Spin push for thin kick elements by cemitch99 · Pull Request #1606 · BLAST-ImpactX/impactx

cemitch99 · 2026-08-10T22:33:22Z

This PR adds the spin push through those Thin elements not associated with fringe fields. The logic for all of these elements uses the same pattern, and only the evaluation of the magnetic or electric fields should differ.

The implementation here is $s$-symmetric. A reversibility test will be included.

This closes Issue #1322

codspeed · 2026-08-10T22:49:11Z

Merging this PR will not alter performance

✅ 58 untouched benchmarks
🆕 10 new benchmarks

Performance Changes

	Benchmark	`BASE`	`HEAD`	Efficiency
🆕	`test_Buncher[nospin]`	N/A	268.7 µs	N/A
🆕	`test_Buncher[spin]`	N/A	1.3 ms	N/A
🆕	`test_Kicker[nospin]`	N/A	110.5 µs	N/A
🆕	`test_Kicker[spin]`	N/A	1.3 ms	N/A
🆕	`test_NonlinearLens[nospin]`	N/A	1.4 ms	N/A
🆕	`test_NonlinearLens[spin]`	N/A	2.6 ms	N/A
🆕	`test_ShortRF[nospin]`	N/A	246.4 µs	N/A
🆕	`test_ShortRF[spin]`	N/A	861.6 µs	N/A
🆕	`test_ThinDipole[nospin]`	N/A	199.2 µs	N/A
🆕	`test_ThinDipole[spin]`	N/A	552.7 µs	N/A

_{Comparing cemitch99:spin_push_thin_kicks (0c24f29) with development (a488422)}

for more information, see https://pre-commit.ci

cemitch99 · 2026-08-11T21:22:23Z

The error is due to a warning associated with a possible loss of precision (double > float conversion) when running w/ SIMD.

CI is correct, this was a bug. Co-authored-by: Axel Huebl <axel.huebl@plasma.ninja>

ax3l · 2026-08-11T23:23:25Z

Thanks for the PR! 🎉

I fixed the bug that CI caught.
Will do a detailed review a bit later :)

Used by compilers and our SIMD implementation to optimize write-backs into memory.

The spin push duplicated the complex sqrt/log/division block from operator() and then called operator(), which recomputed it. That evaluated the most expensive part of the element twice per particle. Hoist it into dF_dzeta() and single-source the momentum kick in apply_kick(), so the phase space push and the spin push share one evaluation. F'(zeta) does not change over the thin kick, so the spin push reuses the same value for the magnetic field and for the kick. Numerically identical: both evaluations had identical inputs, and py - m_kick*dF.m_imag is the same IEEE operation as the previous py + dpy with dpy = -m_kick*dF.m_imag. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

ax3l · 2026-08-12T02:17:34Z

@cemitch99 a general question on the implementation of half-kicks here:
Excluding ShortRF, for the elements here (Kicker, Buncher, NonlinearLens, ThinDipole) the fields depend only on coordinates the kick leaves unchanged (x, y, plus t for Buncher). Like TaperedPL, we could also achieve an exactly reversible kick via midpoint push, that is second order and exactly reversible, too.

midpoint is 1 tbmt_precession_vector + 1 rotate_spin instead of 2 of each. rotate_spin carries a sincos and a sqrt, so it's the expensive half.
In #1609 I also saw a 2x smaller error constant (midpoint vs. trapezoidal rule behavior) for the midpoint scheme.

Is there a reason to prefer the half-kick structure that I'm missing?

The geometric factor -(1 + h*x)/beta_gamma_z in the Thomas-BMT precession vector is a Frenet-Serret Jacobian and needs the design-orbit curvature. Passing h and B_y at half strength and scaling the generator by the full arc length gives the correct half-step contribution for every term that enters linearly, but evaluates (1 + h*x) with half the curvature, an error of order 0.5*x/rc on the field part of the rotation. Reversibility does not catch it: the construction is self-inverting for any value of h. Move the factor of one half from h and B_y onto the arc length, which is equivalent for the linear terms and matches the convention already used by ExactCFbend (h = 1/rc, unscaled fields, generator scaled by the step). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Evaluate the Thomas-BMT generator once, at the momentum halfway through the kick, instead of applying two half rotations around the phase space kick. This is the structure already used by TaperedPL. Both forms are exactly reversible and second order: the reversed element maps p_out back to p_in, so it sees the same midpoint, and the negated strength gives the inverse rotation. They differ at third order, where the midpoint rule has half the error constant of the trapezoidal rule that two half kicks amount to. Per particle this costs 1 tbmt_precession_vector + 1 rotate_spin instead of 2 of each. Measured on the thin Multipole, which has the same structure, that was ~28% off the spin push (~39% of the spin-specific part) and 2x smaller error against a converged reference. Applies to Kicker, Buncher, NonlinearLens and ThinDipole, whose fields depend only on coordinates the kick leaves unchanged. Buncher also changes the energy deviation, so pt is averaged along with the transverse momenta. ShortRF keeps the half kick form: it changes the reference energy, so px/py/pt are renormalized across the kick and a plain average would mix two normalizations. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

ax3l · 2026-08-12T03:25:53Z

In case midpush sounds good to you, I prepared a PR for this branch that you can apply directly:
cemitch99#4

If you merge the linked PR to your fork then the PR here will automatically update.

Performance: Midpoint Spin Push for Thin Kicks

cemitch99 · 2026-08-12T18:48:18Z

@ax3l Ok, ready for re-review.

ax3l

Looks great, thank you! 🎉

Add Kicker spin push.

5df8518

cemitch99 changed the title ~~Add Kicker spin push.~~ Spin push for thin kick elements Aug 10, 2026

cemitch99 requested review from ax3l and qianglbl August 10, 2026 22:34

cemitch99 and others added 9 commits August 10, 2026 16:24

Add ShortRF and update reversibility test.

6e6de1d

[pre-commit.ci] auto fixes from pre-commit.com hooks

5c8fe30

for more information, see https://pre-commit.ci

Clean-up of ShortRF.

c81d008

Add spin push for Buncher.

1a6bc77

Add spin push ThinDipole.

003057d

Add spin push NonlinearLens.

83dc321

Update reversibility tests.

99285f1

[pre-commit.ci] auto fixes from pre-commit.com hooks

5273f2f

for more information, see https://pre-commit.ci

Fix reversibility of ThinDipole.

74cd762

cemitch99 added the component: spin tracking Spin tracking label Aug 11, 2026

cemitch99 added this to the FY26-27 LDRD: future colliders & light sources milestone Aug 11, 2026

cemitch99 added the component: elements Elements/maps/external fields label Aug 11, 2026