10b1450ccSAleksandar Markovic /*
20b1450ccSAleksandar Markovic  *  Test program for MSA instruction MSUBV.B
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.B";
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         { 0xa9a9a9a9a9a9a9a9ULL, 0xa9a9a9a9a9a9a9a9ULL, },
49*391f77aaSAleksandar Markovic         { 0xfefefefefefefefeULL, 0xfefefefefefefefeULL, },
50*391f77aaSAleksandar Markovic         { 0xcacacacacacacacaULL, 0xcacacacacacacacaULL, },
51*391f77aaSAleksandar Markovic         { 0xfdfdfdfdfdfdfdfdULL, 0xfdfdfdfdfdfdfdfdULL, },
52*391f77aaSAleksandar Markovic         { 0xe08b35e08b35e08bULL, 0x35e08b35e08b35e0ULL, },
53*391f77aaSAleksandar Markovic         { 0xfcfcfcfcfcfcfcfcULL, 0xfcfcfcfcfcfcfcfcULL, },
54*391f77aaSAleksandar Markovic         { 0xfcfcfcfcfcfcfcfcULL, 0xfcfcfcfcfcfcfcfcULL, },    /*   8  */
55*391f77aaSAleksandar Markovic         { 0xfcfcfcfcfcfcfcfcULL, 0xfcfcfcfcfcfcfcfcULL, },
56*391f77aaSAleksandar Markovic         { 0xfcfcfcfcfcfcfcfcULL, 0xfcfcfcfcfcfcfcfcULL, },
57*391f77aaSAleksandar Markovic         { 0xfcfcfcfcfcfcfcfcULL, 0xfcfcfcfcfcfcfcfcULL, },
58*391f77aaSAleksandar Markovic         { 0xfcfcfcfcfcfcfcfcULL, 0xfcfcfcfcfcfcfcfcULL, },
59*391f77aaSAleksandar Markovic         { 0xfcfcfcfcfcfcfcfcULL, 0xfcfcfcfcfcfcfcfcULL, },
60*391f77aaSAleksandar Markovic         { 0xfcfcfcfcfcfcfcfcULL, 0xfcfcfcfcfcfcfcfcULL, },
61*391f77aaSAleksandar Markovic         { 0xfcfcfcfcfcfcfcfcULL, 0xfcfcfcfcfcfcfcfcULL, },
62*391f77aaSAleksandar Markovic         { 0xa6a6a6a6a6a6a6a6ULL, 0xa6a6a6a6a6a6a6a6ULL, },    /*  16  */
63*391f77aaSAleksandar Markovic         { 0xa6a6a6a6a6a6a6a6ULL, 0xa6a6a6a6a6a6a6a6ULL, },
64*391f77aaSAleksandar Markovic         { 0xc2c2c2c2c2c2c2c2ULL, 0xc2c2c2c2c2c2c2c2ULL, },
65*391f77aaSAleksandar Markovic         { 0x5050505050505050ULL, 0x5050505050505050ULL, },
66*391f77aaSAleksandar Markovic         { 0xd8d8d8d8d8d8d8d8ULL, 0xd8d8d8d8d8d8d8d8ULL, },
67*391f77aaSAleksandar Markovic         { 0xfafafafafafafafaULL, 0xfafafafafafafafaULL, },
68*391f77aaSAleksandar Markovic         { 0x3caeca3caeca3caeULL, 0xca3caeca3caeca3cULL, },
69*391f77aaSAleksandar Markovic         { 0xa4a4a4a4a4a4a4a4ULL, 0xa4a4a4a4a4a4a4a4ULL, },
70*391f77aaSAleksandar Markovic         { 0xf9f9f9f9f9f9f9f9ULL, 0xf9f9f9f9f9f9f9f9ULL, },    /*  24  */
71*391f77aaSAleksandar Markovic         { 0xf9f9f9f9f9f9f9f9ULL, 0xf9f9f9f9f9f9f9f9ULL, },
72*391f77aaSAleksandar Markovic         { 0x8787878787878787ULL, 0x8787878787878787ULL, },
73*391f77aaSAleksandar Markovic         { 0x4e4e4e4e4e4e4e4eULL, 0x4e4e4e4e4e4e4e4eULL, },
74*391f77aaSAleksandar Markovic         { 0x9292929292929292ULL, 0x9292929292929292ULL, },
75*391f77aaSAleksandar Markovic         { 0xa3a3a3a3a3a3a3a3ULL, 0xa3a3a3a3a3a3a3a3ULL, },
76*391f77aaSAleksandar Markovic         { 0x447d0b447d0b447dULL, 0x0b447d0b447d0b44ULL, },
77*391f77aaSAleksandar Markovic         { 0xf8f8f8f8f8f8f8f8ULL, 0xf8f8f8f8f8f8f8f8ULL, },
78*391f77aaSAleksandar Markovic         { 0xc4c4c4c4c4c4c4c4ULL, 0xc4c4c4c4c4c4c4c4ULL, },    /*  32  */
79*391f77aaSAleksandar Markovic         { 0xc4c4c4c4c4c4c4c4ULL, 0xc4c4c4c4c4c4c4c4ULL, },
80*391f77aaSAleksandar Markovic         { 0x4c4c4c4c4c4c4c4cULL, 0x4c4c4c4c4c4c4c4cULL, },
81*391f77aaSAleksandar Markovic         { 0x9090909090909090ULL, 0x9090909090909090ULL, },
82*391f77aaSAleksandar Markovic         { 0x0000000000000000ULL, 0x0000000000000000ULL, },
83*391f77aaSAleksandar Markovic         { 0x5c5c5c5c5c5c5c5cULL, 0x5c5c5c5c5c5c5c5cULL, },
84*391f77aaSAleksandar Markovic         { 0x7834bc7834bc7834ULL, 0xbc7834bc7834bc78ULL, },
85*391f77aaSAleksandar Markovic         { 0x2828282828282828ULL, 0x2828282828282828ULL, },
86*391f77aaSAleksandar Markovic         { 0x5b5b5b5b5b5b5b5bULL, 0x5b5b5b5b5b5b5b5bULL, },    /*  40  */
87*391f77aaSAleksandar Markovic         { 0x5b5b5b5b5b5b5b5bULL, 0x5b5b5b5b5b5b5b5bULL, },
88*391f77aaSAleksandar Markovic         { 0x7d7d7d7d7d7d7d7dULL, 0x7d7d7d7d7d7d7d7dULL, },
89*391f77aaSAleksandar Markovic         { 0x8e8e8e8e8e8e8e8eULL, 0x8e8e8e8e8e8e8e8eULL, },
90*391f77aaSAleksandar Markovic         { 0xeaeaeaeaeaeaeaeaULL, 0xeaeaeaeaeaeaeaeaULL, },
91*391f77aaSAleksandar Markovic         { 0xc1c1c1c1c1c1c1c1ULL, 0xc1c1c1c1c1c1c1c1ULL, },
92*391f77aaSAleksandar Markovic         { 0x8877998877998877ULL, 0x9988779988779988ULL, },
93*391f77aaSAleksandar Markovic         { 0xf4f4f4f4f4f4f4f4ULL, 0xf4f4f4f4f4f4f4f4ULL, },
94*391f77aaSAleksandar Markovic         { 0xd7822cd7822cd782ULL, 0x2cd7822cd7822cd7ULL, },    /*  48  */
95*391f77aaSAleksandar Markovic         { 0xd7822cd7822cd782ULL, 0x2cd7822cd7822cd7ULL, },
96*391f77aaSAleksandar Markovic         { 0x1936fc1936fc1936ULL, 0xfc1936fc1936fc19ULL, },
97*391f77aaSAleksandar Markovic         { 0xba1064ba1064ba10ULL, 0x64ba1064ba1064baULL, },
98*391f77aaSAleksandar Markovic         { 0xd6e8c4d6e8c4d6e8ULL, 0xc4d6e8c4d6e8c4d6ULL, },
99*391f77aaSAleksandar Markovic         { 0x9d9e9c9d9e9c9d9eULL, 0x9c9d9e9c9d9e9c9dULL, },
100*391f77aaSAleksandar Markovic         { 0x54da5c54da5c54daULL, 0x5c54da5c54da5c54ULL, },
101*391f77aaSAleksandar Markovic         { 0x802cd4802cd4802cULL, 0xd4802cd4802cd480ULL, },
102*391f77aaSAleksandar Markovic         { 0x9c9d9b9c9d9b9c9dULL, 0x9b9c9d9b9c9d9b9cULL, },    /*  56  */
103*391f77aaSAleksandar Markovic         { 0x9c9d9b9c9d9b9c9dULL, 0x9b9c9d9b9c9d9b9cULL, },
104*391f77aaSAleksandar Markovic         { 0x0493750493750493ULL, 0x7504937504937504ULL, },
105*391f77aaSAleksandar Markovic         { 0xb80e62b80e62b80eULL, 0x62b80e62b80e62b8ULL, },
106*391f77aaSAleksandar Markovic         { 0x6802ce6802ce6802ULL, 0xce6802ce6802ce68ULL, },
107*391f77aaSAleksandar Markovic         { 0xd47f29d47f29d47fULL, 0x29d47f29d47f29d4ULL, },
108*391f77aaSAleksandar Markovic         { 0x00d1a100d1a100d1ULL, 0xa100d1a100d1a100ULL, },
109*391f77aaSAleksandar Markovic         { 0xf0f0f0f0f0f0f0f0ULL, 0xf0f0f0f0f0f0f0f0ULL, },
110*391f77aaSAleksandar Markovic         { 0xb00c4c60b06cb7f0ULL, 0xf77f776cecd7f060ULL, },    /*  64  */
111*391f77aaSAleksandar Markovic         { 0x58604c7ca826a4f0ULL, 0xb11e6ee016929090ULL, },
112*391f77aaSAleksandar Markovic         { 0xf81cf804c0e87df0ULL, 0x4436ec3e6ce920a0ULL, },
113*391f77aaSAleksandar Markovic         { 0x786634a810267370ULL, 0xf53f14eebe33c020ULL, },
114*391f77aaSAleksandar Markovic         { 0x20ba34c408e06070ULL, 0xafde0b62e8ee6050ULL, },
115*391f77aaSAleksandar Markovic         { 0x07b6347bdf77af30ULL, 0x6b8d72be2f6d1c40ULL, },
116*391f77aaSAleksandar Markovic         { 0x63ea34bd3a9aa230ULL, 0xad25d0d828d84290ULL, },
117*391f77aaSAleksandar Markovic         { 0x934834f6f477f4c0ULL, 0xc39e78e84b9ade10ULL, },
118*391f77aaSAleksandar Markovic         { 0x3304e07e0c39cdc0ULL, 0x56b6f646a1f16e20ULL, },    /*  72  */
119*391f77aaSAleksandar Markovic         { 0x8f38e0c0675cc0c0ULL, 0x984e54609a5c9470ULL, },
120*391f77aaSAleksandar Markovic         { 0xff949cdcb6fb47c0ULL, 0xa70e305f61233be0ULL, },
121*391f77aaSAleksandar Markovic         { 0xbfcea8bac85c91c0ULL, 0x2cb600377e0d9160ULL, },
122*391f77aaSAleksandar Markovic         { 0x3f18e45e189a8740ULL, 0xddbf28e7d05731e0ULL, },
123*391f77aaSAleksandar Markovic         { 0x6f76e497d277d9d0ULL, 0xf338d0f7f319cd60ULL, },
124*391f77aaSAleksandar Markovic         { 0x2fb0f075e4d823d0ULL, 0x78e0a0cf100323e0ULL, },
125*391f77aaSAleksandar Markovic         { 0x2f4f0c4c60779f0cULL, 0xcfff608f7fff9fe0ULL, },
126*391f77aaSAleksandar Markovic         { 0x379944bc60e9d40cULL, 0x2a66400d7d7a4f60ULL, },    /*  80  */
127*391f77aaSAleksandar Markovic         { 0x4a0b4408801e08acULL, 0x36fc80bb3c7401e0ULL, },
128*391f77aaSAleksandar Markovic         { 0x922d0cb800dcb0acULL, 0xfc5c807628f8dc60ULL, },
129*391f77aaSAleksandar Markovic         { 0xb24a046000c05044ULL, 0x30c080e6c008a460ULL, },
130*391f77aaSAleksandar Markovic         { 0x22a66ce00040c044ULL, 0x208000724030e4e0ULL, },
131*391f77aaSAleksandar Markovic         { 0xcc726c4000808024ULL, 0xe00000de0060dc60ULL, },
132*391f77aaSAleksandar Markovic         { 0xbc5e04c000000024ULL, 0xc00000bc004010e0ULL, },
133*391f77aaSAleksandar Markovic         { 0x7c5cac000000002cULL, 0x0000001c00c0f0e0ULL, },
134*391f77aaSAleksandar Markovic         { 0x9c4424000000002cULL, 0x000000d40080f060ULL, },    /*  88  */
135*391f77aaSAleksandar Markovic         { 0xa8cc2400000000ccULL, 0x0000004c000010e0ULL, },
136*391f77aaSAleksandar Markovic         { 0xc814ac00000000ccULL, 0x000000980000c060ULL, },
137*391f77aaSAleksandar Markovic         { 0x48e8e400000000a4ULL, 0x0000005800004060ULL, },
138*391f77aaSAleksandar Markovic         { 0x08d80c00000000a4ULL, 0x00000008000040e0ULL, },
139*391f77aaSAleksandar Markovic         { 0x30880c0000000084ULL, 0x000000380000c060ULL, },
140*391f77aaSAleksandar Markovic         { 0xf0b8e40000000084ULL, 0x00000070000000e0ULL, },
141*391f77aaSAleksandar Markovic         { 0xf0f04c000000004cULL, 0x000000f0000000e0ULL, },
142*391f77aaSAleksandar Markovic         { 0x709004000000004cULL, 0x000000d000000060ULL, },    /*  96  */
143*391f77aaSAleksandar Markovic         { 0xf0f06c000000004cULL, 0x00000070000000e0ULL, },
144*391f77aaSAleksandar Markovic         { 0x709064000000004cULL, 0x0000005000000060ULL, },
145*391f77aaSAleksandar Markovic         { 0xf0f08c000000004cULL, 0x000000f0000000e0ULL, },
146*391f77aaSAleksandar Markovic         { 0xa0d08c00000000ecULL, 0x0000009000000060ULL, },
147*391f77aaSAleksandar Markovic         { 0xc0708c000000008cULL, 0x000000f0000000e0ULL, },
148*391f77aaSAleksandar Markovic         { 0x80508c000000002cULL, 0x0000009000000060ULL, },
149*391f77aaSAleksandar Markovic         { 0x00f08c00000000ccULL, 0x000000f0000000e0ULL, },
150*391f77aaSAleksandar Markovic         { 0x00906400000000ccULL, 0x000000e000000060ULL, },    /* 104  */
151*391f77aaSAleksandar Markovic         { 0x00f06c00000000ccULL, 0x000000c0000000e0ULL, },
152*391f77aaSAleksandar Markovic         { 0x00900400000000ccULL, 0x0000008000000060ULL, },
153*391f77aaSAleksandar Markovic         { 0x00f04c00000000ccULL, 0x00000000000000e0ULL, },
154*391f77aaSAleksandar Markovic         { 0x00e0c400000000a4ULL, 0x00000000000000e0ULL, },
155*391f77aaSAleksandar Markovic         { 0x00c0ec00000000acULL, 0x00000000000000e0ULL, },
156*391f77aaSAleksandar Markovic         { 0x0080a40000000044ULL, 0x00000000000000e0ULL, },
157*391f77aaSAleksandar Markovic         { 0x00008c000000008cULL, 0x00000000000000e0ULL, },
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_B(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_B(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_B__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_B__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