Reuse output of Join in C backend #1340
Conversation
ricardoV94
force-pushed
the
faster_join_impl_c_backend
branch
2 times, most recently
from
d8efc2f to
9b81d41
Compare
ricardoV94
force-pushed
the
faster_join_impl_c_backend
branch
2 times, most recently
from
c6536ac to
deb4a19
Compare
ricardoV94
changed the title
ricardoV94
force-pushed
the
faster_join_impl_c_backend
branch
2 times, most recently
from
81e22db to
4d3f133
Compare
ricardoV94
force-pushed
the
faster_join_impl_c_backend
branch
from
4d3f133 to
948977d
Compare
Codecov ReportAttention: Patch coverage is
❌ Your patch status has failed because the patch coverage (97.29%) is below the target coverage (100.00%). You can increase the patch coverage or adjust the target coverage. Additional details and impacted files@@ Coverage Diff @@
## main #1340 +/- ##
=======================================
Coverage 82.13% 82.13%
=======================================
Files 211 211
Lines 49773 49734 -39
Branches 8830 8816 -14
=======================================
- Hits 40879 40850 -29
+ Misses 6714 6709 -5
+ Partials 2180 2175 -5
🚀 New features to boost your workflow:
|
There was a problem hiding this comment.
Copilot reviewed 8 out of 8 changed files in this pull request and generated no comments.
Comments suppressed due to low confidence (3)
tests/tensor/test_basic.py:1766
- The exception expectation was changed from IndexError to ValueError. Please verify that all related code and tests are updated accordingly.
with pytest.raises(ValueError):
pytensor/tensor/basic.py:2415
- The error message for out‐of‐bounds axis errors has been updated to match numpy's standard. Confirm that this change is intentional and that no downstream dependencies rely on the old message.
numpy.exceptions.AxisError: axis 2 is out of bounds for array of dimension 2
pytensor/link/numba/dispatch/tensor_basic.py:122
- Switching to axis.item() assumes the axis input is a 0-d array; please ensure that all cases (including negative axes and non-array types) are correctly handled by this update.
return np.concatenate(tensors, axis.item())
948977d to
2e0c79d
Compare
pytensor/tensor/basic.py
Outdated
|
|
||
| Joined tensors must have the same rank | ||
| >>> pt.join(0, x, u) | ||
| Traceback (most recent call last): | ||
| TypeError: Only tensors with the same number of dimensions can be joined. Input ndims were: [2, 1]. | ||
| TypeError: Only tensors with the same number of dimensions can be joined |
There was a problem hiding this comment.
I don't love that we're losing information in the traceback, is that 100% necessary?
pytensor/tensor/basic.py
Outdated
| raise TypeError( | ||
| "Join cannot handle arguments of dimension 0." | ||
| " Use `stack` to join scalar values." | ||
| " Use `stack` to join scalar values and/or increase rank of scalars." |
There was a problem hiding this comment.
Is "rank" the right word here? I only know it in the context of matrices as a synonym for column span.
Maybe "Use stack to join scalar values into higher dimensional objects" ?
|
|
||
| # Most times axis is constant, inline it | ||
| # This is safe to do because the hash of the c_code includes the constant signature |
There was a problem hiding this comment.
What do you mean by "constant signature"?
There was a problem hiding this comment.
The values of the constant are used to determine if a cached c_impl can be reused or not, otherwise it would be dangerous to inline the values like we're doing here for performance
| [old_output] = node.outputs | ||
|
|
||
| if ret.dtype != old_output.dtype: | ||
| ret = ret.astype(old_output.dtype) |
There was a problem hiding this comment.
Is there any savings to doing this check vs just always casting?
There was a problem hiding this comment.
There seems to be some extra work in cast:
def cast(x, dtype: str | np.dtype) -> TensorVariable:
"""Symbolically cast `x` to a Tensor of type `dtype`."""
if isinstance(dtype, str) and dtype == "floatX":
dtype = config.floatX
dtype_name = np.dtype(dtype).name
_x = as_tensor_variable(x)
if _x.type.dtype == dtype_name:
return _x
if _x.type.dtype.startswith("complex") and not dtype_name.startswith("complex"):
raise TypeError(
"Casting from complex to real is ambiguous: consider real(), "
"imag(), angle() or abs()"
)
return _cast_mapping[dtype_name](x)Not super crazy, but I don't mind too much the one check to save those extra cycles
ricardoV94
force-pushed
the
faster_join_impl_c_backend
branch
from
2e0c79d to
e828a98
Compare
Do not normalize constant axis in make_node and fix rewrite that assumed this would always be positive
ricardoV94
force-pushed
the
faster_join_impl_c_backend
branch
from
e828a98 to
8e4bb7e
Compare
When GC is disabled we can get a meaningful speedup by managing the copy to the output buffer ourselves instead of using
PyArray_CONCATENATEwhich doesn't allow passing an output.I didn't circumnvent
PyArray_CONCATENATEaltogether because numpy has clever logic to allocate the output array so that strides are as aligned with the inputs as possible. Copying this logic to the case where we can't reuse the buffer would add quite some complexity to our codebase.Also simplified the implementation by removing the exotic
view_flag, which closes #753Benchmark with new test
Before
After
📚 Documentation preview 📚: https://pytensor--1340.org.readthedocs.build/en/1340/