Sourceme's for frontier

worth maintaining

Grid/allocator/MemoryManager.cc

@@ -63,12 +63,10 @@ void MemoryManager::PrintBytes(void)
   std::cout << " MemoryManager : "<<(total_device>>20)<<" accelerator Mbytes "<<std::endl;
   std::cout << " MemoryManager : "<<(total_host>>20)  <<" cpu         Mbytes "<<std::endl;
   uint64_t cacheBytes;
-  cacheBytes = CacheBytes[Cpu];
-  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" cpu cache Mbytes "<<std::endl;
-  cacheBytes = CacheBytes[Acc];
-  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" acc cache Mbytes "<<std::endl;
-  cacheBytes = CacheBytes[Shared];
-  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" shared cache Mbytes "<<std::endl;
+  cacheBytes = HostCacheBytes();
+  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" cpu alloc cache Mbytes "<<std::endl;
+  cacheBytes = DeviceCacheBytes();
+  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" acc alloc cache Mbytes "<<std::endl;
   
 #ifdef GRID_CUDA
   cuda_mem();

Grid/lattice/Lattice_base.h

@@ -289,7 +289,7 @@ public:
   ///////////////////////////////////////////
   // move constructor
   ///////////////////////////////////////////
-  Lattice(Lattice && r){ 
+  Lattice(Lattice && r) noexcept { 
     this->_grid = r.Grid();
     this->_odata      = r._odata;
     this->_odata_size = r._odata_size;
@@ -330,7 +330,7 @@ public:
   ///////////////////////////////////////////
   // Move assignment possible if same type
   ///////////////////////////////////////////
-  inline Lattice<vobj> & operator = (Lattice<vobj> && r){
+  inline Lattice<vobj> & operator = (Lattice<vobj> && r) noexcept {
 
     resize(0); // deletes if appropriate
     this->_grid       = r.Grid();

Grid/lattice/Lattice_reduction_gpu.h

@@ -438,5 +438,11 @@ inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osite
   result = sumD_gpu_large(lat,osites);
   return result;
 }
+template<class Word> Word checksum_gpu(Word *vec,uint64_t L)
+{
+  Word w;
+  bzero(&w,sizeof(w));
+  return w;
+}
 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_slicesum_core.h

@@ -59,7 +59,7 @@ inline void sliceSumReduction_cub_small(const vobj *Data,
 #if defined(__CUDACC__) && (__CUDACC_VER_MAJOR__ >= 13)
   #define GRID_CUB_SUM_OP ::cuda::std::plus<>{}
 #else
-  #define GRID_CUB_SUM_OP ::cub::Sum()
+  #define GRID_CUB_SUM_OP ::gpucub::Sum()
 #endif
   
   gpuError_t gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p,d_out, rd, d_offsets, d_offsets+1, GRID_CUB_SUM_OP, zero_init, computeStream);

Grid/perfmon/PerfCount.h

@@ -51,8 +51,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #endif
 #ifdef __x86_64__
 #ifdef GRID_CUDA
-//accelerator_inline uint64_t __rdtsc(void) {  return 0; }
-//accelerator_inline uint64_t __rdpmc(int ) {  return 0; }
+accelerator_inline uint64_t __rdtsc(void) {  return 0; }
+accelerator_inline uint64_t __rdpmc(int ) {  return 0; }
 #else
 #include <x86intrin.h>
 #endif
@@ -93,8 +93,7 @@ inline uint64_t cyclecount(void){
 }
 #elif defined __x86_64__
 inline uint64_t cyclecount(void){ 
-  uint64_t ret = __rdtsc();
-  return (uint64_t)ret;
+  return (uint64_t)0;
 }
 #else
 

Grid/simd/Simd.h

@@ -115,10 +115,10 @@ accelerator_inline ComplexF adj(const ComplexF& r ){ return(conjugate(r)); }
 
 #if defined(GRID_CUDA) || defined(GRID_HIP)
 //Provide for convenience
-accelerator_inline std::complex<double> conjugate(const std::complex<double>& r){ return(conj(r)); }
-accelerator_inline std::complex<float>  conjugate(const std::complex<float>& r) { return(conj(r)); }
-accelerator_inline std::complex<double> adj(const std::complex<double>& r)      { return(conj(r)); }
-accelerator_inline std::complex<float>  adj(const std::complex<float>& r)       { return(conj(r)); }
+inline std::complex<double> conjugate(const std::complex<double>& r){ return(conj(r)); }
+inline std::complex<float>  conjugate(const std::complex<float>& r) { return(conj(r)); }
+inline std::complex<double> adj(const std::complex<double>& r)      { return(conj(r)); }
+inline std::complex<float>  adj(const std::complex<float>& r)       { return(conj(r)); }
 #endif
 
 accelerator_inline RealF real(const RealF  & r){ return r; }

systems/Booster/config-command

@@ -8,7 +8,6 @@ LIME=/p/home/jusers/boyle2/juwels/gm2dwf/boyle/
     --disable-gparity \
     --disable-fermion-reps \
     --with-lime=$LIME \
-    --enable-accelerator-cshift \
     --disable-unified \
     CXX=nvcc \
     LDFLAGS="-cudart shared " \

systems/Frontier/config-command

@@ -13,8 +13,8 @@ CLIME=`spack find --paths c-lime@2-3-9 | grep c-lime| cut -c 15-`
 --with-mpfr=/opt/cray/pe/gcc/mpfr/3.1.4/ \
 --disable-fermion-reps \
 CXX=hipcc MPICXX=mpicxx \
-CXXFLAGS="-fPIC -I${ROCM_PATH}/include/ -I${MPICH_DIR}/include -L/lib64 " \
- LDFLAGS="-L/lib64 -L${ROCM_PATH}/lib -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa -lhipblas -lrocblas -lhipfft"
+CXXFLAGS="-fPIC -I${ROCM_PATH}/include/ -I${MPICH_DIR}/include " \
+ LDFLAGS="-L${ROCM_PATH}/lib -L${MPICH_DIR}/lib -lmpi -lmpi_gtl_hsa -lhipblas -lrocblas -lhipfft -lamdhip64"
 
 
 

systems/Frontier/sourceme-rocm6.4.sh

@@ -0,0 +1,14 @@
+
+echo spack
+. /autofs/nccs-svm1_home1/paboyle/spack/share/spack/setup-env.sh
+
+export CLIME=`spack find --paths c-lime | grep ^c-lime | awk '{print $2}' `
+export MPFR=`spack find --paths mpfr    | grep ^mpfr  | awk '{print $2}' `
+export OPENSSL=`spack find --paths openssl | grep openssl | awk  '{print $2}' `
+export GMP=`spack find --paths gmp      | grep ^gmp | awk '{print $2}' `
+
+module load cce/20.0.0
+module load cpe/25.09
+module load rocm/6.4.0
+export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/opt/rocm-6.4.0/lib/llvm/lib/:$LD_LIBRARY_PATH

systems/Frontier/sourceme-rocm7.0.sh

@@ -0,0 +1,10 @@
+
+echo spack
+. /autofs/nccs-svm1_home1/paboyle/spack/share/spack/setup-env.sh
+
+
+module load cce/21.0.0
+module load cpe/26.03
+module load rocm/7.0.2
+export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/opt/rocm-7.0.2/lib/llvm/lib/:$LD_LIBRARY_PATH

systems/Frontier/sourceme-rocm7.2.sh

@@ -0,0 +1,14 @@
+
+echo spack
+. /autofs/nccs-svm1_home1/paboyle/spack/share/spack/setup-env.sh
+
+export CLIME=`spack find --paths c-lime | grep ^c-lime | awk '{print $2}' `
+export MPFR=`spack find --paths mpfr    | grep ^mpfr  | awk '{print $2}' `
+export OPENSSL=`spack find --paths openssl | grep openssl | awk  '{print $2}' `
+export GMP=`spack find --paths gmp      | grep ^gmp | awk '{print $2}' `
+
+module load cce/21.0.0
+module load cpe/26.03
+module load rocm/7.2.0
+export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/opt/rocm-7.2.0/lib/llvm/lib/:$LD_LIBRARY_PATH

systems/Frontier/sourceme.sh

@@ -1,16 +1,10 @@
 
 echo spack
-. /autofs/nccs-svm1_home1/paboyle/Crusher/Grid/spack/share/spack/setup-env.sh
+. /autofs/nccs-svm1_home1/paboyle/spack/share/spack/setup-env.sh
 
-module load amd/7.0.2
-module load cray-fftw
-module load craype-accel-amd-gfx90a
-mkdir $HOME/LD_PATH
-ln -s /opt/rocm-6.4.2/lib/libamdhip* $HOME/LD_PATH
 
-#Ugly hacks to get down level software working on current system
-export LD_LIBRARY_PATH=/opt/cray/libfabric/1.20.1/lib64/:$LD_LIBRARY_PATH
-export LD_LIBRARY_PATH=/opt/gcc/mpfr/3.1.4/lib:$LD_LIBRARY_PATH
-#export LD_LIBRARY_PATH=`pwd`/:$LD_LIBRARY_PATH
-export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/LD_PATH/
-export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/rocm-7.0.2/lib
+module load cce/21.0.0
+module load cpe/26.03
+module load rocm/7.0.2
+export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/opt/rocm-7.0.2/lib/llvm/lib/:$LD_LIBRARY_PATH

tests/Test_fft_memory.cc

@@ -0,0 +1,261 @@
+/*************************************************************************************
+
+ Test_fft_memory.cc
+
+ Memory growth test for PlannedFFT on a spin-colour matrix (propagator) field.
+
+ The test creates a single PlannedFFT object (which allocates FFTW plans once),
+ then repeatedly applies FFT_all_dim to the same propagator 400 times.
+
+ If PlannedFFT is working correctly the RSS should remain flat after the first
+ iteration — no new plans, no new deviceVector allocations beyond the per-call
+ pencil buffer which is freed at the end of each FFT_dim_execute call.
+
+ Build exactly like any other Grid test, e.g.:
+   make Test_fft_memory
+ or compile manually:
+   $(CXX) $(CXXFLAGS) Test_fft_memory.cc -o Test_fft_memory $(LDFLAGS)
+
+*************************************************************************************/
+
+#include <Grid/Grid.h>
+using namespace Grid;
+
+// --------------------------------------------------------------------------
+// Helper: read RSS (resident set size) in kB from /proc/self/status.
+// Returns 0 on platforms where /proc is unavailable.
+// --------------------------------------------------------------------------
+static long getCPURSSKb()
+{
+  long rss = 0;
+  FILE *fp = fopen("/proc/self/status", "r");
+  if (!fp) return -1;
+  char line[256];
+  while (fgets(line, sizeof(line), fp)) {
+    if (strncmp(line, "VmRSS:", 6) == 0) {
+      sscanf(line + 6, "%ld", &rss);
+      break;
+    }
+  }
+  fclose(fp);
+  return rss;
+}
+
+static long getGPUUsedMb()
+{
+#if defined(GRID_CUDA)
+  size_t free_bytes  = 0;
+  size_t total_bytes = 0;
+  cudaError_t err = cudaMemGetInfo(&free_bytes, &total_bytes);
+  if (err != cudaSuccess) return -1;
+  return (long)((total_bytes - free_bytes) / (1024 * 1024));
+ 
+#elif defined(GRID_HIP)
+  size_t free_bytes  = 0;
+  size_t total_bytes = 0;
+  hipError_t err = hipMemGetInfo(&free_bytes, &total_bytes);
+  if (err != hipSuccess) return -1;
+  return (long)((total_bytes - free_bytes) / (1024 * 1024));
+ 
+#else
+  return -1;   // CPU-only build: no GPU to query
+#endif
+}
+
+// ============================================================
+//  Convenience struct — one snapshot of both sides
+// ============================================================
+struct MemSnapshot {
+  long cpu_rss_kb;   // host RSS in kB  (-1 if unavailable)
+  long gpu_used_mb;  // device used in MB (-1 if no GPU)
+};
+ 
+static MemSnapshot takeSnapshot()
+{
+  MemSnapshot s;
+  s.cpu_rss_kb  = getCPURSSKb();
+  s.gpu_used_mb = getGPUUsedMb();
+  return s;
+}
+ 
+// ============================================================
+//  Pretty-print one row of the monitoring table
+// ============================================================
+static void printRow(int iter,
+                     const MemSnapshot &now,
+                     const MemSnapshot &prev)
+{
+  long cpu_delta = (now.cpu_rss_kb  >= 0 && prev.cpu_rss_kb  >= 0)
+                 ? now.cpu_rss_kb  - prev.cpu_rss_kb  : 0;
+  long gpu_delta = (now.gpu_used_mb >= 0 && prev.gpu_used_mb >= 0)
+                 ? now.gpu_used_mb - prev.gpu_used_mb : 0;
+ 
+  // Sign prefix so deltas are unambiguous
+  auto sign = [](long v) -> const char* { return v >= 0 ? "+" : ""; };
+ 
+  std::cout << GridLogMessage
+            << std::setw(6)  << iter
+            << "  CPU: " << std::setw(10) << now.cpu_rss_kb  << " kB"
+            << "  (" << sign(cpu_delta) << std::setw(7) << cpu_delta << " kB)"
+            << "  GPU: " << std::setw(7)  << now.gpu_used_mb << " MB"
+            << "  (" << sign(gpu_delta) << std::setw(5) << gpu_delta << " MB)"
+            << "\n";
+}
+
+// ============================================================
+
+int main(int argc, char **argv)
+{
+  Grid_init(&argc, &argv);
+
+  int threads = GridThread::GetThreads();
+  std::cout << GridLogMessage
+            << "Grid is setup to use " << threads << " threads" << std::endl;
+
+  // ------------------------------------------------------------------
+  // Grid setup — use whatever lattice/mpi/simd was passed on the CLI,
+  // e.g.  --grid 8.8.8.8 --mpi 1.1.1.1
+  // ------------------------------------------------------------------
+  Coordinate latt_size   = GridDefaultLatt();
+  Coordinate simd_layout = GridDefaultSimd(Nd, vComplexD::Nsimd());
+  Coordinate mpi_layout  = GridDefaultMpi();
+
+  GridCartesian GRID(latt_size, simd_layout, mpi_layout);
+
+  int vol = 1;
+  for (int d = 0; d < (int)latt_size.size(); d++) vol *= latt_size[d];
+
+  std::cout << GridLogMessage << "Lattice : ";
+  for (int d = 0; d < Nd; d++) std::cout << latt_size[d] << " ";
+  std::cout << std::endl;
+
+  // ------------------------------------------------------------------
+  // Propagator field: SpinColourMatrix = 12x12 complex, i.e.
+  // LatticePropagatorD (= Lattice<iSpinColourMatrix<vComplexD>>).
+  // This is the standard QCD quark propagator type.
+  // ------------------------------------------------------------------
+  LatticePropagatorD prop(&GRID);
+  
+  // ------------------------------------------------------------------
+  // Fill the propagator with a momentum-space plane wave,
+  // following the pattern from Test_fft.cc.
+  // We set each spin-colour component (a,b) to exp(i * sum_mu p_mu x_mu)
+  // with a fixed momentum p = (1,2,1,2).
+  // ------------------------------------------------------------------
+  Coordinate pvec({1, 2, 1, 2});
+
+  LatticeComplexD phase(&GRID);
+  LatticeComplexD coor(&GRID);
+  ComplexD ci(0.0, 1.0);
+
+  phase = Zero();
+  for (int mu = 0; mu < Nd; mu++) {
+    RealD TwoPiL = M_PI * 2.0 / latt_size[mu];
+    LatticeCoordinate(coor, mu);
+    phase = phase + (TwoPiL * pvec[mu]) * coor;
+  }
+  phase = exp(phase * ci);   // e^{i p.x}
+
+  // Broadcast the phase into every spin-colour matrix entry
+  prop = Zero();
+  prop = prop + phase;
+
+  std::cout << GridLogMessage
+            << "Propagator norm2 = " << norm2(prop) << std::endl;
+
+  // ------------------------------------------------------------------
+  // Baseline snapshot BEFORE PlannedFFT construction
+  // ------------------------------------------------------------------
+  MemSnapshot snap_before_plan = takeSnapshot();
+  std::cout << GridLogMessage
+            << "[mem] Before PlannedFFT construction"
+            << "  CPU: " << snap_before_plan.cpu_rss_kb  << " kB"
+            << "  GPU: " << snap_before_plan.gpu_used_mb << " MB"
+            << std::endl;
+ 
+  // ------------------------------------------------------------------
+  // Create the PlannedFFT — plans are allocated here ONCE for all
+  // dimensions and stored inside the object.
+  // ------------------------------------------------------------------
+  PlannedFFT<iSpinColourMatrix<vComplexD>> plannedFFT(&GRID);
+
+  // ------------------------------------------------------------------
+  // Snapshot AFTER plan construction — this is the true baseline
+  // for the loop, because cufftPlanMany itself grabs device memory.
+  // ------------------------------------------------------------------
+  MemSnapshot snap_after_plan = takeSnapshot();
+  std::cout << GridLogMessage
+            << "[mem] After  PlannedFFT construction"
+            << "  CPU: " << snap_after_plan.cpu_rss_kb  << " kB"
+            << "  GPU: " << snap_after_plan.gpu_used_mb << " MB"
+            << "  (plan overhead:"
+            << "  CPU +" << snap_after_plan.cpu_rss_kb - snap_before_plan.cpu_rss_kb << " kB"
+            << "  GPU +" << snap_after_plan.gpu_used_mb - snap_before_plan.gpu_used_mb << " MB)"
+            << std::endl;
+
+  MemoryManager::Print();
+  // ------------------------------------------------------------------
+  // 400-iteration loop.
+  // Each iteration computes the full 4d forward FFT of `prop`.
+  // We deliberately do NOT cache the result — we always start from
+  // the same `prop` so the FFT is recomputed identically each time.
+  // The point is to watch memory, not correctness.
+  // ------------------------------------------------------------------
+  const int Niter = 40;
+  const int Niter2 = 32;
+
+  // Print header for the memory table
+  std::cout << GridLogMessage
+              << "\n"
+              << std::setw(6)  << "iter"
+              << "  CPU: " << std::setw(10) << "RSS[kB]"
+              << "  (  delta  )"
+              << "  GPU: " << std::setw(7)  << "used[MB]"
+              << "  (delta)"
+              << "\n";
+ 
+  MemSnapshot snap_prev = snap_after_plan;
+
+  for (int i = 0; i < Niter; i++) {
+    std::vector<LatticePropagatorD> G;
+
+    for (int j = 0; j < Niter2; j++) {
+      LatticePropagatorD prop_fft(&GRID);
+
+      // Full 4d forward FFT using the pre-built plans
+      plannedFFT.FFT_all_dim(prop_fft, prop, FFT::forward);
+
+      G.push_back(prop_fft);
+    }
+ 
+    // cudaMemGetInfo reflects the state *after* any pooled frees have
+    // been committed, so this is accurate without an explicit sync —
+    // FFT_dim_execute already calls accelerator_barrier() internally.
+    MemSnapshot snap_now = takeSnapshot();
+    printRow(i, snap_now, snap_prev);
+    MemoryManager::Print();
+    snap_prev = snap_now;
+  }
+
+  // ------------------------------------------------------------------
+  // Summary
+  // ------------------------------------------------------------------
+  MemSnapshot snap_final = takeSnapshot();
+ 
+  long cpu_growth = snap_final.cpu_rss_kb  - snap_after_plan.cpu_rss_kb;
+  long gpu_growth = snap_final.gpu_used_mb - snap_after_plan.gpu_used_mb;
+ 
+  std::cout << GridLogMessage
+            << "\n==== Memory summary (baseline = after plan construction) ====\n"
+            << "  CPU RSS growth over " << Niter << " FFTs : "
+            << cpu_growth << " kB"
+            << (cpu_growth == 0 ? "  OK" : "  *** GROWING ***") << "\n"
+            << "  GPU used growth over " << Niter << " FFTs : "
+            << gpu_growth << " MB"
+            << (gpu_growth == 0 ? "  OK" : "  *** GROWING ***") << "\n"
+            << "  Note: first-call watermark from pool fill is expected and benign.\n"
+            << "        A leak shows as continuous growth beyond iter ~2-3.\n";
+ 
+  Grid_finalize();
+  return 0;
+}

tests/debug/Test_hipfft_bug_fail.cc

@@ -0,0 +1,76 @@
+/*
+ * Isolating the hipfft HIPFFT_PARSE_ERROR on ROCm 7 / hipFFT 1.0.20.
+ *
+ * Tests three orderings with an empty rocFFT cache to find which GPU
+ * operation before plan creation triggers the failure:
+ *   A) hipMalloc only            — hypothesis: passes (no async GPU work)
+ *   B) hipMalloc + hipMemset     — hypothesis: fails  (async GPU work in flight)
+ *   C) hipMalloc + hipMemset     — hypothesis: passes (work completed before plan)
+ *      + hipDeviceSynchronize
+ *
+ * Compile:
+ *   hipcc -o Test_hipfft_bug_fail Test_hipfft_bug_fail.cc -lhipfft
+ *
+ * Run with empty cache:
+ *   rm -rf ~/.cache/
+ *   ./Test_hipfft_bug_fail
+ */
+
+#include <cstdio>
+#include <hipfft/hipfft.h>
+#include <hip/hip_runtime.h>
+
+static const char *res(hipfftResult rv) {
+  return rv == HIPFFT_SUCCESS ? "SUCCESS" : "PARSE_ERROR";
+}
+
+static hipfftResult makePlan(int G, int howmany) {
+  int n[] = {G};
+  hipfftHandle p;
+  size_t workSize = 0;
+  hipfftCreate(&p);
+  hipfftResult rv = hipfftMakePlanMany(p, 1, n,
+                                       nullptr, 1, G, nullptr, 1, G,
+                                       HIPFFT_Z2Z, howmany, &workSize);
+  hipfftDestroy(p);
+  return rv;
+}
+
+int main(void) {
+  hipDeviceProp_t prop;
+  hipGetDeviceProperties(&prop, 0);
+  printf("Device: %s\n", prop.name);
+#ifdef hipfftVersionMinor
+  printf("hipFFT version: %d.%d.%d\n\n",
+         hipfftVersionMajor, hipfftVersionMinor, hipfftVersionPatch);
+#endif
+
+  for (int G : {4, 8, 16, 32}) {
+    int howmany = 512;
+    long nelems = (long)G * howmany;
+    hipfftDoubleComplex *buf = nullptr;
+    hipMalloc(&buf, nelems * sizeof(hipfftDoubleComplex));
+
+    // Tests ordered so each runs before a prior success can populate the cache.
+
+    // B first: hipMalloc + hipMemset (async GPU work in flight)
+    // If this fails, A (no hipMemset) will pass, confirming hipMemset is the trigger.
+    hipMemset(buf, 0, nelems * sizeof(hipfftDoubleComplex));
+    hipfftResult rvB = makePlan(G, howmany);
+    printf("G=%-4d  B) hipMalloc + hipMemset       : %s\n", G, res(rvB));
+
+    // C: hipMalloc + hipMemset + sync — does syncing before plan creation fix it?
+    hipMemset(buf, 0, nelems * sizeof(hipfftDoubleComplex));
+    hipDeviceSynchronize();
+    hipfftResult rvC = makePlan(G, howmany);
+    printf("G=%-4d  C) hipMalloc + hipMemset + sync: %s\n", G, res(rvC));
+
+    // A last: hipMalloc only, no async GPU work — should always pass
+    hipfftResult rvA = makePlan(G, howmany);
+    printf("G=%-4d  A) hipMalloc only             : %s\n\n", G, res(rvA));
+
+    hipFree(buf);
+  }
+
+  return 0;
+}

tests/debug/Test_hipfft_bug_pass.cc

@@ -0,0 +1,61 @@
+/*
+ * Minimal program demonstrating the workaround for the hipfft ROCm 7 bug.
+ *
+ * Workaround: create the hipfft plan BEFORE any hipMalloc.  Plan creation
+ * for G < 32 then succeeds even with an empty rocFFT cache.
+ *
+ * Compile:
+ *   hipcc -o Test_hipfft_bug_pass Test_hipfft_bug_pass.cc -lhipfft
+ *
+ * Run:
+ *   rm -rf ~/.cache/rocfft
+ *   ./Test_hipfft_bug_pass
+ *
+ * Expected: all G values succeed.
+ * Compare with Test_hipfft_bug_fail.cc which uses the opposite ordering.
+ */
+
+#include <cstdio>
+#include <hipfft/hipfft.h>
+#include <hip/hip_runtime.h>
+
+int main(void) {
+  hipDeviceProp_t prop;
+  hipGetDeviceProperties(&prop, 0);
+  printf("Device: %s\n", prop.name);
+#ifdef hipfftVersionMinor
+  printf("hipFFT version: %d.%d.%d\n\n",
+         hipfftVersionMajor, hipfftVersionMinor, hipfftVersionPatch);
+#endif
+
+  for (int G : {8, 16, 32}) {
+    int howmany = 512;
+    int n[] = {G};
+    long nelems = (long)G * howmany;
+
+    // Plan created BEFORE hipMalloc — succeeds for all G
+    hipfftHandle p;
+    size_t workSize = 0;
+    hipfftCreate(&p);
+    hipfftResult rv = hipfftMakePlanMany(p, 1, n,
+                                         nullptr, 1, G, nullptr, 1, G,
+                                         HIPFFT_Z2Z, howmany, &workSize);
+    printf("G=%-4d  plan-then-hipMalloc: %d (%s)\n",
+           G, (int)rv, rv == HIPFFT_SUCCESS ? "HIPFFT_SUCCESS" : "HIPFFT_PARSE_ERROR");
+
+    if (rv == HIPFFT_SUCCESS) {
+      hipfftDoubleComplex *buf = nullptr;
+      hipMalloc(&buf, nelems * sizeof(hipfftDoubleComplex));
+      hipMemset(buf, 0, nelems * sizeof(hipfftDoubleComplex));
+      rv = hipfftExecZ2Z(p, buf, buf, HIPFFT_FORWARD);
+      hipDeviceSynchronize();
+      printf("G=%-4d  execFwd:             %d (%s)\n",
+             G, (int)rv, rv == HIPFFT_SUCCESS ? "HIPFFT_SUCCESS" : "FAILED");
+      hipFree(buf);
+    }
+
+    hipfftDestroy(p);
+  }
+
+  return 0;
+}

tests/debug/Test_hipfft_minimal.cc

@@ -0,0 +1,142 @@
+/*
+ * Minimal reproducer for hipfftMakePlanMany / hipfftPlanMany failures.
+ *
+ * Compile on Frontier (no Grid headers needed):
+ *   hipcc -o Test_hipfft_minimal Test_hipfft_minimal.cc -lhipfft
+ *
+ * Run:
+ *   ./Test_hipfft_minimal
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <hipfft/hipfft.h>
+#include <hip/hip_runtime.h>
+
+static const char *hipfftResultString(hipfftResult r) {
+  switch (r) {
+  case HIPFFT_SUCCESS:                  return "HIPFFT_SUCCESS";
+  case HIPFFT_INVALID_PLAN:             return "HIPFFT_INVALID_PLAN";
+  case HIPFFT_ALLOC_FAILED:             return "HIPFFT_ALLOC_FAILED";
+  case HIPFFT_INVALID_TYPE:             return "HIPFFT_INVALID_TYPE";
+  case HIPFFT_INVALID_VALUE:            return "HIPFFT_INVALID_VALUE";
+  case HIPFFT_INTERNAL_ERROR:           return "HIPFFT_INTERNAL_ERROR";
+  case HIPFFT_EXEC_FAILED:              return "HIPFFT_EXEC_FAILED";
+  case HIPFFT_SETUP_FAILED:             return "HIPFFT_SETUP_FAILED";
+  case HIPFFT_INVALID_SIZE:             return "HIPFFT_INVALID_SIZE";
+  case HIPFFT_UNALIGNED_DATA:           return "HIPFFT_UNALIGNED_DATA";
+  case HIPFFT_INCOMPLETE_PARAMETER_LIST:return "HIPFFT_INCOMPLETE_PARAMETER_LIST";
+  case HIPFFT_INVALID_DEVICE:           return "HIPFFT_INVALID_DEVICE";
+  case HIPFFT_PARSE_ERROR:              return "HIPFFT_PARSE_ERROR";
+  case HIPFFT_NO_WORKSPACE:             return "HIPFFT_NO_WORKSPACE";
+  case HIPFFT_NOT_IMPLEMENTED:          return "HIPFFT_NOT_IMPLEMENTED";
+  case HIPFFT_NOT_SUPPORTED:            return "HIPFFT_NOT_SUPPORTED";
+  default:                              return "UNKNOWN";
+  }
+}
+
+// Plan creation + execution for (G, howmany).
+// Tests two orderings to isolate whether a prior hipMalloc poisons hipfft
+// plan creation for small G on ROCm 7:
+//   A) plan BEFORE hipMalloc  — hypothesis: succeeds
+//   B) hipMalloc BEFORE plan  — hypothesis: fails for G < 32
+static void tryPlanAndExec(int G, long howmany) {
+  int n[] = {G};
+  long nelems = (long)G * howmany;
+
+  printf("--- G=%-4d  howmany=%-10ld  total_elems=%-12ld ---\n",
+         G, howmany, nelems);
+
+  // --- A: create plan first, allocate buffer afterwards ---
+  {
+    hipfftHandle p;
+    size_t workSize = 0;
+    hipfftCreate(&p);
+    hipfftResult rv = hipfftMakePlanMany(p, 1, n,
+                                         nullptr, 1, G, nullptr, 1, G,
+                                         HIPFFT_Z2Z, (int)howmany, &workSize);
+    printf("  plan-first  create : %d (%s)\n", (int)rv, hipfftResultString(rv));
+    if (rv == HIPFFT_SUCCESS) {
+      hipfftDoubleComplex *buf = nullptr;
+      hipMalloc(&buf, nelems * sizeof(hipfftDoubleComplex));
+      hipMemset(buf, 0, nelems * sizeof(hipfftDoubleComplex));
+      rv = hipfftExecZ2Z(p, buf, buf, HIPFFT_FORWARD);
+      hipDeviceSynchronize();
+      printf("  plan-first  execFwd: %d (%s)\n", (int)rv, hipfftResultString(rv));
+      hipFree(buf);
+    }
+    hipfftDestroy(p);
+  }
+
+  // --- B: hipMalloc first, create plan afterwards ---
+  {
+    hipfftDoubleComplex *buf = nullptr;
+    hipMalloc(&buf, nelems * sizeof(hipfftDoubleComplex));
+    hipMemset(buf, 0, nelems * sizeof(hipfftDoubleComplex));
+
+    hipfftHandle p;
+    size_t workSize = 0;
+    hipfftCreate(&p);
+    hipfftResult rv = hipfftMakePlanMany(p, 1, n,
+                                         nullptr, 1, G, nullptr, 1, G,
+                                         HIPFFT_Z2Z, (int)howmany, &workSize);
+    printf("  malloc-first create : %d (%s)\n", (int)rv, hipfftResultString(rv));
+    if (rv == HIPFFT_SUCCESS) {
+      rv = hipfftExecZ2Z(p, buf, buf, HIPFFT_FORWARD);
+      hipDeviceSynchronize();
+      printf("  malloc-first execFwd: %d (%s)\n", (int)rv, hipfftResultString(rv));
+    }
+    hipfftDestroy(p);
+    hipFree(buf);
+  }
+
+  printf("\n");
+}
+
+int main(void) {
+  // Print HIP device info
+  int device = 0;
+  hipGetDevice(&device);
+  hipDeviceProp_t prop;
+  hipGetDeviceProperties(&prop, device);
+  printf("Device %d: %s  warpSize=%d\n\n", device, prop.name, prop.warpSize);
+
+#ifdef hipfftVersionMinor
+  printf("hipFFT version: %d.%d.%d\n\n",
+         hipfftVersionMajor, hipfftVersionMinor, hipfftVersionPatch);
+#endif
+
+  // Original sweep with small howmany (these passed first time)
+  printf("=== Small howmany (original sweep) ===\n\n");
+  for (int G : {4, 8, 12, 16, 24, 32, 48, 64})
+    tryPlanAndExec(G, 512);
+
+  // Grid-realistic howmany values derived from actual lattice geometries.
+  // howmany = Ncomp * product(ldimensions[d] for d != dim)
+  // For LatticeComplexD: Ncomp=1.
+  printf("=== Grid-realistic parameters ===\n\n");
+
+  // --grid 16.16.16.16  4D FFT  (KNOWN TO FAIL in Grid)
+  // Each dim: G=16, Nperp=16^3=4096
+  tryPlanAndExec(16, 4096);
+
+  // --grid 32.32.32.32  4D FFT  (KNOWN TO SUCCEED in Grid)
+  // Each dim: G=32, Nperp=32^3=32768
+  tryPlanAndExec(32, 32768);
+
+  // --grid 32.32.32.32 Ls=8  5D DWF FFT  (KNOWN TO FAIL on dim 0 in Grid)
+  // dim 0: G=8,  Nperp=32^4=1048576
+  tryPlanAndExec(8, 1048576);
+  // dim 1-4: G=32, Nperp=8*32^3=262144
+  tryPlanAndExec(32, 262144);
+
+  // Extra intermediate cases to bracket the failure
+  tryPlanAndExec(16, 1024);
+  tryPlanAndExec(16, 2048);
+  tryPlanAndExec(16, 8192);
+  tryPlanAndExec(8,  4096);
+  tryPlanAndExec(8,  65536);
+  tryPlanAndExec(8,  262144);
+
+  return 0;
+}

tests/debug/Test_hipfft_repro.cc

@@ -0,0 +1,168 @@
+/*
+ * Reproducer for HIPFFT_PARSE_ERROR (error 12) from hipfftMakePlanMany on
+ * ROCm 7 / hipFFT 1.0.20 (Frontier, MI210 login and MI250X compute nodes).
+ *
+ * Observed failure: G < 32 returns HIPFFT_PARSE_ERROR from all three plan
+ * creation APIs (hipfftPlanMany, hipfftMakePlanMany, hipfftPlan1d) when a
+ * device buffer is allocated and zeroed with hipMalloc+hipMemset before the
+ * plan creation call.  G >= 32 succeeds.
+ *
+ * Contrast with Test_hipfft_minimal.cc (plan-first ordering) which passes
+ * for all G even with an empty rocFFT cache.
+ *
+ * Compile on Frontier (no Grid headers needed):
+ *   hipcc -o Test_hipfft_repro Test_hipfft_repro.cc -lhipfft
+ *
+ * Run with empty cache to reproduce the failure:
+ *   rm -rf ~/.cache/rocfft
+ *   ./Test_hipfft_repro
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <hipfft/hipfft.h>
+#include <hip/hip_runtime.h>
+
+static const char *hipfftResultString(hipfftResult r) {
+  switch (r) {
+  case HIPFFT_SUCCESS:                  return "HIPFFT_SUCCESS";
+  case HIPFFT_INVALID_PLAN:             return "HIPFFT_INVALID_PLAN";
+  case HIPFFT_ALLOC_FAILED:             return "HIPFFT_ALLOC_FAILED";
+  case HIPFFT_INVALID_TYPE:             return "HIPFFT_INVALID_TYPE";
+  case HIPFFT_INVALID_VALUE:            return "HIPFFT_INVALID_VALUE";
+  case HIPFFT_INTERNAL_ERROR:           return "HIPFFT_INTERNAL_ERROR";
+  case HIPFFT_EXEC_FAILED:              return "HIPFFT_EXEC_FAILED";
+  case HIPFFT_SETUP_FAILED:             return "HIPFFT_SETUP_FAILED";
+  case HIPFFT_INVALID_SIZE:             return "HIPFFT_INVALID_SIZE";
+  case HIPFFT_UNALIGNED_DATA:           return "HIPFFT_UNALIGNED_DATA";
+  case HIPFFT_INCOMPLETE_PARAMETER_LIST:return "HIPFFT_INCOMPLETE_PARAMETER_LIST";
+  case HIPFFT_INVALID_DEVICE:           return "HIPFFT_INVALID_DEVICE";
+  case HIPFFT_PARSE_ERROR:              return "HIPFFT_PARSE_ERROR";
+  case HIPFFT_NO_WORKSPACE:             return "HIPFFT_NO_WORKSPACE";
+  case HIPFFT_NOT_IMPLEMENTED:          return "HIPFFT_NOT_IMPLEMENTED";
+  case HIPFFT_NOT_SUPPORTED:            return "HIPFFT_NOT_SUPPORTED";
+  default:                              return "UNKNOWN";
+  }
+}
+
+// Plan creation + execution for (G, howmany) using hipfftCreate+hipfftMakePlanMany.
+// This is the path Grid's FFT.h now uses.
+static void tryPlanAndExec(int G, long howmany) {
+  int n[] = {G};
+  long nelems = (long)G * howmany;
+
+  printf("--- G=%-4d  howmany=%-10ld  total_elems=%-12ld ---\n",
+         G, howmany, nelems);
+
+  // Allocate device buffer (hipfftDoubleComplex = 16 bytes each)
+  hipfftDoubleComplex *dbuf = nullptr;
+  hipError_t herr = hipMalloc(&dbuf, nelems * sizeof(hipfftDoubleComplex));
+  if (herr != hipSuccess) {
+    printf("  hipMalloc failed (%d) for %ld elems — skipping\n\n", (int)herr, nelems);
+    return;
+  }
+  hipMemset(dbuf, 0, nelems * sizeof(hipfftDoubleComplex));
+
+  // 1. hipfftPlanMany (one-step, nullptr embed) — current Grid path
+  {
+    hipfftHandle p;
+    hipfftResult rv = hipfftPlanMany(&p, 1, n,
+                                     nullptr, 1, G,
+                                     nullptr, 1, G,
+                                     HIPFFT_Z2Z, (int)howmany);
+    printf("  hipfftPlanMany   create : %d (%s)\n", (int)rv, hipfftResultString(rv));
+    if (rv == HIPFFT_SUCCESS) {
+      rv = hipfftExecZ2Z(p, dbuf, dbuf, HIPFFT_FORWARD);
+      hipDeviceSynchronize();
+      printf("  hipfftPlanMany   execFwd: %d (%s)\n", (int)rv, hipfftResultString(rv));
+      hipfftDestroy(p);
+    }
+  }
+
+  // 2. hipfftCreate + hipfftMakePlanMany (two-step) — also current Grid path
+  {
+    hipfftHandle p;
+    size_t workSize = 0;
+    hipfftResult rc = hipfftCreate(&p);
+    if (rc == HIPFFT_SUCCESS) {
+      hipfftResult rv = hipfftMakePlanMany(p, 1, n,
+                                           nullptr, 1, G,
+                                           nullptr, 1, G,
+                                           HIPFFT_Z2Z, (int)howmany, &workSize);
+      printf("  hipfftMakePlanMany      : %d (%s)  workSize=%zu\n",
+             (int)rv, hipfftResultString(rv), workSize);
+      if (rv == HIPFFT_SUCCESS) {
+        rv = hipfftExecZ2Z(p, dbuf, dbuf, HIPFFT_FORWARD);
+        hipDeviceSynchronize();
+        printf("  hipfftMakePlanMany exec : %d (%s)\n", (int)rv, hipfftResultString(rv));
+      }
+      hipfftDestroy(p);
+    } else {
+      printf("  hipfftCreate            : %d (%s)\n", (int)rc, hipfftResultString(rc));
+    }
+  }
+
+  // 3. hipfftPlan1d (simplest API, batch = howmany)
+  {
+    hipfftHandle p;
+    hipfftResult rv = hipfftPlan1d(&p, G, HIPFFT_Z2Z, (int)howmany);
+    printf("  hipfftPlan1d     create : %d (%s)\n", (int)rv, hipfftResultString(rv));
+    if (rv == HIPFFT_SUCCESS) {
+      rv = hipfftExecZ2Z(p, dbuf, dbuf, HIPFFT_FORWARD);
+      hipDeviceSynchronize();
+      printf("  hipfftPlan1d     execFwd: %d (%s)\n", (int)rv, hipfftResultString(rv));
+      hipfftDestroy(p);
+    }
+  }
+
+  hipFree(dbuf);
+  printf("\n");
+}
+
+int main(void) {
+  // Print HIP device info
+  int device = 0;
+  hipGetDevice(&device);
+  hipDeviceProp_t prop;
+  hipGetDeviceProperties(&prop, device);
+  printf("Device %d: %s  warpSize=%d\n\n", device, prop.name, prop.warpSize);
+
+#ifdef hipfftVersionMinor
+  printf("hipFFT version: %d.%d.%d\n\n",
+         hipfftVersionMajor, hipfftVersionMinor, hipfftVersionPatch);
+#endif
+
+  // Original sweep with small howmany (these passed first time)
+  printf("=== Small howmany (original sweep) ===\n\n");
+  for (int G : {4, 8, 12, 16, 24, 32, 48, 64})
+    tryPlanAndExec(G, 512);
+
+  // Grid-realistic howmany values derived from actual lattice geometries.
+  // howmany = Ncomp * product(ldimensions[d] for d != dim)
+  // For LatticeComplexD: Ncomp=1.
+  printf("=== Grid-realistic parameters ===\n\n");
+
+  // --grid 16.16.16.16  4D FFT  (KNOWN TO FAIL in Grid)
+  // Each dim: G=16, Nperp=16^3=4096
+  tryPlanAndExec(16, 4096);
+
+  // --grid 32.32.32.32  4D FFT  (KNOWN TO SUCCEED in Grid)
+  // Each dim: G=32, Nperp=32^3=32768
+  tryPlanAndExec(32, 32768);
+
+  // --grid 32.32.32.32 Ls=8  5D DWF FFT  (KNOWN TO FAIL on dim 0 in Grid)
+  // dim 0: G=8,  Nperp=32^4=1048576
+  tryPlanAndExec(8, 1048576);
+  // dim 1-4: G=32, Nperp=8*32^3=262144
+  tryPlanAndExec(32, 262144);
+
+  // Extra intermediate cases to bracket the failure
+  tryPlanAndExec(16, 1024);
+  tryPlanAndExec(16, 2048);
+  tryPlanAndExec(16, 8192);
+  tryPlanAndExec(8,  4096);
+  tryPlanAndExec(8,  65536);
+  tryPlanAndExec(8,  262144);
+
+  return 0;
+}