Browse code

make state transition tables global as they are constant and the code is slightly faster that way

Originally committed as revision 6655 to svn://svn.ffmpeg.org/ffmpeg/trunk

Michael Niedermayer authored on 2006/10/11 23:44:17
Showing 3 changed files
... ...
@@ -51,8 +51,10 @@ static const uint8_t lps_range[64][4]= {
51 51
 };
52 52
 
53 53
 uint8_t ff_h264_lps_range[2*65][4];
54
+uint8_t ff_h264_lps_state[2*64];
55
+uint8_t ff_h264_mps_state[2*64];
54 56
 
55
-const uint8_t ff_h264_mps_state[64]= {
57
+static const uint8_t mps_state[64]= {
56 58
   1, 2, 3, 4, 5, 6, 7, 8,
57 59
   9,10,11,12,13,14,15,16,
58 60
  17,18,19,20,21,22,23,24,
... ...
@@ -63,7 +65,7 @@ const uint8_t ff_h264_mps_state[64]= {
63 63
  57,58,59,60,61,62,62,63,
64 64
 };
65 65
 
66
-const uint8_t ff_h264_lps_state[64]= {
66
+static const uint8_t lps_state[64]= {
67 67
   0, 0, 1, 2, 2, 4, 4, 5,
68 68
   6, 7, 8, 9, 9,11,11,12,
69 69
  13,13,15,15,16,16,18,18,
... ...
@@ -121,32 +123,31 @@ void ff_init_cabac_decoder(CABACContext *c, const uint8_t *buf, int buf_size){
121 121
     c->range= 0x1FE<<(CABAC_BITS + 1);
122 122
 }
123 123
 
124
-void ff_init_cabac_states(CABACContext *c,
125
-                          uint8_t const *mps_state, uint8_t const *lps_state, int state_count){
124
+void ff_init_cabac_states(CABACContext *c){
126 125
     int i, j;
127 126
 
128
-    for(i=0; i<state_count; i++){
127
+    for(i=0; i<64; i++){
129 128
         for(j=0; j<4; j++){ //FIXME check if this is worth the 1 shift we save
130 129
             ff_h264_lps_range[2*i+0][j+4]=
131 130
             ff_h264_lps_range[2*i+1][j+4]= lps_range[i][j];
132 131
         }
133 132
 
134
-        c->mps_state[2*i+0]= 2*mps_state[i]+0;
135
-        c->mps_state[2*i+1]= 2*mps_state[i]+1;
133
+        ff_h264_mps_state[2*i+0]= 2*mps_state[i]+0;
134
+        ff_h264_mps_state[2*i+1]= 2*mps_state[i]+1;
136 135
 
137 136
         if( i ){
138 137
 #ifdef BRANCHLESS_CABAC_DECODER
139
-            c->mps_state[-2*i-1]= 2*lps_state[i]+0; //FIXME yes this is not valid C but iam lazy, cleanup welcome
140
-            c->mps_state[-2*i-2]= 2*lps_state[i]+1;
138
+            ff_h264_mps_state[-2*i-1]= 2*lps_state[i]+0; //FIXME yes this is not valid C but iam lazy, cleanup welcome
139
+            ff_h264_mps_state[-2*i-2]= 2*lps_state[i]+1;
141 140
         }else{
142
-            c->mps_state[-2*i-1]= 1;
143
-            c->mps_state[-2*i-2]= 0;
141
+            ff_h264_mps_state[-2*i-1]= 1;
142
+            ff_h264_mps_state[-2*i-2]= 0;
144 143
 #else
145
-            c->lps_state[2*i+0]= 2*lps_state[i]+0;
146
-            c->lps_state[2*i+1]= 2*lps_state[i]+1;
144
+            ff_h264_lps_state[2*i+0]= 2*lps_state[i]+0;
145
+            ff_h264_lps_state[2*i+1]= 2*lps_state[i]+1;
147 146
         }else{
148
-            c->lps_state[2*i+0]= 1;
149
-            c->lps_state[2*i+1]= 0;
147
+            ff_h264_lps_state[2*i+0]= 1;
148
+            ff_h264_lps_state[2*i+1]= 0;
150 149
 #endif
151 150
         }
152 151
     }
... ...
@@ -41,8 +41,6 @@ typedef struct CABACContext{
41 41
 #ifdef STRICT_LIMITS
42 42
     int symCount;
43 43
 #endif
44
-    uint8_t lps_state[2*64];      ///< transIdxLPS
45
-    uint8_t mps_state[2*64];      ///< transIdxMPS
46 44
     const uint8_t *bytestream_start;
47 45
     const uint8_t *bytestream;
48 46
     const uint8_t *bytestream_end;
... ...
@@ -50,15 +48,14 @@ typedef struct CABACContext{
50 50
 }CABACContext;
51 51
 
52 52
 extern uint8_t ff_h264_lps_range[2*65][4];  ///< rangeTabLPS
53
-extern const uint8_t ff_h264_mps_state[64];
54
-extern const uint8_t ff_h264_lps_state[64];
53
+extern uint8_t ff_h264_mps_state[2*64];     ///< transIdxMPS
54
+extern uint8_t ff_h264_lps_state[2*64];     ///< transIdxLPS
55 55
 extern const uint8_t ff_h264_norm_shift[128];
56 56
 
57 57
 
58 58
 void ff_init_cabac_encoder(CABACContext *c, uint8_t *buf, int buf_size);
59 59
 void ff_init_cabac_decoder(CABACContext *c, const uint8_t *buf, int buf_size);
60
-void ff_init_cabac_states(CABACContext *c,
61
-                          uint8_t const *mps_state, uint8_t const *lps_state, int state_count);
60
+void ff_init_cabac_states(CABACContext *c);
62 61
 
63 62
 
64 63
 static inline void put_cabac_bit(CABACContext *c, int b){
... ...
@@ -91,11 +88,11 @@ static void put_cabac(CABACContext *c, uint8_t * const state, int bit){
91 91
 
92 92
     if(bit == ((*state)&1)){
93 93
         c->range -= RangeLPS;
94
-        *state= c->mps_state[*state];
94
+        *state= ff_h264_mps_state[*state];
95 95
     }else{
96 96
         c->low += c->range - RangeLPS;
97 97
         c->range = RangeLPS;
98
-        *state= c->lps_state[*state];
98
+        *state= ff_h264_lps_state[*state];
99 99
     }
100 100
 
101 101
     renorm_cabac_encoder(c);
... ...
@@ -369,11 +366,9 @@ static int get_cabac(CABACContext *c, uint8_t * const state){
369 369
 
370 370
 #define LOW          "0"
371 371
 #define RANGE        "4"
372
-#define LPS_STATE   "12"
373
-#define MPS_STATE   "12+2*64"
374
-#define BYTESTART   "12+4*64"
375
-#define BYTE        "16+4*64"
376
-#define BYTEEND     "20+4*64"
372
+#define BYTESTART   "12"
373
+#define BYTE        "16"
374
+#define BYTEEND     "20"
377 375
 #ifndef BRANCHLESS_CABAC_DECODER
378 376
     asm volatile(
379 377
         "movzbl (%1), %%eax                     \n\t"
... ...
@@ -391,7 +386,7 @@ static int get_cabac(CABACContext *c, uint8_t * const state){
391 391
         "setb %%cl                              \n\t"
392 392
         "shl %%cl, %%edx                        \n\t"
393 393
         "shl %%cl, %%ebx                        \n\t"
394
-        "movzbl "MPS_STATE"(%2, %%eax), %%ecx   \n\t"
394
+        "movzbl "MANGLE(ff_h264_mps_state)"(%%eax), %%ecx   \n\t"
395 395
         "movb %%cl, (%1)                        \n\t"
396 396
 //eax:state ebx:low, edx:range, esi:RangeLPS
397 397
         "test %%bx, %%bx                        \n\t"
... ...
@@ -413,7 +408,7 @@ static int get_cabac(CABACContext *c, uint8_t * const state){
413 413
         "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx   \n\t"
414 414
         "shll %%cl, %%ebx                       \n\t"
415 415
         "shll %%cl, %%edx                       \n\t"
416
-        "movzbl "LPS_STATE"(%2, %%eax), %%ecx   \n\t"
416
+        "movzbl "MANGLE(ff_h264_lps_state)"(%%eax), %%ecx   \n\t"
417 417
         "movb %%cl, (%1)                        \n\t"
418 418
         "addl $1, %%eax                         \n\t"
419 419
         "test %%bx, %%bx                        \n\t"
... ...
@@ -475,7 +470,7 @@ static int get_cabac(CABACContext *c, uint8_t * const state){
475 475
 #endif
476 476
 
477 477
 //eax:state ebx:low edx:mask esi:range
478
-        "movzbl "MPS_STATE"(%2, %%eax), %%ecx   \n\t"
478
+        "movzbl "MANGLE(ff_h264_mps_state)"(%%eax), %%ecx   \n\t"
479 479
         "movb %%cl, (%1)                        \n\t"
480 480
 
481 481
         "movl %%esi, %%edx                      \n\t"
... ...
@@ -523,12 +518,12 @@ static int get_cabac(CABACContext *c, uint8_t * const state){
523 523
 #ifndef BRANCHLESS_CABAC_DECODER
524 524
     if(c->low < c->range){
525 525
         bit= s&1;
526
-        *state= c->mps_state[s];
526
+        *state= ff_h264_mps_state[s];
527 527
         renorm_cabac_decoder_once(c);
528 528
     }else{
529 529
         bit= ff_h264_norm_shift[RangeLPS>>19];
530 530
         c->low -= c->range;
531
-        *state= c->lps_state[s];
531
+        *state= ff_h264_lps_state[s];
532 532
         c->range = RangeLPS<<bit;
533 533
         c->low <<= bit;
534 534
         bit= (s&1)^1;
... ...
@@ -544,7 +539,7 @@ static int get_cabac(CABACContext *c, uint8_t * const state){
544 544
     c->range += (RangeLPS - c->range) & lps_mask;
545 545
 
546 546
     s^=lps_mask;
547
-    *state= c->mps_state[s];
547
+    *state= ff_h264_mps_state[s];
548 548
     bit= s&1;
549 549
 
550 550
     lps_mask= ff_h264_norm_shift[c->range>>(CABAC_BITS+3)];
... ...
@@ -7388,7 +7388,7 @@ static int decode_slice(H264Context *h){
7388 7388
         align_get_bits( &s->gb );
7389 7389
 
7390 7390
         /* init cabac */
7391
-        ff_init_cabac_states( &h->cabac, ff_h264_mps_state, ff_h264_lps_state, 64 );
7391
+        ff_init_cabac_states( &h->cabac);
7392 7392
         ff_init_cabac_decoder( &h->cabac,
7393 7393
                                s->gb.buffer + get_bits_count(&s->gb)/8,
7394 7394
                                ( s->gb.size_in_bits - get_bits_count(&s->gb) + 7)/8);