10b1450ccSAleksandar Markovic /*
20b1450ccSAleksandar Markovic  *  Test program for MSA instruction MSUBV.D
30b1450ccSAleksandar Markovic  *
40b1450ccSAleksandar Markovic  *  Copyright (C) 2019  Wave Computing, Inc.
50b1450ccSAleksandar Markovic  *  Copyright (C) 2019  Aleksandar Markovic <amarkovic@wavecomp.com>
60b1450ccSAleksandar Markovic  *
70b1450ccSAleksandar Markovic  *  This program is free software: you can redistribute it and/or modify
80b1450ccSAleksandar Markovic  *  it under the terms of the GNU General Public License as published by
90b1450ccSAleksandar Markovic  *  the Free Software Foundation, either version 2 of the License, or
100b1450ccSAleksandar Markovic  *  (at your option) any later version.
110b1450ccSAleksandar Markovic  *`
120b1450ccSAleksandar Markovic  *  This program is distributed in the hope that it will be useful,
130b1450ccSAleksandar Markovic  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
140b1450ccSAleksandar Markovic  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
150b1450ccSAleksandar Markovic  *  GNU General Public License for more details.
160b1450ccSAleksandar Markovic  *
170b1450ccSAleksandar Markovic  *  You should have received a copy of the GNU General Public License
180b1450ccSAleksandar Markovic  *  along with this program.  If not, see <https://www.gnu.org/licenses/>.
190b1450ccSAleksandar Markovic  *
200b1450ccSAleksandar Markovic  */
210b1450ccSAleksandar Markovic 
220b1450ccSAleksandar Markovic #include <sys/time.h>
230b1450ccSAleksandar Markovic #include <stdint.h>
240b1450ccSAleksandar Markovic 
250b1450ccSAleksandar Markovic #include "../../../../include/wrappers_msa.h"
260b1450ccSAleksandar Markovic #include "../../../../include/test_inputs_128.h"
270b1450ccSAleksandar Markovic #include "../../../../include/test_utils_128.h"
280b1450ccSAleksandar Markovic 
290b1450ccSAleksandar Markovic #define TEST_COUNT_TOTAL (                                                \
300b1450ccSAleksandar Markovic             (PATTERN_INPUTS_SHORT_COUNT) * (PATTERN_INPUTS_SHORT_COUNT) + \
310b1450ccSAleksandar Markovic             3 * (RANDOM_INPUTS_SHORT_COUNT) * (RANDOM_INPUTS_SHORT_COUNT))
320b1450ccSAleksandar Markovic 
330b1450ccSAleksandar Markovic 
main(void)340b1450ccSAleksandar Markovic int32_t main(void)
350b1450ccSAleksandar Markovic {
360b1450ccSAleksandar Markovic     char *isa_ase_name = "MSA";
370b1450ccSAleksandar Markovic     char *group_name = "Int Multiply";
380b1450ccSAleksandar Markovic     char *instruction_name =  "MSUBV.D";
390b1450ccSAleksandar Markovic     int32_t ret;
400b1450ccSAleksandar Markovic     uint32_t i, j;
410b1450ccSAleksandar Markovic     struct timeval start, end;
420b1450ccSAleksandar Markovic     double elapsed_time;
430b1450ccSAleksandar Markovic 
440b1450ccSAleksandar Markovic     uint64_t b128_result[TEST_COUNT_TOTAL][2];
450b1450ccSAleksandar Markovic     uint64_t b128_expect[TEST_COUNT_TOTAL][2] = {
46*391f77aaSAleksandar Markovic         { 0xffffffffffffffffULL, 0xffffffffffffffffULL, },    /*   0  */
47*391f77aaSAleksandar Markovic         { 0xffffffffffffffffULL, 0xffffffffffffffffULL, },
48*391f77aaSAleksandar Markovic         { 0xaaaaaaaaaaaaaaa9ULL, 0xaaaaaaaaaaaaaaa9ULL, },
49*391f77aaSAleksandar Markovic         { 0xfffffffffffffffeULL, 0xfffffffffffffffeULL, },
50*391f77aaSAleksandar Markovic         { 0xcccccccccccccccaULL, 0xcccccccccccccccaULL, },
51*391f77aaSAleksandar Markovic         { 0xfffffffffffffffdULL, 0xfffffffffffffffdULL, },
52*391f77aaSAleksandar Markovic         { 0xe38e38e38e38e38bULL, 0x38e38e38e38e38e0ULL, },
53*391f77aaSAleksandar Markovic         { 0xfffffffffffffffcULL, 0xfffffffffffffffcULL, },
54*391f77aaSAleksandar Markovic         { 0xfffffffffffffffcULL, 0xfffffffffffffffcULL, },    /*   8  */
55*391f77aaSAleksandar Markovic         { 0xfffffffffffffffcULL, 0xfffffffffffffffcULL, },
56*391f77aaSAleksandar Markovic         { 0xfffffffffffffffcULL, 0xfffffffffffffffcULL, },
57*391f77aaSAleksandar Markovic         { 0xfffffffffffffffcULL, 0xfffffffffffffffcULL, },
58*391f77aaSAleksandar Markovic         { 0xfffffffffffffffcULL, 0xfffffffffffffffcULL, },
59*391f77aaSAleksandar Markovic         { 0xfffffffffffffffcULL, 0xfffffffffffffffcULL, },
60*391f77aaSAleksandar Markovic         { 0xfffffffffffffffcULL, 0xfffffffffffffffcULL, },
61*391f77aaSAleksandar Markovic         { 0xfffffffffffffffcULL, 0xfffffffffffffffcULL, },
62*391f77aaSAleksandar Markovic         { 0xaaaaaaaaaaaaaaa6ULL, 0xaaaaaaaaaaaaaaa6ULL, },    /*  16  */
63*391f77aaSAleksandar Markovic         { 0xaaaaaaaaaaaaaaa6ULL, 0xaaaaaaaaaaaaaaa6ULL, },
64*391f77aaSAleksandar Markovic         { 0x71c71c71c71c71c2ULL, 0x71c71c71c71c71c2ULL, },
65*391f77aaSAleksandar Markovic         { 0x5555555555555550ULL, 0x5555555555555550ULL, },
66*391f77aaSAleksandar Markovic         { 0xddddddddddddddd8ULL, 0xddddddddddddddd8ULL, },
67*391f77aaSAleksandar Markovic         { 0xfffffffffffffffaULL, 0xfffffffffffffffaULL, },
68*391f77aaSAleksandar Markovic         { 0xed097b425ed097aeULL, 0xd097b425ed097b3cULL, },
69*391f77aaSAleksandar Markovic         { 0xaaaaaaaaaaaaaaa4ULL, 0xaaaaaaaaaaaaaaa4ULL, },
70*391f77aaSAleksandar Markovic         { 0xfffffffffffffff9ULL, 0xfffffffffffffff9ULL, },    /*  24  */
71*391f77aaSAleksandar Markovic         { 0xfffffffffffffff9ULL, 0xfffffffffffffff9ULL, },
72*391f77aaSAleksandar Markovic         { 0xe38e38e38e38e387ULL, 0xe38e38e38e38e387ULL, },
73*391f77aaSAleksandar Markovic         { 0x555555555555554eULL, 0x555555555555554eULL, },
74*391f77aaSAleksandar Markovic         { 0x9999999999999992ULL, 0x9999999999999992ULL, },
75*391f77aaSAleksandar Markovic         { 0xaaaaaaaaaaaaaaa3ULL, 0xaaaaaaaaaaaaaaa3ULL, },
76*391f77aaSAleksandar Markovic         { 0xa12f684bda12f67dULL, 0x12f684bda12f6844ULL, },
77*391f77aaSAleksandar Markovic         { 0xfffffffffffffff8ULL, 0xfffffffffffffff8ULL, },
78*391f77aaSAleksandar Markovic         { 0xccccccccccccccc4ULL, 0xccccccccccccccc4ULL, },    /*  32  */
79*391f77aaSAleksandar Markovic         { 0xccccccccccccccc4ULL, 0xccccccccccccccc4ULL, },
80*391f77aaSAleksandar Markovic         { 0x555555555555554cULL, 0x555555555555554cULL, },
81*391f77aaSAleksandar Markovic         { 0x9999999999999990ULL, 0x9999999999999990ULL, },
82*391f77aaSAleksandar Markovic         { 0xa3d70a3d70a3d700ULL, 0xa3d70a3d70a3d700ULL, },
83*391f77aaSAleksandar Markovic         { 0x666666666666665cULL, 0x666666666666665cULL, },
84*391f77aaSAleksandar Markovic         { 0xe93e93e93e93e934ULL, 0x2d82d82d82d82d78ULL, },
85*391f77aaSAleksandar Markovic         { 0x3333333333333328ULL, 0x3333333333333328ULL, },
86*391f77aaSAleksandar Markovic         { 0x666666666666665bULL, 0x666666666666665bULL, },    /*  40  */
87*391f77aaSAleksandar Markovic         { 0x666666666666665bULL, 0x666666666666665bULL, },
88*391f77aaSAleksandar Markovic         { 0x888888888888887dULL, 0x888888888888887dULL, },
89*391f77aaSAleksandar Markovic         { 0x999999999999998eULL, 0x999999999999998eULL, },
90*391f77aaSAleksandar Markovic         { 0x5c28f5c28f5c28eaULL, 0x5c28f5c28f5c28eaULL, },
91*391f77aaSAleksandar Markovic         { 0xccccccccccccccc1ULL, 0xccccccccccccccc1ULL, },
92*391f77aaSAleksandar Markovic         { 0x2d82d82d82d82d77ULL, 0x3e93e93e93e93e88ULL, },
93*391f77aaSAleksandar Markovic         { 0xfffffffffffffff4ULL, 0xfffffffffffffff4ULL, },
94*391f77aaSAleksandar Markovic         { 0xe38e38e38e38e382ULL, 0x38e38e38e38e38d7ULL, },    /*  48  */
95*391f77aaSAleksandar Markovic         { 0xe38e38e38e38e382ULL, 0x38e38e38e38e38d7ULL, },
96*391f77aaSAleksandar Markovic         { 0xd097b425ed097b36ULL, 0x097b425ed097b419ULL, },
97*391f77aaSAleksandar Markovic         { 0xc71c71c71c71c710ULL, 0x71c71c71c71c71baULL, },
98*391f77aaSAleksandar Markovic         { 0x49f49f49f49f49e8ULL, 0x38e38e38e38e38d6ULL, },
99*391f77aaSAleksandar Markovic         { 0xaaaaaaaaaaaaaa9eULL, 0xaaaaaaaaaaaaaa9dULL, },
100*391f77aaSAleksandar Markovic         { 0xf9add3c0ca4587daULL, 0x587e6b74f0329154ULL, },
101*391f77aaSAleksandar Markovic         { 0x8e38e38e38e38e2cULL, 0xe38e38e38e38e380ULL, },
102*391f77aaSAleksandar Markovic         { 0xaaaaaaaaaaaaaa9dULL, 0xaaaaaaaaaaaaaa9cULL, },    /*  56  */
103*391f77aaSAleksandar Markovic         { 0xaaaaaaaaaaaaaa9dULL, 0xaaaaaaaaaaaaaa9cULL, },
104*391f77aaSAleksandar Markovic         { 0x684bda12f684bd93ULL, 0x84bda12f684bda04ULL, },
105*391f77aaSAleksandar Markovic         { 0xc71c71c71c71c70eULL, 0x71c71c71c71c71b8ULL, },
106*391f77aaSAleksandar Markovic         { 0x1111111111111102ULL, 0x7777777777777768ULL, },
107*391f77aaSAleksandar Markovic         { 0xe38e38e38e38e37fULL, 0x38e38e38e38e38d4ULL, },
108*391f77aaSAleksandar Markovic         { 0x781948b0fcd6e9d1ULL, 0xc3f35ba781948b00ULL, },
109*391f77aaSAleksandar Markovic         { 0xfffffffffffffff0ULL, 0xfffffffffffffff0ULL, },
110*391f77aaSAleksandar Markovic         { 0x52ba41969e9c6ff0ULL, 0xcd6802158b677f60ULL, },    /*  64  */
111*391f77aaSAleksandar Markovic         { 0x63129bf5b78505f0ULL, 0x1556f7f61c4e5b90ULL, },
112*391f77aaSAleksandar Markovic         { 0x5a4c8855f350a5f0ULL, 0x6a36586fc42edea0ULL, },
113*391f77aaSAleksandar Markovic         { 0x5e6b001b04d82c70ULL, 0xe819332c365e3f20ULL, },
114*391f77aaSAleksandar Markovic         { 0x6ec35a7a1dc0c270ULL, 0x3008290cc7451b50ULL, },
115*391f77aaSAleksandar Markovic         { 0x37152f411fd35230ULL, 0xc7e3b2957c56b340ULL, },
116*391f77aaSAleksandar Markovic         { 0xcc49f1d861667630ULL, 0x1808e0646811cb90ULL, },
117*391f77aaSAleksandar Markovic         { 0xde8a7f544022c1c0ULL, 0x9886bc9978437610ULL, },
118*391f77aaSAleksandar Markovic         { 0xd5c46bb47bee61c0ULL, 0xed661d132023f920ULL, },    /*  72  */
119*391f77aaSAleksandar Markovic         { 0x6af92e4bbd8185c0ULL, 0x3d8b4ae20bdf1170ULL, },
120*391f77aaSAleksandar Markovic         { 0xe4d44869d87d45c0ULL, 0x6409d23bd9c847e0ULL, },
121*391f77aaSAleksandar Markovic         { 0x6e2e9ce94e99c4c0ULL, 0xc30837db04ed7360ULL, },
122*391f77aaSAleksandar Markovic         { 0x724d14ae60214b40ULL, 0x40eb1297771cd3e0ULL, },
123*391f77aaSAleksandar Markovic         { 0x848da22a3edd96d0ULL, 0xc168eecc874e7e60ULL, },
124*391f77aaSAleksandar Markovic         { 0x0de7f6a9b4fa15d0ULL, 0x2067546bb273a9e0ULL, },
125*391f77aaSAleksandar Markovic         { 0xc233bfd40310460cULL, 0x0d9585bacf54c5e0ULL, },
126*391f77aaSAleksandar Markovic         { 0x061015122724c70cULL, 0x0169d01f7cb17f60ULL, },    /*  80  */
127*391f77aaSAleksandar Markovic         { 0x23dacc726f603aacULL, 0xf3ea8c4eaa8b5ce0ULL, },
128*391f77aaSAleksandar Markovic         { 0xd82df953c25380acULL, 0xba87b7f0f99bbb60ULL, },
129*391f77aaSAleksandar Markovic         { 0x546cb94a0c5e7444ULL, 0x3818c320ce1bdf60ULL, },
130*391f77aaSAleksandar Markovic         { 0xa38f9428761ecf44ULL, 0x63113b9e681b66e0ULL, },
131*391f77aaSAleksandar Markovic         { 0x7dc23fbe59fe7924ULL, 0x156ddd68750e6260ULL, },
132*391f77aaSAleksandar Markovic         { 0x8a17717d36df5b24ULL, 0x36b1f5939596d2e0ULL, },
133*391f77aaSAleksandar Markovic         { 0x7e854cd9a677ce2cULL, 0xf2b6202eb36946e0ULL, },
134*391f77aaSAleksandar Markovic         { 0x246d8d067437a72cULL, 0x04c6347e9c1ff460ULL, },    /*  88  */
135*391f77aaSAleksandar Markovic         { 0xc48a013a554339ccULL, 0xcb81fd31acc4a5e0ULL, },
136*391f77aaSAleksandar Markovic         { 0xb971282c0b508fccULL, 0x20d62d6344ce5060ULL, },
137*391f77aaSAleksandar Markovic         { 0x835f812f0bc6a7a4ULL, 0x17bd6b5a08275460ULL, },
138*391f77aaSAleksandar Markovic         { 0xc0ee1b9557ab4aa4ULL, 0x170471a9d22d5fe0ULL, },
139*391f77aaSAleksandar Markovic         { 0xc6f66d89431f7984ULL, 0x5c6f5a646cad3f60ULL, },
140*391f77aaSAleksandar Markovic         { 0x5ae0b289f6ac0b84ULL, 0x6f9f6bc81fdb6be0ULL, },
141*391f77aaSAleksandar Markovic         { 0x2f584ee03fd2014cULL, 0xa7e34ccbd1bc3fe0ULL, },
142*391f77aaSAleksandar Markovic         { 0x5947927731cb724cULL, 0xf76af1f9a05f4160ULL, },    /*  96  */
143*391f77aaSAleksandar Markovic         { 0x68112ad490e3a34cULL, 0x7f944a22f5d630e0ULL, },
144*391f77aaSAleksandar Markovic         { 0x1cf6705c5faa944cULL, 0x801292d47291e660ULL, },
145*391f77aaSAleksandar Markovic         { 0x5519f2782cb0454cULL, 0x3d691c2dd53919e0ULL, },
146*391f77aaSAleksandar Markovic         { 0xe5c979861aac06ecULL, 0x585247d6e899e160ULL, },
147*391f77aaSAleksandar Markovic         { 0x2450b27896665b8cULL, 0x8276d8ad504f46e0ULL, },
148*391f77aaSAleksandar Markovic         { 0x2716d456a4a5ab2cULL, 0x46e1f3460c71c260ULL, },
149*391f77aaSAleksandar Markovic         { 0x5751460331251dccULL, 0xdc1dc7a4a693abe0ULL, },
150*391f77aaSAleksandar Markovic         { 0x3bf387b7f37473ccULL, 0x8efb4ff7cc92de60ULL, },    /* 104  */
151*391f77aaSAleksandar Markovic         { 0xc3103a3df066c9ccULL, 0x7d3b07351cd59ee0ULL, },
152*391f77aaSAleksandar Markovic         { 0x0d612554557c1fccULL, 0x5dbabfc2ac8ed560ULL, },
153*391f77aaSAleksandar Markovic         { 0x1cd018ef103475ccULL, 0xca277277956f49e0ULL, },
154*391f77aaSAleksandar Markovic         { 0x15d520225c2e79a4ULL, 0x08f2025804e95de0ULL, },
155*391f77aaSAleksandar Markovic         { 0x820f9c65be3ea1acULL, 0x37094edbda6ef1e0ULL, },
156*391f77aaSAleksandar Markovic         { 0x0f18515c62838744ULL, 0xcfbd4b5627d005e0ULL, },
157*391f77aaSAleksandar Markovic         { 0x11d549f26502488cULL, 0x8de999d53cdc99e0ULL, },
1580b1450ccSAleksandar Markovic     };
1590b1450ccSAleksandar Markovic 
1600b1450ccSAleksandar Markovic     reset_msa_registers();
1610b1450ccSAleksandar Markovic 
1620b1450ccSAleksandar Markovic     gettimeofday(&start, NULL);
1630b1450ccSAleksandar Markovic 
1640b1450ccSAleksandar Markovic     for (i = 0; i < PATTERN_INPUTS_SHORT_COUNT; i++) {
1650b1450ccSAleksandar Markovic         for (j = 0; j < PATTERN_INPUTS_SHORT_COUNT; j++) {
1660b1450ccSAleksandar Markovic             do_msa_MSUBV_D(b128_pattern[i], b128_pattern[j],
1670b1450ccSAleksandar Markovic                            b128_result[PATTERN_INPUTS_SHORT_COUNT * i + j]);
1680b1450ccSAleksandar Markovic         }
1690b1450ccSAleksandar Markovic     }
1700b1450ccSAleksandar Markovic 
1710b1450ccSAleksandar Markovic     for (i = 0; i < RANDOM_INPUTS_SHORT_COUNT; i++) {
1720b1450ccSAleksandar Markovic         for (j = 0; j < RANDOM_INPUTS_SHORT_COUNT; j++) {
1730b1450ccSAleksandar Markovic             do_msa_MSUBV_D(b128_random[i], b128_random[j],
1740b1450ccSAleksandar Markovic                            b128_result[((PATTERN_INPUTS_SHORT_COUNT) *
1750b1450ccSAleksandar Markovic                                         (PATTERN_INPUTS_SHORT_COUNT)) +
1760b1450ccSAleksandar Markovic                                        RANDOM_INPUTS_SHORT_COUNT * i + j]);
1770b1450ccSAleksandar Markovic         }
1780b1450ccSAleksandar Markovic     }
1790b1450ccSAleksandar Markovic 
1800b1450ccSAleksandar Markovic     for (i = 0; i < RANDOM_INPUTS_SHORT_COUNT; i++) {
1810b1450ccSAleksandar Markovic         for (j = 0; j < RANDOM_INPUTS_SHORT_COUNT; j++) {
1820b1450ccSAleksandar Markovic             do_msa_MSUBV_D__DDT(b128_random[i], b128_random[j],
1830b1450ccSAleksandar Markovic                                 b128_result[
1840b1450ccSAleksandar Markovic                                     ((PATTERN_INPUTS_SHORT_COUNT) *
1850b1450ccSAleksandar Markovic                                      (PATTERN_INPUTS_SHORT_COUNT)) +
1860b1450ccSAleksandar Markovic                                     ((RANDOM_INPUTS_SHORT_COUNT) *
1870b1450ccSAleksandar Markovic                                      (RANDOM_INPUTS_SHORT_COUNT)) +
1880b1450ccSAleksandar Markovic                                     RANDOM_INPUTS_SHORT_COUNT * i + j]);
1890b1450ccSAleksandar Markovic         }
1900b1450ccSAleksandar Markovic     }
1910b1450ccSAleksandar Markovic 
1920b1450ccSAleksandar Markovic     for (i = 0; i < RANDOM_INPUTS_SHORT_COUNT; i++) {
1930b1450ccSAleksandar Markovic         for (j = 0; j < RANDOM_INPUTS_SHORT_COUNT; j++) {
1940b1450ccSAleksandar Markovic             do_msa_MSUBV_D__DSD(b128_random[i], b128_random[j],
1950b1450ccSAleksandar Markovic                                 b128_result[
1960b1450ccSAleksandar Markovic                                     ((PATTERN_INPUTS_SHORT_COUNT) *
1970b1450ccSAleksandar Markovic                                      (PATTERN_INPUTS_SHORT_COUNT)) +
1980b1450ccSAleksandar Markovic                                     (2 * (RANDOM_INPUTS_SHORT_COUNT) *
1990b1450ccSAleksandar Markovic                                      (RANDOM_INPUTS_SHORT_COUNT)) +
2000b1450ccSAleksandar Markovic                                     RANDOM_INPUTS_SHORT_COUNT * i + j]);
2010b1450ccSAleksandar Markovic         }
2020b1450ccSAleksandar Markovic     }
2030b1450ccSAleksandar Markovic 
2040b1450ccSAleksandar Markovic     gettimeofday(&end, NULL);
2050b1450ccSAleksandar Markovic 
2060b1450ccSAleksandar Markovic     elapsed_time = (end.tv_sec - start.tv_sec) * 1000.0;
2070b1450ccSAleksandar Markovic     elapsed_time += (end.tv_usec - start.tv_usec) / 1000.0;
2080b1450ccSAleksandar Markovic 
2090b1450ccSAleksandar Markovic     ret = check_results_128(isa_ase_name, group_name, instruction_name,
2100b1450ccSAleksandar Markovic                             TEST_COUNT_TOTAL, elapsed_time,
2110b1450ccSAleksandar Markovic                             &b128_result[0][0], &b128_expect[0][0]);
2120b1450ccSAleksandar Markovic 
2130b1450ccSAleksandar Markovic     return ret;
2140b1450ccSAleksandar Markovic }
215