From bdba5b8403fd7816ebf53d2d0634d0dd4724efc4 Mon Sep 17 00:00:00 2001
From: Peter Boyle <paboyle@ph.ed.ac.uk>
Date: Tue, 19 May 2026 21:49:06 -0400
Subject: [PATCH] FFT: use host stack buffer in PlanCreate, not deviceVector

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 Grid/algorithms/FFT.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/Grid/algorithms/FFT.h b/Grid/algorithms/FFT.h
index 406a49b6..8bacc274 100644
--- a/Grid/algorithms/FFT.h
+++ b/Grid/algorithms/FFT.h
@@ -275,8 +275,10 @@ public:
 
       // GPU backends (cuFFT/hipFFT) ignore the buffer pointer at plan creation.
       // CPU FFTW with FFTW_ESTIMATE inspects only alignment and never touches data.
-      deviceVector<scalar> dummy(2);
-      FFTW_scalar *buf = (FFTW_scalar *)&dummy[0];
+      // Use a host stack buffer: a device allocation here triggers a rocFFT RTC
+      // bug on ROCm 7 that causes plan creation to fail for small transform sizes.
+      scalar stack_dummy[2] = {};
+      FFTW_scalar *buf = (FFTW_scalar *)stack_dummy;
 
       {
         FFTW_plan p = FFTW<scalar>::fftw_plan_many_dft(