Sync codeclib bitstream code with upstream ffmpeg code. Build ffmpeg_bitstream.c as a part of the codec lib. Use this codeclib implementation in libffmpegFLAC. Implement adapted version of the unaligned longword reading optimization for coldfire from the libwma version of this code. Speeds up cook decoding by 2-3% on h300 and flac by 25% on h300, also speeds up flac decoding by 2% on c200 (decoding speed of cook on c200 is unchanged).

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@27430 a1c6a512-1295-4272-9138-f99709370657

1 files changed, 205 insertions, 116 deletions

diff --git a/apps/codecs/lib/ffmpeg_bitstream.c b/apps/codecs/lib/ffmpeg_bitstream.c
index 88e3cbfe3a..c879661c47 100644
--- a/apps/codecs/lib/ffmpeg_bitstream.c
+++ b/apps/codecs/lib/ffmpeg_bitstream.c
@@ -2,6 +2,7 @@
  * Common bit i/o utils
  * Copyright (c) 2000, 2001 Fabrice Bellard
  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
+ * Copyright (c) 2010 Loren Merritt
  *
  * alternative bitstream reader & writer by Michael Niedermayer <michaelni@gmx.at>
  *
@@ -22,7 +23,16 @@
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
-#include "ffmpeg_bitstream.h"
+/**
+ * @file
+ * bitstream api.
+ */
+
+//#include "avcodec.h"
+#include "ffmpeg_get_bits.h"
+//#include "put_bits.h"
+
+#define av_log(...)
 
 #ifdef ROCKBOX
 #undef DEBUGF
@@ -36,27 +46,47 @@ const uint8_t ff_log2_run[32]={
  8, 9,10,11,12,13,14,15
 };
 
-/**
- * Same as av_mallocz_static(), but does a realloc.
- *
- * @param[in] ptr The block of memory to reallocate.
- * @param[in] size The requested size.
- * @return Block of memory of requested size.
- * @deprecated. Code which uses ff_realloc_static is broken/misdesigned
- * and should correctly use static arrays
- */
-
+#if 0 // unused in rockbox
+void align_put_bits(PutBitContext *s)
+{
+#ifdef ALT_BITSTREAM_WRITER
+    put_bits(s,(  - s->index) & 7,0);
+#else
+    put_bits(s,s->bit_left & 7,0);
+#endif
+}
 
-void ff_put_string(PutBitContext * pbc, const char *s, int put_zero)
+void ff_put_string(PutBitContext *pb, const char *string, int terminate_string)
 {
-    while(*s){
-        put_bits(pbc, 8, *s);
-        s++;
+    while(*string){
+        put_bits(pb, 8, *string);
+        string++;
     }
-    if(put_zero)
-        put_bits(pbc, 8, 0);
+    if(terminate_string)
+        put_bits(pb, 8, 0);
 }
 
+void ff_copy_bits(PutBitContext *pb, const uint8_t *src, int length)
+{
+    int words= length>>4;
+    int bits= length&15;
+    int i;
+
+    if(length==0) return;
+
+    if(CONFIG_SMALL || words < 16 || put_bits_count(pb)&7){
+        for(i=0; i<words; i++) put_bits(pb, 16, AV_RB16(src + 2*i));
+    }else{
+        for(i=0; put_bits_count(pb)&31; i++)
+            put_bits(pb, 8, src[i]);
+        flush_put_bits(pb);
+        memcpy(put_bits_ptr(pb), src+i, 2*words-i);
+        skip_put_bytes(pb, 2*words-i);
+    }
+
+    put_bits(pb, bits, AV_RB16(src + 2*words)>>(16-bits));
+}
+#endif
 /* VLC decoding */
 
 //#define DEBUG_VLC
@@ -84,118 +114,143 @@ static int alloc_table(VLC *vlc, int size, int use_static)
     index = vlc->table_size;
     vlc->table_size += size;
     if (vlc->table_size > vlc->table_allocated) {
-        if(use_static>1){
+        if(use_static)
+        {
             DEBUGF("init_vlc() used with too little memory : table_size > allocated_memory\n");
+            return -1;
         }
-
+//            abort(); //cant do anything, init_vlc() is used with too little memory
+//        vlc->table_allocated += (1 << vlc->bits);
+//        vlc->table = av_realloc(vlc->table,
+//                                sizeof(VLC_TYPE) * 2 * vlc->table_allocated);
         if (!vlc->table)
             return -1;
     }
     return index;
 }
 
-static int build_table(VLC *vlc, int table_nb_bits,
-                       int nb_codes,
-                       const void *bits, int bits_wrap, int bits_size,
-                       const void *codes, int codes_wrap, int codes_size,
-                       const void *symbols, int symbols_wrap, int symbols_size,
-                       uint32_t code_prefix, int n_prefix, int flags)
-{
-    int i, j, k, n, table_size, table_index, nb, n1, index, code_prefix2, symbol;
+/* 
+static av_always_inline uint32_t bitswap_32(uint32_t x) {
+    return av_reverse[x&0xFF]<<24
+         | av_reverse[(x>>8)&0xFF]<<16
+         | av_reverse[(x>>16)&0xFF]<<8
+         | av_reverse[x>>24];
+}
+*/
+
+typedef struct {
+    uint8_t bits;
+    uint16_t symbol;
+    /** codeword, with the first bit-to-be-read in the msb
+     * (even if intended for a little-endian bitstream reader) */
     uint32_t code;
+} VLCcode;
+
+static int compare_vlcspec(const void *a, const void *b)
+{
+    const VLCcode *sa=a, *sb=b;
+    return (sa->code >> 1) - (sb->code >> 1);
+}
+
+/**
+ * Build VLC decoding tables suitable for use with get_vlc().
+ *
+ * @param vlc            the context to be initted
+ *
+ * @param table_nb_bits  max length of vlc codes to store directly in this table
+ *                       (Longer codes are delegated to subtables.)
+ *
+ * @param nb_codes       number of elements in codes[]
+ *
+ * @param codes          descriptions of the vlc codes
+ *                       These must be ordered such that codes going into the same subtable are contiguous.
+ *                       Sorting by VLCcode.code is sufficient, though not necessary.
+ */
+static int build_table(VLC *vlc, int table_nb_bits, int nb_codes,
+                       VLCcode *codes, int flags)
+{
+    int table_size, table_index, index, symbol, subtable_bits;
+    int i, j, k, n, nb, inc;
+    uint32_t code, code_prefix;
     VLC_TYPE (*table)[2];
 
     table_size = 1 << table_nb_bits;
-    table_index = alloc_table(vlc, table_size, flags & (INIT_VLC_USE_STATIC|INIT_VLC_USE_NEW_STATIC));
+    table_index = alloc_table(vlc, table_size, flags & INIT_VLC_USE_NEW_STATIC);
 #ifdef DEBUG_VLC
-    DEBUGF("new table index=%d size=%d code_prefix=%x n=%d\n",
-           table_index, table_size, code_prefix, n_prefix);
+    av_log(NULL,AV_LOG_DEBUG,"new table index=%d size=%d\n",
+           table_index, table_size);
 #endif
     if (table_index < 0)
         return -1;
     table = &vlc->table[table_index];
 
-    for(i=0;i<table_size;i++) {
+    for (i = 0; i < table_size; i++) {
         table[i][1] = 0; //bits
         table[i][0] = -1; //codes
     }
 
     /* first pass: map codes and compute auxillary table sizes */
-    for(i=0;i<nb_codes;i++) {
-        GET_DATA(n, bits, i, bits_wrap, bits_size);
-        GET_DATA(code, codes, i, codes_wrap, codes_size);
-        /* we accept tables with holes */
-        if (n <= 0)
-            continue;
-        if (!symbols)
-            symbol = i;
-        else
-            GET_DATA(symbol, symbols, i, symbols_wrap, symbols_size);
+    for (i = 0; i < nb_codes; i++) {
+        n = codes[i].bits;
+        code = codes[i].code;
+        symbol = codes[i].symbol;
 #if defined(DEBUG_VLC) && 0
-        DEBUGF("i=%d n=%d code=0x%x\n", i, n, code);
+        av_log(NULL,AV_LOG_DEBUG,"i=%d n=%d code=0x%x\n", i, n, code);
 #endif
-        /* if code matches the prefix, it is in the table */
-        n -= n_prefix;
-        if(flags & INIT_VLC_LE)
-            code_prefix2= code & (n_prefix>=32 ? (int)0xffffffff : (1 << n_prefix)-1);
-        else
-            code_prefix2= code >> n;
-        if (n > 0 && code_prefix2 == (int)code_prefix) {
-            if (n <= table_nb_bits) {
-                /* no need to add another table */
-                j = (code << (table_nb_bits - n)) & (table_size - 1);
-                nb = 1 << (table_nb_bits - n);
-                for(k=0;k<nb;k++) {
-                    if(flags & INIT_VLC_LE)
-                        j = (code >> n_prefix) + (k<<n);
+        if (n <= table_nb_bits) {
+            /* no need to add another table */
+            j = code >> (32 - table_nb_bits);
+            nb = 1 << (table_nb_bits - n);
+            inc = 1;
+/*            if (flags & INIT_VLC_LE) {
+                j = bitswap_32(code);
+                inc = 1 << n;
+            } */
+            for (k = 0; k < nb; k++) {
 #ifdef DEBUG_VLC
-                    DEBUGF("%4x: code=%d n=%d\n",
-                           j, i, n);
+                av_log(NULL, AV_LOG_DEBUG, "%4x: code=%d n=%d\n",
+                       j, i, n);
 #endif
-                    if (table[j][1] /*bits*/ != 0) {
-                        DEBUGF("incorrect codes\n");
-                        return -1;
-                    }
-                    table[j][1] = n; //bits
-                    table[j][0] = symbol;
-                    j++;
+                if (table[j][1] /*bits*/ != 0) {
+                    av_log(NULL, AV_LOG_ERROR, "incorrect codes\n");
+                    return -1;
                 }
-            } else {
-                n -= table_nb_bits;
-                j = (code >> ((flags & INIT_VLC_LE) ? n_prefix : n)) & ((1 << table_nb_bits) - 1);
-#ifdef DEBUG_VLC
-                DEBUGF("%4x: n=%d (subtable)\n",
-                       j, n);
-#endif
-                /* compute table size */
-                n1 = -table[j][1]; //bits
-                if (n > n1)
-                    n1 = n;
-                table[j][1] = -n1; //bits
+                table[j][1] = n; //bits
+                table[j][0] = symbol;
+                j += inc;
             }
-        }
-    }
-
-    /* second pass : fill auxillary tables recursively */
-    for(i=0;i<table_size;i++) {
-        n = table[i][1]; //bits
-        if (n < 0) {
-            n = -n;
-            if (n > table_nb_bits) {
-                n = table_nb_bits;
-                table[i][1] = -n; //bits
+        } else {
+            /* fill auxiliary table recursively */
+            n -= table_nb_bits;
+            code_prefix = code >> (32 - table_nb_bits);
+            subtable_bits = n;
+            codes[i].bits = n;
+            codes[i].code = code << table_nb_bits;
+            for (k = i+1; k < nb_codes; k++) {
+                n = codes[k].bits - table_nb_bits;
+                if (n <= 0)
+                    break;
+                code = codes[k].code;
+                if (code >> (32 - table_nb_bits) != code_prefix)
+                    break;
+                codes[k].bits = n;
+                codes[k].code = code << table_nb_bits;
+                subtable_bits = FFMAX(subtable_bits, n);
             }
-            index = build_table(vlc, n, nb_codes,
-                                bits, bits_wrap, bits_size,
-                                codes, codes_wrap, codes_size,
-                                symbols, symbols_wrap, symbols_size,
-                                (flags & INIT_VLC_LE) ? (code_prefix | (i << n_prefix)) : ((code_prefix << table_nb_bits) | i),
-                                n_prefix + table_nb_bits, flags);
+            subtable_bits = FFMIN(subtable_bits, table_nb_bits);
+            j = /*(flags & INIT_VLC_LE) ? bitswap_32(code_prefix) >> (32 - table_nb_bits) :*/ code_prefix;
+            table[j][1] = -subtable_bits;
+#ifdef DEBUG_VLC
+            av_log(NULL,AV_LOG_DEBUG,"%4x: n=%d (subtable)\n",
+                   j, codes[i].bits + table_nb_bits);
+#endif
+            index = build_table(vlc, subtable_bits, k-i, codes+i, flags);
             if (index < 0)
                 return -1;
             /* note: realloc has been done, so reload tables */
             table = &vlc->table[table_index];
-            table[i][0] = index; //code
+            table[j][0] = index; //code
+            i = k-1;
         }
     }
     return table_index;
@@ -228,49 +283,83 @@ static int build_table(VLC *vlc, int table_nb_bits,
    'use_static' should be set to 1 for tables, which should be freed
    with av_free_static(), 0 if free_vlc() will be used.
 */
+
+/* Rockbox: support for INIT_VLC_LE is currently disabled since none of our
+   codecs use it, there's a LUT based bit reverse function for this commented
+   out above (bitswap_32) and an inline asm version in libtremor/codebook.c
+   if we ever want this */
 int init_vlc_sparse(VLC *vlc, int nb_bits, int nb_codes,
              const void *bits, int bits_wrap, int bits_size,
              const void *codes, int codes_wrap, int codes_size,
              const void *symbols, int symbols_wrap, int symbols_size,
              int flags)
 {
+    VLCcode buf[nb_codes+1]; /* worst case from cook seems to be nb_codes == 607
+                                which would make this about 4.8k... */
+    int i, j, ret;
+
     vlc->bits = nb_bits;
     if(flags & INIT_VLC_USE_NEW_STATIC){
         if(vlc->table_size && vlc->table_size == vlc->table_allocated){
             return 0;
         }else if(vlc->table_size){
-            return -1; // fatal error, we are called on a partially initialized table
+            DEBUGF("fatal error, we are called on a partially initialized table\n");
+            return -1;
+//            abort(); // fatal error, we are called on a partially initialized table
         }
-    }else if(!(flags & INIT_VLC_USE_STATIC)) {
+    }else {
         vlc->table = NULL;
         vlc->table_allocated = 0;
         vlc->table_size = 0;
-    } else {
-        /* Static tables are initially always NULL, return
-           if vlc->table != NULL to avoid double allocation */
-        if(vlc->table)
-            return 0;
     }
 
 #ifdef DEBUG_VLC
-    DEBUGF("build table nb_codes=%d\n", nb_codes);
+    av_log(NULL,AV_LOG_DEBUG,"build table nb_codes=%d\n", nb_codes);
 #endif
 
-    if (build_table(vlc, nb_bits, nb_codes,
-                    bits, bits_wrap, bits_size,
-                    codes, codes_wrap, codes_size,
-                    symbols, symbols_wrap, symbols_size,
-                    0, 0, flags) < 0) {
-        //free(&vlc->table);
+//    buf = av_malloc((nb_codes+1)*sizeof(VLCcode));
+
+//    assert(symbols_size <= 2 || !symbols);
+    j = 0;
+#define COPY(condition)\
+    for (i = 0; i < nb_codes; i++) {\
+        GET_DATA(buf[j].bits, bits, i, bits_wrap, bits_size);\
+        if (!(condition))\
+            continue;\
+        GET_DATA(buf[j].code, codes, i, codes_wrap, codes_size);\
+/*        if (flags & INIT_VLC_LE)*/\
+/*            buf[j].code = bitswap_32(buf[j].code);*/\
+/*        else*/\
+            buf[j].code <<= 32 - buf[j].bits;\
+        if (symbols)\
+            GET_DATA(buf[j].symbol, symbols, i, symbols_wrap, symbols_size)\
+        else\
+            buf[j].symbol = i;\
+        j++;\
+    }
+    COPY(buf[j].bits > nb_bits);
+    // qsort is the slowest part of init_vlc, and could probably be improved or avoided
+    qsort(buf, j, sizeof(VLCcode), compare_vlcspec);
+    COPY(buf[j].bits && buf[j].bits <= nb_bits);
+    nb_codes = j;
+
+    ret = build_table(vlc, nb_bits, nb_codes, buf, flags);
+
+//    av_free(buf);
+    if (ret < 0) {
+//        av_freep(&vlc->table);
         return -1;
     }
-    /* Changed the following condition to be true if table_size > table_allocated. *
-     * This would be more sensible for static tables since we want warnings for    *
-     * memory shortages only.                                                      */
-#ifdef TEST
-    if((flags & INIT_VLC_USE_NEW_STATIC) && vlc->table_size > vlc->table_allocated)
-        DEBUGF("needed %d had %d\n", vlc->table_size, vlc->table_allocated);
-#endif
+    if((flags & INIT_VLC_USE_NEW_STATIC) && vlc->table_size != vlc->table_allocated) {
+        av_log(NULL, AV_LOG_ERROR, "needed %d had %d\n", vlc->table_size, vlc->table_allocated);
+    }
     return 0;
 }
 
+/* not used in rockbox
+void free_vlc(VLC *vlc)
+{
+    av_freep(&vlc->table);
+}
+*/
+