[Growl1234]

• Prepare kernel constants once, skip unused C6 and ATM work, remove unused buffers, and move repeated calculations out of inner loops.
• Support radix-2/3 FFT meshes to avoid excessive padding, and reuse kernel values for opposite wave vectors while handling Nyquist planes correctly.
• Combine symmetric pole contributions and streamline mesh initialization and contraction.

For the supplied 576-atom cell with OMP_NUM_THREADS=24 ./s-dftd3 1.xyz --zero pbe --grad --ewald, wall time has decreased from 96.536 s to 2.726 s.

[codecov]

Codecov Report

❌ Patch coverage is 83.90805% with 14 lines in your changes missing coverage. Please review.
✅ Project coverage is 63.45%. Comparing base (6a0bb18) to head (9142b51).

Files with missing lines	Patch %	Lines
src/dftd3/fourier/kernel.f90	69.56%	1 Missing and 6 partials ⚠️
src/dftd3/disp.f90	70.00%	0 Missing and 3 partials ⚠️
src/dftd3/fourier/fft.f90	95.34%	0 Missing and 2 partials ⚠️
src/dftd3/fourier/ewald.f90	50.00%	0 Missing and 1 partial ⚠️
src/dftd3/fourier/spme.f90	88.88%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #244      +/-   ##
==========================================
+ Coverage   63.43%   63.45%   +0.02%     
==========================================
  Files          47       47              
  Lines        6831     6855      +24     
  Branches     2422     2438      +16     
==========================================
+ Hits         4333     4350      +17     
- Misses       1190     1191       +1     
- Partials     1308     1314       +6     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[awvwgk]

Thank you, looks good to me

[Growl1234]

Side note:

• With radix-2/3 FFT support, --ewald-mesh now rounds up to the smallest size of the form $(2^a)(3^b)$, instead of the next power of two. For example, --ewald-mesh 288 now uses $288^3$ rather than $512^3$, so identical command-line settings may produce different actual meshes.
• Runtime is substantially lower even at a fixed $512^3$ mesh (--ewald-mesh 512; now about 19~20s), thanks to kernel precomputation, symmetry reuse, the revised FFT implementation, and fewer redundant mesh writes.

Now this PR should be ready. There maybe follow-up work worth considering, but I don't expect it to bring significant performance improvement.