[hipFFTW] hipFort bindings and tests by marikurz-amd · Pull Request #295 · ROCm/hipfort

marikurz-amd · 2026-04-10T11:15:16Z

Status: This is a draft. The Fortran API is not complete yet, but can be extended in future work. This PR is opened mainly to generate visibility and avoid duplicate work.

Motivation

This PR starts the addition of FFTW-style Fortran interfaces that allow to call hipFFTW through hipFort. This PR adds a first batch of interfaces, in particular double-precision, legacy-style interfaces for the basic and advanced API. It also adds a CMake structure that builds tests depending on the hipfort::hipfftw target being available or not.

Technical Details

First commit: Adds hipfort_hipfftw.F90 with the first hipFFTW Fortran interfaces (double precision, “legacy” style types; basic and new-array transforms).
Second commit: Adds the advanced *_many interfaces and tidies formatting and comments in that file.
Third commit: Adds F2003 tests for 1D C2C, R2C/C2R round-trip, and batched many cases (including interleaved, padded embed, and many-vs-individual). Updates CMake and package config so the hipfort::hipfftw target is exported and tests can build; updates Makefiles under test/ as needed.

The build only runs a test if the matching C API symbols exist in libhipfftw, so older or partial ROCm installs do not fail the whole test suite.

Test Plan

Configure with ROCm so hip::hipfftw and libhipfftw are found.
Build the hipFFTW piece and the hipfftw_* tests when the required symbols are present.
Run the three tests: hipfftw_c2c, hipfftw_r2c_c2r, and hipfftw_many. Expect some tests to be skipped if a symbol is not in the library yet.
The tests cover the majority of implemented API calls and strive to reduce duplicate testing.

Test Result

ROCm	hipFFTW CMake target	Tests built	CTest summary	hipFFTW tests run¹
7.0.0	skipped	30	30 / 30 passed	0
7.0.1	skipped	30	30 / 30 passed	0
7.0.2	skipped	30	30 / 30 passed	0
7.1.0	yes	30	30 / 30 passed	0
7.1.1	yes	30	30 / 30 passed	0
7.2.0	yes	32	32 / 32 passed	2
7.2.1	yes	32	32 / 32 passed	2

¹ On 7.2.x, CTest registered four hipFFTW-related tests; two binaries were built and executed (hipfftw_c2c, hipfftw_r2c_c2r). All passed.

Submission Checklist

[x ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

…y legacy-type interfaces in double-precision for the basic and new-array transforms.

Add F2003 GPU tests covering 1D C2C, R2C/C2R round-trip, and advanced batched many interfaces (interleaved, padded embed, many-vs-individual). Include corresponding CMake target export and package config integration.

…n built against exports the necessary symbols.

regan-amd

This looks like the right direction to me. Most comments are cosmetic and/or simple suggestions...

I'm assuming upcoming work items include (not necessarily in this order)

expanding test coverage for
- dfftw_plan_dft_2d, dfftw_plan_dft_3d, dfftw_plan_dft (and r2c/c2r equivalent);
- dfftw_execute;
- in-place transforms;
expanding interface to guru and guru64 plan creation subroutines as well along with testing thereof;
covering single-precision as well.

Is that correct? If not, please clarify what's not planned for (and/or what else is).

regan-amd · 2026-05-12T22:47:40Z

+! ==============================================================================
+! hipfort: FORTRAN Interfaces for GPU kernels
+! ==============================================================================
+! Copyright (c) 2020-2022 Advanced Micro Devices, Inc. All rights reserved.


Suggested change

! Copyright (c) 2020-2022 Advanced Micro Devices, Inc. All rights reserved.

! Copyright (c) 2026 Advanced Micro Devices, Inc. All rights reserved.

regan-amd · 2026-05-12T22:55:02Z

+  !>   @details Allocates and initializes an N-dimensional complex-to-complex FFT plan.
+  !>
+  !>   @param[in] rank Number of dimensions.
+  !>   @param[in] n Array of dimensions.


Suggested change

!> @param[in] n Array of dimensions.

!> @param[in] n Array of lengths.

+ same comment elsewhere.

regan-amd · 2026-05-12T23:12:01Z

+  !>  @brief C interface: Create a 2D complex-to-complex FFT plan.
+  !>
+  !>   @details Allocates and initializes a 2D complex-to-complex FFT plan.
+  !>
+  !>   @param[in] n0 Length of the transform in the first dimension.
+  !>   @param[in] n1 Length of the transform in the second dimension.
+  !>   @param[in] in Input array.
+  !>   @param[in] out Output array.
+  !>   @param[in] sign Sign of the exponent in the formula that defines the Fourier transform.
+  !>   @param[in] flags FFTW planner flags.
+  !>   @return Handle to the FFT plan.
+  interface


Maybe we should leave the (private) binding functions undocumented and keep documentation only for the public subroutines of this module? I'm afraid this may trigger some confusion regarding the ordering of dimensions for users, e.g., w.r.t. the meaning of "{first, second, third} dimension" across the various documentation bits in this file... Unless I'm mistaken, these functions are not directly accessible to users of this module so there is no need for them to know much about...

Alternatively, if we feel strongly about keeping the private binders documented, my suggestion would be to change these by a simple comment referring them to the actual C header file hipfft/hipfftw.h if they're ever seeking more information about this. That would

alleviate duplication by design and naturally avoid out-of-sync documentation updates, if any;

implicitly indicate to users/interested readers that they're then leaving the realm of Fortran for C conventions if following the link (thereby abandoning their Fortran-specific default dimension-ordering convention).

regan-amd · 2026-05-12T23:41:53Z

+    implicit none
+    integer(c_int64_t) :: plan
+    integer(c_int) :: rank
+    integer(c_int), dimension(rank), target :: n


Suggested change

integer(c_int), dimension(rank), target :: n

integer(c_int), dimension(rank), target, intent(in) :: n

+ same elsewhere for all subroutines' arguments of non-primary type.

regan-amd · 2026-05-13T17:17:26Z

+  end if
+
+  ! Backward C2R: c2r(r2c(x)) should equal N * x
+  call hipCheck(hipMemcpy(dy, c_loc(hresult_c(1)), Nbytes_c, hipMemcpyHostToDevice))


This copy is unnecessary in the current form given the verifications done above.

regan-amd · 2026-05-13T17:23:27Z

+    endif()
+
+    if(HIPFFTW_HAS_EXECUTE_DFT)
+      hipfort_add_test(hipfftw hipfftw_c2c f2003)


Note: I believe you could easily enable this hipfftw_c2c test for the dfftw_execute subroutine as well (e.g., with pre-processor compilation-time defines): the execution arrays dx_f and dx_y were used at plan creation so dfftw_execute (using the plan only) should operate on those.

That would enable some minimal coverage even for the "1.0.22" version of hipFFT.

regan-amd · 2026-05-13T18:00:31Z

+    integer, intent(inout) :: nfail
+    integer(c_int), parameter :: NX = 4, NY = 6, LDX = 8, howmany = 2
+    integer(c_size_t), parameter :: total_bytes = LDX * NY * howmany * 16
+    complex(c_double_complex), allocatable, target :: hx(:), hresult(:)


Why not use Fortran multi-dimensional allocatable arrays, e.g., with 3 array dimensions in this case (2 for length and 1 for batch)? Wouldn't that make things a little more readable and better align with standard Fortran coding practices? If taking the suggestion, please apply consistently throughout.

regan-amd · 2026-05-13T18:06:29Z

+  call hipCheck(hipMalloc(dx, Nbytes))
+  call hipCheck(hipMalloc(dy, Nbytes))


hipFFTW is supposed to be able to handle host-residing I/O data, and my understanding was that it would be the main use case scenario in fact. If the latter is true for Fortran users as well, I would recommend testing with host-residing I/O data, too (and/or only that if test scopes must be kept small)...
That would significantly simplify the tests and align better with the most relevant use case scenario (in my understanding).

regan-amd · 2026-05-13T18:36:09Z

@@ -0,0 +1,3 @@
+CFLAGS ?= -lhipfft -lhipfftw -lrocfft


Suggested change

CFLAGS ?= -lhipfft -lhipfftw -lrocfft

CFLAGS ?= -lhipfftw -lrocfft

If actually needed, could you clarify what it is needed for? This is surprising to me.

Also, why isn't it FFLAGS?

marikurz-amd added 3 commits December 12, 2025 07:39

Introduces basic hipfort interfaces for the hipFFTW API. Contains onl…

86968f1

…y legacy-type interfaces in double-precision for the basic and new-array transforms.

Add advanced *_many interfaces and improve formatting and descriptions

54f09ad

Add Fortran test programs and build support for hipfftw

a71d5fb

Add F2003 GPU tests covering 1D C2C, R2C/C2R round-trip, and advanced batched many interfaces (interleaved, padded embed, many-vs-individual). Include corresponding CMake target export and package config integration.

marikurz-amd requested a review from cgmb as a code owner April 10, 2026 11:15

marikurz-amd marked this pull request as draft April 10, 2026 11:16

marikurz-amd changed the title ~~Draft: hipFort bindings and tests for hipFFTW~~ hipFFTW: hipFort bindings and tests Apr 10, 2026

marikurz-amd changed the title ~~hipFFTW: hipFort bindings and tests~~ [hipFFTW] hipFort bindings and tests Apr 10, 2026

Add logic to compile hipfftw tests only if the hipFFTW library versio…

ab19b53

…n built against exports the necessary symbols.

regan-amd reviewed May 13, 2026

View reviewed changes

ronlieb self-requested a review May 25, 2026 19:57

ronlieb approved these changes May 25, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[hipFFTW] hipFort bindings and tests#295

[hipFFTW] hipFort bindings and tests#295
marikurz-amd wants to merge 4 commits into
developfrom
users/marikurz/hipfftw

marikurz-amd commented Apr 10, 2026

Uh oh!

regan-amd left a comment

Uh oh!

regan-amd May 12, 2026

Uh oh!

regan-amd May 12, 2026

Uh oh!

regan-amd May 12, 2026

Uh oh!

Uh oh!

regan-amd May 12, 2026

Uh oh!

regan-amd May 13, 2026

Uh oh!

regan-amd May 13, 2026

Uh oh!

regan-amd May 13, 2026

Uh oh!

regan-amd May 13, 2026

Uh oh!

regan-amd May 13, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

	! Copyright (c) 2020-2022 Advanced Micro Devices, Inc. All rights reserved.
	! Copyright (c) 2026 Advanced Micro Devices, Inc. All rights reserved.

	!> @param[in] n Array of dimensions.
	!> @param[in] n Array of lengths.

	integer(c_int), dimension(rank), target :: n
	integer(c_int), dimension(rank), target, intent(in) :: n

		call hipCheck(hipMalloc(dx, Nbytes))
		call hipCheck(hipMalloc(dy, Nbytes))

	CFLAGS ?= -lhipfft -lhipfftw -lrocfft
	CFLAGS ?= -lhipfftw -lrocfft

Conversation

marikurz-amd commented Apr 10, 2026

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

Uh oh!

regan-amd left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants