16 files changed, 344 insertions, 308 deletions

diff --git a/i965_drv_video/i965_defines.h b/i965_drv_video/i965_defines.h
index e4e5b2d..260d820 100644
--- a/i965_drv_video/i965_defines.h
+++ b/i965_drv_video/i965_defines.h
@@ -14,7 +14,7 @@
 #define CMD_STATE_BASE_ADDRESS                  CMD(0, 1, 1)
 #define CMD_STATE_SIP                           CMD(0, 1, 2)
 #define CMD_PIPELINE_SELECT                     CMD(1, 1, 4)
-#define CMD_SAMPLER_PALETTE_LOAD                CMD(3, 3, 2)
+#define CMD_SAMPLER_PALETTE_LOAD                CMD(3, 1, 2)
 
 #define CMD_MEDIA_STATE_POINTERS                CMD(2, 0, 0)
 #define CMD_MEDIA_OBJECT                        CMD(2, 1, 0)
diff --git a/i965_drv_video/i965_drv_video.c b/i965_drv_video/i965_drv_video.c
index d26942d..8f4bfc1 100644
--- a/i965_drv_video/i965_drv_video.c
+++ b/i965_drv_video/i965_drv_video.c
@@ -39,6 +39,7 @@
 
 #include "i965_media.h"
 #include "i965_drv_video.h"
+#include "i965_defines.h"
 
 #define CONFIG_ID_OFFSET                0x01000000
 #define CONTEXT_ID_OFFSET               0x02000000
@@ -47,6 +48,48 @@
 #define IMAGE_ID_OFFSET                 0x0a000000
 #define SUBPIC_ID_OFFSET                0x10000000
 
+enum {
+    I965_SURFACETYPE_RGBA = 1,
+    I965_SURFACETYPE_YUV,
+    I965_SURFACETYPE_INDEXED
+};
+
+/* List of supported subpicture formats */
+typedef struct {
+    unsigned int        type;
+    unsigned int        format;
+    VAImageFormat       va_format;
+    unsigned int        va_flags;
+} i965_subpic_format_map_t;
+
+static const i965_subpic_format_map_t
+i965_subpic_formats_map[I965_MAX_SUBPIC_FORMATS + 1] = {
+    { I965_SURFACETYPE_INDEXED, I965_SURFACEFORMAT_P4A4_UNORM,
+      { VA_FOURCC('I','A','4','4'), VA_MSB_FIRST, 8, },
+      0 },
+    { I965_SURFACETYPE_INDEXED, I965_SURFACEFORMAT_A4P4_UNORM,
+      { VA_FOURCC('A','I','4','4'), VA_MSB_FIRST, 8, },
+      0 },
+};
+
+static const i965_subpic_format_map_t *
+get_subpic_format(const VAImageFormat *va_format)
+{
+    unsigned int i;
+    for (i = 0; i < sizeof(i965_subpic_formats_map)/sizeof(i965_subpic_formats_map[0]); i++) {
+        const i965_subpic_format_map_t * const m = &i965_subpic_formats_map[i];
+        if (m->va_format.fourcc == va_format->fourcc &&
+            (m->type == I965_SURFACETYPE_RGBA ?
+             (m->va_format.byte_order == va_format->byte_order &&
+              m->va_format.red_mask   == va_format->red_mask   &&
+              m->va_format.green_mask == va_format->green_mask &&
+              m->va_format.blue_mask  == va_format->blue_mask  &&
+              m->va_format.alpha_mask == va_format->alpha_mask) : 1))
+            return m;
+    }
+    return NULL;
+}
+
 VAStatus 
 i965_QueryConfigProfiles(VADriverContextP ctx,
                          VAProfile *profile_list,       /* out */
@@ -293,6 +336,7 @@ i965_CreateSurfaces(VADriverContextP ctx,
 
         surfaces[i] = surfaceID;
         obj_surface->status = VASurfaceReady;
+        obj_surface->subpic = VA_INVALID_ID;
         obj_surface->width = width;
         obj_surface->height = height;
         obj_surface->size = SIZE_YUV420(width, height);
@@ -346,6 +390,9 @@ i965_QueryImageFormats(VADriverContextP ctx,
                        VAImageFormat *format_list,      /* out */
                        int *num_formats)                /* out */
 {
+    if (num_formats)
+        *num_formats = 0;
+
     return VA_STATUS_SUCCESS;
 }
 
@@ -371,12 +418,19 @@ i965_QuerySubpictureFormats(VADriverContextP ctx,
                             unsigned int *flags,                /* out */
                             unsigned int *num_formats)          /* out */
 {
-    /*support 2 subpicture formats*/
-    *num_formats = 2;
-    format_list[0].fourcc=FOURCC_IA44;
-    format_list[0].byte_order=VA_LSB_FIRST;
-    format_list[1].fourcc=FOURCC_AI44;
-    format_list[1].byte_order=VA_LSB_FIRST;
+    int n;
+
+    for (n = 0; i965_subpic_formats_map[n].va_format.fourcc != 0; n++) {
+        const i965_subpic_format_map_t * const m = &i965_subpic_formats_map[n];
+        if (format_list)
+            format_list[n] = m->va_format;
+        if (flags)
+            flags[n] = m->va_flags;
+    }
+
+    if (num_formats)
+        *num_formats = n;
+
     return VA_STATUS_SUCCESS;
 }
 
@@ -394,25 +448,27 @@ i965_CreateSubpicture(VADriverContextP ctx,
                       VASubpictureID *subpicture)         /* out */
 {
     struct i965_driver_data *i965 = i965_driver_data(ctx);
-    VAStatus vaStatus = VA_STATUS_SUCCESS;
     VASubpictureID subpicID = NEW_SUBPIC_ID()
 	
     struct object_subpic *obj_subpic = SUBPIC(subpicID);
+    if (!obj_subpic)
+        return VA_STATUS_ERROR_ALLOCATION_FAILED;
+
     struct object_image *obj_image = IMAGE(image);
-    
-    if (NULL == obj_subpic) {
-        vaStatus = VA_STATUS_ERROR_ALLOCATION_FAILED;
-    }
-    if (NULL == obj_image) {
-        vaStatus = VA_STATUS_ERROR_ALLOCATION_FAILED;
-    }
+    if (!obj_image)
+        return VA_STATUS_ERROR_INVALID_IMAGE;
+
+    const i965_subpic_format_map_t * const m = get_subpic_format(&obj_image->image.format);
+    if (!m)
+        return VA_STATUS_ERROR_UNKNOWN; /* XXX: VA_STATUS_ERROR_UNSUPPORTED_FORMAT? */
+
     *subpicture = subpicID;
-    obj_subpic->image = image;
-    obj_subpic->width = obj_image->width;
-    obj_subpic->height = obj_image->height;
-    obj_subpic->bo = obj_image->bo;
-	
-    return vaStatus;
+    obj_subpic->image  = image;
+    obj_subpic->format = m->format;
+    obj_subpic->width  = obj_image->image.width;
+    obj_subpic->height = obj_image->image.height;
+    obj_subpic->bo     = obj_image->bo;
+    return VA_STATUS_SUCCESS;
 }
 
 VAStatus 
@@ -434,27 +490,6 @@ i965_SetSubpictureImage(VADriverContextP ctx,
     return VA_STATUS_SUCCESS;
 }
 
-/*
- * pointer to an array holding the palette data.  The size of the array is
- * num_palette_entries * entry_bytes in size.  The order of the components
- * in the palette is described by the component_order in VASubpicture struct
- */
-VAStatus 
-i965_SetSubpicturePalette(VADriverContextP ctx,
-                          VASubpictureID subpicture,
-                          unsigned char *palette)
-{
-    /*set palette in shader,so the following code is unused*/
-    struct i965_driver_data *i965 = i965_driver_data(ctx);
-    VAStatus vaStatus = VA_STATUS_SUCCESS;
-    struct object_subpic *obj_subpic = SUBPIC(subpicture);
-    if (NULL == obj_subpic) {
-        vaStatus = VA_STATUS_ERROR_ALLOCATION_FAILED;
-    }
-    memcpy(obj_subpic->palette, palette, 3*16);	
-    return VA_STATUS_SUCCESS;
-}
-
 VAStatus 
 i965_SetSubpictureChromakey(VADriverContextP ctx,
                             VASubpictureID subpicture,
@@ -486,8 +521,6 @@ i965_AssociateSubpicture(VADriverContextP ctx,
                          short dest_y,
                          unsigned short dest_width,
                          unsigned short dest_height,
-                         unsigned short width,
-                         unsigned short height,
                          /*
                           * whether to enable chroma-keying or global-alpha
                           * see VA_SUBPICTURE_XXX values
@@ -495,22 +528,24 @@ i965_AssociateSubpicture(VADriverContextP ctx,
                          unsigned int flags)
 {
     struct i965_driver_data *i965 = i965_driver_data(ctx);
-    VAStatus vaStatus = VA_STATUS_SUCCESS;
-    /*only ipicture*/
-	
-    struct object_surface *obj_surface = SURFACE(*target_surfaces);
     struct object_subpic *obj_subpic = SUBPIC(subpicture);
+    int i;
 
-    if (NULL == obj_surface) {
-        vaStatus = VA_STATUS_ERROR_INVALID_CONFIG;
-        return vaStatus;
-    }
-  
-    obj_subpic->dstx = dest_x;
-    obj_subpic->dsty = dest_y;
-
-    obj_surface->subpic = subpicture;
+    obj_subpic->src_rect.x      = src_x;
+    obj_subpic->src_rect.y      = src_y;
+    obj_subpic->src_rect.width  = src_width;
+    obj_subpic->src_rect.height = src_height;
+    obj_subpic->dst_rect.x      = dest_x;
+    obj_subpic->dst_rect.y      = dest_y;
+    obj_subpic->dst_rect.width  = dest_width;
+    obj_subpic->dst_rect.height = dest_height;
 
+    for (i = 0; i < num_surfaces; i++) {
+        struct object_surface *obj_surface = SURFACE(target_surfaces[i]);
+        if (!obj_surface)
+            return VA_STATUS_ERROR_INVALID_SURFACE;
+        obj_surface->subpic = subpicture;
+    }
     return VA_STATUS_SUCCESS;
 }
 
@@ -521,6 +556,16 @@ i965_DeassociateSubpicture(VADriverContextP ctx,
                            VASurfaceID *target_surfaces,
                            int num_surfaces)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
+    int i;
+
+    for (i = 0; i < num_surfaces; i++) {
+        struct object_surface *obj_surface = SURFACE(target_surfaces[i]);
+        if (!obj_surface)
+            return VA_STATUS_ERROR_INVALID_SURFACE;
+        if (obj_surface->subpic == subpicture)
+            obj_surface->subpic = VA_INVALID_ID;
+    }
     return VA_STATUS_SUCCESS;
 }
 
@@ -1031,8 +1076,10 @@ i965_QueryDisplayAttributes(VADriverContextP ctx,
                             VADisplayAttribute *attr_list,	/* out */
                             int *num_attributes)		/* out */
 {
-    /* TODO */
-    return VA_STATUS_ERROR_UNKNOWN;
+    if (num_attributes)
+        *num_attributes = 0;
+
+    return VA_STATUS_SUCCESS;
 }
 
 /* 
@@ -1116,43 +1163,89 @@ i965_destroy_heap(struct object_heap *heap,
 }
 
 
+VAStatus 
+i965_DestroyImage(VADriverContextP ctx, VAImageID image);
 
 VAStatus 
 i965_CreateImage(VADriverContextP ctx,
                  VAImageFormat *format,
                  int width,
                  int height,
-                 VAImage *image)        /* out */
+                 VAImage *out_image)        /* out */
 {
     struct i965_driver_data *i965 = i965_driver_data(ctx);
-    VAStatus va_status;
-    /*we will receive the actual subpicture size from the player,now we assume it is 720*32*/
-    char subpic_buf[width*height];
-    int subpic_size = 720*32;
-    unsigned int img_buf_id;
+    struct object_image *obj_image;
+    VAStatus va_status = VA_STATUS_ERROR_OPERATION_FAILED;
+    VAImageID image_id;
+    unsigned int width2, height2, size2, size;
 
-    image->image_id = NEW_IMAGE_ID();
-    struct object_image *obj_image = IMAGE(image->image_id);
+    out_image->image_id = VA_INVALID_ID;
+    out_image->buf      = VA_INVALID_ID;
 
-    /*assume we got IA44 in format[0]*/
-    image->format = *format;
-	
-    /*create empty buffer*/
-    va_status = i965_CreateBuffer(ctx, 0, VAImageBufferType, 
-	    subpic_size, 1, subpic_buf, &img_buf_id);				
-    assert( VA_STATUS_SUCCESS == va_status );
-    struct object_buffer *obj_buf = BUFFER(img_buf_id);
-
-    image->buf = img_buf_id;
-    image->width = width;
-    image->height = height;
-	
-    obj_image->width = width;
-    obj_image->height = height;
-    obj_image->size = subpic_size;
-    obj_image->bo = obj_buf->buffer_store->bo;
-	
+    image_id = NEW_IMAGE_ID();
+    if (image_id == VA_INVALID_ID)
+        return VA_STATUS_ERROR_ALLOCATION_FAILED;
+
+    obj_image = IMAGE(image_id);
+    if (!obj_image)
+        return VA_STATUS_ERROR_ALLOCATION_FAILED;
+    obj_image->bo         = NULL;
+    obj_image->palette    = NULL;
+
+    VAImage * const image = &obj_image->image;
+    image->image_id       = image_id;
+    image->buf            = VA_INVALID_ID;
+
+    size    = width * height;
+    width2  = (width  + 1) / 2;
+    height2 = (height + 1) / 2;
+    size2   = width2 * height2;
+
+    image->num_palette_entries = 0;
+    image->entry_bytes         = 0;
+    memset(image->component_order, 0, sizeof(image->component_order));
+
+    switch (format->fourcc) {
+    case VA_FOURCC('I','A','4','4'):
+    case VA_FOURCC('A','I','4','4'):
+        image->num_planes = 1;
+        image->pitches[0] = width;
+        image->offsets[0] = 0;
+        image->data_size  = image->offsets[0] + image->pitches[0] * height;
+        image->num_palette_entries = 16;
+        image->entry_bytes         = 3;
+        image->component_order[0]  = 'R';
+        image->component_order[1]  = 'G';
+        image->component_order[2]  = 'B';
+        break;
+    default:
+        goto error;
+    }
+
+    va_status = i965_CreateBuffer(ctx, 0, VAImageBufferType,
+                                  image->data_size, 1, NULL, &image->buf);
+    if (va_status != VA_STATUS_SUCCESS)
+        goto error;
+
+    obj_image->bo = BUFFER(image->buf)->buffer_store->bo;
+
+    if (image->num_palette_entries > 0 && image->entry_bytes > 0) {
+        obj_image->palette = malloc(image->num_palette_entries * sizeof(obj_image->palette));
+        if (!obj_image->palette)
+            goto error;
+    }
+
+    image->image_id             = image_id;
+    image->format               = *format;
+    image->width                = width;
+    image->height               = height;
+
+    *out_image                  = *image;
     return VA_STATUS_SUCCESS;
+
+ error:
+    i965_DestroyImage(ctx, image_id);
+    return va_status;
 }
 
 VAStatus i965_DeriveImage(VADriverContextP ctx,
@@ -1175,18 +1268,48 @@ i965_DestroyImage(VADriverContextP ctx, VAImageID image)
     struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct object_image *obj_image = IMAGE(image); 
 
-    i965_DestroyBuffer(ctx, image);				
-	
+    if (!obj_image)
+        return VA_STATUS_SUCCESS;
+
+    if (obj_image->image.buf != VA_INVALID_ID) {
+        i965_DestroyBuffer(ctx, obj_image->image.buf);
+        obj_image->image.buf = VA_INVALID_ID;
+    }
+
+    if (obj_image->palette) {
+        free(obj_image->palette);
+        obj_image->palette = NULL;
+    }
+
     i965_destroy_image(&i965->image_heap, (struct object_base *)obj_image);
 	
     return VA_STATUS_SUCCESS;
 }
 
+/*
+ * pointer to an array holding the palette data.  The size of the array is
+ * num_palette_entries * entry_bytes in size.  The order of the components
+ * in the palette is described by the component_order in VASubpicture struct
+ */
 VAStatus 
 i965_SetImagePalette(VADriverContextP ctx,
                      VAImageID image,
                      unsigned char *palette)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
+    unsigned int i;
+
+    struct object_image *obj_image = IMAGE(image);
+    if (!obj_image)
+        return VA_STATUS_ERROR_INVALID_IMAGE;
+
+    if (!obj_image->palette)
+        return VA_STATUS_ERROR_ALLOCATION_FAILED; /* XXX: unpaletted/error */
+
+    for (i = 0; i < obj_image->image.num_palette_entries; i++)
+        obj_image->palette[i] = (((unsigned int)palette[3*i + 0] << 16) |
+                                 ((unsigned int)palette[3*i + 1] <<  8) |
+                                  (unsigned int)palette[3*i + 2]);
     return VA_STATUS_SUCCESS;
 }
 
@@ -1274,7 +1397,7 @@ i965_PutSurface(VADriverContextP ctx,
                             srcx, srcy, srcw, srch,
                             destx, desty, destw, desth);
     obj_surface = SURFACE(surface);
-    if(obj_surface->subpic != 0) {	
+    if(obj_surface->subpic != VA_INVALID_ID) {	
 	i965_render_put_subpic(ctx, surface,
                            srcx, srcy, srcw, srch,
                            destx, desty, destw, desth);
@@ -1317,8 +1440,8 @@ __vaDriverInit_0_31(  VADriverContextP ctx )
     struct i965_driver_data *i965;
     int result;
 
-    ctx->version_major = 0;
-    ctx->version_minor = 29;
+    ctx->version_major = VA_MAJOR_VERSION;
+    ctx->version_minor = VA_MINOR_VERSION;
     ctx->max_profiles = I965_MAX_PROFILES;
     ctx->max_entrypoints = I965_MAX_ENTRYPOINTS;
     ctx->max_attributes = I965_MAX_CONFIG_ATTRIBUTES;
@@ -1361,7 +1484,6 @@ __vaDriverInit_0_31(  VADriverContextP ctx )
     ctx->vtable.vaCreateSubpicture = i965_CreateSubpicture;
     ctx->vtable.vaDestroySubpicture = i965_DestroySubpicture;
     ctx->vtable.vaSetSubpictureImage = i965_SetSubpictureImage;
-    //ctx->vtable.vaSetSubpicturePalette = i965_SetSubpicturePalette;
     ctx->vtable.vaSetSubpictureChromakey = i965_SetSubpictureChromakey;
     ctx->vtable.vaSetSubpictureGlobalAlpha = i965_SetSubpictureGlobalAlpha;
     ctx->vtable.vaAssociateSubpicture = i965_AssociateSubpicture;
diff --git a/i965_drv_video/i965_drv_video.h b/i965_drv_video/i965_drv_video.h
index 1771b0d..c95c799 100644
--- a/i965_drv_video/i965_drv_video.h
+++ b/i965_drv_video/i965_drv_video.h
@@ -42,7 +42,7 @@
 #define I965_MAX_ENTRYPOINTS                    5
 #define I965_MAX_CONFIG_ATTRIBUTES              10
 #define I965_MAX_IMAGE_FORMATS                  10
-#define I965_MAX_SUBPIC_FORMATS                 4
+#define I965_MAX_SUBPIC_FORMATS                 2
 #define I965_MAX_DISPLAY_ATTRIBUTES             4
 #define I965_STR_VENDOR                         "i965 Driver 0.1"
 
@@ -91,7 +91,7 @@ struct object_surface
     struct object_base base;
     VASurfaceStatus status;
     VASubpictureID subpic;
-	int width;
+    int width;
     int height;
     int size;
     dri_bo *bo;
@@ -106,24 +106,25 @@ struct object_buffer
     int size_element;
     VABufferType type;
 };
+
 struct object_image 
 {
     struct object_base base;
-    int width;
-    int height;
-    int size;
+    VAImage image;
     dri_bo *bo;
+    unsigned int *palette;
 };
+
 struct object_subpic 
 {
     struct object_base base;
     VAImageID image;
-	int dstx;
-	int dsty;
-	int width;
-	int height;
-	unsigned char palette[3][16];
-	dri_bo *bo;
+    VARectangle src_rect;
+    VARectangle dst_rect;
+    unsigned int format;
+    int width;
+    int height;
+    dri_bo *bo;
 };
 
 
diff --git a/i965_drv_video/i965_media_mpeg2.c b/i965_drv_video/i965_media_mpeg2.c
index 17b4679..6a14e9e 100644
--- a/i965_drv_video/i965_media_mpeg2.c
+++ b/i965_drv_video/i965_media_mpeg2.c
@@ -868,7 +868,7 @@ i965_media_mpeg2_objects(VADriverContextP ctx, struct decode_state *decode_state
                   I915_GEM_DOMAIN_SAMPLER, 0, 
                   slice_param->slice_data_offset + (slice_param->macroblock_offset >> 3));
         OUT_BATCH(ctx, 
-                  ((0 << 24) |     
+                  ((slice_param->slice_horizontal_position << 24) |     
                    (slice_param->slice_vertical_position << 16) |
                    (127 << 8) | 
                    (slice_param->macroblock_offset & 0x7)));
diff --git a/i965_drv_video/i965_render.c b/i965_drv_video/i965_render.c
index e897dcd..f3e89c2 100644
--- a/i965_drv_video/i965_render.c
+++ b/i965_drv_video/i965_render.c
@@ -70,7 +70,7 @@ static const unsigned int ps_subpic_kernel_static[][4] =
 {
 #include "shaders/render/exa_wm_xy.g4b"
 #include "shaders/render/exa_wm_src_affine.g4b"
-#include "shaders/render/exa_wm_blend_subpicture.g4b"
+#include "shaders/render/exa_wm_src_sample_argb.g4b"
 #include "shaders/render/exa_wm_write.g4b"
 };
 
@@ -92,7 +92,7 @@ static const unsigned int ps_subpic_kernel_static_gen5[][4] =
 {
 #include "shaders/render/exa_wm_xy.g4b.gen5"
 #include "shaders/render/exa_wm_src_affine.g4b.gen5"
-#include "shaders/render/exa_wm_blend_subpicture.g4b.gen5"
+#include "shaders/render/exa_wm_src_sample_argb.g4b.gen5"
 #include "shaders/render/exa_wm_write.g4b.gen5"
 };
 
@@ -580,12 +580,13 @@ i965_render_src_surface_state(VADriverContextP ctx,
     render_state->wm.surface[index] = ss_bo;
     render_state->wm.sampler_count++;
 }
+
 static void
 i965_subpic_render_src_surface_state(VADriverContextP ctx, 
                               int index,
                               dri_bo *region,
                               unsigned long offset,
-                              int w, int h)
+                              int w, int h, int format)
 {
     struct i965_driver_data *i965 = i965_driver_data(ctx);  
     struct i965_render_state *render_state = &i965->render_state;
@@ -601,7 +602,7 @@ i965_subpic_render_src_surface_state(VADriverContextP ctx,
     ss = ss_bo->virtual;
     memset(ss, 0, sizeof(*ss));
     ss->ss0.surface_type = I965_SURFACE_2D;
-    ss->ss0.surface_format = I965_SURFACEFORMAT_R8_UNORM;
+    ss->ss0.surface_format = format;
     ss->ss0.writedisable_alpha = 0;
     ss->ss0.writedisable_red = 0;
     ss->ss0.writedisable_green = 0;
@@ -677,8 +678,8 @@ i965_subpic_render_src_surfaces_state(VADriverContextP ctx,
     region = obj_surface->bo;
     subpic_region = obj_image->bo;
     /*subpicture surface*/
-    i965_subpic_render_src_surface_state(ctx, 1, subpic_region, 0, obj_image->width, obj_image->height);     
-    i965_subpic_render_src_surface_state(ctx, 2, subpic_region, 0, obj_image->width, obj_image->height);
+    i965_subpic_render_src_surface_state(ctx, 1, subpic_region, 0, obj_subpic->width, obj_subpic->height, obj_subpic->format);     
+    i965_subpic_render_src_surface_state(ctx, 2, subpic_region, 0, obj_subpic->width, obj_subpic->height, obj_subpic->format);     
 }
 
 static void
@@ -788,58 +789,57 @@ i965_render_binding_table(VADriverContextP ctx)
 
 static void 
 i965_subpic_render_upload_vertex(VADriverContextP ctx,
-                          VASurfaceID surface,
-                          short srcx,
-                          short srcy,
-                          unsigned short srcw,
-                          unsigned short srch,
-                          short destx,
-                          short desty,
-                          unsigned short destw,
-                          unsigned short desth)
+                                 VASurfaceID surface,
+                                 const VARectangle *output_rect)
 {    
-	struct i965_driver_data *i965 = i965_driver_data(ctx);
+    struct i965_driver_data  *i965         = i965_driver_data(ctx);
     struct i965_render_state *render_state = &i965->render_state;
-    struct intel_region *dest_region = render_state->draw_region;
-    struct object_surface *obj_surface;
-    struct object_subpic *obj_subpic;
-    float *vb;
-    float src_scale_x, src_scale_y;
-    int i, width, height;
-    obj_surface = SURFACE(surface);
-    obj_subpic = SUBPIC(obj_surface->subpic);
-    assert(obj_surface);
-    assert(obj_subpic);
-    
-    int box_x1 = dest_region->x + obj_subpic->dstx;
-    int box_y1 = dest_region->y + obj_subpic->dsty;
-    int box_x2 = box_x1 + obj_subpic->width;
-    int box_y2 = box_y1 + obj_subpic->height;
-
-    width = obj_surface->width;
-    height = obj_surface->height;
-    src_scale_x = ((float)srcw / width) / (float)destw;
-    src_scale_y = ((float)srch / height) / (float)desth;
+    struct object_surface    *obj_surface  = SURFACE(surface);
+    struct object_subpic     *obj_subpic   = SUBPIC(obj_surface->subpic);
+
+    const float psx = (float)obj_surface->width  / (float)obj_subpic->width;
+    const float psy = (float)obj_surface->height / (float)obj_subpic->height;
+    const float ssx = (float)output_rect->width  / (float)obj_surface->width;
+    const float ssy = (float)output_rect->height / (float)obj_surface->height;
+    const float sx  = psx * ssx;
+    const float sy  = psy * ssy;
+    float *vb, tx1, tx2, ty1, ty2, x1, x2, y1, y2;
+    int i = 0;
+
+    VARectangle dst_rect;
+    dst_rect.x      = output_rect->x + sx * (float)obj_subpic->dst_rect.x;
+    dst_rect.y      = output_rect->y + sx * (float)obj_subpic->dst_rect.y;
+    dst_rect.width  = sx * (float)obj_subpic->dst_rect.width;
+    dst_rect.height = sy * (float)obj_subpic->dst_rect.height;
 
     dri_bo_map(render_state->vb.vertex_buffer, 1);
     assert(render_state->vb.vertex_buffer->virtual);
     vb = render_state->vb.vertex_buffer->virtual;
-    /*vertex covers the full subpicture*/
-    i = 0;
-    vb[i++] = 1;
-    vb[i++] = 1;
-    vb[i++] = (float)box_x2;
-    vb[i++] = (float)box_y2;
-    
-    vb[i++] = 0.0;
-    vb[i++] = 1;
-    vb[i++] = (float)box_x1;
-    vb[i++] = (float)box_y2;
 
-    vb[i++] = 0.0;
-    vb[i++] = 0.0;
-    vb[i++] = (float)box_x1;
-    vb[i++] = (float)box_y1;
+    tx1 = (float)obj_subpic->src_rect.x / (float)obj_subpic->width;
+    ty1 = (float)obj_subpic->src_rect.y / (float)obj_subpic->height;
+    tx2 = (float)(obj_subpic->src_rect.x + obj_subpic->src_rect.width) / (float)obj_subpic->width;
+    ty2 = (float)(obj_subpic->src_rect.y + obj_subpic->src_rect.height) / (float)obj_subpic->height;
+
+    x1 = (float)dst_rect.x;
+    y1 = (float)dst_rect.y;
+    x2 = (float)(dst_rect.x + dst_rect.width);
+    y2 = (float)(dst_rect.y + dst_rect.height);
+
+    vb[i++] = tx2;
+    vb[i++] = ty2;
+    vb[i++] = x2;
+    vb[i++] = y2;
+
+    vb[i++] = tx1;
+    vb[i++] = ty2;
+    vb[i++] = x1;
+    vb[i++] = y2;
+
+    vb[i++] = tx1;
+    vb[i++] = ty1;
+    vb[i++] = x1;
+    vb[i++] = y1;
     dri_bo_unmap(render_state->vb.vertex_buffer);
 }
 
@@ -947,9 +947,13 @@ i965_subpic_render_state_setup(VADriverContextP ctx,
     i965_render_cc_viewport(ctx);
     i965_subpic_render_cc_unit(ctx);
     i965_render_binding_table(ctx);
-    i965_subpic_render_upload_vertex(ctx, surface,
-	    srcx, srcy, srcw, srch,
-	    destx, desty, destw, desth);
+
+    VARectangle output_rect;
+    output_rect.x      = destx;
+    output_rect.y      = desty;
+    output_rect.width  = destw;
+    output_rect.height = desth;
+    i965_subpic_render_upload_vertex(ctx, surface, &output_rect);
 }
 
 
@@ -1162,21 +1166,31 @@ i965_render_vertex_elements(VADriverContextP ctx)
     }
 }
 
-void
-i965_render_upload_palette(VADriverContextP ctx)
+static void
+i965_render_upload_image_palette(
+    VADriverContextP ctx,
+    VAImageID        image_id,
+    unsigned int     alpha
+)
 {
-    BEGIN_BATCH(ctx, 17);
-    OUT_BATCH(ctx, CMD_SAMPLER_PALETTE_LOAD | 15);
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
+    unsigned int i;
+
+    struct object_image *obj_image = IMAGE(image_id);
+    assert(obj_image);
+
+    if (obj_image->image.num_palette_entries == 0)
+        return;
+
+    BEGIN_BATCH(ctx, 1 + obj_image->image.num_palette_entries);
+    OUT_BATCH(ctx, CMD_SAMPLER_PALETTE_LOAD | (obj_image->image.num_palette_entries - 1));
     /*fill palette*/
     //int32_t out[16]; //0-23:color 23-31:alpha
-    int32_t i,c;
-    for(i = 0; i < 16; i ++){
-        c = i*16; //16 colors
-        OUT_BATCH(ctx,c<<24/*alpha*/|c<<16/*R*/|c<<8/*G*/|c/*B*/);//c<<24/*alpha*/|c<<16/*R*/|c<<8/*G*/|c/*B*/);
-    }
-
+    for (i = 0; i < obj_image->image.num_palette_entries; i++)
+        OUT_BATCH(ctx, (alpha << 24) | obj_image->palette[i]);
     ADVANCE_BATCH(ctx);
 }
+
 static void
 i965_render_startup(VADriverContextP ctx)
 {
@@ -1284,7 +1298,6 @@ i965_subpic_render_pipeline_setup(VADriverContextP ctx)
     i965_render_binding_table_pointers(ctx);
     i965_render_constant_color(ctx);
     i965_render_pipelined_pointers(ctx);
-    //i965_render_upload_palette(ctx);
     i965_render_urb_layout(ctx);
     i965_render_cs_urb_layout(ctx);
     i965_render_drawing_rectangle(ctx);
@@ -1412,11 +1425,17 @@ i965_render_put_subpic(VADriverContextP ctx,
                         unsigned short destw,
                         unsigned short desth)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
+    struct object_surface *obj_surface = SURFACE(surface);
+    struct object_subpic *obj_subpic = SUBPIC(obj_surface->subpic);
+    assert(obj_subpic);
+
     i965_render_initialize(ctx);
     i965_subpic_render_state_setup(ctx, surface,
 	    srcx, srcy, srcw, srch,
 	    destx, desty, destw, desth);
     i965_subpic_render_pipeline_setup(ctx);
+    i965_render_upload_image_palette(ctx, obj_subpic->image, 0xff);
     intel_batchbuffer_flush(ctx);
 }
 
diff --git a/i965_drv_video/intel_driver.h b/i965_drv_video/intel_driver.h
index f768703..84df0aa 100644
--- a/i965_drv_video/intel_driver.h
+++ b/i965_drv_video/intel_driver.h
@@ -43,6 +43,7 @@ struct intel_batchbuffer;
 
 #define ALIGN(i, n)    (((i) + (n) - 1) & ~((n) - 1))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
 
 #define SET_BLOCKED_SIGSET()   do {     \
         sigset_t bl_mask;               \
diff --git a/i965_drv_video/shaders/render/Makefile.am b/i965_drv_video/shaders/render/Makefile.am
index f38c234..fb9b11e 100644
--- a/i965_drv_video/shaders/render/Makefile.am
+++ b/i965_drv_video/shaders/render/Makefile.am
@@ -7,27 +7,27 @@ INTEL_G4A =				\
 	exa_sf.g4a			\
 	exa_wm_xy.g4a			\
 	exa_wm_src_affine.g4a		\
+	exa_wm_src_sample_argb.g4a	\
 	exa_wm_src_sample_planar.g4a	\
 	exa_wm_yuv_rgb.g4a		\
-	exa_wm_blend_subpicture.g4a		\
 	exa_wm_write.g4a
 
 INTEL_G4B =				\
 	exa_sf.g4b			\
 	exa_wm_xy.g4b			\
 	exa_wm_src_affine.g4b		\
+	exa_wm_src_sample_argb.g4b	\
 	exa_wm_src_sample_planar.g4b	\
 	exa_wm_yuv_rgb.g4b		\
-	exa_wm_blend_subpicture.g4b		\
 	exa_wm_write.g4b
 
 INTEL_G4B_GEN5 =				\
 	exa_sf.g4b.gen5				\
 	exa_wm_xy.g4b.gen5			\
 	exa_wm_src_affine.g4b.gen5		\
+	exa_wm_src_sample_argb.g4b.gen5	\
 	exa_wm_src_sample_planar.g4b.gen5	\
 	exa_wm_yuv_rgb.g4b.gen5			\
-	exa_wm_blend_subpicture.g4b.gen5	\
 	exa_wm_write.g4b.gen5
 
 EXTRA_DIST = $(INTEL_G4I)	\
diff --git a/i965_drv_video/shaders/render/exa_wm_blend_subpicture.g4b b/i965_drv_video/shaders/render/exa_wm_blend_subpicture.g4b
deleted file mode 100644
index bbc75c0..0000000
--- a/i965_drv_video/shaders/render/exa_wm_blend_subpicture.g4b
+++ /dev/null
@@ -1,15 +0,0 @@
-   { 0x00000201, 0x20080061, 0x00000000, 0x0000e000 },
-   { 0x01800031, 0x22001d29, 0x008d0000, 0x02520001 },
-   { 0x01800031, 0x21c01d29, 0x008d0000, 0x02520201 },
-   { 0x01800031, 0x22401d29, 0x008d0000, 0x02520401 },
-   { 0x00600001, 0x228003bd, 0x008d01c0, 0x00000000 },
-   { 0x00600001, 0x22a003bd, 0x008d01e0, 0x00000000 },
-   { 0x00600040, 0x21c07fbd, 0x008d01c0, 0x3f000000 },
-   { 0x00600040, 0x21e07fbd, 0x008d01e0, 0x3f000000 },
-   { 0x00600040, 0x22007fbd, 0x008d0200, 0x3f000000 },
-   { 0x00600040, 0x22207fbd, 0x008d0220, 0x3f000000 },
-   { 0x00600040, 0x22407fbd, 0x008d0240, 0x3f000000 },
-   { 0x00600040, 0x22607fbd, 0x008d0260, 0x3f000000 },
-   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
-   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
-   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
diff --git a/i965_drv_video/shaders/render/exa_wm_blend_subpicture.g4b.gen5 b/i965_drv_video/shaders/render/exa_wm_blend_subpicture.g4b.gen5
deleted file mode 100644
index 0e7534f..0000000
--- a/i965_drv_video/shaders/render/exa_wm_blend_subpicture.g4b.gen5
+++ /dev/null
@@ -1,15 +0,0 @@
-   { 0x00000201, 0x20080061, 0x00000000, 0x0000e000 },
-   { 0x01800031, 0x22001d29, 0x208d0000, 0x0a2a0001 },
-   { 0x01800031, 0x21c01d29, 0x208d0000, 0x0a2a0201 },
-   { 0x01800031, 0x22401d29, 0x208d0000, 0x0a2a0401 },
-   { 0x00600001, 0x228003bd, 0x008d01c0, 0x00000000 },
-   { 0x00600001, 0x22a003bd, 0x008d01e0, 0x00000000 },
-   { 0x00600040, 0x21c07fbd, 0x008d01c0, 0x3f000000 },
-   { 0x00600040, 0x21e07fbd, 0x008d01e0, 0x3f000000 },
-   { 0x00600040, 0x22007fbd, 0x008d0200, 0x3f000000 },
-   { 0x00600040, 0x22207fbd, 0x008d0220, 0x3f000000 },
-   { 0x00600040, 0x22407fbd, 0x008d0240, 0x3f000000 },
-   { 0x00600040, 0x22607fbd, 0x008d0260, 0x3f000000 },
-   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
-   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
-   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
diff --git a/i965_drv_video/shaders/render/exa_wm_blend_subpicture.g4a b/i965_drv_video/shaders/render/exa_wm_src_sample_argb.g4a
index 2d2cd31..c20f53f 100644
--- a/i965_drv_video/shaders/render/exa_wm_blend_subpicture.g4a
+++ b/i965_drv_video/shaders/render/exa_wm_src_sample_argb.g4a
@@ -25,7 +25,8 @@
  *    Keith Packard <keithp@keithp.com>
  */
 
-/* Sample the src surface in planar format */
+/* Sample the src surface */
+
 include(`exa_wm.g4i')
 
 /* prepare sampler read back gX register, which would be written back to output */
@@ -33,51 +34,14 @@ include(`exa_wm.g4i')
 /* use simd16 sampler, param 0 is u, param 1 is v. */
 /* 'payload' loading, assuming tex coord start from g4 */
 
-/* load r */
-mov (1) g0.8<1>UD	0x0000e000UD { align1 mask_disable };
+/* load argb */
+mov (1) g0.8<1>UD	0x00000000UD { align1 mask_disable };
 
 /* src_msg will be copied with g0, as it contains send desc */
 /* emit sampler 'send' cmd */
-
-
-/* sample Y */
-send (16) src_msg_ind		/* msg reg index */
-	src_sample_g<1>UW 	/* readback */
-	g0<8,8,1>UW		/* copy to msg start reg*/
-	sampler (1,0,F)		/* sampler message description, (binding_table,sampler_index,datatype)*/
-	mlen 5 rlen 2 {align1};
-
-/* sample U (Cr) */
-send (16) src_msg_ind		/* msg reg index */
-	src_sample_r<1>UW 	/* readback */
-	g0<8,8,1>UW		/* copy to msg start reg*/
-	sampler (1,2,F)		/* sampler message description, (binding_table,sampler_index,datatype)
-				/* here(src->dst) we should use src_sampler and src_surface */
-	mlen 5 rlen 2 { align1 };   /* required message len 5, readback len 8 */
-	
-/* sample V (Cb) */
 send (16) src_msg_ind		/* msg reg index */
-	src_sample_b<1>UW 	/* readback */
+	src_sample_base<1>UW 	/* readback */
 	g0<8,8,1>UW		/* copy to msg start reg*/
-	sampler (1,4,F)		/* sampler message description, (binding_table,sampler_index,datatype)
+	sampler (1,0,F)		/* sampler message description, (binding_table,sampler_index,datatype)
 				/* here(src->dst) we should use src_sampler and src_surface */
-	mlen 5 rlen 2 { align1 };   /* required message len 5, readback len 8 */
-
-/*extract alpha value, alpha value stores in one float with color value, because the color value is small,
-we can neglect it when doing alpha blend*/
-mov (8) src_sample_a<1>F g14<8,8,1>F { align1 };
-mov (8) src_sample_a_23<1>F g15<8,8,1>F { align1 };
-
-
-/*set subtitle color*/
-add (8) g14<1>F g14<8,8,1>F 0.5F { align1 };
-add (8) g15<1>F g15<8,8,1>F 0.5F { align1 };
-add (8) g16<1>F g16<8,8,1>F 0.5F { align1 };
-add (8) g17<1>F g17<8,8,1>F 0.5F { align1 };
-add (8) g18<1>F g18<8,8,1>F 0.5F { align1 };
-add (8) g19<1>F g19<8,8,1>F 0.5F { align1 };
-
-nop;
-nop;
-nop;
-
+	mlen 5 rlen 8 { align1 };   /* required message len 5, readback len 8 */
diff --git a/i965_drv_video/shaders/render/exa_wm_src_sample_argb.g4b b/i965_drv_video/shaders/render/exa_wm_src_sample_argb.g4b
new file mode 100644
index 0000000..c5b9274
--- /dev/null
+++ b/i965_drv_video/shaders/render/exa_wm_src_sample_argb.g4b
@@ -0,0 +1,2 @@
+   { 0x00000201, 0x20080061, 0x00000000, 0x00000000 },
+   { 0x01800031, 0x21c01d29, 0x008d0000, 0x02580001 },
diff --git a/i965_drv_video/shaders/render/exa_wm_src_sample_argb.g4b.gen5 b/i965_drv_video/shaders/render/exa_wm_src_sample_argb.g4b.gen5
new file mode 100644
index 0000000..f8cb41e
--- /dev/null
+++ b/i965_drv_video/shaders/render/exa_wm_src_sample_argb.g4b.gen5
@@ -0,0 +1,2 @@
+   { 0x00000201, 0x20080061, 0x00000000, 0x00000000 },
+   { 0x01800031, 0x21c01d29, 0x208d0000, 0x0a8a0001 },
diff --git a/src/va.c b/src/va.c
index 2eaa966..6bc9c06 100644
--- a/src/va.c
+++ b/src/va.c
@@ -313,6 +313,8 @@ const char *vaErrorStr(VAStatus error_status)
             return "resolution not supported";
         case VA_STATUS_ERROR_UNIMPLEMENTED:
             return "the requested function is not implemented";
+        case VA_STATUS_ERROR_SURFACE_IN_DISPLAYING:
+            return "surface is in displaying (may by overlay)" ;
         case VA_STATUS_ERROR_UNKNOWN:
             return "unknown libva error";
     }
diff --git a/src/va.h b/src/va.h
index d3f901b..525e192 100644..100755
--- a/src/va.h
+++ b/src/va.h
@@ -127,7 +127,8 @@ typedef int VAStatus;	/* Return status type from functions */
 #define VA_STATUS_ERROR_FLAG_NOT_SUPPORTED      0x00000011
 #define VA_STATUS_ERROR_INVALID_PARAMETER	0x00000012
 #define VA_STATUS_ERROR_RESOLUTION_NOT_SUPPORTED 0x00000013
-#define VA_STATUS_ERROR_UNIMPLEMENTED           0x00000014    
+#define VA_STATUS_ERROR_UNIMPLEMENTED           0x00000014
+#define VA_STATUS_ERROR_SURFACE_IN_DISPLAYING   0x00000015
 #define VA_STATUS_ERROR_UNKNOWN			0xFFFFFFFF
 
 /*
@@ -569,6 +570,7 @@ typedef struct _VASliceParameterBufferMPEG2
     unsigned int slice_data_offset;/* the offset to the first byte of slice data */
     unsigned int slice_data_flag; /* see VA_SLICE_DATA_FLAG_XXX defintions */
     unsigned int macroblock_offset;/* the offset to the first bit of MB from the first byte of slice data */
+    unsigned int slice_horizontal_position;
     unsigned int slice_vertical_position;
     int quantiser_scale_code;
     int intra_slice_flag;
diff --git a/src/x11/dri1_util.c b/src/x11/dri1_util.c
index b3db5b4..3e67e2a 100644
--- a/src/x11/dri1_util.c
+++ b/src/x11/dri1_util.c
@@ -107,7 +107,6 @@ isDRI1Connected(VADriverContextP ctx, char **driver_name)
     
     dri_state->fd = drmOpenOnce(NULL, BusID, &newlyopened);
     XFree(BusID);
-    assert(dri_state->fd >= 0);
 
     if (dri_state->fd < 0)
         goto err_out1;
diff --git a/src/x11/va_x11.c b/src/x11/va_x11.c
index 9de904e..5908fbe 100644
--- a/src/x11/va_x11.c
+++ b/src/x11/va_x11.c
@@ -42,26 +42,6 @@
 
 static VADisplayContextP pDisplayContexts = NULL;
 
-static void va_errorMessage(const char *msg, ...)
-{
-    va_list args;
-
-    fprintf(stderr, "libva error: ");
-    va_start(args, msg);
-    vfprintf(stderr, msg, args);
-    va_end(args);
-}
-
-static void va_infoMessage(const char *msg, ...)
-{
-    va_list args;
-
-    fprintf(stderr, "libva: ");
-    va_start(args, msg);
-    vfprintf(stderr, msg, args);
-    va_end(args);
-}
-
 static int va_DisplayContextIsValid (
     VADisplayContextP pDisplayContext
 )
@@ -132,49 +112,21 @@ static VAStatus va_NVCTRL_GetDriverName (
 )
 {
     VADriverContextP ctx = pDisplayContext->pDriverContext;
-    VAStatus vaStatus = VA_STATUS_ERROR_UNKNOWN;
-    int direct_capable;
-    int driver_major;
-    int driver_minor;
-    int driver_patch;
-    Bool result = True;
-    char *nvidia_driver_name = NULL;
+    int direct_capable, driver_major, driver_minor, driver_patch;
+    Bool result;
 
-    if (result)
-    {
-        result = VA_NVCTRLQueryDirectRenderingCapable(ctx->x11_dpy, ctx->x11_screen, &direct_capable);
-        if (!result)
-        {
-            va_errorMessage("VA_NVCTRLQueryDirectRenderingCapable failed\n");
-        }
-    }
-    if (result)
-    {
-        result = direct_capable;
-        if (!result)
-        {
-            va_errorMessage("VA_NVCTRLQueryDirectRenderingCapable returned false\n");
-        }
-    }
-    if (result)
-    {
-        result = VA_NVCTRLGetClientDriverName(ctx->x11_dpy, ctx->x11_screen, &driver_major, &driver_minor,
-                                              &driver_patch, &nvidia_driver_name);
-        if (!result)
-        {
-            va_errorMessage("VA_NVCTRLGetClientDriverName returned false\n");
-        }
-    }
-    if (result)
-    {
-        vaStatus = VA_STATUS_SUCCESS;
-        va_infoMessage("va_NVCTRL_GetDriverName: %d.%d.%d %s (screen %d)\n",
-                       driver_major, driver_minor, driver_patch,
-                       nvidia_driver_name, ctx->x11_screen);
-	if (driver_name)
-            *driver_name = nvidia_driver_name;
-    }
-    return vaStatus;
+    result = VA_NVCTRLQueryDirectRenderingCapable(ctx->x11_dpy, ctx->x11_screen,
+                                                  &direct_capable);
+    if (!result || !direct_capable)
+        return VA_STATUS_ERROR_UNKNOWN;
+
+    result = VA_NVCTRLGetClientDriverName(ctx->x11_dpy, ctx->x11_screen,
+                                          &driver_major, &driver_minor,
+                                          &driver_patch, driver_name);
+    if (!result)
+        return VA_STATUS_ERROR_UNKNOWN;
+
+    return VA_STATUS_SUCCESS;
 }
 
 static VAStatus va_DisplayContextGetDriverName (