[Ffmpeg-devel] [RFC] Fraps v2 support
Kostya
kostya.shishkov
Sat Nov 4 07:04:12 CET 2006
On Fri, Nov 03, 2006 at 09:02:49PM +0100, Michael Niedermayer wrote:
> Hi
>
> On Fri, Nov 03, 2006 at 03:00:51PM +0200, Kostya wrote:
> > On Thu, Nov 02, 2006 at 07:52:23PM +0100, Michael Niedermayer wrote:
> > > Hi
> > >
> > > On Thu, Nov 02, 2006 at 08:24:46PM +0200, Kostya wrote:
> > > > Please test if patch attached confirms to $subj. I have only a bit of one
> > > > sample here.
> > >
> > > [...]
> > >
> > > > + cur_node = 510; // for futher use in huffman decoding
> > > > + /* we have built Huffman table and are ready to decode plane */
> > > > + /* XXX this bitreader ignores every 32th bit so lavc bitreader cannot be used*/
> > >
> > > hmm i can belive it ... are you sure that they ignore every 32th bit??
> > > if its really true then id suggest that you use a temporary buffer and
> > > convert that 31/32bit block stream to a normal one, i guess that this
> > > with the get_vlc() code from lavc is still faster ...
> >
> > Sorry, I misunderstood it. New patch use default bitreader on bswapped() buffer
> > (and ALT_BITSTREAM_READER_LE returns bits in reverse order)
> > As a side effect decoder spends ~40% time less on decoding.
>
> [...]
> > #define FPS_TAG MKTAG('F', 'P', 'S', 'x')
> >
> > +/* symbol for Huffman tree node */
> > +#define HNODE -1
> > +
> > /**
> > + * Huffman node
> > + * FIXME one day this should belong to one general framework
> > + */
> > +typedef struct Node{
> > + int sym;
> > + int count;
> > + int n0;
> > + int n1;
> > +}Node;
>
> sym,n0,n1 dont need an int
Changed to int16_t
>
> [...]
> > +/**
> > + * Comparator - our nodes should ascend by count
> > + * but with preserved symbol order
> > + */
> > +static int huff_cmp(const Node *a, const Node *b){
> > + if(a->count == b->count){
> > + return (a->sym > b->sym) ? 1 : -1;
> > + }
> > + return (a->count > b->count) ? 1 : -1;
>
> return (a->count - b->count)*256 + a->sym - b->sym;
Changed
>
> [...]
> > + for(i = 0; i < 256; i++){
> > + s->nodes[i].sym = i;
> > + s->nodes[i].count = LE_32(src);
> > + s->nodes[i].n0 = s->nodes[i].n1 = -1;
> > + src += 4;
> > + }
> > + size -= 1024;
> > + qsort(s->nodes, 256, sizeof(Node), huff_cmp);
> > + cur_node = 256;
> > + // FIXME how it will handle nodes with zero count?
> > + for(i = 0; i < 511; i += 2){
> > + s->nodes[cur_node].sym = HNODE;
> > + s->nodes[cur_node].count = s->nodes[i].count + s->nodes[i+1].count;
> > + s->nodes[cur_node].n0 = i;
> > + s->nodes[cur_node].n1 = i + 1;
>
> it doesnt seem that n1 is ever changed, it will always be n0+1 so its
> redundant
I also thought so, changed
> and n0,sym will either be -1,0..255 or 0..510,-1 so they could be merged
> into one variable
They could but that would hurt readability.
> > + for(j = cur_node; j > 0; j--){
> > + if(s->nodes[j].count >= s->nodes[j - 1].count) break;
> > + tmp = s->nodes[j];
> > + s->nodes[j] = s->nodes[j - 1];
> > + s->nodes[j - 1] = tmp;
>
> (FF)SWAP()
Changed
>
>
> > + }
>
> there is a possible problem here, and that is that the sum of the 2 counts
> can overflow, if it does then this loop will reorder too many elements
> and by doing that trash the n0/n1 pointers which then a few lines later
> could lead to a stack overflow due the recursive vlc building being
> run on a tree which isnt one as it contains a loop from the corruption
Added an explicit check for that
>
> > + cur_node++;
>
> i is always 2*cur_node-512 so maybe one of them could be set based on the
> other, like
> int i= 2*cur_node-512;
> at the top of the loop maybe thats more readable, iam not sure though ...
Also left for readability
>
>
> [...]
> > + s->tmpbuf = av_realloc(s->tmpbuf, offs[1] - offs[0] - 1024);
> > + if(fraps2_decode_plane(s, f->data[0], f->linesize[0], avctx->width,
> > + avctx->height, buf + offs[0], offs[1] - offs[0], 0) < 0) {
> > + av_log(avctx, AV_LOG_ERROR, "Error decoding plane 0\n");
> > + return -1;
> > + }
> > + s->tmpbuf = av_realloc(s->tmpbuf, offs[2] - offs[1] - 1024);
> > + if(fraps2_decode_plane(s, f->data[1], f->linesize[1], avctx->width/2,
> > + avctx->height/2, buf + offs[1], offs[2] - offs[1], 1) < 0) {
> > + av_log(avctx, AV_LOG_ERROR, "Error decoding plane 1\n");
> > + return -1;
> > + }
> > + s->tmpbuf = av_realloc(s->tmpbuf, buf_size - offs[2] - 1024);
> > + if(fraps2_decode_plane(s, f->data[2], f->linesize[1], avctx->width/2,
> > + avctx->height/2, buf + offs[2], buf_size - offs[2], 1) < 0) {
> > + av_log(avctx, AV_LOG_ERROR, "Error decoding plane 2\n");
> > + return -1;
> > + }
>
> code triplification, imho make a loop out of the 3
Done with additional support for grayscale Fraps (v4, it unpacks 1 plane instead of 3)
>
> [...]
>
> --
> Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
>
> In the past you could go to a library and read, borrow or copy any book
> Today you'd get arrested for mere telling someone where the library is
>
-------------- next part --------------
Index: libavcodec/fraps.c
===================================================================
--- libavcodec/fraps.c (revision 6830)
+++ libavcodec/fraps.c (working copy)
@@ -1,6 +1,7 @@
/*
* Fraps FPS1 decoder
* Copyright (c) 2005 Roine Gustafsson
+ * Copyright (c) 2006 Konstantin Shishkov
*
* This file is part of FFmpeg.
*
@@ -24,24 +25,42 @@
* @file fraps.c
* Lossless Fraps 'FPS1' decoder
* @author Roine Gustafsson <roine at users sf net>
+ * @author Konstantin Shishkov
*
* Only decodes version 0 and 1 files.
* Codec algorithm for version 0 is taken from Transcode <www.transcoding.org>
*
- * Version 2 files, which are the most commonly found Fraps files, cannot be
- * decoded yet.
+ * Version 2 files support by Konstantin Shishkov
*/
#include "avcodec.h"
+#include "bitstream.h"
+#include "dsputil.h"
#define FPS_TAG MKTAG('F', 'P', 'S', 'x')
+/* symbol for Huffman tree node */
+#define HNODE -1
+
/**
+ * Huffman node
+ * FIXME one day this should belong to one general framework
+ */
+typedef struct Node{
+ int16_t sym;
+ int16_t n0;
+ int count;
+}Node;
+
+/**
* local variable storage
*/
typedef struct FrapsContext{
AVCodecContext *avctx;
AVFrame frame;
+ Node nodes[512];
+ uint8_t *tmpbuf;
+ DSPContext dsp;
} FrapsContext;
@@ -60,12 +79,112 @@
s->avctx = avctx;
s->frame.data[0] = NULL;
+ s->tmpbuf = NULL;
+ dsputil_init(&s->dsp, avctx);
+
return 0;
}
+/**
+ * Comparator - our nodes should ascend by count
+ * but with preserved symbol order
+ */
+static int huff_cmp(const Node *a, const Node *b){
+ return (a->count - b->count)*256 + a->sym - b->sym;
+}
+static void get_tree_codes(uint32_t *bits, int16_t *lens, Node *nodes, int node, uint32_t pfx, int pl)
+{
+ int s;
+
+ s = nodes[node].sym;
+ if(s != HNODE){
+ bits[s] = pfx;
+ lens[s] = pl;
+ }else{
+ pfx <<= 1;
+ pl++;
+ get_tree_codes(bits, lens, nodes, nodes[node].n0, pfx, pl);
+ pfx |= 1;
+ get_tree_codes(bits, lens, nodes, nodes[node].n0+1, pfx, pl);
+ }
+}
+
+static int build_huff_tree(VLC *vlc, Node *nodes)
+{
+ uint32_t bits[256];
+ int16_t lens[256];
+
+ get_tree_codes(bits, lens, nodes, 510, 0, 0);
+ return init_vlc(vlc, 9, 256, lens, 2, 2, bits, 4, 4, 0);
+}
+
+
/**
+ * decode Fraps v2 packed plane
+ */
+static int fraps2_decode_plane(FrapsContext *s, uint8_t *dst, int stride, int w,
+ int h, uint8_t *src, int size, int Uoff)
+{
+ int i, j;
+ int cur_node;
+ GetBitContext gb;
+ VLC vlc;
+ int64_t sum = 0;
+
+ for(i = 0; i < 256; i++){
+ s->nodes[i].sym = i;
+ s->nodes[i].count = LE_32(src);
+ s->nodes[i].n0 = -2;
+ src += 4;
+ sum += s->nodes[i].count;
+ }
+ size -= 1024;
+
+ if(sum >> 31) {
+ av_log(s->avctx, AV_LOG_ERROR, "Too high symbol frequencies. Tree construction is not possible\n");
+ return -1;
+ }
+ qsort(s->nodes, 256, sizeof(Node), huff_cmp);
+ cur_node = 256;
+ // FIXME how it will handle nodes with zero count?
+ for(i = 0; i < 511; i += 2){
+ s->nodes[cur_node].sym = HNODE;
+ s->nodes[cur_node].count = s->nodes[i].count + s->nodes[i+1].count;
+ s->nodes[cur_node].n0 = i;
+ for(j = cur_node; j > 0; j--){
+ if(s->nodes[j].count >= s->nodes[j - 1].count) break;
+ FFSWAP(Node, s->nodes[j], s->nodes[j - 1]);
+ }
+ cur_node++;
+ }
+ if(build_huff_tree(&vlc, s->nodes) < 0){
+ av_log(s->avctx, AV_LOG_ERROR, "Error building tree\n");
+ return -1;
+ }
+ /* we have built Huffman table and are ready to decode plane */
+
+ /* convert bits so they may be used by standard bitreader */
+ s->dsp.bswap_buf(s->tmpbuf, src, size >> 2);
+
+ init_get_bits(&gb, s->tmpbuf, size * 8);
+ for(j = 0; j < h; j++){
+ for(i = 0; i < w; i++){
+ dst[i] = get_vlc2(&gb, vlc.table, 9, 3);
+ /* lines are stored as deltas between previous lines
+ * and we need to add 0x80 to the first lines of chroma planes
+ */
+ if(j) dst[i] += dst[i - stride];
+ else if(Uoff) dst[i] += 0x80;
+ }
+ dst += stride;
+ }
+ free_vlc(&vlc);
+ return 0;
+}
+
+/**
* decode a frame
* @param avctx codec context
* @param data output AVFrame
@@ -86,16 +205,18 @@
unsigned int x, y;
uint32_t *buf32;
uint32_t *luma1,*luma2,*cb,*cr;
+ uint32_t offs[4];
+ int i, ss, planes;
header = LE_32(buf);
version = header & 0xff;
header_size = (header & (1<<30))? 8 : 4; /* bit 30 means pad to 8 bytes */
- if (version > 1) {
+ if (version > 2 && version != 4) {
av_log(avctx, AV_LOG_ERROR,
"This file is encoded with Fraps version %d. " \
- "This codec can only decode version 0 and 1.\n", version);
+ "This codec can only decode version 0, 1, 2 and 4.\n", version);
return -1;
}
@@ -187,30 +308,50 @@
break;
case 2:
+ case 4:
/**
- * Fraps v2 sub-header description. All numbers are little-endian:
- * (this is all guesswork)
- *
- * 0: DWORD 'FPSx'
- * 4: DWORD 0x00000010 unknown, perhaps flags
- * 8: DWORD off_2 offset to plane 2
- * 12: DWORD off_3 offset to plane 3
- * 16: 256xDWORD freqtbl_1 frequency table for plane 1
- * 1040: plane_1
- * ...
- * off_2: 256xDWORD freqtbl_2 frequency table for plane 2
- * plane_2
- * ...
- * off_3: 256xDWORD freqtbl_3 frequency table for plane 3
- * plane_3
+ * Fraps v2 is Huffman-coded YUV420 planes
+ * Fraps v4 is the same except it works in grayscale
*/
- if ((BE_32(buf) != FPS_TAG)||(buf_size < (3*1024 + 8))) {
+ avctx->pix_fmt = (version == 2) ? PIX_FMT_YUV420P : PIX_FMT_GRAY8;
+ planes = (version == 2) ? 3 : 1;
+ f->reference = 1;
+ f->buffer_hints = FF_BUFFER_HINTS_VALID |
+ FF_BUFFER_HINTS_PRESERVE |
+ FF_BUFFER_HINTS_REUSABLE;
+ if (avctx->reget_buffer(avctx, f)) {
+ av_log(avctx, AV_LOG_ERROR, "reget_buffer() failed\n");
+ return -1;
+ }
+ /* skip frame */
+ if(buf_size == 8) {
+ f->pict_type = FF_P_TYPE;
+ f->key_frame = 0;
+ break;
+ }
+ f->pict_type = FF_I_TYPE;
+ f->key_frame = 1;
+ if ((LE_32(buf) != FPS_TAG)||(buf_size < (planes*1024 + 24))) {
av_log(avctx, AV_LOG_ERROR, "Fraps: error in data stream\n");
return -1;
}
-
- /* NOT FINISHED */
-
+ for(i = 0; i < planes; i++) {
+ offs[i] = LE_32(buf + 4 + i * 4);
+ if(offs[i] >= buf_size || (i && offs[i] <= offs[i - 1] + 1024)) {
+ av_log(avctx, AV_LOG_ERROR, "Fraps: plane %i offset is out of bounds\n", i);
+ return -1;
+ }
+ }
+ offs[planes] = buf_size;
+ for(i = 0; i < planes; i++){
+ ss = !!i;
+ s->tmpbuf = av_realloc(s->tmpbuf, offs[i + 1] - offs[i] - 1024);
+ if(fraps2_decode_plane(s, f->data[i], f->linesize[i], avctx->width >> ss,
+ avctx->height >> ss, buf + offs[i], offs[i + 1] - offs[i], ss) < 0) {
+ av_log(avctx, AV_LOG_ERROR, "Error decoding plane 0\n");
+ return -1;
+ }
+ }
break;
}
@@ -233,6 +374,7 @@
if (s->frame.data[0])
avctx->release_buffer(avctx, &s->frame);
+ av_freep(&s->tmpbuf);
return 0;
}
More information about the ffmpeg-devel
mailing list