ENH: Check for floating point exceptions in dot #28442

rtimpe · 2025-03-06T20:59:31Z

Currently dot does not check for floating point exceptions, while similar functions (such as multiply and matmul) do. Add these checks for consistency.

Note that certain types of floating point exceptions are still not caught:

Multiplications of 0 with nan or inf are ignored (see BUG: np.dot result changes based on shape when the input contains nan #27902)
Integer overflows are ignored. These aren't true floating point exceptions, but the behavior is inconsistent with scalar multiply, i.e.

>>> np.int16(32000) * np.int16(3)
<python-input-7>:1: RuntimeWarning: overflow encountered in scalar multiply
  np.int16(32000) * np.int16(3)
np.int16(30464)
>>> np.dot(np.int16(32000), np.int16(3))
np.int16(30464)

Fixes #14925

Currently dot does not check for floating point exceptions, while similar functions (such as multiply and matmul) do. Add these checks for consistency. Note that certain types of floating point exceptions are still not caught: - Multiplications of 0 with nan or inf are ignored (see numpy#27902) - Integer overflows are ignored. These aren't true floating point exceptions, but the behavior is inconsistent with scalar multiply, i.e. ``` >>> np.int16(32000) * np.int16(3) <python-input-7>:1: RuntimeWarning: overflow encountered in scalar multiply np.int16(32000) * np.int16(3) np.int16(30464) >>> np.dot(np.int16(32000), np.int16(3)) np.int16(30464) ``` Fixes numpy#14925

seberg

Thanks for looking into it, seems like a good small improvement. Could note it as a very brief "change" in the relase note, although I don't think it mattre much.

seberg · 2025-03-09T11:18:59Z

numpy/_core/src/multiarray/multiarraymodule.c

+    }
+
+    return 0;
+}


Please use PyUFunc_GiveFloatingpointErrors, you can find the rest of the pattern elsewhere (e.g. in scalar math).

seberg · 2025-03-09T11:20:41Z

numpy/_core/src/multiarray/multiarraymodule.c

@@ -1008,11 +1022,17 @@ PyArray_MatrixProduct2(PyObject *op1, PyObject *op2, PyArrayObject* out)
        return NULL;
    }

+    npy_clear_floatstatus();


I would prefer pairing this with the GIL release (i.e. very targeted around the place of interest).

That probably means moving it into cblas_matrixproduct.

I moved this to cblas_matrixproduct but I'm not sure if it's quite what you want. Do you want this immediately before (or after?) the GIL release, or just closer to where the actual work is being done?

seberg · 2025-03-09T11:21:34Z

numpy/_core/src/multiarray/multiarraymodule.c

-        return cblas_matrixproduct(typenum, ap1, ap2, out);
+        PyObject *res = cblas_matrixproduct(typenum, ap1, ap2, out);
+        if (check_fperr("dot") < 0) {
+            return NULL;


Just to note, this is also incorrect of course: You would have to throw away res here, but you are leaking it.

rtimpe · 2025-03-11T05:11:16Z

I'm not sure what's causing the CI failure, but it doesn't seem to be related to my PR. Or am I missing something?

mattip · 2025-03-11T06:03:53Z

I restarted the s390x build.

seberg

Thanks, LGTM! A small nit that we should extend the test a bit, I think.

Not super important here, but you could add a change or improvement release note if you like (see docs/release/upcoming_changes/README.rst).

(Unrelated... I am half wondering if we couldn't fold most of this to use the matmul ufunc as a whole or just its loops. That would also have the information that no floating point errors can happen. But I am not worried about it much, matmuls on integers isn't all that common.)

seberg · 2025-03-14T09:34:38Z

numpy/_core/tests/test_multiarray.py

@@ -3345,6 +3345,29 @@ def test_dot(self):
        a.dot(b=b, out=c)
        assert_equal(c, np.dot(a, b))

+    @pytest.mark.skipif(IS_WASM, reason="no wasm fp exception support")
+    def test_dot_errstate(self):


Small nit: Could you parametrize this to include half and/or longdouble (just because they currently take the other code paths and can set the warnings).

The point is that double/float should effectively always take the first branch, that way we should cover both error paths.

rtimpe · 2025-03-15T19:53:50Z

(Unrelated... I am half wondering if we couldn't fold most of this to use the matmul ufunc as a whole or just its loops. That would also have the information that no floating point errors can happen. But I am not worried about it much, matmuls on integers isn't all that common.)

I can look into this, if it would be of interest. I picked this bug as a good way to get started with numpy development, and I was thinking of applying similar fixes to some of the other functions in the multiarray module (assuming they have similar bugs, which I haven't looked into yet). Do you have a preference on what would be more useful?

seberg · 2025-03-16T11:19:38Z

Thanks for the follow-up, let's give this a shot!

* Fixes scipygh-22720. * This patch adds a few floating point exception handling shims to compensate for the upstream change for `np.dot` at numpy/numpy#28442, merged five days ago. This allows the Scipy testsuite to pass again locally on x86_64 Linux against NumPy `main`.

tylerjereddy · 2025-03-21T17:05:34Z

Note: required a small number of shims downstream, but not too bad I don't think.

* BUG: Check for floating point exceptions in dot Currently dot does not check for floating point exceptions, while similar functions (such as multiply and matmul) do. Add these checks for consistency. Note that certain types of floating point exceptions are still not caught: - Multiplications of 0 with nan or inf are ignored (see numpy#27902) - Integer overflows are ignored. These aren't true floating point exceptions, but the behavior is inconsistent with scalar multiply, i.e. ``` >>> np.int16(32000) * np.int16(3) <python-input-7>:1: RuntimeWarning: overflow encountered in scalar multiply np.int16(32000) * np.int16(3) np.int16(30464) >>> np.dot(np.int16(32000), np.int16(3)) np.int16(30464) ``` Fixes numpy#14925 * pr comments * parametrize test, add release note * use generic typenames

* Fixes scipygh-22720. * This patch adds a few floating point exception handling shims to compensate for the upstream change for `np.dot` at numpy/numpy#28442, merged five days ago. This allows the Scipy testsuite to pass again locally on x86_64 Linux against NumPy `main`.

github-actions bot added the 00 - Bug label Mar 6, 2025

rtimpe force-pushed the dot-invalid branch from 166fb82 to 7541ae1 Compare March 6, 2025 21:58

seberg reviewed Mar 9, 2025

View reviewed changes

melissawm added this to NumPy first-time contributor PRs Mar 10, 2025

melissawm moved this to Pending authors' response in NumPy first-time contributor PRs Mar 10, 2025

pr comments

70d08a6

rtimpe force-pushed the dot-invalid branch from a307daa to 70d08a6 Compare March 11, 2025 00:23

seberg reviewed Mar 14, 2025

View reviewed changes

parametrize test, add release note

21fefc0

use generic typenames

7e1390a

seberg changed the title ~~BUG: Check for floating point exceptions in dot~~ ENH: Check for floating point exceptions in dot Mar 16, 2025

seberg merged commit 432d9b2 into numpy:main Mar 16, 2025
73 checks passed

github-project-automation bot moved this from Pending authors' response to Completed in NumPy first-time contributor PRs Mar 16, 2025

tylerjereddy mentioned this pull request Mar 21, 2025

MAINT: compensate for dot exceptions scipy/scipy#22726

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

ENH: Check for floating point exceptions in dot #28442

ENH: Check for floating point exceptions in dot #28442

Uh oh!

rtimpe commented Mar 6, 2025

Uh oh!

seberg left a comment

Uh oh!

seberg Mar 9, 2025

Uh oh!

seberg Mar 9, 2025

Uh oh!

rtimpe Mar 11, 2025

Uh oh!

seberg Mar 9, 2025

Uh oh!

rtimpe commented Mar 11, 2025

Uh oh!

mattip commented Mar 11, 2025

Uh oh!

seberg left a comment

Uh oh!

seberg Mar 14, 2025

Uh oh!

rtimpe commented Mar 15, 2025

Uh oh!

seberg commented Mar 16, 2025

Uh oh!

Uh oh!

tylerjereddy commented Mar 21, 2025

Uh oh!

Uh oh!

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Uh oh!

ENH: Check for floating point exceptions in dot #28442

ENH: Check for floating point exceptions in dot #28442

Uh oh!

Conversation

rtimpe commented Mar 6, 2025

Uh oh!

seberg left a comment

Choose a reason for hiding this comment

Uh oh!

seberg Mar 9, 2025

Choose a reason for hiding this comment

Uh oh!

seberg Mar 9, 2025

Choose a reason for hiding this comment

Uh oh!

rtimpe Mar 11, 2025

Choose a reason for hiding this comment

Uh oh!

seberg Mar 9, 2025

Choose a reason for hiding this comment

Uh oh!

rtimpe commented Mar 11, 2025

Uh oh!

mattip commented Mar 11, 2025

Uh oh!

seberg left a comment

Choose a reason for hiding this comment

Uh oh!

seberg Mar 14, 2025

Choose a reason for hiding this comment

Uh oh!

rtimpe commented Mar 15, 2025

Uh oh!

seberg commented Mar 16, 2025

Uh oh!

Uh oh!

tylerjereddy commented Mar 21, 2025

Uh oh!

Uh oh!

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.