File openjpeg-svn470-t1-flags-mmx.patch of Package openjpeg
x
1
diff -urN -x '*.orig' -x '*.rej' -x '*~' -x '.*' OpenJPEG.orig/libopenjpeg/t1.c OpenJPEG.patched/libopenjpeg/t1.c2
--- OpenJPEG.orig/libopenjpeg/t1.c 2007-11-13 13:52:05.000000000 -06003
+++ OpenJPEG.patched/libopenjpeg/t1.c 2007-11-14 01:09:40.000000000 -06004
5
#include "opj_includes.h"6
#include "t1_luts.h"7
8
+/* Don't use MMX on amd64 */9
+/* Note that merely including mmintrin.h, even if we don't use it, changes the code gcc */10
+/* outputs on amd64, and it is measurably slower. A bug in gcc? */11
+#ifdef __amd64__12
+#undef __MMX__13
+#endif14
+15
+#ifdef __MMX__16
+#include <mmintrin.h>17
+#endif18
+19
/** @defgroup T1 T1 - Implementation of the tier-1 coding */20
/*@{*/21
22
23
static char t1_getspb(int f);24
static short t1_getnmsedec_sig(int x, int bitpos);25
static short t1_getnmsedec_ref(int x, int bitpos);26
-#ifdef __amd64__27
+#if defined(__amd64__) || defined(__MMX__)28
static INLINE void t1_updateflags(flag_t *flagsp, int s, int stride);29
#else30
static void t1_updateflags(flag_t *flagsp, int s, int stride);31
32
}33
34
#else35
+#ifdef __MMX__36
+37
+static void t1_updateflags(flag_t *flagsp, int s, int stride) {38
+ static const __v4hi mod[] = {39
+ {T1_SIG_SE, T1_SIG_E, T1_SIG_NE, 0},40
+ {T1_SIG_SE, T1_SIG_E|T1_SGN_E, T1_SIG_NE, 0},41
+ {T1_SIG_S, T1_SIG, T1_SIG_N, 0},42
+ {T1_SIG_S|T1_SGN_S, T1_SIG, T1_SIG_N|T1_SGN_N, 0},43
+ {T1_SIG_SW, T1_SIG_W, T1_SIG_NW, 0},44
+ {T1_SIG_SW, T1_SIG_W|T1_SGN_W, T1_SIG_NW, 0}45
+ };46
+47
+ __m64 tmp1 = *(__m64*)((void*)&flagsp[-1 - stride]);48
+ __m64 tmp2 = *(__m64*)((void*)&flagsp[-1 ]);49
+ __m64 tmp3 = *(__m64*)((void*)&flagsp[-1 + stride]);50
+51
+ tmp1 = _mm_or_si64(tmp1, mod[s]);52
+ tmp2 = _mm_or_si64(tmp2, mod[s+2]);53
+ tmp3 = _mm_or_si64(tmp3, mod[s+4]);54
+55
+ *(__m64*)((void*)&flagsp[-1 - stride]) = tmp1;56
+ *(__m64*)((void*)&flagsp[-1 ]) = tmp2;57
+ *(__m64*)((void*)&flagsp[-1 + stride]) = tmp3;58
+}59
+60
+#else61
62
static void t1_updateflags(flag_t *flagsp, int s, int stride) {63
static const flag_t mod[] = {64
65
}66
67
#endif68
+#endif69
70
static void t1_enc_sigpass_step(71
opj_t1_t *t1,72
73
| ((int64)(T1_SIG | T1_VISIT | T1_SIG_OTH)<<48);74
agg = !tmp;75
#else76
+ int* flagsp = (int*)&t1->flags[(k+1) + (i+1)*(t1->h+2)];77
+ agg = flagsp[1];78
if (cblksty & J2K_CCP_CBLKSTY_VSC) {79
- agg = !(t1->flags[(k+1) + (i+1)*(t1->h+2)] & (T1_SIG | T1_VISIT | T1_SIG_OTH)80
- || t1->flags[(k+2) + (i+1)*(t1->h+2)] & (T1_SIG | T1_VISIT | T1_SIG_OTH)81
- || t1->flags[(k+3) + (i+1)*(t1->h+2)] & (T1_SIG | T1_VISIT | T1_SIG_OTH)82
- || (t1->flags[(k+4) + (i+1)*(t1->h+2)] 83
- & (~(T1_SIG_S | T1_SIG_SE | T1_SIG_SW | T1_SGN_S))) & (T1_SIG | T1_VISIT | T1_SIG_OTH));84
- } else {85
- agg = !(t1->flags[(k+1) + (i+1)*(t1->h+2)] & (T1_SIG | T1_VISIT | T1_SIG_OTH)86
- || t1->flags[(k+2) + (i+1)*(t1->h+2)] & (T1_SIG | T1_VISIT | T1_SIG_OTH)87
- || t1->flags[(k+3) + (i+1)*(t1->h+2)] & (T1_SIG | T1_VISIT | T1_SIG_OTH)88
- || t1->flags[(k+4) + (i+1)*(t1->h+2)] & (T1_SIG | T1_VISIT | T1_SIG_OTH));89
+ agg &= ~((T1_SIG_S|T1_SIG_SE|T1_SIG_SW|T1_SGN_S)<<16);90
}91
+ agg |= flagsp[0];92
+ agg &= (T1_SIG|T1_VISIT|T1_SIG_OTH)|(T1_SIG|T1_VISIT|T1_SIG_OTH)<<16;93
+ agg = !agg;94
#endif95
} else {96
agg = 0;97
98
memset(t1->data,0,datasize * sizeof(int));99
100
flagssize=(h+2) * (w+2);101
-#ifdef __amd64__102
+#if defined(__amd64__) || defined(__MMX__)103
/* 64 bit SIMD/SWAR in t1_updateflags requires one short of headroom104
because three shorts = 48 bits. */105
++flagssize;106
107
int correction = 3;108
type = ((bpno < (cblk->numbps - 4)) && (passtype < 2) && (cblksty & J2K_CCP_CBLKSTY_LAZY)) ? T1_TYPE_RAW : T1_TYPE_MQ;109
110
+#if !defined(__amd64__) && defined(__MMX__)111
+ _mm_empty();112
+#endif113
switch (passtype) {114
case 0:115
t1_enc_sigpass(t1, bpno, orient, &nmsedec, type, cblksty);116
117
mqc_segmark_enc(mqc);118
break;119
}120
+#if !defined(__amd64__) && defined(__MMX__)121
+ _mm_empty();122
+#endif123
124
/* fixed_quality */125
cumwmsedec += t1_getwmsedec(nmsedec, compno, level, orient, bpno, qmfbid, stepsize, numcomps);126
127
mqc_setstate(mqc, T1_CTXNO_AGG, 0, 3);128
mqc_setstate(mqc, T1_CTXNO_ZC, 0, 4);129
130
+#if !defined(__amd64__) && defined(__MMX__)131
+ _mm_empty();132
+#endif133
for (segno = 0; segno < cblk->numsegs; ++segno) {134
opj_tcd_seg_t *seg = &cblk->segs[segno];135
136
137
}138
}139
}140
+#if !defined(__amd64__) && defined(__MMX__)141
+ _mm_empty();142
+#endif143
}144
145
/* ----------------------------------------------------------------------- */146