diff --git a/src/aticonfig.c b/src/aticonfig.c
index 9823652..9630559 100644
--- a/src/aticonfig.c
+++ b/src/aticonfig.c
@@ -147,6 +147,7 @@ #   define ShadowFB      PublicOption[AT
 #   define SWCursor      PublicOption[ATI_OPTION_SWCURSOR].value.bool
 #   define AccelMethod   PublicOption[ATI_OPTION_ACCELMETHOD].value.str
 #   define RenderAccel   PublicOption[ATI_OPTION_RENDER_ACCEL].value.bool
+#   define UpDownAccel   PublicOption[ATI_OPTION_UPDOWN_ACCEL].value.bool
 #   define LCDSync       PrivateOption[ATI_OPTION_LCDSYNC].value.bool
 
 #   define ReferenceClock \
@@ -314,6 +315,10 @@ #if defined(USE_EXA)
         pATI->RenderAccelEnabled = FALSE;
         if (pATI->useEXA && RenderAccel)
             pATI->RenderAccelEnabled = TRUE;
+
+        pATI->UpDownAccelEnabled = FALSE;
+        if (pATI->useEXA && UpDownAccel)
+            pATI->UpDownAccelEnabled = TRUE;
 #endif
     }
 
diff --git a/src/atidri.c b/src/atidri.c
index 2b80859..7c0e64c 100644
--- a/src/atidri.c
+++ b/src/atidri.c
@@ -814,7 +814,7 @@ static Bool ATIDRIAgpInit( ScreenPtr pSc
    pATIDRIServer->agpSize = ATI_DEFAULT_AGP_SIZE;
    pATIDRIServer->agpMode = ATI_DEFAULT_AGP_MODE;
    pATIDRIServer->bufferSize = ATI_DEFAULT_BUFFER_SIZE;
-   pATIDRIServer->ringSize = 16; /* 16 kB ring */
+   pATIDRIServer->ringSize = 32; /* 32 kB ring */
 
    if ( drmAgpAcquire( pATI->drmFD ) < 0 ) {
       xf86DrvMsg( pScreen->myNum, X_WARNING, "[agp] AGP not available\n" );
@@ -978,7 +978,7 @@ static Bool ATIDRIPciInit( ScreenPtr pSc
    ATIDRIServerInfoPtr pATIDRIServer = pATI->pDRIServerInfo;
 
    pATIDRIServer->bufferSize = ATI_DEFAULT_BUFFER_SIZE;
-   pATIDRIServer->ringSize = 16; /* 16 kB ring */
+   pATIDRIServer->ringSize = 32; /* 32 kB ring */
 
    if ( !ATIDRISetBufSize( pScreen, (unsigned)(-1) ) )
       return FALSE;
diff --git a/src/atilock.c b/src/atilock.c
index fd93128..83ca948 100644
--- a/src/atilock.c
+++ b/src/atilock.c
@@ -82,8 +82,15 @@ #endif /* AVOID_CPIO */
 #ifdef XF86DRI_DEVEL
 
 	if (pATI->irq > 0)
+	{
+	    /* Enable VBLANK interrupt - handled by DRM */
 	    outr(CRTC_INT_CNTL, (inr(CRTC_INT_CNTL) & ~CRTC_INT_ACKS) | 
-		 CRTC_VBLANK_INT_EN); /* Enable VBLANK interrupt - handled by DRM */
+		 CRTC_VBLANK_INT_EN);
+
+	    /* Enable BM_EOL interrupt - handled by DRM */
+	    outr(CRTC_INT_CNTL, (inr(CRTC_INT_CNTL) & ~CRTC_INT_ACKS) | 
+		 CRTC_BUSMASTER_EOL_INT_EN);
+	}
 
 #endif /* XF86DRI_DEVEL */
 
diff --git a/src/atimach64exa.c b/src/atimach64exa.c
index 6a9c473..dbc9833 100644
--- a/src/atimach64exa.c
+++ b/src/atimach64exa.c
@@ -69,6 +69,7 @@ #include "atiregs.h"
 #ifdef XF86DRI_DEVEL
 #include "mach64_dri.h"
 #include "mach64_sarea.h"
+#include "mach64_common.h"
 #endif
 
 #ifdef USE_EXA
@@ -388,19 +389,73 @@ (
 static void Mach64DoneSolid(PixmapPtr pPixmap) { }
 
 /*
- * Memcpy-based UTS.
+ * DMA BitBlt based UTS/DFS.
+ *
+ * EXA hits more optimized paths when it does not have to fallback because of
+ * missing UTS/DFS. Fallback to local memcpy-based UTS/DFS when dmablit fails.
  */
 static Bool
 Mach64UploadToScreen(PixmapPtr pDst, int x, int y, int w, int h,
     char *src, int src_pitch)
 {
+    ScrnInfoPtr pScreenInfo = xf86Screens[pDst->drawable.pScreen->myNum];
+    ATIPtr pATI = ATIPTR(pScreenInfo);
+
     char  *dst        = pDst->devPrivate.ptr;
+    int    dst_offset = exaGetPixmapOffset(pDst);
     int    dst_pitch  = exaGetPixmapPitch(pDst);
 
     int bpp    = pDst->drawable.bitsPerPixel;
     int cpp    = (bpp + 7) / 8;
     int wBytes = w * cpp;
 
+#ifdef XF86DRI_DEVEL
+
+    drmMach64SysBlit blit;
+    int ret;
+
+    dst_offset += (x * cpp) + (y * dst_pitch);
+
+    if (!pATI->UpDownAccelEnabled)
+        goto uts_fallback;
+
+    /* Fallback for small or narrow pixmaps:
+     * memcpy-based UTS gets 155 MB/s on average, the following allows only for
+     * pixmaps large enough to get an average throughput of 125 MB/s which is
+     * close to what memcpy gets, while offloading the CPU.
+     */
+    if (h <= 64 || wBytes <= 256)
+        goto uts_fallback;
+
+    /* Byte alignments should match */
+    if ((((unsigned long)src & 3) != (dst_offset & 3)) ||
+        ((src_pitch & 3) != (dst_pitch & 3)))
+        goto uts_fallback;
+
+    /* Pixmap transfer should require at most 2K DMA descriptors */
+    if (h > 1024 || wBytes > 4096)
+        goto uts_fallback;
+
+    ATIDRISync(pScreenInfo);
+
+    blit.h = h;
+    blit.w_bytes = wBytes;
+    blit.fb_addr = dst_offset;
+    blit.fb_pitch = dst_pitch;
+    blit.mem_addr = src;
+    blit.mem_pitch = src_pitch;
+    blit.to_fb = 1;
+
+    ret = drmCommandWrite(pATI->drmFD, DRM_MACH64_SYSBLIT, &blit,
+			  sizeof(blit));
+
+    if (ret == 0)
+        return TRUE;
+
+#endif /* XF86DRI_DEVEL */
+
+uts_fallback: /* memcpy-based UTS */
+
     exaWaitSync(pDst->drawable.pScreen);
 
     dst += (x * cpp) + (y * dst_pitch);
@@ -414,20 +469,68 @@ Mach64UploadToScreen(PixmapPtr pDst, int
     return TRUE;
 }
 
-/*
- * Memcpy-based DFS.
- */
 static Bool
 Mach64DownloadFromScreen(PixmapPtr pSrc, int x, int y, int w, int h,
     char *dst, int dst_pitch)
 {
+    ScrnInfoPtr pScreenInfo = xf86Screens[pSrc->drawable.pScreen->myNum];
+    ATIPtr pATI = ATIPTR(pScreenInfo);
+
     char  *src        = pSrc->devPrivate.ptr;
+    int    src_offset = exaGetPixmapOffset(pSrc);
     int    src_pitch  = exaGetPixmapPitch(pSrc);
 
     int bpp    = pSrc->drawable.bitsPerPixel;
     int cpp    = (bpp + 7) / 8;
     int wBytes = w * cpp;
 
+#ifdef XF86DRI_DEVEL
+
+    drmMach64SysBlit blit;
+    int ret;
+
+    src_offset += (x * cpp) + (y * src_pitch);
+
+    if (!pATI->UpDownAccelEnabled)
+        goto dfs_fallback;
+
+    /* Fallback for small or narrow pixmaps:
+     * memcpy-based DFS gets 6.5 MB/s on average, the following allows for
+     * a wide range of short or big pixmaps for which dmablit gets around
+     * 13 MB/s and 72 MB/s respectively.
+     */
+    if (w * h <= 24 * 24 || h <= 0 || wBytes <= 64)
+        goto dfs_fallback;
+
+    /* Byte alignments should match */
+    if ((((unsigned long)dst & 3) != (src_offset & 3)) ||
+        ((src_pitch & 3) != (dst_pitch & 3)))
+        goto dfs_fallback;
+
+    /* Pixmap transfer should require at most 2K DMA descriptors */
+    if (h > 1024 || wBytes > 4096)
+        goto dfs_fallback;
+
+    ATIDRISync(pScreenInfo);
+
+    blit.h = h;
+    blit.w_bytes = wBytes;
+    blit.fb_addr = src_offset;
+    blit.fb_pitch = src_pitch;
+    blit.mem_addr = dst;
+    blit.mem_pitch = dst_pitch;
+    blit.to_fb = 0;
+
+    ret = drmCommandWrite(pATI->drmFD, DRM_MACH64_SYSBLIT, &blit,
+			  sizeof(blit));
+
+    if (ret == 0)
+        return TRUE;
+
+#endif /* XF86DRI_DEVEL */
+
+dfs_fallback: /* memcpy-based DFS */
+
     exaWaitSync(pSrc->drawable.pScreen);
 
     src += (x * cpp) + (y * src_pitch);
@@ -659,9 +762,6 @@ Bool ATIMach64ExaInit(ScreenPtr pScreen)
     pExa->Copy = Mach64Copy;
     pExa->DoneCopy = Mach64DoneCopy;
 
-    /* EXA hits more optimized paths when it does not have to fallback because
-     * of missing UTS/DFS, hook memcpy-based UTS/DFS.
-     */
     pExa->UploadToScreen = Mach64UploadToScreen;
     pExa->DownloadFromScreen = Mach64DownloadFromScreen;
 
diff --git a/src/atioption.c b/src/atioption.c
index 42fbd56..c755304 100644
--- a/src/atioption.c
+++ b/src/atioption.c
@@ -220,6 +220,13 @@ #endif /* TV_OUT */
         FALSE
     },
     {
+        ATI_OPTION_UPDOWN_ACCEL,
+        "UpDownAccel",
+        OPTV_BOOLEAN,
+        {0, },
+        FALSE
+    },
+    {
         -1,
         NULL,
         OPTV_NONE,
diff --git a/src/atioption.h b/src/atioption.h
index 4bdfb04..1d0201f 100644
--- a/src/atioption.h
+++ b/src/atioption.h
@@ -64,7 +64,8 @@ #endif /* TV_OUT */
     ATI_OPTION_SHADOW_FB,
     ATI_OPTION_SWCURSOR,
     ATI_OPTION_ACCELMETHOD,
-    ATI_OPTION_RENDER_ACCEL
+    ATI_OPTION_RENDER_ACCEL,
+    ATI_OPTION_UPDOWN_ACCEL
 } ATIPublicOptionType;
 
 #ifdef TV_OUT
diff --git a/src/atiscreen.c b/src/atiscreen.c
index e849941..c6f1b0c 100644
--- a/src/atiscreen.c
+++ b/src/atiscreen.c
@@ -612,6 +612,33 @@ #ifdef XF86DRI_DEVEL
 
 #endif /* XF86DRI_DEVEL */
 
+#ifdef USE_EXA
+
+    if (pATI->useEXA) {
+
+        /* Check for UTS/DFS acceleration after IRQ installation */
+        if (pATI->UpDownAccelEnabled) {
+
+#ifdef XF86DRI_DEVEL
+
+            if (!pATI->directRenderingEnabled || (pATI->irq <= 0))
+
+#endif /* XF86DRI_DEVEL */
+
+            {
+                xf86DrvMsg(pScreen->myNum, X_INFO,
+                    "Data transfer acceleration between the host memory and "
+                    "the framebuffer requires enabling DRM and IRQ.\n");
+                pATI->UpDownAccelEnabled = FALSE;
+            }
+        }
+
+        xf86DrvMsg(pScreen->myNum, X_INFO, "Data transfer acceleration %s\n",
+                   pATI->UpDownAccelEnabled ? "enabled" : "disabled");
+    }
+
+#endif /* USE_EXA */
+
     return TRUE;
 }
 
diff --git a/src/atistruct.h b/src/atistruct.h
index 39a7e98..03c361a 100644
--- a/src/atistruct.h
+++ b/src/atistruct.h
@@ -309,6 +309,7 @@ #ifdef USE_XAA
 #endif
 #ifdef USE_EXA
     Bool RenderAccelEnabled;
+    Bool UpDownAccelEnabled;
     Mach64ContextRegs3D m3d;
 #endif
 
diff --git a/src/mach64_common.h b/src/mach64_common.h
index f1f765a..3f783e1 100644
--- a/src/mach64_common.h
+++ b/src/mach64_common.h
@@ -46,6 +46,7 @@ #define DRM_MACH64_VERTEX         0x05
 #define DRM_MACH64_BLIT           0x06
 #define DRM_MACH64_FLUSH          0x07
 #define DRM_MACH64_GETPARAM       0x08
+#define DRM_MACH64_SYSBLIT        0x09
 
 /* Buffer flags for clears
  */
@@ -127,4 +128,17 @@ typedef struct {
 #define MACH64_PARAM_FRAMES_QUEUED 1
 #define MACH64_PARAM_IRQ_NR       2
 
+typedef struct drm_mach64_sysblit {
+   unsigned int h;
+   unsigned int w_bytes;
+   
+   unsigned int fb_addr;
+   unsigned int fb_pitch;
+   
+   char        *mem_addr;
+   unsigned int mem_pitch;
+   
+   int to_fb;
+} drmMach64SysBlit;
+
 #endif /* __MACH64_COMMON_H__ */