100l (forgot alpha & ppc)

Originally committed as revision 2611 to svn://svn.ffmpeg.org/ffmpeg/trunk

1 files changed, 64 insertions, 13 deletions

diff --git a/libavcodec/alpha/mpegvideo_alpha.c b/libavcodec/alpha/mpegvideo_alpha.c
index 6b720373c2..f64fb7472a 100644
--- a/libavcodec/alpha/mpegvideo_alpha.c
+++ b/libavcodec/alpha/mpegvideo_alpha.c
@@ -21,7 +21,7 @@
 #include "../dsputil.h"
 #include "../mpegvideo.h"
 
-static void dct_unquantize_h263_axp(MpegEncContext *s, DCTELEM *block,
+static void dct_unquantize_h263_intra_axp(MpegEncContext *s, DCTELEM *block,
                                     int n, int qscale)
 {
     int i, n_coeffs;
@@ -35,19 +35,15 @@ static void dct_unquantize_h263_axp(MpegEncContext *s, DCTELEM *block,
     /* This mask kills spill from negative subwords to the next subword.  */ 
     correction = WORD_VEC((qmul - 1) + 1); /* multiplication / addition */
 
-    if (s->mb_intra) {
-        if (!s->h263_aic) {
-            if (n < 4) 
-                block0 = block[0] * s->y_dc_scale;
-            else
-                block0 = block[0] * s->c_dc_scale;
-        } else {
-	    qadd = 0;
-	}
-        n_coeffs = 63; // does not always use zigzag table 
+    if (!s->h263_aic) {
+        if (n < 4) 
+            block0 = block[0] * s->y_dc_scale;
+        else
+            block0 = block[0] * s->c_dc_scale;
     } else {
-        n_coeffs = s->intra_scantable.raster_end[s->block_last_index[n]];
+        qadd = 0;
     }
+    n_coeffs = 63; // does not always use zigzag table 
 
     for(i = 0; i <= n_coeffs; block += 4, i += 4) {
         uint64_t levels, negmask, zeros, add;
@@ -90,7 +86,62 @@ static void dct_unquantize_h263_axp(MpegEncContext *s, DCTELEM *block,
         orig_block[0] = block0;
 }
 
+static void dct_unquantize_h263_inter_axp(MpegEncContext *s, DCTELEM *block,
+                                    int n, int qscale)
+{
+    int i, n_coeffs;
+    uint64_t qmul, qadd;
+    uint64_t correction;
+    DCTELEM *orig_block = block;
+    DCTELEM block0;
+
+    qadd = WORD_VEC((qscale - 1) | 1);
+    qmul = qscale << 1;
+    /* This mask kills spill from negative subwords to the next subword.  */ 
+    correction = WORD_VEC((qmul - 1) + 1); /* multiplication / addition */
+
+    n_coeffs = s->intra_scantable.raster_end[s->block_last_index[n]];
+
+    for(i = 0; i <= n_coeffs; block += 4, i += 4) {
+        uint64_t levels, negmask, zeros, add;
+
+        levels = ldq(block);
+        if (levels == 0)
+            continue;
+
+#ifdef __alpha_max__
+        /* I don't think the speed difference justifies runtime
+           detection.  */
+        negmask = maxsw4(levels, -1); /* negative -> ffff (-1) */
+        negmask = minsw4(negmask, 0); /* positive -> 0000 (0) */
+#else
+        negmask = cmpbge(WORD_VEC(0x7fff), levels);
+        negmask &= (negmask >> 1) | (1 << 7);
+        negmask = zap(-1, negmask);
+#endif
+
+        zeros = cmpbge(0, levels);
+        zeros &= zeros >> 1;
+        /* zeros |= zeros << 1 is not needed since qadd <= 255, so
+           zapping the lower byte suffices.  */
+
+        levels *= qmul;
+        levels -= correction & (negmask << 16);
+
+        /* Negate qadd for negative levels.  */
+        add = qadd ^ negmask;
+        add += WORD_VEC(0x0001) & negmask;
+        /* Set qadd to 0 for levels == 0.  */
+        add = zap(add, zeros);
+
+        levels += add;
+
+        stq(levels, block);
+    }
+}
+
 void MPV_common_init_axp(MpegEncContext *s)
 {
-    s->dct_unquantize_h263 = dct_unquantize_h263_axp;
+    s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_axp;
+    s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_axp;
 }