20
20
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23
#include "../dsputil.h"
25
25
#include "gcc_fixes.h"
27
#include "dsputil_altivec.h"
30
#include <sys/sysctl.h>
31
#else /* CONFIG_DARWIN */
33
#include <exec/exec.h>
34
#include <interfaces/exec.h>
35
#include <proto/exec.h>
36
#else /* __AMIGAOS4__ */
40
static sigjmp_buf jmpbuf;
41
static volatile sig_atomic_t canjump = 0;
43
static void sigill_handler (int sig)
46
signal (sig, SIG_DFL);
51
siglongjmp (jmpbuf, 1);
53
#endif /* CONFIG_DARWIN */
54
#endif /* __AMIGAOS4__ */
27
#include "dsputil_ppc.h"
28
#include "util_altivec.h"
56
30
int sad16_x2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
59
int s __attribute__((aligned(16)));
60
const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0);
33
DECLARE_ALIGNED_16(int, s);
34
const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
61
35
vector unsigned char *tv;
62
36
vector unsigned char pix1v, pix2v, pix2iv, avgv, t5;
63
37
vector unsigned int sad;
103
77
int sad16_y2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
106
int s __attribute__((aligned(16)));
107
const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0);
80
DECLARE_ALIGNED_16(int, s);
81
const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
108
82
vector unsigned char *tv;
109
83
vector unsigned char pix1v, pix2v, pix3v, avgv, t5;
110
84
vector unsigned int sad;
163
137
int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
166
int s __attribute__((aligned(16)));
140
DECLARE_ALIGNED_16(int, s);
167
141
uint8_t *pix3 = pix2 + line_size;
168
const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0);
169
const_vector unsigned short two = (const_vector unsigned short)vec_splat_u16(2);
142
const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
143
const vector unsigned short two = (const vector unsigned short)vec_splat_u16(2);
170
144
vector unsigned char *tv, avgv, t5;
171
145
vector unsigned char pix1v, pix2v, pix3v, pix2iv, pix3iv;
172
146
vector unsigned short pix2lv, pix2hv, pix2ilv, pix2ihv;
218
192
pix3iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[1]));
221
Note that Altivec does have vec_avg, but this works on vector pairs
195
Note that AltiVec does have vec_avg, but this works on vector pairs
222
196
and rounds up. We could do avg(avg(a,b),avg(c,d)), but the rounding
223
197
would mean that, for example, avg(3,0,0,1) = 2, when it should be 1.
224
198
Instead, we have to split the pixel vectors into vectors of shorts,
264
238
int sad16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
267
int s __attribute__((aligned(16)));
268
const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
241
DECLARE_ALIGNED_16(int, s);
242
const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
269
243
vector unsigned char perm1, perm2, *pix1v, *pix2v;
270
244
vector unsigned char t1, t2, t3,t4, t5;
271
245
vector unsigned int sad;
306
280
int sad8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
309
int s __attribute__((aligned(16)));
310
const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
283
DECLARE_ALIGNED_16(int, s);
284
const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
311
285
vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
312
286
vector unsigned char t1, t2, t3,t4, t5;
313
287
vector unsigned int sad;
387
361
int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
390
int s __attribute__((aligned(16)));
391
const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
364
DECLARE_ALIGNED_16(int, s);
365
const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
392
366
vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
393
367
vector unsigned char t1, t2, t3,t4, t5;
394
368
vector unsigned int sum;
443
417
int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
446
int s __attribute__((aligned(16)));
447
const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
420
DECLARE_ALIGNED_16(int, s);
421
const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
448
422
vector unsigned char perm1, perm2, *pix1v, *pix2v;
449
423
vector unsigned char t1, t2, t3,t4, t5;
450
424
vector unsigned int sum;
845
819
blockv, temp1, temp2;
846
820
register vector unsigned short
847
821
pixelssum1, pixelssum2, temp3;
848
register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
849
register const_vector unsigned short vcone = (const_vector unsigned short)vec_splat_u16(1);
850
register const_vector unsigned short vctwo = (const_vector unsigned short)vec_splat_u16(2);
822
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
823
register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
824
register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
852
826
temp1 = vec_ld(0, pixels);
853
827
temp2 = vec_ld(16, pixels);
922
896
register vector unsigned short
923
897
pixelssum1, pixelssum2, temp3,
924
898
pixelssum3, pixelssum4, temp4;
925
register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
926
register const_vector unsigned short vctwo = (const_vector unsigned short)vec_splat_u16(2);
899
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
900
register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
928
902
POWERPC_PERF_START_COUNT(altivec_put_pixels16_xy2_num, 1);
1004
978
register vector unsigned short
1005
979
pixelssum1, pixelssum2, temp3,
1006
980
pixelssum3, pixelssum4, temp4;
1007
register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
1008
register const_vector unsigned short vcone = (const_vector unsigned short)vec_splat_u16(1);
1009
register const_vector unsigned short vctwo = (const_vector unsigned short)vec_splat_u16(2);
981
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
982
register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
983
register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
1011
985
POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
1078
1052
int hadamard8_diff8x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
1079
1053
POWERPC_PERF_DECLARE(altivec_hadamard8_diff8x8_num, 1);
1081
register const_vector unsigned char vzero =
1082
(const_vector unsigned char)vec_splat_u8(0);
1055
register const vector unsigned char vzero =
1056
(const vector unsigned char)vec_splat_u8(0);
1083
1057
register vector signed short temp0, temp1, temp2, temp3, temp4,
1084
1058
temp5, temp6, temp7;
1085
1059
POWERPC_PERF_START_COUNT(altivec_hadamard8_diff8x8_num, 1);
1087
register const_vector signed short vprod1 =(const_vector signed short)
1061
register const vector signed short vprod1 =(const vector signed short)
1088
1062
AVV( 1,-1, 1,-1, 1,-1, 1,-1);
1089
register const_vector signed short vprod2 =(const_vector signed short)
1063
register const vector signed short vprod2 =(const vector signed short)
1090
1064
AVV( 1, 1,-1,-1, 1, 1,-1,-1);
1091
register const_vector signed short vprod3 =(const_vector signed short)
1065
register const vector signed short vprod3 =(const vector signed short)
1092
1066
AVV( 1, 1, 1, 1,-1,-1,-1,-1);
1093
register const_vector unsigned char perm1 = (const_vector unsigned char)
1067
register const vector unsigned char perm1 = (const vector unsigned char)
1094
1068
AVV(0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
1095
1069
0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D);
1096
register const_vector unsigned char perm2 = (const_vector unsigned char)
1070
register const vector unsigned char perm2 = (const vector unsigned char)
1097
1071
AVV(0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
1098
1072
0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B);
1099
register const_vector unsigned char perm3 = (const_vector unsigned char)
1073
register const vector unsigned char perm3 = (const vector unsigned char)
1100
1074
AVV(0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
1101
1075
0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07);
1226
1200
temp5S REG_v(v13),
1227
1201
temp6S REG_v(v14),
1228
1202
temp7S REG_v(v15);
1229
register const_vector unsigned char vzero REG_v(v31)=
1230
(const_vector unsigned char)vec_splat_u8(0);
1203
register const vector unsigned char vzero REG_v(v31)=
1204
(const vector unsigned char)vec_splat_u8(0);
1232
register const_vector signed short vprod1 REG_v(v16)=
1233
(const_vector signed short)AVV( 1,-1, 1,-1, 1,-1, 1,-1);
1234
register const_vector signed short vprod2 REG_v(v17)=
1235
(const_vector signed short)AVV( 1, 1,-1,-1, 1, 1,-1,-1);
1236
register const_vector signed short vprod3 REG_v(v18)=
1237
(const_vector signed short)AVV( 1, 1, 1, 1,-1,-1,-1,-1);
1238
register const_vector unsigned char perm1 REG_v(v19)=
1239
(const_vector unsigned char)
1206
register const vector signed short vprod1 REG_v(v16)=
1207
(const vector signed short)AVV( 1,-1, 1,-1, 1,-1, 1,-1);
1208
register const vector signed short vprod2 REG_v(v17)=
1209
(const vector signed short)AVV( 1, 1,-1,-1, 1, 1,-1,-1);
1210
register const vector signed short vprod3 REG_v(v18)=
1211
(const vector signed short)AVV( 1, 1, 1, 1,-1,-1,-1,-1);
1212
register const vector unsigned char perm1 REG_v(v19)=
1213
(const vector unsigned char)
1240
1214
AVV(0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
1241
1215
0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D);
1242
register const_vector unsigned char perm2 REG_v(v20)=
1243
(const_vector unsigned char)
1216
register const vector unsigned char perm2 REG_v(v20)=
1217
(const vector unsigned char)
1244
1218
AVV(0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
1245
1219
0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B);
1246
register const_vector unsigned char perm3 REG_v(v21)=
1247
(const_vector unsigned char)
1220
register const vector unsigned char perm3 REG_v(v21)=
1221
(const vector unsigned char)
1248
1222
AVV(0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
1249
1223
0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07);
1422
int has_altivec(void)
1426
extern struct ExecIFace *IExec;
1428
IExec->GetCPUInfoTags(GCIT_VectorUnit, &result, TAG_DONE);
1429
if (result == VECTORTYPE_ALTIVEC) return 1;
1431
#else /* __AMIGAOS4__ */
1433
#ifdef CONFIG_DARWIN
1434
int sels[2] = {CTL_HW, HW_VECTORUNIT};
1436
size_t len = sizeof(has_vu);
1439
err = sysctl(sels, 2, &has_vu, &len, NULL, 0);
1441
if (err == 0) return (has_vu != 0);
1442
#else /* CONFIG_DARWIN */
1443
/* no Darwin, do it the brute-force way */
1444
/* this is borrowed from the libmpeg2 library */
1446
signal (SIGILL, sigill_handler);
1447
if (sigsetjmp (jmpbuf, 1)) {
1448
signal (SIGILL, SIG_DFL);
1452
asm volatile ("mtspr 256, %0\n\t"
1453
"vand %%v0, %%v0, %%v0"
1457
signal (SIGILL, SIG_DFL);
1461
#endif /* CONFIG_DARWIN */
1463
#endif /* __AMIGAOS4__ */
1466
1396
static void vorbis_inverse_coupling_altivec(float *mag, float *ang,