Profiling PyTensor function — PyTensor dev documentation (original) (raw)
Note
This method replaces the old ProfileMode
. Do not use ProfileMode
anymore.
Besides checking for errors, another important task is to profile your code in terms of speed and/or memory usage. You can profile your functions using either of the following two options:
- Use the PyTensor flag config.profile to enable profiling.
- To enable the memory profiler use the PyTensor flag:config.profile_memory in addition to config.profile.
- Moreover, to enable the profiling of PyTensor rewrite phases, use the PyTensor flag: config.profile_optimizer in addition to config.profile.
- You can also use the PyTensor flags
profiling__n_apply
,profiling__n_ops
andprofiling__min_memory_size
to modify the quantity of information printed.
- Pass the argument
profile=True
to the functionpytensor.function
and then callf.profile.summary()
for a single function.- Use this option when you want to profile not all the functions but only one or more specific function(s).
- You can also combine the profile results of many functions:
The profiler will output one profile per PyTensor function and produce a profile that is the sum of the printed profiles. Each profile contains four sections: global info, class info, Op
s info and Apply
node info.
In the global section, the “Message” is the name of the PyTensor function. pytensor.function() has an optional parameter name
that defaults to None
. Change it to something else to help profile many PyTensor functions. In that section, we also see the number of times the function was called (1) and the total time spent in all those calls. The time spent in Function.vm.__call__()
and in thunks is useful to understand PyTensor’s overhead.
Also, we see the time spent in the two parts of the compilation process: rewriting (i.e. modifying the graph to make it more stable/faster) and the linking (i.e. compile C code and make the Python callable returned bypytensor.function()).
The class, Op
s and Apply
nodes sections have the same information: i.e. information about the Apply
nodes that ran. The Op
s section takes the information from the Apply
section and merges it with the Apply
nodes that have exactly the same Op
. If two Apply
nodes in the graph have two Op
s that compare equal, they will be merged. Some Op
s, like Elemwise
, will not compare equal if their parameters differ, so the class section will merge moreApply
nodes than the Op
s section.
Note that the profile also shows which Op
s were run with C implementation.
Developers wishing to optimize the performance of their graph should focus on the worst offending Op
s and Apply
nodes–either by optimizing an implementation, providing a missing C implementation, or by writing a graph rewrite that eliminates the offending Op
altogether.
Here is some example output when PyTensor’s rewrites are disabled. With all rewrites enabled, there would be only one Op
remaining in the graph.
To run the example:
PYTENSOR_FLAGS=optimizer_excluding=fusion:inplace,profile=True python doc/tutorial/profiling_example.py
The output:
Function profiling
Message: None Time in 1 calls to Function.call: 5.698204e-05s Time in Function.vm.call: 1.192093e-05s (20.921%) Time in thunks: 6.198883e-06s (10.879%) Total compile time: 3.642474e+00s PyTensor rewrite time: 7.326508e-02s PyTensor validate time: 3.712177e-04s PyTensor Linker time (includes C, CUDA code generation/compiling): 9.584920e-01s
Class
<% time> <sum %>
Ops
<% time> <sum %>
Apply
<% time> <sum %>