* altivec and pix_norm patch by Brian Foley

Originally committed as revision 1269 to svn://svn.ffmpeg.org/ffmpeg/trunk

6 files changed, 79 insertions, 51 deletions

diff --git a/libavcodec/dsputil.c b/libavcodec/dsputil.c
index a8e060b66e..9039d2625f 100644
--- a/libavcodec/dsputil.c
+++ b/libavcodec/dsputil.c
@@ -144,6 +144,31 @@ static int pix_norm1_c(UINT8 * pix, int line_size)
 }
 
 
+static int pix_norm_c(UINT8 * pix1, UINT8 * pix2, int line_size)
+{
+    int s, i, j;
+    UINT32 *sq = squareTbl + 256;
+
+    s = 0;
+    for (i = 0; i < 16; i++) {
+        for (j = 0; j < 16; j += 8) {
+            s += sq[pix1[0] - pix2[0]];
+            s += sq[pix1[1] - pix2[1]];
+            s += sq[pix1[2] - pix2[2]];
+            s += sq[pix1[3] - pix2[3]];
+            s += sq[pix1[4] - pix2[4]];
+            s += sq[pix1[5] - pix2[5]];
+            s += sq[pix1[6] - pix2[6]];
+            s += sq[pix1[7] - pix2[7]];
+            pix1 += 8;
+            pix2 += 8;
+        }
+        pix1 += line_size - 16;
+        pix2 += line_size - 16;
+    }
+    return s;
+}
+
 static void get_pixels_c(DCTELEM *restrict block, const UINT8 *pixels, int line_size)
 {
     int i;
@@ -1404,6 +1429,7 @@ void dsputil_init(DSPContext* c, unsigned mask)
     c->clear_blocks = clear_blocks_c;
     c->pix_sum = pix_sum_c;
     c->pix_norm1 = pix_norm1_c;
+    c->pix_norm = pix_norm_c;
 
     /* TODO [0] 16  [1] 8 */
     c->pix_abs16x16     = pix_abs16x16_c;
diff --git a/libavcodec/dsputil.h b/libavcodec/dsputil.h
index 4f036b05d6..0b1d34f908 100644
--- a/libavcodec/dsputil.h
+++ b/libavcodec/dsputil.h
@@ -98,6 +98,7 @@ typedef struct DSPContext {
     void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
     int (*pix_sum)(UINT8 * pix, int line_size);
     int (*pix_norm1)(UINT8 * pix, int line_size);
+    int (*pix_norm)(UINT8 * pix1, UINT8 * pix2, int line_size);
 
     /* maybe create an array for 16/8 functions */
     op_pixels_func put_pixels_tab[2][4];
diff --git a/libavcodec/motion_est.c b/libavcodec/motion_est.c
index 95e67f66a5..b1a3c61f73 100644
--- a/libavcodec/motion_est.c
+++ b/libavcodec/motion_est.c
@@ -63,31 +63,6 @@ static int pix_dev(UINT8 * pix, int line_size, int mean)
     return s;
 }
 
-static int pix_norm(UINT8 * pix1, UINT8 * pix2, int line_size)
-{
-    int s, i, j;
-    UINT32 *sq = squareTbl + 256;
-
-    s = 0;
-    for (i = 0; i < 16; i++) {
-	for (j = 0; j < 16; j += 8) {
-	    s += sq[pix1[0] - pix2[0]];
-	    s += sq[pix1[1] - pix2[1]];
-	    s += sq[pix1[2] - pix2[2]];
-	    s += sq[pix1[3] - pix2[3]];
-	    s += sq[pix1[4] - pix2[4]];
-	    s += sq[pix1[5] - pix2[5]];
-	    s += sq[pix1[6] - pix2[6]];
-	    s += sq[pix1[7] - pix2[7]];
-	    pix1 += 8;
-	    pix2 += 8;
-	}
-	pix1 += line_size - 16;
-	pix2 += line_size - 16;
-    }
-    return s;
-}
-
 static inline void no_motion_search(MpegEncContext * s,
 				    int *mx_ptr, int *my_ptr)
 {
@@ -1137,7 +1112,7 @@ void ff_estimate_p_frame_motion(MpegEncContext * s,
     
     varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
     // FIXME: MMX OPTIMIZE
-    vard = (pix_norm(pix, ppix, s->linesize)+128)>>8;
+    vard = (s->dsp.pix_norm(pix, ppix, s->linesize)+128)>>8;
 
 //printf("%d %d %d %X %X %X\n", s->mb_width, mb_x, mb_y,(int)s, (int)s->mb_var, (int)s->mc_mb_var); fflush(stdout);
     s->mb_var   [s->mb_width * mb_y + mb_x] = varc;
diff --git a/libavcodec/ppc/dsputil_altivec.c b/libavcodec/ppc/dsputil_altivec.c
index 1e877d550d..ed34a2d92b 100644
--- a/libavcodec/ppc/dsputil_altivec.c
+++ b/libavcodec/ppc/dsputil_altivec.c
@@ -137,13 +137,11 @@ int pix_abs16x16_xy2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
     vector unsigned char pix1v, pix2v, pix3v, pix2iv, pix3iv;
     vector unsigned short pix2lv, pix2hv, pix2ilv, pix2ihv;
     vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;
-    vector unsigned short avghv, avglv, two, shift_mask;
+    vector unsigned short avghv, avglv, two;
     vector unsigned short t1, t2, t3, t4;
     vector unsigned int sad;
     vector signed int sumdiffs;
 
-    shift_mask = (vector unsigned short) (0x3fff, 0x3fff, 0x3fff, 0x3fff,
-                                          0x3fff, 0x3fff, 0x3fff, 0x3fff);
     zero = vec_splat_u8(0);
     two = vec_splat_u16(2);
     sad = vec_splat_u32(0);
@@ -205,11 +203,8 @@ int pix_abs16x16_xy2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
         t3 = vec_add(pix3hv, pix3ihv);
         t4 = vec_add(pix3lv, pix3ilv);
 
-        avghv = vec_add(vec_add(t1, t3), two);
-        avghv= vec_and(vec_srl(avghv, two), shift_mask);
-
-        avglv = vec_add(vec_add(t2, t4), two);
-        avglv = vec_and(vec_srl(avglv, two), shift_mask);
+        avghv = vec_sr(vec_add(vec_add(t1, t3), two), two);
+        avglv = vec_sr(vec_add(vec_add(t2, t4), two), two);
 
         /* Pack the shorts back into a result */
         avgv = vec_pack(avghv, avglv);
@@ -323,17 +318,10 @@ int pix_norm1_altivec(uint8_t *pix, int line_size)
     int s, i;
     vector unsigned char *tv, zero;
     vector unsigned char pixv;
-    vector unsigned short pixlv, pixhv, zeros;
     vector unsigned int sv;
     vector signed int sum;
-    vector unsigned char perm_stoint_h = (vector unsigned char)
-        (16, 16, 0, 1, 16, 16, 2, 3, 16, 16, 4, 5, 16, 16, 6, 7);
-    
-    vector unsigned char perm_stoint_l = (vector unsigned char)
-        (16, 16, 8, 9, 16, 16, 10, 11, 16, 16, 12, 13, 16, 16, 14, 15);
         
     zero = vec_splat_u8(0);
-    zeros = vec_splat_u16(0);
     sv = vec_splat_u32(0);
     
     s = 0;
@@ -342,14 +330,8 @@ int pix_norm1_altivec(uint8_t *pix, int line_size)
         tv = (vector unsigned char *) pix;
         pixv = vec_perm(tv[0], tv[1], vec_lvsl(0, pix));
 
-        /* Split them into two vectors of shorts */
-        pixhv = (vector unsigned short) vec_mergeh(zero, pixv);
-        pixlv = (vector unsigned short) vec_mergel(zero, pixv);
-
-        
-        /* Square the values and add them to our sum */
-        sv = vec_msum(pixhv, pixhv, sv);
-        sv = vec_msum(pixlv, pixlv, sv);
+        /* Square the values, and add them to our sum */
+        sv = vec_msum(pixv, pixv, sv);
 
         pix += line_size;
     }
@@ -361,6 +343,48 @@ int pix_norm1_altivec(uint8_t *pix, int line_size)
     return s;
 }
 
+
+int pix_norm_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
+{
+    int s, i;
+    vector unsigned char *tv, zero;
+    vector unsigned char pix1v, pix2v, t5;
+    vector unsigned int sv;
+    vector signed int sum;
+
+    zero = vec_splat_u8(0);
+    sv = vec_splat_u32(0);
+    s = 0;
+    for (i = 0; i < 16; i++) {
+        /* Read in the potentially unaligned pixels */
+        tv = (vector unsigned char *) pix1;
+        pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
+
+        tv = (vector unsigned char *) pix2;
+        pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix2));
+
+        /*
+           Since we want to use unsigned chars, we can take advantage
+           of the fact that abs(a-b)^2 = (a-b)^2.
+        */
+        
+        /* Calculate a sum of abs differences vector */
+        t5 = vec_sub(vec_max(pix1v, pix2v), vec_min(pix1v, pix2v));
+
+        /* Square the values and add them to our sum */
+        sv = vec_msum(t5, t5, sv);
+        
+        pix1 += line_size;
+        pix2 += line_size;
+    }
+    /* Sum up the four partial sums, and put the result into s */
+    sum = vec_sums((vector signed int) sv, (vector signed int) zero);
+    sum = vec_splat(sum, 3);
+    vec_ste(sum, 0, &s);
+    return s;
+}
+
+
 int pix_sum_altivec(UINT8 * pix, int line_size)
 {
 
diff --git a/libavcodec/ppc/dsputil_altivec.h b/libavcodec/ppc/dsputil_altivec.h
index d4d259d9ed..94fe3a0238 100644
--- a/libavcodec/ppc/dsputil_altivec.h
+++ b/libavcodec/ppc/dsputil_altivec.h
@@ -23,6 +23,7 @@ extern int pix_abs16x16_xy2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 extern int pix_abs16x16_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
 extern int pix_abs8x8_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
 extern int pix_norm1_altivec(uint8_t *pix, int line_size);
+extern int pix_norm_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
 extern int pix_sum_altivec(UINT8 * pix, int line_size);
 extern void diff_pixels_altivec(DCTELEM* block, const UINT8* s1, const UINT8* s2, int stride);
 extern void get_pixels_altivec(DCTELEM* block, const UINT8 * pixels, int line_size);
diff --git a/libavcodec/ppc/dsputil_ppc.c b/libavcodec/ppc/dsputil_ppc.c
index a515cf3685..c4dae6cb3b 100644
--- a/libavcodec/ppc/dsputil_ppc.c
+++ b/libavcodec/ppc/dsputil_ppc.c
@@ -34,10 +34,11 @@ void dsputil_init_ppc(DSPContext* c, unsigned mask)
         // Altivec specific optimisations
         c->pix_abs16x16_x2 = pix_abs16x16_x2_altivec;
         c->pix_abs16x16_y2 = pix_abs16x16_y2_altivec;
-	c->pix_abs16x16_xy2 = pix_abs16x16_xy2_altivec;
-	c->pix_abs16x16 = pix_abs16x16_altivec;
+        c->pix_abs16x16_xy2 = pix_abs16x16_xy2_altivec;
+        c->pix_abs16x16 = pix_abs16x16_altivec;
         c->pix_abs8x8 = pix_abs8x8_altivec;
         c->pix_norm1 = pix_norm1_altivec;
+        c->pix_norm = pix_norm_altivec;
         c->pix_sum = pix_sum_altivec;
         c->diff_pixels = diff_pixels_altivec;
         c->get_pixels = get_pixels_altivec;