snow: split snow in snowdec and snowenc

The common non inlined code goes in snow.c, the common inlined code in
snow.h, tables move in snowdata.h (included only by snow.c)

1 files changed, 614 insertions, 0 deletions

diff --git a/libavcodec/snow.h b/libavcodec/snow.h
index db61b875c4..f43deb1d72 100644
--- a/libavcodec/snow.h
+++ b/libavcodec/snow.h
@@ -25,6 +25,10 @@
 #include "dsputil.h"
 #include "dwt.h"
 
+#include "rangecoder.h"
+#include "mathops.h"
+#include "mpegvideo.h"
+
 #define MID_STATE 128
 
 #define MAX_PLANES 4
@@ -36,6 +40,138 @@
 
 #define LOG2_OBMC_MAX 8
 #define OBMC_MAX (1<<(LOG2_OBMC_MAX))
+typedef struct BlockNode{
+    int16_t mx;
+    int16_t my;
+    uint8_t ref;
+    uint8_t color[3];
+    uint8_t type;
+//#define TYPE_SPLIT    1
+#define BLOCK_INTRA   1
+#define BLOCK_OPT     2
+//#define TYPE_NOCOLOR  4
+    uint8_t level; //FIXME merge into type?
+}BlockNode;
+
+static const BlockNode null_block= { //FIXME add border maybe
+    .color= {128,128,128},
+    .mx= 0,
+    .my= 0,
+    .ref= 0,
+    .type= 0,
+    .level= 0,
+};
+
+#define LOG2_MB_SIZE 4
+#define MB_SIZE (1<<LOG2_MB_SIZE)
+#define ENCODER_EXTRA_BITS 4
+#define HTAPS_MAX 8
+
+typedef struct x_and_coeff{
+    int16_t x;
+    uint16_t coeff;
+} x_and_coeff;
+
+typedef struct SubBand{
+    int level;
+    int stride;
+    int width;
+    int height;
+    int qlog;        ///< log(qscale)/log[2^(1/6)]
+    DWTELEM *buf;
+    IDWTELEM *ibuf;
+    int buf_x_offset;
+    int buf_y_offset;
+    int stride_line; ///< Stride measured in lines, not pixels.
+    x_and_coeff * x_coeff;
+    struct SubBand *parent;
+    uint8_t state[/*7*2*/ 7 + 512][32];
+}SubBand;
+
+typedef struct Plane{
+    int width;
+    int height;
+    SubBand band[MAX_DECOMPOSITIONS][4];
+
+    int htaps;
+    int8_t hcoeff[HTAPS_MAX/2];
+    int diag_mc;
+    int fast_mc;
+
+    int last_htaps;
+    int8_t last_hcoeff[HTAPS_MAX/2];
+    int last_diag_mc;
+}Plane;
+
+typedef struct SnowContext{
+    AVClass *class;
+    AVCodecContext *avctx;
+    RangeCoder c;
+    DSPContext dsp;
+    DWTContext dwt;
+    AVFrame new_picture;
+    AVFrame input_picture;              ///< new_picture with the internal linesizes
+    AVFrame current_picture;
+    AVFrame last_picture[MAX_REF_FRAMES];
+    uint8_t *halfpel_plane[MAX_REF_FRAMES][4][4];
+    AVFrame mconly_picture;
+//     uint8_t q_context[16];
+    uint8_t header_state[32];
+    uint8_t block_state[128 + 32*128];
+    int keyframe;
+    int always_reset;
+    int version;
+    int spatial_decomposition_type;
+    int last_spatial_decomposition_type;
+    int temporal_decomposition_type;
+    int spatial_decomposition_count;
+    int last_spatial_decomposition_count;
+    int temporal_decomposition_count;
+    int max_ref_frames;
+    int ref_frames;
+    int16_t (*ref_mvs[MAX_REF_FRAMES])[2];
+    uint32_t *ref_scores[MAX_REF_FRAMES];
+    DWTELEM *spatial_dwt_buffer;
+    IDWTELEM *spatial_idwt_buffer;
+    int colorspace_type;
+    int chroma_h_shift;
+    int chroma_v_shift;
+    int spatial_scalability;
+    int qlog;
+    int last_qlog;
+    int lambda;
+    int lambda2;
+    int pass1_rc;
+    int mv_scale;
+    int last_mv_scale;
+    int qbias;
+    int last_qbias;
+#define QBIAS_SHIFT 3
+    int b_width;
+    int b_height;
+    int block_max_depth;
+    int last_block_max_depth;
+    Plane plane[MAX_PLANES];
+    BlockNode *block;
+#define ME_CACHE_SIZE 1024
+    int me_cache[ME_CACHE_SIZE];
+    int me_cache_generation;
+    slice_buffer sb;
+    int memc_only;
+
+    MpegEncContext m; // needed for motion estimation, should not be used for anything else, the idea is to eventually make the motion estimation independent of MpegEncContext, so this will be removed then (FIXME/XXX)
+
+    uint8_t *scratchbuf;
+}SnowContext;
+
+/* Tables */
+extern const uint8_t * const obmc_tab[4];
+#ifdef __sgi
+// Avoid a name clash on SGI IRIX
+#undef qexp
+#endif
+extern uint8_t qexp[QROOT];
+extern int scale_mv_ref[MAX_REF_FRAMES][MAX_REF_FRAMES];
 
 /* C bits used by mmx/sse2/altivec */
 
@@ -75,4 +211,482 @@ static av_always_inline void snow_horizontal_compose_liftS_lead_out(int i, IDWTE
         }
 }
 
+/* common code */
+
+int ff_snow_common_init(AVCodecContext *avctx);
+int ff_snow_common_init_after_header(AVCodecContext *avctx);
+void ff_snow_common_end(SnowContext *s);
+void ff_snow_release_buffer(AVCodecContext *avctx);
+void ff_snow_reset_contexts(SnowContext *s);
+int ff_snow_alloc_blocks(SnowContext *s);
+int ff_snow_frame_start(SnowContext *s);
+void ff_snow_pred_block(SnowContext *s, uint8_t *dst, uint8_t *tmp, int stride,
+                     int sx, int sy, int b_w, int b_h, BlockNode *block,
+                     int plane_index, int w, int h);
+/* common inline functions */
+//XXX doublecheck all of them should stay inlined
+
+static inline void snow_set_blocks(SnowContext *s, int level, int x, int y, int l, int cb, int cr, int mx, int my, int ref, int type){
+    const int w= s->b_width << s->block_max_depth;
+    const int rem_depth= s->block_max_depth - level;
+    const int index= (x + y*w) << rem_depth;
+    const int block_w= 1<<rem_depth;
+    BlockNode block;
+    int i,j;
+
+    block.color[0]= l;
+    block.color[1]= cb;
+    block.color[2]= cr;
+    block.mx= mx;
+    block.my= my;
+    block.ref= ref;
+    block.type= type;
+    block.level= level;
+
+    for(j=0; j<block_w; j++){
+        for(i=0; i<block_w; i++){
+            s->block[index + i + j*w]= block;
+        }
+    }
+}
+
+static inline void pred_mv(SnowContext *s, int *mx, int *my, int ref,
+                           const BlockNode *left, const BlockNode *top, const BlockNode *tr){
+    if(s->ref_frames == 1){
+        *mx = mid_pred(left->mx, top->mx, tr->mx);
+        *my = mid_pred(left->my, top->my, tr->my);
+    }else{
+        const int *scale = scale_mv_ref[ref];
+        *mx = mid_pred((left->mx * scale[left->ref] + 128) >>8,
+                       (top ->mx * scale[top ->ref] + 128) >>8,
+                       (tr  ->mx * scale[tr  ->ref] + 128) >>8);
+        *my = mid_pred((left->my * scale[left->ref] + 128) >>8,
+                       (top ->my * scale[top ->ref] + 128) >>8,
+                       (tr  ->my * scale[tr  ->ref] + 128) >>8);
+    }
+}
+
+static av_always_inline int same_block(BlockNode *a, BlockNode *b){
+    if((a->type&BLOCK_INTRA) && (b->type&BLOCK_INTRA)){
+        return !((a->color[0] - b->color[0]) | (a->color[1] - b->color[1]) | (a->color[2] - b->color[2]));
+    }else{
+        return !((a->mx - b->mx) | (a->my - b->my) | (a->ref - b->ref) | ((a->type ^ b->type)&BLOCK_INTRA));
+    }
+}
+
+//FIXME name cleanup (b_w, block_w, b_width stuff)
+//XXX should we really inline it?
+static av_always_inline void add_yblock(SnowContext *s, int sliced, slice_buffer *sb, IDWTELEM *dst, uint8_t *dst8, const uint8_t *obmc, int src_x, int src_y, int b_w, int b_h, int w, int h, int dst_stride, int src_stride, int obmc_stride, int b_x, int b_y, int add, int offset_dst, int plane_index){
+    const int b_width = s->b_width  << s->block_max_depth;
+    const int b_height= s->b_height << s->block_max_depth;
+    const int b_stride= b_width;
+    BlockNode *lt= &s->block[b_x + b_y*b_stride];
+    BlockNode *rt= lt+1;
+    BlockNode *lb= lt+b_stride;
+    BlockNode *rb= lb+1;
+    uint8_t *block[4];
+    int tmp_step= src_stride >= 7*MB_SIZE ? MB_SIZE : MB_SIZE*src_stride;
+    uint8_t *tmp = s->scratchbuf;
+    uint8_t *ptmp;
+    int x,y;
+
+    if(b_x<0){
+        lt= rt;
+        lb= rb;
+    }else if(b_x + 1 >= b_width){
+        rt= lt;
+        rb= lb;
+    }
+    if(b_y<0){
+        lt= lb;
+        rt= rb;
+    }else if(b_y + 1 >= b_height){
+        lb= lt;
+        rb= rt;
+    }
+
+    if(src_x<0){ //FIXME merge with prev & always round internal width up to *16
+        obmc -= src_x;
+        b_w += src_x;
+        if(!sliced && !offset_dst)
+            dst -= src_x;
+        src_x=0;
+    }else if(src_x + b_w > w){
+        b_w = w - src_x;
+    }
+    if(src_y<0){
+        obmc -= src_y*obmc_stride;
+        b_h += src_y;
+        if(!sliced && !offset_dst)
+            dst -= src_y*dst_stride;
+        src_y=0;
+    }else if(src_y + b_h> h){
+        b_h = h - src_y;
+    }
+
+    if(b_w<=0 || b_h<=0) return;
+
+    assert(src_stride > 2*MB_SIZE + 5);
+
+    if(!sliced && offset_dst)
+        dst += src_x + src_y*dst_stride;
+    dst8+= src_x + src_y*src_stride;
+//    src += src_x + src_y*src_stride;
+
+    ptmp= tmp + 3*tmp_step;
+    block[0]= ptmp;
+    ptmp+=tmp_step;
+    ff_snow_pred_block(s, block[0], tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h);
+
+    if(same_block(lt, rt)){
+        block[1]= block[0];
+    }else{
+        block[1]= ptmp;
+        ptmp+=tmp_step;
+        ff_snow_pred_block(s, block[1], tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h);
+    }
+
+    if(same_block(lt, lb)){
+        block[2]= block[0];
+    }else if(same_block(rt, lb)){
+        block[2]= block[1];
+    }else{
+        block[2]= ptmp;
+        ptmp+=tmp_step;
+        ff_snow_pred_block(s, block[2], tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h);
+    }
+
+    if(same_block(lt, rb) ){
+        block[3]= block[0];
+    }else if(same_block(rt, rb)){
+        block[3]= block[1];
+    }else if(same_block(lb, rb)){
+        block[3]= block[2];
+    }else{
+        block[3]= ptmp;
+        ff_snow_pred_block(s, block[3], tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h);
+    }
+    if(sliced){
+        s->dwt.inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
+    }else{
+        for(y=0; y<b_h; y++){
+            //FIXME ugly misuse of obmc_stride
+            const uint8_t *obmc1= obmc + y*obmc_stride;
+            const uint8_t *obmc2= obmc1+ (obmc_stride>>1);
+            const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);
+            const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
+            for(x=0; x<b_w; x++){
+                int v=   obmc1[x] * block[3][x + y*src_stride]
+                        +obmc2[x] * block[2][x + y*src_stride]
+                        +obmc3[x] * block[1][x + y*src_stride]
+                        +obmc4[x] * block[0][x + y*src_stride];
+
+                v <<= 8 - LOG2_OBMC_MAX;
+                if(FRAC_BITS != 8){
+                    v >>= 8 - FRAC_BITS;
+                }
+                if(add){
+                    v += dst[x + y*dst_stride];
+                    v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;
+                    if(v&(~255)) v= ~(v>>31);
+                    dst8[x + y*src_stride] = v;
+                }else{
+                    dst[x + y*dst_stride] -= v;
+                }
+            }
+        }
+    }
+}
+
+static av_always_inline void predict_slice(SnowContext *s, IDWTELEM *buf, int plane_index, int add, int mb_y){
+    Plane *p= &s->plane[plane_index];
+    const int mb_w= s->b_width  << s->block_max_depth;
+    const int mb_h= s->b_height << s->block_max_depth;
+    int x, y, mb_x;
+    int block_size = MB_SIZE >> s->block_max_depth;
+    int block_w    = plane_index ? block_size/2 : block_size;
+    const uint8_t *obmc  = plane_index ? obmc_tab[s->block_max_depth+1] : obmc_tab[s->block_max_depth];
+    const int obmc_stride= plane_index ? block_size : 2*block_size;
+    int ref_stride= s->current_picture.linesize[plane_index];
+    uint8_t *dst8= s->current_picture.data[plane_index];
+    int w= p->width;
+    int h= p->height;
+
+    if(s->keyframe || (s->avctx->debug&512)){
+        if(mb_y==mb_h)
+            return;
+
+        if(add){
+            for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++){
+                for(x=0; x<w; x++){
+                    int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
+                    v >>= FRAC_BITS;
+                    if(v&(~255)) v= ~(v>>31);
+                    dst8[x + y*ref_stride]= v;
+                }
+            }
+        }else{
+            for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++){
+                for(x=0; x<w; x++){
+                    buf[x + y*w]-= 128<<FRAC_BITS;
+                }
+            }
+        }
+
+        return;
+    }
+
+    for(mb_x=0; mb_x<=mb_w; mb_x++){
+        add_yblock(s, 0, NULL, buf, dst8, obmc,
+                   block_w*mb_x - block_w/2,
+                   block_w*mb_y - block_w/2,
+                   block_w, block_w,
+                   w, h,
+                   w, ref_stride, obmc_stride,
+                   mb_x - 1, mb_y - 1,
+                   add, 1, plane_index);
+    }
+}
+
+static av_always_inline void predict_plane(SnowContext *s, IDWTELEM *buf, int plane_index, int add){
+    const int mb_h= s->b_height << s->block_max_depth;
+    int mb_y;
+    for(mb_y=0; mb_y<=mb_h; mb_y++)
+        predict_slice(s, buf, plane_index, add, mb_y);
+}
+
+static inline void set_blocks(SnowContext *s, int level, int x, int y, int l, int cb, int cr, int mx, int my, int ref, int type){
+    const int w= s->b_width << s->block_max_depth;
+    const int rem_depth= s->block_max_depth - level;
+    const int index= (x + y*w) << rem_depth;
+    const int block_w= 1<<rem_depth;
+    BlockNode block;
+    int i,j;
+
+    block.color[0]= l;
+    block.color[1]= cb;
+    block.color[2]= cr;
+    block.mx= mx;
+    block.my= my;
+    block.ref= ref;
+    block.type= type;
+    block.level= level;
+
+    for(j=0; j<block_w; j++){
+        for(i=0; i<block_w; i++){
+            s->block[index + i + j*w]= block;
+        }
+    }
+}
+
+static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
+    const int offset[3]= {
+          y*c->  stride + x,
+        ((y*c->uvstride + x)>>1),
+        ((y*c->uvstride + x)>>1),
+    };
+    int i;
+    for(i=0; i<3; i++){
+        c->src[0][i]= src [i];
+        c->ref[0][i]= ref [i] + offset[i];
+    }
+    assert(!ref_index);
+}
+
+
+/* bitstream functions */
+
+extern const int8_t quant3bA[256];
+
+#define QEXPSHIFT (7-FRAC_BITS+8) //FIXME try to change this to 0
+
+static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed){
+    int i;
+
+    if(v){
+        const int a= FFABS(v);
+        const int e= av_log2(a);
+        const int el= FFMIN(e, 10);
+        put_rac(c, state+0, 0);
+
+        for(i=0; i<el; i++){
+            put_rac(c, state+1+i, 1);  //1..10
+        }
+        for(; i<e; i++){
+            put_rac(c, state+1+9, 1);  //1..10
+        }
+        put_rac(c, state+1+FFMIN(i,9), 0);
+
+        for(i=e-1; i>=el; i--){
+            put_rac(c, state+22+9, (a>>i)&1); //22..31
+        }
+        for(; i>=0; i--){
+            put_rac(c, state+22+i, (a>>i)&1); //22..31
+        }
+
+        if(is_signed)
+            put_rac(c, state+11 + el, v < 0); //11..21
+    }else{
+        put_rac(c, state+0, 1);
+    }
+}
+
+static inline int get_symbol(RangeCoder *c, uint8_t *state, int is_signed){
+    if(get_rac(c, state+0))
+        return 0;
+    else{
+        int i, e, a;
+        e= 0;
+        while(get_rac(c, state+1 + FFMIN(e,9))){ //1..10
+            e++;
+        }
+
+        a= 1;
+        for(i=e-1; i>=0; i--){
+            a += a + get_rac(c, state+22 + FFMIN(i,9)); //22..31
+        }
+
+        e= -(is_signed && get_rac(c, state+11 + FFMIN(e,10))); //11..21
+        return (a^e)-e;
+    }
+}
+
+static inline void put_symbol2(RangeCoder *c, uint8_t *state, int v, int log2){
+    int i;
+    int r= log2>=0 ? 1<<log2 : 1;
+
+    assert(v>=0);
+    assert(log2>=-4);
+
+    while(v >= r){
+        put_rac(c, state+4+log2, 1);
+        v -= r;
+        log2++;
+        if(log2>0) r+=r;
+    }
+    put_rac(c, state+4+log2, 0);
+
+    for(i=log2-1; i>=0; i--){
+        put_rac(c, state+31-i, (v>>i)&1);
+    }
+}
+
+static inline int get_symbol2(RangeCoder *c, uint8_t *state, int log2){
+    int i;
+    int r= log2>=0 ? 1<<log2 : 1;
+    int v=0;
+
+    assert(log2>=-4);
+
+    while(get_rac(c, state+4+log2)){
+        v+= r;
+        log2++;
+        if(log2>0) r+=r;
+    }
+
+    for(i=log2-1; i>=0; i--){
+        v+= get_rac(c, state+31-i)<<i;
+    }
+
+    return v;
+}
+
+static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, int orientation){
+    const int w= b->width;
+    const int h= b->height;
+    int x,y;
+
+    int run, runs;
+    x_and_coeff *xc= b->x_coeff;
+    x_and_coeff *prev_xc= NULL;
+    x_and_coeff *prev2_xc= xc;
+    x_and_coeff *parent_xc= parent ? parent->x_coeff : NULL;
+    x_and_coeff *prev_parent_xc= parent_xc;
+
+    runs= get_symbol2(&s->c, b->state[30], 0);
+    if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
+    else           run= INT_MAX;
+
+    for(y=0; y<h; y++){
+        int v=0;
+        int lt=0, t=0, rt=0;
+
+        if(y && prev_xc->x == 0){
+            rt= prev_xc->coeff;
+        }
+        for(x=0; x<w; x++){
+            int p=0;
+            const int l= v;
+
+            lt= t; t= rt;
+
+            if(y){
+                if(prev_xc->x <= x)
+                    prev_xc++;
+                if(prev_xc->x == x + 1)
+                    rt= prev_xc->coeff;
+                else
+                    rt=0;
+            }
+            if(parent_xc){
+                if(x>>1 > parent_xc->x){
+                    parent_xc++;
+                }
+                if(x>>1 == parent_xc->x){
+                    p= parent_xc->coeff;
+                }
+            }
+            if(/*ll|*/l|lt|t|rt|p){
+                int context= av_log2(/*FFABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1));
+
+                v=get_rac(&s->c, &b->state[0][context]);
+                if(v){
+                    v= 2*(get_symbol2(&s->c, b->state[context + 2], context-4) + 1);
+                    v+=get_rac(&s->c, &b->state[0][16 + 1 + 3 + quant3bA[l&0xFF] + 3*quant3bA[t&0xFF]]);
+
+                    xc->x=x;
+                    (xc++)->coeff= v;
+                }
+            }else{
+                if(!run){
+                    if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
+                    else           run= INT_MAX;
+                    v= 2*(get_symbol2(&s->c, b->state[0 + 2], 0-4) + 1);
+                    v+=get_rac(&s->c, &b->state[0][16 + 1 + 3]);
+
+                    xc->x=x;
+                    (xc++)->coeff= v;
+                }else{
+                    int max_run;
+                    run--;
+                    v=0;
+
+                    if(y) max_run= FFMIN(run, prev_xc->x - x - 2);
+                    else  max_run= FFMIN(run, w-x-1);
+                    if(parent_xc)
+                        max_run= FFMIN(max_run, 2*parent_xc->x - x - 1);
+                    x+= max_run;
+                    run-= max_run;
+                }
+            }
+        }
+        (xc++)->x= w+1; //end marker
+        prev_xc= prev2_xc;
+        prev2_xc= xc;
+
+        if(parent_xc){
+            if(y&1){
+                while(parent_xc->x != parent->width+1)
+                    parent_xc++;
+                parent_xc++;
+                prev_parent_xc= parent_xc;
+            }else{
+                parent_xc= prev_parent_xc;
+            }
+        }
+    }
+
+    (xc++)->x= w+1; //end marker
+}
+
 #endif /* AVCODEC_SNOW_H */