1*c0a19f7bSAleksandar Markovic /*
2*c0a19f7bSAleksandar Markovic  *  Test program for MSA instruction DPSUB_U.W
3*c0a19f7bSAleksandar Markovic  *
4*c0a19f7bSAleksandar Markovic  *  Copyright (C) 2019  Wave Computing, Inc.
5*c0a19f7bSAleksandar Markovic  *  Copyright (C) 2019  Aleksandar Markovic <amarkovic@wavecomp.com>
6*c0a19f7bSAleksandar Markovic  *
7*c0a19f7bSAleksandar Markovic  *  This program is free software: you can redistribute it and/or modify
8*c0a19f7bSAleksandar Markovic  *  it under the terms of the GNU General Public License as published by
9*c0a19f7bSAleksandar Markovic  *  the Free Software Foundation, either version 2 of the License, or
10*c0a19f7bSAleksandar Markovic  *  (at your option) any later version.
11*c0a19f7bSAleksandar Markovic  *`
12*c0a19f7bSAleksandar Markovic  *  This program is distributed in the hope that it will be useful,
13*c0a19f7bSAleksandar Markovic  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
14*c0a19f7bSAleksandar Markovic  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15*c0a19f7bSAleksandar Markovic  *  GNU General Public License for more details.
16*c0a19f7bSAleksandar Markovic  *
17*c0a19f7bSAleksandar Markovic  *  You should have received a copy of the GNU General Public License
18*c0a19f7bSAleksandar Markovic  *  along with this program.  If not, see <https://www.gnu.org/licenses/>.
19*c0a19f7bSAleksandar Markovic  *
20*c0a19f7bSAleksandar Markovic  */
21*c0a19f7bSAleksandar Markovic 
22*c0a19f7bSAleksandar Markovic #include <sys/time.h>
23*c0a19f7bSAleksandar Markovic #include <stdint.h>
24*c0a19f7bSAleksandar Markovic 
25*c0a19f7bSAleksandar Markovic #include "../../../../include/wrappers_msa.h"
26*c0a19f7bSAleksandar Markovic #include "../../../../include/test_inputs_128.h"
27*c0a19f7bSAleksandar Markovic #include "../../../../include/test_utils_128.h"
28*c0a19f7bSAleksandar Markovic 
29*c0a19f7bSAleksandar Markovic #define TEST_COUNT_TOTAL (                                                \
30*c0a19f7bSAleksandar Markovic             (PATTERN_INPUTS_SHORT_COUNT) * (PATTERN_INPUTS_SHORT_COUNT) + \
31*c0a19f7bSAleksandar Markovic             3 * (RANDOM_INPUTS_SHORT_COUNT) * (RANDOM_INPUTS_SHORT_COUNT))
32*c0a19f7bSAleksandar Markovic 
33*c0a19f7bSAleksandar Markovic 
main(void)34*c0a19f7bSAleksandar Markovic int32_t main(void)
35*c0a19f7bSAleksandar Markovic {
36*c0a19f7bSAleksandar Markovic     char *isa_ase_name = "MSA";
37*c0a19f7bSAleksandar Markovic     char *group_name = "Int Dot Product";
38*c0a19f7bSAleksandar Markovic     char *instruction_name =  "DPSUB_U.W";
39*c0a19f7bSAleksandar Markovic     int32_t ret;
40*c0a19f7bSAleksandar Markovic     uint32_t i, j;
41*c0a19f7bSAleksandar Markovic     struct timeval start, end;
42*c0a19f7bSAleksandar Markovic     double elapsed_time;
43*c0a19f7bSAleksandar Markovic 
44*c0a19f7bSAleksandar Markovic     uint64_t b128_result[TEST_COUNT_TOTAL][2];
45*c0a19f7bSAleksandar Markovic     uint64_t b128_expect[TEST_COUNT_TOTAL][2] = {
46*c0a19f7bSAleksandar Markovic         { 0x0003fffe0003fffeULL, 0x0003fffe0003fffeULL, },    /*   0  */
47*c0a19f7bSAleksandar Markovic         { 0x0003fffe0003fffeULL, 0x0003fffe0003fffeULL, },
48*c0a19f7bSAleksandar Markovic         { 0xaab15552aab15552ULL, 0xaab15552aab15552ULL, },
49*c0a19f7bSAleksandar Markovic         { 0x0007fffc0007fffcULL, 0x0007fffc0007fffcULL, },
50*c0a19f7bSAleksandar Markovic         { 0x6671999466719994ULL, 0x6671999466719994ULL, },
51*c0a19f7bSAleksandar Markovic         { 0x000bfffa000bfffaULL, 0x000bfffa000bfffaULL, },
52*c0a19f7bSAleksandar Markovic         { 0xe39c1c6b8e4771c0ULL, 0x38f1c715e39c1c6bULL, },
53*c0a19f7bSAleksandar Markovic         { 0x000ffff8000ffff8ULL, 0x000ffff8000ffff8ULL, },
54*c0a19f7bSAleksandar Markovic         { 0x000ffff8000ffff8ULL, 0x000ffff8000ffff8ULL, },    /*   8  */
55*c0a19f7bSAleksandar Markovic         { 0x000ffff8000ffff8ULL, 0x000ffff8000ffff8ULL, },
56*c0a19f7bSAleksandar Markovic         { 0x000ffff8000ffff8ULL, 0x000ffff8000ffff8ULL, },
57*c0a19f7bSAleksandar Markovic         { 0x000ffff8000ffff8ULL, 0x000ffff8000ffff8ULL, },
58*c0a19f7bSAleksandar Markovic         { 0x000ffff8000ffff8ULL, 0x000ffff8000ffff8ULL, },
59*c0a19f7bSAleksandar Markovic         { 0x000ffff8000ffff8ULL, 0x000ffff8000ffff8ULL, },
60*c0a19f7bSAleksandar Markovic         { 0x000ffff8000ffff8ULL, 0x000ffff8000ffff8ULL, },
61*c0a19f7bSAleksandar Markovic         { 0x000ffff8000ffff8ULL, 0x000ffff8000ffff8ULL, },
62*c0a19f7bSAleksandar Markovic         { 0xaabd554caabd554cULL, 0xaabd554caabd554cULL, },    /*  16  */
63*c0a19f7bSAleksandar Markovic         { 0xaabd554caabd554cULL, 0xaabd554caabd554cULL, },
64*c0a19f7bSAleksandar Markovic         { 0xc730e384c730e384ULL, 0xc730e384c730e384ULL, },
65*c0a19f7bSAleksandar Markovic         { 0x556aaaa0556aaaa0ULL, 0x556aaaa0556aaaa0ULL, },
66*c0a19f7bSAleksandar Markovic         { 0x445bbbb0445bbbb0ULL, 0x445bbbb0445bbbb0ULL, },
67*c0a19f7bSAleksandar Markovic         { 0x0017fff40017fff4ULL, 0x0017fff40017fff4ULL, },
68*c0a19f7bSAleksandar Markovic         { 0x427812ea0994f678ULL, 0x7b5bda06427812eaULL, },
69*c0a19f7bSAleksandar Markovic         { 0xaac55548aac55548ULL, 0xaac55548aac55548ULL, },
70*c0a19f7bSAleksandar Markovic         { 0x001bfff2001bfff2ULL, 0x001bfff2001bfff2ULL, },    /*  24  */
71*c0a19f7bSAleksandar Markovic         { 0x001bfff2001bfff2ULL, 0x001bfff2001bfff2ULL, },
72*c0a19f7bSAleksandar Markovic         { 0x8e55c70e8e55c70eULL, 0x8e55c70e8e55c70eULL, },
73*c0a19f7bSAleksandar Markovic         { 0x5572aa9c5572aa9cULL, 0x5572aa9c5572aa9cULL, },
74*c0a19f7bSAleksandar Markovic         { 0xcceb3324cceb3324ULL, 0xcceb3324cceb3324ULL, },
75*c0a19f7bSAleksandar Markovic         { 0xaac95546aac95546ULL, 0xaac95546aac95546ULL, },
76*c0a19f7bSAleksandar Markovic         { 0x4bf95ec12f87d088ULL, 0x686b424f4bf95ec1ULL, },
77*c0a19f7bSAleksandar Markovic         { 0x001ffff0001ffff0ULL, 0x001ffff0001ffff0ULL, },
78*c0a19f7bSAleksandar Markovic         { 0x6689998866899988ULL, 0x6689998866899988ULL, },    /*  32  */
79*c0a19f7bSAleksandar Markovic         { 0x6689998866899988ULL, 0x6689998866899988ULL, },
80*c0a19f7bSAleksandar Markovic         { 0x557aaa98557aaa98ULL, 0x557aaa98557aaa98ULL, },
81*c0a19f7bSAleksandar Markovic         { 0xccf33320ccf33320ULL, 0xccf33320ccf33320ULL, },
82*c0a19f7bSAleksandar Markovic         { 0x8547ae008547ae00ULL, 0x8547ae008547ae00ULL, },
83*c0a19f7bSAleksandar Markovic         { 0x335cccb8335cccb8ULL, 0x335cccb8335cccb8ULL, },
84*c0a19f7bSAleksandar Markovic         { 0x4fd016ac0b8c5af0ULL, 0x94149f344fd016acULL, },
85*c0a19f7bSAleksandar Markovic         { 0x99c6665099c66650ULL, 0x99c6665099c66650ULL, },
86*c0a19f7bSAleksandar Markovic         { 0x3360ccb63360ccb6ULL, 0x3360ccb63360ccb6ULL, },    /*  40  */
87*c0a19f7bSAleksandar Markovic         { 0x3360ccb63360ccb6ULL, 0x3360ccb63360ccb6ULL, },
88*c0a19f7bSAleksandar Markovic         { 0xef1d10faef1d10faULL, 0xef1d10faef1d10faULL, },
89*c0a19f7bSAleksandar Markovic         { 0xccfb331cccfb331cULL, 0xccfb331cccfb331cULL, },
90*c0a19f7bSAleksandar Markovic         { 0x7b1051d47b1051d4ULL, 0x7b1051d47b1051d4ULL, },
91*c0a19f7bSAleksandar Markovic         { 0x6695998266959982ULL, 0x6695998266959982ULL, },
92*c0a19f7bSAleksandar Markovic         { 0x2db26bff1ca17d10ULL, 0x3ec38e212db26bffULL, },
93*c0a19f7bSAleksandar Markovic         { 0x002fffe8002fffe8ULL, 0x002fffe8002fffe8ULL, },
94*c0a19f7bSAleksandar Markovic         { 0xe3c01c598e6b71aeULL, 0x3915c703e3c01c59ULL, },    /*  48  */
95*c0a19f7bSAleksandar Markovic         { 0xe3c01c598e6b71aeULL, 0x3915c703e3c01c59ULL, },
96*c0a19f7bSAleksandar Markovic         { 0x26202f4f97e86832ULL, 0xb459a11526202f4fULL, },
97*c0a19f7bSAleksandar Markovic         { 0xc75038ca1ca6e374ULL, 0x71fb8e1ec75038caULL, },
98*c0a19f7bSAleksandar Markovic         { 0xe3c382bef4d671acULL, 0xd2b3609ae3c382beULL, },
99*c0a19f7bSAleksandar Markovic         { 0xaae0553baae2553aULL, 0xaae15539aae0553bULL, },
100*c0a19f7bSAleksandar Markovic         { 0xd3f7192e919b0636ULL, 0x4f3b0fb0d3f7192eULL, },
101*c0a19f7bSAleksandar Markovic         { 0x8e7071ac391dc700ULL, 0xe3c71c548e7071acULL, },
102*c0a19f7bSAleksandar Markovic         { 0xaae45539aae65538ULL, 0xaae55537aae45539ULL, },    /*  56  */
103*c0a19f7bSAleksandar Markovic         { 0xaae45539aae65538ULL, 0xaae55537aae45539ULL, },
104*c0a19f7bSAleksandar Markovic         { 0x133197974c16b408ULL, 0xda4ed07913319797ULL, },
105*c0a19f7bSAleksandar Markovic         { 0xc75838c61caee370ULL, 0x72038e1ac75838c6ULL, },
106*c0a19f7bSAleksandar Markovic         { 0x114e886aaae8eed0ULL, 0x77b55536114e886aULL, },
107*c0a19f7bSAleksandar Markovic         { 0xe3cc1c538e7771a8ULL, 0x3921c6fde3cc1c53ULL, },
108*c0a19f7bSAleksandar Markovic         { 0x9e4574d135fa3272ULL, 0xcdadd3a19e4574d1ULL, },
109*c0a19f7bSAleksandar Markovic         { 0x003fffe0003fffe0ULL, 0x003fffe0003fffe0ULL, },
110*c0a19f7bSAleksandar Markovic         { 0xe77c016cdd7daa5cULL, 0xe98945eb8a373037ULL, },    /*  64  */
111*c0a19f7bSAleksandar Markovic         { 0x60fd93dc8ef99616ULL, 0xdba475fe3c075522ULL, },
112*c0a19f7bSAleksandar Markovic         { 0x67ae7b204335bcd8ULL, 0xc7121f747860f089ULL, },
113*c0a19f7bSAleksandar Markovic         { 0x17bb0f0ee8fd8c96ULL, 0x972fe12d34478253ULL, },
114*c0a19f7bSAleksandar Markovic         { 0x913ca17e9a797850ULL, 0x894b1140e617a73eULL, },
115*c0a19f7bSAleksandar Markovic         { 0x99ae5a31e83a45a7ULL, 0xff24684bc96dd9adULL, },
116*c0a19f7bSAleksandar Markovic         { 0xefeea5a7437774caULL, 0x6ac04bfdaf344b68ULL, },
117*c0a19f7bSAleksandar Markovic         { 0x8175363e76faed37ULL, 0xfc38b88657b1e4aaULL, },
118*c0a19f7bSAleksandar Markovic         { 0x88261d822b3713f9ULL, 0xe7a661fc940b8011ULL, },    /*  72  */
119*c0a19f7bSAleksandar Markovic         { 0xde6668f88674431cULL, 0x534245ae79d1f1ccULL, },
120*c0a19f7bSAleksandar Markovic         { 0xf331d070b3932dbbULL, 0xb25f4d6d0200af03ULL, },
121*c0a19f7bSAleksandar Markovic         { 0x985e1b88f3e41a1cULL, 0x31e873ed7002566dULL, },
122*c0a19f7bSAleksandar Markovic         { 0x486aaf7699abe9daULL, 0x020635a62be8e837ULL, },
123*c0a19f7bSAleksandar Markovic         { 0xd9f1400dcd2f6247ULL, 0x937ea22fd4668179ULL, },
124*c0a19f7bSAleksandar Markovic         { 0x7f1d8b250d804ea8ULL, 0x1307c8af426828e3ULL, },
125*c0a19f7bSAleksandar Markovic         { 0x4be6ff9b22ca7983ULL, 0x7b2e358e09e10fdfULL, },
126*c0a19f7bSAleksandar Markovic         { 0x3d0470dbf4d6b86fULL, 0x548567e8f5250450ULL, },    /*  80  */
127*c0a19f7bSAleksandar Markovic         { 0x00d897321b41b715ULL, 0x02517c05df66c875ULL, },
128*c0a19f7bSAleksandar Markovic         { 0x991ec80ea3b5c306ULL, 0xa18dc9b22cff8e2fULL, },
129*c0a19f7bSAleksandar Markovic         { 0x44850796bb133f8dULL, 0xdc2a4cc591614211ULL, },
130*c0a19f7bSAleksandar Markovic         { 0x192b30fc8866f607ULL, 0x97e8c289d36e61aaULL, },
131*c0a19f7bSAleksandar Markovic         { 0x0058689e9fcad43dULL, 0xfe7a0cc7a239bc40ULL, },
132*c0a19f7bSAleksandar Markovic         { 0xb8bc4cc2b8296867ULL, 0xccf01b9e1a7e74adULL, },
133*c0a19f7bSAleksandar Markovic         { 0x61014864181c5d2cULL, 0x4c8bc05ea1b0cc11ULL, },
134*c0a19f7bSAleksandar Markovic         { 0xec0d0e4af547db74ULL, 0x2d758eed74a13bb5ULL, },    /*  88  */
135*c0a19f7bSAleksandar Markovic         { 0x03e797060056a10fULL, 0xc1a1d5f8579892eaULL, },
136*c0a19f7bSAleksandar Markovic         { 0x9a3ca5d4a8548905ULL, 0xfd2bfd1807c0081aULL, },
137*c0a19f7bSAleksandar Markovic         { 0x4820b48cf1454f6bULL, 0xe982ac5dfb74445aULL, },
138*c0a19f7bSAleksandar Markovic         { 0x7eec2fbcb0c3c941ULL, 0x9d1459e9d27d4766ULL, },
139*c0a19f7bSAleksandar Markovic         { 0x020a22e0debbd140ULL, 0x4fbb0ef3a9e0453bULL, },
140*c0a19f7bSAleksandar Markovic         { 0xe8df4a9ccb0c350bULL, 0x37b3761e2e442cffULL, },
141*c0a19f7bSAleksandar Markovic         { 0x7c3604df51731065ULL, 0xd9add64be7d81e17ULL, },
142*c0a19f7bSAleksandar Markovic         { 0x35a1aacf3f24481fULL, 0x900caa26ecaf303bULL, },    /*  96  */
143*c0a19f7bSAleksandar Markovic         { 0x7f0fd7311d2a2997ULL, 0x5e11155ee03d0362ULL, },
144*c0a19f7bSAleksandar Markovic         { 0x7959c1ef0ab6e6c3ULL, 0x41695f03ff01377bULL, },
145*c0a19f7bSAleksandar Markovic         { 0x89d8f6a1bc2ded57ULL, 0x29ed46aadb5c8a3cULL, },
146*c0a19f7bSAleksandar Markovic         { 0x01ec800ecaa24ac8ULL, 0xf32ccdbb9c58b788ULL, },
147*c0a19f7bSAleksandar Markovic         { 0xffd7297c53176782ULL, 0x4acc984953e0cc00ULL, },
148*c0a19f7bSAleksandar Markovic         { 0x04316ff6e9707c3dULL, 0xd5f54b0b0ac9f7e0ULL, },
149*c0a19f7bSAleksandar Markovic         { 0xffe6fc76421c7405ULL, 0x8f42f98ab98b12e9ULL, },
150*c0a19f7bSAleksandar Markovic         { 0xa75ea33ed2e809e1ULL, 0xb6fdbf643abee85cULL, },    /* 104  */
151*c0a19f7bSAleksandar Markovic         { 0xc75019063471bcc9ULL, 0x05bcd250f1d0ad42ULL, },
152*c0a19f7bSAleksandar Markovic         { 0x300d94eaa78224eaULL, 0x615cfa00370a0c2aULL, },
153*c0a19f7bSAleksandar Markovic         { 0xaa1a04f419d03dccULL, 0x8fe0ca60107a1a34ULL, },
154*c0a19f7bSAleksandar Markovic         { 0x5f0bb18ad9b000d4ULL, 0xd3ed3780ee630840ULL, },
155*c0a19f7bSAleksandar Markovic         { 0x25e24aa388dc4d8cULL, 0x40c1586349788fbaULL, },
156*c0a19f7bSAleksandar Markovic         { 0x0ec344de11f41ac8ULL, 0xed9aea2a99a95e8aULL, },
157*c0a19f7bSAleksandar Markovic         { 0x02499bebf3ac5a24ULL, 0xecb186c0e06045b8ULL, },
158*c0a19f7bSAleksandar Markovic     };
159*c0a19f7bSAleksandar Markovic 
160*c0a19f7bSAleksandar Markovic     reset_msa_registers();
161*c0a19f7bSAleksandar Markovic 
162*c0a19f7bSAleksandar Markovic     gettimeofday(&start, NULL);
163*c0a19f7bSAleksandar Markovic 
164*c0a19f7bSAleksandar Markovic     for (i = 0; i < PATTERN_INPUTS_SHORT_COUNT; i++) {
165*c0a19f7bSAleksandar Markovic         for (j = 0; j < PATTERN_INPUTS_SHORT_COUNT; j++) {
166*c0a19f7bSAleksandar Markovic             do_msa_DPSUB_U_W(b128_pattern[i], b128_pattern[j],
167*c0a19f7bSAleksandar Markovic                              b128_result[PATTERN_INPUTS_SHORT_COUNT * i + j]);
168*c0a19f7bSAleksandar Markovic         }
169*c0a19f7bSAleksandar Markovic     }
170*c0a19f7bSAleksandar Markovic 
171*c0a19f7bSAleksandar Markovic     for (i = 0; i < RANDOM_INPUTS_SHORT_COUNT; i++) {
172*c0a19f7bSAleksandar Markovic         for (j = 0; j < RANDOM_INPUTS_SHORT_COUNT; j++) {
173*c0a19f7bSAleksandar Markovic             do_msa_DPSUB_U_W(b128_random[i], b128_random[j],
174*c0a19f7bSAleksandar Markovic                              b128_result[((PATTERN_INPUTS_SHORT_COUNT) *
175*c0a19f7bSAleksandar Markovic                                           (PATTERN_INPUTS_SHORT_COUNT)) +
176*c0a19f7bSAleksandar Markovic                                          RANDOM_INPUTS_SHORT_COUNT * i + j]);
177*c0a19f7bSAleksandar Markovic         }
178*c0a19f7bSAleksandar Markovic     }
179*c0a19f7bSAleksandar Markovic 
180*c0a19f7bSAleksandar Markovic     for (i = 0; i < RANDOM_INPUTS_SHORT_COUNT; i++) {
181*c0a19f7bSAleksandar Markovic         for (j = 0; j < RANDOM_INPUTS_SHORT_COUNT; j++) {
182*c0a19f7bSAleksandar Markovic             do_msa_DPSUB_U_W__DDT(b128_random[i], b128_random[j],
183*c0a19f7bSAleksandar Markovic                                   b128_result[
184*c0a19f7bSAleksandar Markovic                                       ((PATTERN_INPUTS_SHORT_COUNT) *
185*c0a19f7bSAleksandar Markovic                                        (PATTERN_INPUTS_SHORT_COUNT)) +
186*c0a19f7bSAleksandar Markovic                                       ((RANDOM_INPUTS_SHORT_COUNT) *
187*c0a19f7bSAleksandar Markovic                                        (RANDOM_INPUTS_SHORT_COUNT)) +
188*c0a19f7bSAleksandar Markovic                                       RANDOM_INPUTS_SHORT_COUNT * i + j]);
189*c0a19f7bSAleksandar Markovic         }
190*c0a19f7bSAleksandar Markovic     }
191*c0a19f7bSAleksandar Markovic 
192*c0a19f7bSAleksandar Markovic     for (i = 0; i < RANDOM_INPUTS_SHORT_COUNT; i++) {
193*c0a19f7bSAleksandar Markovic         for (j = 0; j < RANDOM_INPUTS_SHORT_COUNT; j++) {
194*c0a19f7bSAleksandar Markovic             do_msa_DPSUB_U_W__DSD(b128_random[i], b128_random[j],
195*c0a19f7bSAleksandar Markovic                                   b128_result[
196*c0a19f7bSAleksandar Markovic                                       ((PATTERN_INPUTS_SHORT_COUNT) *
197*c0a19f7bSAleksandar Markovic                                        (PATTERN_INPUTS_SHORT_COUNT)) +
198*c0a19f7bSAleksandar Markovic                                       (2 * (RANDOM_INPUTS_SHORT_COUNT) *
199*c0a19f7bSAleksandar Markovic                                        (RANDOM_INPUTS_SHORT_COUNT)) +
200*c0a19f7bSAleksandar Markovic                                       RANDOM_INPUTS_SHORT_COUNT * i + j]);
201*c0a19f7bSAleksandar Markovic         }
202*c0a19f7bSAleksandar Markovic     }
203*c0a19f7bSAleksandar Markovic 
204*c0a19f7bSAleksandar Markovic     gettimeofday(&end, NULL);
205*c0a19f7bSAleksandar Markovic 
206*c0a19f7bSAleksandar Markovic     elapsed_time = (end.tv_sec - start.tv_sec) * 1000.0;
207*c0a19f7bSAleksandar Markovic     elapsed_time += (end.tv_usec - start.tv_usec) / 1000.0;
208*c0a19f7bSAleksandar Markovic 
209*c0a19f7bSAleksandar Markovic     ret = check_results_128(isa_ase_name, group_name, instruction_name,
210*c0a19f7bSAleksandar Markovic                             TEST_COUNT_TOTAL, elapsed_time,
211*c0a19f7bSAleksandar Markovic                             &b128_result[0][0], &b128_expect[0][0]);
212*c0a19f7bSAleksandar Markovic 
213*c0a19f7bSAleksandar Markovic     return ret;
214*c0a19f7bSAleksandar Markovic }
215