diff --git a/src/atidri.c b/src/atidri.c
index 01d449a..8b4ede3 100644
--- a/src/atidri.c
+++ b/src/atidri.c
@@ -785,6 +785,7 @@ static Bool ATIDRIAgpInit( ScreenPtr pSc
    pATIDRIServer->agpMode = ATI_DEFAULT_AGP_MODE;
    pATIDRIServer->bufferSize = ATI_DEFAULT_BUFFER_SIZE;
    pATIDRIServer->ringSize = 16; /* 16 kB ring */
+   pATIDRIServer->sgSize = 32;   /* 32 kB sg list */
 
    if ( drmAgpAcquire( pATI->drmFD ) < 0 ) {
       xf86DrvMsg( pScreen->myNum, X_WARNING, "[agp] AGP not available\n" );
@@ -817,6 +818,9 @@ static Bool ATIDRIAgpInit( ScreenPtr pSc
    xf86DrvMsg(pScreen->myNum, X_INFO,
 	      "[agp] Using %d kB for DMA descriptor ring\n", pATIDRIServer->ringSize);
 
+   xf86DrvMsg(pScreen->myNum, X_INFO,
+	      "[agp] Using %d kB for DMA descriptor sg-list\n", pATIDRIServer->sgSize);
+
    if (pATI->OptionBufferSize) {
       if (pATI->OptionBufferSize < 1 || pATI->OptionBufferSize > pATIDRIServer->agpSize  ) {
 	 xf86DrvMsg( pScreen->myNum, X_ERROR, "[agp] Illegal DMA buffers size: %d MB\n",
@@ -844,8 +848,12 @@ static Bool ATIDRIAgpInit( ScreenPtr pSc
    pATIDRIServer->ringStart   = pATIDRIServer->agpOffset;
    pATIDRIServer->ringMapSize = pATIDRIServer->ringSize*1024; /* ringSize is in kB */
 
+   /* Reserve space for the DMA descriptor sg-list */
+   pATIDRIServer->sgStart   = pATIDRIServer->ringStart + pATIDRIServer->ringMapSize;
+   pATIDRIServer->sgMapSize = pATIDRIServer->sgSize*1024; /* sgSize is in kB */
+
    /* Reserve space for the vertex buffer */
-   pATIDRIServer->bufferStart   = pATIDRIServer->ringStart + pATIDRIServer->ringMapSize;
+   pATIDRIServer->bufferStart   = pATIDRIServer->sgStart + pATIDRIServer->sgMapSize;
    pATIDRIServer->bufferMapSize = pATIDRIServer->bufferSize*1024*1024;
     
    /* Reserve the rest for AGP textures */
@@ -880,6 +888,27 @@ static Bool ATIDRIAgpInit( ScreenPtr pSc
 	       "[agp] Ring mapped at 0x%08lx\n",
 	       (unsigned long)pATIDRIServer->ringMap );
 
+   /* Map DMA descriptor sg-list */
+   if ( drmAddMap( pATI->drmFD, pATIDRIServer->sgStart, pATIDRIServer->sgMapSize,
+		   DRM_AGP, DRM_RESTRICTED, &pATIDRIServer->sgHandle ) < 0 ) {
+      xf86DrvMsg( pScreen->myNum, X_ERROR,
+		  "[agp] Could not add sg-list mapping\n" );
+      return FALSE;
+   }
+   xf86DrvMsg( pScreen->myNum, X_INFO,
+	       "[agp] sg-list handle = 0x%08x\n",
+	       pATIDRIServer->sgHandle );
+
+   if ( drmMap( pATI->drmFD, pATIDRIServer->sgHandle,
+		pATIDRIServer->sgMapSize, &pATIDRIServer->sgMap ) < 0 ) {
+      xf86DrvMsg( pScreen->myNum, X_ERROR,
+		  "[agp] Could not map sg-list\n" );
+      return FALSE;
+   }
+   xf86DrvMsg( pScreen->myNum, X_INFO,
+	       "[agp] Ring mapped at 0x%08lx\n",
+	       (unsigned long)pATIDRIServer->sgMap );
+
    /* Map vertex buffers */
    if ( drmAddMap( pATI->drmFD, pATIDRIServer->bufferStart, pATIDRIServer->bufferMapSize,
 		   DRM_AGP, 0, &pATIDRIServer->bufferHandle ) < 0 ) {
@@ -1012,6 +1041,7 @@ static Bool ATIDRIKernelInit( ScreenPtr 
    info.fb_offset		= pATI->LinearBase;
    info.mmio_offset		= pATIDRIServer->regsHandle;
    info.ring_offset		= pATIDRIServer->ringHandle;
+   info.sg_offset		= pATIDRIServer->sgHandle;
    info.buffers_offset		= pATIDRIServer->bufferHandle;
    info.agp_textures_offset	= pATIDRIServer->agpTexHandle;
 
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..662796b 100644
--- a/src/atimach64exa.c
+++ b/src/atimach64exa.c
@@ -57,6 +57,7 @@ #include "config.h"
 #endif
 
 #include <string.h>
+#include <sys/time.h>
 
 #include "ati.h"
 #include "atichip.h"
@@ -387,24 +388,135 @@ (
 
 static void Mach64DoneSolid(PixmapPtr pPixmap) { }
 
-/*
- * Memcpy-based UTS.
- */
+#if 1
+typedef unsigned long cairo_perf_ticks_t;
+
+typedef struct _cairo_perf_timer
+{
+    struct timeval tv_start;
+    struct timeval tv_stop;
+} cairo_perf_timer_t;
+
+static cairo_perf_timer_t timer;
+
+void
+cairo_perf_timer_start (void) {
+    gettimeofday (&timer.tv_start, NULL);
+}
+
+void
+cairo_perf_timer_stop (void) {
+    gettimeofday (&timer.tv_stop, NULL);
+}
+
+cairo_perf_ticks_t
+cairo_perf_timer_elapsed (void) {
+    cairo_perf_ticks_t ticks;
+
+    ticks = (timer.tv_stop.tv_sec - timer.tv_start.tv_sec) * 1000000;
+    ticks += (timer.tv_stop.tv_usec - timer.tv_start.tv_usec);
+
+    return ticks;
+}
+
+static double up = 0.0, dn = 0.0;
+static unsigned ut = 0, dt = 0;
+#endif
+
+#include "drm.h"
+#include "mach64_drm.h"
+
 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;
 
+    drm_mach64_sysblit_t blit;
+    int ret;
+
+    cairo_perf_ticks_t ticks_blit;
+    double t;
+    int v = h;
+
+#ifdef XF86DRI_DEVEL
+
+    dst_offset += (x * cpp) + (y * dst_pitch);
+
+    if (!pATI->directRenderingEnabled)
+        goto uts_fallback;
+
+    if (pATI->irq <= 0)
+        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;
+
+    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;
+
     exaWaitSync(pDst->drawable.pScreen);
 
+    cairo_perf_timer_start();
+
+    ret = drmCommandWrite(pATI->drmFD, DRM_MACH64_SYSBLIT, &blit,
+			  sizeof(blit));
+
+    cairo_perf_timer_stop();
+
+    ticks_blit = cairo_perf_timer_elapsed();
+
+#if 0
+    t = 1.0 * v * wBytes / ticks_blit;
+
+    if (ret == 0) {
+    ut++;
+    up += t;
+    ErrorF("Up (%4d,%4d) %6.1f MB/s %9lu %6.1f MB/s\n", wBytes, v, t, ticks_blit, up / ut);
+    }
+
+    ErrorF("Up ret=%d\n", ret);
+#endif
+
+    if (ret == 0)
+        return TRUE;
+
+#endif /* XF86DRI_DEVEL */
+
+uts_fallback: /* memcpy-based UTS */
+
     dst += (x * cpp) + (y * dst_pitch);
 
+    exaWaitSync(pDst->drawable.pScreen);
+
     while (h--) {
         memcpy(dst, src, wBytes);
         src += src_pitch;
@@ -414,24 +526,97 @@ 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;
 
+    drm_mach64_sysblit_t blit;
+    int ret;
+
+    int v = h;
+    cairo_perf_ticks_t ticks_blit;
+    double t;
+
+#ifdef XF86DRI_DEVEL
+
+    src_offset += (x * cpp) + (y * src_pitch);
+
+    if (!pATI->directRenderingEnabled)
+        goto dfs_fallback;
+
+    if (pATI->irq <= 0)
+        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;
+
+    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;
+
     exaWaitSync(pSrc->drawable.pScreen);
 
+    cairo_perf_timer_start();
+
+    ret = drmCommandWrite(pATI->drmFD, DRM_MACH64_SYSBLIT, &blit,
+			  sizeof(blit));
+
+    cairo_perf_timer_stop();
+
+    ticks_blit = cairo_perf_timer_elapsed();
+
+#if 0
+    t = 1.0 * v * wBytes / ticks_blit;
+
+    if (ret == 0) {
+    dt++;
+    dn += t;
+    ErrorF("Dn (%4d,%4d) %6.1f MB/s %9lu %6.1f MB/s\n", wBytes, v, t, ticks_blit, dn / dt);
+    }
+
+    ErrorF("Dn ret=%d\n", ret);
+#endif
+
+    if (ret == 0)
+        return TRUE;
+
+#endif /* XF86DRI_DEVEL */
+
+dfs_fallback: /* memcpy-based DFS */
+
     src += (x * cpp) + (y * src_pitch);
 
+    exaWaitSync(pSrc->drawable.pScreen);
+
     while (h--) {
         memcpy(dst, src, wBytes);
         src += src_pitch;
diff --git a/src/mach64_common.h b/src/mach64_common.h
index 1fb765a..d213658 100644
--- a/src/mach64_common.h
+++ b/src/mach64_common.h
@@ -92,6 +92,7 @@ typedef struct {
    unsigned long fb_offset;
    unsigned long mmio_offset;
    unsigned long ring_offset;
+   unsigned long sg_offset;
    unsigned long buffers_offset;
    unsigned long agp_textures_offset;
 } drmMach64Init;
diff --git a/src/mach64_dri.h b/src/mach64_dri.h
index 7061931..8c78523 100644
--- a/src/mach64_dri.h
+++ b/src/mach64_dri.h
@@ -47,14 +47,21 @@ typedef struct {
 
    /* DMA descriptor ring */
    unsigned long     ringStart;        /* Offset into AGP space */
-   drm_handle_t         ringHandle;       /* Handle from drmAddMap */
+   drm_handle_t      ringHandle;       /* Handle from drmAddMap */
    drmSize           ringMapSize;      /* Size of map */
    int               ringSize;         /* Size of ring (in kB) */
    drmAddress        ringMap;          /* Map */
 
+   /* DMA descriptor sg-list */
+   unsigned long     sgStart;          /* Offset into AGP space */
+   drm_handle_t      sgHandle;         /* Handle from drmAddMap */
+   drmSize           sgMapSize;        /* Size of map */
+   int               sgSize;           /* Size of sg-list (in kB) */
+   drmAddress        sgMap;            /* Map */
+
    /* vertex buffer data */
    unsigned long     bufferStart;      /* Offset into AGP space */
-   drm_handle_t         bufferHandle;     /* Handle from drmAddMap */
+   drm_handle_t      bufferHandle;     /* Handle from drmAddMap */
    drmSize           bufferMapSize;    /* Size of map */
    int               bufferSize;       /* Size of buffers (in MB) */
    drmAddress        bufferMap;        /* Map */
@@ -64,7 +71,7 @@ typedef struct {
 
    /* AGP Texture data */
    unsigned long     agpTexStart;      /* Offset into AGP space */
-   drm_handle_t         agpTexHandle;     /* Handle from drmAddMap */
+   drm_handle_t      agpTexHandle;     /* Handle from drmAddMap */
    drmSize           agpTexMapSize;    /* Size of map */
    int               agpTexSize;       /* Size of AGP tex space (in MB) */
    drmAddress        agpTexMap;        /* Map */