feat(pw/fft): CPE-DFTI sticks FFT as a factory-selected backend (FFT_SWDFTI)

[A-006]

What

Add a separate FFT backend FFT_SWDFTI that accelerates FFT_CPU's local 1D
sticks FFTs on the Sunway CPEs via the swFFT xMath-SACA DFTI API. It is selected
through the FFT_Bundle factory, so FFT_CPU itself stays free of any DFTI
#ifdef.

How

• New backend source/source_base/module_fft/fft_swdfti.{h,cpp}:
  FFT_SWDFTI<double> : public FFT_CPU<double>, overriding only:

  • fftzfor/fftzbac — batched 1D-z on CPE
  • fftxyfor/fftxybac — strided 1D-x on CPE (y stays on FFTW)
  • setupFFT — builds the DFTI descriptors after the base FFTW plans

  Non-xprime / disabled cases delegate to FFT_CPU. Runtime toggle
  ABACUS_NO_DFTI=1.

• FFT_Bundle factory: device "cpu" (double) instantiates FFT_SWDFTI
  when built with __SWDFTI, else FFT_CPU — the single backend-selection point.

• fft_cpu.h: members private → protected so the subclass can reuse
  plans/dims.

• CMake: USE_SWDFTI option (default ON under USE_SW) compiles
  fft_swdfti.cpp and defines __SWDFTI; adds -mieee for IEEE FP under
  USE_SW; links the objcopy-isolated libswfft_xmath_iso.a (avoids the
  fftw_* symbol hijack).

Safety / correctness

• Guarded so that OFF => byte-identical to develop (verified: fft_cpu /
  fft_bundle compile clean at USE_SW=OFF, fft_swdfti excluded from the
  baseline, cmake reconfigures clean).
• Measured (4xGaAs, ecut 60, 54^3 grid): veff_pw 1.7-1.8x, scales with
  np; energy bit-identical.

Files

• source/source_base/module_fft/fft_swdfti.{h,cpp} (new)
• source/source_base/module_fft/fft_bundle.cpp
• source/source_base/module_fft/fft_cpu.h
• source/source_basis/module_pw/CMakeLists.txt
• CMakeLists.txt

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

This PR adds an optional Sunway-specific CPU FFT backend that accelerates local 1D “sticks” FFTs using the swFFT xMath DFTI API, and wires it into the FFT factory and build system.

Changes:

• Introduces FFT_SWDFTI (derived from FFT_CPU) that offloads z/x 1D FFTs to DFTI while keeping other steps on FFTW.
• Updates FFT_Bundle to instantiate FFT_SWDFTI when __SWDFTI is enabled.
• Extends CMake to add the new source and configure Sunway build flags and linking for the isolated swFFT archive.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
source/source_basis/module_pw/CMakeLists.txt	Conditionally compiles the new SWDFTI FFT implementation into the build.
source/source_base/module_fft/fft_swdfti.h	Declares the new FFT_SWDFTI backend class.
source/source_base/module_fft/fft_swdfti.cpp	Implements SWDFTI setup/compute paths for z/x FFTs and falls back to FFTW as needed.
source/source_base/module_fft/fft_cpu.h	Exposes internals to allow the SWDFTI subclass to reuse FFT_CPU plans/dimensions.
CMakeLists.txt	Adds Sunway compile flag checks, the USE_SWDFTI option, and changes SW linking to an isolated swFFT archive.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[A-006]

Addressed the Copilot review (commit b780c63):

• cleanFFT descriptor leak — the z/x DFTI descriptors are now freed via DftiFreeDescriptor before the handles are nulled (null-guarded).
• Non-thread-safe init — DftiInitAthread is now a std::call_once one-time CPE spawn instead of a plain static-int guard.
• Unconditional iso-archive link — libswfft_xmath_iso.a is linked only when USE_SWDFTI is ON, with a clear FATAL_ERROR if the archive is missing.
• Self-contained header — fft_swdfti.h now includes <complex> explicitly.
• private → protected in fft_cpu.h — replied inline: intentional and minimal-scope for the single FFT_SWDFTI subclass; happy to narrow to a friend if preferred.