sync to ffmpeg build 4699

CVS patchset: 6090
CVS date: 2004/02/01 05:31:16

3 files changed, 79 insertions, 36 deletions

diff --git a/src/libffmpeg/libavcodec/alpha/dsputil_alpha.c b/src/libffmpeg/libavcodec/alpha/dsputil_alpha.c
index 82ff7db66..496f46120 100644
--- a/src/libffmpeg/libavcodec/alpha/dsputil_alpha.c
+++ b/src/libffmpeg/libavcodec/alpha/dsputil_alpha.c
@@ -39,11 +39,11 @@ void get_pixels_mvi(DCTELEM *restrict block,
                     const uint8_t *restrict pixels, int line_size);
 void diff_pixels_mvi(DCTELEM *block, const uint8_t *s1, const uint8_t *s2,
                      int stride);
-int pix_abs8x8_mvi(uint8_t *pix1, uint8_t *pix2, int line_size);
+int pix_abs8x8_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h);
 int pix_abs16x16_mvi_asm(uint8_t *pix1, uint8_t *pix2, int line_size);
-int pix_abs16x16_x2_mvi(uint8_t *pix1, uint8_t *pix2, int line_size);
-int pix_abs16x16_y2_mvi(uint8_t *pix1, uint8_t *pix2, int line_size);
-int pix_abs16x16_xy2_mvi(uint8_t *pix1, uint8_t *pix2, int line_size);
+int pix_abs16x16_x2_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h);
+int pix_abs16x16_y2_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h);
+int pix_abs16x16_xy2_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h);
 
 #if 0
 /* These functions were the base for the optimized assembler routines,
@@ -290,11 +290,6 @@ static int sad16x16_mvi(void *s, uint8_t *a, uint8_t *b, int stride)
     return pix_abs16x16_mvi_asm(a, b, stride);
 }
 
-static int sad8x8_mvi(void *s, uint8_t *a, uint8_t *b, int stride)
-{
-    return pix_abs8x8_mvi(a, b, stride);
-}
-
 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx)
 {
     c->put_pixels_tab[0][0] = put_pixels16_axp_asm;
@@ -347,12 +342,13 @@ void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx)
         c->get_pixels       = get_pixels_mvi;
         c->diff_pixels      = diff_pixels_mvi;
         c->sad[0]           = sad16x16_mvi;
-        c->sad[1]           = sad8x8_mvi;
-        c->pix_abs8x8       = pix_abs8x8_mvi;
-        c->pix_abs16x16     = pix_abs16x16_mvi_asm;
-        c->pix_abs16x16_x2  = pix_abs16x16_x2_mvi;
-        c->pix_abs16x16_y2  = pix_abs16x16_y2_mvi;
-        c->pix_abs16x16_xy2 = pix_abs16x16_xy2_mvi;
+        c->sad[1]           = pix_abs8x8_mvi;
+//        c->pix_abs[0][0]    = pix_abs16x16_mvi_asm; //FIXME function arguments for the asm must be fixed
+        c->pix_abs[0][0]    = sad16x16_mvi;
+        c->pix_abs[1][0]    = pix_abs8x8_mvi;
+        c->pix_abs[0][1]    = pix_abs16x16_x2_mvi;
+        c->pix_abs[0][2]    = pix_abs16x16_y2_mvi;
+        c->pix_abs[0][3]    = pix_abs16x16_xy2_mvi;
     }
 
     put_pixels_clamped_axp_p = c->put_pixels_clamped;
diff --git a/src/libffmpeg/libavcodec/alpha/motion_est_alpha.c b/src/libffmpeg/libavcodec/alpha/motion_est_alpha.c
index 804e1d2b6..8b8a0a25c 100644
--- a/src/libffmpeg/libavcodec/alpha/motion_est_alpha.c
+++ b/src/libffmpeg/libavcodec/alpha/motion_est_alpha.c
@@ -84,10 +84,9 @@ static inline uint64_t avg4(uint64_t l1, uint64_t l2, uint64_t l3, uint64_t l4)
     return r1 + r2;
 }
 
-int pix_abs8x8_mvi(uint8_t *pix1, uint8_t *pix2, int line_size)
+int pix_abs8x8_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 {
     int result = 0;
-    int h = 8;
 
     if ((size_t) pix2 & 0x7) {
         /* works only when pix2 is actually unaligned */
@@ -160,10 +159,9 @@ int pix_abs16x16_mvi(uint8_t *pix1, uint8_t *pix2, int line_size)
 }
 #endif
 
-int pix_abs16x16_x2_mvi(uint8_t *pix1, uint8_t *pix2, int line_size)
+int pix_abs16x16_x2_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 {
     int result = 0;
-    int h = 16;
     uint64_t disalign = (size_t) pix2 & 0x7;
 
     switch (disalign) {
@@ -234,10 +232,9 @@ int pix_abs16x16_x2_mvi(uint8_t *pix1, uint8_t *pix2, int line_size)
     return result;
 }
 
-int pix_abs16x16_y2_mvi(uint8_t *pix1, uint8_t *pix2, int line_size)
+int pix_abs16x16_y2_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 {
     int result = 0;
-    int h = 16;
 
     if ((size_t) pix2 & 0x7) {
         uint64_t t, p2_l, p2_r;
@@ -288,10 +285,9 @@ int pix_abs16x16_y2_mvi(uint8_t *pix1, uint8_t *pix2, int line_size)
     return result;
 }
 
-int pix_abs16x16_xy2_mvi(uint8_t *pix1, uint8_t *pix2, int line_size)
+int pix_abs16x16_xy2_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 {
     int result = 0;
-    int h = 16;
     
     uint64_t p1_l, p1_r;
     uint64_t p2_l, p2_r, p2_x;
diff --git a/src/libffmpeg/libavcodec/alpha/mpegvideo_alpha.c b/src/libffmpeg/libavcodec/alpha/mpegvideo_alpha.c
index 6b720373c..f64fb7472 100644
--- a/src/libffmpeg/libavcodec/alpha/mpegvideo_alpha.c
+++ b/src/libffmpeg/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;
 }