[FFmpeg-devel] [PATCH] RV40 Decoder - 1/3 - Loop Filter
Kostya
kostya.shishkov
Thu Nov 27 08:23:21 CET 2008
$subj
Fixed several bugs in chroma loop filtering introduced during
refactoring, now it's bitexact again.
-------------- next part --------------
Index: libavcodec/rv40.c
===================================================================
--- libavcodec/rv40.c (revision 15920)
+++ libavcodec/rv40.c (working copy)
@@ -389,7 +389,251 @@
alpha, beta, beta2, chroma, edge);
}
+enum RV40BlockPos{
+ POS_CUR,
+ POS_TOP,
+ POS_LEFT,
+ POS_BOTTOM,
+};
+
+#define MASK_CUR 0x0001
+#define MASK_RIGHT 0x0008
+#define MASK_BOTTOM 0x0010
+#define MASK_TOP 0x1000
+#define MASK_Y_TOP_ROW 0x000F
+#define MASK_Y_LAST_ROW 0xF000
+#define MASK_Y_LEFT_COL 0x1111
+#define MASK_Y_RIGHT_COL 0x8888
+#define MASK_C_TOP_ROW 0x0003
+#define MASK_C_LAST_ROW 0x000C
+#define MASK_C_LEFT_COL 0x0005
+#define MASK_C_RIGHT_COL 0x000A
+
+static const int neighbour_offs_x[4] = { 0, 0, -1, 0 };
+static const int neighbour_offs_y[4] = { 0, -1, 0, 1 };
+
/**
+ * RV40 loop filtering function
+ */
+static void rv40_loop_filter(RV34DecContext *r)
+{
+ MpegEncContext *s = &r->s;
+ int mb_pos;
+ int i, j, k;
+ uint8_t *Y, *C;
+ int alpha, beta, betaY, betaC;
+ int q;
+ int mbtype[4]; ///< current macroblock and its neighbours types
+ /**
+ * flags indicating that macroblock can be filtered with strong filter
+ * it is set only for intra coded MB and MB with DCs coded separately
+ */
+ int mb_strong[4];
+ int clip[4]; ///< MB filter clipping value calculated from filtering strength
+ /**
+ * coded block patterns for luma part of current macroblock and its neighbours
+ * Format:
+ * LSB corresponds to the top left block,
+ * each nibble represents one row of subblocks.
+ */
+ int cbp[4];
+ /**
+ * coded block patterns for chroma part of current macroblock and its neighbours
+ * Format is the same as for luma with two subblocks in a row.
+ */
+ int uvcbp[4][2];
+ /**
+ * This mask represents the pattern of luma subblocks that should be filtered
+ * in addition to the coded ones because because they lie at the edge of
+ * 8x8 block with different enough motion vectors
+ */
+ int mvmasks[4];
+
+ for(s->mb_y = 0; s->mb_y < s->mb_height; s->mb_y++){
+ mb_pos = s->mb_y * s->mb_stride;
+ for(s->mb_x = 0; s->mb_x < s->mb_width; s->mb_x++, mb_pos++){
+ int mbtype = s->current_picture_ptr->mb_type[mb_pos];
+ if(IS_INTRA(mbtype) || IS_SEPARATE_DC(mbtype))
+ r->cbp_luma [mb_pos] = 0xFFFF;
+ if(IS_INTRA(mbtype))
+ r->cbp_chroma[mb_pos] = 0xFF;
+ }
+ }
+ for(s->mb_y = 0; s->mb_y < s->mb_height; s->mb_y++){
+ mb_pos = s->mb_y * s->mb_stride;
+ for(s->mb_x = 0; s->mb_x < s->mb_width; s->mb_x++, mb_pos++){
+ int y_h_deblock, y_v_deblock;
+ int c_v_deblock[2], c_h_deblock[2];
+ int clip_left;
+ int avail[4];
+ int y_to_deblock, c_to_deblock[2];
+
+ ff_init_block_index(s);
+ ff_update_block_index(s);
+ Y = s->dest[0];
+ q = s->current_picture_ptr->qscale_table[mb_pos];
+ alpha = rv40_alpha_tab[q];
+ beta = rv40_beta_tab [q];
+ betaY = betaC = beta * 3;
+ if(s->width * s->height <= 176*144)
+ betaY += beta;
+
+ avail[0] = 1;
+ avail[1] = s->mb_y;
+ avail[2] = s->mb_x;
+ avail[3] = s->mb_y < s->mb_height - 1;
+ for(i = 0; i < 4; i++){
+ if(avail[i]){
+ int pos = mb_pos + neighbour_offs_x[i] + neighbour_offs_y[i]*s->mb_stride;
+ mvmasks[i] = r->deblock_coefs[pos];
+ mbtype [i] = s->current_picture_ptr->mb_type[pos];
+ cbp [i] = r->cbp_luma[pos];
+ uvcbp[i][0] = r->cbp_chroma[pos] & 0xF;
+ uvcbp[i][1] = r->cbp_chroma[pos] >> 4;
+ }else{
+ mvmasks[i] = 0;
+ mbtype [i] = mbtype[0];
+ cbp [i] = 0;
+ uvcbp[i][0] = uvcbp[i][1] = 0;
+ }
+ mb_strong[i] = IS_INTRA(mbtype[i]) || IS_SEPARATE_DC(mbtype[i]);
+ clip[i] = rv40_filter_clip_tbl[mb_strong[i] + 1][q];
+ }
+ y_to_deblock = cbp[POS_CUR]
+ | (cbp[POS_BOTTOM] << 16)
+ | mvmasks[POS_CUR]
+ | (mvmasks[POS_BOTTOM] << 16);
+ /* This pattern contains bits signalling that horizontal edges of
+ * the current block can be filtered.
+ * That happens when either of adjacent subblocks is coded or lies on
+ * the edge of 8x8 blocks with motion vectors differing by more than
+ * 3/4 pel in any component (any edge orientation for some reason).
+ */
+ y_h_deblock = y_to_deblock
+ | ((cbp[POS_CUR] << 4) & ~MASK_Y_TOP_ROW)
+ | ((cbp[POS_TOP] & MASK_Y_LAST_ROW) >> 12);
+ /* This pattern contains bits signalling that vertical edges of
+ * the current block can be filtered.
+ * That happens when either of adjacent subblocks is coded or lies on
+ * the edge of 8x8 blocks with motion vectors differing by more than
+ * 3/4 pel in any component (any edge orientation for some reason).
+ */
+ y_v_deblock = y_to_deblock
+ | ((cbp[POS_CUR] << 1) & ~MASK_Y_LEFT_COL)
+ | ((cbp[POS_LEFT] & MASK_Y_RIGHT_COL) >> 3);
+ if(!s->mb_x)
+ y_v_deblock &= ~MASK_Y_LEFT_COL;
+ if(!s->mb_y)
+ y_h_deblock &= ~MASK_Y_TOP_ROW;
+ if(s->mb_y == s->mb_height - 1 || (mb_strong[POS_CUR] || mb_strong[POS_BOTTOM]))
+ y_h_deblock &= ~(MASK_Y_TOP_ROW << 16);
+ /* Calculating chroma patterns is similar and easier since there is
+ * no motion vector pattern for them.
+ */
+ for(i = 0; i < 2; i++){
+ c_to_deblock[i] = (uvcbp[POS_BOTTOM][i] << 4) | uvcbp[POS_CUR][i];
+ c_v_deblock[i] = c_to_deblock[i]
+ | ((uvcbp[POS_CUR] [i] << 1) & ~MASK_C_LEFT_COL)
+ | ((uvcbp[POS_LEFT][i] & MASK_C_RIGHT_COL) >> 1);
+ c_h_deblock[i] = c_to_deblock[i]
+ | ((uvcbp[POS_TOP][i] & MASK_C_LAST_ROW) >> 2)
+ | (uvcbp[POS_CUR][i] << 2);
+ if(!s->mb_x)
+ c_v_deblock[i] &= ~MASK_C_LEFT_COL;
+ if(!s->mb_y)
+ c_h_deblock[i] &= ~MASK_C_TOP_ROW;
+ if(s->mb_y == s->mb_height - 1 || mb_strong[POS_CUR] || mb_strong[POS_BOTTOM])
+ c_h_deblock[i] &= ~(MASK_C_TOP_ROW << 4);
+ }
+
+ for(j = 0; j < 16; j += 4){
+ Y = s->dest[0] + j * s->linesize;
+ for(i = 0; i < 4; i++, Y += 4){
+ int ij = i + j;
+ int clip_cur = y_to_deblock & (MASK_CUR << ij) ? clip[POS_CUR] : 0;
+ int dither = j ? ij : i*4;
+
+ // if bottom block is coded then we can filter its top edge
+ // (or bottom edge of this block, which is the same)
+ if(y_h_deblock & (MASK_BOTTOM << ij)){
+ rv40_h_loop_filter(Y+4*s->linesize, s->linesize, dither,
+ y_to_deblock & (MASK_BOTTOM << ij) ? clip[POS_CUR] : 0,
+ clip_cur,
+ alpha, beta, betaY, 0, 0);
+ }
+ // filter left block edge in ordinary mode (with low filtering strength)
+ if(y_v_deblock & (MASK_CUR << ij) && (i || !(mb_strong[POS_CUR] || mb_strong[POS_LEFT]))){
+ if(!i)
+ clip_left = (cbp[POS_LEFT] | mvmasks[POS_LEFT]) & (MASK_RIGHT << j) ? clip[POS_LEFT] : 0;
+ else
+ clip_left = y_to_deblock & (MASK_CUR << (ij-1)) ? clip[POS_CUR] : 0;
+ rv40_v_loop_filter(Y, s->linesize, dither,
+ clip_cur,
+ clip_left,
+ alpha, beta, betaY, 0, 0);
+ }
+ // filter top edge of the current macroblock when filtering strength is high
+ if(!j && y_h_deblock & (MASK_CUR << i) && (mb_strong[POS_CUR] || mb_strong[POS_TOP])){
+ rv40_h_loop_filter(Y, s->linesize, dither,
+ clip_cur,
+ (cbp[POS_TOP] | mvmasks[POS_TOP]) & (MASK_TOP << i) ? clip[POS_TOP] : 0,
+ alpha, beta, betaY, 0, 1);
+ }
+ // filter left block edge in edge mode (with high filtering strength)
+ if(y_v_deblock & (MASK_CUR << ij) && !i && (mb_strong[POS_CUR] || mb_strong[POS_LEFT])){
+ clip_left = (cbp[POS_LEFT] | mvmasks[POS_LEFT]) & (MASK_RIGHT << j) ? clip[POS_LEFT] : 0;
+ rv40_v_loop_filter(Y, s->linesize, dither,
+ clip_cur,
+ clip_left,
+ alpha, beta, betaY, 0, 1);
+ }
+ }
+ }
+ for(k = 0; k < 2; k++){
+ for(j = 0; j < 2; j++){
+ C = s->dest[k+1] + j * 4 * s->uvlinesize;
+ for(i = 0; i < 2; i++, C += 4){
+ int ij = i + j*2;
+ int clip_cur = c_to_deblock[k] & (MASK_CUR << ij) ? clip[POS_CUR] : 0;
+ if(c_h_deblock[k] & (MASK_CUR << (ij+2))){
+ int clip_bot = c_to_deblock[k] & (MASK_CUR << (ij+2)) ? clip[POS_CUR] : 0;
+ rv40_h_loop_filter(C+4*s->uvlinesize, s->uvlinesize, i*8,
+ clip_bot,
+ clip_cur,
+ alpha, beta, betaC, 1, 0);
+ }
+ if((c_v_deblock[k] & (MASK_CUR << ij)) && (i || !(mb_strong[POS_CUR] || mb_strong[POS_LEFT]))){
+ if(!i)
+ clip_left = uvcbp[POS_LEFT][k] & (MASK_CUR << (2*j+1)) ? clip[POS_LEFT] : 0;
+ else
+ clip_left = c_to_deblock[k] & (MASK_CUR << (ij-1)) ? clip[POS_CUR] : 0;
+ rv40_v_loop_filter(C, s->uvlinesize, j*8,
+ clip_cur,
+ clip_left,
+ alpha, beta, betaC, 1, 0);
+ }
+ if(!j && c_h_deblock[k] & (MASK_CUR << ij) && (mb_strong[POS_CUR] || mb_strong[POS_TOP])){
+ int clip_top = uvcbp[POS_TOP][k] & (MASK_CUR << (ij+2)) ? clip[POS_TOP] : 0;
+ rv40_h_loop_filter(C, s->uvlinesize, i*8,
+ clip_cur,
+ clip_top,
+ alpha, beta, betaC, 1, 1);
+ }
+ if(c_v_deblock[k] & (MASK_CUR << ij) && !i && (mb_strong[POS_CUR] || mb_strong[POS_LEFT])){
+ clip_left = uvcbp[POS_LEFT][k] & (MASK_CUR << (2*j+1)) ? clip[POS_LEFT] : 0;
+ rv40_v_loop_filter(C, s->uvlinesize, j*8,
+ clip_cur,
+ clip_left,
+ alpha, beta, betaC, 1, 1);
+ }
+ }
+ }
+ }
+ }
+ }
+}
+
+/**
* Initialize decoder.
*/
static av_cold int rv40_decode_init(AVCodecContext *avctx)
@@ -403,6 +647,7 @@
r->parse_slice_header = rv40_parse_slice_header;
r->decode_intra_types = rv40_decode_intra_types;
r->decode_mb_info = rv40_decode_mb_info;
+ r->loop_filter = rv40_loop_filter;
r->luma_dc_quant_i = rv40_luma_dc_quant[0];
r->luma_dc_quant_p = rv40_luma_dc_quant[1];
return 0;
More information about the ffmpeg-devel
mailing list