[amontoison]

Summary

Binds the hipFFT Xt API in hipfort_hipfft, as plain bind(C) interfaces, regenerated from hipfftXt.h and hipfftMp.h (which include hipfft.h):

• the multi-GPU descriptor API: hipfftXtSetGPUs, hipfftXtMalloc, hipfftXtMemcpy, hipfftXtFree, hipfftXtExecDescriptor{,C2C,R2C,C2R,Z2Z,D2Z,Z2D};
• hipfftXtMakePlanMany, hipfftXtGetSizeMany and hipfftXtExec on hipDataType;
• the callback routines (hipfftXtSetCallback, hipfftXtClearCallback, hipfftXtSetCallbackSharedSize, hipfftXtSetJITCallback);
• the multi-process routines of hipfftMp.h: hipfftMpAttachComm, hipfftXtSetDistribution, hipfftXtSetSubformatDefault.

hipXtDesc and hipLibXtDesc are bind(C) derived types in a new hipfort_hipfft_types module; the Xt enums are in hipfort_hipfft_enums.

rocFFT and hipFFTW need no change: rocFFT has no Xt routines (its multi-GPU API, rocfft_field_* / rocfft_brick_*, is already bound), and hipfftw.h has no Xt API.

Notes

• The hipfftMp.h routines need an MPI-enabled hipFFT. They are compiled only with HIPFFT_MPI_ENABLE, and the ROCm packages do not export them. A bind(C) interface adds no reference until a program calls it, so other programs link as before; the symbol test skips these three. hipfftMpAttachComm takes a pointer to a C MPI_Comm (build it with MPI_Comm_f2c), which hipFFT keeps, so it has to outlive the plan.
• Descriptors are owned by hipFFT. Keep the type(c_ptr) that hipfftXtMalloc returns, c_f_pointer it to a type(hipLibXtDesc), pointer, and pass that pointer to hipfftXtExecDescriptor* / hipfftXtFree, never a copy of the type.
• AMD limits, not binding limits: the descriptor routines require a plan over two or more GPUs, unbatched 1D is not implemented there, and a multi-GPU plan writes one work size per GPU (pass the first element of an array with one slot per GPU).
• CUDA backend: the Xt and Mp routines are HIP-only (#ifndef USE_CUDA_NAMES), like every routine without a verified cufft twin.

Documentation

docs/tutorials/hipfft-examples.rst gains three sections: Xt plans on hipDataType, multi-GPU transforms (the descriptor workflow, ownership, one work size per GPU, the AMD limits, with the descriptor test as the example), and multi-process transforms with MPI (the C MPI_Comm helper and the lifetime rule).

Tests

• hipfft_xt_makeplanmany_c: hipfftXtGetSizeMany + hipfftXtMakePlanMany + hipfftXtExec, batched C2C round trip on one GPU.
• hipfft_xt_descriptor_z: 2D Z2Z round trip through XtSetGPUs / XtMalloc / XtMemcpy / XtExecDescriptorZ2Z / XtFree on two GPUs; it passes trivially with fewer than two.
• test_hipfft (symbol test) links every new routine.

[amontoison]

@malcolmroberts Is it fine for you?

[Abuudiii]

Tested on 2 GPUs (gfx1201, ROCm 7.2.4): hipfft_xt_descriptor_z passes, and all 33 hipFFT tests pass.