Merge branch 'master' into drm-gem

Conflicts:

	configure.ac
	src/reg_dumper/Makefile.am

6 files changed, 776 insertions, 738 deletions

diff --git a/src/xvmc/I810XvMC.c b/src/xvmc/I810XvMC.c
index 03be2514..ce5395ac 100644
--- a/src/xvmc/I810XvMC.c
+++ b/src/xvmc/I810XvMC.c
@@ -130,7 +130,6 @@ Status XvMCCreateContext(Display *display, XvPortID port,
 			 int surface_type_id, int width, int height, int flags,
 			 XvMCContext *context) {  
   i810XvMCContext *pI810XvMC;
-  char busIdString[10];
   int priv_count;
   uint *priv_data;
   uint magic;
@@ -217,7 +216,6 @@ Status XvMCCreateContext(Display *display, XvPortID port,
   /* Open DRI Device */
   if((pI810XvMC->fd = drmOpen("i810",NULL)) < 0) {
     printf("DRM Device for i810 could not be opened.\n");
-    free(busIdString);
     free(pI810XvMC);
     return BadAccess;
   } /* !pI810XvMC->fd */
diff --git a/src/xvmc/i915_structs.h b/src/xvmc/i915_structs.h
index a25d7349..b5185cb7 100644
--- a/src/xvmc/i915_structs.h
+++ b/src/xvmc/i915_structs.h
@@ -28,6 +28,8 @@
 #ifndef _I915_STRUCTS_H
 #define _I915_STRUCTS_H
 
+#include <stdint.h>
+
 /* MI_INSTRUCTION */
 #define CMD_MI          0x00
 
@@ -352,7 +354,7 @@ struct i915_3dstate_dest_buffer_variables_mpeg
         unsigned rcontrol : 1;
         unsigned decode_mode : 2;
     } dw1;
-        
+
     struct {
         unsigned pad0 : 1;
         unsigned picture_coding_type : 2;
@@ -372,6 +374,15 @@ struct i915_3dstate_dest_buffer_variables_mpeg
     } dw2;
 };
 
+struct i915_mc_static_indirect_state_buffer {
+    struct i915_3dstate_buffer_info dest_y;
+    struct i915_3dstate_buffer_info dest_u;
+    struct i915_3dstate_buffer_info dest_v;
+    struct i915_3dstate_dest_buffer_variables dest_buf;
+    struct i915_3dstate_dest_buffer_variables_mpeg dest_buf_mpeg;
+    struct i915_3dstate_buffer_info corr;
+};
+
 #define MAP_MAP0        0x0001
 #define MAP_MAP1        0x0002
 #define MAP_MAP2        0x0004
@@ -432,7 +443,18 @@ struct i915_3dstate_map_state
         unsigned map_mask : 16;
         unsigned pad0 : 16;
     } dw1;
-//    struct texture_map *tms;
+};
+
+struct i915_mc_map_state {
+    struct i915_3dstate_map_state y_map;
+    struct texture_map y_forward;
+    struct texture_map y_backward;
+    struct i915_3dstate_map_state u_map;
+    struct texture_map u_forward;
+    struct texture_map u_backward;
+    struct i915_3dstate_map_state v_map;
+    struct texture_map v_forward;
+    struct texture_map v_backward;
 };
 
 #define SAMPLER_SAMPLER0        0x0001
@@ -532,6 +554,9 @@ struct i915_3dstate_sampler_state
         unsigned sampler_masker : 16;
         unsigned pad0 : 16;
     } dw1;
+    /* we always use two samplers for mc */
+    struct texture_sampler sampler0;
+    struct texture_sampler sampler1;
 };
 
 struct arithmetic_inst
@@ -641,17 +666,48 @@ union shader_inst
     struct declaration_inst d;
 };
 
+struct i915_3dstate_pixel_shader_header {
+    unsigned length : 9;
+    unsigned pad0 : 6;
+    unsigned retain : 1;
+    unsigned opcode : 13;
+    unsigned type : 3;
+};
+
 struct i915_3dstate_pixel_shader_program
 {
-    struct {
-        unsigned length : 9;
-        unsigned pad0 : 6;
-        unsigned retain : 1;
-        unsigned opcode : 13;
-        unsigned type : 3;
-    } dw0;
-
-    // union shader_inst *insts;
+    struct i915_3dstate_pixel_shader_header shader0;
+    /* mov oC, c0.0000 */
+    uint32_t inst0[3];
+
+    struct i915_3dstate_pixel_shader_header shader1;
+    /* dcl t0.xy */
+    /* dcl t1.xy */
+    /* dcl_2D s0 */
+    /* texld r0, t0, s0 */
+    /* mov oC, r0 */
+    uint32_t inst1[3*5];
+
+    struct i915_3dstate_pixel_shader_header shader2;
+    /* dcl t2.xy */
+    /* dcl t3.xy */
+    /* dcl_2D s1 */
+    /* texld r0, t2, s1 */
+    /* mov oC, r0 */
+    uint32_t inst2[3*5];
+
+    struct i915_3dstate_pixel_shader_header shader3;
+    /* dcl t0.xy */
+    /* dcl t1.xy */
+    /* dcl t2.xy */
+    /* dcl t3.xy */
+    /* dcl_2D s0 */
+    /* dcl_2D s1 */
+    /* texld r0, t0, s0 */
+    /* texld r0, t2, s1 */
+    /* add r0, r0, r1*/
+    /* mov oC, r0 */
+    uint32_t inst3[3*10];
 };
 
 #define REG_CR0         0x00000001
@@ -707,7 +763,8 @@ struct i915_3dstate_pixel_shader_constants
     struct {
         unsigned reg_mask;
     } dw1;
-    // struct shader_constant *consts;
+    /* we only need one constant */
+    struct shader_constant value;
 };
 
 #define BLOCK_SIS       0x01
diff --git a/src/xvmc/i915_xvmc.c b/src/xvmc/i915_xvmc.c
index 519de5a3..c32073a1 100644
--- a/src/xvmc/i915_xvmc.c
+++ b/src/xvmc/i915_xvmc.c
@@ -39,9 +39,9 @@
                                  SIZE_Y420(surface->width, surface->height))
 
 typedef union {
-    short s[4];
-    uint  u[2];
-} su_t;
+    int16_t component[2];
+    int32_t v;
+} vector_t;
 
 #if 0
 static int findOverlap(unsigned int width, unsigned int height,
@@ -81,577 +81,6 @@ static int findOverlap(unsigned int width, unsigned int height,
 }
 #endif
 
-static void i915_flush(int map, int render)
-{
-    struct i915_mi_flush mi_flush;
-
-    memset(&mi_flush, 0, sizeof(mi_flush));
-    mi_flush.dw0.type = CMD_MI;
-    mi_flush.dw0.opcode = OPC_MI_FLUSH;
-    mi_flush.dw0.map_cache_invalidate = map;
-    mi_flush.dw0.render_cache_flush_inhibit = render;
-
-    intelBatchbufferData(&mi_flush, sizeof(mi_flush), 0);
-}
-
-/* for MC picture rendering */
-static void i915_mc_static_indirect_state_buffer(XvMCContext *context,
-	XvMCSurface *surface,
-	unsigned int picture_structure,
-	unsigned int flags,
-	unsigned int picture_coding_type)
-{
-    struct i915_3dstate_buffer_info *buffer_info;
-    struct i915_3dstate_dest_buffer_variables *dest_buffer_variables;
-    struct i915_3dstate_dest_buffer_variables_mpeg *dest_buffer_variables_mpeg;
-    i915XvMCSurface *pI915Surface = (i915XvMCSurface *)surface->privData;
-    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
-    unsigned int w = surface->width;
-
-    /* 3DSTATE_BUFFER_INFO */
-    /* DEST Y */
-    buffer_info = (struct i915_3dstate_buffer_info *)pI915XvMC->sis.map;
-    memset(buffer_info, 0, sizeof(*buffer_info));
-    buffer_info->dw0.type = CMD_3D;
-    buffer_info->dw0.opcode = OPC_3DSTATE_BUFFER_INFO;
-    buffer_info->dw0.length = 1;
-    buffer_info->dw1.aux_id = 0;
-    buffer_info->dw1.buffer_id = BUFFERID_COLOR_BACK;
-    buffer_info->dw1.fence_regs = 0;    /* disabled */ /* FIXME: tiled y for performance */
-    buffer_info->dw1.tiled_surface = 0; /* linear */
-    buffer_info->dw1.walk = TILEWALK_XMAJOR;
-    buffer_info->dw1.pitch = (pI915Surface->yStride >> 2);      /* in DWords */
-    buffer_info->dw2.base_address = (YOFFSET(pI915Surface) >> 2);    /* starting DWORD address */
-
-    /* DEST U */
-    ++buffer_info;
-    memset(buffer_info, 0, sizeof(*buffer_info));
-    buffer_info->dw0.type = CMD_3D;
-    buffer_info->dw0.opcode = OPC_3DSTATE_BUFFER_INFO;
-    buffer_info->dw0.length = 1;
-    buffer_info->dw1.aux_id = 0;
-    buffer_info->dw1.buffer_id = BUFFERID_COLOR_AUX;
-    buffer_info->dw1.fence_regs = 0;
-    buffer_info->dw1.tiled_surface = 0;
-    buffer_info->dw1.walk = TILEWALK_XMAJOR;
-    buffer_info->dw1.pitch = (pI915Surface->uvStride >> 2);      /* in DWords */
-    buffer_info->dw2.base_address = (UOFFSET(pI915Surface) >> 2);      /* starting DWORD address */
-
-    /* DEST V */
-    ++buffer_info;
-    memset(buffer_info, 0, sizeof(*buffer_info));
-    buffer_info->dw0.type = CMD_3D;
-    buffer_info->dw0.opcode = OPC_3DSTATE_BUFFER_INFO;
-    buffer_info->dw0.length = 1;
-    buffer_info->dw1.aux_id = 1;
-    buffer_info->dw1.buffer_id = BUFFERID_COLOR_AUX;
-    buffer_info->dw1.fence_regs = 0;
-    buffer_info->dw1.tiled_surface = 0;
-    buffer_info->dw1.walk = TILEWALK_XMAJOR;
-    buffer_info->dw1.pitch = (pI915Surface->uvStride >> 2);      /* in Dwords */
-    buffer_info->dw2.base_address = (VOFFSET(pI915Surface) >> 2);      /* starting DWORD address */
-
-    /* 3DSTATE_DEST_BUFFER_VARIABLES */
-    dest_buffer_variables = (struct i915_3dstate_dest_buffer_variables *)(++buffer_info);
-    memset(dest_buffer_variables, 0, sizeof(*dest_buffer_variables));
-    dest_buffer_variables->dw0.type = CMD_3D;
-    dest_buffer_variables->dw0.opcode = OPC_3DSTATE_DEST_BUFFER_VARIABLES;
-    dest_buffer_variables->dw0.length = 0;
-    dest_buffer_variables->dw1.dest_v_bias = 8; /* 0.5 */
-    dest_buffer_variables->dw1.dest_h_bias = 8; /* 0.5 */
-    dest_buffer_variables->dw1.color_fmt = COLORBUFFER_8BIT;
-    dest_buffer_variables->dw1.v_ls = 0;
-    dest_buffer_variables->dw1.v_ls_offset = 0;
-
-    if ((picture_structure & XVMC_FRAME_PICTURE) == XVMC_FRAME_PICTURE) {
-        ;
-    } else if ((picture_structure & XVMC_FRAME_PICTURE) == XVMC_TOP_FIELD) {
-        dest_buffer_variables->dw1.v_ls = 1;
-    } else if ((picture_structure & XVMC_FRAME_PICTURE) == XVMC_BOTTOM_FIELD) {
-        dest_buffer_variables->dw1.v_ls = 1;
-        dest_buffer_variables->dw1.v_ls_offset = 1;
-    }
-
-    /* 3DSTATE_DEST_BUFFER_VARIABLES_MPEG */
-    dest_buffer_variables_mpeg = (struct i915_3dstate_dest_buffer_variables_mpeg *)(++dest_buffer_variables);
-    memset(dest_buffer_variables_mpeg, 0, sizeof(*dest_buffer_variables_mpeg));
-    dest_buffer_variables_mpeg->dw0.type = CMD_3D;
-    dest_buffer_variables_mpeg->dw0.opcode = OPC_3DSTATE_DEST_BUFFER_VARIABLES_MPEG;
-    dest_buffer_variables_mpeg->dw0.length = 1;
-    dest_buffer_variables_mpeg->dw1.decode_mode = MPEG_DECODE_MC;
-    dest_buffer_variables_mpeg->dw1.rcontrol = 0;               /* for MPEG-1/MPEG-2 */
-    dest_buffer_variables_mpeg->dw1.bidir_avrg_control = 0;     /* for MPEG-1/MPEG-2/MPEG-4 */
-    dest_buffer_variables_mpeg->dw1.abort_on_error = 1;
-    dest_buffer_variables_mpeg->dw1.intra8 = 0;         /* 16-bit formatted correction data */
-    dest_buffer_variables_mpeg->dw1.tff = 1;
-
-    if (picture_structure & XVMC_FRAME_PICTURE) {
-        ;
-    } else if (picture_structure & XVMC_TOP_FIELD) {
-        if (flags & XVMC_SECOND_FIELD)
-            dest_buffer_variables_mpeg->dw1.tff = 0;
-        else
-            dest_buffer_variables_mpeg->dw1.tff = 1;
-    } else if (picture_structure & XVMC_BOTTOM_FIELD) {
-        if (flags & XVMC_SECOND_FIELD)
-            dest_buffer_variables_mpeg->dw1.tff = 1;
-        else
-            dest_buffer_variables_mpeg->dw1.tff = 0;
-    }
-
-    dest_buffer_variables_mpeg->dw1.v_subsample_factor = MC_SUB_1V;
-    dest_buffer_variables_mpeg->dw1.h_subsample_factor = MC_SUB_1H;
-    dest_buffer_variables_mpeg->dw1.picture_width = (w >> 4);     /* in macroblocks */
-    dest_buffer_variables_mpeg->dw2.picture_coding_type = picture_coding_type;
-
-    /* 3DSATE_BUFFER_INFO */
-    /* CORRECTION DATA */
-    buffer_info = (struct i915_3dstate_buffer_info *)(++dest_buffer_variables_mpeg);
-    memset(buffer_info, 0, sizeof(*buffer_info));
-    buffer_info->dw0.type = CMD_3D;
-    buffer_info->dw0.opcode = OPC_3DSTATE_BUFFER_INFO;
-    buffer_info->dw0.length = 1;
-    buffer_info->dw1.aux_id = 0;
-    buffer_info->dw1.buffer_id = BUFFERID_MC_INTRA_CORR;
-    buffer_info->dw1.aux_id = 0;
-    buffer_info->dw1.fence_regs = 0;
-    buffer_info->dw1.tiled_surface = 0;
-    buffer_info->dw1.walk = 0;
-    buffer_info->dw1.pitch = 0;
-    buffer_info->dw2.base_address = (pI915XvMC->corrdata.offset >> 2);  /* starting DWORD address */
-}
-
-static void i915_mc_map_state_buffer(XvMCContext *context,
-                                       i915XvMCSurface *privTarget,
-                                       i915XvMCSurface *privPast,
-                                       i915XvMCSurface *privFuture)
-{
-    struct i915_3dstate_map_state *map_state;
-    struct texture_map *tm;
-    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
-    unsigned int w = context->width, h = context->height;
-
-    /* 3DSATE_MAP_STATE: Y */
-    map_state = (struct i915_3dstate_map_state *)pI915XvMC->msb.map;
-    memset(map_state, 0, sizeof(*map_state));
-    map_state->dw0.type = CMD_3D;
-    map_state->dw0.opcode = OPC_3DSTATE_MAP_STATE;
-    map_state->dw0.retain = 1;
-    map_state->dw0.length = 6;
-    map_state->dw1.map_mask = MAP_MAP0 | MAP_MAP1;
-
-    /* texture map: Forward (Past) */
-    tm = (struct texture_map *)(++map_state);
-    memset(tm, 0, sizeof(*tm));
-    tm->tm0.v_ls_offset = 0;
-    tm->tm0.v_ls = 0;
-    tm->tm0.base_address = (YOFFSET(privPast) >> 2);
-    tm->tm1.tile_walk = TILEWALK_XMAJOR;        /* FIXME: tiled y for performace */
-    tm->tm1.tiled_surface = 0;
-    tm->tm1.utilize_fence_regs = 0;
-    tm->tm1.texel_fmt = 0;      /* 8bit */
-    tm->tm1.surface_fmt = 1;    /* 8bit */
-    tm->tm1.width = w - 1;
-    tm->tm1.height = h - 1;
-    tm->tm2.depth = 0;
-    tm->tm2.max_lod = 0;
-    tm->tm2.cube_face = 0;
-    tm->tm2.pitch = (privPast->yStride >> 2) - 1;       /* in DWords - 1 */
-
-    /* texture map: Backward (Future) */
-    ++tm;
-    memset(tm, 0, sizeof(*tm));
-    tm->tm0.v_ls_offset = 0;
-    tm->tm0.v_ls = 0;
-    tm->tm0.base_address = (YOFFSET(privFuture) >> 2);
-    tm->tm1.tile_walk = TILEWALK_XMAJOR;
-    tm->tm1.tiled_surface = 0;
-    tm->tm1.utilize_fence_regs = 0;
-    tm->tm1.texel_fmt = 0;      /* 8bit */
-    tm->tm1.surface_fmt = 1;    /* 8bit */
-    tm->tm1.width = w - 1;
-    tm->tm1.height = h - 1;
-    tm->tm2.depth = 0;
-    tm->tm2.max_lod = 0;
-    tm->tm2.cube_face = 0;
-    tm->tm2.pitch = (privFuture->yStride >> 2) - 1;
-
-    /* 3DSATE_MAP_STATE: U */
-    map_state = (struct i915_3dstate_map_state *)(++tm);
-    memset(map_state, 0, sizeof(*map_state));
-    map_state->dw0.type = CMD_3D;
-    map_state->dw0.opcode = OPC_3DSTATE_MAP_STATE;
-    map_state->dw0.retain = 1;
-    map_state->dw0.length = 6;
-    map_state->dw1.map_mask = MAP_MAP0 | MAP_MAP1;
-
-    /* texture map: Forward */
-    tm = (struct texture_map *)(++map_state);
-    memset(tm, 0, sizeof(*tm));
-    tm->tm0.v_ls_offset = 0;
-    tm->tm0.v_ls = 0;
-    tm->tm0.base_address = (UOFFSET(privPast) >> 2);
-    tm->tm1.tile_walk = TILEWALK_XMAJOR;
-    tm->tm1.tiled_surface = 0;
-    tm->tm1.utilize_fence_regs = 0;
-    tm->tm1.texel_fmt = 0;      /* 8bit */
-    tm->tm1.surface_fmt = 1;    /* 8bit */
-    tm->tm1.width = (w >> 1) - 1;
-    tm->tm1.height = (h >> 1) - 1;
-    tm->tm2.depth = 0;
-    tm->tm2.max_lod = 0;
-    tm->tm2.cube_face = 0;
-    tm->tm2.pitch = (privPast->uvStride >> 2) - 1;       /* in DWords - 1 */
-
-    /* texture map: Backward */
-    ++tm;
-    memset(tm, 0, sizeof(*tm));
-    tm->tm0.v_ls_offset = 0;
-    tm->tm0.v_ls = 0;
-    tm->tm0.base_address = (UOFFSET(privFuture) >> 2);
-    tm->tm1.tile_walk = TILEWALK_XMAJOR;
-    tm->tm1.tiled_surface = 0;
-    tm->tm1.utilize_fence_regs = 0;
-    tm->tm1.texel_fmt = 0;
-    tm->tm1.surface_fmt = 1;
-    tm->tm1.width = (w >> 1) - 1;
-    tm->tm1.height = (h >> 1) - 1;
-    tm->tm2.depth = 0;
-    tm->tm2.max_lod = 0;
-    tm->tm2.cube_face = 0;
-    tm->tm2.pitch = (privFuture->uvStride >> 2) - 1;
-
-    /* 3DSATE_MAP_STATE: V */
-    map_state = (struct i915_3dstate_map_state *)(++tm);
-    memset(map_state, 0, sizeof(*map_state));
-    map_state->dw0.type = CMD_3D;
-    map_state->dw0.opcode = OPC_3DSTATE_MAP_STATE;
-    map_state->dw0.retain = 1;
-    map_state->dw0.length = 6;
-    map_state->dw1.map_mask = MAP_MAP0 | MAP_MAP1;
-
-    /* texture map: Forward */
-    tm = (struct texture_map *)(++map_state);
-    memset(tm, 0, sizeof(*tm));
-    tm->tm0.v_ls_offset = 0;
-    tm->tm0.v_ls = 0;
-    tm->tm0.base_address = (VOFFSET(privPast) >> 2);
-    tm->tm1.tile_walk = TILEWALK_XMAJOR;
-    tm->tm1.tiled_surface = 0;
-    tm->tm1.utilize_fence_regs = 0;
-    tm->tm1.texel_fmt = 0;
-    tm->tm1.surface_fmt = 1;
-    tm->tm1.width = (w >> 1) - 1;
-    tm->tm1.height = (h >> 1) - 1;
-    tm->tm2.depth = 0;
-    tm->tm2.max_lod = 0;
-    tm->tm2.cube_face = 0;
-    tm->tm2.pitch = (privPast->uvStride >> 2) - 1;       /* in DWords - 1 */
-
-    /* texture map: Backward */
-    ++tm;
-    memset(tm, 0, sizeof(*tm));
-    tm->tm0.v_ls_offset = 0;
-    tm->tm0.v_ls = 0;
-    tm->tm0.base_address = (VOFFSET(privFuture) >> 2);
-    tm->tm1.tile_walk = TILEWALK_XMAJOR;
-    tm->tm1.tiled_surface = 0;
-    tm->tm1.utilize_fence_regs = 0;
-    tm->tm1.texel_fmt = 0;
-    tm->tm1.surface_fmt = 1;
-    tm->tm1.width = (w >> 1) - 1;
-    tm->tm1.height = (h >> 1) - 1;
-    tm->tm2.depth = 0;
-    tm->tm2.max_lod = 0;
-    tm->tm2.cube_face = 0;
-    tm->tm2.pitch = (privFuture->uvStride >> 2) - 1;
-}
-
-static void i915_mc_load_sis_msb_buffers(XvMCContext *context)
-{
-    struct i915_3dstate_load_indirect *load_indirect;
-    sis_state *sis = NULL;
-    msb_state *msb = NULL;
-    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
-    void *base = NULL;
-    unsigned int size;
-    int mem_select = 1;
-
-    /* 3DSTATE_LOAD_INDIRECT */
-    size = sizeof(*load_indirect) + sizeof(*sis) + sizeof(*msb);
-    base = calloc(1, size);
-    load_indirect = (struct i915_3dstate_load_indirect *)base;
-    load_indirect->dw0.type = CMD_3D;
-    load_indirect->dw0.opcode = OPC_3DSTATE_LOAD_INDIRECT;
-    load_indirect->dw0.block_mask = BLOCK_SIS | BLOCK_MSB;
-    load_indirect->dw0.length = (size >> 2) - 2;
-
-    if (pI915XvMC->deviceID == PCI_CHIP_I915_G ||
-        pI915XvMC->deviceID == PCI_CHIP_I915_GM ||
-        pI915XvMC->deviceID == PCI_CHIP_I945_G ||
-        pI915XvMC->deviceID == PCI_CHIP_I945_GM)
-        mem_select = 0;
-
-    load_indirect->dw0.mem_select = mem_select;
-
-    /* SIS */
-    sis = (sis_state *)(++load_indirect);
-    sis->dw0.valid = 1;
-    sis->dw0.force = 1;
-    sis->dw1.length = 16; // 4 * 3 + 2 + 3 - 1
-
-    if (mem_select)
-        sis->dw0.buffer_address = (pI915XvMC->sis.offset >> 2);
-    else
-        sis->dw0.buffer_address = (pI915XvMC->sis.bus_addr >> 2);
-
-    /* MSB */
-    msb = (msb_state *)(++sis);
-    msb->dw0.valid = 1;
-    msb->dw0.force = 1;
-    msb->dw1.length = 23; // 3 * 8 - 1
-
-    if (mem_select)
-        msb->dw0.buffer_address = (pI915XvMC->msb.offset >> 2);
-    else
-        msb->dw0.buffer_address = (pI915XvMC->msb.bus_addr >> 2);
-
-    intelBatchbufferData(base, size, 0);
-    free(base);
-}
-
-static void i915_mc_mpeg_set_origin(XvMCContext *context, XvMCMacroBlock *mb)
-{
-    struct i915_3dmpeg_set_origin set_origin;
-
-    /* 3DMPEG_SET_ORIGIN */
-    memset(&set_origin, 0, sizeof(set_origin));
-    set_origin.dw0.type = CMD_3D;
-    set_origin.dw0.opcode = OPC_3DMPEG_SET_ORIGIN;
-    set_origin.dw0.length = 0;
-    set_origin.dw1.h_origin = mb->x;
-    set_origin.dw1.v_origin = mb->y;
-
-    intelBatchbufferData(&set_origin, sizeof(set_origin), 0);
-}
-
-static void i915_mc_mpeg_macroblock_ipicture(XvMCContext *context, XvMCMacroBlock *mb)
-{
-    struct i915_3dmpeg_macroblock_ipicture macroblock_ipicture;
-
-    /* 3DMPEG_MACROBLOCK_IPICTURE */
-    memset(&macroblock_ipicture, 0, sizeof(macroblock_ipicture));
-    macroblock_ipicture.dw0.type = CMD_3D;
-    macroblock_ipicture.dw0.opcode = OPC_3DMPEG_MACROBLOCK_IPICTURE;
-    macroblock_ipicture.dw0.dct_type = (mb->dct_type == XVMC_DCT_TYPE_FIELD);
-
-    intelBatchbufferData(&macroblock_ipicture, sizeof(macroblock_ipicture), 0);
-}
-
-#if 0
-static void i915_mc_mpeg_macroblock_0mv(XvMCContext *context, XvMCMacroBlock *mb)
-{
-    struct i915_3dmpeg_macroblock_0mv macroblock_0mv;
-
-    /* 3DMPEG_MACROBLOCK(0mv) */
-    memset(&macroblock_0mv, 0, sizeof(macroblock_0mv));
-    macroblock_0mv.header.dw0.type = CMD_3D;
-    macroblock_0mv.header.dw0.opcode = OPC_3DMPEG_MACROBLOCK;
-    macroblock_0mv.header.dw0.length = 0;
-    macroblock_0mv.header.dw1.mb_intra = 1;     /* should be 1 */
-    macroblock_0mv.header.dw1.forward = 0;      /* should be 0 */
-    macroblock_0mv.header.dw1.backward = 0;     /* should be 0 */
-    macroblock_0mv.header.dw1.h263_4mv = 0;     /* should be 0 */
-    macroblock_0mv.header.dw1.dct_type = (mb->dct_type == XVMC_DCT_TYPE_FIELD);
-
-/*
-    if (!mb->coded_block_pattern)
-        macroblock_0mv.header.dw1.dct_type = XVMC_DCT_TYPE_FRAME;
-*/
-
-    macroblock_0mv.header.dw1.motion_type = 0; // (mb->motion_type & 0x3);
-    macroblock_0mv.header.dw1.vertical_field_select = 0; // mb->motion_vertical_field_select & 0xf;
-    macroblock_0mv.header.dw1.coded_block_pattern = mb->coded_block_pattern;
-    macroblock_0mv.header.dw1.skipped_macroblocks = 0;
-
-    intelBatchbufferData(&macroblock_0mv, sizeof(macroblock_0mv), 0);
-}
-#endif
-
-static void i915_mc_mpeg_macroblock_1fbmv(XvMCContext *context, XvMCMacroBlock *mb)
-{
-    struct i915_3dmpeg_macroblock_1fbmv macroblock_1fbmv;
-
-    /* Motion Vectors */
-    su_t fmv;
-    su_t bmv;
-
-    /* 3DMPEG_MACROBLOCK(1fbmv) */
-    memset(&macroblock_1fbmv, 0, sizeof(macroblock_1fbmv));
-    macroblock_1fbmv.header.dw0.type = CMD_3D;
-    macroblock_1fbmv.header.dw0.opcode = OPC_3DMPEG_MACROBLOCK;
-    macroblock_1fbmv.header.dw0.length = 2;
-    macroblock_1fbmv.header.dw1.mb_intra = 0;   /* should be 0 */
-    macroblock_1fbmv.header.dw1.forward = ((mb->macroblock_type & XVMC_MB_TYPE_MOTION_FORWARD) ? 1 : 0);
-    macroblock_1fbmv.header.dw1.backward = ((mb->macroblock_type & XVMC_MB_TYPE_MOTION_BACKWARD) ? 1 : 0);
-    macroblock_1fbmv.header.dw1.h263_4mv = 0;   /* should be 0 */
-    macroblock_1fbmv.header.dw1.dct_type = (mb->dct_type == XVMC_DCT_TYPE_FIELD);
-
-    if (!(mb->coded_block_pattern & 0x3f))
-        macroblock_1fbmv.header.dw1.dct_type = XVMC_DCT_TYPE_FRAME;
-
-    macroblock_1fbmv.header.dw1.motion_type = (mb->motion_type & 0x03);
-    macroblock_1fbmv.header.dw1.vertical_field_select = (mb->motion_vertical_field_select & 0x0f);
-    macroblock_1fbmv.header.dw1.coded_block_pattern = mb->coded_block_pattern;
-    macroblock_1fbmv.header.dw1.skipped_macroblocks = 0;
-
-    fmv.s[0] = mb->PMV[0][0][0];
-    fmv.s[1] = mb->PMV[0][0][1];
-    bmv.s[0] = mb->PMV[0][1][0];
-    bmv.s[1] = mb->PMV[0][1][1];
-
-    macroblock_1fbmv.dw2 = fmv.u[0];
-    macroblock_1fbmv.dw3 = bmv.u[0];
-
-    intelBatchbufferData(&macroblock_1fbmv, sizeof(macroblock_1fbmv), 0);
-}
-
-static void i915_mc_mpeg_macroblock_2fbmv(XvMCContext *context, XvMCMacroBlock *mb, unsigned int ps)
-{
-    struct i915_3dmpeg_macroblock_2fbmv macroblock_2fbmv;
-
-    /* Motion Vectors */
-    su_t fmv;
-    su_t bmv;
-
-    /* 3DMPEG_MACROBLOCK(2fbmv) */
-    memset(&macroblock_2fbmv, 0, sizeof(macroblock_2fbmv));
-    macroblock_2fbmv.header.dw0.type = CMD_3D;
-    macroblock_2fbmv.header.dw0.opcode = OPC_3DMPEG_MACROBLOCK;
-    macroblock_2fbmv.header.dw0.length = 4;
-    macroblock_2fbmv.header.dw1.mb_intra = 0;   /* should be 0 */
-    macroblock_2fbmv.header.dw1.forward = ((mb->macroblock_type & XVMC_MB_TYPE_MOTION_FORWARD) ? 1 : 0);
-    macroblock_2fbmv.header.dw1.backward = ((mb->macroblock_type & XVMC_MB_TYPE_MOTION_BACKWARD) ? 1 : 0);
-    macroblock_2fbmv.header.dw1.h263_4mv = 0;   /* should be 0 */
-    macroblock_2fbmv.header.dw1.dct_type = (mb->dct_type == XVMC_DCT_TYPE_FIELD);
-
-    if (!(mb->coded_block_pattern & 0x3f))
-        macroblock_2fbmv.header.dw1.dct_type = XVMC_DCT_TYPE_FRAME;
-
-    macroblock_2fbmv.header.dw1.motion_type = (mb->motion_type & 0x03);
-    macroblock_2fbmv.header.dw1.vertical_field_select = (mb->motion_vertical_field_select & 0x0f);
-    macroblock_2fbmv.header.dw1.coded_block_pattern = mb->coded_block_pattern;
-    macroblock_2fbmv.header.dw1.skipped_macroblocks = 0;
-
-    fmv.s[0] = mb->PMV[0][0][0];
-    fmv.s[1] = mb->PMV[0][0][1];
-    fmv.s[2] = mb->PMV[1][0][0];
-    fmv.s[3] = mb->PMV[1][0][1];
-    bmv.s[0] = mb->PMV[0][1][0];
-    bmv.s[1] = mb->PMV[0][1][1];
-    bmv.s[2] = mb->PMV[1][1][0];
-    bmv.s[3] = mb->PMV[1][1][1];
-
-    if ((ps & XVMC_FRAME_PICTURE) == XVMC_FRAME_PICTURE) {
-        if ((mb->motion_type & 3) == XVMC_PREDICTION_FIELD) {
-            fmv.s[0] = mb->PMV[0][0][0];
-            fmv.s[1] = mb->PMV[0][0][1] >> 1;
-            fmv.s[2] = mb->PMV[1][0][0];
-            fmv.s[3] = mb->PMV[1][0][1] >> 1;
-            bmv.s[0] = mb->PMV[0][1][0];
-            bmv.s[1] = mb->PMV[0][1][1] >> 1;
-            bmv.s[2] = mb->PMV[1][1][0];
-            bmv.s[3] = mb->PMV[1][1][1] >> 1;
-        } else if ((mb->motion_type & 3) == XVMC_PREDICTION_DUAL_PRIME) {
-            fmv.s[0] = mb->PMV[0][0][0];
-            fmv.s[1] = mb->PMV[0][0][1] >> 1;
-            fmv.s[2] = mb->PMV[0][0][0];
-            fmv.s[3] = mb->PMV[0][0][1] >> 1;  // MPEG2 MV[0][1] isn't used
-            bmv.s[0] = mb->PMV[1][0][0];
-            bmv.s[1] = mb->PMV[1][0][1] >> 1;
-            bmv.s[2] = mb->PMV[1][1][0];
-            bmv.s[3] = mb->PMV[1][1][1] >> 1;
-        }
-    }
-
-    macroblock_2fbmv.dw2 = fmv.u[0];
-    macroblock_2fbmv.dw3 = bmv.u[0];
-    macroblock_2fbmv.dw4 = fmv.u[1];
-    macroblock_2fbmv.dw5 = bmv.u[1];
-
-    intelBatchbufferData(&macroblock_2fbmv, sizeof(macroblock_2fbmv), 0);
-}
-
-/* for MC context initialization */
-static void i915_mc_sampler_state_buffer(XvMCContext *context)
-{
-    struct i915_3dstate_sampler_state *sampler_state;
-    struct texture_sampler *ts;
-    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
-
-    /* 3DSATE_SAMPLER_STATE */
-    sampler_state = (struct i915_3dstate_sampler_state *)pI915XvMC->ssb.map;
-    memset(sampler_state, 0, sizeof(*sampler_state));
-    sampler_state->dw0.type = CMD_3D;
-    sampler_state->dw0.opcode = OPC_3DSTATE_SAMPLER_STATE;
-    sampler_state->dw0.length = 6;
-    sampler_state->dw1.sampler_masker = SAMPLER_SAMPLER0 | SAMPLER_SAMPLER1;
-
-    /* Sampler 0 */
-    ts = (struct texture_sampler *)(++sampler_state);
-    memset(ts, 0, sizeof(*ts));
-    ts->ts0.reverse_gamma = 0;
-    ts->ts0.planar2packet = 0;
-    ts->ts0.color_conversion = 0;
-    ts->ts0.chromakey_index = 0;
-    ts->ts0.base_level = 0;
-    ts->ts0.mip_filter = MIPFILTER_NONE;        /* NONE */
-    ts->ts0.mag_filter = MAPFILTER_LINEAR;      /* LINEAR */
-    ts->ts0.min_filter = MAPFILTER_LINEAR;      /* LINEAR */
-    ts->ts0.lod_bias = 0;       /* 0.0 */
-    ts->ts0.shadow_enable = 0;
-    ts->ts0.max_anisotropy = ANISORATIO_2;
-    ts->ts0.shadow_function = PREFILTEROP_ALWAYS;
-    ts->ts1.min_lod = 0;        /* 0.0 Maximum Mip Level */
-    ts->ts1.kill_pixel = 0;
-    ts->ts1.keyed_texture_filter = 0;
-    ts->ts1.chromakey_enable = 0;
-    ts->ts1.tcx_control = TEXCOORDMODE_CLAMP;
-    ts->ts1.tcy_control = TEXCOORDMODE_CLAMP;
-    ts->ts1.tcz_control = TEXCOORDMODE_CLAMP;
-    ts->ts1.normalized_coor = 0;
-    ts->ts1.map_index = 0;
-    ts->ts1.east_deinterlacer = 0;
-    ts->ts2.default_color = 0;
-
-    /* Sampler 1 */
-    ++ts;
-    memset(ts, 0, sizeof(*ts));
-    ts->ts0.reverse_gamma = 0;
-    ts->ts0.planar2packet = 0;
-    ts->ts0.color_conversion = 0;
-    ts->ts0.chromakey_index = 0;
-    ts->ts0.base_level = 0;
-    ts->ts0.mip_filter = MIPFILTER_NONE;        /* NONE */
-    ts->ts0.mag_filter = MAPFILTER_LINEAR;      /* LINEAR */
-    ts->ts0.min_filter = MAPFILTER_LINEAR;      /* LINEAR */
-    ts->ts0.lod_bias = 0;       /* 0.0 */
-    ts->ts0.shadow_enable = 0;
-    ts->ts0.max_anisotropy = ANISORATIO_2;
-    ts->ts0.shadow_function = PREFILTEROP_ALWAYS;
-    ts->ts1.min_lod = 0;        /* 0.0 Maximum Mip Level */
-    ts->ts1.kill_pixel = 0;
-    ts->ts1.keyed_texture_filter = 0;
-    ts->ts1.chromakey_enable = 0;
-    ts->ts1.tcx_control = TEXCOORDMODE_CLAMP;
-    ts->ts1.tcy_control = TEXCOORDMODE_CLAMP;
-    ts->ts1.tcz_control = TEXCOORDMODE_CLAMP;
-    ts->ts1.normalized_coor = 0;
-    ts->ts1.map_index = 1;
-    ts->ts1.east_deinterlacer = 0;
-    ts->ts2.default_color = 0;
-}
-
 static void i915_inst_arith(unsigned int *inst,
                             unsigned int op,
                             unsigned int dest,
@@ -695,191 +124,245 @@ static void i915_inst_texld(unsigned int *inst,
    *inst = T2_MBZ;
 }
 
-static void i915_mc_pixel_shader_program_buffer(XvMCContext *context)
+static void i915_emit_batch(void *data, int size, int flag)
+{
+    intelBatchbufferData(data, size, flag);
+}
+
+/* one time context initialization buffer */
+static uint32_t *one_time_load_state_imm1;
+static uint32_t *one_time_load_indirect;
+static int one_time_load_state_imm1_size, one_time_load_indirect_size;
+
+/* load indirect buffer for mc rendering */
+static uint32_t *mc_render_load_indirect;
+static int mc_render_load_indirect_size;
+
+static void i915_mc_one_time_context_init(XvMCContext *context)
 {
-    struct i915_3dstate_pixel_shader_program *pixel_shader_program;
-    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
-    unsigned int *inst;
     unsigned int dest, src0, src1, src2;
+    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
+    int i;
+    struct i915_3dstate_sampler_state *sampler_state;
+    struct i915_3dstate_pixel_shader_program *pixel_shader_program;
+    struct i915_3dstate_pixel_shader_constants *pixel_shader_constants;
 
-    /* Shader 0 */
+    /* sampler static state */
+    sampler_state = (struct i915_3dstate_sampler_state *)pI915XvMC->ssb.map;
+    /* pixel shader static state */
     pixel_shader_program = (struct i915_3dstate_pixel_shader_program *)pI915XvMC->psp.map;
+    /* pixel shader contant static state */
+    pixel_shader_constants = (struct i915_3dstate_pixel_shader_constants *)pI915XvMC->psc.map;
+
+    memset(sampler_state, 0, sizeof(*sampler_state));
+    sampler_state->dw0.type = CMD_3D;
+    sampler_state->dw0.opcode = OPC_3DSTATE_SAMPLER_STATE;
+    sampler_state->dw0.length = 6;
+    sampler_state->dw1.sampler_masker = SAMPLER_SAMPLER0 | SAMPLER_SAMPLER1;
+
+    sampler_state->sampler0.ts0.reverse_gamma = 0;
+    sampler_state->sampler0.ts0.planar2packet = 0;
+    sampler_state->sampler0.ts0.color_conversion = 0;
+    sampler_state->sampler0.ts0.chromakey_index = 0;
+    sampler_state->sampler0.ts0.base_level = 0;
+    sampler_state->sampler0.ts0.mip_filter = MIPFILTER_NONE;        /* NONE */
+    sampler_state->sampler0.ts0.mag_filter = MAPFILTER_LINEAR;      /* LINEAR */
+    sampler_state->sampler0.ts0.min_filter = MAPFILTER_LINEAR;      /* LINEAR */
+    sampler_state->sampler0.ts0.lod_bias = 0;       /* 0.0 */
+    sampler_state->sampler0.ts0.shadow_enable = 0;
+    sampler_state->sampler0.ts0.max_anisotropy = ANISORATIO_2;
+    sampler_state->sampler0.ts0.shadow_function = PREFILTEROP_ALWAYS;
+    sampler_state->sampler0.ts1.min_lod = 0;        /* 0.0 Maximum Mip Level */
+    sampler_state->sampler0.ts1.kill_pixel = 0;
+    sampler_state->sampler0.ts1.keyed_texture_filter = 0;
+    sampler_state->sampler0.ts1.chromakey_enable = 0;
+    sampler_state->sampler0.ts1.tcx_control = TEXCOORDMODE_CLAMP;
+    sampler_state->sampler0.ts1.tcy_control = TEXCOORDMODE_CLAMP;
+    sampler_state->sampler0.ts1.tcz_control = TEXCOORDMODE_CLAMP;
+    sampler_state->sampler0.ts1.normalized_coor = 0;
+    sampler_state->sampler0.ts1.map_index = 0;
+    sampler_state->sampler0.ts1.east_deinterlacer = 0;
+    sampler_state->sampler0.ts2.default_color = 0;
+
+    sampler_state->sampler1.ts0.reverse_gamma = 0;
+    sampler_state->sampler1.ts0.planar2packet = 0;
+    sampler_state->sampler1.ts0.color_conversion = 0;
+    sampler_state->sampler1.ts0.chromakey_index = 0;
+    sampler_state->sampler1.ts0.base_level = 0;
+    sampler_state->sampler1.ts0.mip_filter = MIPFILTER_NONE;        /* NONE */
+    sampler_state->sampler1.ts0.mag_filter = MAPFILTER_LINEAR;      /* LINEAR */
+    sampler_state->sampler1.ts0.min_filter = MAPFILTER_LINEAR;      /* LINEAR */
+    sampler_state->sampler1.ts0.lod_bias = 0;       /* 0.0 */
+    sampler_state->sampler1.ts0.shadow_enable = 0;
+    sampler_state->sampler1.ts0.max_anisotropy = ANISORATIO_2;
+    sampler_state->sampler1.ts0.shadow_function = PREFILTEROP_ALWAYS;
+    sampler_state->sampler1.ts1.min_lod = 0;        /* 0.0 Maximum Mip Level */
+    sampler_state->sampler1.ts1.kill_pixel = 0;
+    sampler_state->sampler1.ts1.keyed_texture_filter = 0;
+    sampler_state->sampler1.ts1.chromakey_enable = 0;
+    sampler_state->sampler1.ts1.tcx_control = TEXCOORDMODE_CLAMP;
+    sampler_state->sampler1.ts1.tcy_control = TEXCOORDMODE_CLAMP;
+    sampler_state->sampler1.ts1.tcz_control = TEXCOORDMODE_CLAMP;
+    sampler_state->sampler1.ts1.normalized_coor = 0;
+    sampler_state->sampler1.ts1.map_index = 1;
+    sampler_state->sampler1.ts1.east_deinterlacer = 0;
+    sampler_state->sampler1.ts2.default_color = 0;
+
     memset(pixel_shader_program, 0, sizeof(*pixel_shader_program));
-    pixel_shader_program->dw0.type = CMD_3D;
-    pixel_shader_program->dw0.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
-    pixel_shader_program->dw0.retain = 1;
-    pixel_shader_program->dw0.length = 2;
-    /* mov oC, c0.0000 */
-    inst = (unsigned int*)(++pixel_shader_program);
+    pixel_shader_program->shader0.type = CMD_3D;
+    pixel_shader_program->shader0.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
+    pixel_shader_program->shader0.retain = 1;
+    pixel_shader_program->shader0.length = 2; /* 1 inst */
+    i = 0;
+
     dest = UREG(REG_TYPE_OC, 0);
     src0 = UREG(REG_TYPE_CONST, 0);
     src1 = 0;
     src2 = 0;
-    i915_inst_arith(inst, A0_MOV, dest, A0_DEST_CHANNEL_ALL,
-                    A0_DEST_SATURATE, src0, src1, src2);
-    inst += 3;
-
-    /* Shader 1 */
-    pixel_shader_program = (struct i915_3dstate_pixel_shader_program *)inst;
-    memset(pixel_shader_program, 0, sizeof(*pixel_shader_program));
-    pixel_shader_program->dw0.type = CMD_3D;
-    pixel_shader_program->dw0.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
-    pixel_shader_program->dw0.retain = 1;
-    pixel_shader_program->dw0.length = 14;
+    i915_inst_arith(&pixel_shader_program->inst0[i], A0_MOV,
+	    dest, A0_DEST_CHANNEL_ALL, A0_DEST_SATURATE, src0, src1, src2);
+
+    pixel_shader_program->shader1.type = CMD_3D;
+    pixel_shader_program->shader1.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
+    pixel_shader_program->shader1.retain = 1;
+    pixel_shader_program->shader1.length = 14; /* 5 inst */
+    i = 0;
     /* dcl t0.xy */
-    inst = (unsigned int*)(++pixel_shader_program);
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX0, D0_CHANNEL_XY);
+    i915_inst_decl(&pixel_shader_program->inst1[i], REG_TYPE_T, T_TEX0, D0_CHANNEL_XY);
+    i+=3;
     /* dcl t1.xy */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX1, D0_CHANNEL_XY);
+    i915_inst_decl(&pixel_shader_program->inst1[i], REG_TYPE_T, T_TEX1, D0_CHANNEL_XY);
     /* dcl_2D s0 */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_S, 0, D0_SAMPLE_TYPE_2D);
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst1[i], REG_TYPE_S, 0, D0_SAMPLE_TYPE_2D);
     /* texld r0, t0, s0 */
-    inst += 3;
+    i += 3;
     dest = UREG(REG_TYPE_R, 0);
     src0 = UREG(REG_TYPE_T, 0); /* COORD */
     src1 = UREG(REG_TYPE_S, 0); /* SAMPLER */
-    i915_inst_texld(inst, T0_TEXLD, dest, src0, src1);
+    i915_inst_texld(&pixel_shader_program->inst1[i], T0_TEXLD, dest, src0, src1);
     /* mov oC, r0 */
-    inst += 3;
+    i += 3;
     dest = UREG(REG_TYPE_OC, 0);
     src0 = UREG(REG_TYPE_R, 0);
     src1 = src2 = 0;
-    i915_inst_arith(inst, A0_MOV, dest, A0_DEST_CHANNEL_ALL,
+    i915_inst_arith(&pixel_shader_program->inst1[i], A0_MOV, dest, A0_DEST_CHANNEL_ALL,
                     A0_DEST_SATURATE, src0, src1, src2);
-    inst += 3;
 
-    /* Shader 2 */
-    pixel_shader_program = (struct i915_3dstate_pixel_shader_program *)inst;
-    memset(pixel_shader_program, 0, sizeof(*pixel_shader_program));
-    pixel_shader_program->dw0.type = CMD_3D;
-    pixel_shader_program->dw0.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
-    pixel_shader_program->dw0.retain = 1;
-    pixel_shader_program->dw0.length = 14;
+
+    pixel_shader_program->shader2.type = CMD_3D;
+    pixel_shader_program->shader2.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
+    pixel_shader_program->shader2.retain = 1;
+    pixel_shader_program->shader2.length = 14; /* 5 inst */
+    i = 0;
     /* dcl t2.xy */
-    inst = (unsigned int*)(++pixel_shader_program);
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX2, D0_CHANNEL_XY);
+    i915_inst_decl(&pixel_shader_program->inst2[i], REG_TYPE_T, T_TEX2, D0_CHANNEL_XY);
     /* dcl t3.xy */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX3, D0_CHANNEL_XY);
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst2[i], REG_TYPE_T, T_TEX3, D0_CHANNEL_XY);
     /* dcl_2D s1 */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_S, 1, D0_SAMPLE_TYPE_2D);
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst2[i], REG_TYPE_S, 1, D0_SAMPLE_TYPE_2D);
     /* texld r0, t2, s1 */
-    inst += 3;
+    i += 3;
     dest = UREG(REG_TYPE_R, 0);
     src0 = UREG(REG_TYPE_T, 2); /* COORD */
     src1 = UREG(REG_TYPE_S, 1); /* SAMPLER */
-    i915_inst_texld(inst, T0_TEXLD, dest, src0, src1);
+    i915_inst_texld(&pixel_shader_program->inst2[i], T0_TEXLD, dest, src0, src1);
     /* mov oC, r0 */
-    inst += 3;
+    i += 3;
     dest = UREG(REG_TYPE_OC, 0);
     src0 = UREG(REG_TYPE_R, 0);
     src1 = src2 = 0;
-    i915_inst_arith(inst, A0_MOV, dest, A0_DEST_CHANNEL_ALL,
+    i915_inst_arith(&pixel_shader_program->inst2[i], A0_MOV, dest, A0_DEST_CHANNEL_ALL,
                     A0_DEST_SATURATE, src0, src1, src2);
-    inst += 3;
 
     /* Shader 3 */
-    pixel_shader_program = (struct i915_3dstate_pixel_shader_program *)inst;
-    memset(pixel_shader_program, 0, sizeof(*pixel_shader_program));
-    pixel_shader_program->dw0.type = CMD_3D;
-    pixel_shader_program->dw0.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
-    pixel_shader_program->dw0.retain = 1;
-    pixel_shader_program->dw0.length = 29;
+    pixel_shader_program->shader3.type = CMD_3D;
+    pixel_shader_program->shader3.opcode = OPC_3DSTATE_PIXEL_SHADER_PROGRAM;
+    pixel_shader_program->shader3.retain = 1;
+    pixel_shader_program->shader3.length = 29; /* 10 inst */
+    i = 0;
     /* dcl t0.xy */
-    inst = (unsigned int*)(++pixel_shader_program);
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX0, D0_CHANNEL_XY);
+    i915_inst_decl(&pixel_shader_program->inst3[i], REG_TYPE_T, T_TEX0, D0_CHANNEL_XY);
     /* dcl t1.xy */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX1, D0_CHANNEL_XY);
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst3[i], REG_TYPE_T, T_TEX1, D0_CHANNEL_XY);
     /* dcl t2.xy */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX2, D0_CHANNEL_XY);
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst3[i], REG_TYPE_T, T_TEX2, D0_CHANNEL_XY);
     /* dcl t3.xy */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_T, T_TEX3, D0_CHANNEL_XY);
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst3[i], REG_TYPE_T, T_TEX3, D0_CHANNEL_XY);
     /* dcl_2D s0 */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_S, 0, D0_SAMPLE_TYPE_2D);
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst3[i], REG_TYPE_S, 0, D0_SAMPLE_TYPE_2D);
     /* dcl_2D s1 */
-    inst += 3;
-    i915_inst_decl(inst, REG_TYPE_S, 1, D0_SAMPLE_TYPE_2D);
+    i += 3;
+    i915_inst_decl(&pixel_shader_program->inst3[i], REG_TYPE_S, 1, D0_SAMPLE_TYPE_2D);
     /* texld r0, t0, s0 */
-    inst += 3;
+    i += 3;
     dest = UREG(REG_TYPE_R, 0);
     src0 = UREG(REG_TYPE_T, 0); /* COORD */
     src1 = UREG(REG_TYPE_S, 0); /* SAMPLER */
-    i915_inst_texld(inst, T0_TEXLD, dest, src0, src1);
+    i915_inst_texld(&pixel_shader_program->inst3[i], T0_TEXLD, dest, src0, src1);
     /* texld r1, t2, s1 */
-    inst += 3;
+    i += 3;
     dest = UREG(REG_TYPE_R, 1);
     src0 = UREG(REG_TYPE_T, 2); /* COORD */
     src1 = UREG(REG_TYPE_S, 1); /* SAMPLER */
-    i915_inst_texld(inst, T0_TEXLD, dest, src0, src1);
+    i915_inst_texld(&pixel_shader_program->inst3[i], T0_TEXLD, dest, src0, src1);
     /* add r0, r0, r1 */
-    inst += 3;
+    i += 3;
     dest = UREG(REG_TYPE_R, 0);
     src0 = UREG(REG_TYPE_R, 0);
     src1 = UREG(REG_TYPE_R, 1);
     src2 = 0;
-    i915_inst_arith(inst, A0_ADD, dest, A0_DEST_CHANNEL_ALL,
+    i915_inst_arith(&pixel_shader_program->inst3[i], A0_ADD, dest, A0_DEST_CHANNEL_ALL,
                     0 /* A0_DEST_SATURATE */, src0, src1, src2);
     /* mul oC, r0, c0 */
-    inst += 3;
+    i += 3;
     dest = UREG(REG_TYPE_OC, 0);
     src0 = UREG(REG_TYPE_R, 0);
     src1 = UREG(REG_TYPE_CONST, 0);
     src2 = 0;
-    i915_inst_arith(inst, A0_MUL, dest, A0_DEST_CHANNEL_ALL,
+    i915_inst_arith(&pixel_shader_program->inst3[i], A0_MUL, dest, A0_DEST_CHANNEL_ALL,
                     A0_DEST_SATURATE, src0, src1, src2);
-    inst += 3;
-}
-
-static void i915_mc_pixel_shader_constants_buffer(XvMCContext *context)
-{
-    struct i915_3dstate_pixel_shader_constants *pixel_shader_constants;
-    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
-    float *value;
 
-    pixel_shader_constants = (struct i915_3dstate_pixel_shader_constants *)pI915XvMC->psc.map;
     memset(pixel_shader_constants, 0, sizeof(*pixel_shader_constants));
     pixel_shader_constants->dw0.type = CMD_3D;
     pixel_shader_constants->dw0.opcode = OPC_3DSTATE_PIXEL_SHADER_CONSTANTS;
     pixel_shader_constants->dw0.length = 4;
     pixel_shader_constants->dw1.reg_mask = REG_CR0;
-    value = (float *)(++pixel_shader_constants);
-    *(value++) = 0.5;
-    *(value++) = 0.5;
-    *(value++) = 0.5;
-    *(value++) = 0.5;
+    pixel_shader_constants->value.x = 0.5;
+    pixel_shader_constants->value.y = 0.5;
+    pixel_shader_constants->value.z = 0.5;
+    pixel_shader_constants->value.w = 0.5;
+
 }
 
-static void i915_mc_one_time_state_initialization(XvMCContext *context)
+static void i915_mc_one_time_state_init(XvMCContext *context)
 {
-    struct i915_3dstate_load_state_immediate_1 *load_state_immediate_1 = NULL;
     struct s3_dword *s3 = NULL;
     struct s6_dword *s6 = NULL;
-    struct i915_3dstate_load_indirect *load_indirect = NULL;
     dis_state *dis = NULL;
     ssb_state *ssb = NULL;
     psp_state *psp = NULL;
     psc_state *psc = NULL;
     i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
-    unsigned int size;
-    void *base = NULL;
-    int mem_select = 1;
+    struct i915_3dstate_load_state_immediate_1 *load_state_immediate_1;
+    struct i915_3dstate_load_indirect *load_indirect;
+    int mem_select;
 
     /* 3DSTATE_LOAD_STATE_IMMEDIATE_1 */
-    size = sizeof(*load_state_immediate_1) + sizeof(*s3) + sizeof(*s6);
-    base = calloc(1, size);
-    load_state_immediate_1 = (struct i915_3dstate_load_state_immediate_1 *)base;
+    one_time_load_state_imm1_size = sizeof(*load_state_immediate_1) + sizeof(*s3) + sizeof(*s6);
+    one_time_load_state_imm1 = calloc(1, one_time_load_state_imm1_size);
+    load_state_immediate_1 = (struct i915_3dstate_load_state_immediate_1 *)one_time_load_state_imm1;
     load_state_immediate_1->dw0.type = CMD_3D;
     load_state_immediate_1->dw0.opcode = OPC_3DSTATE_LOAD_STATE_IMMEDIATE_1;
     load_state_immediate_1->dw0.load_s3 = 1;
     load_state_immediate_1->dw0.load_s6 = 1;
-    load_state_immediate_1->dw0.length = (size >> 2) - 2;
+    load_state_immediate_1->dw0.length = (one_time_load_state_imm1_size >> 2) - 2;
 
     s3 = (struct s3_dword *)(++load_state_immediate_1);
     s3->set0_pcd = 1;
@@ -905,33 +388,31 @@ static void i915_mc_one_time_state_initialization(XvMCContext *context)
     s6->color_buffer_write = 1;
     s6->triangle_pv = 0;
 
-    intelBatchbufferData(base, size, 0);
-    free(base);
-
     /* 3DSTATE_LOAD_INDIRECT */
-    size = sizeof(*load_indirect) + sizeof(*dis) + sizeof(*ssb) + sizeof(*psp) + sizeof(*psc);
-    base = calloc(1, size);
-    load_indirect = (struct i915_3dstate_load_indirect *)base;
+    one_time_load_indirect_size = sizeof(*load_indirect) + sizeof(*dis) + sizeof(*ssb) + sizeof(*psp) + sizeof(*psc);
+    one_time_load_indirect = calloc(1, one_time_load_indirect_size);
+    load_indirect = (struct i915_3dstate_load_indirect *)one_time_load_indirect;
     load_indirect->dw0.type = CMD_3D;
     load_indirect->dw0.opcode = OPC_3DSTATE_LOAD_INDIRECT;
     load_indirect->dw0.block_mask = BLOCK_DIS | BLOCK_SSB | BLOCK_PSP | BLOCK_PSC;
-    load_indirect->dw0.length = (size >> 2) - 2;
+    load_indirect->dw0.length = (one_time_load_indirect_size >> 2) - 2;
 
     if (pI915XvMC->deviceID == PCI_CHIP_I915_G ||
-        pI915XvMC->deviceID == PCI_CHIP_I915_GM ||
-        pI915XvMC->deviceID == PCI_CHIP_I945_G ||
-        pI915XvMC->deviceID == PCI_CHIP_I945_GM)
-        mem_select = 0;
+        pI915XvMC->deviceID == PCI_CHIP_I915_GM)
+	mem_select = 0; /* use physical address */
+    else
+	mem_select = 1; /* use gfx address */
 
     load_indirect->dw0.mem_select = mem_select;
 
-    /* DIS */
+
+    /* Dynamic indirect state buffer */
     dis = (dis_state *)(++load_indirect);
     dis->dw0.valid = 0;
     dis->dw0.reset = 0;
     dis->dw0.buffer_address = 0;
 
-    /* SSB */
+    /* Sample state buffer */
     ssb = (ssb_state *)(++dis);
     ssb->dw0.valid = 1;
     ssb->dw0.force = 1;
@@ -940,20 +421,20 @@ static void i915_mc_one_time_state_initialization(XvMCContext *context)
     if (mem_select)
         ssb->dw0.buffer_address = (pI915XvMC->ssb.offset >> 2);
     else
-        ssb->dw0.buffer_address = (pI915XvMC->ssb.bus_addr >> 2);
+	ssb->dw0.buffer_address = (pI915XvMC->ssb.bus_addr >> 2);
 
-    /* PSP */
+    /* Pixel shader program buffer */
     psp = (psp_state *)(++ssb);
     psp->dw0.valid = 1;
     psp->dw0.force = 1;
     psp->dw1.length = 66; /* 4 + 16 + 16 + 31 - 1 */
 
     if (mem_select)
-        psp->dw0.buffer_address = (pI915XvMC->psp.offset >> 2);
+	psp->dw0.buffer_address = (pI915XvMC->psp.offset >> 2);
     else
-        psp->dw0.buffer_address = (pI915XvMC->psp.bus_addr >> 2);
+	psp->dw0.buffer_address = (pI915XvMC->psp.bus_addr >> 2);
 
-    /* PSC */
+    /* Pixel shader constant buffer */
     psc = (psc_state *)(++psp);
     psc->dw0.valid = 1;
     psc->dw0.force = 1;
@@ -963,9 +444,483 @@ static void i915_mc_one_time_state_initialization(XvMCContext *context)
         psc->dw0.buffer_address = (pI915XvMC->psc.offset >> 2);
     else
         psc->dw0.buffer_address = (pI915XvMC->psc.bus_addr >> 2);
+}
 
-    intelBatchbufferData(base, size, 0);
-    free(base);
+static void i915_mc_one_time_state_emit(void)
+{
+    i915_emit_batch(one_time_load_state_imm1, one_time_load_state_imm1_size, 0);
+    i915_emit_batch(one_time_load_indirect, one_time_load_indirect_size, 0);
+}
+
+static void i915_mc_static_indirect_state_init(XvMCContext *context)
+{
+    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
+    struct i915_mc_static_indirect_state_buffer *buffer_info =
+	(struct i915_mc_static_indirect_state_buffer *)pI915XvMC->sis.map;
+
+    memset(buffer_info, 0, sizeof(*buffer_info));
+    /* dest Y */
+    buffer_info->dest_y.dw0.type = CMD_3D;
+    buffer_info->dest_y.dw0.opcode = OPC_3DSTATE_BUFFER_INFO;
+    buffer_info->dest_y.dw0.length = 1;
+    buffer_info->dest_y.dw1.aux_id = 0;
+    buffer_info->dest_y.dw1.buffer_id = BUFFERID_COLOR_BACK;
+    buffer_info->dest_y.dw1.fence_regs = 0;    /* disabled */ /* FIXME: tiled y for performance */
+    buffer_info->dest_y.dw1.tiled_surface = 0; /* linear */
+    buffer_info->dest_y.dw1.walk = TILEWALK_XMAJOR;
+
+    /* dest U */
+    buffer_info->dest_u.dw0.type = CMD_3D;
+    buffer_info->dest_u.dw0.opcode = OPC_3DSTATE_BUFFER_INFO;
+    buffer_info->dest_u.dw0.length = 1;
+    buffer_info->dest_u.dw1.aux_id = 0;
+    buffer_info->dest_u.dw1.buffer_id = BUFFERID_COLOR_AUX;
+    buffer_info->dest_u.dw1.fence_regs = 0;
+    buffer_info->dest_u.dw1.tiled_surface = 0;
+    buffer_info->dest_u.dw1.walk = TILEWALK_XMAJOR;
+
+    /* dest V */
+    buffer_info->dest_v.dw0.type = CMD_3D;
+    buffer_info->dest_v.dw0.opcode = OPC_3DSTATE_BUFFER_INFO;
+    buffer_info->dest_v.dw0.length = 1;
+    buffer_info->dest_v.dw1.aux_id = 1;
+    buffer_info->dest_v.dw1.buffer_id = BUFFERID_COLOR_AUX;
+    buffer_info->dest_v.dw1.fence_regs = 0;
+    buffer_info->dest_v.dw1.tiled_surface = 0;
+    buffer_info->dest_v.dw1.walk = TILEWALK_XMAJOR;
+
+    buffer_info->dest_buf.dw0.type = CMD_3D;
+    buffer_info->dest_buf.dw0.opcode = OPC_3DSTATE_DEST_BUFFER_VARIABLES;
+    buffer_info->dest_buf.dw0.length = 0;
+    buffer_info->dest_buf.dw1.dest_v_bias = 8; /* 0.5 */
+    buffer_info->dest_buf.dw1.dest_h_bias = 8; /* 0.5 */
+    buffer_info->dest_buf.dw1.color_fmt = COLORBUFFER_8BIT;
+    buffer_info->dest_buf.dw1.v_ls = 0; /* fill later */
+    buffer_info->dest_buf.dw1.v_ls_offset = 0; /* fill later */
+
+    buffer_info->dest_buf_mpeg.dw0.type = CMD_3D;
+    buffer_info->dest_buf_mpeg.dw0.opcode = OPC_3DSTATE_DEST_BUFFER_VARIABLES_MPEG;
+    buffer_info->dest_buf_mpeg.dw0.length = 1;
+    buffer_info->dest_buf_mpeg.dw1.decode_mode = MPEG_DECODE_MC;
+    buffer_info->dest_buf_mpeg.dw1.rcontrol = 0;               /* for MPEG-1/MPEG-2 */
+    buffer_info->dest_buf_mpeg.dw1.bidir_avrg_control = 0;     /* for MPEG-1/MPEG-2/MPEG-4 */
+    buffer_info->dest_buf_mpeg.dw1.abort_on_error = 1;
+    buffer_info->dest_buf_mpeg.dw1.intra8 = 0;         /* 16-bit formatted correction data */
+    buffer_info->dest_buf_mpeg.dw1.tff = 1; /* fill later */
+
+    buffer_info->dest_buf_mpeg.dw1.v_subsample_factor = MC_SUB_1V;
+    buffer_info->dest_buf_mpeg.dw1.h_subsample_factor = MC_SUB_1H;
+
+    buffer_info->corr.dw0.type = CMD_3D;
+    buffer_info->corr.dw0.opcode = OPC_3DSTATE_BUFFER_INFO;
+    buffer_info->corr.dw0.length = 1;
+    buffer_info->corr.dw1.aux_id = 0;
+    buffer_info->corr.dw1.buffer_id = BUFFERID_MC_INTRA_CORR;
+    buffer_info->corr.dw1.aux_id = 0;
+    buffer_info->corr.dw1.fence_regs = 0;
+    buffer_info->corr.dw1.tiled_surface = 0;
+    buffer_info->corr.dw1.walk = 0;
+    buffer_info->corr.dw1.pitch = 0;
+    buffer_info->corr.dw2.base_address = (pI915XvMC->corrdata.offset >> 2);  /* starting DWORD address */
+}
+
+static void i915_mc_static_indirect_state_set(XvMCContext *context, XvMCSurface *dest,
+	unsigned int picture_structure, unsigned int flags, unsigned int picture_coding_type)
+{
+    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
+    i915XvMCSurface *pI915Surface = (i915XvMCSurface *)dest->privData;
+    struct i915_mc_static_indirect_state_buffer *buffer_info =
+	(struct i915_mc_static_indirect_state_buffer *)pI915XvMC->sis.map;
+    unsigned int w = dest->width;
+
+    buffer_info->dest_y.dw1.pitch = (pI915Surface->yStride >> 2);      /* in DWords */
+    buffer_info->dest_y.dw2.base_address = (YOFFSET(pI915Surface) >> 2);    /* starting DWORD address */
+    buffer_info->dest_u.dw1.pitch = (pI915Surface->uvStride >> 2);      /* in DWords */
+    buffer_info->dest_u.dw2.base_address = (UOFFSET(pI915Surface) >> 2);      /* starting DWORD address */
+    buffer_info->dest_v.dw1.pitch = (pI915Surface->uvStride >> 2);      /* in Dwords */
+    buffer_info->dest_v.dw2.base_address = (VOFFSET(pI915Surface) >> 2);      /* starting DWORD address */
+
+    if ((picture_structure & XVMC_FRAME_PICTURE) == XVMC_FRAME_PICTURE) {
+        ;
+    } else if ((picture_structure & XVMC_FRAME_PICTURE) == XVMC_TOP_FIELD) {
+        buffer_info->dest_buf.dw1.v_ls = 1;
+    } else if ((picture_structure & XVMC_FRAME_PICTURE) == XVMC_BOTTOM_FIELD) {
+        buffer_info->dest_buf.dw1.v_ls = 1;
+        buffer_info->dest_buf.dw1.v_ls_offset = 1;
+    }
+
+    if (picture_structure & XVMC_FRAME_PICTURE) {
+        ;
+    } else if (picture_structure & XVMC_TOP_FIELD) {
+        if (flags & XVMC_SECOND_FIELD)
+            buffer_info->dest_buf_mpeg.dw1.tff = 0;
+        else
+            buffer_info->dest_buf_mpeg.dw1.tff = 1;
+    } else if (picture_structure & XVMC_BOTTOM_FIELD) {
+        if (flags & XVMC_SECOND_FIELD)
+            buffer_info->dest_buf_mpeg.dw1.tff = 1;
+        else
+            buffer_info->dest_buf_mpeg.dw1.tff = 0;
+    }
+
+    buffer_info->dest_buf_mpeg.dw1.picture_width = (dest->width >> 4);     /* in macroblocks */
+    buffer_info->dest_buf_mpeg.dw2.picture_coding_type = picture_coding_type;
+}
+
+static void i915_mc_map_state_init(XvMCContext *context)
+{
+    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
+    unsigned int w = context->width;
+    unsigned int h = context->height;
+    struct i915_mc_map_state *map_state;
+
+    map_state = (struct i915_mc_map_state *)pI915XvMC->msb.map;
+
+    memset(map_state, 0, sizeof(*map_state));
+
+    /* 3DSATE_MAP_STATE: Y */
+    map_state->y_map.dw0.type = CMD_3D;
+    map_state->y_map.dw0.opcode = OPC_3DSTATE_MAP_STATE;
+    map_state->y_map.dw0.retain = 1;
+    map_state->y_map.dw0.length = 6;
+    map_state->y_map.dw1.map_mask = MAP_MAP0 | MAP_MAP1;
+
+    /* Y Forward (Past) */
+    map_state->y_forward.tm0.v_ls_offset = 0;
+    map_state->y_forward.tm0.v_ls = 0;
+    map_state->y_forward.tm1.tile_walk = TILEWALK_XMAJOR;
+    map_state->y_forward.tm1.tiled_surface = 0;
+    map_state->y_forward.tm1.utilize_fence_regs = 0;
+    map_state->y_forward.tm1.texel_fmt = 0;      /* 8bit */
+    map_state->y_forward.tm1.surface_fmt = 1;    /* 8bit */
+    map_state->y_forward.tm1.width = w - 1;
+    map_state->y_forward.tm1.height = h - 1;
+    map_state->y_forward.tm2.depth = 0;
+    map_state->y_forward.tm2.max_lod = 0;
+    map_state->y_forward.tm2.cube_face = 0;
+
+    /* Y Backward (Future) */
+    map_state->y_backward.tm0.v_ls_offset = 0;
+    map_state->y_backward.tm0.v_ls = 0;
+    map_state->y_backward.tm1.tile_walk = TILEWALK_XMAJOR;
+    map_state->y_backward.tm1.tiled_surface = 0;
+    map_state->y_backward.tm1.utilize_fence_regs = 0;
+    map_state->y_backward.tm1.texel_fmt = 0;      /* 8bit */
+    map_state->y_backward.tm1.surface_fmt = 1;    /* 8bit */
+    map_state->y_backward.tm1.width = w - 1;
+    map_state->y_backward.tm1.height = h - 1;
+    map_state->y_backward.tm2.depth = 0;
+    map_state->y_backward.tm2.max_lod = 0;
+    map_state->y_backward.tm2.cube_face = 0;
+
+    /* 3DSATE_MAP_STATE: U */
+    map_state->u_map.dw0.type = CMD_3D;
+    map_state->u_map.dw0.opcode = OPC_3DSTATE_MAP_STATE;
+    map_state->u_map.dw0.retain = 1;
+    map_state->u_map.dw0.length = 6;
+    map_state->u_map.dw1.map_mask = MAP_MAP0 | MAP_MAP1;
+
+    /* U Forward */
+    map_state->u_forward.tm0.v_ls_offset = 0;
+    map_state->u_forward.tm0.v_ls = 0;
+    map_state->u_forward.tm1.tile_walk = TILEWALK_XMAJOR;
+    map_state->u_forward.tm1.tiled_surface = 0;
+    map_state->u_forward.tm1.utilize_fence_regs = 0;
+    map_state->u_forward.tm1.texel_fmt = 0;      /* 8bit */
+    map_state->u_forward.tm1.surface_fmt = 1;    /* 8bit */
+    map_state->u_forward.tm1.width = (w >> 1) - 1;
+    map_state->u_forward.tm1.height = (h >> 1) - 1;
+    map_state->u_forward.tm2.depth = 0;
+    map_state->u_forward.tm2.max_lod = 0;
+    map_state->u_forward.tm2.cube_face = 0;
+
+    /* U Backward */
+    map_state->u_backward.tm0.v_ls_offset = 0;
+    map_state->u_backward.tm0.v_ls = 0;
+    map_state->u_backward.tm1.tile_walk = TILEWALK_XMAJOR;
+    map_state->u_backward.tm1.tiled_surface = 0;
+    map_state->u_backward.tm1.utilize_fence_regs = 0;
+    map_state->u_backward.tm1.texel_fmt = 0;
+    map_state->u_backward.tm1.surface_fmt = 1;
+    map_state->u_backward.tm1.width = (w >> 1) - 1;
+    map_state->u_backward.tm1.height = (h >> 1) - 1;
+    map_state->u_backward.tm2.depth = 0;
+    map_state->u_backward.tm2.max_lod = 0;
+    map_state->u_backward.tm2.cube_face = 0;
+
+    /* 3DSATE_MAP_STATE: V */
+    map_state->v_map.dw0.type = CMD_3D;
+    map_state->v_map.dw0.opcode = OPC_3DSTATE_MAP_STATE;
+    map_state->v_map.dw0.retain = 1;
+    map_state->v_map.dw0.length = 6;
+    map_state->v_map.dw1.map_mask = MAP_MAP0 | MAP_MAP1;
+
+    /* V Forward */
+    map_state->v_forward.tm0.v_ls_offset = 0;
+    map_state->v_forward.tm0.v_ls = 0;
+    map_state->v_forward.tm1.tile_walk = TILEWALK_XMAJOR;
+    map_state->v_forward.tm1.tiled_surface = 0;
+    map_state->v_forward.tm1.utilize_fence_regs = 0;
+    map_state->v_forward.tm1.texel_fmt = 0;
+    map_state->v_forward.tm1.surface_fmt = 1;
+    map_state->v_forward.tm1.width = (w >> 1) - 1;
+    map_state->v_forward.tm1.height = (h >> 1) - 1;
+    map_state->v_forward.tm2.depth = 0;
+    map_state->v_forward.tm2.max_lod = 0;
+    map_state->v_forward.tm2.cube_face = 0;
+
+    /* V Backward */
+    map_state->v_backward.tm0.v_ls_offset = 0;
+    map_state->v_backward.tm0.v_ls = 0;
+    map_state->v_backward.tm1.tile_walk = TILEWALK_XMAJOR;
+    map_state->v_backward.tm1.tiled_surface = 0;
+    map_state->v_backward.tm1.utilize_fence_regs = 0;
+    map_state->v_backward.tm1.texel_fmt = 0;
+    map_state->v_backward.tm1.surface_fmt = 1;
+    map_state->v_backward.tm1.width = (w >> 1) - 1;
+    map_state->v_backward.tm1.height = (h >> 1) - 1;
+    map_state->v_backward.tm2.depth = 0;
+    map_state->v_backward.tm2.max_lod = 0;
+    map_state->v_backward.tm2.cube_face = 0;
+}
+
+static void i915_mc_map_state_set(XvMCContext *context,
+	i915XvMCSurface *privPast,
+	i915XvMCSurface *privFuture)
+{
+    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
+    struct i915_mc_map_state *map_state;
+
+    map_state = (struct i915_mc_map_state *)pI915XvMC->msb.map;
+
+    map_state->y_forward.tm0.base_address = (YOFFSET(privPast) >> 2);
+    map_state->y_forward.tm2.pitch = (privPast->yStride >> 2) - 1;       /* in DWords - 1 */
+    map_state->y_backward.tm0.base_address = (YOFFSET(privFuture) >> 2);
+    map_state->y_backward.tm2.pitch = (privFuture->yStride >> 2) - 1;
+    map_state->u_forward.tm0.base_address = (UOFFSET(privPast) >> 2);
+    map_state->u_forward.tm2.pitch = (privPast->uvStride >> 2) - 1;       /* in DWords - 1 */
+    map_state->u_backward.tm0.base_address = (UOFFSET(privFuture) >> 2);
+    map_state->u_backward.tm2.pitch = (privFuture->uvStride >> 2) - 1;
+    map_state->v_forward.tm0.base_address = (VOFFSET(privPast) >> 2);
+    map_state->v_forward.tm2.pitch = (privPast->uvStride >> 2) - 1;       /* in DWords - 1 */
+    map_state->v_backward.tm0.base_address = (VOFFSET(privFuture) >> 2);
+    map_state->v_backward.tm2.pitch = (privFuture->uvStride >> 2) - 1;
+}
+
+static void i915_flush(int map, int render)
+{
+    struct i915_mi_flush mi_flush;
+
+    memset(&mi_flush, 0, sizeof(mi_flush));
+    mi_flush.dw0.type = CMD_MI;
+    mi_flush.dw0.opcode = OPC_MI_FLUSH;
+    mi_flush.dw0.map_cache_invalidate = map;
+    mi_flush.dw0.render_cache_flush_inhibit = render;
+
+    intelBatchbufferData(&mi_flush, sizeof(mi_flush), 0);
+}
+
+static void i915_mc_load_indirect_render_init(XvMCContext *context)
+{
+    i915XvMCContext *pI915XvMC = (i915XvMCContext *)context->privData;
+    sis_state *sis;
+    msb_state *msb;
+    struct i915_3dstate_load_indirect *load_indirect;
+    int mem_select;
+
+    mc_render_load_indirect_size = sizeof(*load_indirect) + sizeof(*sis)
+					+ sizeof(*msb);
+    mc_render_load_indirect = calloc(1, mc_render_load_indirect_size);
+
+    load_indirect = (struct i915_3dstate_load_indirect *)mc_render_load_indirect;
+    load_indirect->dw0.type = CMD_3D;
+    load_indirect->dw0.opcode = OPC_3DSTATE_LOAD_INDIRECT;
+    load_indirect->dw0.block_mask = BLOCK_SIS | BLOCK_MSB;
+    load_indirect->dw0.length = (mc_render_load_indirect_size >> 2) - 2;
+
+    if (pI915XvMC->deviceID == PCI_CHIP_I915_G ||
+        pI915XvMC->deviceID == PCI_CHIP_I915_GM)
+        mem_select = 0;
+    else
+	mem_select = 1;
+
+    load_indirect->dw0.mem_select = mem_select;
+
+    /* Static Indirect state buffer (dest buffer info) */
+    sis = (sis_state *)(++load_indirect);
+    sis->dw0.valid = 1;
+    sis->dw0.force = 1;
+    sis->dw1.length = 16; /* 4 * 3 + 2 + 3 - 1 */
+
+    if (mem_select)
+	sis->dw0.buffer_address = (pI915XvMC->sis.offset >> 2);
+    else
+	sis->dw0.buffer_address = (pI915XvMC->sis.bus_addr >> 2);
+
+    /* Map state buffer (reference buffer info) */
+    msb = (msb_state *)(++sis);
+    msb->dw0.valid = 1;
+    msb->dw0.force = 1;
+    msb->dw1.length = 23; /* 3 * 8 - 1 */
+
+    if (mem_select)
+	msb->dw0.buffer_address = (pI915XvMC->msb.offset >> 2);
+    else
+	msb->dw0.buffer_address = (pI915XvMC->msb.bus_addr >> 2);
+}
+
+static void i915_mc_load_indirect_render_emit(void)
+{
+    i915_emit_batch(mc_render_load_indirect, mc_render_load_indirect_size, 0);
+}
+
+static void i915_mc_mpeg_set_origin(XvMCContext *context, XvMCMacroBlock *mb)
+{
+    struct i915_3dmpeg_set_origin set_origin;
+
+    /* 3DMPEG_SET_ORIGIN */
+    memset(&set_origin, 0, sizeof(set_origin));
+    set_origin.dw0.type = CMD_3D;
+    set_origin.dw0.opcode = OPC_3DMPEG_SET_ORIGIN;
+    set_origin.dw0.length = 0;
+    set_origin.dw1.h_origin = mb->x;
+    set_origin.dw1.v_origin = mb->y;
+
+    intelBatchbufferData(&set_origin, sizeof(set_origin), 0);
+}
+
+static void i915_mc_mpeg_macroblock_ipicture(XvMCContext *context, XvMCMacroBlock *mb)
+{
+    struct i915_3dmpeg_macroblock_ipicture macroblock_ipicture;
+
+    /* 3DMPEG_MACROBLOCK_IPICTURE */
+    memset(&macroblock_ipicture, 0, sizeof(macroblock_ipicture));
+    macroblock_ipicture.dw0.type = CMD_3D;
+    macroblock_ipicture.dw0.opcode = OPC_3DMPEG_MACROBLOCK_IPICTURE;
+    macroblock_ipicture.dw0.dct_type = (mb->dct_type == XVMC_DCT_TYPE_FIELD);
+
+    intelBatchbufferData(&macroblock_ipicture, sizeof(macroblock_ipicture), 0);
+}
+
+#if 0
+static void i915_mc_mpeg_macroblock_0mv(XvMCContext *context, XvMCMacroBlock *mb)
+{
+    struct i915_3dmpeg_macroblock_0mv macroblock_0mv;
+
+    /* 3DMPEG_MACROBLOCK(0mv) */
+    memset(&macroblock_0mv, 0, sizeof(macroblock_0mv));
+    macroblock_0mv.header.dw0.type = CMD_3D;
+    macroblock_0mv.header.dw0.opcode = OPC_3DMPEG_MACROBLOCK;
+    macroblock_0mv.header.dw0.length = 0;
+    macroblock_0mv.header.dw1.mb_intra = 1;     /* should be 1 */
+    macroblock_0mv.header.dw1.forward = 0;      /* should be 0 */
+    macroblock_0mv.header.dw1.backward = 0;     /* should be 0 */
+    macroblock_0mv.header.dw1.h263_4mv = 0;     /* should be 0 */
+    macroblock_0mv.header.dw1.dct_type = (mb->dct_type == XVMC_DCT_TYPE_FIELD);
+
+/*
+    if (!mb->coded_block_pattern)
+        macroblock_0mv.header.dw1.dct_type = XVMC_DCT_TYPE_FRAME;
+*/
+
+    macroblock_0mv.header.dw1.motion_type = 0; // (mb->motion_type & 0x3);
+    macroblock_0mv.header.dw1.vertical_field_select = 0; // mb->motion_vertical_field_select & 0xf;
+    macroblock_0mv.header.dw1.coded_block_pattern = mb->coded_block_pattern;
+    macroblock_0mv.header.dw1.skipped_macroblocks = 0;
+
+    intelBatchbufferData(&macroblock_0mv, sizeof(macroblock_0mv), 0);
+}
+#endif
+
+static void i915_mc_mpeg_macroblock_1fbmv(XvMCContext *context, XvMCMacroBlock *mb)
+{
+    struct i915_3dmpeg_macroblock_1fbmv macroblock_1fbmv;
+    vector_t mv0[2];
+
+    /* 3DMPEG_MACROBLOCK(1fbmv) */
+    memset(&macroblock_1fbmv, 0, sizeof(macroblock_1fbmv));
+    macroblock_1fbmv.header.dw0.type = CMD_3D;
+    macroblock_1fbmv.header.dw0.opcode = OPC_3DMPEG_MACROBLOCK;
+    macroblock_1fbmv.header.dw0.length = 2;
+    macroblock_1fbmv.header.dw1.mb_intra = 0;   /* should be 0 */
+    macroblock_1fbmv.header.dw1.forward = ((mb->macroblock_type & XVMC_MB_TYPE_MOTION_FORWARD) ? 1 : 0);
+    macroblock_1fbmv.header.dw1.backward = ((mb->macroblock_type & XVMC_MB_TYPE_MOTION_BACKWARD) ? 1 : 0);
+    macroblock_1fbmv.header.dw1.h263_4mv = 0;   /* should be 0 */
+    macroblock_1fbmv.header.dw1.dct_type = (mb->dct_type == XVMC_DCT_TYPE_FIELD);
+
+    if (!(mb->coded_block_pattern & 0x3f))
+        macroblock_1fbmv.header.dw1.dct_type = XVMC_DCT_TYPE_FRAME;
+
+    macroblock_1fbmv.header.dw1.motion_type = (mb->motion_type & 0x03);
+    macroblock_1fbmv.header.dw1.vertical_field_select = (mb->motion_vertical_field_select & 0x0f);
+    macroblock_1fbmv.header.dw1.coded_block_pattern = mb->coded_block_pattern;
+    macroblock_1fbmv.header.dw1.skipped_macroblocks = 0;
+
+    mv0[0].component[0] = mb->PMV[0][0][0];
+    mv0[0].component[1] = mb->PMV[0][0][1];
+    mv0[1].component[0] = mb->PMV[0][1][0];
+    mv0[1].component[1] = mb->PMV[0][1][1];
+
+    macroblock_1fbmv.dw2 = mv0[0].v;
+    macroblock_1fbmv.dw3 = mv0[1].v;
+
+    intelBatchbufferData(&macroblock_1fbmv, sizeof(macroblock_1fbmv), 0);
+}
+
+static void i915_mc_mpeg_macroblock_2fbmv(XvMCContext *context, XvMCMacroBlock *mb, unsigned int ps)
+{
+    struct i915_3dmpeg_macroblock_2fbmv macroblock_2fbmv;
+    vector_t mv0[2];
+    vector_t mv1[2];
+
+    /* 3DMPEG_MACROBLOCK(2fbmv) */
+    memset(&macroblock_2fbmv, 0, sizeof(macroblock_2fbmv));
+    macroblock_2fbmv.header.dw0.type = CMD_3D;
+    macroblock_2fbmv.header.dw0.opcode = OPC_3DMPEG_MACROBLOCK;
+    macroblock_2fbmv.header.dw0.length = 4;
+    macroblock_2fbmv.header.dw1.mb_intra = 0;   /* should be 0 */
+    macroblock_2fbmv.header.dw1.forward = ((mb->macroblock_type & XVMC_MB_TYPE_MOTION_FORWARD) ? 1 : 0);
+    macroblock_2fbmv.header.dw1.backward = ((mb->macroblock_type & XVMC_MB_TYPE_MOTION_BACKWARD) ? 1 : 0);
+    macroblock_2fbmv.header.dw1.h263_4mv = 0;   /* should be 0 */
+    macroblock_2fbmv.header.dw1.dct_type = (mb->dct_type == XVMC_DCT_TYPE_FIELD);
+
+    if (!(mb->coded_block_pattern & 0x3f))
+        macroblock_2fbmv.header.dw1.dct_type = XVMC_DCT_TYPE_FRAME;
+
+    macroblock_2fbmv.header.dw1.motion_type = (mb->motion_type & 0x03);
+    macroblock_2fbmv.header.dw1.vertical_field_select = (mb->motion_vertical_field_select & 0x0f);
+    macroblock_2fbmv.header.dw1.coded_block_pattern = mb->coded_block_pattern;
+    macroblock_2fbmv.header.dw1.skipped_macroblocks = 0;
+
+    mv0[0].component[0] = mb->PMV[0][0][0];
+    mv0[0].component[1] = mb->PMV[0][0][1];
+    mv0[1].component[0] = mb->PMV[0][1][0];
+    mv0[1].component[1] = mb->PMV[0][1][1];
+    mv1[0].component[0] = mb->PMV[1][0][0];
+    mv1[0].component[1] = mb->PMV[1][0][1];
+    mv1[1].component[0] = mb->PMV[1][1][0];
+    mv1[1].component[1] = mb->PMV[1][1][1];
+
+    if ((ps & XVMC_FRAME_PICTURE) == XVMC_FRAME_PICTURE) {
+        if ((mb->motion_type & 3) == XVMC_PREDICTION_FIELD) {
+            mv0[0].component[1] = mb->PMV[0][0][1] >> 1;
+            mv0[1].component[1] = mb->PMV[0][1][1] >> 1;
+            mv1[0].component[1] = mb->PMV[1][0][1] >> 1;
+            mv1[1].component[1] = mb->PMV[1][1][1] >> 1;
+        } else if ((mb->motion_type & 3) == XVMC_PREDICTION_DUAL_PRIME) {
+            mv0[0].component[1] = mb->PMV[0][0][1] >> 1;
+            mv0[1].component[1] = mb->PMV[0][1][1] >> 1;  // MPEG2 MV[0][1] isn't used
+            mv1[0].component[1] = mb->PMV[1][0][1] >> 1;
+            mv1[1].component[1] = mb->PMV[1][1][1] >> 1;
+        }
+    }
+
+    macroblock_2fbmv.dw2 = mv0[0].v;
+    macroblock_2fbmv.dw3 = mv0[1].v;
+    macroblock_2fbmv.dw4 = mv1[0].v;
+    macroblock_2fbmv.dw5 = mv1[1].v;
+
+    intelBatchbufferData(&macroblock_2fbmv, sizeof(macroblock_2fbmv), 0);
 }
 
 #if 0
@@ -1646,9 +1601,7 @@ static Status i915_xvmc_mc_create_context(Display *display, XvMCContext *context
     pI915XvMC->psc.size = tmpComm->psc.size;
 
     if (pI915XvMC->deviceID == PCI_CHIP_I915_G ||
-        pI915XvMC->deviceID == PCI_CHIP_I915_GM ||
-        pI915XvMC->deviceID == PCI_CHIP_I945_G ||
-        pI915XvMC->deviceID == PCI_CHIP_I945_GM) {
+        pI915XvMC->deviceID == PCI_CHIP_I915_GM) {
         pI915XvMC->sis.bus_addr = tmpComm->sis.bus_addr;
         pI915XvMC->ssb.bus_addr = tmpComm->ssb.bus_addr;
         pI915XvMC->msb.bus_addr = tmpComm->msb.bus_addr;
@@ -1683,6 +1636,16 @@ static Status i915_xvmc_mc_create_context(Display *display, XvMCContext *context
     pI915XvMC->last_flip = 0;
     pI915XvMC->port = context->port;
     pI915XvMC->ref = 1;
+
+    /* pre-init state buffers */
+    i915_mc_one_time_context_init(context);
+    i915_mc_one_time_state_init(context);
+
+    i915_mc_static_indirect_state_init(context);
+
+    i915_mc_map_state_init(context);
+
+    i915_mc_load_indirect_render_init(context);
     return Success;
 }
 
@@ -1695,6 +1658,10 @@ static int i915_xvmc_mc_destroy_context(Display *display, XvMCContext *context)
 
     /* Pass Control to the X server to destroy the drm_context_t */
     i915_release_resource(display,context);
+
+    free(one_time_load_state_imm1);
+    free(one_time_load_indirect);
+    free(mc_render_load_indirect);
     return Success;
 }
 
@@ -1850,9 +1817,8 @@ static int i915_xvmc_mc_render_surface(Display *display, XvMCContext *context,
     if (!(privTarget = target_surface->privData))
         return XvMCBadSurface;
 
-    /* Test For YV12 Surface */
-    if (context->surface_type_id != FOURCC_YV12) {
-        XVMC_ERR("HWMC only possible on YV12 Surfaces.");
+    if (context->surface_type_id >= SURFACE_TYPE_MAX) {
+        XVMC_ERR("Unsupprted surface_type_id %d.", context->surface_type_id);
         return BadValue;
     }
 
@@ -1938,16 +1904,15 @@ static int i915_xvmc_mc_render_surface(Display *display, XvMCContext *context,
     // i915_mc_invalidate_subcontext_buffers(context, BLOCK_SIS | BLOCK_DIS | BLOCK_SSB
     // | BLOCK_MSB | BLOCK_PSP | BLOCK_PSC);
 
-    i915_mc_sampler_state_buffer(context);
-    i915_mc_pixel_shader_program_buffer(context);
-    i915_mc_pixel_shader_constants_buffer(context);
-    i915_mc_one_time_state_initialization(context);
+    i915_mc_one_time_state_emit();
+
+    i915_mc_static_indirect_state_set(context, target_surface, picture_structure,
+	    flags, picture_coding_type);
+    /* setup reference surfaces */
+    i915_mc_map_state_set(context, privPast, privFuture);
+
+    i915_mc_load_indirect_render_emit();
 
-    i915_mc_static_indirect_state_buffer(context, target_surface,
-                                         picture_structure, flags,
-                                         picture_coding_type);
-    i915_mc_map_state_buffer(context, privTarget, privPast, privFuture);
-    i915_mc_load_sis_msb_buffers(context);
     i915_mc_mpeg_set_origin(context, &macroblock_array->macro_blocks[first_macroblock]);
 
     for (i = first_macroblock; i < (num_macroblocks + first_macroblock); i++) {
diff --git a/src/xvmc/intel_xvmc.c b/src/xvmc/intel_xvmc.c
index ae357aa5..8fabb35e 100644
--- a/src/xvmc/intel_xvmc.c
+++ b/src/xvmc/intel_xvmc.c
@@ -98,9 +98,17 @@ unsigned int mb_bytes_420[] = {
     768  /* 111111 */
 };
 
+int DEBUG;
+
 static int error_base;
 static int event_base;
 
+static void intel_xvmc_debug_init(void)
+{
+    if (getenv("INTEL_XVMC_DEBUG"))
+	DEBUG = 1;
+}
+
 /* locking */
 static void intel_xvmc_try_heavy_lock(drm_context_t ctx)
 {
@@ -282,6 +290,8 @@ Status XvMCCreateContext(Display *display, XvPortID port,
         return BadValue;
     }
 
+    intel_xvmc_debug_init();
+
     /* Open DRI Device */
     if((fd = drmOpen("i915", NULL)) < 0) {
         XVMC_ERR("DRM Device could not be opened.");
@@ -588,7 +598,7 @@ Status XvMCCreateBlocks(Display *display, XvMCContext *context,
  */
 Status XvMCDestroyBlocks(Display *display, XvMCBlockArray *block)
 {
-    if (!display || block)
+    if (!display || !block)
         return BadValue;
 
     if (block->blocks)
diff --git a/src/xvmc/intel_xvmc.h b/src/xvmc/intel_xvmc.h
index 31196238..c4dfd770 100644
--- a/src/xvmc/intel_xvmc.h
+++ b/src/xvmc/intel_xvmc.h
@@ -38,6 +38,7 @@
 #include <string.h>
 #include <assert.h>
 #include <signal.h>
+#include <stdint.h>
 
 #include <xf86drm.h>
 #include "i830_common.h"
@@ -56,7 +57,7 @@
 
 #include "intel_batchbuffer.h"
 
-#define DEBUG 0
+extern int DEBUG;
 
 #define XVMC_ERR(s, arg...)					\
     do {							\
diff --git a/src/xvmc/intel_xvmc_dump.c b/src/xvmc/intel_xvmc_dump.c
index 12fc52a5..419bd0d8 100644
--- a/src/xvmc/intel_xvmc_dump.c
+++ b/src/xvmc/intel_xvmc_dump.c
@@ -116,18 +116,25 @@ void intel_xvmc_dump_render(XvMCContext *context, unsigned int picture_structure
 	    fprintf(fp, "intra ");
 	fprintf(fp, ")  ");
 	fprintf(fp, "mc type ");
-	if (mb->motion_type & XVMC_PREDICTION_FIELD)
-	    fprintf(fp, "(field)  ");
-	else if (mb->motion_type & XVMC_PREDICTION_FRAME)
-	    fprintf(fp, "(frame)  ");
-	else if (mb->motion_type & XVMC_PREDICTION_DUAL_PRIME)
-	    fprintf(fp, "(dual-prime)  ");
-	else if (mb->motion_type & XVMC_PREDICTION_16x8)
-	    fprintf(fp, "(16x8)  ");
-	else if (mb->motion_type & XVMC_PREDICTION_4MV)
-	    fprintf(fp, "(4MV)  ");
-	else
-	    fprintf(fp, "(none)  ");
+	if (picture_structure == XVMC_FRAME_PICTURE) {
+	    if (mb->motion_type & XVMC_PREDICTION_FIELD)
+		fprintf(fp, "(field)  ");
+	    else if (mb->motion_type & XVMC_PREDICTION_FRAME)
+		fprintf(fp, "(frame)  ");
+	    else if (mb->motion_type & XVMC_PREDICTION_DUAL_PRIME)
+		fprintf(fp, "(dual-prime)  ");
+	    else
+		fprintf(fp, "(unknown %d)  ", mb->motion_type);
+	} else { /* field */
+	    if (mb->motion_type & XVMC_PREDICTION_FIELD)
+		fprintf(fp, "(field)  ");
+	    else if (mb->motion_type & XVMC_PREDICTION_DUAL_PRIME)
+		fprintf(fp, "(dual-prime)  ");
+	    else if (mb->motion_type & XVMC_PREDICTION_16x8)
+		fprintf(fp, "(16x8)  ");
+	    else
+		fprintf(fp, "(unknown %d)  ", mb->motion_type);
+	}
 
 	if (mb->dct_type == XVMC_DCT_TYPE_FRAME)
 	    fprintf(fp, "dct type (frame)  ");