Benchmarking

benchmark/bench_tfphasespace.py

@@ -9,13 +9,15 @@
 
 import tensorflow as tf
 
+import numpy as np
+
 from tfphasespace import tfphasespace
 
 import os, sys
 
 sys.path.append(os.path.dirname(__file__))
 
-from .monitoring import Timer
+from monitoring import Timer
 
 # to play around with optimization, no big effect though
 NUM_PARALLEL_EXEC_UNITS = 1
@@ -26,17 +28,16 @@
 B_AT_REST = tf.stack((0.0, 0.0, 0.0, B_MASS), axis=-1)
 PION_MASS = 139.6
 
-N_EVENTS = 1000000
-CHUNK_SIZE = N_EVENTS
+N_EVENTS = 100000000
+CHUNK_SIZE = 1000000
 
 N_EVENTS_VAR = tf.Variable(initial_value=N_EVENTS)
 CHUNK_SIZE_VAR = tf.Variable(initial_value=CHUNK_SIZE)
 
 samples = [tfphasespace.generate(B_AT_REST,
                                  [PION_MASS, PION_MASS, PION_MASS],
                                  CHUNK_SIZE_VAR)
-           for _ in range(0, N_EVENTS, CHUNK_SIZE)
-           ]
+           for _ in range(0, N_EVENTS, CHUNK_SIZE)]
 
 sess = tf.Session(
         config=config
@@ -51,10 +52,15 @@ def test_three_body():
         print("Initial run (may takes more time than consequent runs)")
         do_run()  # to get rid of initial overhead
     print("starting benchmark")
-    with Timer(verbose=True):
+    with Timer(verbose=True) as timer:
         CHUNK_SIZE_VAR.load(CHUNK_SIZE + 1, session=sess)  # +1 to make sure we're not using any trivial caching
         samples = do_run()
+        tot_samples = sum(sample[0].shape[0] for sample in samples)
+        if np.any(samples[0][0] == samples[1][0]):
+            raise ValueError("You're generating the same sample!")
+        print("Total number of generated samples", tot_samples)
         print("Shape of one particle momentum", samples[0][1][0].shape)
+    print("Time per sample: {} ms".format(timer.elapsed/tot_samples))
 
 
 def do_run():

benchmark/bench_tgenphasespace.cxx

@@ -5,7 +5,7 @@
 #include "TLorentzVector.h"
 #include "TGenPhaseSpace.h"
 
-Int_t N_EVENTS = 1000000;
+Int_t N_EVENTS = 100000000;
 
 int bench_tgenphasespace()
 {