portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2026-07-18 08:03:27 +01:00
Code Releases Activity

tests/debug: extend hipfft reproducer with Grid-realistic howmany and exec tests

Browse Source 
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Peter Boyle
2026-05-21 12:34:30 -04:00
co-authored by Claude Sonnet 4.6
parent 303a4d26e5
commit 74e0f846cb
1 changed files with 85 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/debug/Test_hipfft_minimal.cc
+85 -49
View File
@@ -35,59 +35,77 @@ static const char *hipfftResultString(hipfftResult r) {
} }
} }
// Try both hipfftPlanMany and hipfftCreate+hipfftMakePlanMany for given G and howmany. // Plan creation + execution for (G, howmany) using hipfftCreate+hipfftMakePlanMany.
static void tryPlans(int G, int howmany) { // This is the path Grid's FFT.h now uses.
static void tryPlanAndExec(int G, long howmany) {
int n[] = {G}; int n[] = {G};
long nelems = (long)G * howmany;
printf("--- G=%-4d howmany=%-6d ---\n", G, howmany); printf("--- G=%-4d howmany=%-10ld total_elems=%-12ld ---\n",
G, howmany, nelems);
// 1. hipfftPlanMany (one-step) // 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; hipfftHandle p;
hipfftResult rv = hipfftPlanMany(&p, 1, n, hipfftResult rv = hipfftPlanMany(&p, 1, n,
nullptr, 1, G, nullptr, 1, G,
nullptr, 1, G, nullptr, 1, G,
HIPFFT_Z2Z, howmany); HIPFFT_Z2Z, (int)howmany);
printf(" hipfftPlanMany : %d (%s)\n", (int)rv, hipfftResultString(rv)); printf(" hipfftPlanMany create : %d (%s)\n", (int)rv, hipfftResultString(rv));
if (rv == HIPFFT_SUCCESS) hipfftDestroy(p); if (rv == HIPFFT_SUCCESS) {
} rv = hipfftExecZ2Z(p, dbuf, dbuf, HIPFFT_FORWARD);
hipDeviceSynchronize();
// 2. hipfftPlanMany with inembed=n (old Grid behaviour) printf(" hipfftPlanMany execFwd: %d (%s)\n", (int)rv, hipfftResultString(rv));
{
hipfftHandle p;
hipfftResult rv = hipfftPlanMany(&p, 1, n,
n, 1, G,
n, 1, G,
HIPFFT_Z2Z, howmany);
printf(" hipfftPlanMany(inembed=n): %d (%s)\n", (int)rv, hipfftResultString(rv));
if (rv == HIPFFT_SUCCESS) hipfftDestroy(p);
}
// 3. hipfftCreate + hipfftMakePlanMany (two-step, nullptr embed)
{
hipfftHandle p;
size_t workSize = 0;
hipfftResult rc = hipfftCreate(&p);
printf(" hipfftCreate : %d (%s)\n", (int)rc, hipfftResultString(rc));
if (rc == HIPFFT_SUCCESS) {
hipfftResult rv = hipfftMakePlanMany(p, 1, n,
nullptr, 1, G,
nullptr, 1, G,
HIPFFT_Z2Z, howmany, &workSize);
printf(" hipfftMakePlanMany : %d (%s) workSize=%zu\n",
(int)rv, hipfftResultString(rv), workSize);
hipfftDestroy(p); hipfftDestroy(p);
} }
} }
// 4. hipfftPlan1d (simplest API) // 2. hipfftCreate + hipfftMakePlanMany (two-step) — also current Grid path
{ {
hipfftHandle p; hipfftHandle p;
hipfftResult rv = hipfftPlan1d(&p, G, HIPFFT_Z2Z, howmany); size_t workSize = 0;
printf(" hipfftPlan1d : %d (%s)\n", (int)rv, hipfftResultString(rv)); hipfftResult rc = hipfftCreate(&p);
if (rv == HIPFFT_SUCCESS) hipfftDestroy(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"); printf("\n");
} }
@@ -99,24 +117,42 @@ int main(void) {
hipGetDeviceProperties(&prop, device); hipGetDeviceProperties(&prop, device);
printf("Device %d: %s warpSize=%d\n\n", device, prop.name, prop.warpSize); printf("Device %d: %s warpSize=%d\n\n", device, prop.name, prop.warpSize);
// Print hipFFT version if available
#ifdef hipfftVersionMinor #ifdef hipfftVersionMinor
printf("hipFFT version: %d.%d.%d\n\n", printf("hipFFT version: %d.%d.%d\n\n",
hipfftVersionMajor, hipfftVersionMinor, hipfftVersionPatch); hipfftVersionMajor, hipfftVersionMinor, hipfftVersionPatch);
#endif #endif
// Sweep over transform sizes G with a fixed representative howmany // Original sweep with small howmany (these passed first time)
// (512 = typical Nperp*Ncomp for a small lattice) printf("=== Small howmany (original sweep) ===\n\n");
const int howmany = 512; for (int G : {4, 8, 12, 16, 24, 32, 48, 64})
for (int G : {4, 8, 12, 16, 24, 32, 48, 64}) { tryPlanAndExec(G, 512);
tryPlans(G, howmany);
}
// Also try the exact parameters from the failing Grid FFT // Grid-realistic howmany values derived from actual lattice geometries.
printf("=== Grid-specific parameters ===\n\n"); // howmany = Ncomp * product(ldimensions[d] for d != dim)
tryPlans(8, 512); // Ls=8, small 4D lattice // For LatticeComplexD: Ncomp=1.
tryPlans(16, 512); // 16^4 printf("=== Grid-realistic parameters ===\n\n");
tryPlans(32, 512); // 32^4 (known to work)
// --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; return 0;
} }