void problem_free(bench_problem *p)
{
if (p->outphys && p->outphys != p->inphys)
bench_free(p->outphys);
if (p->inphys)
bench_free(p->inphys);
tensor_destroy(p->sz);
tensor_destroy(p->vecsz);
}
void problem_destroy(bench_problem *p)
{
BENCH_ASSERT(p);
problem_free(p);
if (p->k)
bench_free(p->k);
bench_free(p->pstring);
bench_free(p);
}
static void alloc_local(ptrdiff_t nreal, int inplace)
{
bench_free(local_in);
if (local_out != local_in) bench_free(local_out);
local_in = local_out = 0;
if (nreal > 0) {
ptrdiff_t i;
local_in = (bench_real*) bench_malloc(nreal * sizeof(bench_real));
if (inplace)
local_out = local_in;
else
local_out = (bench_real*) bench_malloc(nreal * sizeof(bench_real));
for (i = 0; i < nreal; ++i) local_in[i] = local_out[i] = 0.0;
}
}
void verify_rdft2(bench_problem *p, int rounds, double tol, errors *e)
{
C *inA, *inB, *inC, *outA, *outB, *outC, *tmp;
int n, vecn, N;
dofft_rdft2_closure k;
BENCH_ASSERT(p->kind == PROBLEM_REAL);
if (!FINITE_RNK(p->sz->rnk) || !FINITE_RNK(p->vecsz->rnk))
return; /* give up */
k.k.apply = rdft2_apply;
k.k.recopy_input = 0;
k.p = p;
if (rounds == 0)
rounds = 20; /* default value */
n = tensor_sz(p->sz);
vecn = tensor_sz(p->vecsz);
N = n * vecn;
inA = (C *) bench_malloc(N * sizeof(C));
inB = (C *) bench_malloc(N * sizeof(C));
inC = (C *) bench_malloc(N * sizeof(C));
outA = (C *) bench_malloc(N * sizeof(C));
outB = (C *) bench_malloc(N * sizeof(C));
outC = (C *) bench_malloc(N * sizeof(C));
tmp = (C *) bench_malloc(N * sizeof(C));
e->i = impulse(&k.k, n, vecn, inA, inB, inC, outA, outB, outC,
tmp, rounds, tol);
e->l = linear(&k.k, 1, N, inA, inB, inC, outA, outB, outC,
tmp, rounds, tol);
e->s = 0.0;
if (p->sign < 0)
e->s = dmax(e->s, tf_shift(&k.k, 1, p->sz, n, vecn, p->sign,
inA, inB, outA, outB,
tmp, rounds, tol, TIME_SHIFT));
else
e->s = dmax(e->s, tf_shift(&k.k, 1, p->sz, n, vecn, p->sign,
inA, inB, outA, outB,
tmp, rounds, tol, FREQ_SHIFT));
if (!p->in_place && !p->destroy_input)
preserves_input(&k.k, p->sign < 0 ? mkreal : mkhermitian1,
N, inA, inB, outB, rounds);
bench_free(tmp);
bench_free(outC);
bench_free(outB);
bench_free(outA);
bench_free(inC);
bench_free(inB);
bench_free(inA);
}
void accuracy_dft(bench_problem *p, int rounds, int impulse_rounds,
double t[6])
{
dofft_dft_closure k;
int n;
C *a, *b;
BENCH_ASSERT(p->kind == PROBLEM_COMPLEX);
BENCH_ASSERT(p->sz->rnk == 1);
BENCH_ASSERT(p->vecsz->rnk == 0);
k.k.apply = dft_apply;
k.k.recopy_input = 0;
k.p = p;
n = tensor_sz(p->sz);
a = (C *) bench_malloc(n * sizeof(C));
b = (C *) bench_malloc(n * sizeof(C));
accuracy_test(&k.k, 0, p->sign, n, a, b, rounds, impulse_rounds, t);
bench_free(b);
bench_free(a);
}
void speed(const char *param)
{
double *t;
int iter, k;
bench_problem *p;
double tmin, y;
t = (double *) bench_malloc(time_repeat * sizeof(double));
p = problem_parse(param);
BENCH_ASSERT(can_do(p));
problem_alloc(p);
problem_zero(p);
timer_start();
setup(p);
p->setup_time = timer_stop();
start_over:
for (iter = 1; iter < (1<<30); iter *= 2) {
tmin = 1.0e20;
for (k = 0; k < time_repeat; ++k) {
timer_start();
doit(iter, p);
y = timer_stop();
if (y < 0) /* yes, it happens */
goto start_over;
t[k] = y;
if (y < tmin)
tmin = y;
}
if (tmin >= time_min)
goto done;
}
goto start_over; /* this also happens */
done:
done(p);
for (k = 0; k < time_repeat; ++k) {
t[k] /= iter;
}
report(p, t, time_repeat);
problem_destroy(p);
bench_free(t);
return;
}
void accuracy_r2r(bench_problem *p, int rounds, int impulse_rounds,
double t[6])
{
dofft_r2r_closure k;
int n, n0 = 1;
C *a, *b;
aconstrain constrain = 0;
BENCH_ASSERT(p->kind == PROBLEM_R2R);
BENCH_ASSERT(p->sz->rnk == 1);
BENCH_ASSERT(p->vecsz->rnk == 0);
k.k.apply = r2r_apply;
k.k.recopy_input = 0;
k.p = p;
n = tensor_sz(p->sz);
switch (p->k[0]) {
case R2R_R2HC: constrain = mkreal; n0 = n; break;
case R2R_HC2R: constrain = mkhermitian1; n0 = n; break;
case R2R_REDFT00: constrain = mkre00; n0 = 2*(n-1); break;
case R2R_RODFT00: constrain = mkro00; n0 = 2*(n+1); break;
case R2R_REDFT01: constrain = mkre01; n0 = 4*n; break;
case R2R_REDFT10: constrain = mkre10; n0 = 4*n; break;
case R2R_RODFT01: constrain = mkro01; n0 = 4*n; break;
case R2R_RODFT10: constrain = mkio10; n0 = 4*n; break;
case R2R_REDFT11: constrain = mkre11; n0 = 8*n; break;
case R2R_RODFT11: constrain = mkro11; n0 = 8*n; break;
default: BENCH_ASSERT(0); /* not yet implemented */
}
k.n0 = n0;
a = (C *) bench_malloc(n0 * sizeof(C));
b = (C *) bench_malloc(n0 * sizeof(C));
accuracy_test(&k.k, constrain, -1, n0, a, b, rounds, impulse_rounds, t);
bench_free(b);
bench_free(a);
}
void accuracy_rdft2(bench_problem *p, int rounds, int impulse_rounds,
double t[6])
{
dofft_rdft2_closure k;
int n;
C *a, *b;
BENCH_ASSERT(p->kind == PROBLEM_REAL);
BENCH_ASSERT(p->sz->rnk == 1);
BENCH_ASSERT(p->vecsz->rnk == 0);
k.k.apply = rdft2_apply;
k.k.recopy_input = 0;
k.p = p;
n = tensor_sz(p->sz);
a = (C *) bench_malloc(n * sizeof(C));
b = (C *) bench_malloc(n * sizeof(C));
accuracy_test(&k.k, p->sign < 0 ? mkreal : mkhermitian1, p->sign,
n, a, b, rounds, impulse_rounds, t);
bench_free(b);
bench_free(a);
}
void verify_dft(bench_problem *p, int rounds, double tol, errors *e)
{
C *inA, *inB, *inC, *outA, *outB, *outC, *tmp;
int n, vecn, N;
dofft_dft_closure k;
BENCH_ASSERT(p->kind == PROBLEM_COMPLEX);
k.k.apply = dft_apply;
k.k.recopy_input = 0;
k.p = p;
if (rounds == 0)
rounds = 20; /* default value */
n = tensor_sz(p->sz);
vecn = tensor_sz(p->vecsz);
N = n * vecn;
inA = (C *) bench_malloc(N * sizeof(C));
inB = (C *) bench_malloc(N * sizeof(C));
inC = (C *) bench_malloc(N * sizeof(C));
outA = (C *) bench_malloc(N * sizeof(C));
outB = (C *) bench_malloc(N * sizeof(C));
outC = (C *) bench_malloc(N * sizeof(C));
tmp = (C *) bench_malloc(N * sizeof(C));
e->i = impulse(&k.k, n, vecn, inA, inB, inC, outA, outB, outC,
tmp, rounds, tol);
e->l = linear(&k.k, 0, N, inA, inB, inC, outA, outB, outC,
tmp, rounds, tol);
e->s = 0.0;
e->s = dmax(e->s, tf_shift(&k.k, 0, p->sz, n, vecn, p->sign,
inA, inB, outA, outB,
tmp, rounds, tol, TIME_SHIFT));
e->s = dmax(e->s, tf_shift(&k.k, 0, p->sz, n, vecn, p->sign,
inA, inB, outA, outB,
tmp, rounds, tol, FREQ_SHIFT));
if (!p->in_place && !p->destroy_input)
preserves_input(&k.k, 0, N, inA, inB, outB, rounds);
bench_free(tmp);
bench_free(outC);
bench_free(outB);
bench_free(outA);
bench_free(inC);
bench_free(inB);
bench_free(inA);
}
static void alloc_rnk(int rnk_) {
rnk = rnk_;
bench_free(local_ni);
if (rnk == 0)
local_ni = 0;
else
local_ni = (ptrdiff_t *) bench_malloc(sizeof(ptrdiff_t) * rnk
* (8 + n_pes * 4));
local_starti = local_ni + rnk;
local_no = local_ni + 2 * rnk;
local_starto = local_ni + 3 * rnk;
istrides = local_ni + 4 * rnk;
ostrides = local_ni + 5 * rnk;
total_ni = local_ni + 6 * rnk;
total_no = local_ni + 7 * rnk;
all_local_ni = local_ni + 8 * rnk;
all_local_starti = local_ni + (8 + n_pes) * rnk;
all_local_no = local_ni + (8 + 2 * n_pes) * rnk;
all_local_starto = local_ni + (8 + 3 * n_pes) * rnk;
}
static void do_scatter_in(bench_real *in)
{
bench_real *ali;
int i;
if (all_local_in_alloc) {
bench_free(all_local_in);
all_local_in = (bench_real*) bench_malloc(iNtot*sizeof(bench_real));
all_local_in_alloc = 0;
}
ali = all_local_in;
for (i = 0; i < n_pes; ++i) {
copy_block_in(ali,
rnk, all_local_ni + i * rnk,
all_local_starti + i * rnk,
vn, istrides, vn,
in);
ali += isend_cnt[i];
}
MPI_Scatterv(all_local_in, isend_cnt, isend_off, BENCH_MPI_TYPE,
local_in, isend_cnt[my_pe], BENCH_MPI_TYPE,
0, MPI_COMM_WORLD);
}
static void do_gather_out(bench_real *out)
{
bench_real *alo;
int i;
if (all_local_out_alloc) {
bench_free(all_local_out);
all_local_out = (bench_real*) bench_malloc(oNtot*sizeof(bench_real));
all_local_out_alloc = 0;
}
MPI_Gatherv(local_out, orecv_cnt[my_pe], BENCH_MPI_TYPE,
all_local_out, orecv_cnt, orecv_off, BENCH_MPI_TYPE,
0, MPI_COMM_WORLD);
MPI_Bcast(all_local_out, oNtot, BENCH_MPI_TYPE, 0, MPI_COMM_WORLD);
alo = all_local_out;
for (i = 0; i < n_pes; ++i) {
copy_block_out(alo,
rnk, all_local_no + i * rnk,
all_local_starto + i * rnk,
vn, ostrides, vn,
out);
alo += orecv_cnt[i];
}
}
void verify_r2r(bench_problem *p, int rounds, double tol, errors *e)
{
R *inA, *inB, *inC, *outA, *outB, *outC, *tmp;
info nfo;
int n, vecn, N;
double impulse_amp = 1.0;
dim_stuff *d;
int i;
if (rounds == 0)
rounds = 20; /* default value */
n = tensor_sz(p->sz);
vecn = tensor_sz(p->vecsz);
N = n * vecn;
d = (dim_stuff *) bench_malloc(sizeof(dim_stuff) * p->sz->rnk);
for (i = 0; i < p->sz->rnk; ++i) {
int n0, i0, k0;
trigfun ti, ts;
d[i].n = n0 = p->sz->dims[i].n;
if (p->k[i] > R2R_DHT)
n0 = 2 * (n0 + (p->k[i] == R2R_REDFT00 ? -1 :
(p->k[i] == R2R_RODFT00 ? 1 : 0)));
switch (p->k[i]) {
case R2R_R2HC:
i0 = k0 = 0;
ti = realhalf;
ts = coshalf;
break;
case R2R_DHT:
i0 = k0 = 0;
ti = unity;
ts = cos00;
break;
case R2R_HC2R:
i0 = k0 = 0;
ti = unity;
ts = cos00;
break;
case R2R_REDFT00:
i0 = k0 = 0;
ti = ts = cos00;
break;
case R2R_REDFT01:
i0 = k0 = 0;
ti = ts = cos01;
break;
case R2R_REDFT10:
i0 = k0 = 0;
ti = cos10; impulse_amp *= 2.0;
ts = cos00;
break;
case R2R_REDFT11:
i0 = k0 = 0;
ti = cos11; impulse_amp *= 2.0;
ts = cos01;
break;
case R2R_RODFT00:
i0 = k0 = 1;
ti = sin00; impulse_amp *= 2.0;
ts = cos00;
break;
case R2R_RODFT01:
i0 = 1; k0 = 0;
ti = sin01; impulse_amp *= n == 1 ? 1.0 : 2.0;
ts = cos01;
break;
case R2R_RODFT10:
i0 = 0; k0 = 1;
ti = sin10; impulse_amp *= 2.0;
ts = cos00;
break;
case R2R_RODFT11:
i0 = k0 = 0;
ti = sin11; impulse_amp *= 2.0;
ts = cos01;
break;
default:
BENCH_ASSERT(0);
return;
}
d[i].n0 = n0;
d[i].i0 = i0;
d[i].k0 = k0;
d[i].ti = ti;
d[i].ts = ts;
}
inA = (R *) bench_malloc(N * sizeof(R));
inB = (R *) bench_malloc(N * sizeof(R));
inC = (R *) bench_malloc(N * sizeof(R));
outA = (R *) bench_malloc(N * sizeof(R));
outB = (R *) bench_malloc(N * sizeof(R));
outC = (R *) bench_malloc(N * sizeof(R));
tmp = (R *) bench_malloc(N * sizeof(R));
nfo.p = p;
nfo.probsz = p->sz;
nfo.totalsz = tensor_append(p->vecsz, nfo.probsz);
nfo.pckdsz = verify_pack(nfo.totalsz, 1);
nfo.pckdvecsz = verify_pack(p->vecsz, tensor_sz(nfo.probsz));
e->i = rimpulse(d, impulse_amp, n, vecn, &nfo,
inA, inB, inC, outA, outB, outC, tmp, rounds, tol);
e->l = rlinear(N, &nfo, inA, inB, inC, outA, outB, outC, tmp, rounds,tol);
e->s = t_shift(n, vecn, &nfo, inA, inB, outA, outB, tmp,
rounds, tol, d);
/* grr, verify-lib.c:preserves_input() only works for complex */
if (!p->in_place && !p->destroy_input) {
bench_tensor *totalsz_swap, *pckdsz_swap;
totalsz_swap = tensor_copy_swapio(nfo.totalsz);
pckdsz_swap = tensor_copy_swapio(nfo.pckdsz);
for (i = 0; i < rounds; ++i) {
rarand(inA, N);
dofft(&nfo, inA, outB);
cpyr((R *) nfo.p->in, totalsz_swap, inB, pckdsz_swap);
racmp(inB, inA, N, "preserves_input", 0.0);
}
tensor_destroy(totalsz_swap);
tensor_destroy(pckdsz_swap);
}
tensor_destroy(nfo.totalsz);
tensor_destroy(nfo.pckdsz);
tensor_destroy(nfo.pckdvecsz);
bench_free(tmp);
bench_free(outC);
bench_free(outB);
bench_free(outA);
bench_free(inC);
bench_free(inB);
bench_free(inA);
bench_free(d);
}
void speed(const char *param, int setup_only)
{
double *t;
int iter = 0, k;
bench_problem *p;
double tmin, y;
t = (double *) bench_malloc(time_repeat * sizeof(double));
for (k = 0; k < time_repeat; ++k)
t[k] = 0;
p = problem_parse(param);
BENCH_ASSERT(can_do(p));
if (!no_speed_allocation) {
problem_alloc(p);
problem_zero(p);
}
timer_start(LIBBENCH_TIMER);
setup(p);
p->setup_time = bench_cost_postprocess(timer_stop(LIBBENCH_TIMER));
/* reset the input to zero again, because the planner in paranoid
mode sets it to random values, thus making the benchmark
diverge. */
if (!no_speed_allocation)
problem_zero(p);
if (setup_only)
goto done;
start_over:
for (iter = 1; iter < (1<<30); iter *= 2) {
tmin = 1.0e20;
for (k = 0; k < time_repeat; ++k) {
timer_start(LIBBENCH_TIMER);
doit(iter, p);
y = bench_cost_postprocess(timer_stop(LIBBENCH_TIMER));
if (y < 0) /* yes, it happens */
goto start_over;
t[k] = y;
if (y < tmin)
tmin = y;
}
if (tmin >= time_min)
goto done;
}
goto start_over; /* this also happens */
done:
done(p);
if (iter)
for (k = 0; k < time_repeat; ++k)
t[k] /= iter;
else
for (k = 0; k < time_repeat; ++k)
t[k] = 0;
report(p, t, time_repeat);
if (!no_speed_allocation)
problem_destroy(p);
bench_free(t);
return;
}
int bench_main(int argc, char *argv[])
{
double tmin = 0.0;
double tol;
int repeat = 0;
int rounds = 10;
int iarounds = 0;
int arounds = 1; /* this is too low for precise results */
int c;
int index;
char *short_options = make_short_options(long_options);
check_alignment(&tol);
report = report_verbose; /* default */
verbose = 0;
tol = SINGLE_PRECISION ? 1.0e-3 : 1.0e-10;
bench_srand(1);
while ((c = getopt_long (argc, argv, short_options,
long_options, &index)) != -1) {
switch (c) {
case 't' :
tmin = strtod(optarg, 0);
break;
case 'r':
repeat = atoi(optarg);
break;
case 's':
timer_init(tmin, repeat);
speed(optarg);
break;
case 'd':
report_can_do(optarg);
break;
case 'o':
useropt(optarg);
break;
case 'v':
if (optarg)
verbose = atoi(optarg);
else
++verbose;
break;
case 'y':
verify(optarg, rounds, tol);
break;
case 'a':
accuracy(optarg, arounds, iarounds);
break;
case 'i':
report_info(optarg);
break;
case 'I':
report_info_all();
break;
case 'h':
usage(argv[0], long_options);
break;
case 300: /* --report-mflops */
report = report_mflops;
break;
case 310: /* --report-time */
report = report_time;
break;
case 320: /* --report-benchmark */
report = report_benchmark;
break;
case 330: /* --report-verbose */
report = report_verbose;
break;
case 400: /* --print-time-min */
timer_init(tmin, repeat);
ovtpvt("%g\n", time_min);
break;
case 401: /* --verify-rounds */
rounds = atoi(optarg);
break;
case 402: /* --print-precision */
if (SINGLE_PRECISION)
ovtpvt("single\n");
else if (LDOUBLE_PRECISION)
ovtpvt("long-double\n");
else if (DOUBLE_PRECISION)
ovtpvt("double\n");
else
ovtpvt("unknown %d\n", sizeof(bench_real));
break;
case 403: /* --verify-tolerance */
tol = strtod(optarg, 0);
break;
case 404: /* --random-seed */
bench_srand(atoi(optarg));
break;
case 405: /* --accuracy-rounds */
arounds = atoi(optarg);
break;
case 406: /* --impulse-accuracy-rounds */
iarounds = atoi(optarg);
break;
case '?':
/* `getopt_long' already printed an error message. */
break;
default:
abort ();
}
}
/* assume that any remaining arguments are problems to be
benchmarked */
while (optind < argc) {
timer_init(tmin, repeat);
speed(argv[optind++]);
}
cleanup();
bench_free(short_options);
return 0;
}
void done(struct problem *p)
{
UNUSED(p);
if (p->rank == 1)
bench_free(WSAVE);
}
void tensor_destroy(bench_tensor *sz)
{
bench_free0(sz->dims);
bench_free(sz);
}
static const char *parsetensor(const char *s, bench_tensor **tp,
r2r_kind_t **k)
{
struct dimlist *l = 0, *m;
bench_tensor *t;
int rnk = 0;
L1:
m = (struct dimlist *)bench_malloc(sizeof(struct dimlist));
/* nconc onto l */
m->cdr = l; l = m;
++rnk;
s = parseint(s, &m->car.n);
if (*s == ':') {
/* read input stride */
++s;
s = parseint(s, &m->car.is);
if (*s == ':') {
/* read output stride */
++s;
s = parseint(s, &m->car.os);
} else {
/* default */
m->car.os = m->car.is;
}
} else {
m->car.is = 0;
m->car.os = 0;
}
if (*s == 'f' || *s == 'F') {
m->k = R2R_R2HC;
++s;
}
else if (*s == 'b' || *s == 'B') {
m->k = R2R_HC2R;
++s;
}
else if (*s == 'h' || *s == 'H') {
m->k = R2R_DHT;
++s;
}
else if (*s == 'e' || *s == 'E' || *s == 'o' || *s == 'O') {
char c = *(s++);
int ab;
s = parseint(s, &ab);
if (c == 'e' || c == 'E') {
if (ab == 0)
m->k = R2R_REDFT00;
else if (ab == 1)
m->k = R2R_REDFT01;
else if (ab == 10)
m->k = R2R_REDFT10;
else if (ab == 11)
m->k = R2R_REDFT11;
else
BENCH_ASSERT(0);
}
else {
if (ab == 0)
m->k = R2R_RODFT00;
else if (ab == 1)
m->k = R2R_RODFT01;
else if (ab == 10)
m->k = R2R_RODFT10;
else if (ab == 11)
m->k = R2R_RODFT11;
else
BENCH_ASSERT(0);
}
}
else
m->k = R2R_R2HC;
if (*s == 'x' || *s == 'X') {
++s;
goto L1;
}
/* now we have a dimlist. Build bench_tensor, etc. */
if (k && rnk > 0) {
int i;
*k = (r2r_kind_t *) bench_malloc(sizeof(r2r_kind_t) * rnk);
for (m = l, i = rnk - 1; i >= 0; --i, m = m->cdr) {
BENCH_ASSERT(m);
(*k)[i] = m->k;
}
}
t = mktensor(rnk);
while (--rnk >= 0) {
bench_iodim *d = t->dims + rnk;
BENCH_ASSERT(l);
m = l; l = m->cdr;
d->n = m->car.n;
d->is = m->car.is;
d->os = m->car.os;
bench_free(m);
}
*tp = t;
return s;
}
