When direct rendering is enabled, use hostdata blits to transfer data from

    system memory to video RAM, which should reduce CPU usage especially
    with larger videos. Can be disabled via Option "DMAForXv" if there
    should be any stability issues, but it's been stable for me during a
    week of testing. Based on a patch by Nikolaus Meine
    <meine@tnt.uni-hannover.de>. Probably fix endianness issues in some
    newer XVideo code, untested.
Also use hostdata blits to transfer RENDER image data to video RAM to avoid
    idling the accelerator engine. Increases RENDER performance
    significantly for me.
These changes were only tested on an M9 in a Titanium PowerBook but should
    work with all Radeons where direct rendering is supported.

1 files changed, 219 insertions, 66 deletions

diff --git a/src/radeon_video.c b/src/radeon_video.c
index bdabc316..f21e3af5 100644
--- a/src/radeon_video.c
+++ b/src/radeon_video.c
@@ -1941,23 +1941,96 @@ static struct {
 
 static void
 RADEONCopyData(
+  ScrnInfoPtr pScrn,
   unsigned char *src,
   unsigned char *dst,
   int srcPitch,
   int dstPitch,
   int h,
-  int w
+  int w,
+  int bpp
 ){
-    w <<= 1;
-    while(h--) {
-	memcpy(dst, src, w);
-	src += srcPitch;
-	dst += dstPitch;
+#ifdef XF86DRI
+    RADEONInfoPtr info = RADEONPTR(pScrn);
+
+    if ( info->directRenderingEnabled && info->DMAForXv )
+    {
+	CARD8 *buf;
+	CARD32 bufPitch;
+	unsigned int hpass;
+
+	/* Get the byte-swapping right for big endian systems */
+	if ( bpp == 2 )
+	{
+	    w *= 2;
+	    bpp = 1;
+	}
+
+	while ( buf = RADEONHostDataBlit( pScrn, bpp, w, dstPitch,
+					  &bufPitch, &dst, &h, &hpass ) )
+	{
+	    RADEONHostDataBlitCopyPass( buf, src, hpass, bufPitch, srcPitch );
+	    src += hpass * srcPitch;
+	}
+
+	FLUSH_RING();
+
+	return;
+    }
+    else
+#endif /* XF86DRI */
+    {
+#if X_BYTE_ORDER == X_BIG_ENDIAN
+	unsigned char *RADEONMMIO = info->MMIO;
+	if ( bpp == 2 )
+	{
+	    OUTREG(RADEON_SURFACE_CNTL, info->ModeReg.surface_cntl
+					& ~(RADEON_NONSURF_AP0_SWP_32BPP
+					| RADEON_NONSURF_AP0_SWP_16BPP));
+	}
+	else /* bpp == 4 */
+	{
+	    OUTREG(RADEON_SURFACE_CNTL, (info->ModeReg.surface_cntl
+					| RADEON_NONSURF_AP0_SWP_32BPP)
+					& ~RADEON_NONSURF_AP0_SWP_16BPP);
+	}
+#endif
+
+	w *= 2;
+
+	while (h--)
+	{
+	    memcpy(dst, src, w);
+	    src += srcPitch;
+	    dst += dstPitch;
+	}
+
+#if X_BYTE_ORDER == X_BIG_ENDIAN
+	/* restore byte swapping */
+	OUTREG(RADEON_SURFACE_CNTL, info->ModeReg.surface_cntl);
+#endif
     }
 }
 
+static void RADEON_420_422(
+    unsigned int *d,
+    unsigned char *s1,
+    unsigned char *s2,
+    unsigned char *s3,
+    unsigned int n
+)
+{
+    while ( n ) {
+	*(d++) = s1[0] | (s1[1] << 16) | (s3[0] << 8) | (s2[0] << 24);
+	s1+=2; s2++; s3++;
+	n--;
+    }
+}
+
+
 static void
 RADEONCopyRGB24Data(
+  ScrnInfoPtr pScrn,
   unsigned char *src,
   unsigned char *dst,
   int srcPitch,
@@ -1968,22 +2041,68 @@ RADEONCopyRGB24Data(
     CARD32 *dptr;
     CARD8 *sptr;
     int i,j;
-    
+#ifdef XF86DRI
+    RADEONInfoPtr info = RADEONPTR(pScrn);
 
-    for(j=0;j<h;j++){
-	dptr=(CARD32 *)(dst+j*dstPitch);
-        sptr=src+j*srcPitch;
-    
-    	for(i=w;i>0;i--){
-	dptr[0]=((sptr[0])<<24)|((sptr[1])<<16)|(sptr[2]);
-	dptr++;
-	sptr+=3;
+    if ( info->directRenderingEnabled && info->DMAForXv )
+    {
+	CARD32 bufPitch;
+	unsigned int hpass;
+
+	while ( dptr = ( CARD32* )RADEONHostDataBlit( pScrn, 4, w, dstPitch,
+						      &bufPitch, &dst, &h,
+						      &hpass ) )
+	{
+	    for( j = 0; j < hpass; j++ )
+	    {
+		sptr = src;
+
+		for ( i = 0 ; i < w; i++ )
+		{
+		    *dptr++ = ( ( *sptr++ ) << 24 ) | ( ( *sptr++ ) << 16 ) |
+			      ( *sptr++ );
+		}
+
+		src += hpass * srcPitch;
+		dptr += hpass * bufPitch;
+	    }
+	}
+
+	FLUSH_RING();
+
+	return;
+    }
+    else
+#endif /* XF86DRI */
+    {
+#if X_BYTE_ORDER == X_BIG_ENDIAN
+	unsigned char *RADEONMMIO = info->MMIO;
+	OUTREG(RADEON_SURFACE_CNTL, (info->ModeReg.surface_cntl
+				   | RADEON_NONSURF_AP0_SWP_32BPP)
+				  & ~RADEON_NONSURF_AP0_SWP_16BPP);
+#endif
+
+	for(j=0;j<h;j++){
+	    dptr=(CARD32 *)(dst+j*dstPitch);
+	    sptr=src+j*srcPitch;
+
+	    for(i=w;i>0;i--){
+	      dptr[0]=((sptr[0])<<24)|((sptr[1])<<16)|(sptr[2]);
+	      dptr++;
+	      sptr+=3;
+	    }
 	}
+
+#if X_BYTE_ORDER == X_BIG_ENDIAN
+	/* restore byte swapping */
+	OUTREG(RADEON_SURFACE_CNTL, info->ModeReg.surface_cntl);
+#endif
     }
 }
 
 static void
 RADEONCopyMungedData(
+   ScrnInfoPtr pScrn,
    unsigned char *src1,
    unsigned char *src2,
    unsigned char *src3,
@@ -1994,37 +2113,86 @@ RADEONCopyMungedData(
    int h,
    int w
 ){
-   CARD32 *dst;
-   CARD8 *s1, *s2, *s3;
-   int i, j;
-
-   w >>= 1;
-
-   for(j = 0; j < h; j++) {
-	dst = (pointer)dst1;
-	s1 = src1;  s2 = src2;  s3 = src3;
-	i = w;
-	while(i > 4) {
-	   dst[0] = s1[0] | (s1[1] << 16) | (s3[0] << 8) | (s2[0] << 24);
-	   dst[1] = s1[2] | (s1[3] << 16) | (s3[1] << 8) | (s2[1] << 24);
-	   dst[2] = s1[4] | (s1[5] << 16) | (s3[2] << 8) | (s2[2] << 24);
-	   dst[3] = s1[6] | (s1[7] << 16) | (s3[3] << 8) | (s2[3] << 24);
-	   dst += 4; s2 += 4; s3 += 4; s1 += 8;
-	   i -= 4;
-	}
-	while(i--) {
-	   dst[0] = s1[0] | (s1[1] << 16) | (s3[0] << 8) | (s2[0] << 24);
-	   dst++; s2++; s3++;
-	   s1 += 2;
+#ifdef XF86DRI
+    RADEONInfoPtr info = RADEONPTR(pScrn);
+
+    if ( info->directRenderingEnabled && info->DMAForXv )
+    {
+	CARD8 *buf;
+	CARD32 y = 0, bufPitch;
+	unsigned int hpass;
+
+	while ( buf = RADEONHostDataBlit( pScrn, 4, w/2, dstPitch,
+					  &bufPitch, &dst1, &h, &hpass ) )
+	{
+	    while ( hpass-- )
+	    {
+		RADEON_420_422( (unsigned int *) buf, src1, src2, src3,
+				bufPitch / 4 );
+		src1 += srcPitch;
+		if ( y & 1 )
+		{
+		    src2 += srcPitch2;
+		    src3 += srcPitch2;
+		}
+		buf += bufPitch;
+		y++;
+	    }
 	}
 
-	dst1 += dstPitch;
-	src1 += srcPitch;
-	if(j & 1) {
-	    src2 += srcPitch2;
-	    src3 += srcPitch2;
+	FLUSH_RING();
+    }
+    else
+#endif /* XF86DRI */
+    {
+	CARD32 *dst;
+	CARD8 *s1, *s2, *s3;
+	int i, j;
+
+#if X_BYTE_ORDER == X_BIG_ENDIAN
+	unsigned char *RADEONMMIO = info->MMIO;
+	OUTREG(RADEON_SURFACE_CNTL, (info->ModeReg.surface_cntl
+				   | RADEON_NONSURF_AP0_SWP_32BPP)
+				  & ~RADEON_NONSURF_AP0_SWP_16BPP);
+#endif
+
+	w /= 2;
+
+	for( j = 0; j < h; j++ )
+	{
+	    dst = (pointer)dst1;
+	    s1 = src1;  s2 = src2;  s3 = src3;
+	    i = w;
+	    while( i > 4 )
+	    {
+		dst[0] = s1[0] | (s1[1] << 16) | (s3[0] << 8) | (s2[0] << 24);
+		dst[1] = s1[2] | (s1[3] << 16) | (s3[1] << 8) | (s2[1] << 24);
+		dst[2] = s1[4] | (s1[5] << 16) | (s3[2] << 8) | (s2[2] << 24);
+		dst[3] = s1[6] | (s1[7] << 16) | (s3[3] << 8) | (s2[3] << 24);
+		dst += 4; s2 += 4; s3 += 4; s1 += 8;
+		i -= 4;
+	    }
+	    while( i-- )
+	    {
+		dst[0] = s1[0] | (s1[1] << 16) | (s3[0] << 8) | (s2[0] << 24);
+		dst++; s2++; s3++;
+		s1 += 2;
+	    }
+
+	    dst1 += dstPitch;
+	    src1 += srcPitch;
+	    if( j & 1 )
+	    {
+		src2 += srcPitch2;
+		src3 += srcPitch2;
+	    }
 	}
-   }
+
+#if X_BYTE_ORDER == X_BIG_ENDIAN
+	/* restore byte swapping */
+	OUTREG(RADEON_SURFACE_CNTL, info->ModeReg.surface_cntl);
+#endif
+    }
 }
 
 
@@ -2472,7 +2640,7 @@ RADEONPutImage(
 	break;
    case FOURCC_YV12:
    case FOURCC_I420:
-	dstPitch = ((width << 1) + 15) & ~15;
+	dstPitch = ((width << 1) + 63) & ~63;
 	new_size = ((dstPitch * height) + bpp - 1) / bpp;
 	srcPitch = (width + 3) & ~3;
 	s2offset = srcPitch * height;
@@ -2482,7 +2650,7 @@ RADEONPutImage(
    case FOURCC_UYVY:
    case FOURCC_YUY2:
    default:
-	dstPitch = ((width << 1) + 15) & ~15;
+	dstPitch = ((width << 1) + 63) & ~63;
 	new_size = ((dstPitch * height) + bpp - 1) / bpp;
 	srcPitch = (width << 1);
 	break;
@@ -2521,14 +2689,9 @@ RADEONPutImage(
 	   s3offset = tmp;
 	}
 	nlines = ((((yb + 0xffff) >> 16) + 1) & ~1) - top;
-#if X_BYTE_ORDER == X_BIG_ENDIAN
-	OUTREG(RADEON_SURFACE_CNTL, (info->ModeReg.surface_cntl |
-				     RADEON_NONSURF_AP0_SWP_32BPP)
-				    & ~RADEON_NONSURF_AP0_SWP_16BPP);
-#endif
-	RADEONCopyMungedData(buf + (top * srcPitch) + left, buf + s2offset,
-			     buf + s3offset, dst_start, srcPitch, srcPitch2,
-			     dstPitch, nlines, npixels);
+	RADEONCopyMungedData(pScrn, buf + (top * srcPitch) + left,
+			     buf + s2offset, buf + s3offset, dst_start,
+			     srcPitch, srcPitch2, dstPitch, nlines, npixels);
 	break;
     case FOURCC_RGBT16:
     case FOURCC_RGB16:
@@ -2539,32 +2702,22 @@ RADEONPutImage(
 	buf += (top * srcPitch) + left;
 	nlines = ((yb + 0xffff) >> 16) - top;
 	dst_start += left;
-#if X_BYTE_ORDER == X_BIG_ENDIAN
-	OUTREG(RADEON_SURFACE_CNTL, info->ModeReg.surface_cntl &
-				    ~(RADEON_NONSURF_AP0_SWP_32BPP
-				      | RADEON_NONSURF_AP0_SWP_16BPP));
-#endif
-	RADEONCopyData(buf, dst_start, srcPitch, dstPitch, nlines, npixels);
+	RADEONCopyData(pScrn, buf, dst_start, srcPitch, dstPitch, nlines, npixels, 2);
 	break;
     case FOURCC_RGBA32:
 	buf += (top * srcPitch) + left*4;
 	nlines = ((yb + 0xffff) >> 16) - top;
 	dst_start += left*4;
-	RADEONCopyData(buf, dst_start, srcPitch, dstPitch, nlines, 2*npixels); 
+	RADEONCopyData(pScrn, buf, dst_start, srcPitch, dstPitch, nlines, npixels, 4);
     	break;
     case FOURCC_RGB24:
 	buf += (top * srcPitch) + left*3;
 	nlines = ((yb + 0xffff) >> 16) - top;
 	dst_start += left*4;
-	RADEONCopyRGB24Data(buf, dst_start, srcPitch, dstPitch, nlines, npixels); 
+	RADEONCopyRGB24Data(pScrn, buf, dst_start, srcPitch, dstPitch, nlines, npixels);
     	break;
     }
 
-#if X_BYTE_ORDER == X_BIG_ENDIAN
-    /* restore byte swapping */
-    OUTREG(RADEON_SURFACE_CNTL, info->ModeReg.surface_cntl);
-#endif
-
     /* update cliplist */
     if(!REGION_EQUAL(pScrn->pScreen, &pPriv->clip, clipBoxes))
     {