rv6: change loop filter to a version used in newer binary specifications

[nihav.git] / nihav-realmedia / src / codecs / rv60dsp.rs

diff --git a/nihav-realmedia/src/codecs/rv60dsp.rs b/nihav-realmedia/src/codecs/rv60dsp.rs
index 062557b2895296203f83d22206f2230896571e2e..8995b48196e8e2f7440c23eb2c9c3f352b2dd37b 100644 (file)
--- a/nihav-realmedia/src/codecs/rv60dsp.rs
+++ b/nihav-realmedia/src/codecs/rv60dsp.rs
@@ -170,10 +170,10 @@ fn filter_luma_edge(dst: &mut [u8], mut offset: usize, step: usize, stride: usiz
     let str_q = strength!(diff_p1p0[0] + diff_p1p0[1] + diff_p1p0[2] + diff_p1p0[3], lim2);
     if str_p + str_q > 2 {
         let msum = ((mode1 + mode2 + str_q + str_p) >> 1) as i16;
-        let (maxprod, weak) = if (str_q == 1) || (str_p == 1) { (512, true) } else { (384, false) };
+        let (maxprod, weak) = if (str_q == 1) || (str_p == 1) { (384, true) } else { (256, false) };
         for y in 0..4 {
             let diff_p0q0 = diff!(dst, offset, offset - step);
-            if (diff_p0q0 != 0) && (lim1 * diff_p0q0.abs() < maxprod) {
+            if (diff_p0q0 != 0) && (((lim1 * diff_p0q0.abs()) & !0x7F) <= maxprod) {
                 let diff_q1q2 = diff!(dst, offset - 2 * step, offset - 3 * step);
                 let diff_p1p2 = diff!(dst, offset +     step, offset + 2 * step);
                 let delta = if weak {
@@ -214,10 +214,10 @@ fn filter_chroma_edge(dst: &mut [u8], mut offset: usize, step: usize, stride: us
     let str_p = strength!(diff_p, lim2);
     if str_p + str_q > 2 {
         let msum = ((mode1 + mode2 + str_q + str_p) >> 1) as i16;
-        let (maxprod, weak) = if (str_q == 1) || (str_p == 1) { (512, true) } else { (384, false) };
+        let (maxprod, weak) = if (str_q == 1) || (str_p == 1) { (384, true) } else { (256, false) };
         for _ in 0..2 {
             let diff_pq = diff!(dst, offset, offset - step);
-            if (diff_pq != 0) && (lim1 * diff_pq.abs() < maxprod) {
+            if (diff_pq != 0) && (((lim1 * diff_pq.abs()) & !0x7F) <= maxprod) {
                 let delta = if weak {
                         clip_symm((diff_pq + 1) >> 1, msum >> 1)
                     } else {
@@ -411,7 +411,7 @@ impl RV60DSP {
         blk[off +  0 * step] = ((tm0 + tt0 + 64) >> 7) as i16;
         blk[off +  1 * step] = ((tm4 + tt1 + 64) >> 7) as i16;
         blk[off +  2 * step] = ((tm6 + tt2 + 64) >> 7) as i16;
-        blk[off +  3 * step] = ((tm4 + tt3 + 64) >> 7) as i16;
+        blk[off +  3 * step] = ((tm2 + tt3 + 64) >> 7) as i16;
         blk[off +  4 * step] = ((tm3 + tt4 + 64) >> 7) as i16;
         blk[off +  5 * step] = ((tm7 + tt5 + 64) >> 7) as i16;
         blk[off +  6 * step] = ((tm5 + tt6 + 64) >> 7) as i16;
@@ -581,46 +581,46 @@ impl RV60DSP {
                   xpos: usize, ypos: usize, top_str: &[u8], left_str: &[u8], dblkpos: usize) {
         if xpos > 0 {
             if ypos > 0 {
-                let str_l = left_str[dblkpos - dparams.dblkstride];
-                let str_r = left_str[dblkpos];
-                if (str_l | str_r) != 0 {
+                let str_l = left_str[dblkpos - dparams.dblkstride - 1];
+                let str_r = left_str[dblkpos - dparams.dblkstride];
+                if ((str_l | str_r) & 3) != 0 {
                     self.deblock_edge4_ver(frame, xpos, ypos - 4, str_l, str_r, dparams.deblock_chroma);
                 }
             }
             {
-                let str_l = left_str[dblkpos];
-                let str_r = left_str[dblkpos + dparams.dblkstride];
-                if (str_l | str_r) != 0 {
+                let str_l = left_str[dblkpos - 1];
+                let str_r = left_str[dblkpos];
+                if ((str_l | str_r) & 3) != 0 {
                     self.deblock_edge4_ver(frame, xpos, ypos + 0, str_l, str_r, dparams.deblock_chroma);
                 }
             }
-            if ypos + 4 >= dparams.height {
-                let str_l = left_str[dblkpos + dparams.dblkstride];
-                let str_r = left_str[dblkpos + dparams.dblkstride * 2];
-                if (str_l | str_r) != 0 {
+            if ypos + 8 >= dparams.height {
+                let str_l = left_str[dblkpos + dparams.dblkstride - 1];
+                let str_r = left_str[dblkpos + dparams.dblkstride];
+                if ((str_l | str_r) & 3) != 0 {
                     self.deblock_edge4_ver(frame, xpos, ypos + 4, str_l, str_r, dparams.deblock_chroma);
                 }
             }
         }
         if ypos > 0 {
             if xpos > 0 {
-                let str_t = top_str[dblkpos - 1];
-                let str_d = top_str[dblkpos];
-                if (str_t | str_d) != 0 {
+                let str_t = top_str[dblkpos - dparams.dblkstride - 1];
+                let str_d = top_str[dblkpos - 1];
+                if ((str_t | str_d) & 3) != 0 {
                     self.deblock_edge4_hor(frame, xpos - 4, ypos, str_t, str_d, dparams.deblock_chroma);
                 }
             }
             {
-                let str_t = top_str[dblkpos];
-                let str_d = top_str[dblkpos + 1];
-                if (str_t | str_d) != 0 {
+                let str_t = top_str[dblkpos - dparams.dblkstride];
+                let str_d = top_str[dblkpos];
+                if ((str_t | str_d) & 3) != 0 {
                     self.deblock_edge4_hor(frame, xpos + 0, ypos, str_t, str_d, dparams.deblock_chroma);
                 }
             }
-            if xpos + 4 >= dparams.width {
-                let str_t = top_str[dblkpos + 1];
-                let str_d = top_str[dblkpos + 2];
-                if (str_t | str_d) != 0 {
+            if xpos + 8 >= dparams.width {
+                let str_t = top_str[dblkpos - dparams.dblkstride + 1];
+                let str_d = top_str[dblkpos + 1];
+                if ((str_t | str_d) & 3) != 0 {
                     self.deblock_edge4_hor(frame, xpos + 4, ypos, str_t, str_d, dparams.deblock_chroma);
                 }
             }