Skip to content

ENH: Re-enable VXE from build targets for sin/cos #27665

Merged
seberg merged 2 commits intonumpy:mainfrom
r-devulap:enable-vxe
Nov 22, 2024
Merged

ENH: Re-enable VXE from build targets for sin/cos #27665
seberg merged 2 commits intonumpy:mainfrom
r-devulap:enable-vxe

Conversation

@r-devulap
Copy link
Member

@r-devulap r-devulap commented Oct 29, 2024 •
edited by seiko2plus
Loading

As discussed in the optimization meeting, @seiko2plus wanted to split VSX and VXE into two separate PR's.

For clarification, SIMD optimizations for sine and cosine functions on both ppc64 and z/Architecture (IBM Z) were disabled by #25781 to bypass CI tests. This PR aims to re-enable optimizations for z/Architecture after addressing the following runtime errors, while #27627 re-enabled ppc64 optimizations.

numpy/_core/tests/test_umath_accuracy.py:84: AssertionError
=========================== short test summary info ============================
FAILED numpy/_core/tests/test_umath.py::TestSpecialFloats::test_sincos_values[f] - AssertionError: 
FAILED numpy/_core/tests/test_umath.py::TestSpecialFloats::test_sincos_errors[inf-f-sin] - AssertionError: FloatingPointError not raised by sin
FAILED numpy/_core/tests/test_umath.py::TestSpecialFloats::test_sincos_errors[inf-f-cos] - AssertionError: FloatingPointError not raised by cos
FAILED numpy/_core/tests/test_umath.py::TestSpecialFloats::test_sincos_errors[-inf-f-sin] - AssertionError: FloatingPointError not raised by sin
FAILED numpy/_core/tests/test_umath.py::TestSpecialFloats::test_sincos_errors[-inf-f-cos] - AssertionError: FloatingPointError not raised by cos
FAILED numpy/_core/tests/test_umath.py::TestAVXFloat32Transcendental::test_sincos_float32 - AssertionError: Arrays are not almost equal up to 2 ULP (max difference is ...
FAILED numpy/_core/tests/test_umath.py::TestAVXFloat32Transcendental::test_strided_float32 - AssertionError: Arrays are not equal to 2 ULP (max is 1.05785e+09)
FAILED numpy/_core/tests/test_umath_accuracy.py::TestAccuracy::test_validate_fp16_transcendentals[cos] - AssertionError: Arrays are not almost equal up to 1 ULP (max difference is ...
FAILED numpy/_core/tests/test_umath_accuracy.py::TestAccuracy::test_validate_fp16_transcendentals[sin] - AssertionError: Arrays are not almost equal up to 1 ULP (max difference is ...
= 9 failed, 22213 passed, 1559 skipped, 26186 deselected, 7 xfailed, 2 xpassed in 555.12s (0:09:15) =

@r-devulap r-devulap added component: SIMD Issues in SIMD (fast instruction sets) code or machinery HWY features related to google Highway labels Oct 29, 2024
@seberg seberg requested a review from seiko2plus November 7, 2024 09:41
seiko2plus
seiko2plus reviewed Nov 9, 2024
View reviewed changes
numpy/_core/src/umath/loops_trigonometric.dispatch.cpp
"SIMD in general.");
npy_uint64 simd_maski;
hn::StoreMaskBits(f32, simd_mask, (uint8_t*)&simd_maski);
hn::StoreMaskBits(f32, simd_mask, (uint8_t *)&simd_maski);
Copy link
Member

@seiko2plus seiko2plus Nov 9, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A better approach is needed for libc fallback to properly handle scalable extensions. One potential improvement could be introducing a new intrinsic, such as bool TestBit(MASK, int pos), which would also be endian-friendly. For now, I applied a quick fix below to address a big-endian bug, as passing a uint64_t as a uint8_t array obviously lead to accessing garbage bytes instead.

seiko2plus
seiko2plus reviewed Nov 9, 2024
View reviewed changes
@seiko2plus
Copy link
Member

seiko2plus commented Nov 9, 2024

@r-devulap, Oh, my editor automatically reformatted the entire source code based on the NumPy clang-format configuration. Would you prefer that I revert these changes, or is this formatting acceptable for you?

@seiko2plus
Copy link
Member

seiko2plus commented Nov 9, 2024

All CI tests for s390x have passed.

@seberg
Copy link
Member

seberg commented Nov 22, 2024

@r-devulap if you are happy about the changes please just put it in (I don't think the reformatting matters too much -- it's not that much).
2.2.0 is going to be branched very soon, it it seems reasonable to want to have this in.

@r-devulap
Copy link
Member Author

r-devulap commented Nov 22, 2024

@r-devulap, Oh, my editor automatically reformatted the entire source code based on the NumPy clang-format configuration. Would you prefer that I revert these changes, or is this formatting acceptable for you?

Yeah no worries. This looks better anyways.

@r-devulap
Copy link
Member Author

r-devulap commented Nov 22, 2024

@r-devulap if you are happy about the changes please just put it in (I don't think the reformatting matters too much -- it's not that much). 2.2.0 is going to be branched very soon, it it seems reasonable to want to have this in.

yeah makes sense. I had to fix some rebase problems. But once the CI passes it should be good to go in.

@seberg
Copy link
Member

seberg commented Nov 22, 2024

Let's put this in then, thanks @r-devulap and @seiko2plus.

@seberg seberg merged commit be54bff into numpy:main Nov 22, 2024
@Sunny-Anand
Copy link

Sunny-Anand commented Dec 13, 2024

Not sure if this is having any impact here. google/highway#2409

@Sunny-Anand
Copy link

Sunny-Anand commented Dec 13, 2024

I can confirm that trying to install numpy>= 2.0.0 for python3.10 breaks on UBI-8 image.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@seiko2plus seiko2plus seiko2plus left review comments

Assignees

No one assigned

Labels

01 - Enhancement component: SIMD Issues in SIMD (fast instruction sets) code or machinery HWY features related to google Highway

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@r-devulap @seiko2plus @seberg @Sunny-Anand

Comments