diff-delta: produce optimal pack data
authorNicolas Pitre <nico@cam.org>
Wed, 22 Feb 2006 01:45:36 +0000 (20:45 -0500)
committerJunio C Hamano <junkio@cox.net>
Thu, 2 Mar 2006 05:51:00 +0000 (21:51 -0800)
Indexing based on adler32 has a match precision based on the block size
(currently 16).  Lowering the block size would produce smaller deltas
but the indexing memory and computing cost increases significantly.

For optimal delta result the indexing block size should be 3 with an
increment of 1 (instead of 16 and 16).  With such low params the adler32
becomes a clear overhead increasing the time for git-repack by a factor
of 3.  And with such small blocks the adler 32 is not very useful as the
whole of the block bits can be used directly.

This patch replaces the adler32 with an open coded index value based on
3 characters directly.  This gives sufficient bits for hashing and
allows for optimal delta with reasonable CPU cycles.

The resulting packs are 6% smaller on average.  The increase in CPU time
is about 25%.  But this cost is now hidden by the delta reuse patch
while the saving on data transfers is always there.

Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
diff-delta.c

index 2ed5984..27f83a0 100644 (file)
 
 #include <stdlib.h>
 #include <string.h>
-#include <zlib.h>
 #include "delta.h"
 
 
-/* block size: min = 16, max = 64k, power of 2 */
-#define BLK_SIZE 16
-
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-
-#define GR_PRIME 0x9e370001
-#define HASH(v, shift) (((unsigned int)(v) * GR_PRIME) >> (shift))
-
 struct index {
        const unsigned char *ptr;
-       unsigned int val;
        struct index *next;
 };
 
@@ -42,21 +32,21 @@ static struct index ** delta_index(const unsigned char *buf,
                                   unsigned long bufsize,
                                   unsigned int *hash_shift)
 {
-       unsigned int hsize, hshift, entries, blksize, i;
+       unsigned long hsize;
+       unsigned int hshift, i;
        const unsigned char *data;
        struct index *entry, **hash;
        void *mem;
 
        /* determine index hash size */
-       entries = (bufsize + BLK_SIZE - 1) / BLK_SIZE;
-       hsize = entries / 4;
-       for (i = 4; (1 << i) < hsize && i < 16; i++);
+       hsize = bufsize / 4;
+       for (i = 8; (1 << i) < hsize && i < 16; i++);
        hsize = 1 << i;
-       hshift = 32 - i;
+       hshift = i - 8;
        *hash_shift = hshift;
 
        /* allocate lookup index */
-       mem = malloc(hsize * sizeof(*hash) + entries * sizeof(*entry));
+       mem = malloc(hsize * sizeof(*hash) + bufsize * sizeof(*entry));
        if (!mem)
                return NULL;
        hash = mem;
@@ -64,17 +54,12 @@ static struct index ** delta_index(const unsigned char *buf,
        memset(hash, 0, hsize * sizeof(*hash));
 
        /* then populate it */
-       data = buf + entries * BLK_SIZE - BLK_SIZE;
-       blksize = bufsize - (data - buf);
-       while (data >= buf) {
-               unsigned int val = adler32(0, data, blksize);
-               i = HASH(val, hshift);
-               entry->ptr = data;
-               entry->val = val;
+       data = buf + bufsize - 2;
+       while (data > buf) {
+               entry->ptr = --data;
+               i = data[0] ^ data[1] ^ (data[2] << hshift);
                entry->next = hash[i];
                hash[i] = entry++;
-               blksize = BLK_SIZE;
-               data -= BLK_SIZE;
        }
 
        return hash;
@@ -141,29 +126,27 @@ void *diff_delta(void *from_buf, unsigned long from_size,
 
        while (data < top) {
                unsigned int moff = 0, msize = 0;
-               unsigned int blksize = MIN(top - data, BLK_SIZE);
-               unsigned int val = adler32(0, data, blksize);
-               i = HASH(val, hash_shift);
-               for (entry = hash[i]; entry; entry = entry->next) {
-                       const unsigned char *ref = entry->ptr;
-                       const unsigned char *src = data;
-                       unsigned int ref_size = ref_top - ref;
-                       if (entry->val != val)
-                               continue;
-                       if (ref_size > top - src)
-                               ref_size = top - src;
-                       while (ref_size && *src++ == *ref) {
-                               ref++;
-                               ref_size--;
-                       }
-                       ref_size = ref - entry->ptr;
-                       if (ref_size > msize) {
-                               /* this is our best match so far */
-                               moff = entry->ptr - ref_data;
-                               msize = ref_size;
-                               if (msize >= 0x10000) {
-                                       msize = 0x10000;
+               if (data + 2 < top) {
+                       i = data[0] ^ data[1] ^ (data[2] << hash_shift);
+                       for (entry = hash[i]; entry; entry = entry->next) {
+                               const unsigned char *ref = entry->ptr;
+                               const unsigned char *src = data;
+                               unsigned int ref_size = ref_top - ref;
+                               if (ref_size > top - src)
+                                       ref_size = top - src;
+                               if (ref_size > 0x10000)
+                                       ref_size = 0x10000;
+                               if (ref_size <= msize)
                                        break;
+                               while (ref_size && *src++ == *ref) {
+                                       ref++;
+                                       ref_size--;
+                               }
+                               ref_size = ref - entry->ptr;
+                               if (msize < ref - entry->ptr) {
+                                       /* this is our best match so far */
+                                       msize = ref - entry->ptr;
+                                       moff = entry->ptr - ref_data;
                                }
                        }
                }