diff --git a/drivers/ddk/core.S b/drivers/ddk/core.S
index 8f4a11bbe9..4e68712a90 100644
--- a/drivers/ddk/core.S
+++ b/drivers/ddk/core.S
@@ -51,6 +51,8 @@
     .global _SetScreen
     .global _SysMsgBoardStr
 
+    .global _UserAlloc
+
     .global _WaitEvent
 
 
@@ -101,6 +103,8 @@
     .def  _SetKeyboardData;   .scl  2;  .type 32; .endef
     .def  _SysMsgBoardStr;    .scl  2;  .type 32; .endef
 
+    .def  _UserAlloc;         .scl  2;  .type 32; .endef
+
     .def  _WaitEvent;         .scl  2;  .type 32; .endef
 
 
@@ -153,6 +157,7 @@ _SetKeyboardData:
 _SetScreen:
 _SysMsgBoardStr:
 
+_UserAlloc:
 _WaitEvent:
 
         ret
@@ -209,5 +214,7 @@ _WaitEvent:
         .ascii " -export:SetScreen"            # stdcall
         .ascii " -export:SysMsgBoardStr"       # stdcall
 
+        .ascii " -export:UserAlloc"            # stdcall
+
         .ascii " -export:WaitEvent"            # stdcall
 
diff --git a/drivers/video/drm/radeon/r600_blit_kms.c b/drivers/video/drm/radeon/r600_blit_kms.c
index 69a3f096b9..cd3e8b7df7 100644
--- a/drivers/video/drm/radeon/r600_blit_kms.c
+++ b/drivers/video/drm/radeon/r600_blit_kms.c
@@ -843,3 +843,132 @@ void r600_kms_blit_copy(struct radeon_device *rdev,
 	}
 }
 
+
+void r600_kms_video_blit(struct radeon_device *rdev,
+            u64 src_gpu_addr, int dstx, int dsty, int w, int h, int pitch)
+{
+    u64 vb_gpu_addr;
+    u32 *vb;
+
+    DRM_DEBUG("emitting video copy\n");
+    vb = (u32 *)(rdev->r600_blit.vb_ib->ptr + rdev->r600_blit.vb_used);
+
+    if ((rdev->r600_blit.vb_used + 48) > rdev->r600_blit.vb_total) {
+              // WARN_ON(1);
+    }
+
+    vb[0] = i2f(dstx);
+    vb[1] = i2f(dsty);
+    vb[2] = 0;
+    vb[3] = 0;
+
+    vb[4] = i2f(dstx);
+    vb[5] = i2f(dsty+h);
+    vb[6] = 0;
+    vb[7] = i2f(h);
+
+    vb[8] = i2f(dstx + w);
+    vb[9] = i2f(dsty + h);
+    vb[10] = i2f(w);
+    vb[11] = i2f(h);
+
+    /* src 9 */
+    set_tex_resource(rdev, FMT_8_8_8_8,
+                     w, h, pitch/4, src_gpu_addr);
+    /* 5 */
+    cp_set_surface_sync(rdev,
+                        PACKET3_TC_ACTION_ENA, pitch * h, src_gpu_addr);
+
+    /* dst 23 */
+    set_render_target(rdev, COLOR_8_8_8_8,
+                     1024, 768, rdev->mc.vram_start);
+
+            /* scissors 12  */
+    set_scissors(rdev, 0, 0, 1024, 768);
+
+    /* Vertex buffer setup 14 */
+    vb_gpu_addr = rdev->r600_blit.vb_ib->gpu_addr + rdev->r600_blit.vb_used;
+    set_vtx_resource(rdev, vb_gpu_addr);
+
+    /* draw 10 */
+    draw_auto(rdev);
+
+    /* 5 */
+    cp_set_surface_sync(rdev,
+                PACKET3_CB_ACTION_ENA | PACKET3_CB0_DEST_BASE_ENA,
+                1024*4*768, rdev->mc.vram_start);
+
+    /* 78 ring dwords per loop */
+    vb += 12;
+    rdev->r600_blit.vb_used += 12 * 4;
+
+}
+
+extern struct radeon_device *main_device;
+
+int r600_video_blit(uint64_t src_offset, int  x, int y,
+                    int w, int h, int pitch)
+{
+    int r;
+    struct radeon_device *rdev = main_device;
+
+    mutex_lock(&rdev->r600_blit.mutex);
+    rdev->r600_blit.vb_ib = NULL;
+    r = r600_blit_prepare_copy(rdev, h*pitch);
+    if (r) {
+//       if (rdev->r600_blit.vb_ib)
+//           radeon_ib_free(rdev, &rdev->r600_blit.vb_ib);
+        mutex_unlock(&rdev->r600_blit.mutex);
+        return r;
+    }
+
+
+    r600_kms_video_blit(rdev, src_offset,x,y,w,h,pitch);
+    r600_blit_done_copy(rdev, NULL);
+    mutex_unlock(&rdev->r600_blit.mutex);
+};
+
+int r600_create_video(int w, int h, u32_t *outp)
+{
+    int r;
+    struct radeon_device *rdev = main_device;
+    struct radeon_bo *sobj = NULL;
+    uint64_t saddr;
+    void   *uaddr;
+
+    size_t size;
+    size_t pitch;
+
+    pitch = radeon_align_pitch(rdev, w, 32, false) * 4;
+
+    size = pitch * h;
+    r = radeon_bo_create(rdev, size, PAGE_SIZE, true,
+                         RADEON_GEM_DOMAIN_GTT, &sobj);
+    if (r) {
+        goto fail;
+    }
+    r = radeon_bo_reserve(sobj, false);
+    if (unlikely(r != 0))
+        goto fail;
+    r = radeon_bo_pin(sobj, RADEON_GEM_DOMAIN_GTT, &saddr);
+//   radeon_bo_unreserve(sobj);
+    if (r) {
+        goto fail;
+    }
+
+    r = radeon_bo_user_map(sobj, &uaddr);
+    if (r) {
+        goto fail;
+    }
+
+    ((uint64_t*)outp)[0] = saddr;
+    outp[2] = uaddr;
+    outp[3] = pitch;
+
+    dbgprintf("Create video surface %x, mapped at %x pitch %d\n",
+              (uint32_t)saddr, uaddr, pitch);
+    return 0;
+
+fail:
+    return -1;
+};
diff --git a/drivers/video/drm/radeon/radeon.h b/drivers/video/drm/radeon/radeon.h
index ed7c0c9f92..1da24645d5 100644
--- a/drivers/video/drm/radeon/radeon.h
+++ b/drivers/video/drm/radeon/radeon.h
@@ -340,6 +340,7 @@ struct radeon_bo {
 	struct ttm_bo_kmap_obj		kmap;
     unsigned                    pin_count;
     void                       *kptr;
+    void                       *uptr;
     u32                         cpu_addr;
     u32                         tiling_flags;
     u32                         pitch;
diff --git a/drivers/video/drm/radeon/radeon_asic.c b/drivers/video/drm/radeon/radeon_asic.c
index 23a98729b0..3412a781d9 100644
--- a/drivers/video/drm/radeon/radeon_asic.c
+++ b/drivers/video/drm/radeon/radeon_asic.c
@@ -165,6 +165,7 @@ static struct radeon_asic r100_asic = {
 	.hpd_sense = &r100_hpd_sense,
 	.hpd_set_polarity = &r100_hpd_set_polarity,
 	.ioctl_wait_idle = NULL,
+	.gui_idle = &r100_gui_idle,
 };
 
 static struct radeon_asic r200_asic = {
@@ -203,6 +204,7 @@ static struct radeon_asic r200_asic = {
 	.hpd_sense = &r100_hpd_sense,
 	.hpd_set_polarity = &r100_hpd_set_polarity,
 	.ioctl_wait_idle = NULL,
+	.gui_idle = &r100_gui_idle,
 };
 
 static struct radeon_asic r300_asic = {
@@ -241,6 +243,7 @@ static struct radeon_asic r300_asic = {
 	.hpd_sense = &r100_hpd_sense,
 	.hpd_set_polarity = &r100_hpd_set_polarity,
 	.ioctl_wait_idle = NULL,
+	.gui_idle = &r100_gui_idle,
 };
 
 static struct radeon_asic r300_asic_pcie = {
@@ -278,6 +281,7 @@ static struct radeon_asic r300_asic_pcie = {
 	.hpd_sense = &r100_hpd_sense,
 	.hpd_set_polarity = &r100_hpd_set_polarity,
 	.ioctl_wait_idle = NULL,
+	.gui_idle = &r100_gui_idle,
 };
 
 static struct radeon_asic r420_asic = {
diff --git a/drivers/video/drm/radeon/radeon_benchmark.c b/drivers/video/drm/radeon/radeon_benchmark.c
index d92582e695..27ac83edfc 100644
--- a/drivers/video/drm/radeon/radeon_benchmark.c
+++ b/drivers/video/drm/radeon/radeon_benchmark.c
@@ -48,7 +48,7 @@ void radeon_benchmark_move(struct radeon_device *rdev, unsigned bsize,
     ENTER();
 
 	size = bsize;
-    n = 2; //1024;
+    n = 4; //1024;
 
     dbgprintf("source domain %x\n", sdomain);
 
@@ -99,12 +99,14 @@ void radeon_benchmark_move(struct radeon_device *rdev, unsigned bsize,
 			if (r) {
 				goto out_cleanup;
 			}
+		}
+
 			r = radeon_fence_wait(fence, false);
 			if (r) {
 				goto out_cleanup;
 			}
 			radeon_fence_unref(&fence);
-		}
+
         end_jiffies = GetTimerTicks();
 		time = end_jiffies - start_jiffies;
 		time = jiffies_to_msecs(time);
@@ -128,12 +130,14 @@ void radeon_benchmark_move(struct radeon_device *rdev, unsigned bsize,
 		if (r) {
 			goto out_cleanup;
 		}
+	}
+
 		r = radeon_fence_wait(fence, false);
 		if (r) {
 			goto out_cleanup;
 		}
 		radeon_fence_unref(&fence);
-	}
+
     end_jiffies = GetTimerTicks();
 	time = end_jiffies - start_jiffies;
 	time = jiffies_to_msecs(time);
@@ -176,10 +180,10 @@ out_cleanup:
 
 void radeon_benchmark(struct radeon_device *rdev)
 {
-   radeon_benchmark_move(rdev, 8192*4096, RADEON_GEM_DOMAIN_GTT,
+   radeon_benchmark_move(rdev, 4096*4096, RADEON_GEM_DOMAIN_GTT,
                  RADEON_GEM_DOMAIN_VRAM);
-   radeon_benchmark_move(rdev, 8192*4096, RADEON_GEM_DOMAIN_VRAM,
+   radeon_benchmark_move(rdev, 4096*4096, RADEON_GEM_DOMAIN_VRAM,
                  RADEON_GEM_DOMAIN_GTT);
-   radeon_benchmark_move(rdev, 8192*4096, RADEON_GEM_DOMAIN_VRAM,
+   radeon_benchmark_move(rdev, 4096*4096, RADEON_GEM_DOMAIN_VRAM,
                  RADEON_GEM_DOMAIN_VRAM);
 }
diff --git a/drivers/video/drm/radeon/radeon_device.c b/drivers/video/drm/radeon/radeon_device.c
index 8398bd8119..d7c6da459f 100644
--- a/drivers/video/drm/radeon/radeon_device.c
+++ b/drivers/video/drm/radeon/radeon_device.c
@@ -974,6 +974,12 @@ static struct pci_device_id pciidlist[] = {
 #define SRV_ENUM_MODES  1
 #define SRV_SET_MODE    2
 
+#define SRV_CREATE_VIDEO 9
+#define SRV_BLIT_VIDEO   10
+
+int r600_video_blit(uint64_t src_offset, int  x, int y,
+                    int w, int h, int pitch);
+
 int _stdcall display_handler(ioctl_t *io)
 {
     int    retval = -1;
@@ -1016,6 +1022,18 @@ int _stdcall display_handler(ioctl_t *io)
                 retval = set_user_mode((videomode_t*)inp);
             };
             break;
+
+        case SRV_CREATE_VIDEO:
+            retval = r600_create_video(inp[0], inp[1], outp);
+            break;
+
+        case SRV_BLIT_VIDEO:
+            r600_video_blit( ((uint64_t*)inp)[0], inp[2], inp[3],
+                    inp[4], inp[5], inp[6]);
+
+            retval = 0;
+            break;
+
     };
 
     return retval;
diff --git a/drivers/video/drm/radeon/radeon_irq_kms.c b/drivers/video/drm/radeon/radeon_irq_kms.c
index 80e7e7301e..9a87383a99 100644
--- a/drivers/video/drm/radeon/radeon_irq_kms.c
+++ b/drivers/video/drm/radeon/radeon_irq_kms.c
@@ -32,7 +32,7 @@
 #include "radeon.h"
 #include "atom.h"
 
-static struct radeon_device *main_device;
+struct radeon_device *main_device;
 
 #if 0
 
diff --git a/drivers/video/drm/radeon/radeon_object_kos.c b/drivers/video/drm/radeon/radeon_object_kos.c
index c1269b4c79..4440f3b609 100644
--- a/drivers/video/drm/radeon/radeon_object_kos.c
+++ b/drivers/video/drm/radeon/radeon_object_kos.c
@@ -294,6 +294,48 @@ int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
     return 0;
 }
 
+int radeon_bo_user_map(struct radeon_bo *bo, void **ptr)
+{
+    bool is_iomem;
+
+    if (bo->uptr) {
+        if (ptr) {
+            *ptr = bo->uptr;
+        }
+        return 0;
+    }
+
+    if(bo->domain & RADEON_GEM_DOMAIN_VRAM)
+    {
+        return -1;
+    }
+    else
+    {
+        bo->uptr = UserAlloc(bo->tbo.num_pages << PAGE_SHIFT);
+        if(bo->uptr)
+        {
+            u32_t *src, *dst;
+            int count;
+            src =  &((u32_t*)page_tabs)[(u32_t)bo->kptr >> 12];
+            dst =  &((u32_t*)page_tabs)[(u32_t)bo->uptr >> 12];
+            count = bo->tbo.num_pages;
+
+            while(count--)
+            {
+              *dst++ = (0xFFFFF000 & *src++) | 0x207 ; // map as shared page
+            };
+        }
+        else
+            return -1;
+    }
+
+    if (ptr) {
+        *ptr = bo->uptr;
+    }
+
+    return 0;
+}
+
 void radeon_bo_kunmap(struct radeon_bo *bo)
 {
     if (bo->kptr == NULL)