Improve performance of CAReduce in Numba backend #1109
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This Op does not really fit the CAReduce API, as it requires an extra bit of information (number of elements in the axis) during the loop. A better solution will be a fused Elemwise+CAReduce
ricardoV94
force-pushed
the
better_numba_careduce
branch
from
bfa16dd to
2bc894a
Compare
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #1109 +/- ##
==========================================
- Coverage 82.12% 82.10% -0.03%
==========================================
Files 183 183
Lines 48111 48030 -81
Branches 8667 8658 -9
==========================================
- Hits 39510 39433 -77
+ Misses 6435 6434 -1
+ Partials 2166 2163 -3
|
ricardoV94
force-pushed
the
better_numba_careduce
branch
from
2bc894a to
79e8109
Compare
|
Here is a direct comparison of C and numba backends for the non C-contiguous case: import numpy as np
import pytensor
c_contiguous = False
for transpose_in_graph in (True, False):
rng = np.random.default_rng(123)
N = 256
x_test = rng.uniform(size=(N, N, N))
transpose_axis = (0, 1, 2) if c_contiguous else (2, 0, 1)
if not transpose_in_graph:
x_test = x_test.transpose(transpose_axis)
x = pytensor.shared(x_test, name="x", shape=x_test.shape, borrow=True)
if transpose_in_graph:
x = x.transpose(transpose_axis)
out = x.sum(axis=0)
c_fn = pytensor.function([], out, mode="FAST_COMPILE")
numba_fn = pytensor.function([], out, mode="NUMBA").vm.jit_fn
np.testing.assert_allclose(c_fn(), numba_fn()[0])
print(f"{transpose_in_graph=}")
%timeit c_fn()
%timeit numba_fn()
# transpose_in_graph=True
# 33.7 ms ± 2.25 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
# 188 ms ± 4.05 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
# transpose_in_graph=False
# 33 ms ± 1.15 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
# 103 ms ± 1.96 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)Airect numba implementation shows the same bad performance. import numpy as np
import numba
c_contiguous = False
rng = np.random.default_rng(123)
N = 256
x_test = rng.uniform(size=(N, N, N))
transpose_axis = (0, 1, 2) if c_contiguous else (2, 0, 1)
x_test = x_test.transpose(transpose_axis)
out_dtype = np.float64
@numba.njit(fastmath=True, boundscheck=False)
def careduce_add(x):
x_shape = x.shape
res_shape = (x_shape[1], x_shape[2])
res = np.full((x_shape[1], x_shape[2]), np.asarray(0.0).item(), dtype=out_dtype)
for i0 in range(x_shape[0]):
for i1 in range(x_shape[1]):
for i2 in range(x_shape[2]):
res[i1, i2] += x[i0, i1, i2]
return res
np.testing.assert_allclose(careduce_add(x_test), np.sum(x_test, 0))
%timeit careduce_add(x_test)
# 136 ms ± 1.57 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) |
ricardoV94
force-pushed
the
better_numba_careduce
branch
from
79e8109 to
6268d99
Compare
AlexAndorra
approved these changes
Nov 29, 2024
There was a problem hiding this comment.
Thanks for the walk through in the comparison @ricardoV94 , definitely interesting
|
Numba doing badly on the non-contiguous case is all due to loop ordering. LLVM doesn't reorder based on strides :( Anyway this PR improves overall, better old speeds where just due to chance when the reduced loop was the one with smallest strides |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Closes #935
Closes #931
The implementation for multiple axes no longer operates one axis at a time. Here are the benchmarks for the Sum test before and after this PR:
Note that we have a special dispatch for
Sum(axes=None)introduced in #92, so the changes are not reflected in that benchmark. I temporarily disabled the special dispatch, to confirm that case is still improved:Because it is still a bit slower, and this is the most common reduction, I decided to leave the special case.
Numba doesn't seem to optimize non-contiguous arrays very well. The C backend implementation with explicit loop reordering written in #971 does not show such a penalty.
Finally we also see an improvement in the slowest case of the pre-existing numba-logsumexp benchmark: