File openjpeg-20070821svn-t1-updateflags-x86_64.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 2007-08-23 05:53:17.000000000 -05003
+++ OpenJPEG.patched/libopenjpeg/t1.c 2007-08-23 05:56:33.000000000 -05004
5
static char t1_getspb(int f);6
static short t1_getnmsedec_sig(int x, int bitpos);7
static short t1_getnmsedec_ref(int x, int bitpos);8
+#ifdef __amd64__9
+static INLINE void t1_updateflags(flag_t *flagsp, int s, int stride);10
+#else11
static void t1_updateflags(flag_t *flagsp, int s, int stride);12
+#endif13
/**14
Encode significant pass15
*/16
17
return lut_nmsedec_ref0[x & ((1 << T1_NMSEDEC_BITS) - 1)];18
}19
20
+#ifdef __amd64__21
+22
+/* On 64 bit platforms we can set three flags at a time. (SWAR) */23
+/* FIXME: Assumes little endian? */24
+25
+#define VEC(x,y,z) (int64)(x)|((int64)(y)<<16)|((int64)(z)<<32)26
+27
+static void t1_updateflags(flag_t *flagsp, int s, int stride) {28
+ static const int64 mod[] = {29
+ VEC(T1_SIG_SE, T1_SIG_E, T1_SIG_NE),30
+ VEC(T1_SIG_SE, T1_SIG_E|T1_SGN_E, T1_SIG_NE),31
+ VEC(T1_SIG_S, T1_SIG, T1_SIG_N),32
+ VEC(T1_SIG_S|T1_SGN_S, T1_SIG, T1_SIG_N|T1_SGN_N),33
+ VEC(T1_SIG_SW, T1_SIG_W, T1_SIG_NW),34
+ VEC(T1_SIG_SW, T1_SIG_W|T1_SGN_W, T1_SIG_NW)35
+ };36
+37
+ int64 tmp1 = *(int64*)((void*)&flagsp[-1 - stride]);38
+ int64 tmp2 = *(int64*)((void*)&flagsp[-1 ]);39
+ int64 tmp3 = *(int64*)((void*)&flagsp[-1 + stride]);40
+41
+ tmp1 |= mod[s];42
+ tmp2 |= mod[s+2];43
+ tmp3 |= mod[s+4];44
+45
+ *(int64*)((void*)&flagsp[-1 - stride]) = tmp1;46
+ *(int64*)((void*)&flagsp[-1 ]) = tmp2;47
+ *(int64*)((void*)&flagsp[-1 + stride]) = tmp3;48
+}49
+50
+#else51
+52
static void t1_updateflags(flag_t *flagsp, int s, int stride) {53
static const flag_t mod[] = {54
T1_SIG_E, T1_SIG_E|T1_SGN_E,55
56
flagsp[ 1 + stride] |= T1_SIG_NW;57
}58
59
+#endif60
+61
static void t1_enc_sigpass_step(62
opj_t1_t *t1,63
flag_t *flagsp,64
65
for (i = 0; i < t1->w; ++i) {66
if (k + 3 < t1->h) {67
#ifdef __amd64__68
+ /* 64 bit SWAR */69
+ /* FIXME: Assumes little endian? */70
int64 tmp = *((int64*)&t1->flags[(k+1) + (i+1)*(t1->h+2)]);71
if (cblksty & J2K_CCP_CBLKSTY_VSC) {72
tmp &= ~((int64)(T1_SIG_S | T1_SIG_SE | T1_SIG_SW | T1_SGN_S)<<48);73
74
memset(t1->data,0,datasize * sizeof(int));75
76
flagssize=(h+2) * (w+2);77
+#ifdef __amd64__78
+ /* 64 bit SIMD/SWAR in t1_updateflags requires one short of headroom79
+ because three shorts = 48 bits. */80
+ ++flagssize;81
+#endif82
83
if(flagssize > t1->flagssize){84
opj_aligned_free(t1->flags);85