AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)

Originally committed as revision 12131 to svn://svn.mplayerhq.hu/mplayer/trunk/postproc

3 files changed, 222 insertions, 10 deletions

diff --git a/postproc/swscale.c b/postproc/swscale.c
index 38284a19f5..dd10521582 100644
--- a/postproc/swscale.c
+++ b/postproc/swscale.c
@@ -1041,6 +1041,21 @@ static inline void initFilter(int16_t **outFilter, int16_t **filterPos, int *out
 		if(min>minFilterSize) minFilterSize= min;
 	}
 
+        if (flags & SWS_CPU_CAPS_ALTIVEC) {
+          // we can handle the special case 4,
+          // so we don't want to go to the full 8
+          if (minFilterSize < 5)
+            filterAlign = 4;
+
+          // we really don't want to waste our time
+          // doing useless computation, so fall-back on
+          // the scalar C code for very small filter.
+          // vectorizing is worth it only if you have
+          // decent-sized vector.
+          if (minFilterSize < 3)
+            filterAlign = 1;
+        }
+
 	ASSERT(minFilterSize > 0)
 	filterSize= (minFilterSize +(filterAlign-1)) & (~(filterAlign-1));
 	ASSERT(filterSize > 0)
@@ -1947,7 +1962,10 @@ SwsContext *sws_getContext(int srcW, int srcH, int origSrcFormat, int dstW, int
 
 	/* precalculate horizontal scaler filter coefficients */
 	{
-		const int filterAlign= (flags & SWS_CPU_CAPS_MMX) ? 4 : 1;
+		const int filterAlign=
+		  (flags & SWS_CPU_CAPS_MMX) ? 4 :
+		  (flags & SWS_CPU_CAPS_ALTIVEC) ? 8 :
+		  1;
 
 		initFilter(&c->hLumFilter, &c->hLumFilterPos, &c->hLumFilterSize, c->lumXInc,
 				 srcW      ,       dstW, filterAlign, 1<<14,
@@ -1976,14 +1994,20 @@ SwsContext *sws_getContext(int srcW, int srcH, int origSrcFormat, int dstW, int
 
 
 	/* precalculate vertical scaler filter coefficients */
-	initFilter(&c->vLumFilter, &c->vLumFilterPos, &c->vLumFilterSize, c->lumYInc,
-			srcH      ,        dstH, 1, (1<<12)-4,
-			(flags&SWS_BICUBLIN) ? (flags|SWS_BICUBIC)  : flags,
-			srcFilter->lumV, dstFilter->lumV);
-	initFilter(&c->vChrFilter, &c->vChrFilterPos, &c->vChrFilterSize, c->chrYInc,
-			c->chrSrcH, c->chrDstH, 1, (1<<12)-4,
-			(flags&SWS_BICUBLIN) ? (flags|SWS_BILINEAR) : flags,
-			srcFilter->chrV, dstFilter->chrV);
+	{
+		const int filterAlign=
+		  (flags & SWS_CPU_CAPS_ALTIVEC) ? 8 :
+		  1;
+
+		initFilter(&c->vLumFilter, &c->vLumFilterPos, &c->vLumFilterSize, c->lumYInc,
+				srcH      ,        dstH, filterAlign, (1<<12)-4,
+				(flags&SWS_BICUBLIN) ? (flags|SWS_BICUBIC)  : flags,
+				srcFilter->lumV, dstFilter->lumV);
+		initFilter(&c->vChrFilter, &c->vChrFilterPos, &c->vChrFilterSize, c->chrYInc,
+				c->chrSrcH, c->chrDstH, filterAlign, (1<<12)-4,
+				(flags&SWS_BICUBLIN) ? (flags|SWS_BILINEAR) : flags,
+				srcFilter->chrV, dstFilter->chrV);
+	}
 
 	// Calculate Buffer Sizes so that they won't run out while handling these damn slices
 	c->vLumBufSize= c->vLumFilterSize;
diff --git a/postproc/swscale_altivec_template.c b/postproc/swscale_altivec_template.c
index 0f6a3cda34..5cd70683a5 100644
--- a/postproc/swscale_altivec_template.c
+++ b/postproc/swscale_altivec_template.c
@@ -20,7 +20,19 @@
     Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 */
 
-static const vector unsigned int altivec_vectorShiftInt19 = {19, 19, 19, 19};
+#ifdef CONFIG_DARWIN
+static const vector signed int vzero =
+  (vector signed int)(0, 0, 0, 0);
+static const vector unsigned int altivec_vectorShiftInt19 =
+  (vector unsigned int)(19, 19, 19, 19);
+#else
+static const vector signed int vzero =
+  (vector signed int){0,0,0,0};
+static const vector unsigned int altivec_vectorShiftInt19 =
+  (vector unsigned int){19, 19, 19, 19};
+
+#endif
+
 static inline void
 altivec_packIntArrayToCharArray(int *val, uint8_t* dest, int dstW) {
   register int i;
@@ -201,3 +213,175 @@ yuv2yuvX_altivec_real(int16_t *lumFilter, int16_t **lumSrc, int lumFilterSize,
     altivec_packIntArrayToCharArray(v,vDest,chrDstW);
   }
 }
+
+static inline void hScale_altivec_real(int16_t *dst, int dstW, uint8_t *src, int srcW, int xInc, int16_t *filter, int16_t *filterPos, int filterSize) {
+  register int i;
+  int __attribute__ ((aligned (16))) tempo[4];
+
+  if (filterSize % 4) {
+    for(i=0; i<dstW; i++) {
+      register int j;
+      register int srcPos = filterPos[i];
+      register int val = 0;
+      for(j=0; j<filterSize; j++) {
+	val += ((int)src[srcPos + j])*filter[filterSize*i + j];
+      }
+      dst[i] = MIN(MAX(0, val>>7), (1<<15)-1);
+    }
+  }
+  else
+  switch (filterSize) {
+  case 4:
+    {
+      for(i=0; i<dstW; i++) {
+	register int j;
+	register int srcPos = filterPos[i];
+
+	vector unsigned char src_v0 = vec_ld(srcPos, src);
+	vector unsigned char src_v1;
+	if ((((int)src + srcPos)% 16) > 12) {
+	  src_v1 = vec_ld(srcPos + 16, src);
+	}
+	vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
+
+	vector signed short src_v = // vec_unpackh sign-extends...
+	  (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
+	// now put our elements in the even slots
+	src_v = vec_mergeh(src_v, (vector signed short)vzero);
+
+	vector signed short filter_v = vec_ld(i << 3, filter);
+        // the 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2)
+
+        // the neat trick : we only care for half the elements,
+        // high or low depending on (i<<3)%16 (it's 0 or 8 here),
+        // and we're going to use vec_mule, so we chose
+        // carefully how to "unpack" the elements into the even slots
+	if ((i << 3) % 16)
+	  filter_v = vec_mergel(filter_v,(vector signed short)vzero);
+	else
+	  filter_v = vec_mergeh(filter_v,(vector signed short)vzero);
+
+	vector signed int val_vEven = vec_mule(src_v, filter_v);
+	vector signed int val_s = vec_sums(val_vEven, vzero);
+	vec_st(val_s, 0, tempo);
+	dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
+      }
+    }
+    break;
+
+  case 8:
+    {
+      for(i=0; i<dstW; i++) {
+	register int srcPos = filterPos[i];
+
+	vector unsigned char src_v0 = vec_ld(srcPos, src);
+	vector unsigned char src_v1;
+	if ((((int)src + srcPos)% 16) > 8) {
+	  src_v1 = vec_ld(srcPos + 16, src);
+	}
+	vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
+
+	vector signed short src_v = // vec_unpackh sign-extends...
+	  (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
+	vector signed short filter_v = vec_ld(i << 4, filter);
+        // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2)
+
+	vector signed int val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
+	vector signed int val_s = vec_sums(val_v, vzero);
+	vec_st(val_s, 0, tempo);
+	dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
+      }
+    }
+    break;
+
+  case 16:
+    {
+      for(i=0; i<dstW; i++) {
+	register int srcPos = filterPos[i];
+
+	vector unsigned char src_v0 = vec_ld(srcPos, src);
+	vector unsigned char src_v1 = vec_ld(srcPos + 16, src);
+	vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
+
+	vector signed short src_vA = // vec_unpackh sign-extends...
+	  (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
+	vector signed short src_vB = // vec_unpackh sign-extends...
+	  (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
+
+	vector signed short filter_v0 = vec_ld(i << 5, filter);
+       	vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
+        // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2)
+
+	vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
+	vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc);
+
+	vector signed int val_s = vec_sums(val_v, vzero);
+
+	vec_st(val_s, 0, tempo);
+	dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
+      }
+    }
+    break;
+    
+  default:
+    {
+      for(i=0; i<dstW; i++) {
+	register int j;
+	register int srcPos = filterPos[i];
+
+        vector signed int val_v = (vector signed int)vzero;
+	vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter);
+        vector unsigned char permF = vec_lvsl((i * 2 * filterSize), filter);
+
+        vector unsigned char src_v0 = vec_ld(srcPos, src);
+        vector unsigned char permS = vec_lvsl(srcPos, src);
+
+        for (j = 0 ; j < filterSize - 15; j += 16) {
+          vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src);
+          vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
+          
+          vector signed short src_vA = // vec_unpackh sign-extends...
+            (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
+          vector signed short src_vB = // vec_unpackh sign-extends...
+            (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
+          
+          vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
+          vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter);
+          vector signed short filter_v0 = vec_perm(filter_v0R, filter_v1R, permF);
+          vector signed short filter_v1 = vec_perm(filter_v1R, filter_v2R, permF);
+          
+          vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
+          val_v = vec_msums(src_vB, filter_v1, val_acc);
+
+          filter_v0R = filter_v2R;
+          src_v0 = src_v1;
+        }
+
+        if (j < (filterSize-7)) {
+          // loading src_v0 is useless, it's already done above
+          //vector unsigned char src_v0 = vec_ld(srcPos + j, src);
+          vector unsigned char src_v1;
+          if ((((int)src + srcPos)% 16) > 8) {
+            src_v1 = vec_ld(srcPos + j + 16, src);
+          }
+          vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
+          
+          vector signed short src_v = // vec_unpackh sign-extends...
+            (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
+          // loading filter_v0R is useless, it's already done above
+          //vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter);
+          vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
+          vector signed short filter_v = vec_perm(filter_v0R, filter_v1R, permF);
+          
+          val_v = vec_msums(src_v, filter_v, val_v);
+        }
+
+        vector signed int val_s = vec_sums(val_v, vzero);
+          
+        vec_st(val_s, 0, tempo);
+        dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);        
+      }
+      
+    }
+  }
+}
diff --git a/postproc/swscale_template.c b/postproc/swscale_template.c
index 6a8117f2fb..2d266f5655 100644
--- a/postproc/swscale_template.c
+++ b/postproc/swscale_template.c
@@ -2154,6 +2154,9 @@ static inline void RENAME(hScale)(int16_t *dst, int dstW, uint8_t *src, int srcW
 		);
 	}
 #else
+#ifdef HAVE_ALTIVEC
+	hScale_altivec_real(dst, dstW, src, srcW, xInc, filter, filterPos, filterSize);
+#else
 	int i;
 	for(i=0; i<dstW; i++)
 	{
@@ -2171,6 +2174,7 @@ static inline void RENAME(hScale)(int16_t *dst, int dstW, uint8_t *src, int srcW
 //		dst[i] = val>>7;
 	}
 #endif
+#endif
 }
       // *** horizontal scale Y line to temp buffer
 static inline void RENAME(hyscale)(uint16_t *dst, int dstWidth, uint8_t *src, int srcW, int xInc,