void test_correctness(int n)
{
int howmany;
fftw_plan validated_plan_forward, validated_plan_backward;
WHEN_VERBOSE(1,
my_printf("Testing correctness for n = %d...", n);
my_fflush(stdout));
/* produce a good plan */
validated_plan_forward =
fftw_create_plan(n, FFTW_FORWARD, measure_flag | wisdom_flag);
validated_plan_backward =
fftw_create_plan(n, FFTW_BACKWARD, measure_flag | wisdom_flag);
for (howmany = 1; howmany <= MAX_HOWMANY; ++howmany)
test_in_place_both(n, howmany, howmany,
validated_plan_forward,
validated_plan_backward);
fftw_destroy_plan(validated_plan_forward);
fftw_destroy_plan(validated_plan_backward);
if (!(wisdom_flag & FFTW_USE_WISDOM) && chk_mem_leak)
fftw_check_memory_leaks();
WHEN_VERBOSE(1, my_printf("OK\n"));
}
void test_speed_nd_aux(struct size sz,
fftw_direction dir, int flags, int specific)
{
fftw_complex *in;
fftwnd_plan plan;
double t;
fftw_time begin, end;
int i, N;
/* only bench in-place multi-dim transforms */
flags |= FFTW_IN_PLACE;
N = 1;
for (i = 0; i < sz.rank; ++i)
N *= (sz.narray[i]);
in = (fftw_complex *) fftw_malloc(N * howmany_fields *
sizeof(fftw_complex));
if (specific) {
begin = fftw_get_time();
plan = fftwnd_create_plan_specific(sz.rank, sz.narray, dir,
speed_flag | flags
| wisdom_flag | no_vector_flag,
in, howmany_fields, 0, 1);
} else {
begin = fftw_get_time();
plan = fftwnd_create_plan(sz.rank, sz.narray,
dir, speed_flag | flags
| wisdom_flag | no_vector_flag);
}
end = fftw_get_time();
CHECK(plan != NULL, "can't create plan");
t = fftw_time_to_sec(fftw_time_diff(end, begin));
WHEN_VERBOSE(2, printf("time for planner: %f s\n", t));
WHEN_VERBOSE(2, printf("\n"));
WHEN_VERBOSE(2, (fftwnd_print_plan(plan)));
WHEN_VERBOSE(2, printf("\n"));
FFTW_TIME_FFT(fftwnd(plan, howmany_fields,
in, howmany_fields, 1, 0, 0, 0),
in, N * howmany_fields, t);
fftwnd_destroy_plan(plan);
WHEN_VERBOSE(1, printf("time for one fft: %s", smart_sprint_time(t)));
WHEN_VERBOSE(1, printf(" (%s/point)\n", smart_sprint_time(t / N)));
WHEN_VERBOSE(1, printf("\"mflops\" = 5 (N log2 N) / (t in microseconds)"
" = %f\n", howmany_fields * mflops(t, N)));
fftw_free(in);
WHEN_VERBOSE(1, printf("\n"));
}
void test_out_of_place(int n, int istride, int ostride,
int howmany, fftw_direction dir,
fftw_plan validated_plan,
int specific)
{
/* one-dim. out-of-place transforms will never be supported in MPI */
WHEN_VERBOSE(2, my_printf("N/A\n"));
}
void test_speed_aux(int n, fftw_direction dir, int flags, int specific)
{
fftw_complex *in, *out;
fftw_plan plan;
double t, t0;
fftw_time begin, end;
in = (fftw_complex *) fftw_malloc(n * howmany_fields
* sizeof(fftw_complex));
out = (fftw_complex *) fftw_malloc(n * howmany_fields
* sizeof(fftw_complex));
if (specific) {
begin = fftw_get_time();
plan = fftw_create_plan_specific(n, dir,
speed_flag | flags
| wisdom_flag | no_vector_flag,
in, howmany_fields,
out, howmany_fields);
end = fftw_get_time();
} else {
begin = fftw_get_time();
plan = fftw_create_plan(n, dir, speed_flag | flags
| wisdom_flag | no_vector_flag);
end = fftw_get_time();
}
CHECK(plan != NULL, "can't create plan");
t = fftw_time_to_sec(fftw_time_diff(end, begin));
WHEN_VERBOSE(2, printf("time for planner: %f s\n", t));
WHEN_VERBOSE(2, fftw_print_plan(plan));
FFTW_TIME_FFT(fftw(plan, howmany_fields,
in, howmany_fields, 1, out, howmany_fields, 1),
in, n * howmany_fields, t0);
FFTW_TIME_FFT(fftw_threads(nthreads, plan, howmany_fields,
in, howmany_fields, 1, out, howmany_fields, 1),
in, n * howmany_fields, t);
fftw_destroy_plan(plan);
WHEN_VERBOSE(1, printf("time for one fft (uniprocessor): %s\n", smart_sprint_time(t0)));
WHEN_VERBOSE(1, printf("time for one fft (%d threads): %s", nthreads, smart_sprint_time(t)));
WHEN_VERBOSE(1, printf(" (%s/point)\n", smart_sprint_time(t / n)));
WHEN_VERBOSE(1, printf("\"mflops\" = 5 (n log2 n) / (t in microseconds)"
" = %f\n", howmany_fields * mflops(t, n)));
WHEN_VERBOSE(1, printf("parallel speedup: %f\n", t0 / t));
fftw_free(in);
fftw_free(out);
WHEN_VERBOSE(1, printf("\n"));
}
void test_in_place_both(int n, int istride, int howmany,
fftw_plan validated_plan_forward,
fftw_plan validated_plan_backward)
{
WHEN_VERBOSE(2,
my_printf("TEST CORRECTNESS (in place, FFTW_FORWARD, %s) "
"n = %d istride = %d howmany = %d\n",
SPECIFICP(0),
n, istride, howmany));
test_in_place(n, istride, howmany, FFTW_FORWARD,
validated_plan_forward, 0);
WHEN_VERBOSE(2,
my_printf("TEST CORRECTNESS (in place, FFTW_BACKWARD, %s) "
"n = %d istride = %d howmany = %d\n",
SPECIFICP(0),
n, istride, howmany));
test_in_place(n, istride, howmany, FFTW_BACKWARD,
validated_plan_backward, 0);
}
void test_in_place_both(int n, int istride, int howmany,
fftw_plan validated_plan_forward,
fftw_plan validated_plan_backward)
{
int specific;
for (specific = 0; specific <= 1; ++specific) {
WHEN_VERBOSE(2,
printf("TEST CORRECTNESS (in place, FFTW_FORWARD, %s) "
"n = %d istride = %d howmany = %d\n",
SPECIFICP(specific),
n, istride, howmany));
test_in_place(n, istride, howmany, FFTW_FORWARD,
validated_plan_forward, specific);
WHEN_VERBOSE(2,
printf("TEST CORRECTNESS (in place, FFTW_BACKWARD, %s) "
"n = %d istride = %d howmany = %d\n",
SPECIFICP(specific),
n, istride, howmany));
test_in_place(n, istride, howmany, FFTW_BACKWARD,
validated_plan_backward, specific);
}
}
void test_correctness(int n)
{
int istride, ostride, howmany;
fftw_plan validated_plan_forward, validated_plan_backward;
WHEN_VERBOSE(1,
printf("Testing correctness for n = %d...", n);
fflush(stdout));
/* produce a *good* plan (validated by Ergun's test procedure) */
validated_plan_forward =
fftw_create_plan(n, FFTW_FORWARD, measure_flag | wisdom_flag);
test_ergun(n, FFTW_FORWARD, validated_plan_forward);
validated_plan_backward =
fftw_create_plan(n, FFTW_BACKWARD, measure_flag | wisdom_flag);
test_ergun(n, FFTW_BACKWARD, validated_plan_backward);
for (istride = 1; istride <= MAX_STRIDE; ++istride)
for (ostride = 1; ostride <= MAX_STRIDE; ++ostride)
for (howmany = 1; howmany <= MAX_HOWMANY; ++howmany)
test_out_of_place_both(n, istride, ostride, howmany,
validated_plan_forward,
validated_plan_backward);
for (istride = 1; istride <= MAX_STRIDE; ++istride)
for (howmany = 1; howmany <= MAX_HOWMANY; ++howmany)
test_in_place_both(n, istride, howmany,
validated_plan_forward,
validated_plan_backward);
fftw_destroy_plan(validated_plan_forward);
fftw_destroy_plan(validated_plan_backward);
if (!(wisdom_flag & FFTW_USE_WISDOM) && chk_mem_leak)
fftw_check_memory_leaks();
WHEN_VERBOSE(1, printf("OK\n"));
}
/* Same as test_ergun, but for multi-dimensional transforms: */
void testnd_ergun(int rank, int *n, fftw_direction dir, fftwnd_plan plan)
{
fftw_complex *inA, *inB, *inC, *outA, *outB, *outC;
fftw_complex *tmp;
fftw_complex impulse;
int N, n_before, n_after, dim;
int i, which_impulse;
int rounds = 20;
FFTW_TRIG_REAL twopin;
N = 1;
for (dim = 0; dim < rank; ++dim)
N *= n[dim];
inA = (fftw_complex *) fftw_malloc(N * sizeof(fftw_complex));
inB = (fftw_complex *) fftw_malloc(N * sizeof(fftw_complex));
inC = (fftw_complex *) fftw_malloc(N * sizeof(fftw_complex));
outA = (fftw_complex *) fftw_malloc(N * sizeof(fftw_complex));
outB = (fftw_complex *) fftw_malloc(N * sizeof(fftw_complex));
outC = (fftw_complex *) fftw_malloc(N * sizeof(fftw_complex));
tmp = (fftw_complex *) fftw_malloc(N * sizeof(fftw_complex));
WHEN_VERBOSE(2,
printf("Validating plan, N = %d, dir = %s\n", N,
dir == FFTW_FORWARD ? "FORWARD" : "BACKWARD"));
/* test 1: check linearity */
for (i = 0; i < rounds; ++i) {
fftw_complex alpha, beta;
c_re(alpha) = DRAND();
c_im(alpha) = DRAND();
c_re(beta) = DRAND();
c_im(beta) = DRAND();
fill_random(inA, N);
fill_random(inB, N);
fftwnd(plan, 1, inA, 1, N, outA, 1, N);
fftwnd(plan, 1, inB, 1, N, outB, 1, N);
array_scale(outA, alpha, N);
array_scale(outB, beta, N);
array_add(tmp, outA, outB, N);
array_scale(inA, alpha, N);
array_scale(inB, beta, N);
array_add(inC, inA, inB, N);
fftwnd(plan, 1, inC, 1, N, outC, 1, N);
array_compare(outC, tmp, N);
}
/*
* test 2: check that the unit impulse is transformed properly -- we
* need to test both the real and imaginary impulses
*/
for (which_impulse = 0; which_impulse < 2; ++which_impulse) {
if (which_impulse == 0) { /* real impulse */
c_re(impulse) = 1.0;
c_im(impulse) = 0.0;
} else { /* imaginary impulse */
c_re(impulse) = 0.0;
c_im(impulse) = 1.0;
}
for (i = 0; i < N; ++i) {
/* impulse */
c_re(inA[i]) = 0.0;
c_im(inA[i]) = 0.0;
/* transform of the impulse */
outA[i] = impulse;
}
inA[0] = impulse;
for (i = 0; i < rounds; ++i) {
fill_random(inB, N);
array_sub(inC, inA, inB, N);
fftwnd(plan, 1, inB, 1, N, outB, 1, N);
fftwnd(plan, 1, inC, 1, N, outC, 1, N);
array_add(tmp, outB, outC, N);
array_compare(tmp, outA, N);
}
}
/* test 3: check the time-shift property */
/* the paper performs more tests, but this code should be fine too */
/* -- we have to check shifts in each dimension */
n_before = 1;
n_after = N;
for (dim = 0; dim < rank; ++dim) {
int n_cur = n[dim];
n_after /= n_cur;
twopin = FFTW_K2PI / (FFTW_TRIG_REAL) n_cur;
for (i = 0; i < rounds; ++i) {
int j, jb, ja;
fill_random(inA, N);
array_rol(inB, inA, n_cur, n_before, n_after);
fftwnd(plan, 1, inA, 1, N, outA, 1, N);
fftwnd(plan, 1, inB, 1, N, outB, 1, N);
for (jb = 0; jb < n_before; ++jb)
for (j = 0; j < n_cur; ++j) {
FFTW_TRIG_REAL s = dir * FFTW_TRIG_SIN(j * twopin);
FFTW_TRIG_REAL c = FFTW_TRIG_COS(j * twopin);
for (ja = 0; ja < n_after; ++ja) {
c_re(tmp[(jb * n_cur + j) * n_after + ja]) =
c_re(outB[(jb * n_cur + j) * n_after + ja]) * c
- c_im(outB[(jb * n_cur + j) * n_after + ja]) * s;
c_im(tmp[(jb * n_cur + j) * n_after + ja]) =
c_re(outB[(jb * n_cur + j) * n_after + ja]) * s
+ c_im(outB[(jb * n_cur + j) * n_after + ja]) * c;
}
}
array_compare(tmp, outA, N);
}
n_before *= n_cur;
}
WHEN_VERBOSE(2, printf("Validation done\n"));
fftw_free(tmp);
fftw_free(outC);
fftw_free(outB);
fftw_free(outA);
fftw_free(inC);
fftw_free(inB);
fftw_free(inA);
}
void test_in_place(int n, int istride, int howmany, fftw_direction dir,
fftw_plan validated_plan, int specific)
{
int local_n, local_start, local_n_after_transform,
local_start_after_transform, total_local_size;
fftw_complex *in1, *work = NULL, *in2, *out2;
fftw_mpi_plan plan;
int i;
int flags = measure_flag | wisdom_flag | FFTW_IN_PLACE;
if (specific) {
WHEN_VERBOSE(2, my_printf("N/A\n"));
return;
}
if (coinflip())
flags |= FFTW_THREADSAFE;
plan = fftw_mpi_create_plan(MPI_COMM_WORLD, n, dir, flags);
fftw_mpi_local_sizes(plan, &local_n, &local_start,
&local_n_after_transform,
&local_start_after_transform,
&total_local_size);
in1 = (fftw_complex *) fftw_malloc(total_local_size
* sizeof(fftw_complex) * howmany);
if (coinflip()) {
WHEN_VERBOSE(2, my_printf("w/work..."));
work = (fftw_complex *) fftw_malloc(total_local_size
* sizeof(fftw_complex) * howmany);
}
in2 = (fftw_complex *) fftw_malloc(n * sizeof(fftw_complex) * howmany);
out2 = (fftw_complex *) fftw_malloc(n * sizeof(fftw_complex) * howmany);
/* generate random inputs */
for (i = 0; i < n * howmany; ++i) {
c_re(in2[i]) = DRAND();
c_im(in2[i]) = DRAND();
}
for (i = 0; i < local_n * howmany; ++i) {
c_re(in1[i]) = c_re(in2[i + local_start*howmany]);
c_im(in1[i]) = c_im(in2[i + local_start*howmany]);
}
/* fft-ize */
fftw_mpi(plan, howmany, in1, work);
fftw_mpi_destroy_plan(plan);
fftw(validated_plan, howmany, in2, howmany, 1, out2, howmany, 1);
CHECK(compute_error_complex(in1, 1,
out2 + local_start_after_transform*howmany, 1,
howmany*local_n_after_transform) < TOLERANCE,
"test_in_place: wrong answer");
WHEN_VERBOSE(2, my_printf("OK\n"));
fftw_free(in1);
fftw_free(work);
fftw_free(in2);
fftw_free(out2);
}
void testnd_in_place(int rank, int *n, fftwnd_plan validated_plan,
int alternate_api, int specific)
{
int istride, ostride, howmany;
int N, dim, i, j, k;
int nc, nhc, nr;
fftw_real *in1, *out3;
fftw_complex *in2, *out1, *out2;
fftwnd_plan p, ip;
int flags = measure_flag | wisdom_flag | FFTW_IN_PLACE;
if (coinflip())
flags |= FFTW_THREADSAFE;
N = nc = nr = nhc = 1;
for (dim = 0; dim < rank; ++dim)
N *= n[dim];
if (rank > 0) {
nr = n[rank - 1];
nc = N / nr;
nhc = nr / 2 + 1;
}
in1 = (fftw_real *) fftw_malloc(2 * nhc * nc * MAX_STRIDE * sizeof(fftw_real));
out3 = in1;
out1 = (fftw_complex *) in1;
in2 = (fftw_complex *) fftw_malloc(N * sizeof(fftw_complex));
out2 = (fftw_complex *) fftw_malloc(N * sizeof(fftw_complex));
if (alternate_api && specific && (rank == 2 || rank == 3)) {
if (rank == 2) {
p = rfftw2d_create_plan_specific(n[0], n[1],
FFTW_REAL_TO_COMPLEX, flags,
in1, MAX_STRIDE, 0, 0);
ip = rfftw2d_create_plan_specific(n[0], n[1],
FFTW_COMPLEX_TO_REAL, flags,
in1, MAX_STRIDE, 0, 0);
} else {
p = rfftw3d_create_plan_specific(n[0], n[1], n[2],
FFTW_REAL_TO_COMPLEX, flags,
in1, MAX_STRIDE, 0, 0);
ip = rfftw3d_create_plan_specific(n[0], n[1], n[2],
FFTW_COMPLEX_TO_REAL, flags,
in1, MAX_STRIDE, 0, 0);
}
} else if (specific) {
p = rfftwnd_create_plan_specific(rank, n, FFTW_REAL_TO_COMPLEX,
flags,
in1, MAX_STRIDE, in1, MAX_STRIDE);
ip = rfftwnd_create_plan_specific(rank, n, FFTW_COMPLEX_TO_REAL,
flags,
in1, MAX_STRIDE, in1, MAX_STRIDE);
} else if (alternate_api && (rank == 2 || rank == 3)) {
if (rank == 2) {
p = rfftw2d_create_plan(n[0], n[1], FFTW_REAL_TO_COMPLEX,
flags);
ip = rfftw2d_create_plan(n[0], n[1], FFTW_COMPLEX_TO_REAL,
flags);
} else {
p = rfftw3d_create_plan(n[0], n[1], n[2], FFTW_REAL_TO_COMPLEX,
flags);
ip = rfftw3d_create_plan(n[0], n[1], n[2], FFTW_COMPLEX_TO_REAL,
flags);
}
} else {
p = rfftwnd_create_plan(rank, n, FFTW_REAL_TO_COMPLEX, flags);
ip = rfftwnd_create_plan(rank, n, FFTW_COMPLEX_TO_REAL, flags);
}
CHECK(p != NULL && ip != NULL, "can't create plan");
for (i = 0; i < nc * nhc * 2 * MAX_STRIDE; ++i)
out3[i] = 0;
for (istride = 1; istride <= MAX_STRIDE; ++istride) {
/* generate random inputs */
for (i = 0; i < nc; ++i)
for (j = 0; j < nr; ++j) {
c_re(in2[i * nr + j]) = DRAND();
c_im(in2[i * nr + j]) = 0.0;
for (k = 0; k < istride; ++k)
in1[(i * nhc * 2 + j) * istride + k]
= c_re(in2[i * nr + j]);
}
fftwnd(validated_plan, 1, in2, 1, 1, out2, 1, 1);
howmany = ostride = istride;
WHEN_VERBOSE(2, printf("\n testing in-place stride %d...",
istride));
if (howmany != 1 || istride != 1 || ostride != 1 || coinflip())
rfftwnd_real_to_complex(p, howmany, in1, istride, 1,
out1, ostride, 1);
else
rfftwnd_one_real_to_complex(p, in1, NULL);
for (i = 0; i < nc; ++i)
for (k = 0; k < howmany; ++k)
CHECK(compute_error_complex(out1 + i * nhc * ostride + k,
ostride,
out2 + i * nr, 1,
nhc) < TOLERANCE,
"in-place (r2c): wrong answer");
if (howmany != 1 || istride != 1 || ostride != 1 || coinflip())
rfftwnd_complex_to_real(ip, howmany, out1, ostride, 1,
out3, istride, 1);
else
rfftwnd_one_complex_to_real(ip, out1, NULL);
for (i = 0; i < nc * nhc * 2 * istride; ++i)
out3[i] *= 1.0 / N;
for (i = 0; i < nc; ++i)
for (k = 0; k < howmany; ++k)
CHECK(compute_error(out3 + i * nhc * 2 * istride + k,
istride,
(fftw_real *) (in2 + i * nr), 2,
nr) < TOLERANCE,
"in-place (c2r): wrong answer (check 2)");
}
rfftwnd_destroy_plan(p);
rfftwnd_destroy_plan(ip);
fftw_free(out2);
fftw_free(in2);
fftw_free(in1);
}
void test_speed_nd_aux(struct size sz,
fftw_direction dir, int flags, int specific)
{
int local_nx, local_x_start, local_ny_after_transpose,
local_y_start_after_transpose, total_local_size;
fftw_complex *in, *work;
fftwnd_plan plan = 0;
fftwnd_mpi_plan mpi_plan;
double t, t0 = 0.0;
int i, N;
if (sz.rank < 2)
return;
/* only bench in-place multi-dim transforms */
flags |= FFTW_IN_PLACE;
N = 1;
for (i = 0; i < sz.rank; ++i)
N *= (sz.narray[i]);
if (specific) {
return;
} else {
if (io_okay && !only_parallel)
plan = fftwnd_create_plan(sz.rank, sz.narray,
dir, speed_flag | flags
| wisdom_flag | no_vector_flag);
mpi_plan = fftwnd_mpi_create_plan(MPI_COMM_WORLD, sz.rank, sz.narray,
dir, speed_flag | flags
| wisdom_flag | no_vector_flag);
}
CHECK(mpi_plan != NULL, "can't create plan");
fftwnd_mpi_local_sizes(mpi_plan, &local_nx, &local_x_start,
&local_ny_after_transpose,
&local_y_start_after_transpose,
&total_local_size);
if (io_okay && !only_parallel)
in = (fftw_complex *) fftw_malloc(N * howmany_fields *
sizeof(fftw_complex));
else
in = (fftw_complex *) fftw_malloc(total_local_size * howmany_fields *
sizeof(fftw_complex));
work = (fftw_complex *) fftw_malloc(total_local_size * howmany_fields *
sizeof(fftw_complex));
if (io_okay && !only_parallel) {
FFTW_TIME_FFT(fftwnd(plan, howmany_fields,
in, howmany_fields, 1, 0, 0, 0),
in, N * howmany_fields, t0);
fftwnd_destroy_plan(plan);
WHEN_VERBOSE(1, printf("time for one fft (uniprocessor): %s\n",
smart_sprint_time(t0)));
}
MPI_TIME_FFT(fftwnd_mpi(mpi_plan, howmany_fields,
in, NULL, FFTW_NORMAL_ORDER),
in, total_local_size * howmany_fields, t);
if (io_okay) {
WHEN_VERBOSE(1, printf("NORMAL: time for one fft (%d cpus): %s",
ncpus, smart_sprint_time(t)));
WHEN_VERBOSE(1, printf(" (%s/point)\n", smart_sprint_time(t / N)));
WHEN_VERBOSE(1, printf("NORMAL: \"mflops\" = 5 (N log2 N) / "
"(t in microseconds)"
" = %f\n", howmany_fields * mflops(t, N)));
if (!only_parallel)
WHEN_VERBOSE(1, printf("NORMAL: parallel speedup: %f\n", t0 / t));
}
MPI_TIME_FFT(fftwnd_mpi(mpi_plan, howmany_fields,
in, NULL, FFTW_TRANSPOSED_ORDER),
in, total_local_size * howmany_fields, t);
if (io_okay) {
WHEN_VERBOSE(1, printf("TRANSP.: time for one fft (%d cpus): %s",
ncpus, smart_sprint_time(t)));
WHEN_VERBOSE(1, printf(" (%s/point)\n", smart_sprint_time(t / N)));
WHEN_VERBOSE(1, printf("TRANSP.: \"mflops\" = 5 (N log2 N) / "
"(t in microseconds)"
" = %f\n", howmany_fields * mflops(t, N)));
if (!only_parallel)
WHEN_VERBOSE(1, printf("TRANSP.: parallel speedup: %f\n", t0 / t));
}
MPI_TIME_FFT(fftwnd_mpi(mpi_plan, howmany_fields,
in, work, FFTW_NORMAL_ORDER),
in, total_local_size * howmany_fields, t);
if (io_okay) {
WHEN_VERBOSE(1, printf("NORMAL,w/WORK: time for one fft (%d cpus): %s",
ncpus, smart_sprint_time(t)));
WHEN_VERBOSE(1, printf(" (%s/point)\n", smart_sprint_time(t / N)));
WHEN_VERBOSE(1, printf("NORMAL,w/WORK: \"mflops\" = 5 (N log2 N) / "
"(t in microseconds)"
" = %f\n", howmany_fields * mflops(t, N)));
if (!only_parallel)
WHEN_VERBOSE(1, printf("NORMAL,w/WORK: parallel speedup: %f\n", t0 / t));
}
MPI_TIME_FFT(fftwnd_mpi(mpi_plan, howmany_fields,
in, work, FFTW_TRANSPOSED_ORDER),
in, total_local_size * howmany_fields, t);
if (io_okay) {
WHEN_VERBOSE(1, printf("TRANSP.,w/WORK: time for one fft (%d cpus): %s",
ncpus, smart_sprint_time(t)));
WHEN_VERBOSE(1, printf(" (%s/point)\n", smart_sprint_time(t / N)));
WHEN_VERBOSE(1, printf("TRANSP.,w/WORK: \"mflops\" = 5 (N log2 N) / "
"(t in microseconds)"
" = %f\n", howmany_fields * mflops(t, N)));
if (!only_parallel)
WHEN_VERBOSE(1, printf("TRANSP.,w/WORK: parallel speedup: %f\n", t0 / t));
}
fftwnd_mpi_destroy_plan(mpi_plan);
fftw_free(in);
fftw_free(work);
WHEN_VERBOSE(1, my_printf("\n"));
}
void test_planner(int rank)
{
/*
* create and destroy many plans, at random. Check the
* garbage-collecting allocator of twiddle factors
*/
int i, dim;
int r, s;
fftw_plan p[PLANNER_TEST_SIZE];
fftwnd_plan pnd[PLANNER_TEST_SIZE];
int *narr, maxdim;
chk_mem_leak = 0;
verbose--;
please_wait();
if (rank < 1)
rank = 1;
narr = (int *) fftw_malloc(rank * sizeof(int));
maxdim = (int) pow(8192.0, 1.0/rank);
for (i = 0; i < PLANNER_TEST_SIZE; ++i) {
p[i] = (fftw_plan) 0;
pnd[i] = (fftwnd_plan) 0;
}
for (i = 0; i < PLANNER_TEST_SIZE * PLANNER_TEST_SIZE; ++i) {
r = rand();
if (r < 0)
r = -r;
r = r % PLANNER_TEST_SIZE;
for (dim = 0; dim < rank; ++dim) {
do {
s = rand();
if (s < 0)
s = -s;
s = s % maxdim + 1;
} while (s == 0);
narr[dim] = s;
}
if (rank == 1) {
if (p[r])
fftw_destroy_plan(p[r]);
p[r] = fftw_create_plan(narr[0], random_dir(), measure_flag |
wisdom_flag);
if (paranoid && narr[0] < 200)
test_correctness(narr[0]);
}
if (pnd[r])
fftwnd_destroy_plan(pnd[r]);
pnd[r] = fftwnd_create_plan(rank, narr,
random_dir(), measure_flag |
wisdom_flag);
if (i % (PLANNER_TEST_SIZE * PLANNER_TEST_SIZE / 20) == 0) {
WHEN_VERBOSE(0, printf("test planner: so far so good\n"));
WHEN_VERBOSE(0, printf("test planner: iteration %d out of %d\n",
i, PLANNER_TEST_SIZE * PLANNER_TEST_SIZE));
}
}
for (i = 0; i < PLANNER_TEST_SIZE; ++i) {
if (p[i])
fftw_destroy_plan(p[i]);
if (pnd[i])
fftwnd_destroy_plan(pnd[i]);
}
fftw_free(narr);
verbose++;
chk_mem_leak = 1;
}
void testnd_out_of_place(int rank, int *n, fftwnd_plan validated_plan)
{
int istride, ostride;
int N, dim, i, j, k;
int nc, nhc, nr;
fftw_real *in1, *out3;
fftw_complex *in2, *out1, *out2;
fftwnd_plan p, ip;
int flags = measure_flag | wisdom_flag;
if (coinflip())
flags |= FFTW_THREADSAFE;
N = nc = nr = nhc = 1;
for (dim = 0; dim < rank; ++dim)
N *= n[dim];
if (rank > 0) {
nr = n[rank - 1];
nc = N / nr;
nhc = nr / 2 + 1;
}
in1 = (fftw_real *) fftw_malloc(N * MAX_STRIDE * sizeof(fftw_real));
out3 = (fftw_real *) fftw_malloc(N * MAX_STRIDE * sizeof(fftw_real));
out1 = (fftw_complex *) fftw_malloc(nhc * nc * MAX_STRIDE
* sizeof(fftw_complex));
in2 = (fftw_complex *) fftw_malloc(N * sizeof(fftw_complex));
out2 = (fftw_complex *) fftw_malloc(N * sizeof(fftw_complex));
p = rfftwnd_create_plan(rank, n, FFTW_REAL_TO_COMPLEX, flags);
ip = rfftwnd_create_plan(rank, n, FFTW_COMPLEX_TO_REAL, flags);
CHECK(p != NULL && ip != NULL, "can't create plan");
for (istride = 1; istride <= MAX_STRIDE; ++istride) {
/* generate random inputs */
for (i = 0; i < nc; ++i)
for (j = 0; j < nr; ++j) {
c_re(in2[i * nr + j]) = DRAND();
c_im(in2[i * nr + j]) = 0.0;
for (k = 0; k < istride; ++k)
in1[(i * nr + j) * istride + k]
= c_re(in2[i * nr + j]);
}
for (i = 0; i < N * istride; ++i)
out3[i] = 0.0;
fftwnd(validated_plan, 1, in2, 1, 1, out2, 1, 1);
for (ostride = 1; ostride <= MAX_STRIDE; ++ostride) {
int howmany = (istride < ostride) ? istride : ostride;
WHEN_VERBOSE(2, printf("\n testing stride %d/%d...",
istride, ostride));
if (howmany != 1 || istride != 1 || ostride != 1 || coinflip())
rfftwnd_real_to_complex(p, howmany, in1, istride, 1,
out1, ostride, 1);
else
rfftwnd_one_real_to_complex(p, in1, out1);
for (i = 0; i < nc; ++i)
for (k = 0; k < howmany; ++k)
CHECK(compute_error_complex(out1 + i * nhc * ostride + k,
ostride,
out2 + i * nr, 1,
nhc) < TOLERANCE,
"out-of-place (r2c): wrong answer");
if (howmany != 1 || istride != 1 || ostride != 1 || coinflip())
rfftwnd_complex_to_real(ip, howmany, out1, ostride, 1,
out3, istride, 1);
else
rfftwnd_one_complex_to_real(ip, out1, out3);
for (i = 0; i < N * istride; ++i)
out3[i] *= 1.0 / N;
if (istride == howmany)
CHECK(compute_error(out3, 1, in1, 1, N * istride)
< TOLERANCE, "out-of-place (c2r): wrong answer");
for (i = 0; i < nc; ++i)
for (k = 0; k < howmany; ++k)
CHECK(compute_error(out3 + i * nr * istride + k,
istride,
(fftw_real *) (in2 + i * nr), 2,
nr) < TOLERANCE,
"out-of-place (c2r): wrong answer (check 2)");
}
}
rfftwnd_destroy_plan(p);
rfftwnd_destroy_plan(ip);
fftw_free(out3);
fftw_free(out2);
fftw_free(in2);
fftw_free(out1);
fftw_free(in1);
}
void test_planner(int rank)
{
/*
* create and destroy many plans, at random. Check the
* garbage-collecting allocator of twiddle factors
*/
int i, dim;
int r, s;
fftw_mpi_plan p[PLANNER_TEST_SIZE];
fftwnd_mpi_plan pnd[PLANNER_TEST_SIZE];
int *narr, maxdim;
chk_mem_leak = 0;
verbose--;
please_wait();
if (rank < 1)
rank = 1;
narr = (int *) fftw_malloc(rank * sizeof(int));
for (i = 0; i < PLANNER_TEST_SIZE; ++i) {
p[i] = (fftw_mpi_plan) 0;
pnd[i] = (fftwnd_mpi_plan) 0;
}
if (PLANNER_TEST_SIZE >= 8) {
p[0] = fftw_mpi_create_plan(MPI_COMM_WORLD, 1024, FFTW_FORWARD, 0);
p[1] = fftw_mpi_create_plan(MPI_COMM_WORLD, 1024, FFTW_FORWARD, 0);
p[2] = fftw_mpi_create_plan(MPI_COMM_WORLD, 1024, FFTW_BACKWARD, 0);
p[3] = fftw_mpi_create_plan(MPI_COMM_WORLD, 1024, FFTW_BACKWARD, 0);
p[4] = fftw_mpi_create_plan(MPI_COMM_WORLD, 1024, FFTW_FORWARD, 0);
p[5] = fftw_mpi_create_plan(MPI_COMM_WORLD, 1024, FFTW_FORWARD, 0);
p[6] = fftw_mpi_create_plan(MPI_COMM_WORLD, 1024, FFTW_BACKWARD, 0);
p[7] = fftw_mpi_create_plan(MPI_COMM_WORLD, 1024, FFTW_BACKWARD, 0);
}
maxdim = (int) pow(8192, 1.0/rank);
for (i = 0; i < PLANNER_TEST_SIZE * PLANNER_TEST_SIZE; ++i) {
r = rand();
if (r < 0)
r = -r;
r = r % PLANNER_TEST_SIZE;
for (dim = 0; dim < rank; ++dim) {
do {
s = rand();
if (s < 0)
s = -s;
s = s % maxdim + 1;
} while (s == 0);
narr[dim] = s;
}
if (rank == 1) {
if (p[r])
fftw_mpi_destroy_plan(p[r]);
p[r] = fftw_mpi_create_plan(MPI_COMM_WORLD,
narr[0], random_dir(),
measure_flag | wisdom_flag);
}
if (rank > 1) {
if (pnd[r])
fftwnd_mpi_destroy_plan(pnd[r]);
pnd[r] = fftwnd_mpi_create_plan(MPI_COMM_WORLD, rank, narr,
random_dir(), measure_flag |
wisdom_flag);
}
if (i % (PLANNER_TEST_SIZE * PLANNER_TEST_SIZE / 20) == 0) {
WHEN_VERBOSE(0, my_printf("test planner: so far so good\n"));
WHEN_VERBOSE(0, my_printf("test planner: iteration %d"
" out of %d\n",
i, PLANNER_TEST_SIZE * PLANNER_TEST_SIZE));
}
}
for (i = 0; i < PLANNER_TEST_SIZE; ++i) {
if (p[i])
fftw_mpi_destroy_plan(p[i]);
if (pnd[i])
fftwnd_mpi_destroy_plan(pnd[i]);
}
fftw_free(narr);
verbose++;
chk_mem_leak = 1;
}
/*
* This is a real (as opposed to complex) variation of the FFT tester
* described in
*
* Funda Ergün. Testing multivariate linear functions: Overcoming the
* generator bottleneck. In Proceedings of the Twenty-Seventh Annual
* ACM Symposium on the Theory of Computing, pages 407-416, Las Vegas,
* Nevada, 29 May--1 June 1995.
*/
void test_ergun(int n, fftw_direction dir, fftw_plan plan)
{
fftw_real *inA, *inB, *inC, *outA, *outB, *outC;
fftw_real *inA1, *inB1;
fftw_real *tmp;
int i;
int rounds = 20;
FFTW_TRIG_REAL twopin = FFTW_K2PI / (FFTW_TRIG_REAL) n;
inA = (fftw_real *) fftw_malloc(n * sizeof(fftw_real));
inB = (fftw_real *) fftw_malloc(n * sizeof(fftw_real));
inA1 = (fftw_real *) fftw_malloc(n * sizeof(fftw_real));
inB1 = (fftw_real *) fftw_malloc(n * sizeof(fftw_real));
inC = (fftw_real *) fftw_malloc(n * sizeof(fftw_real));
outA = (fftw_real *) fftw_malloc(n * sizeof(fftw_real));
outB = (fftw_real *) fftw_malloc(n * sizeof(fftw_real));
outC = (fftw_real *) fftw_malloc(n * sizeof(fftw_real));
tmp = (fftw_real *) fftw_malloc(n * sizeof(fftw_real));
WHEN_VERBOSE(2,
printf("Validating plan, n = %d, dir = %s\n", n,
dir == FFTW_REAL_TO_COMPLEX ?
"REAL_TO_COMPLEX" : "COMPLEX_TO_REAL"));
/* test 1: check linearity */
for (i = 0; i < rounds; ++i) {
fftw_real alpha, beta;
alpha = DRAND();
beta = DRAND();
rfill_random(inA, n);
rfill_random(inB, n);
rarray_scale(inA1, inA, alpha, n);
rarray_scale(inB1, inB, beta, n);
rarray_add(inC, inA1, inB1, n);
rfftw_out_of_place(plan, n, inA, outA);
rfftw_out_of_place(plan, n, inB, outB);
rarray_scale(outA, outA, alpha, n);
rarray_scale(outB, outB, beta, n);
rarray_add(tmp, outA, outB, n);
rfftw_out_of_place(plan, n, inC, outC);
rarray_compare(outC, tmp, n);
}
/* test 2: check that the unit impulse is transformed properly */
for (i = 0; i < n; ++i) {
/* impulse */
inA[i] = 0.0;
/* transform of the impulse */
if (2 * i <= n)
outA[i] = 1.0;
else
outA[i] = 0.0;
}
inA[0] = 1.0;
if (dir == FFTW_REAL_TO_COMPLEX) {
for (i = 0; i < rounds; ++i) {
rfill_random(inB, n);
rarray_sub(inC, inA, inB, n);
rfftw_out_of_place(plan, n, inB, outB);
rfftw_out_of_place(plan, n, inC, outC);
rarray_add(tmp, outB, outC, n);
rarray_compare(tmp, outA, n);
}
} else {
for (i = 0; i < rounds; ++i) {
rfill_random(outB, n);
rarray_sub(outC, outA, outB, n);
rfftw_out_of_place(plan, n, outB, inB);
rfftw_out_of_place(plan, n, outC, inC);
rarray_add(tmp, inB, inC, n);
rarray_scale(tmp, tmp, 1.0 / ((double) n), n);
rarray_compare(tmp, inA, n);
}
}
/* test 3: check the time-shift property */
/* the paper performs more tests, but this code should be fine too */
if (dir == FFTW_REAL_TO_COMPLEX) {
for (i = 0; i < rounds; ++i) {
int j;
rfill_random(inA, n);
rarray_rol(inB, inA, n, 1, 1);
rfftw_out_of_place(plan, n, inA, outA);
rfftw_out_of_place(plan, n, inB, outB);
tmp[0] = outB[0];
for (j = 1; 2 * j < n; ++j) {
FFTW_TRIG_REAL s = dir * FFTW_TRIG_SIN(j * twopin);
FFTW_TRIG_REAL c = FFTW_TRIG_COS(j * twopin);
tmp[j] = outB[j] * c - outB[n - j] * s;
tmp[n - j] = outB[j] * s + outB[n - j] * c;
}
if (2 * j == n)
tmp[j] = -outB[j];
rarray_compare(tmp, outA, n);
}
} else {
for (i = 0; i < rounds; ++i) {
int j;
rfill_random(inA, n);
inB[0] = inA[0];
for (j = 1; 2 * j < n; ++j) {
FFTW_TRIG_REAL s = dir * FFTW_TRIG_SIN(j * twopin);
FFTW_TRIG_REAL c = FFTW_TRIG_COS(j * twopin);
inB[j] = inA[j] * c - inA[n - j] * s;
inB[n - j] = inA[j] * s + inA[n - j] * c;
}
if (2 * j == n)
inB[j] = -inA[j];
rfftw_out_of_place(plan, n, inA, outA);
rfftw_out_of_place(plan, n, inB, outB);
rarray_rol(tmp, outA, n, 1, 1);
rarray_compare(tmp, outB, n);
}
}
WHEN_VERBOSE(2, printf("Validation done\n"));
fftw_free(tmp);
fftw_free(outC);
fftw_free(outB);
fftw_free(outA);
fftw_free(inC);
fftw_free(inB1);
fftw_free(inA1);
fftw_free(inB);
fftw_free(inA);
}
static void fftw_plan_hook_function(fftw_plan p)
{
WHEN_VERBOSE(3, printf("Validating tentative plan\n"));
WHEN_VERBOSE(3, fftw_print_plan(p));
test_ergun(p->n, p->dir, p);
}
/*
* Implementation of the FFT tester described in
*
* Funda Ergün. Testing multivariate linear functions: Overcoming the
* generator bottleneck. In Proceedings of the Twenty-Seventh Annual
* ACM Symposium on the Theory of Computing, pages 407-416, Las Vegas,
* Nevada, 29 May--1 June 1995.
*/
void test_ergun(int n, fftw_direction dir, fftw_plan plan)
{
fftw_complex *inA, *inB, *inC, *outA, *outB, *outC;
fftw_complex *tmp;
fftw_complex impulse;
int i;
int rounds = 20;
FFTW_TRIG_REAL twopin = FFTW_K2PI / (FFTW_TRIG_REAL) n;
inA = (fftw_complex *) fftw_malloc(n * sizeof(fftw_complex));
inB = (fftw_complex *) fftw_malloc(n * sizeof(fftw_complex));
inC = (fftw_complex *) fftw_malloc(n * sizeof(fftw_complex));
outA = (fftw_complex *) fftw_malloc(n * sizeof(fftw_complex));
outB = (fftw_complex *) fftw_malloc(n * sizeof(fftw_complex));
outC = (fftw_complex *) fftw_malloc(n * sizeof(fftw_complex));
tmp = (fftw_complex *) fftw_malloc(n * sizeof(fftw_complex));
WHEN_VERBOSE(2,
printf("Validating plan, n = %d, dir = %s\n", n,
dir == FFTW_FORWARD ? "FORWARD" : "BACKWARD"));
/* test 1: check linearity */
for (i = 0; i < rounds; ++i) {
fftw_complex alpha, beta;
c_re(alpha) = DRAND();
c_im(alpha) = DRAND();
c_re(beta) = DRAND();
c_im(beta) = DRAND();
fill_random(inA, n);
fill_random(inB, n);
fftw_out_of_place(plan, n, inA, outA);
fftw_out_of_place(plan, n, inB, outB);
array_scale(outA, alpha, n);
array_scale(outB, beta, n);
array_add(tmp, outA, outB, n);
array_scale(inA, alpha, n);
array_scale(inB, beta, n);
array_add(inC, inA, inB, n);
fftw_out_of_place(plan, n, inC, outC);
array_compare(outC, tmp, n);
}
/* test 2: check that the unit impulse is transformed properly */
c_re(impulse) = 1.0;
c_im(impulse) = 0.0;
for (i = 0; i < n; ++i) {
/* impulse */
c_re(inA[i]) = 0.0;
c_im(inA[i]) = 0.0;
/* transform of the impulse */
outA[i] = impulse;
}
inA[0] = impulse;
for (i = 0; i < rounds; ++i) {
fill_random(inB, n);
array_sub(inC, inA, inB, n);
fftw_out_of_place(plan, n, inB, outB);
fftw_out_of_place(plan, n, inC, outC);
array_add(tmp, outB, outC, n);
array_compare(tmp, outA, n);
}
/* test 3: check the time-shift property */
/* the paper performs more tests, but this code should be fine too */
for (i = 0; i < rounds; ++i) {
int j;
fill_random(inA, n);
array_rol(inB, inA, n, 1, 1);
fftw_out_of_place(plan, n, inA, outA);
fftw_out_of_place(plan, n, inB, outB);
for (j = 0; j < n; ++j) {
FFTW_TRIG_REAL s = dir * FFTW_TRIG_SIN(j * twopin);
FFTW_TRIG_REAL c = FFTW_TRIG_COS(j * twopin);
c_re(tmp[j]) = c_re(outB[j]) * c - c_im(outB[j]) * s;
c_im(tmp[j]) = c_re(outB[j]) * s + c_im(outB[j]) * c;
}
array_compare(tmp, outA, n);
}
WHEN_VERBOSE(2, printf("Validation done\n"));
fftw_free(tmp);
fftw_free(outC);
fftw_free(outB);
fftw_free(outA);
fftw_free(inC);
fftw_free(inB);
fftw_free(inA);
}
void testnd_in_place(int rank, int *n, fftw_direction dir,
fftwnd_plan validated_plan,
int alternate_api, int specific, int force_buffered)
{
int local_nx, local_x_start, local_ny_after_transpose,
local_y_start_after_transpose, total_local_size;
int istride;
int N, dim, i;
fftw_complex *in1, *work = 0, *in2;
fftwnd_mpi_plan p = 0;
int flags = measure_flag | wisdom_flag | FFTW_IN_PLACE;
if (specific || rank < 2)
return;
if (coinflip())
flags |= FFTW_THREADSAFE;
if (force_buffered)
flags |= FFTWND_FORCE_BUFFERED;
N = 1;
for (dim = 0; dim < rank; ++dim)
N *= n[dim];
if (alternate_api && (rank == 2 || rank == 3)) {
if (rank == 2)
p = fftw2d_mpi_create_plan(MPI_COMM_WORLD,
n[0], n[1], dir, flags);
else
p = fftw3d_mpi_create_plan(MPI_COMM_WORLD,
n[0], n[1], n[2], dir, flags);
}
else /* standard api */
p = fftwnd_mpi_create_plan(MPI_COMM_WORLD, rank, n, dir, flags);
fftwnd_mpi_local_sizes(p, &local_nx, &local_x_start,
&local_ny_after_transpose,
&local_y_start_after_transpose,
&total_local_size);
in1 = (fftw_complex *) fftw_malloc(total_local_size * MAX_STRIDE
* sizeof(fftw_complex));
if (coinflip()) {
WHEN_VERBOSE(1, my_printf("w/work..."));
work = (fftw_complex *) fftw_malloc(total_local_size * MAX_STRIDE
* sizeof(fftw_complex));
}
in2 = (fftw_complex *) fftw_malloc(N * sizeof(fftw_complex));
for (istride = 1; istride <= MAX_STRIDE; ++istride) {
/* generate random inputs */
for (i = 0; i < N; ++i) {
c_re(in2[i]) = DRAND();
c_im(in2[i]) = DRAND();
}
for (i = 0; i < local_nx * (N/n[0]); ++i) {
int j;
for (j = 0; j < istride; ++j) {
c_re(in1[i * istride + j]) = c_re((in2 + local_x_start
* (N/n[0])) [i]);
c_im(in1[i * istride + j]) = c_im((in2 + local_x_start
* (N/n[0])) [i]);
}
}
fftwnd_mpi(p, istride, in1, work, FFTW_NORMAL_ORDER);
fftwnd(validated_plan, 1, in2, 1, 1, NULL, 0, 0);
for (i = 0; i < istride; ++i)
CHECK(compute_error_complex(in1 + i, istride,
in2 + local_x_start * (N/n[0]),
1, local_nx * (N/n[0])) < TOLERANCE,
"testnd_in_place: wrong answer");
}
fftwnd_mpi_destroy_plan(p);
fftw_free(in2);
fftw_free(work);
fftw_free(in1);
}
void test_in_place(int n, int istride, int howmany, fftw_direction dir,
fftw_plan validated_plan, int specific)
{
fftw_complex *in1, *in2, *out2;
fftw_plan plan;
int i, j;
int flags = measure_flag | wisdom_flag | FFTW_IN_PLACE;
if (coinflip())
flags |= FFTW_THREADSAFE;
in1 = (fftw_complex *) fftw_malloc(istride * n * sizeof(fftw_complex) * howmany);
in2 = (fftw_complex *) fftw_malloc(n * sizeof(fftw_complex) * howmany);
out2 = (fftw_complex *) fftw_malloc(n * sizeof(fftw_complex) * howmany);
if (!specific)
plan = fftw_create_plan(n, dir, flags);
else
plan = fftw_create_plan_specific(n, dir, flags,
in1, istride,
(fftw_complex *) NULL, 0);
/* generate random inputs */
for (i = 0; i < n * howmany; ++i) {
c_re(in1[i * istride]) = c_re(in2[i]) = DRAND();
c_im(in1[i * istride]) = c_im(in2[i]) = DRAND();
}
/*
* fill in other positions of the array, to make sure that
* fftw doesn't overwrite them
*/
for (j = 1; j < istride; ++j)
for (i = 0; i < n * howmany; ++i) {
c_re(in1[i * istride + j]) = i * istride + j;
c_im(in1[i * istride + j]) = i * istride - j;
}
CHECK(plan != NULL, "can't create plan");
WHEN_VERBOSE(2, fftw_print_plan(plan));
/* fft-ize */
if (howmany != 1 || istride != 1 || coinflip())
fftw(plan, howmany, in1, istride, n * istride,
(fftw_complex *) NULL, 0, 0);
else
fftw_one(plan, in1, NULL);
fftw_destroy_plan(plan);
/* check for overwriting */
for (j = 1; j < istride; ++j)
for (i = 0; i < n * howmany; ++i)
CHECK(c_re(in1[i * istride + j]) == i * istride + j &&
c_im(in1[i * istride + j]) == i * istride - j,
"input has been overwritten");
for (i = 0; i < howmany; ++i) {
fftw(validated_plan, 1, in2 + n * i, 1, n, out2 + n * i, 1, n);
}
CHECK(compute_error_complex(in1, istride, out2, 1, n * howmany) < TOLERANCE,
"test_in_place: wrong answer");
WHEN_VERBOSE(2, printf("OK\n"));
fftw_free(in1);
fftw_free(in2);
fftw_free(out2);
}
void test_speed_aux(int n, fftw_direction dir, int flags, int specific)
{
int local_n, local_start, local_n_after_transform,
local_start_after_transform, total_local_size, nalloc;
fftw_complex *in, *work;
fftw_plan plan = 0;
fftw_mpi_plan mpi_plan;
double t, t0 = 0.0;
if (specific || !(flags & FFTW_IN_PLACE))
return;
if (io_okay && !only_parallel)
plan = fftw_create_plan(n, dir, speed_flag | flags
| wisdom_flag | no_vector_flag);
mpi_plan = fftw_mpi_create_plan(MPI_COMM_WORLD, n, dir,
speed_flag | flags
| wisdom_flag | no_vector_flag);
CHECK(mpi_plan, "failed to create plan!");
fftw_mpi_local_sizes(mpi_plan, &local_n, &local_start,
&local_n_after_transform,
&local_start_after_transform,
&total_local_size);
if (io_okay && !only_parallel)
nalloc = n;
else
nalloc = total_local_size;
in = (fftw_complex *) fftw_malloc(nalloc * howmany_fields
* sizeof(fftw_complex));
work = (fftw_complex *) fftw_malloc(nalloc * howmany_fields
* sizeof(fftw_complex));
if (io_okay) {
WHEN_VERBOSE(2, fftw_mpi_print_plan(mpi_plan));
}
if (io_okay && !only_parallel) {
FFTW_TIME_FFT(fftw(plan, howmany_fields,
in, howmany_fields, 1, work, 1, 0),
in, n * howmany_fields, t0);
fftw_destroy_plan(plan);
WHEN_VERBOSE(1, printf("time for one fft (uniprocessor): %s\n", smart_sprint_time(t0)));
}
MPI_TIME_FFT(fftw_mpi(mpi_plan, howmany_fields, in, NULL),
in, total_local_size * howmany_fields, t);
if (io_okay) {
WHEN_VERBOSE(1, printf("time for one fft (%d cpus): %s", ncpus, smart_sprint_time(t)));
WHEN_VERBOSE(1, printf(" (%s/point)\n", smart_sprint_time(t / n)));
WHEN_VERBOSE(1, printf("\"mflops\" = 5 (n log2 n) / (t in microseconds)"
" = %f\n", howmany_fields * mflops(t, n)));
if (!only_parallel)
WHEN_VERBOSE(1, printf("parallel speedup: %f\n", t0 / t));
}
MPI_TIME_FFT(fftw_mpi(mpi_plan, howmany_fields, in, work),
in, total_local_size * howmany_fields, t);
if (io_okay) {
WHEN_VERBOSE(1, printf("w/WORK: time for one fft (%d cpus): %s", ncpus, smart_sprint_time(t)));
WHEN_VERBOSE(1, printf(" (%s/point)\n", smart_sprint_time(t / n)));
WHEN_VERBOSE(1, printf("w/WORK: \"mflops\" = 5 (n log2 n) / (t in microseconds)"
" = %f\n", howmany_fields * mflops(t, n)));
if (!only_parallel)
WHEN_VERBOSE(1, printf("w/WORK: parallel speedup: %f\n", t0 / t));
}
fftw_free(in);
fftw_free(work);
fftw_mpi_destroy_plan(mpi_plan);
WHEN_VERBOSE(1, my_printf("\n"));
}
void test_in_place(int n, int istride,
int howmany, fftw_direction dir,
fftw_plan validated_plan, int specific)
{
fftw_complex *in2, *out2;
fftw_real *in1, *out1, *out3;
fftw_plan plan;
int i, j;
int ostride = istride;
int flags = measure_flag | wisdom_flag | FFTW_IN_PLACE;
if (coinflip())
flags |= FFTW_THREADSAFE;
in1 = (fftw_real *) fftw_malloc(istride * n * sizeof(fftw_real) * howmany);
in2 = (fftw_complex *) fftw_malloc(n * sizeof(fftw_complex));
out1 = in1;
out2 = (fftw_complex *) fftw_malloc(n * sizeof(fftw_complex));
out3 = (fftw_real *) fftw_malloc(n * sizeof(fftw_real));
if (!specific)
plan = rfftw_create_plan(n, dir, flags);
else
plan = rfftw_create_plan_specific(n, dir, flags,
in1, istride, out1, ostride);
CHECK(plan != NULL, "can't create plan");
/* generate random inputs */
fill_random(in1, n, istride);
for (j = 1; j < howmany; ++j)
for (i = 0; i < n; ++i)
in1[(j * n + i) * istride] = in1[i * istride];
/* copy random inputs to complex array for comparison with fftw: */
if (dir == FFTW_REAL_TO_COMPLEX)
for (i = 0; i < n; ++i) {
c_re(in2[i]) = in1[i * istride];
c_im(in2[i]) = 0.0;
} else {
int n2 = (n + 1) / 2;
c_re(in2[0]) = in1[0];
c_im(in2[0]) = 0.0;
for (i = 1; i < n2; ++i) {
c_re(in2[i]) = in1[i * istride];
c_im(in2[i]) = in1[(n - i) * istride];
}
if (n2 * 2 == n) {
c_re(in2[n2]) = in1[n2 * istride];
c_im(in2[n2]) = 0.0;
++i;
}
for (; i < n; ++i) {
c_re(in2[i]) = c_re(in2[n - i]);
c_im(in2[i]) = -c_im(in2[n - i]);
}
}
/*
* fill in other positions of the array, to make sure that
* rfftw doesn't overwrite them
*/
for (j = 1; j < istride; ++j)
for (i = 0; i < n * howmany; ++i)
in1[i * istride + j] = i * istride + j;
WHEN_VERBOSE(2, rfftw_print_plan(plan));
/* fft-ize */
if (howmany != 1 || istride != 1 || coinflip())
rfftw(plan, howmany, in1, istride, n * istride, 0, 0, 0);
else
rfftw_one(plan, in1, NULL);
rfftw_destroy_plan(plan);
/* check for overwriting */
for (j = 1; j < ostride; ++j)
for (i = 0; i < n * howmany; ++i)
CHECK(out1[i * ostride + j] == i * ostride + j,
"output has been overwritten");
fftw(validated_plan, 1, in2, 1, n, out2, 1, n);
if (dir == FFTW_REAL_TO_COMPLEX) {
int n2 = (n + 1) / 2;
out3[0] = c_re(out2[0]);
for (i = 1; i < n2; ++i) {
out3[i] = c_re(out2[i]);
out3[n - i] = c_im(out2[i]);
}
if (n2 * 2 == n)
out3[n2] = c_re(out2[n2]);
} else {
for (i = 0; i < n; ++i)
out3[i] = c_re(out2[i]);
}
for (j = 0; j < howmany; ++j)
CHECK(compute_error(out1 + j * n * ostride, ostride, out3, 1, n)
< TOLERANCE,
"test_in_place: wrong answer");
WHEN_VERBOSE(2, printf("OK\n"));
fftw_free(in1);
fftw_free(in2);
fftw_free(out2);
fftw_free(out3);
}
