• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*  Copyright (C) 2011 IBM
2 
3  Author: Maynard Johnson <maynardj@us.ibm.com>
4 
5  This program is free software; you can redistribute it and/or
6  modify it under the terms of the GNU General Public License as
7  published by the Free Software Foundation; either version 2 of the
8  License, or (at your option) any later version.
9 
10  This program is distributed in the hope that it will be useful, but
11  WITHOUT ANY WARRANTY; without even the implied warranty of
12  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  General Public License for more details.
14 
15  You should have received a copy of the GNU General Public License
16  along with this program; if not, write to the Free Software
17  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
18  02111-1307, USA.
19 
20  The GNU General Public License is contained in the file COPYING.
21  */
22 
23 #ifdef HAS_VSX
24 
25 #include <stdio.h>
26 #include <stdint.h>
27 #include <stdlib.h>
28 #include <string.h>
29 #include <malloc.h>
30 #include <altivec.h>
31 #include <math.h>
32 
33 #ifndef __powerpc64__
34 typedef uint32_t HWord_t;
35 #else
36 typedef uint64_t HWord_t;
37 #endif /* __powerpc64__ */
38 
39 typedef unsigned char Bool;
40 #define True 1
41 #define False 0
42 register HWord_t r14 __asm__ ("r14");
43 register HWord_t r15 __asm__ ("r15");
44 register HWord_t r16 __asm__ ("r16");
45 register HWord_t r17 __asm__ ("r17");
46 register double f14 __asm__ ("fr14");
47 register double f15 __asm__ ("fr15");
48 register double f16 __asm__ ("fr16");
49 register double f17 __asm__ ("fr17");
50 
51 static volatile unsigned int div_flags, div_xer;
52 
53 #define ALLCR "cr0","cr1","cr2","cr3","cr4","cr5","cr6","cr7"
54 
55 #define SET_CR(_arg) \
56       __asm__ __volatile__ ("mtcr  %0" : : "b"(_arg) : ALLCR );
57 
58 #define SET_XER(_arg) \
59       __asm__ __volatile__ ("mtxer %0" : : "b"(_arg) : "xer" );
60 
61 #define GET_CR(_lval) \
62       __asm__ __volatile__ ("mfcr %0"  : "=b"(_lval) )
63 
64 #define GET_XER(_lval) \
65       __asm__ __volatile__ ("mfxer %0" : "=b"(_lval) )
66 
67 #define GET_CR_XER(_lval_cr,_lval_xer) \
68    do { GET_CR(_lval_cr); GET_XER(_lval_xer); } while (0)
69 
70 #define SET_CR_ZERO \
71       SET_CR(0)
72 
73 #define SET_XER_ZERO \
74       SET_XER(0)
75 
76 #define SET_CR_XER_ZERO \
77    do { SET_CR_ZERO; SET_XER_ZERO; } while (0)
78 
79 #define SET_FPSCR_ZERO \
80    do { double _d = 0.0; \
81         __asm__ __volatile__ ("mtfsf 0xFF, %0" : : "f"(_d) ); \
82    } while (0)
83 
84 
85 typedef void (*test_func_t)(void);
86 typedef struct test_table test_table_t;
87 
88 
89 /* These functions below that construct a table of floating point
90  * values were lifted from none/tests/ppc32/jm-insns.c.
91  */
92 
93 #if defined (DEBUG_ARGS_BUILD)
94 #define AB_DPRINTF(fmt, args...) do { fprintf(stderr, fmt , ##args); } while (0)
95 #else
96 #define AB_DPRINTF(fmt, args...) do { } while (0)
97 #endif
98 
register_farg(void * farg,int s,uint16_t _exp,uint64_t mant)99 static inline void register_farg (void *farg,
100                                   int s, uint16_t _exp, uint64_t mant)
101 {
102    uint64_t tmp;
103 
104    tmp = ((uint64_t)s << 63) | ((uint64_t)_exp << 52) | mant;
105    *(uint64_t *)farg = tmp;
106    AB_DPRINTF("%d %03x %013llx => %016llx %0e\n",
107               s, _exp, mant, *(uint64_t *)farg, *(double *)farg);
108 }
109 
110 
111 typedef struct fp_test_args {
112    int fra_idx;
113    int frb_idx;
114 } fp_test_args_t;
115 
116 
117 fp_test_args_t two_arg_fp_tests[] = {
118                                      {8, 8},
119                                      {8, 14},
120                                      {15, 16},
121                                      {8, 5},
122                                      {8, 4},
123                                      {8, 7},
124                                      {8, 9},
125                                      {8, 11},
126                                      {14, 8},
127                                      {14, 14},
128                                      {14, 6},
129                                      {14, 5},
130                                      {14, 4},
131                                      {14, 7},
132                                      {14, 9},
133                                      {14, 11},
134                                      {6, 8},
135                                      {6, 14},
136                                      {6, 6},
137                                      {6, 5},
138                                      {6, 4},
139                                      {6, 7},
140                                      {6, 9},
141                                      {6, 11},
142                                      {5, 8},
143                                      {5, 14},
144                                      {5, 6},
145                                      {5, 5},
146                                      {5, 4},
147                                      {5, 7},
148                                      {5, 9},
149                                      {5, 11},
150                                      {4, 8},
151                                      {4, 14},
152                                      {4, 6},
153                                      {4, 5},
154                                      {4, 1},
155                                      {4, 7},
156                                      {4, 9},
157                                      {4, 11},
158                                      {7, 8},
159                                      {7, 14},
160                                      {7, 6},
161                                      {7, 5},
162                                      {7, 4},
163                                      {7, 7},
164                                      {7, 9},
165                                      {7, 11},
166                                      {10, 8},
167                                      {10, 14},
168                                      {12, 6},
169                                      {12, 5},
170                                      {10, 4},
171                                      {10, 7},
172                                      {10, 9},
173                                      {10, 11},
174                                      {12, 8 },
175                                      {12, 14},
176                                      {12, 6},
177                                      {15, 16},
178                                      {15, 16},
179                                      {9, 11},
180                                      {11, 11},
181                                      {11, 12},
182                                      {16, 18},
183                                      {17, 16},
184                                      {19, 19},
185                                      {19, 18}
186 };
187 
188 
189 static int nb_special_fargs;
190 static double * spec_fargs;
191 static float * spec_sp_fargs;
192 
build_special_fargs_table(void)193 static void build_special_fargs_table(void)
194 {
195 /*
196   Entry  Sign Exp   fraction                  Special value
197    0      0   3fd   0x8000000000000ULL         Positive finite number
198    1      0   404   0xf000000000000ULL         ...
199    2      0   001   0x8000000b77501ULL         ...
200    3      0   7fe   0x800000000051bULL         ...
201    4      0   012   0x3214569900000ULL         ...
202    5      0   000   0x0000000000000ULL         +0.0 (+zero)
203    6      1   000   0x0000000000000ULL         -0.0 (-zero)
204    7      0   7ff   0x0000000000000ULL         +infinity
205    8      1   7ff   0x0000000000000ULL         -infinity
206    9      0   7ff   0x7FFFFFFFFFFFFULL         +SNaN
207    10     1   7ff   0x7FFFFFFFFFFFFULL         -SNaN
208    11     0   7ff   0x8000000000000ULL         +QNaN
209    12     1   7ff   0x8000000000000ULL         -QNaN
210    13     1   000   0x8340000078000ULL         Denormalized val (zero exp and non-zero fraction)
211    14     1   40d   0x0650f5a07b353ULL         Negative finite number
212    15     0   412   0x32585a9900000ULL         A few more positive finite numbers
213    16     0   413   0x82511a2000000ULL         ...
214    17  . . . . . . . . . . . . . . . . . . . . . . .
215    18  . . . . . . . . . . . . . . . . . . . . . . .
216    19  . . . . . . . . . . . . . . . . . . . . . . .
217 */
218 
219    uint64_t mant;
220    uint16_t _exp;
221    int s;
222    int j, i = 0;
223 
224    if (spec_fargs)
225       return;
226 
227    spec_fargs = malloc( 20 * sizeof(double) );
228    spec_sp_fargs = malloc( 20 * sizeof(float) );
229 
230    // #0
231    s = 0;
232    _exp = 0x3fd;
233    mant = 0x8000000000000ULL;
234    register_farg(&spec_fargs[i++], s, _exp, mant);
235 
236    // #1
237    s = 0;
238    _exp = 0x404;
239    mant = 0xf000000000000ULL;
240    register_farg(&spec_fargs[i++], s, _exp, mant);
241 
242    // #2
243    s = 0;
244    _exp = 0x001;
245    mant = 0x8000000b77501ULL;
246    register_farg(&spec_fargs[i++], s, _exp, mant);
247 
248    // #3
249    s = 0;
250    _exp = 0x7fe;
251    mant = 0x800000000051bULL;
252    register_farg(&spec_fargs[i++], s, _exp, mant);
253 
254    // #4
255    s = 0;
256    _exp = 0x012;
257    mant = 0x3214569900000ULL;
258    register_farg(&spec_fargs[i++], s, _exp, mant);
259 
260 
261    /* Special values */
262    /* +0.0      : 0 0x000 0x0000000000000 */
263    // #5
264    s = 0;
265    _exp = 0x000;
266    mant = 0x0000000000000ULL;
267    register_farg(&spec_fargs[i++], s, _exp, mant);
268 
269    /* -0.0      : 1 0x000 0x0000000000000 */
270    // #6
271    s = 1;
272    _exp = 0x000;
273    mant = 0x0000000000000ULL;
274    register_farg(&spec_fargs[i++], s, _exp, mant);
275 
276    /* +infinity : 0 0x7FF 0x0000000000000  */
277    // #7
278    s = 0;
279    _exp = 0x7FF;
280    mant = 0x0000000000000ULL;
281    register_farg(&spec_fargs[i++], s, _exp, mant);
282 
283    /* -infinity : 1 0x7FF 0x0000000000000 */
284    // #8
285    s = 1;
286    _exp = 0x7FF;
287    mant = 0x0000000000000ULL;
288    register_farg(&spec_fargs[i++], s, _exp, mant);
289 
290    /* +QNaN     : 0 0x7FF 0x7FFFFFFFFFFFF */
291    // #9
292    s = 0;
293    _exp = 0x7FF;
294    mant = 0x7FFFFFFFFFFFFULL;
295    register_farg(&spec_fargs[i++], s, _exp, mant);
296 
297    /* -QNaN     : 1 0x7FF 0x7FFFFFFFFFFFF */
298    // #10
299    s = 1;
300    _exp = 0x7FF;
301    mant = 0x7FFFFFFFFFFFFULL;
302    register_farg(&spec_fargs[i++], s, _exp, mant);
303 
304    /* +SNaN     : 0 0x7FF 0x8000000000000 */
305    // #11
306    s = 0;
307    _exp = 0x7FF;
308    mant = 0x8000000000000ULL;
309    register_farg(&spec_fargs[i++], s, _exp, mant);
310 
311    /* -SNaN     : 1 0x7FF 0x8000000000000 */
312    // #12
313    s = 1;
314    _exp = 0x7FF;
315    mant = 0x8000000000000ULL;
316    register_farg(&spec_fargs[i++], s, _exp, mant);
317 
318    /* denormalized value */
319    // #13
320    s = 1;
321    _exp = 0x000;
322    mant = 0x8340000078000ULL;
323    register_farg(&spec_fargs[i++], s, _exp, mant);
324 
325    /* Negative finite number */
326    // #14
327    s = 1;
328    _exp = 0x40d;
329    mant = 0x0650f5a07b353ULL;
330    register_farg(&spec_fargs[i++], s, _exp, mant);
331 
332    /* A few positive finite numbers ... */
333    // #15
334    s = 0;
335    _exp = 0x412;
336    mant = 0x32585a9900000ULL;
337    register_farg(&spec_fargs[i++], s, _exp, mant);
338 
339    // #16
340    s = 0;
341    _exp = 0x413;
342    mant = 0x82511a2000000ULL;
343    register_farg(&spec_fargs[i++], s, _exp, mant);
344 
345    // #17
346    s = 0;
347    _exp = 0x403;
348    mant = 0x12ef5a9300000ULL;
349    register_farg(&spec_fargs[i++], s, _exp, mant);
350 
351    // #18
352    s = 0;
353    _exp = 0x405;
354    mant = 0x14bf5d2300000ULL;
355    register_farg(&spec_fargs[i++], s, _exp, mant);
356 
357    // #19
358    s = 0;
359    _exp = 0x409;
360    mant = 0x76bf982440000ULL;
361    register_farg(&spec_fargs[i++], s, _exp, mant);
362 
363    nb_special_fargs = i;
364    for (j = 0; j < i; j++) {
365       spec_sp_fargs[j] = spec_fargs[j];
366    }
367 }
368 
369 
370 struct test_table
371 {
372    test_func_t test_category;
373    char * name;
374 };
375 
376 /*  Type of input for floating point operations.*/
377 typedef enum {
378    SINGLE_TEST,
379    DOUBLE_TEST
380 } precision_type_t;
381 
382 typedef enum {
383    VX_SCALAR_CONV_TO_WORD,
384    VX_CONV_TO_SINGLE,
385    VX_CONV_TO_DOUBLE,
386    VX_ESTIMATE,
387    VX_DEFAULT
388 } vx_fp_test_type;
389 
390 static vector unsigned int vec_out, vec_inA, vec_inB;
391 
392 /* This function is for checking the reciprocal and reciprocal square root
393  * estimate instructions.
394  */
check_estimate(precision_type_t type,Bool is_rsqrte,int idx,int output_vec_idx)395 Bool check_estimate(precision_type_t type, Bool is_rsqrte, int idx, int output_vec_idx)
396 {
397    /* Technically, the number of bits of precision for xvredp and xvrsqrtedp is
398     * 14 bits (14 = log2 16384).  However, the VEX emulation of these instructions
399     * does an actual reciprocal calculation versus estimation, so the answer we get back from
400     * valgrind can easily differ from the estimate in the lower bits (within the 14 bits of
401     * precision) and the estimate may still be within expected tolerances.  On top of that,
402     * we can't count on these estimates always being the same across implementations.
403     * For example, with the fre[s] instruction (which should be correct to within one part
404     * in 256 -- i.e., 8 bits of precision) . . . When approximating the value 1.0111_1111_1111,
405     * one implementation could return 1.0111_1111_0000 and another implementation could return
406     * 1.1000_0000_0000.  Both estimates meet the 1/256 accuracy requirement, but share only a
407     * single bit in common.
408     *
409     * The upshot is we can't validate the VEX output for these instructions by comparing against
410     * stored bit patterns.  We must check that the result is within expected tolerances.
411     */
412 
413 
414    /* A mask to be used for validation as a last resort.
415     * Only use 12 bits of precision for reasons discussed above.
416     */
417 #define VSX_RECIP_ESTIMATE_MASK_DP 0xFFFFFF0000000000ULL
418 #define VSX_RECIP_ESTIMATE_MASK_SP 0xFFFFFF00
419 
420    Bool result = False;
421    Bool dp_test = type == DOUBLE_TEST;
422    double src_dp, res_dp;
423    float src_sp, res_sp;
424    src_dp = res_dp = 0;
425    src_sp = res_sp = 0;
426 #define SRC (dp_test ? src_dp : src_sp)
427 #define RES (dp_test ? res_dp : res_sp)
428    Bool src_is_negative = False;
429    Bool res_is_negative = False;
430    unsigned long long * dst_dp = NULL;
431    unsigned int * dst_sp = NULL;
432    if (dp_test) {
433       unsigned long long * src_dp_ull;
434       dst_dp = (unsigned long long *) &vec_out;
435       src_dp = spec_fargs[idx];
436       src_dp_ull = (unsigned long long *) &src_dp;
437       src_is_negative = (*src_dp_ull & 0x8000000000000000ULL) ? True : False;
438       res_is_negative = (dst_dp[output_vec_idx] & 0x8000000000000000ULL) ? True : False;
439       memcpy(&res_dp, &dst_dp[output_vec_idx], 8);
440    } else {
441       unsigned int * src_sp_uint;
442       dst_sp = (unsigned int *) &vec_out;
443       src_sp = spec_sp_fargs[idx];
444       src_sp_uint = (unsigned int *) &src_sp;
445       src_is_negative = (*src_sp_uint & 0x80000000) ? True : False;
446       res_is_negative = (dst_sp[output_vec_idx] & 0x80000000) ? True : False;
447       memcpy(&res_sp, &dst_sp[output_vec_idx], 4);
448    }
449 
450    // Below are common rules for xvre{d|s}p and xvrsqrte{d|s}p
451    if (isnan(SRC))
452       return isnan(RES);
453    if (fpclassify(SRC) == FP_ZERO)
454       return isinf(RES);
455    if (!src_is_negative && isinf(SRC))
456       return !res_is_negative && (fpclassify(RES) == FP_ZERO);
457    if (is_rsqrte) {
458       if (src_is_negative)
459          return isnan(RES);
460    } else {
461       if (src_is_negative && isinf(SRC))
462          return res_is_negative && (fpclassify(RES) == FP_ZERO);
463    }
464    if (dp_test) {
465       double calc_diff;
466       double real_diff;
467       double recip_divisor;
468       double div_result;
469       double calc_diff_tmp;
470 
471       if (is_rsqrte)
472          recip_divisor = sqrt(src_dp);
473       else
474          recip_divisor = src_dp;
475 
476       div_result = 1.0/recip_divisor;
477       calc_diff_tmp = recip_divisor * 16384.0;
478       if (isnormal(calc_diff_tmp)) {
479          calc_diff = fabs(1.0/calc_diff_tmp);
480          real_diff = fabs(res_dp - div_result);
481          result = ( ( res_dp == div_result )
482                   || ( real_diff <= calc_diff ) );
483       } else {
484          /* Unable to compute theoretical difference, so we fall back to masking out
485           * un-precise bits.
486           */
487          unsigned long long * div_result_dp = (unsigned long long *) &div_result;
488          result = (dst_dp[output_vec_idx] & VSX_RECIP_ESTIMATE_MASK_DP) == (*div_result_dp & VSX_RECIP_ESTIMATE_MASK_DP);
489       }
490       /* For debug use . . .
491          if (!result) {
492              unsigned long long * dv = &div_result;
493              unsigned long long * rd = &real_diff;
494              unsigned long long * cd = &calc_diff;
495              printf("\n\t {actual div_result: %016llx; real_diff:  %016llx; calc_diff:  %016llx}\n",
496        *dv, *rd, *cd);
497           }
498        */
499    } else {  // single precision test (only have xvrsqrtesp, since xvresp was implemented in stage 2)
500       float calc_diff;
501       float real_diff;
502       float div_result;
503       float calc_diff_tmp;
504       float recip_divisor = sqrt(src_sp);
505 
506       div_result = 1.0/recip_divisor;
507       calc_diff_tmp = recip_divisor * 16384.0;
508       if (isnormal(calc_diff_tmp)) {
509          calc_diff = fabsf(1.0/calc_diff_tmp);
510          real_diff = fabsf(res_sp - div_result);
511          result = ( ( res_sp == div_result )
512                   || ( real_diff <= calc_diff ) );
513       } else {
514          /* Unable to compute theoretical difference, so we fall back to masking out
515           * un-precise bits.
516           */
517          unsigned int * div_result_sp = (unsigned int *) &div_result;
518          result = (dst_sp[output_vec_idx] & VSX_RECIP_ESTIMATE_MASK_SP) == (*div_result_sp & VSX_RECIP_ESTIMATE_MASK_SP);
519       }
520       /* For debug use . . .
521          if (!result) {
522              unsigned long long * dv = &div_result;
523              unsigned long long * rd = &real_diff;
524              unsigned long long * cd = &calc_diff;
525              printf("\n\t {actual div_result: %016llx; real_diff:  %016llx; calc_diff:  %016llx}\n",
526        *dv, *rd, *cd);
527           }
528        */
529    }
530    return result;
531 }
532 
533 typedef struct vx_fp_test
534 {
535    test_func_t test_func;
536    const char * name;
537    fp_test_args_t * targs;
538    int num_tests;
539    precision_type_t precision;
540    vx_fp_test_type type;
541    const char * op;
542 } vx_fp_test_t;
543 
544 
545 static Bool do_dot;
546 
test_xvredp(void)547 static void test_xvredp(void)
548 {
549    __asm__ __volatile__ ("xvredp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
550 }
551 
test_xsredp(void)552 static void test_xsredp(void)
553 {
554    __asm__ __volatile__ ("xsredp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
555 }
556 
test_xvrsqrtedp(void)557 static void test_xvrsqrtedp(void)
558 {
559    __asm__ __volatile__ ("xvrsqrtedp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
560 }
561 
test_xsrsqrtedp(void)562 static void test_xsrsqrtedp(void)
563 {
564    __asm__ __volatile__ ("xsrsqrtedp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
565 }
566 
test_xvrsqrtesp(void)567 static void test_xvrsqrtesp(void)
568 {
569    __asm__ __volatile__ ("xvrsqrtesp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
570 }
571 
test_xstsqrtdp(void)572 static void test_xstsqrtdp(void)
573 {
574    __asm__ __volatile__ ("xstsqrtdp   cr1, %x0" : : "wa" (vec_inB));
575 }
576 
test_xvtsqrtdp(void)577 static void test_xvtsqrtdp(void)
578 {
579    __asm__ __volatile__ ("xvtsqrtdp   cr1, %x0" : : "wa" (vec_inB));
580 }
581 
test_xvtsqrtsp(void)582 static void test_xvtsqrtsp(void)
583 {
584    __asm__ __volatile__ ("xvtsqrtsp   cr1, %x0" : : "wa" (vec_inB));
585 }
586 
test_xvsqrtdp(void)587 static void test_xvsqrtdp(void)
588 {
589    __asm__ __volatile__ ("xvsqrtdp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
590 }
591 
test_xvsqrtsp(void)592 static void test_xvsqrtsp(void)
593 {
594    __asm__ __volatile__ ("xvsqrtsp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
595 }
596 
test_xvtdivdp(void)597 static void test_xvtdivdp(void)
598 {
599    __asm__ __volatile__ ("xvtdivdp   cr1, %x0, %x1" : : "wa" (vec_inA), "wa" (vec_inB));
600 }
601 
test_xvtdivsp(void)602 static void test_xvtdivsp(void)
603 {
604    __asm__ __volatile__ ("xvtdivsp   cr1, %x0, %x1" : : "wa" (vec_inA), "wa" (vec_inB));
605 }
606 
test_xscvdpsp(void)607 static void test_xscvdpsp(void)
608 {
609    __asm__ __volatile__ ("xscvdpsp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
610 }
611 
test_xscvdpuxws(void)612 static void test_xscvdpuxws(void)
613 {
614    __asm__ __volatile__ ("xscvdpuxws   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
615 }
616 
test_xscvspdp(void)617 static void test_xscvspdp(void)
618 {
619    __asm__ __volatile__ ("xscvspdp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
620 }
621 
test_xvcvdpsp(void)622 static void test_xvcvdpsp(void)
623 {
624    __asm__ __volatile__ ("xvcvdpsp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
625 }
626 
test_xvcvdpuxds(void)627 static void test_xvcvdpuxds(void)
628 {
629    __asm__ __volatile__ ("xvcvdpuxds   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
630 }
631 
test_xvcvdpuxws(void)632 static void test_xvcvdpuxws(void)
633 {
634    __asm__ __volatile__ ("xvcvdpuxws   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
635 }
636 
test_xvcvspdp(void)637 static void test_xvcvspdp(void)
638 {
639    __asm__ __volatile__ ("xvcvspdp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
640 }
641 
test_xvcvspsxds(void)642 static void test_xvcvspsxds(void)
643 {
644    __asm__ __volatile__ ("xvcvspsxds   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
645 }
646 
test_xvcvspuxds(void)647 static void test_xvcvspuxds(void)
648 {
649    __asm__ __volatile__ ("xvcvspuxds   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
650 }
651 
test_xvcvdpsxds(void)652 static void test_xvcvdpsxds(void)
653 {
654    __asm__ __volatile__ ("xvcvdpsxds   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
655 }
656 
test_xvcvspuxws(void)657 static void test_xvcvspuxws(void)
658 {
659    __asm__ __volatile__ ("xvcvspuxws   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
660 }
661 
test_xvcvsxddp(void)662 static void test_xvcvsxddp(void)
663 {
664    __asm__ __volatile__ ("xvcvsxddp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
665 }
666 
test_xvcvuxddp(void)667 static void test_xvcvuxddp(void)
668 {
669    __asm__ __volatile__ ("xvcvuxddp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
670 }
671 
test_xvcvsxdsp(void)672 static void test_xvcvsxdsp(void)
673 {
674    __asm__ __volatile__ ("xvcvsxdsp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
675 }
676 
test_xvcvuxdsp(void)677 static void test_xvcvuxdsp(void)
678 {
679    __asm__ __volatile__ ("xvcvuxdsp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
680 }
681 
test_xvcvsxwdp(void)682 static void test_xvcvsxwdp(void)
683 {
684    __asm__ __volatile__ ("xvcvsxwdp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
685 }
686 
test_xvcvuxwdp(void)687 static void test_xvcvuxwdp(void)
688 {
689    __asm__ __volatile__ ("xvcvuxwdp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
690 }
691 
test_xvcvsxwsp(void)692 static void test_xvcvsxwsp(void)
693 {
694    __asm__ __volatile__ ("xvcvsxwsp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
695 }
696 
test_xvcvuxwsp(void)697 static void test_xvcvuxwsp(void)
698 {
699    __asm__ __volatile__ ("xvcvuxwsp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
700 }
701 
test_xsrdpic(void)702 static void test_xsrdpic(void)
703 {
704    __asm__ __volatile__ ("xsrdpic   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
705 }
706 
test_xsrdpiz(void)707 static void test_xsrdpiz(void)
708 {
709    __asm__ __volatile__ ("xsrdpiz   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
710 }
711 
test_xsrdpi(void)712 static void test_xsrdpi(void)
713 {
714    __asm__ __volatile__ ("xsrdpi   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
715 }
716 
test_xvabsdp(void)717 static void test_xvabsdp(void)
718 {
719    __asm__ __volatile__ ("xvabsdp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
720 }
721 
test_xvnabsdp(void)722 static void test_xvnabsdp(void)
723 {
724    __asm__ __volatile__ ("xvnabsdp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
725 }
726 
test_xvnegdp(void)727 static void test_xvnegdp(void)
728 {
729    __asm__ __volatile__ ("xvnegdp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
730 }
731 
test_xvabssp(void)732 static void test_xvabssp(void)
733 {
734    __asm__ __volatile__ ("xvabssp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
735 }
736 
test_xvnabssp(void)737 static void test_xvnabssp(void)
738 {
739    __asm__ __volatile__ ("xvnabssp   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
740 }
741 
test_xvrdpi(void)742 static void test_xvrdpi(void)
743 {
744    __asm__ __volatile__ ("xvrdpi   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
745 }
746 
test_xvrdpic(void)747 static void test_xvrdpic(void)
748 {
749    __asm__ __volatile__ ("xvrdpic   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
750 }
751 
test_xvrdpim(void)752 static void test_xvrdpim(void)
753 {
754    __asm__ __volatile__ ("xvrdpim   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
755 }
756 
test_xvrdpip(void)757 static void test_xvrdpip(void)
758 {
759    __asm__ __volatile__ ("xvrdpip   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
760 }
761 
test_xvrdpiz(void)762 static void test_xvrdpiz(void)
763 {
764    __asm__ __volatile__ ("xvrdpiz   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
765 }
766 
test_xvrspi(void)767 static void test_xvrspi(void)
768 {
769    __asm__ __volatile__ ("xvrspi   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
770 }
771 
test_xvrspic(void)772 static void test_xvrspic(void)
773 {
774    __asm__ __volatile__ ("xvrspic   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
775 }
776 
test_xvrspim(void)777 static void test_xvrspim(void)
778 {
779    __asm__ __volatile__ ("xvrspim   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
780 }
781 
test_xvrspip(void)782 static void test_xvrspip(void)
783 {
784    __asm__ __volatile__ ("xvrspip   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
785 }
786 
test_xvrspiz(void)787 static void test_xvrspiz(void)
788 {
789    __asm__ __volatile__ ("xvrspiz   %x0, %x1" : "=wa" (vec_out): "wa" (vec_inB));
790 }
791 
792 static vx_fp_test_t
793 vsx_one_fp_arg_tests[] = {
794                                 { &test_xvredp, "xvredp", NULL, 18, DOUBLE_TEST, VX_ESTIMATE, "1/x"},
795                                 { &test_xsredp, "xsredp", NULL, 18, DOUBLE_TEST, VX_ESTIMATE, "1/x"},
796                                 { &test_xvrsqrtedp, "xvrsqrtedp", NULL, 18, DOUBLE_TEST, VX_ESTIMATE, "1/x-sqrt"},
797                                 { &test_xsrsqrtedp, "xsrsqrtedp", NULL, 18, DOUBLE_TEST, VX_ESTIMATE, "1/x-sqrt"},
798                                 { &test_xvrsqrtesp, "xvrsqrtesp", NULL, 18, SINGLE_TEST, VX_ESTIMATE, "1/x-sqrt"},
799                                 { &test_xvsqrtdp, "xvsqrtdp", NULL, 18, DOUBLE_TEST, VX_DEFAULT, "sqrt"},
800                                 { &test_xvsqrtsp, "xvsqrtsp", NULL, 18, SINGLE_TEST, VX_DEFAULT, "sqrt"},
801                                 { &test_xscvdpsp, "xscvdpsp", NULL, 20, DOUBLE_TEST, VX_CONV_TO_SINGLE, "conv"},
802                                 { &test_xscvdpuxws, "xscvdpuxws", NULL, 20, DOUBLE_TEST, VX_SCALAR_CONV_TO_WORD, "conv"},
803                                 { &test_xscvspdp, "xscvspdp", NULL, 20, SINGLE_TEST, VX_CONV_TO_DOUBLE, "conv"},
804                                 { &test_xvcvdpsp, "xvcvdpsp", NULL, 20, DOUBLE_TEST, VX_CONV_TO_SINGLE, "conv"},
805                                 { &test_xvcvdpuxds, "xvcvdpuxds", NULL, 20, DOUBLE_TEST, VX_CONV_TO_DOUBLE, "conv"},
806                                 { &test_xvcvdpuxws, "xvcvdpuxws", NULL, 20, DOUBLE_TEST, VX_CONV_TO_SINGLE, "conv"},
807                                 { &test_xvcvspdp, "xvcvspdp", NULL, 20, SINGLE_TEST, VX_CONV_TO_DOUBLE, "conv"},
808                                 { &test_xvcvspsxds, "xvcvspsxds", NULL, 20, SINGLE_TEST, VX_CONV_TO_DOUBLE, "conv"},
809                                 { &test_xvcvdpsxds, "xvcvdpsxds", NULL, 20, DOUBLE_TEST, VX_CONV_TO_DOUBLE, "conv"},
810                                 { &test_xvcvspuxds, "xvcvspuxds", NULL, 20, SINGLE_TEST, VX_CONV_TO_DOUBLE, "conv"},
811                                 { &test_xvcvspuxws, "xvcvspuxws", NULL, 20, SINGLE_TEST, VX_CONV_TO_SINGLE, "conv"},
812                                 { &test_xsrdpic, "xsrdpic", NULL, 20, DOUBLE_TEST, VX_CONV_TO_DOUBLE, "round"},
813                                 { &test_xsrdpiz, "xsrdpiz", NULL, 20, DOUBLE_TEST, VX_CONV_TO_DOUBLE, "round"},
814                                 { &test_xsrdpi, "xsrdpi", NULL, 20, DOUBLE_TEST, VX_CONV_TO_DOUBLE, "round"},
815                                 { &test_xvabsdp, "xvabsdp", NULL, 20, DOUBLE_TEST, VX_DEFAULT, "abs"},
816                                 { &test_xvnabsdp, "xvnabsdp", NULL, 20, DOUBLE_TEST, VX_DEFAULT, "nabs"},
817                                 { &test_xvnegdp, "xvnegdp", NULL, 20, DOUBLE_TEST, VX_DEFAULT, "neg"},
818                                 { &test_xvabssp, "xvabssp", NULL, 20, SINGLE_TEST, VX_DEFAULT, "abs"},
819                                 { &test_xvnabssp, "xvnabssp", NULL, 20, SINGLE_TEST, VX_DEFAULT, "nabs"},
820                                 { &test_xvrdpi,  "xvrdpi",  NULL, 20, DOUBLE_TEST, VX_CONV_TO_DOUBLE, "round"},
821                                 { &test_xvrdpic, "xvrdpic", NULL, 20, DOUBLE_TEST, VX_CONV_TO_DOUBLE, "round"},
822                                 { &test_xvrdpim, "xvrdpim", NULL, 20, DOUBLE_TEST, VX_CONV_TO_DOUBLE, "round"},
823                                 { &test_xvrdpip, "xvrdpip", NULL, 20, DOUBLE_TEST, VX_CONV_TO_DOUBLE, "round"},
824                                 { &test_xvrdpiz, "xvrdpiz", NULL, 20, DOUBLE_TEST, VX_CONV_TO_DOUBLE, "round"},
825                                 { &test_xvrspi,  "xvrspi",  NULL, 20, SINGLE_TEST, VX_CONV_TO_SINGLE, "round"},
826                                 { &test_xvrspic, "xvrspic", NULL, 20, SINGLE_TEST, VX_CONV_TO_SINGLE, "round"},
827                                 { &test_xvrspim, "xvrspim", NULL, 20, SINGLE_TEST, VX_CONV_TO_SINGLE, "round"},
828                                 { &test_xvrspip, "xvrspip", NULL, 20, SINGLE_TEST, VX_CONV_TO_SINGLE, "round"},
829                                 { &test_xvrspiz, "xvrspiz", NULL, 20, SINGLE_TEST, VX_CONV_TO_SINGLE, "round"},
830                                 { NULL, NULL, NULL, 0, 0, 0, NULL}
831 };
832 
833 static vx_fp_test_t
834 vx_tdivORtsqrt_tests[] = {
835                           { &test_xstsqrtdp, "xstsqrtdp", NULL, 20, DOUBLE_TEST, VX_DEFAULT, "test-sqrt"},
836                           { &test_xvtsqrtdp, "xvtsqrtdp", NULL, 20, DOUBLE_TEST, VX_DEFAULT, "test-sqrt"},
837                           { &test_xvtsqrtsp, "xvtsqrtsp", NULL, 20, SINGLE_TEST, VX_DEFAULT, "test-sqrt"},
838                           { &test_xvtdivdp, "xvtdivdp", two_arg_fp_tests, 68, DOUBLE_TEST, VX_DEFAULT, "test-div"},
839                           { &test_xvtdivsp, "xvtdivsp", two_arg_fp_tests, 68, SINGLE_TEST, VX_DEFAULT, "test-div"},
840                           { NULL, NULL, NULL, 0 , 0, 0, NULL}
841 };
842 
843 static unsigned long long doubleWord[] = { 0,
844                                   0xffffffff00000000LL,
845                                   0x00000000ffffffffLL,
846                                   0xffffffffffffffffLL,
847                                   0x89abcde123456789LL,
848                                   0x0102030405060708LL,
849                                   0x00000000a0b1c2d3LL,
850                                   0x1111222233334444LL
851 };
852 
853 static unsigned int singleWord[] = {0,
854                                   0xffff0000,
855                                   0x0000ffff,
856                                   0xffffffff,
857                                   0x89a73522,
858                                   0x01020304,
859                                   0x0000abcd,
860                                   0x11223344
861 };
862 
863 typedef struct vx_intToFp_test
864 {
865    test_func_t test_func;
866    const char * name;
867    void * targs;
868    int num_tests;
869    precision_type_t precision;
870    vx_fp_test_type type;
871 } vx_intToFp_test_t;
872 
873 static vx_intToFp_test_t
874 intToFp_tests[] = {
875                    { test_xvcvsxddp, "xvcvsxddp", (void *)doubleWord, 8, DOUBLE_TEST, VX_CONV_TO_DOUBLE },
876                    { test_xvcvuxddp, "xvcvuxddp", (void *)doubleWord, 8, DOUBLE_TEST, VX_CONV_TO_DOUBLE },
877                    { test_xvcvsxdsp, "xvcvsxdsp", (void *)doubleWord, 8, DOUBLE_TEST, VX_CONV_TO_SINGLE },
878                    { test_xvcvuxdsp, "xvcvuxdsp", (void *)doubleWord, 8, DOUBLE_TEST, VX_CONV_TO_SINGLE },
879                    { test_xvcvsxwdp, "xvcvsxwdp", (void *)singleWord, 8, SINGLE_TEST, VX_CONV_TO_DOUBLE },
880                    { test_xvcvuxwdp, "xvcvuxwdp", (void *)singleWord, 8, SINGLE_TEST, VX_CONV_TO_DOUBLE },
881                    { test_xvcvsxwsp, "xvcvsxwsp", (void *)singleWord, 8, SINGLE_TEST, VX_CONV_TO_SINGLE },
882                    { test_xvcvuxwsp, "xvcvuxwsp", (void *)singleWord, 8, SINGLE_TEST, VX_CONV_TO_SINGLE },
883                    { NULL, NULL, NULL, 0, 0 }
884 };
885 
886 static Bool do_OE;
887 typedef enum {
888    DIV_BASE = 1,
889    DIV_OE = 2,
890    DIV_DOT = 4,
891 } div_type_t;
892 /* Possible divde type combinations are:
893  *   - base
894  *   - base+dot
895  *   - base+OE
896  *   - base+OE+dot
897  */
898 #ifdef __powerpc64__
test_divdeu(void)899 static void test_divdeu(void)
900 {
901    int divdeu_type = DIV_BASE;
902    if (do_OE)
903       divdeu_type |= DIV_OE;
904    if (do_dot)
905       divdeu_type |= DIV_DOT;
906 
907    switch (divdeu_type) {
908       case 1:
909         SET_CR_XER_ZERO;
910          __asm__ __volatile__ ("divdeu %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15));
911          GET_CR_XER(div_flags, div_xer);
912          break;
913       case 3:
914         SET_CR_XER_ZERO;
915          __asm__ __volatile__ ("divdeuo %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15));
916          GET_CR_XER(div_flags, div_xer);
917          break;
918       case 5:
919         SET_CR_XER_ZERO;
920          __asm__ __volatile__ ("divdeu. %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15));
921          GET_CR_XER(div_flags, div_xer);
922          break;
923       case 7:
924         SET_CR_XER_ZERO;
925          __asm__ __volatile__ ("divdeuo. %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15));
926          GET_CR_XER(div_flags, div_xer);
927          break;
928       default:
929          fprintf(stderr, "Invalid divdeu type. Exiting\n");
930          exit(1);
931    }
932 }
933 #endif
934 
test_divwe(void)935 static void test_divwe(void)
936 {
937    int divwe_type = DIV_BASE;
938    if (do_OE)
939       divwe_type |= DIV_OE;
940    if (do_dot)
941       divwe_type |= DIV_DOT;
942 
943    switch (divwe_type) {
944       case 1:
945         SET_CR_XER_ZERO;
946          __asm__ __volatile__ ("divwe %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15));
947          GET_CR_XER(div_flags, div_xer);
948          break;
949       case 3:
950         SET_CR_XER_ZERO;
951          __asm__ __volatile__ ("divweo %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15));
952          GET_CR_XER(div_flags, div_xer);
953          break;
954       case 5:
955         SET_CR_XER_ZERO;
956          __asm__ __volatile__ ("divwe. %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15));
957          GET_CR_XER(div_flags, div_xer);
958          break;
959       case 7:
960         SET_CR_XER_ZERO;
961          __asm__ __volatile__ ("divweo. %0, %1, %2" : "=r" (r17) : "r" (r14),"r" (r15));
962          GET_CR_XER(div_flags, div_xer);
963          break;
964       default:
965          fprintf(stderr, "Invalid divweu type. Exiting\n");
966          exit(1);
967    }
968 }
969 
970 
971 typedef struct simple_test {
972    test_func_t test_func;
973    char * name;
974    precision_type_t precision;
975 } simple_test_t;
976 
977 
setup_sp_fp_args(fp_test_args_t * targs,Bool swap_inputs)978 static void setup_sp_fp_args(fp_test_args_t * targs, Bool swap_inputs)
979 {
980    int a_idx, b_idx, i;
981    void * inA, * inB;
982    void * vec_src = swap_inputs ? &vec_out : &vec_inB;
983 
984    for (i = 0; i < 4; i++) {
985       a_idx = targs->fra_idx;
986       b_idx = targs->frb_idx;
987       inA = (void *)&spec_sp_fargs[a_idx];
988       inB = (void *)&spec_sp_fargs[b_idx];
989       // copy single precision FP  into vector element i
990       memcpy(((void *)&vec_inA) + (i * 4), inA, 4);
991       memcpy(vec_src + (i * 4), inB, 4);
992       targs++;
993    }
994 }
995 
setup_dp_fp_args(fp_test_args_t * targs,Bool swap_inputs)996 static void setup_dp_fp_args(fp_test_args_t * targs, Bool swap_inputs)
997 {
998    int a_idx, b_idx, i;
999    void * inA, * inB;
1000    void * vec_src = swap_inputs ? (void *)&vec_out : (void *)&vec_inB;
1001 
1002    for (i = 0; i < 2; i++) {
1003       a_idx = targs->fra_idx;
1004       b_idx = targs->frb_idx;
1005       inA = (void *)&spec_fargs[a_idx];
1006       inB = (void *)&spec_fargs[b_idx];
1007       // copy double precision FP  into vector element i
1008       memcpy(((void *)&vec_inA) + (i * 8), inA, 8);
1009       memcpy(vec_src + (i * 8), inB, 8);
1010       targs++;
1011    }
1012 }
1013 
1014 #define VX_NOT_CMP_OP 0xffffffff
print_vector_fp_result(unsigned int cc,vx_fp_test_t * test_group,int i,Bool print_vec_out)1015 static void print_vector_fp_result(unsigned int cc, vx_fp_test_t * test_group, int i, Bool print_vec_out)
1016 {
1017    int a_idx, b_idx, k;
1018    char * name = malloc(20);
1019    int dp = test_group->precision == DOUBLE_TEST ? 1 : 0;
1020    int loops = dp ? 2 : 4;
1021    fp_test_args_t * targs = &test_group->targs[i];
1022    unsigned long long * frA_dp, * frB_dp, * dst_dp;
1023    unsigned int * frA_sp, *frB_sp, * dst_sp;
1024    strcpy(name, test_group->name);
1025    printf("#%d: %s%s ", dp? i/2 : i/4, name, (do_dot ? "." : ""));
1026    for (k = 0; k < loops; k++) {
1027       a_idx = targs->fra_idx;
1028       b_idx = targs->frb_idx;
1029       if (k)
1030          printf(" AND ");
1031       if (dp) {
1032          frA_dp = (unsigned long long *)&spec_fargs[a_idx];
1033          frB_dp = (unsigned long long *)&spec_fargs[b_idx];
1034          printf("%016llx %s %016llx", *frA_dp, test_group->op, *frB_dp);
1035       } else {
1036          frA_sp = (unsigned int *)&spec_sp_fargs[a_idx];
1037          frB_sp = (unsigned int *)&spec_sp_fargs[b_idx];
1038          printf("%08x %s %08x", *frA_sp, test_group->op, *frB_sp);
1039       }
1040       targs++;
1041    }
1042    if (cc != VX_NOT_CMP_OP)
1043       printf(" ? cc=%x", cc);
1044 
1045    if (print_vec_out) {
1046       if (dp) {
1047          dst_dp = (unsigned long long *) &vec_out;
1048          printf(" => %016llx %016llx\n", dst_dp[0], dst_dp[1]);
1049       } else {
1050          dst_sp = (unsigned int *) &vec_out;
1051          printf(" => %08x %08x %08x %08x\n", dst_sp[0], dst_sp[1], dst_sp[2], dst_sp[3]);
1052       }
1053    } else {
1054       printf("\n");
1055    }
1056    free(name);
1057 }
1058 
1059 
1060 
test_vsx_one_fp_arg(void)1061 static void test_vsx_one_fp_arg(void)
1062 {
1063    test_func_t func;
1064    int k;
1065    k = 0;
1066    build_special_fargs_table();
1067 
1068    while ((func = vsx_one_fp_arg_tests[k].test_func)) {
1069       int idx, i;
1070       vx_fp_test_t test_group = vsx_one_fp_arg_tests[k];
1071       Bool estimate = (test_group.type == VX_ESTIMATE);
1072       Bool dp = (test_group.precision == DOUBLE_TEST) ? True : False;
1073       Bool is_sqrt = (strstr(test_group.name, "sqrt")) ? True : False;
1074       Bool is_scalar = (strstr(test_group.name, "xs")) ? True : False;
1075       Bool sparse_sp = False;
1076       int stride = dp ? 2 : 4;
1077       int loops = is_scalar ? 1 : stride;
1078       stride = is_scalar ? 1: stride;
1079 
1080       /* For conversions of single to double, the 128-bit input register is sparsely populated:
1081        *    |___ SP___|_Unused_|___SP___|__Unused__|   // for vector op
1082        *                     or
1083        *    |___ SP___|_Unused_|_Unused_|__Unused__|   // for scalar op
1084        *
1085        * For the vector op case, we need to adjust stride from '4' to '2', since
1086        * we'll only be loading two values per loop into the input register.
1087        */
1088       if (!dp && !is_scalar && test_group.type == VX_CONV_TO_DOUBLE) {
1089          sparse_sp = True;
1090          stride = 2;
1091       }
1092 
1093       for (i = 0; i < test_group.num_tests; i+=stride) {
1094          unsigned int * pv;
1095          void * inB;
1096 
1097          pv = (unsigned int *)&vec_out;
1098          // clear vec_out
1099          for (idx = 0; idx < 4; idx++, pv++)
1100             *pv = 0;
1101 
1102          if (dp) {
1103             int j;
1104             unsigned long long * frB_dp, *dst_dp;
1105             for (j = 0; j < loops; j++) {
1106                inB = (void *)&spec_fargs[i + j];
1107                // copy double precision FP into vector element i
1108                memcpy(((void *)&vec_inB) + (j * 8), inB, 8);
1109             }
1110             // execute test insn
1111             (*func)();
1112             dst_dp = (unsigned long long *) &vec_out;
1113             printf("#%d: %s ", i/stride, test_group.name);
1114             for (j = 0; j < loops; j++) {
1115                if (j)
1116                   printf("; ");
1117                frB_dp = (unsigned long long *)&spec_fargs[i + j];
1118                printf("%s(%016llx)", test_group.op, *frB_dp);
1119                if (estimate) {
1120                   Bool res = check_estimate(DOUBLE_TEST, is_sqrt, i + j, j);
1121                   printf(" ==> %s)", res ? "PASS" : "FAIL");
1122                   /* For debugging . . .
1123                    printf(" ==> %s (res=%016llx)", res ? "PASS" : "FAIL", dst_dp[j]);
1124                    */
1125                } else {
1126                   vx_fp_test_type type = test_group.type;
1127                   switch (type) {
1128                      case VX_SCALAR_CONV_TO_WORD:
1129                         printf(" = %016llx", dst_dp[j] & 0x00000000ffffffffULL);
1130                         break;
1131                      case VX_CONV_TO_SINGLE:
1132                         printf(" = %016llx", dst_dp[j] & 0xffffffff00000000ULL);
1133                         break;
1134                      default:  // For VX_CONV_TO_DOUBLE and non-convert instructions . . .
1135                         printf(" = %016llx", dst_dp[j]);
1136                   }
1137                }
1138             }
1139             printf("\n");
1140          } else {
1141             int j, skip_slot;
1142             unsigned int * frB_sp, * dst_sp = NULL;
1143             unsigned long long * dst_dp = NULL;
1144             if (sparse_sp) {
1145                skip_slot = 1;
1146                loops = 2;
1147             } else {
1148                skip_slot = 0;
1149             }
1150             for (j = 0; j < loops; j++) {
1151                inB = (void *)&spec_sp_fargs[i + j];
1152                // copy single precision FP into vector element i
1153                if (skip_slot && j > 0)
1154                   memcpy(((void *)&vec_inB) + ((j + j) * 4), inB, 4);
1155                else
1156                   memcpy(((void *)&vec_inB) + (j * 4), inB, 4);
1157             }
1158             // execute test insn
1159             (*func)();
1160             if (test_group.type == VX_CONV_TO_DOUBLE)
1161                dst_dp = (unsigned long long *) &vec_out;
1162             else
1163                dst_sp = (unsigned int *) &vec_out;
1164             // print result
1165             printf("#%d: %s ", i/stride, test_group.name);
1166             for (j = 0; j < loops; j++) {
1167                if (j)
1168                   printf("; ");
1169                frB_sp = (unsigned int *)&spec_sp_fargs[i + j];
1170                printf("%s(%08x)", test_group.op, *frB_sp);
1171                if (estimate) {
1172                   Bool res = check_estimate(SINGLE_TEST, is_sqrt, i + j, j);
1173                   printf(" ==> %s)", res ? "PASS" : "FAIL");
1174                } else {
1175                   if (test_group.type == VX_CONV_TO_DOUBLE)
1176                         printf(" = %016llx", dst_dp[j]);
1177                   else
1178                   /* Special case: Current VEX implementation for fsqrts (single precision)
1179                    * uses the same implementation as that used for double precision fsqrt.
1180                    * However, I've found that for xvsqrtsp, the result from that implementation
1181                    * may be off by the two LSBs.  Generally, even this small inaccuracy can cause the
1182                    * output to appear very different if you end up with a carry.  But for the given
1183                    * inputs in this testcase, we can simply mask out these bits.
1184                    */
1185                      printf(" = %08x", is_sqrt ? (dst_sp[j] & 0xfffffffc) : dst_sp[j]);
1186                }
1187             }
1188             printf("\n");
1189          }
1190       }
1191       k++;
1192       printf( "\n" );
1193    }
1194 }
1195 
test_int_to_fp_convert(void)1196 static void test_int_to_fp_convert(void)
1197 {
1198    test_func_t func;
1199    int k;
1200    k = 0;
1201 
1202    while ((func = intToFp_tests[k].test_func)) {
1203       int idx, i;
1204       vx_intToFp_test_t test_group = intToFp_tests[k];
1205       Bool dp = (test_group.precision == DOUBLE_TEST) ? True : False;
1206       Bool sparse_sp = False;
1207       int stride = dp ? 2 : 4;
1208       int loops = stride;
1209 
1210       /* For conversions of single to double, the 128-bit input register is sparsely populated:
1211        *    |___ int___|_Unused_|___int___|__Unused__|   // for vector op
1212        *                     or
1213        * We need to adjust stride from '4' to '2', since we'll only be loading
1214        * two values per loop into the input register.
1215        */
1216       if (!dp && test_group.type == VX_CONV_TO_DOUBLE) {
1217          sparse_sp = True;
1218          stride = 2;
1219       }
1220 
1221       for (i = 0; i < test_group.num_tests; i+=stride) {
1222          unsigned int * pv;
1223          void * inB;
1224 
1225          pv = (unsigned int *)&vec_out;
1226          // clear vec_out
1227          for (idx = 0; idx < 4; idx++, pv++)
1228             *pv = 0;
1229 
1230          if (dp) {
1231             int j;
1232             unsigned long long  *dst_dw, * targs = test_group.targs;
1233             for (j = 0; j < loops; j++) {
1234                inB = (void *)&targs[i + j];
1235                // copy doubleword into vector element i
1236                memcpy(((void *)&vec_inB) + (j * 8), inB, 8);
1237             }
1238             // execute test insn
1239             (*func)();
1240             dst_dw = (unsigned long long *) &vec_out;
1241             printf("#%d: %s ", i/stride, test_group.name);
1242             for (j = 0; j < loops; j++) {
1243                if (j)
1244                   printf("; ");
1245                printf("conv(%016llx)", targs[i + j]);
1246 
1247                if (test_group.type == VX_CONV_TO_SINGLE)
1248                   printf(" = %016llx", dst_dw[j] & 0xffffffff00000000ULL);
1249                else
1250                   printf(" = %016llx", dst_dw[j]);
1251             }
1252             printf("\n");
1253          } else {
1254             int j, skip_slot;
1255             unsigned int * dst_sp = NULL;
1256             unsigned int * targs = test_group.targs;
1257             unsigned long long * dst_dp = NULL;
1258             if (sparse_sp) {
1259                skip_slot = 1;
1260                loops = 2;
1261             } else {
1262                skip_slot = 0;
1263             }
1264             for (j = 0; j < loops; j++) {
1265                inB = (void *)&targs[i + j];
1266                // copy single word into vector element i
1267                if (skip_slot && j > 0)
1268                   memcpy(((void *)&vec_inB) + ((j + j) * 4), inB, 4);
1269                else
1270                   memcpy(((void *)&vec_inB) + (j * 4), inB, 4);
1271             }
1272             // execute test insn
1273             (*func)();
1274             if (test_group.type == VX_CONV_TO_DOUBLE)
1275                dst_dp = (unsigned long long *) &vec_out;
1276             else
1277                dst_sp = (unsigned int *) &vec_out;
1278             // print result
1279             printf("#%d: %s ", i/stride, test_group.name);
1280             for (j = 0; j < loops; j++) {
1281                if (j)
1282                   printf("; ");
1283                printf("conv(%08x)", targs[i + j]);
1284                if (test_group.type == VX_CONV_TO_DOUBLE)
1285                   printf(" = %016llx", dst_dp[j]);
1286                else
1287                   printf(" = %08x", dst_sp[j]);
1288             }
1289             printf("\n");
1290          }
1291       }
1292       k++;
1293       printf( "\n" );
1294    }
1295 }
1296 
1297 
1298 
1299 // The div doubleword test data
1300 signed long long div_dw_tdata[13][2] = {
1301                                        { 4, -4 },
1302                                        { 4, -3 },
1303                                        { 4, 4 },
1304                                        { 4, -5 },
1305                                        { 3, 8 },
1306                                        { 0x8000000000000000ULL, 0xa },
1307                                        { 0x50c, -1 },
1308                                        { 0x50c, -4096 },
1309                                        { 0x1234fedc, 0x8000a873 },
1310                                        { 0xabcd87651234fedcULL, 0xa123b893 },
1311                                        { 0x123456789abdcULL, 0 },
1312                                        { 0, 2 },
1313                                        { 0x77, 0xa3499 }
1314 };
1315 #define dw_tdata_len (sizeof(div_dw_tdata)/sizeof(signed long long)/2)
1316 
1317 // The div word test data
1318 unsigned int div_w_tdata[6][2] = {
1319                               { 0, 2 },
1320                               { 2, 0 },
1321                               { 0x7abc1234, 0xf0000000 },
1322                               { 0xfabc1234, 5 },
1323                               { 77, 66 },
1324                               { 5, 0xfabc1234 },
1325 };
1326 #define w_tdata_len (sizeof(div_w_tdata)/sizeof(unsigned int)/2)
1327 
1328 typedef struct div_ext_test
1329 {
1330    test_func_t test_func;
1331    const char *name;
1332    int num_tests;
1333    div_type_t div_type;
1334    precision_type_t precision;
1335 } div_ext_test_t;
1336 
1337 static div_ext_test_t div_tests[] = {
1338 #ifdef __powerpc64__
1339                                    { &test_divdeu, "divdeu", dw_tdata_len, DIV_BASE, DOUBLE_TEST },
1340                                    { &test_divdeu, "divdeuo", dw_tdata_len, DIV_OE, DOUBLE_TEST },
1341 #endif
1342                                    { &test_divwe, "divwe", w_tdata_len, DIV_BASE, SINGLE_TEST },
1343                                    { &test_divwe, "divweo", w_tdata_len, DIV_OE, SINGLE_TEST },
1344                                    { NULL, NULL, 0, 0, 0 }
1345 };
1346 
test_div_extensions(void)1347 static void test_div_extensions(void)
1348 {
1349    test_func_t func;
1350    int k;
1351    k = 0;
1352 
1353    while ((func = div_tests[k].test_func)) {
1354       int i, repeat = 1;
1355       div_ext_test_t test_group = div_tests[k];
1356       do_dot = False;
1357 
1358 again:
1359       for (i = 0; i < test_group.num_tests; i++) {
1360          unsigned int condreg;
1361 
1362          if (test_group.div_type == DIV_OE)
1363             do_OE = True;
1364          else
1365             do_OE = False;
1366 
1367          if (test_group.precision == DOUBLE_TEST) {
1368             r14 = div_dw_tdata[i][0];
1369             r15 = div_dw_tdata[i][1];
1370          } else {
1371             r14 = div_w_tdata[i][0];
1372             r15 = div_w_tdata[i][1];
1373          }
1374          // execute test insn
1375          (*func)();
1376          condreg = (div_flags & 0xf0000000) >> 28;
1377          printf("#%d: %s%s: ", i, test_group.name, do_dot ? "." : "");
1378          if (test_group.precision == DOUBLE_TEST) {
1379             printf("0x%016llx0000000000000000 / 0x%016llx = 0x%016llx;",
1380                    div_dw_tdata[i][0], div_dw_tdata[i][1], (signed long long) r17);
1381          } else {
1382             printf("0x%08x00000000 / 0x%08x = 0x%08x;",
1383                    div_w_tdata[i][0], div_w_tdata[i][1], (unsigned int) r17);
1384          }
1385          printf(" CR=%x; XER=%x\n", condreg, div_xer);
1386       }
1387       printf("\n");
1388       if (repeat) {
1389          repeat = 0;
1390          do_dot = True;
1391          goto again;
1392       }
1393       k++;
1394       printf( "\n" );
1395    }
1396 }
1397 
1398 
test_vx_tdivORtsqrt(void)1399 static void test_vx_tdivORtsqrt(void)
1400 {
1401    test_func_t func;
1402    int k, crx;
1403    unsigned int flags;
1404    k = 0;
1405    do_dot = False;
1406    build_special_fargs_table();
1407 
1408    while ((func = vx_tdivORtsqrt_tests[k].test_func)) {
1409       int idx, i;
1410       vx_fp_test_t test_group = vx_tdivORtsqrt_tests[k];
1411       Bool dp = (test_group.precision == DOUBLE_TEST) ? True : False;
1412       Bool is_scalar = (strstr(test_group.name, "xs")) ? True : False;
1413       Bool two_args = test_group.targs ?  True : False;
1414       int stride = dp ? 2 : 4;
1415       int loops = is_scalar ? 1 : stride;
1416       stride = is_scalar ? 1: stride;
1417 
1418       for (i = 0; i < test_group.num_tests; i+=stride) {
1419          unsigned int * pv;
1420          void * inB;
1421 
1422          pv = (unsigned int *)&vec_out;
1423          // clear vec_out
1424          for (idx = 0; idx < 4; idx++, pv++)
1425             *pv = 0;
1426 
1427          if (dp) {
1428             int j;
1429             unsigned long long * frB_dp;
1430             if (two_args) {
1431                setup_dp_fp_args(&test_group.targs[i], False);
1432             } else {
1433                for (j = 0; j < loops; j++) {
1434                   inB = (void *)&spec_fargs[i + j];
1435                   // copy double precision FP into vector element i
1436                   memcpy(((void *)&vec_inB) + (j * 8), inB, 8);
1437                }
1438             }
1439             // execute test insn
1440             // Must do set/get of CRs immediately before/after calling the asm func
1441             // to avoid CRs being modified by other instructions.
1442             SET_FPSCR_ZERO;
1443             SET_CR_XER_ZERO;
1444             (*func)();
1445             GET_CR(flags);
1446             // assumes using CR1
1447             crx = (flags & 0x0f000000) >> 24;
1448             if (two_args) {
1449                print_vector_fp_result(crx, &test_group, i, False/*do not print vec_out*/);
1450             } else {
1451                printf("#%d: %s ", i/stride, test_group.name);
1452                for (j = 0; j < loops; j++) {
1453                   if (j)
1454                      printf("; ");
1455                   frB_dp = (unsigned long long *)&spec_fargs[i + j];
1456                   printf("%s(%016llx)", test_group.op, *frB_dp);
1457                }
1458                printf( " ? %x (CRx)\n", crx);
1459             }
1460          } else {
1461             int j;
1462             unsigned int * frB_sp;
1463             if (two_args) {
1464                setup_sp_fp_args(&test_group.targs[i], False);
1465             } else {
1466                for (j = 0; j < loops; j++) {
1467                   inB = (void *)&spec_sp_fargs[i + j];
1468                   // copy single precision FP into vector element i
1469                   memcpy(((void *)&vec_inB) + (j * 4), inB, 4);
1470                }
1471             }
1472             // execute test insn
1473             SET_FPSCR_ZERO;
1474             SET_CR_XER_ZERO;
1475             (*func)();
1476             GET_CR(flags);
1477             crx = (flags & 0x0f000000) >> 24;
1478             // print result
1479             if (two_args) {
1480                print_vector_fp_result(crx, &test_group, i, False/*do not print vec_out*/);
1481             } else {
1482                printf("#%d: %s ", i/stride, test_group.name);
1483                for (j = 0; j < loops; j++) {
1484                   if (j)
1485                      printf("; ");
1486                   frB_sp = (unsigned int *)&spec_sp_fargs[i + j];
1487                   printf("%s(%08x)", test_group.op, *frB_sp);
1488                }
1489                printf( " ? %x (CRx)\n", crx);
1490             }
1491          }
1492       }
1493       k++;
1494       printf( "\n" );
1495    }
1496 }
1497 
1498 
test_ftsqrt(void)1499 static void test_ftsqrt(void)
1500 {
1501    int i, crx;
1502    unsigned int flags;
1503    unsigned long long * frbp;
1504    build_special_fargs_table();
1505 
1506 
1507    for (i = 0; i < nb_special_fargs; i++) {
1508       f14 = spec_fargs[i];
1509       frbp = (unsigned long long *)&spec_fargs[i];
1510       SET_FPSCR_ZERO;
1511       SET_CR_XER_ZERO;
1512       __asm__ __volatile__ ("ftsqrt           cr1, %0" : : "d" (f14));
1513       GET_CR(flags);
1514       crx = (flags & 0x0f000000) >> 24;
1515       printf( "ftsqrt: %016llx ? %x (CRx)\n", *frbp, crx);
1516    }
1517    printf( "\n" );
1518 }
1519 
1520 static void
test_popcntw(void)1521 test_popcntw(void)
1522 {
1523 #ifdef __powerpc64__
1524    uint64_t res;
1525    unsigned long long src = 0x9182736405504536ULL;
1526    r14 = src;
1527    __asm__ __volatile__ ("popcntw          %0, %1" : "=r" (res): "r" (r14));
1528    printf("popcntw: 0x%llx => 0x%016llx\n", (unsigned long long)src, (unsigned long long)res);
1529 #else
1530    uint32_t res;
1531    unsigned int src = 0x9182730E;
1532    r14 = src;
1533    __asm__ __volatile__ ("popcntw          %0, %1" : "=r" (res): "r" (r14));
1534    printf("popcntw: 0x%x => 0x%08x\n", src, (int)res);
1535 #endif
1536    printf( "\n" );
1537 }
1538 
1539 
1540 static test_table_t
1541          all_tests[] =
1542 {
1543 
1544                     { &test_vsx_one_fp_arg,
1545                       "Test VSX vector and scalar single argument instructions"} ,
1546                     { &test_int_to_fp_convert,
1547                       "Test VSX vector integer to float conversion instructions" },
1548                     { &test_div_extensions,
1549                        "Test div extensions" },
1550                     { &test_ftsqrt,
1551                        "Test ftsqrt instruction" },
1552                     { &test_vx_tdivORtsqrt,
1553                        "Test vector and scalar tdiv and tsqrt instructions" },
1554                     { &test_popcntw,
1555                        "Test popcntw instruction" },
1556                     { NULL, NULL }
1557 };
1558 #endif // HAS_VSX
1559 
main(int argc,char * argv[])1560 int main(int argc, char *argv[])
1561 {
1562 #ifdef HAS_VSX
1563 
1564    test_table_t aTest;
1565    test_func_t func;
1566    int i = 0;
1567 
1568    while ((func = all_tests[i].test_category)) {
1569       aTest = all_tests[i];
1570       printf( "%s\n", aTest.name );
1571       (*func)();
1572       i++;
1573    }
1574    if (spec_fargs)
1575      free(spec_fargs);
1576    if (spec_sp_fargs)
1577      free(spec_sp_fargs);
1578 
1579 #endif // HAS _VSX
1580 
1581    return 0;
1582 }
1583