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"));
}
/*
* Same as fftw_measure_runtime, except for fftwnd plan.
*/
double fftwnd_measure_runtime(fftwnd_plan plan,
fftw_complex *in, int istride,
fftw_complex *out, int ostride)
{
fftw_time begin, end, start;
double t, tmax, tmin;
int i, iter;
int n;
int repeat;
if (plan->rank == 0)
return 0.0;
n = 1;
for (i = 0; i < plan->rank; ++i)
n *= plan->n[i];
iter = 1;
for (;;) {
tmin = 1.0E10;
tmax = -1.0E10;
init_test_array(in, istride, n);
start = fftw_get_time();
/* repeat the measurement FFTW_TIME_REPEAT times */
for (repeat = 0; repeat < FFTW_TIME_REPEAT; ++repeat) {
begin = fftw_get_time();
for (i = 0; i < iter; ++i) {
fftwnd(plan, 1, in, istride, 0, out, ostride, 0);
}
end = fftw_get_time();
t = fftw_time_to_sec(fftw_time_diff(end, begin));
if (t < tmin)
tmin = t;
if (t > tmax)
tmax = t;
/* do not run for too long */
t = fftw_time_to_sec(fftw_time_diff(end, start));
if (t > FFTW_TIME_LIMIT)
break;
}
if (tmin >= FFTW_TIME_MIN)
break;
iter *= 2;
}
tmin /= (double) iter;
tmax /= (double) iter;
return tmin;
}
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"));
}
/*
* The timer keeps doubling the number of iterations
* until the program runs for more than FFTW_TIME_MIN
*/
double fftw_measure_runtime(fftw_plan plan)
{
FFTW_COMPLEX *in, *out;
fftw_time begin, end;
double t;
int i, iter;
int n;
n = plan->n;
iter = 1;
retry:
in = (FFTW_COMPLEX *) fftw_malloc(n * sizeof(FFTW_COMPLEX));
out = (FFTW_COMPLEX *) fftw_malloc(n * sizeof(FFTW_COMPLEX));
begin = fftw_get_time();
for (i = 0; i < iter; ++i) {
int j;
/* generate random inputs */
for (j = 0; j < n; ++j) {
c_re(in[j]) = 1.0;
c_im(in[j]) = 32.432;
}
fftw(plan, 1, in, 1, 0, out, 1, 0);
}
end = fftw_get_time();
t = fftw_time_to_sec(fftw_time_diff(end,begin));
fftw_free(in);
fftw_free(out);
if (t < FFTW_TIME_MIN) {
iter *= 2;
/*
* See D. E. Knuth, Structured Programming with GOTO Statements,
* Computing Surveys (6), December 1974, for a justification
* of this `goto' in the `n + 1/2' loop.
*/
goto retry;
}
return t / (double)iter;
}
