1 /* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil ; -*- */
3 * (C) 2012 by Argonne National Laboratory.
4 * See COPYRIGHT in top-level directory.
8 * Tests that basic optimizations are performed on vector of vector datatypes.
9 * As the "leaf" element is a large block (when properly optimized), the
10 * performance of an MPI datatype should be nearly as good (if not better)
11 * than manual packing (the threshold used in this test is *very* forgiving).
12 * This test may be run with one process.
14 * If PACK_IS_NATIVE is defined, MPI_Pack stores exactly the same bytes as the
15 * user would pack manually; in that case, there is a consistency check.
23 #include "mpitestconf.h"
26 /* MPICH (as of 6/2012) packs the native bytes */
27 #define PACK_IS_NATIVE
31 static int verbose = 0;
33 int main(int argc, char **argv)
35 int vcount = 16, vblock = vcount * vcount / 2, vstride = 2 * vcount * vblock;
36 int v2stride, typesize, packsize, i, position, errs = 0;
37 char *inbuf, *outbuf, *outbuf2;
38 MPI_Datatype ft1type, ft2type, ft3type;
39 MPI_Datatype ftopttype;
42 double tpack, tmanual, tpackopt;
45 MPI_Init(&argc, &argv);
47 MPI_Type_contiguous(6, MPI_FLOAT, &ft1type);
48 MPI_Type_size(ft1type, &typesize);
49 v2stride = vcount * vcount * vcount * vcount * typesize;
50 MPI_Type_vector(vcount, vblock, vstride, ft1type, &ft2type);
51 MPI_Type_create_hvector(2, 1, v2stride, ft2type, &ft3type);
52 MPI_Type_commit(&ft3type);
53 MPI_Type_free(&ft1type);
54 MPI_Type_free(&ft2type);
55 #if defined(MPICH) && defined(PRINT_DATATYPE_INTERNALS)
56 /* To use MPIDU_Datatype_debug to print the datatype internals,
57 * you must configure MPICH with --enable-g=log */
59 printf("Original datatype:\n");
60 MPIDU_Datatype_debug(ft3type, 10);
63 /* The same type, but without using the contiguous type */
64 MPI_Type_vector(vcount, 6 * vblock, 6 * vstride, MPI_FLOAT, &ft2type);
65 MPI_Type_create_hvector(2, 1, v2stride, ft2type, &ftopttype);
66 MPI_Type_commit(&ftopttype);
67 MPI_Type_free(&ft2type);
68 #if defined(MPICH) && defined(PRINT_DATATYPE_INTERNALS)
70 printf("\n\nMerged datatype:\n");
71 MPIDU_Datatype_debug(ftopttype, 10);
75 MPI_Type_get_extent(ft3type, &lb, &extent);
76 MPI_Type_size(ft3type, &typesize);
78 MPI_Pack_size(1, ft3type, MPI_COMM_WORLD, &packsize);
80 inbuf = (char *) malloc(extent);
81 outbuf = (char *) malloc(packsize);
82 outbuf2 = (char *) malloc(packsize);
84 fprintf(stderr, "Unable to allocate %ld for inbuf\n", (long) extent);
85 MPI_Abort(MPI_COMM_WORLD, 1);
88 fprintf(stderr, "Unable to allocate %ld for outbuf\n", (long) packsize);
89 MPI_Abort(MPI_COMM_WORLD, 1);
92 fprintf(stderr, "Unable to allocate %ld for outbuf2\n", (long) packsize);
93 MPI_Abort(MPI_COMM_WORLD, 1);
95 for (i = 0; i < extent; i++) {
99 /* Warm up the code and data */
100 MPI_Pack(inbuf, 1, ft3type, outbuf, packsize, &position, MPI_COMM_WORLD);
102 /* Pack using the vector of vector of contiguous */
104 for (ntry = 0; ntry < 5; ntry++) {
107 MPI_Pack(inbuf, 1, ft3type, outbuf, packsize, &position, MPI_COMM_WORLD);
108 t1 = MPI_Wtime() - t0;
112 MPI_Type_free(&ft3type);
114 /* Pack using vector of vector with big blocks (same type map) */
116 for (ntry = 0; ntry < 5; ntry++) {
119 MPI_Pack(inbuf, 1, ftopttype, outbuf, packsize, &position, MPI_COMM_WORLD);
120 t1 = MPI_Wtime() - t0;
124 MPI_Type_free(&ftopttype);
126 /* User (manual) packing code.
127 * Note that we exploit the fact that the vector type contains vblock
128 * instances of a contiguous type of size 24, or equivalently a
129 * single block of 24*vblock bytes.
132 for (ntry = 0; ntry < 5; ntry++) {
133 const char *ppe = (const char *) inbuf;
137 for (k = 0; k < 2; k++) { /* hvector count; blocksize is 1 */
138 const char *ptr = ppe;
139 for (j = 0; j < vcount; j++) { /* vector count */
140 memcpy(outbuf2 + position, ptr, 24 * vblock);
142 position += 24 * vblock;
146 t1 = MPI_Wtime() - t0;
150 /* Check on correctness */
151 #ifdef PACK_IS_NATIVE
152 if (memcmp(outbuf, outbuf2, position) != 0) {
153 printf("Panic - pack buffers differ\n");
159 printf("Bytes packed = %d\n", position);
160 printf("MPI_Pack time = %e, opt version = %e, manual pack time = %e\n",
161 tpack, tpackopt, tmanual);
164 /* A factor of 4 is extremely generous, especially since the test suite
165 * no longer builds any of the tests with optimization */
166 if (4 * tmanual < tpack) {
168 printf("MPI_Pack time = %e, manual pack time = %e\n", tpack, tmanual);
169 printf("MPI_Pack time should be less than 4 times the manual time\n");
170 printf("For most informative results, be sure to compile this test with optimization\n");
172 if (4 * tmanual < tpackopt) {
174 printf("MPI_Pack with opt = %e, manual pack time = %e\n", tpackopt, tmanual);
175 printf("MPI_Pack time should be less than 4 times the manual time\n");
176 printf("For most informative results, be sure to compile this test with optimization\n");
179 printf(" Found %d errors\n", errs);
182 printf(" No Errors\n");