merge

examples/advection_gpu.py

@@ -27,7 +27,7 @@ if __name__=='__main__':
     GHOSTS    = 0 
     NSCALARS  = 0
 
-    f_resolution = (129,129,129)[:DIM]
+    f_resolution = (257,257,257)[:DIM]
     v_resolution = f_resolution
     #v_resolution = (33,33,33)[:DIM]
     ghosts       = (GHOSTS,)*DIM
@@ -128,7 +128,7 @@ if __name__=='__main__':
         else:
             print 'OK'
 
-    autotuner_config = AutotunerConfig(autotuner_flag=AutotunerFlags.MEASURE, prune_threshold=1.20, override_cache=False)
+    autotuner_config = AutotunerConfig(autotuner_flag=AutotunerFlags.PATIENT, prune_threshold=1.20, override_cache=False)
 
     method = {
             Backend: Backend.OPENCL,
@@ -139,7 +139,11 @@ if __name__=='__main__':
                 'use_builtin_copy':True,
                 'stretching': {
                     'formulation': StretchingFormulation.GRAD_UW,
+<<<<<<< HEAD
                     'order':64
+=======
+                    'order':16
+>>>>>>> f03a24d1fe54c00f58e70368a14ef9043deb4581
                 }
             }
         }

hysop/codegen/base/statistics.py

@@ -143,7 +143,7 @@ class TimedWorkStatistics(WorkStatistics):
         return self.global_mem_transactions()/self.duration
     def local_mem_throughput(self):
         return self.local_mem_transactions()/self.duration
-    def achieved_mem_throughput(self):
+    def total_mem_throughput(self):
         return self.total_mem_transactions()/self.duration
 
     def _init(self):

hysop/tools/callback.py

@@ -152,6 +152,8 @@ class ComputeInterface(MemInterface):
         if (timed_stats is not None):
             if timed_stats.global_mem_throughput()>0:
                 s += '  throughput={}'.format(bdw2str(timed_stats.global_mem_throughput()))
+                if timed_stats.global_mem_throughput() < timed_stats.total_mem_throughput():
+                    s+= ' (tot={})'.format(bdw2str(timed_stats.total_mem_throughput()))
             for (op_category, ops_per_second) in timed_stats.ops_per_second().iteritems():
                 s += '  {}'.format(unit2str(ops_per_second,op_category,decimal=True,rounded=0))