/* * Copyright (c) 1997-1999, 2003 Massachusetts Institute of Technology * * This program is free software; you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation; either version 2 of the License, or * (at your option) any later version. * * This program is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with this program; if not, write to the Free Software * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA * */ #include #include #include #include #include #include "fftw-int.h" #include "rfftw_mpi.h" #include "test_main.h" #define my_printf if (io_okay) printf #define my_fprintf if (io_okay) fprintf #define my_fflush if (io_okay) fflush int ncpus = 1; int my_cpu = 0; char fftw_prefix[] = "rfftw_mpi"; /************************************************* * Speed tests *************************************************/ #define MPI_TIME_FFT(fft,a,n,t) \ { \ double ts,te; \ double total_t; \ int iters = 1,iter; \ zero_arr((n), (a)); \ do { \ MPI_Barrier(MPI_COMM_WORLD); \ ts = MPI_Wtime(); \ for (iter = 0; iter < iters; ++iter) fft; \ te = MPI_Wtime(); \ t = (total_t = (te - ts)) / iters; \ iters *= 2; \ } while (total_t < 2.0); \ } void zero_arr(int n, fftw_real * a) { int i; for (i = 0; i < n; ++i) a[i] = 0.0; } void test_speed_aux(int n, fftw_direction dir, int flags, int specific) { fftw_real *in, *out; fftw_plan plan; double t; fftw_time begin, end; return; /* one-dim transforms not supported yet in MPI */ in = (fftw_real *) fftw_malloc(n * howmany_fields * sizeof(fftw_real)); out = (fftw_real *) fftw_malloc(n * howmany_fields * sizeof(fftw_real)); if (specific) { begin = fftw_get_time(); plan = rfftw_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 = rfftw_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, rfftw_print_plan(plan)); FFTW_TIME_FFT(rfftw(plan, howmany_fields, in, howmany_fields, 1, out, howmany_fields, 1), in, n * howmany_fields, t); rfftw_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\" = 5/2 (n log2 n) / (t in microseconds)" " = %f\n", 0.5 * howmany_fields * mflops(t, n))); fftw_free(in); fftw_free(out); WHEN_VERBOSE(1, printf("\n")); } 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_real *in, *work; rfftwnd_plan plan = 0; rfftwnd_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 - 1; ++i) N *= sz.narray[i]; N *= (sz.narray[i] + 2); if (specific) { return; } else { if (io_okay) plan = rfftwnd_create_plan(sz.rank, sz.narray, dir, speed_flag | flags | wisdom_flag | no_vector_flag); mpi_plan = rfftwnd_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"); rfftwnd_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) in = (fftw_real *) fftw_malloc(N * howmany_fields * sizeof(fftw_real)); else in = (fftw_real *) fftw_malloc(total_local_size * howmany_fields * sizeof(fftw_real)); work = (fftw_real *) fftw_malloc(total_local_size * howmany_fields * sizeof(fftw_real)); if (io_okay) { if (dir == FFTW_REAL_TO_COMPLEX) { FFTW_TIME_FFT(rfftwnd_real_to_complex(plan, howmany_fields, in, howmany_fields, 1, 0, 0, 0), in, N * howmany_fields, t0); } else { FFTW_TIME_FFT(rfftwnd_complex_to_real(plan, howmany_fields, (fftw_complex *) in, howmany_fields, 1, 0, 0, 0), in, N * howmany_fields, t0); } } rfftwnd_destroy_plan(plan); WHEN_VERBOSE(1, my_printf("time for one fft (uniprocessor): %s\n", smart_sprint_time(t0))); MPI_TIME_FFT(rfftwnd_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\" = 5/2 (N log2 N) / " "(t in microseconds)" " = %f\n", 0.5*howmany_fields*mflops(t, N))); WHEN_VERBOSE(1, printf("NORMAL: parallel speedup: %f\n", t0 / t)); } MPI_TIME_FFT(rfftwnd_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\" = 5/2 (N log2 N) / " "(t in microseconds)" " = %f\n", 0.5*howmany_fields*mflops(t, N))); WHEN_VERBOSE(1, printf("TRANSP.: parallel speedup: %f\n", t0 / t)); } MPI_TIME_FFT(rfftwnd_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\" = 5/2 (N log2 N) / " "(t in microseconds)" " = %f\n", 0.5*howmany_fields*mflops(t, N))); WHEN_VERBOSE(1, printf("NORMAL,w/WORK: parallel speedup: %f\n", t0 / t)); } MPI_TIME_FFT(rfftwnd_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\" = 5/2 (N log2 N) / " "(t in microseconds)" " = %f\n", 0.5*howmany_fields*mflops(t, N))); WHEN_VERBOSE(1, printf("TRANSP.,w/WORK: parallel speedup: %f\n", t0 / t)); } rfftwnd_mpi_destroy_plan(mpi_plan); fftw_free(in); fftw_free(work); WHEN_VERBOSE(1, my_printf("\n")); } /************************************************* * correctness tests *************************************************/ double compute_error(fftw_real * A, int astride, fftw_real * B, int bstride, int n) { /* compute the relative error */ double error = 0.0; int i; for (i = 0; i < n; ++i) { double a; double mag; a = fabs(A[i * astride] - B[i * bstride]); mag = 0.5 * (fabs(A[i * astride]) + fabs(B[i * bstride]))+TOLERANCE; a /= mag; if (a > error) error = a; #ifdef HAVE_ISNAN CHECK(!isnan(a), "NaN in answer"); #endif } return error; } 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_in_place(int n, int istride, int howmany, fftw_direction dir, fftw_plan validated_plan, int specific) { /* one-dim. transforms are not supported yet in MPI */ WHEN_VERBOSE(2, my_printf("N/A\n")); } void test_out_of_place_both(int n, int istride, int ostride, int howmany, fftw_plan validated_plan_forward, fftw_plan validated_plan_backward) { } void test_in_place_both(int n, int istride, int howmany, fftw_plan validated_plan_forward, fftw_plan validated_plan_backward) { WHEN_VERBOSE(2, 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, 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_correctness(int n) { } /************************************************* * multi-dimensional correctness tests *************************************************/ void testnd_out_of_place(int rank, int *n, fftwnd_plan validated_plan) { } void testnd_in_place(int rank, int *n, fftwnd_plan validated_plan, int alternate_api, int specific) { int local_nx, local_x_start, local_ny_after_transpose, local_y_start_after_transpose, total_local_size; int istride, ostride, howmany; int N, dim, i, j, k; int nc, nhc, nr; fftw_real *in1, *out3, *work = 0; fftw_complex *in2, *out1, *out2; rfftwnd_mpi_plan p = 0, ip = 0; int flags = measure_flag | wisdom_flag | FFTW_IN_PLACE; if (specific || rank < 2) return; 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; } if (alternate_api && (rank == 2 || rank == 3)) { if (rank == 2) { p = rfftw2d_mpi_create_plan(MPI_COMM_WORLD, n[0], n[1], FFTW_REAL_TO_COMPLEX, flags); ip = rfftw2d_mpi_create_plan(MPI_COMM_WORLD, n[0], n[1], FFTW_COMPLEX_TO_REAL, flags); } else { p = rfftw3d_mpi_create_plan(MPI_COMM_WORLD, n[0], n[1], n[2], FFTW_REAL_TO_COMPLEX, flags); ip = rfftw3d_mpi_create_plan(MPI_COMM_WORLD, n[0], n[1], n[2], FFTW_COMPLEX_TO_REAL, flags); } } else { p = rfftwnd_mpi_create_plan(MPI_COMM_WORLD, rank, n, FFTW_REAL_TO_COMPLEX, flags); ip = rfftwnd_mpi_create_plan(MPI_COMM_WORLD, rank, n, FFTW_COMPLEX_TO_REAL, flags); } CHECK(p != NULL && ip != NULL, "can't create plan"); rfftwnd_mpi_local_sizes(p, &local_nx, &local_x_start, &local_ny_after_transpose, &local_y_start_after_transpose, &total_local_size); in1 = (fftw_real *) fftw_malloc(total_local_size * MAX_STRIDE * sizeof(fftw_real)); if (coinflip()) { WHEN_VERBOSE(1, my_printf("w/work...")); work = (fftw_real *) fftw_malloc(total_local_size * 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)); for (i = 0; i < total_local_size * 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 (i = 0; i < local_nx * (nc / n[0]); ++i) for (j = 0; j < nr; ++j) { for (k = 0; k < istride; ++k) in1[(i * nhc * 2 + j) * istride + k] = c_re((in2 + local_x_start * (N/n[0])) [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)); rfftwnd_mpi(p, howmany, in1, work, FFTW_NORMAL_ORDER); for (i = 0; i < local_nx * (nc / n[0]); ++i) for (k = 0; k < howmany; ++k) CHECK(compute_error_complex(out1 + i * nhc * ostride + k, ostride, out2 + local_x_start*(N/n[0]) + i * nr, 1, nhc) < TOLERANCE, "in-place (r2c): wrong answer"); rfftwnd_mpi(ip, howmany, in1, work, FFTW_NORMAL_ORDER); for (i = 0; i < total_local_size * istride; ++i) out3[i] *= 1.0 / N; for (i = 0; i < local_nx * (nc / n[0]); ++i) for (k = 0; k < howmany; ++k) CHECK(compute_error(out3 + i * nhc * 2 * istride + k, istride, (fftw_real *) (in2 + local_x_start*(N/n[0]) + i * nr), 2, nr) < TOLERANCE, "in-place (c2r): wrong answer (check 2)"); } rfftwnd_mpi_destroy_plan(p); rfftwnd_mpi_destroy_plan(ip); fftw_free(work); fftw_free(out2); fftw_free(in2); fftw_free(in1); } void testnd_correctness(struct size sz, fftw_direction dir, int alt_api, int specific, int force_buf) { fftwnd_plan validated_plan; if (dir != FFTW_FORWARD) return; if (force_buf) return; validated_plan = fftwnd_create_plan(sz.rank, sz.narray, dir, measure_flag | wisdom_flag); CHECK(validated_plan != NULL, "can't create plan"); testnd_in_place(sz.rank, sz.narray, validated_plan, alt_api, specific); fftwnd_destroy_plan(validated_plan); } /************************************************* * planner tests *************************************************/ 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; rfftwnd_mpi_plan pnd[PLANNER_TEST_SIZE]; int *narr, maxdim; chk_mem_leak = 0; verbose--; please_wait(); if (rank < 2) rank = 2; narr = (int *) fftw_malloc(rank * sizeof(int)); for (i = 0; i < PLANNER_TEST_SIZE; ++i) { pnd[i] = (rfftwnd_mpi_plan) 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 (pnd[r]) rfftwnd_mpi_destroy_plan(pnd[r]); pnd[r] = rfftwnd_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 (pnd[i]) rfftwnd_mpi_destroy_plan(pnd[i]); } fftw_free(narr); verbose++; chk_mem_leak = 1; } /************************************************* * test initialization *************************************************/ void test_init(int *argc, char ***argv) { unsigned int seed; MPI_Init(argc,argv); MPI_Comm_size(MPI_COMM_WORLD,&ncpus); MPI_Comm_rank(MPI_COMM_WORLD,&my_cpu); /* Only process 0 gets to do I/O: */ io_okay = my_cpu == 0; /* Make sure all processes use the same seed for random numbers: */ seed = time(NULL); MPI_Bcast(&seed, 1, MPI_INT, 0, MPI_COMM_WORLD); srand(seed); fftw_die_hook = fftw_mpi_die; /* call MPI_Abort on failure */ } void test_finish(void) { MPI_Finalize(); } void enter_paranoid_mode(void) { } /* in MPI, only process 0 is guaranteed to have access to the argument list */ int get_option(int argc, char **argv, char *argval, int argval_maxlen) { int c; int arglen; if (io_okay) { c = default_get_option(argc,argv,argval,argval_maxlen); arglen = strlen(argval) + 1; } MPI_Bcast(&c, 1, MPI_INT, 0, MPI_COMM_WORLD); MPI_Bcast(&arglen, 1, MPI_INT, 0, MPI_COMM_WORLD); MPI_Bcast(argval, arglen, MPI_CHAR, 0, MPI_COMM_WORLD); return c; }