Author: Maciej Fijalkowski <fij...@gmail.com> Branch: extradoc Changeset: r3718:a84b87f60299 Date: 2011-06-16 21:52 +0200 http://bitbucket.org/pypy/extradoc/changeset/a84b87f60299/
Log: a para about warming up (and a bench file in-progress) diff --git a/talk/iwtc11/benchmarks/new_result.txt b/talk/iwtc11/benchmarks/new_result.txt --- a/talk/iwtc11/benchmarks/new_result.txt +++ b/talk/iwtc11/benchmarks/new_result.txt @@ -35,3 +35,34 @@ NoBorderImagePadded(range): 1.13030288219 +- 0.00631341257207 NoBorderImage: 2.17767724991 +- 0.0252138424223 NoBorderImage(iter): 1.45121192932 +- 0.0274908528137 +NoBorderImage(range): 1.33527753353 +- 0.00819077114708 +sobel(NoBorderImagePadded): 1.01328015327 +- 0.0191050555554 +sobel_uint8(NoBorderImagePadded): 1.10731520653 +- 0.0325015176484 + +gcc -O2 +sqrt(float): 0.984 +- 0.00547722557505 +sqrt(int): 0.796 +- 0.00894427191 +sqrt(Fix16): 0.12 +- 0.0 +conv3(1e6): 0.754 +- 0.00547722557505 +conv5(1e6): 1.01 +- 0.01 +conv3(1e5): 0.666 +- 0.00547722557505 +conv5(1e5): 0.98 +- 0.0122474487139 +conv3x3(3): 0.268 +- 0.004472135955 +conv3x3(1000): 0.24 +- 0.0 +dilate3x3(1000): 0.26 +- 0.0 +sobel_magnitude: 0.19 +- 0.0122474487139 + +gcc -O3 -march=native -fno-tree-vectorize +sqrt(float): 0.982 +- 0.0109544511501 +sqrt(int): 0.788 +- 0.004472135955 +sqrt(Fix16): 0.12 +- 0.0 +conv3(1e6): 0.73 +- 0.00707106781187 +conv5(1e6): 0.758 +- 0.0192353840617 +conv3(1e5): 0.572 +- 0.004472135955 +conv5(1e5): 0.636 +- 0.00547722557505 +conv3x3(3): 0.276 +- 0.00894427191 +conv3x3(1000): 0.24 +- 0.0 +dilate3x3(1000): 0.252 +- 0.004472135955 +sobel_magnitude: 0.182 +- 0.004472135955 + +python2.7 diff --git a/talk/iwtc11/paper.tex b/talk/iwtc11/paper.tex --- a/talk/iwtc11/paper.tex +++ b/talk/iwtc11/paper.tex @@ -907,6 +907,11 @@ We also run PyPy with loop peeling optimization and without (but otherwise identical). +For all JITted run (PyPy and Psyco) 3 iterations were run to warm up the JIT, +followed by 10 iterations for averaging. For GCC and CPython, 5 iterations +were run. In all cases, the standard deviation is very low, making benchmarks +very well reproducible. + XXX discuss some more, explain why gcc is faster XXX we need Psyco numbers _______________________________________________ pypy-commit mailing list pypy-commit@python.org http://mail.python.org/mailman/listinfo/pypy-commit
