{'SLOW':'1', 'FAST':'0', 'FEATURES':'1', 'PARMERGE':'0', 'ARSP':'0'},
+        #{'SLOW':'0', 'FAST':'1', 'FEATURES':'1', 'PARMERGE':'0', 'ARSP':'0'},
+        {'SLOW':'0', 'FAST':'1', 'FEATURES':'0', 'PARMERGE':'1', 'ARSP':'0'},
         #{'SLOW':'0', 'FAST':'1', 'FEATURES':'1', 'PARMERGE':'1', 'ARSP':'0'},
         #{'SLOW':'0', 'FAST':'1', 'FEATURES':'0', 'PARMERGE':'1', 'ARSP':'1'},
@@ -60 +69 @@
 rand_seed = 7
 granularity = 1 # constant for now
-img_size = 1024
+img_size = 2048
 
 check_pyconf_file(pyconf_file)
@@ -93 +102 @@
 
 
+if check_results and eval_perf:
+    print("*** Warning: check_results and eval_perf modes are both set\n")
+if eval_perf and use_valgrind:
+    print("*** Warning: using valgrind while eval_perf mode is set\n")
+if eval_perf and num_app_runs != 1:
+    print("*** Warning: using eval_perf with num_app_runs != 1\n")
+if check_results and num_internal_runs != 1:
+    print("*** Warning: using check_results with num_internal_runs != 1\n")
+
+
 
 def gen_random_image(filename, x, y, granularity, density, seed):
@@ -136 +155 @@
                     break
             if not line_with_key:
-                f.write(line)
+                if "#define MCA_VERBOSE_LEVEL" in line:
+                    if eval_perf:
+                        verb_level = 1
+                    else:
+                        verb_level = 2
+                    f.write("#define MCA_VERBOSE_LEVEL %d\n" % verb_level)  
+                else:
+                    f.write(line)
 
         f.close()
@@ -179 +205 @@
     my_chdir(top_path)
     cmd = ['make']
+    #if eval_perf:
+    #    cmd.extend(['IGNORE_ASSERT=true'])
     print_and_call(cmd)
     my_chdir(scripts_path)
@@ -205 +233 @@
         for nthreads in threads:
             perf_array[fconfig][img_basename][nthreads] = {}
-            for run in range(num_runs):
+            for run in range(num_app_runs):
                 if not os.path.exists(ref_bmpfile):
                     bmpfile = ref_bmpfile
@@ -223 +251 @@
     
                 cmd.extend([short_path(binary_file), '-n', str(nthreads), '-i', short_path(image)])
+
+                if num_internal_runs > 1:
+                    cmd.extend(['-r', str(num_internal_runs)])
                 
                 if check_results:
@@ -236 +267 @@
 
                 # if performance evaluation, get timing measurements
+                # Only the last application run is considered
                 if eval_perf:
                     for line in outlines:
@@ -247 +279 @@
                             step = match.group(1)
                             value = tokens[len(tokens) - 1]
-                            if step in perf_array[fconfig][img_basename][nthreads]:
-                                # Accumulating times over the num_runs runs
-                                perf_array[fconfig][img_basename][nthreads][step] += int(value)
-                            else:
-                                perf_array[fconfig][img_basename][nthreads][step] = int(value)
+                            perf_array[fconfig][img_basename][nthreads][step] = int(value)
 
 
@@ -315 +343 @@
                 for step in sorted(perf_array[fconfig][img_basename][nthreads].keys()):
                     # Average time for each step
-                    file.write("[STEP_%s]   %d\n" % (step, perf_array[fconfig][img_basename][nthreads][step] / num_runs))
+                    file.write("[STEP_%s]   %d\n" % (step, perf_array[fconfig][img_basename][nthreads][step]))
 
         img_idx += 1