[FFmpeg-devel] [PATCH] RV40 Loop Filter, hopefully final version
Kostya
kostya.shishkov
Sat Nov 22 15:11:15 CET 2008
This has been tested and verified to be bitexact if correct
MC functions are used.
-------------- next part --------------
Index: rv40.c
===================================================================
--- rv40.c (revision 15827)
+++ rv40.c (working copy)
@@ -285,7 +285,374 @@
}
}
+static inline void rv40_adaptive_loop_filter(uint8_t *src, const int step,
+ const int stride, const int dmode,
+ const int lim_q1, const int lim_p1,
+ const int alpha,
+ const int beta, const int beta2,
+ const int chroma, const int edge)
+{
+ int diff_p1p0[4], diff_q1q0[4], diff_p1p2[4], diff_q1q2[4];
+ int s0 = 0, s1 = 0, s2 = 0, s3 = 0;
+ uint8_t *ptr;
+ int flag_strong0 = 1, flag_strong1 = 1;
+ int strength0 = 3, strength1 = 3;
+ int i;
+ int lims;
+
+ for(i = 0, ptr = src; i < 4; i++, ptr += stride){
+ diff_p1p0[i] = ptr[-2*step] - ptr[-1*step];
+ diff_q1q0[i] = ptr[ 1*step] - ptr[ 0*step];
+ s0 += diff_p1p0[i];
+ s1 += diff_q1q0[i];
+ }
+ if(FFABS(s0) >= (beta<<2))
+ strength0 = 1;
+ if(FFABS(s1) >= (beta<<2))
+ strength1 = 1;
+ if(strength0 + strength1 <= 2)
+ return;
+
+ for(i = 0, ptr = src; i < 4; i++, ptr += stride){
+ diff_p1p2[i] = ptr[-2*step] - ptr[-3*step];
+ diff_q1q2[i] = ptr[ 1*step] - ptr[ 2*step];
+ s2 += diff_p1p2[i];
+ s3 += diff_q1q2[i];
+ }
+
+ if(!edge){
+ flag_strong0 = flag_strong1 = 0;
+ }else{
+ flag_strong0 = (strength0 == 3) && (FFABS(s2) < beta2);
+ flag_strong1 = (strength1 == 3) && (FFABS(s3) < beta2);
+ }
+
+ lims = (lim_q1 + lim_p1 + strength0 + strength1) >> 1;
+ if(flag_strong0 && flag_strong1){ /* strong filtering */
+ for(i = 0; i < 4; i++, src += stride){
+ int sflag, p0, q0, p1, q1;
+ int t = src[0*step] - src[-1*step];
+
+ if(!t) continue;
+ sflag = (alpha * FFABS(t)) >> 7;
+ if(sflag > 1) continue;
+
+ p0 = (25*src[-3*step] + 26*src[-2*step]
+ + 26*src[-1*step]
+ + 26*src[ 0*step] + 25*src[ 1*step] + rv40_dither_l[dmode + i]) >> 7;
+ q0 = (25*src[-2*step] + 26*src[-1*step]
+ + 26*src[ 0*step]
+ + 26*src[ 1*step] + 25*src[ 2*step] + rv40_dither_r[dmode + i]) >> 7;
+ if(sflag){
+ p0 = av_clip(p0, src[-1*step] - lims, src[-1*step] + lims);
+ q0 = av_clip(q0, src[ 0*step] - lims, src[ 0*step] + lims);
+ }
+ p1 = (25*src[-4*step] + 26*src[-3*step]
+ + 26*src[-2*step]
+ + 26*p0 + 25*src[ 0*step] + rv40_dither_l[dmode + i]) >> 7;
+ q1 = (25*src[-1*step] + 26*q0
+ + 26*src[ 1*step]
+ + 26*src[ 2*step] + 25*src[ 3*step] + rv40_dither_r[dmode + i]) >> 7;
+ if(sflag){
+ p1 = av_clip(p1, src[-2*step] - lims, src[-2*step] + lims);
+ q1 = av_clip(q1, src[ 1*step] - lims, src[ 1*step] + lims);
+ }
+ src[-2*step] = p1;
+ src[-1*step] = p0;
+ src[ 0*step] = q0;
+ src[ 1*step] = q1;
+ if(!chroma){
+ src[-3*step] = (25*src[-1*step] + 26*src[-2*step] + 51*src[-3*step] + 26*src[-4*step] + 64) >> 7;
+ src[ 2*step] = (25*src[ 0*step] + 26*src[ 1*step] + 51*src[ 2*step] + 26*src[ 3*step] + 64) >> 7;
+ }
+ }
+ }else if(strength0 == 3 && strength1 == 3){
+ for(i = 0; i < 4; i++, src += stride)
+ rv40_weak_loop_filter(src, step, 1, 1, alpha, beta, lims, lim_q1, lim_p1,
+ diff_p1p0[i], diff_q1q0[i], diff_p1p2[i], diff_q1q2[i]);
+ }else{
+ for(i = 0; i < 4; i++, src += stride)
+ rv40_weak_loop_filter(src, step, strength0==3, strength1==3,
+ alpha, beta, lims>>1, lim_q1>>1, lim_p1>>1,
+ diff_p1p0[i], diff_q1q0[i], diff_p1p2[i], diff_q1q2[i]);
+ }
+}
+
+static void rv40_v_loop_filter(uint8_t *src, int stride, int dmode,
+ int lim_q1, int lim_p1,
+ int alpha, int beta, int beta2, int chroma, int edge){
+ rv40_adaptive_loop_filter(src, 1, stride, dmode, lim_q1, lim_p1,
+ alpha, beta, beta2, chroma, edge);
+}
+static void rv40_h_loop_filter(uint8_t *src, int stride, int dmode,
+ int lim_q1, int lim_p1,
+ int alpha, int beta, int beta2, int chroma, int edge){
+ rv40_adaptive_loop_filter(src, stride, 1, dmode, lim_q1, lim_p1,
+ 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
+ /**
+ * macroblock filtering strength
+ * 2 for intra coded MB and MB with DCs coded separately, 1 otherwise
+ */
+ int strength[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 btype = s->current_picture_ptr->mb_type[mb_pos];
+ if(IS_INTRA(btype) || IS_SEPARATE_DC(btype))
+ r->cbp_luma [mb_pos] = 0xFFFF;
+ if(IS_INTRA(btype))
+ 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];
+
+ 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;
+ }
+ strength[i] = (IS_INTRA(mbtype[i]) || IS_SEPARATE_DC(mbtype[i])) ? 2 : 1;
+ clip[i] = rv40_filter_clip_tbl[strength[i]][q];
+ }
+ /* 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 = cbp[POS_CUR]
+ | ((cbp[POS_BOTTOM] & MASK_Y_TOP_ROW) << 16)
+ | ((cbp[POS_CUR] << 4) & ~MASK_Y_TOP_ROW)
+ | ((cbp[POS_TOP] & MASK_Y_LAST_ROW) >> 12)
+ | mvmasks[POS_CUR]
+ | ((mvmasks[POS_BOTTOM] & MASK_Y_TOP_ROW) << 16);
+ /* 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 = cbp[POS_CUR]
+ | ((cbp[POS_CUR] << 1) & ~MASK_Y_LEFT_COL)
+ | ((cbp[POS_LEFT] & MASK_Y_RIGHT_COL) >> 3)
+ | mvmasks[POS_CUR];
+ 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 || (strength[POS_CUR] == 2 || strength[POS_BOTTOM] == 2))
+ y_h_deblock &= ~(MASK_Y_TOP_ROW << 16);
+ /* We should add motion vector mask and the bottom CBP to the current
+ * CBP since it's checked in this combination afterwards.
+ */
+ cbp[POS_CUR] = cbp[POS_CUR]
+ | (cbp[POS_BOTTOM] << 16)
+ | mvmasks[POS_CUR]
+ | ((mvmasks[POS_BOTTOM] & 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_v_deblock[i] = ((uvcbp[POS_CUR][i] << 1) & ~MASK_C_RIGHT_COL)
+ | ((uvcbp[POS_LEFT][i] & MASK_C_RIGHT_COL) >> 1)
+ | ((uvcbp[POS_BOTTOM][i] & MASK_C_TOP_ROW) << 4)
+ | uvcbp[POS_CUR][i];
+ c_h_deblock[i] = ((uvcbp[POS_BOTTOM][i] & MASK_C_TOP_ROW) << 4)
+ | uvcbp[POS_CUR][i]
+ | ((uvcbp[POS_TOP][i] & MASK_C_LAST_ROW) >> 2)
+ | ((uvcbp[POS_BOTTOM][i] & MASK_C_TOP_ROW) << 6)
+ | (uvcbp[POS_CUR][i] << 2);
+ uvcbp[POS_CUR][i] = ((uvcbp[POS_BOTTOM][i] & MASK_C_TOP_ROW) << 4) | uvcbp[POS_CUR][i];
+ 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 || strength[POS_CUR] == 2 || strength[POS_BOTTOM] == 2)
+ 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 = cbp[POS_CUR] & (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,
+ cbp[POS_CUR] & (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 || !(strength[POS_CUR] == 2 || strength[POS_LEFT] == 2))){
+ if(!i){
+ clip_left = (cbp[POS_LEFT] | mvmasks[POS_LEFT]) & (MASK_RIGHT << j) ? clip[POS_LEFT] : 0;
+ }else{
+ clip_left = cbp[POS_CUR] & (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)
+ && (strength[POS_CUR] == 2 || strength[POS_TOP] == 2)){
+ 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
+ && (strength[POS_CUR] == 2 || strength[POS_LEFT] == 2)){
+ 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] + j * 4 * s->uvlinesize;
+ for(i = 0; i < 2; i++, C += 4){
+ int ij = i + j*2;
+ int clip_cur = uvcbp[POS_CUR][k] & (MASK_CUR << ij) ? clip[POS_CUR] : 0;
+ if(c_h_deblock[k] & (MASK_CUR << (ij+2))){
+ int clip_bot = uvcbp[POS_CUR][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 || !(strength[POS_CUR] == 2 || strength[POS_LEFT] == 2))){
+ if(!i){
+ clip_left = uvcbp[POS_LEFT][k] & (MASK_CUR << (2*j+1)) ? clip[POS_LEFT] : 0;
+ }else{
+ clip_left = uvcbp[POS_CUR][k] & (MASK_CUR << (ij-1)) ? clip[POS_CUR] : 0;
+ }
+ rv40_v_loop_filter(C, s->uvlinesize, j*8,
+ clip_cur,
+ clip_left,
+ alpha, beta, betaY, 1, 0);
+ }
+ if(!j && c_h_deblock[k] & (MASK_CUR << ij)
+ && (strength[POS_CUR] == 2 || strength[POS_TOP] == 2)){
+ 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 && (strength[POS_CUR] == 2 || strength[POS_LEFT] == 2)){
+ 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, betaY, 1, 1);
+ }
+ }
+ }
+ }
+ }
+ }
+}
+
+/**
* Initialize decoder.
*/
static av_cold int rv40_decode_init(AVCodecContext *avctx)
@@ -299,6 +666,7 @@
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