[g3dvl] give mv their own vertex buffer back

10 files changed, 246 insertions, 151 deletions

diff --git a/src/gallium/auxiliary/vl/vl_defines.h b/src/gallium/auxiliary/vl/vl_defines.h
index 668991f904f..7568db027e6 100644
--- a/src/gallium/auxiliary/vl/vl_defines.h
+++ b/src/gallium/auxiliary/vl/vl_defines.h
@@ -35,4 +35,7 @@
 #define BLOCK_WIDTH 8
 #define BLOCK_HEIGHT 8
 
+#define VL_MAX_PLANES 3
+#define VL_MAX_REF_FRAMES 2
+
 #endif
diff --git a/src/gallium/auxiliary/vl/vl_mc.c b/src/gallium/auxiliary/vl/vl_mc.c
index 707a4a27077..2624d0502c1 100644
--- a/src/gallium/auxiliary/vl/vl_mc.c
+++ b/src/gallium/auxiliary/vl/vl_mc.c
@@ -612,17 +612,12 @@ prepare_pipe_4_rendering(struct vl_mc_buffer *buffer)
 }
 
 void
-vl_mc_render_ref(struct vl_mc_buffer *buffer, struct pipe_sampler_view *ref,
-                 unsigned not_empty_start_instance, unsigned not_empty_num_instances,
-                 unsigned empty_start_instance, unsigned empty_num_instances)
+vl_mc_render_ref(struct vl_mc_buffer *buffer, struct pipe_sampler_view *ref)
 {
    struct vl_mc *renderer;
 
    assert(buffer && ref);
 
-   if (not_empty_num_instances == 0 && empty_num_instances == 0)
-      return;
-
    prepare_pipe_4_rendering(buffer);
 
    renderer = buffer->renderer;
@@ -633,24 +628,19 @@ vl_mc_render_ref(struct vl_mc_buffer *buffer, struct pipe_sampler_view *ref,
    renderer->pipe->set_fragment_sampler_views(renderer->pipe, 1, &ref);
    renderer->pipe->bind_fragment_sampler_states(renderer->pipe, 1, &renderer->sampler_ref);
 
-   if (not_empty_num_instances > 0)
-      util_draw_arrays_instanced(renderer->pipe, PIPE_PRIM_QUADS, 0, 4,
-                                 not_empty_start_instance, not_empty_num_instances);
-
-   if (empty_num_instances > 0)
-      util_draw_arrays_instanced(renderer->pipe, PIPE_PRIM_QUADS, 0, 4,
-                                 empty_start_instance, empty_num_instances);
+   util_draw_arrays_instanced(renderer->pipe, PIPE_PRIM_QUADS, 0, 4, 0,
+                              renderer->buffer_width / MACROBLOCK_WIDTH *
+                              renderer->buffer_height / MACROBLOCK_HEIGHT);
 }
 
 void
-vl_mc_render_ycbcr(struct vl_mc_buffer *buffer,
-                   unsigned not_empty_start_instance, unsigned not_empty_num_instances)
+vl_mc_render_ycbcr(struct vl_mc_buffer *buffer, unsigned num_instances)
 {
    struct vl_mc *renderer;
 
    assert(buffer);
 
-   if (not_empty_num_instances == 0)
+   if (num_instances == 0)
       return;
 
    prepare_pipe_4_rendering(buffer);
@@ -663,6 +653,5 @@ vl_mc_render_ycbcr(struct vl_mc_buffer *buffer,
    renderer->pipe->set_fragment_sampler_views(renderer->pipe, 1, &buffer->source);
    renderer->pipe->bind_fragment_sampler_states(renderer->pipe, 1, &renderer->sampler_ycbcr);
 
-   util_draw_arrays_instanced(renderer->pipe, PIPE_PRIM_QUADS, 0, 4,
-                              not_empty_start_instance, not_empty_num_instances);
+   util_draw_arrays_instanced(renderer->pipe, PIPE_PRIM_QUADS, 0, 4, 0, num_instances);
 }
diff --git a/src/gallium/auxiliary/vl/vl_mc.h b/src/gallium/auxiliary/vl/vl_mc.h
index e5b16b5b9da..bc2b0e7f149 100644
--- a/src/gallium/auxiliary/vl/vl_mc.h
+++ b/src/gallium/auxiliary/vl/vl_mc.h
@@ -75,11 +75,8 @@ void vl_mc_cleanup_buffer(struct vl_mc_buffer *buffer);
 
 void vl_mc_set_surface(struct vl_mc_buffer *buffer, struct pipe_surface *surface);
 
-void vl_mc_render_ref(struct vl_mc_buffer *buffer, struct pipe_sampler_view *ref,
-                      unsigned not_empty_start_instance, unsigned not_empty_num_instances,
-                      unsigned empty_start_instance, unsigned empty_num_instances);
+void vl_mc_render_ref(struct vl_mc_buffer *buffer, struct pipe_sampler_view *ref);
 
-void vl_mc_render_ycbcr(struct vl_mc_buffer *buffer,
-                        unsigned not_empty_start_instance, unsigned not_empty_num_instances);
+void vl_mc_render_ycbcr(struct vl_mc_buffer *buffer, unsigned num_instances);
 
 #endif /* vl_mc_h */
diff --git a/src/gallium/auxiliary/vl/vl_mpeg12_decoder.c b/src/gallium/auxiliary/vl/vl_mpeg12_decoder.c
index 6d10cd16e32..1d8f0f92c37 100644
--- a/src/gallium/auxiliary/vl/vl_mpeg12_decoder.c
+++ b/src/gallium/auxiliary/vl/vl_mpeg12_decoder.c
@@ -293,6 +293,7 @@ vl_mpeg12_destroy(struct pipe_video_decoder *decoder)
       dec->pipe->delete_vertex_elements_state(dec->pipe, dec->ves_mv[i]);
 
    pipe_resource_reference(&dec->quads.buffer, NULL);
+   pipe_resource_reference(&dec->pos.buffer, NULL);
 
    FREE(dec);
 }
@@ -389,15 +390,9 @@ vl_mpeg12_create_buffer(struct pipe_video_decoder *decoder)
    buffer->base.add_macroblocks = vl_mpeg12_buffer_add_macroblocks;
    buffer->base.unmap = vl_mpeg12_buffer_unmap;
 
-   buffer->vertex_bufs.individual.quad.stride = dec->quads.stride;
-   buffer->vertex_bufs.individual.quad.buffer_offset = dec->quads.buffer_offset;
-   pipe_resource_reference(&buffer->vertex_bufs.individual.quad.buffer, dec->quads.buffer);
-
-   buffer->vertex_bufs.individual.stream = vl_vb_init(&buffer->vertex_stream, dec->pipe,
-                                                      dec->base.width / MACROBLOCK_WIDTH *
-                                                      dec->base.height / MACROBLOCK_HEIGHT);
-   if (!buffer->vertex_bufs.individual.stream.buffer)
-      goto error_vertex_stream;
+   vl_vb_init(&buffer->vertex_stream, dec->pipe,
+              dec->base.width / MACROBLOCK_WIDTH,
+              dec->base.height / MACROBLOCK_HEIGHT);
 
    formats[0] = formats[1] = formats[2] =dec->mc_source_format;
    buffer->mc_source = vl_video_buffer_init(dec->base.context, dec->pipe,
@@ -461,7 +456,9 @@ vl_mpeg12_decoder_flush_buffer(struct pipe_video_decode_buffer *buffer,
    struct pipe_sampler_view **sv[2];
    struct pipe_surface **surfaces;
 
-   unsigned ne_start, ne_num, e_start, e_num;
+   struct pipe_vertex_buffer vb[3];
+
+   unsigned num_instances;
    unsigned i, j;
 
    assert(buf);
@@ -474,9 +471,10 @@ vl_mpeg12_decoder_flush_buffer(struct pipe_video_decode_buffer *buffer,
 
    surfaces = dst->get_surfaces(dst);
 
-   vl_vb_restart(&buf->vertex_stream, &ne_start, &ne_num, &e_start, &e_num);
+   num_instances = vl_vb_restart(&buf->vertex_stream);
 
-   dec->pipe->set_vertex_buffers(dec->pipe, 2, buf->vertex_bufs.all);
+   vb[0] = dec->quads;
+   vb[1] = dec->pos;
 
    for (i = 0; i < VL_MAX_PLANES; ++i) {
       vl_mc_set_surface(&buf->mc[i], surfaces[i]);
@@ -484,18 +482,25 @@ vl_mpeg12_decoder_flush_buffer(struct pipe_video_decode_buffer *buffer,
       for (j = 0; j < 2; ++j) {
          if (sv[j] == NULL) continue;
 
+         vb[2] = vl_vb_get_mv(&buf->vertex_stream, j);;
+         dec->pipe->set_vertex_buffers(dec->pipe, 3, vb);
+
          dec->pipe->bind_vertex_elements_state(dec->pipe, dec->ves_mv[j]);
-         vl_mc_render_ref(&buf->mc[i], sv[j][i], ne_start, ne_num, e_start, e_num);
+         vl_mc_render_ref(&buf->mc[i], sv[j][i]);
       }
+   }
 
-      dec->pipe->bind_vertex_elements_state(dec->pipe, dec->ves_eb[i]);
+   vb[1] = vl_vb_get_ycbcr(&buf->vertex_stream);
+   dec->pipe->set_vertex_buffers(dec->pipe, 2, vb);
 
+   for (i = 0; i < VL_MAX_PLANES; ++i) {
+      dec->pipe->bind_vertex_elements_state(dec->pipe, dec->ves_eb[i]);
       if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
-         vl_idct_flush(i == 0 ? &dec->idct_y : &dec->idct_c, &buf->idct[i], ne_num);
-
-      vl_mc_render_ycbcr(&buf->mc[i], ne_start, ne_num);
+         vl_idct_flush(i == 0 ? &dec->idct_y : &dec->idct_c, &buf->idct[i], num_instances);
 
+      vl_mc_render_ycbcr(&buf->mc[i], num_instances);
    }
+
    dec->pipe->flush(dec->pipe, fence);
 }
 
@@ -503,11 +508,10 @@ static void
 vl_mpeg12_decoder_clear_buffer(struct pipe_video_decode_buffer *buffer)
 {
    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer *)buffer;
-   unsigned ne_start, ne_num, e_start, e_num;
 
    assert(buf);
 
-   vl_vb_restart(&buf->vertex_stream, &ne_start, &ne_num, &e_start, &e_num);
+   vl_vb_restart(&buf->vertex_stream);
 }
 
 static bool
@@ -691,18 +695,24 @@ vl_create_mpeg12_decoder(struct pipe_video_context *context,
    dec->base.flush_buffer = vl_mpeg12_decoder_flush_buffer;
    dec->base.clear_buffer = vl_mpeg12_decoder_clear_buffer;
 
+   dec->base.width = align(width, MACROBLOCK_WIDTH);
+   dec->base.height = align(height, MACROBLOCK_HEIGHT);
+
    dec->pipe = pipe;
 
    dec->quads = vl_vb_upload_quads(dec->pipe, 2, 2);
+   dec->pos = vl_vb_upload_pos(
+      dec->pipe,
+      dec->base.width / MACROBLOCK_WIDTH,
+      dec->base.height / MACROBLOCK_HEIGHT
+   );
+
    for (i = 0; i < VL_MAX_PLANES; ++i)
       dec->ves_eb[i] = vl_vb_get_ves_eb(dec->pipe, i);
 
    for (i = 0; i < 2; ++i)
       dec->ves_mv[i] = vl_vb_get_ves_mv(dec->pipe, i);
 
-   dec->base.width = align(width, MACROBLOCK_WIDTH);
-   dec->base.height = align(height, MACROBLOCK_HEIGHT);
-
    /* TODO: Implement 422, 444 */
    assert(dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_420);
    dec->empty_block_mask = &const_empty_block_mask_420;
diff --git a/src/gallium/auxiliary/vl/vl_mpeg12_decoder.h b/src/gallium/auxiliary/vl/vl_mpeg12_decoder.h
index ffb9e56a13c..2bd5591b463 100644
--- a/src/gallium/auxiliary/vl/vl_mpeg12_decoder.h
+++ b/src/gallium/auxiliary/vl/vl_mpeg12_decoder.h
@@ -52,6 +52,8 @@ struct vl_mpeg12_decoder
    enum pipe_format mc_source_format;
 
    struct pipe_vertex_buffer quads;
+   struct pipe_vertex_buffer pos;
+
    void *ves_eb[VL_MAX_PLANES];
    void *ves_mv[2];
 
@@ -71,14 +73,6 @@ struct vl_mpeg12_buffer
    struct pipe_video_buffer *idct_intermediate;
    struct pipe_video_buffer *mc_source;
 
-   union
-   {
-      struct pipe_vertex_buffer all[2];
-      struct {
-         struct pipe_vertex_buffer quad, stream;
-      } individual;
-   } vertex_bufs;
-
    struct vl_idct_buffer idct[VL_MAX_PLANES];
    struct vl_mc_buffer mc[VL_MAX_PLANES];
 
diff --git a/src/gallium/auxiliary/vl/vl_vertex_buffers.c b/src/gallium/auxiliary/vl/vl_vertex_buffers.c
index 84dfc9eccf6..c923686b06e 100644
--- a/src/gallium/auxiliary/vl/vl_vertex_buffers.c
+++ b/src/gallium/auxiliary/vl/vl_vertex_buffers.c
@@ -37,7 +37,11 @@ struct vl_vertex_stream
    uint8_t dct_type_field;
    uint8_t dummy[2];
    uint8_t eb[3][2][2];
-   struct vertex4s mv[4];
+};
+
+struct vl_mv_vertex_stream
+{
+   struct vertex4s mv[2];
 };
 
 /* vertices for a quad covering a block */
@@ -96,6 +100,52 @@ vl_vb_upload_quads(struct pipe_context *pipe, unsigned blocks_x, unsigned blocks
    return quad;
 }
 
+struct pipe_vertex_buffer
+vl_vb_upload_pos(struct pipe_context *pipe, unsigned width, unsigned height)
+{
+   struct pipe_vertex_buffer pos;
+   struct pipe_transfer *buf_transfer;
+   struct vertex2s *v;
+
+   unsigned x, y;
+
+   assert(pipe);
+
+   /* create buffer */
+   pos.stride = sizeof(struct vertex2s);
+   pos.buffer_offset = 0;
+   pos.buffer = pipe_buffer_create
+   (
+      pipe->screen,
+      PIPE_BIND_VERTEX_BUFFER,
+      PIPE_USAGE_STATIC,
+      sizeof(struct vertex2s) * width * height
+   );
+
+   if(!pos.buffer)
+      return pos;
+
+   /* and fill it */
+   v = pipe_buffer_map
+   (
+      pipe,
+      pos.buffer,
+      PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD,
+      &buf_transfer
+   );
+
+   for ( y = 0; y < height; ++y) {
+      for ( x = 0; x < width; ++x, ++v) {
+         v->x = x;
+         v->y = y;
+      }
+   }
+
+   pipe_buffer_unmap(pipe, buf_transfer);
+
+   return pos;
+}
+
 static struct pipe_vertex_element
 vl_vb_get_quad_vertex_element(void)
 {
@@ -159,50 +209,83 @@ vl_vb_get_ves_mv(struct pipe_context *pipe, int motionvector)
 {
    struct pipe_vertex_element vertex_elems[NUM_VS_INPUTS];
 
+   assert(pipe);
+
    memset(&vertex_elems, 0, sizeof(vertex_elems));
    vertex_elems[VS_I_RECT] = vl_vb_get_quad_vertex_element();
 
-   assert(pipe);
-
    /* Position element */
    vertex_elems[VS_I_VPOS].src_format = PIPE_FORMAT_R16G16_SSCALED;
 
+   vl_vb_element_helper(&vertex_elems[VS_I_VPOS], 1, 1);
+
    /* motion vector TOP element */
-   vertex_elems[VS_I_MV_TOP].src_offset = offsetof(struct vl_vertex_stream, mv[motionvector * 2]);
    vertex_elems[VS_I_MV_TOP].src_format = PIPE_FORMAT_R16G16B16A16_SSCALED;
 
    /* motion vector BOTTOM element */
    vertex_elems[VS_I_MV_BOTTOM].src_format = PIPE_FORMAT_R16G16B16A16_SSCALED;
 
-   vl_vb_element_helper(&vertex_elems[VS_I_VPOS], NUM_VS_INPUTS - 1, 1);
+   vl_vb_element_helper(&vertex_elems[VS_I_MV_TOP], 2, 2);
 
    return pipe->create_vertex_elements_state(pipe, NUM_VS_INPUTS, vertex_elems);
 }
 
-struct pipe_vertex_buffer
-vl_vb_init(struct vl_vertex_buffer *buffer, struct pipe_context *pipe, unsigned size)
+void
+vl_vb_init(struct vl_vertex_buffer *buffer, struct pipe_context *pipe, unsigned width, unsigned height)
 {
-   struct pipe_vertex_buffer buf;
+   unsigned i;
 
    assert(buffer);
 
-   buffer->size = size;
-   buffer->num_not_empty = 0;
-   buffer->num_empty = 0;
+   buffer->width = width;
+   buffer->height = height;
+   buffer->num_instances = 0;
 
-   buf.stride = sizeof(struct vl_vertex_stream);
-   buf.buffer_offset = 0;
-   buf.buffer = pipe_buffer_create
+   buffer->resource = pipe_buffer_create
    (
       pipe->screen,
       PIPE_BIND_VERTEX_BUFFER,
       PIPE_USAGE_STREAM,
-      sizeof(struct vl_vertex_stream) * size
+      sizeof(struct vl_vertex_stream) * width * height
    );
 
-   pipe_resource_reference(&buffer->resource, buf.buffer);
+   for (i = 0; i < VL_MAX_REF_FRAMES; ++i) {
+      buffer->mv[i].resource = pipe_buffer_create
+      (
+         pipe->screen,
+         PIPE_BIND_VERTEX_BUFFER,
+         PIPE_USAGE_STREAM,
+         sizeof(struct vl_mv_vertex_stream) * width * height
+      );
+   }
 
    vl_vb_map(buffer, pipe);
+}
+
+struct pipe_vertex_buffer
+vl_vb_get_ycbcr(struct vl_vertex_buffer *buffer)
+{
+   struct pipe_vertex_buffer buf;
+
+   assert(buffer);
+
+   buf.stride = sizeof(struct vl_vertex_stream);
+   buf.buffer_offset = 0;
+   buf.buffer = buffer->resource;
+
+   return buf;
+}
+
+struct pipe_vertex_buffer
+vl_vb_get_mv(struct vl_vertex_buffer *buffer, int motionvector)
+{
+   struct pipe_vertex_buffer buf;
+
+   assert(buffer);
+
+   buf.stride = sizeof(struct vl_mv_vertex_stream);
+   buf.buffer_offset = 0;
+   buf.buffer = buffer->mv[motionvector].resource;
 
    return buf;
 }
@@ -210,110 +293,119 @@ vl_vb_init(struct vl_vertex_buffer *buffer, struct pipe_context *pipe, unsigned
 void
 vl_vb_map(struct vl_vertex_buffer *buffer, struct pipe_context *pipe)
 {
+   unsigned i;
+
    assert(buffer && pipe);
 
-   buffer->start = pipe_buffer_map
+   buffer->buffer = pipe_buffer_map
    (
       pipe,
       buffer->resource,
       PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD,
       &buffer->transfer
    );
-   buffer->end = buffer->start + buffer->resource->width0 / sizeof(struct vl_vertex_stream);
+
+   for (i = 0; i < VL_MAX_REF_FRAMES; ++i) {
+      buffer->mv[i].vertex_stream = pipe_buffer_map
+      (
+         pipe,
+         buffer->mv[i].resource,
+         PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD,
+         &buffer->mv[i].transfer
+      );
+   }
+
 }
 
 static void
-get_motion_vectors(struct pipe_mpeg12_macroblock *mb, struct vertex4s mv[4])
+get_motion_vectors(enum pipe_mpeg12_motion_type mo_type, struct pipe_motionvector *src, struct vertex4s dst[2])
 {
-   if (mb->mo_type == PIPE_MPEG12_MOTION_TYPE_FRAME) {
-      mv[0].x = mv[1].x = mb->mv[0].top.x;
-      mv[0].y = mv[1].y = mb->mv[0].top.y;
-      mv[0].z = mv[1].z = 0;
-
-      mv[2].x = mv[3].x = mb->mv[1].top.x;
-      mv[2].y = mv[3].y = mb->mv[1].top.y;
-      mv[2].z = mv[3].z = 0;
+   if (mo_type == PIPE_MPEG12_MOTION_TYPE_FRAME) {
+      dst[0].x = dst[1].x = src->top.x;
+      dst[0].y = dst[1].y = src->top.y;
+      dst[0].z = dst[1].z = 0;
 
    } else {
-      mv[0].x = mb->mv[0].top.x;
-      mv[0].y = mb->mv[0].top.y;
-      mv[0].z = mb->mv[0].top.field_select ? 3 : 1;
-
-      mv[1].x = mb->mv[0].bottom.x;
-      mv[1].y = mb->mv[0].bottom.y;
-      mv[1].z = mb->mv[0].bottom.field_select ? 3 : 1;
+      dst[0].x = src->top.x;
+      dst[0].y = src->top.y;
+      dst[0].z = src->top.field_select ? 3 : 1;
 
-      mv[2].x = mb->mv[1].top.x;
-      mv[2].y = mb->mv[1].top.y;
-      mv[2].z = mb->mv[1].top.field_select ? 3 : 1;
-
-      mv[3].x = mb->mv[1].bottom.x;
-      mv[3].y = mb->mv[1].bottom.y;
-      mv[3].z = mb->mv[1].bottom.field_select ? 3 : 1;
+      dst[1].x = src->bottom.x;
+      dst[1].y = src->bottom.y;
+      dst[1].z = src->bottom.field_select ? 3 : 1;
    }
 
-   mv[0].w = mv[1].w = mb->mv[0].wheight;
-   mv[2].w = mv[3].w = mb->mv[1].wheight;
+   dst[0].w = src->top.wheight;
+   dst[1].w = src->bottom.wheight;
 }
 
 void
 vl_vb_add_block(struct vl_vertex_buffer *buffer, struct pipe_mpeg12_macroblock *mb,
                 const unsigned (*empty_block_mask)[3][2][2])
 {
-   struct vl_vertex_stream *stream;
    unsigned i, j, k;
+   unsigned mv_pos;
 
    assert(buffer);
    assert(mb);
-   assert(buffer->num_not_empty + buffer->num_empty < buffer->size);
+   assert(buffer->num_instances < buffer->width * buffer->height);
 
-   if(mb->cbp)
-      stream = buffer->start + buffer->num_not_empty++;
-   else
-      stream = buffer->end - ++buffer->num_empty;
+   if(mb->cbp) {
+      struct vl_vertex_stream *stream;
+      stream = buffer->buffer + buffer->num_instances++;
 
-   stream->pos.x = mb->mbx;
-   stream->pos.y = mb->mby;
+      stream->pos.x = mb->mbx;
+      stream->pos.y = mb->mby;
 
-   for ( i = 0; i < 3; ++i)
-      for ( j = 0; j < 2; ++j)
-         for ( k = 0; k < 2; ++k)
-            stream->eb[i][j][k] = !(mb->cbp & (*empty_block_mask)[i][j][k]);
+      for ( i = 0; i < 3; ++i)
+         for ( j = 0; j < 2; ++j)
+            for ( k = 0; k < 2; ++k)
+               stream->eb[i][j][k] = !(mb->cbp & (*empty_block_mask)[i][j][k]);
 
-   stream->dct_type_field = mb->dct_type == PIPE_MPEG12_DCT_TYPE_FIELD;
-   stream->mb_type_intra = mb->dct_intra;
+      stream->dct_type_field = mb->dct_type == PIPE_MPEG12_DCT_TYPE_FIELD;
+      stream->mb_type_intra = mb->dct_intra;
+   }
 
-   get_motion_vectors(mb, stream->mv);
+   mv_pos = mb->mbx + mb->mby * buffer->width;
+   get_motion_vectors(mb->mo_type, &mb->mv[0], buffer->mv[0].vertex_stream[mv_pos].mv);
+   get_motion_vectors(mb->mo_type, &mb->mv[1], buffer->mv[1].vertex_stream[mv_pos].mv);
 }
 
 void
 vl_vb_unmap(struct vl_vertex_buffer *buffer, struct pipe_context *pipe)
 {
+   unsigned i;
+
    assert(buffer && pipe);
 
    pipe_buffer_unmap(pipe, buffer->transfer);
+   for (i = 0; i < VL_MAX_REF_FRAMES; ++i) {
+      pipe_buffer_unmap(pipe, buffer->mv[i].transfer);
+   }
 }
 
-void
-vl_vb_restart(struct vl_vertex_buffer *buffer,
-              unsigned *not_empty_start_instance, unsigned *not_empty_num_instances,
-              unsigned *empty_start_instance, unsigned *empty_num_instances)
+unsigned
+vl_vb_restart(struct vl_vertex_buffer *buffer)
 {
-   assert(buffer);
+   unsigned num_instances;
 
-   *not_empty_start_instance = 0;
-   *not_empty_num_instances = buffer->num_not_empty;
-   *empty_start_instance = buffer->size - buffer->num_empty;
-   *empty_num_instances = buffer->num_empty;
+   assert(buffer);
 
-   buffer->num_not_empty = 0;
-   buffer->num_empty = 0;
+   num_instances = buffer->num_instances;
+   buffer->num_instances = 0;
+   return num_instances;
 }
 
 void
 vl_vb_cleanup(struct vl_vertex_buffer *buffer)
 {
+   unsigned i;
+
    assert(buffer);
 
    pipe_resource_reference(&buffer->resource, NULL);
+
+   for (i = 0; i < VL_MAX_REF_FRAMES; ++i) {
+      pipe_resource_reference(&buffer->mv[i].resource, NULL);
+   }
 }
diff --git a/src/gallium/auxiliary/vl/vl_vertex_buffers.h b/src/gallium/auxiliary/vl/vl_vertex_buffers.h
index 62f7bf00508..bafaff7a311 100644
--- a/src/gallium/auxiliary/vl/vl_vertex_buffers.h
+++ b/src/gallium/auxiliary/vl/vl_vertex_buffers.h
@@ -30,6 +30,7 @@
 #include <pipe/p_state.h>
 #include <pipe/p_video_state.h>
 
+#include "vl_defines.h"
 #include "vl_types.h"
 
 /* vertex buffers act as a todo list
@@ -54,25 +55,36 @@ enum VS_INPUT
 
 struct vl_vertex_buffer
 {
-   unsigned size;
-   unsigned num_not_empty;
-   unsigned num_empty;
+   unsigned width, height;
+   unsigned num_instances;
+
    struct pipe_resource *resource;
    struct pipe_transfer *transfer;
-   struct vl_vertex_stream *start;
-   struct vl_vertex_stream *end;
+   struct vl_vertex_stream *buffer;
+
+   struct {
+      struct pipe_resource       *resource;
+      struct pipe_transfer       *transfer;
+      struct vl_mv_vertex_stream *vertex_stream;
+   } mv[VL_MAX_REF_FRAMES];
 };
 
 struct pipe_vertex_buffer vl_vb_upload_quads(struct pipe_context *pipe,
                                              unsigned blocks_x, unsigned blocks_y);
 
+struct pipe_vertex_buffer vl_vb_upload_pos(struct pipe_context *pipe, unsigned width, unsigned height);
+
 void *vl_vb_get_ves_eb(struct pipe_context *pipe, int component);
 
 void *vl_vb_get_ves_mv(struct pipe_context *pipe, int motionvector);
 
-struct pipe_vertex_buffer vl_vb_init(struct vl_vertex_buffer *buffer,
-                                     struct pipe_context *pipe,
-                                     unsigned max_blocks);
+void vl_vb_init(struct vl_vertex_buffer *buffer,
+                struct pipe_context *pipe,
+                unsigned width, unsigned height);
+
+struct pipe_vertex_buffer vl_vb_get_ycbcr(struct vl_vertex_buffer *buffer);
+
+struct pipe_vertex_buffer vl_vb_get_mv(struct vl_vertex_buffer *buffer, int motionvector);
 
 void vl_vb_map(struct vl_vertex_buffer *buffer, struct pipe_context *pipe);
 
@@ -81,9 +93,7 @@ void vl_vb_add_block(struct vl_vertex_buffer *buffer, struct pipe_mpeg12_macrobl
 
 void vl_vb_unmap(struct vl_vertex_buffer *buffer, struct pipe_context *pipe);
 
-void vl_vb_restart(struct vl_vertex_buffer *buffer,
-                   unsigned *not_empty_start_instance, unsigned *not_empty_num_instances,
-                   unsigned *empty_start_instance, unsigned *empty_num_instances);
+unsigned vl_vb_restart(struct vl_vertex_buffer *buffer);
 
 void vl_vb_cleanup(struct vl_vertex_buffer *buffer);
 
diff --git a/src/gallium/auxiliary/vl/vl_video_buffer.h b/src/gallium/auxiliary/vl/vl_video_buffer.h
index f5c424cf296..960acd28060 100644
--- a/src/gallium/auxiliary/vl/vl_video_buffer.h
+++ b/src/gallium/auxiliary/vl/vl_video_buffer.h
@@ -31,7 +31,7 @@
 #include <pipe/p_context.h>
 #include <pipe/p_video_context.h>
 
-#define VL_MAX_PLANES 3
+#include "vl_defines.h"
 
 /**
  * implementation of a planar ycbcr buffer
diff --git a/src/gallium/include/pipe/p_video_state.h b/src/gallium/include/pipe/p_video_state.h
index 72a27938847..2cf5ea4d975 100644
--- a/src/gallium/include/pipe/p_video_state.h
+++ b/src/gallium/include/pipe/p_video_state.h
@@ -69,10 +69,13 @@ struct pipe_macroblock
    enum pipe_video_codec codec;
 };
 
-struct pipe_mpeg12_motionvector
+struct pipe_motionvector
 {
-   signed x, y;
-   bool field_select;
+   struct {
+      signed x, y;
+      bool field_select;
+      unsigned wheight:8;
+   } top, bottom;
 };
 
 struct pipe_mpeg12_macroblock
@@ -84,10 +87,7 @@ struct pipe_mpeg12_macroblock
    enum pipe_mpeg12_motion_type mo_type;
    bool dct_intra;
    enum pipe_mpeg12_dct_type dct_type;
-   struct {
-      unsigned wheight:8;
-      struct pipe_mpeg12_motionvector top, bottom;
-   } mv[2];
+   struct pipe_motionvector mv[2];
    unsigned cbp;
    short *blocks;
 };
diff --git a/src/gallium/state_trackers/xorg/xvmc/surface.c b/src/gallium/state_trackers/xorg/xvmc/surface.c
index f22d315c90d..9752497eb30 100644
--- a/src/gallium/state_trackers/xorg/xvmc/surface.c
+++ b/src/gallium/state_trackers/xorg/xvmc/surface.c
@@ -114,28 +114,28 @@ MacroBlocksToPipe(struct pipe_screen *screen,
 
       switch (xvmc_mb->macroblock_type & (XVMC_MB_TYPE_MOTION_FORWARD | XVMC_MB_TYPE_MOTION_BACKWARD)) {
       case XVMC_MB_TYPE_MOTION_FORWARD:
-         mb->mv[0].wheight = 255;
-         mb->mv[1].wheight = 0;
+         mb->mv[0].top.wheight = mb->mv[0].bottom.wheight = 255;
+         mb->mv[1].top.wheight = mb->mv[1].bottom.wheight = 0;
          break;
 
       case (XVMC_MB_TYPE_MOTION_FORWARD | XVMC_MB_TYPE_MOTION_BACKWARD):
-         mb->mv[0].wheight = 127;
-         mb->mv[1].wheight = 127;
+         mb->mv[0].top.wheight = mb->mv[0].bottom.wheight = 127;
+         mb->mv[1].top.wheight = mb->mv[1].bottom.wheight = 127;
          break;
 
       case XVMC_MB_TYPE_MOTION_BACKWARD:
-         mb->mv[0].wheight = 0;
-         mb->mv[1].wheight = 255;
+         mb->mv[0].top.wheight = mb->mv[0].bottom.wheight = 0;
+         mb->mv[1].top.wheight = mb->mv[1].bottom.wheight = 255;
          break;
 
       default:
          /* workaround for xines xxmc video out plugin */
          if (!(xvmc_mb->macroblock_type & ~XVMC_MB_TYPE_PATTERN)) {
-            mb->mv[0].wheight = 255;
-            mb->mv[1].wheight = 0;
+            mb->mv[0].top.wheight = mb->mv[0].bottom.wheight = 255;
+            mb->mv[1].top.wheight = mb->mv[1].bottom.wheight = 0;
          } else {
-            mb->mv[0].wheight = 0;
-            mb->mv[1].wheight = 0;
+            mb->mv[0].top.wheight = mb->mv[0].bottom.wheight = 0;
+            mb->mv[1].top.wheight = mb->mv[1].bottom.wheight = 0;
          }
          break;
       }