i965_drv_video: store kernel info in the corresponding context

13 files changed, 166 insertions, 154 deletions

diff --git a/i965_drv_video/gen6_vme.c b/i965_drv_video/gen6_vme.c
index e68a98e..1273ff7 100644
--- a/i965_drv_video/gen6_vme.c
+++ b/i965_drv_video/gen6_vme.c
@@ -74,8 +74,6 @@ static struct i965_kernel gen6_vme_kernels[] = {
     }
 };
 
-#define	GEN6_VME_KERNEL_NUMBER ARRAY_ELEMS(gen6_vme_kernels)
-
 static void
 gen6_vme_set_common_surface_tiling(struct i965_surface_state *ss, unsigned int tiling)
 {
@@ -358,7 +356,7 @@ static VAStatus gen6_vme_interface_setup(VADriverContextP ctx,
 
     for (i = 0; i < GEN6_VME_KERNEL_NUMBER; i++) {
         struct i965_kernel *kernel;
-        kernel = &gen6_vme_kernels[i];
+        kernel = &vme_context->vme_kernels[i];
         assert(sizeof(*desc) == 32);
         /*Setup the descritor table*/
         memset(desc, 0, sizeof(*desc));
@@ -708,9 +706,11 @@ Bool gen6_vme_context_init(VADriverContextP ctx, struct gen6_vme_context *vme_co
     struct i965_driver_data *i965 = i965_driver_data(ctx);
     int i;
 
+    memcpy(vme_context->vme_kernels, gen6_vme_kernels, sizeof(vme_context->vme_kernels));
+
     for (i = 0; i < GEN6_VME_KERNEL_NUMBER; i++) {
         /*Load kernel into GPU memory*/	
-        struct i965_kernel *kernel = &gen6_vme_kernels[i];
+        struct i965_kernel *kernel = &vme_context->vme_kernels[i];
 
         kernel->bo = dri_bo_alloc(i965->intel.bufmgr, 
                                   kernel->name, 
@@ -749,7 +749,7 @@ Bool gen6_vme_context_destroy(struct gen6_vme_context *vme_context)
 
     for (i = 0; i < GEN6_VME_KERNEL_NUMBER; i++) {
         /*Load kernel into GPU memory*/	
-        struct i965_kernel *kernel = &gen6_vme_kernels[i];
+        struct i965_kernel *kernel = &vme_context->vme_kernels[i];
 
         dri_bo_unreference(kernel->bo);
         kernel->bo = NULL;
diff --git a/i965_drv_video/gen6_vme.h b/i965_drv_video/gen6_vme.h
index a63bc18..71b14c1 100644
--- a/i965_drv_video/gen6_vme.h
+++ b/i965_drv_video/gen6_vme.h
@@ -38,6 +38,8 @@
 #define MAX_INTERFACE_DESC_GEN6      32
 #define MAX_MEDIA_SURFACES_GEN6      34
 
+#define GEN6_VME_KERNEL_NUMBER          2
+
 struct encode_state;
 struct gen6_encoder_context;
 
@@ -77,6 +79,8 @@ struct gen6_vme_context
         unsigned int size_block; /* in bytes */
         unsigned int pitch;
     } vme_output;
+
+    struct i965_kernel vme_kernels[GEN6_VME_KERNEL_NUMBER];
 };
 
 VAStatus gen6_vme_pipeline(VADriverContextP ctx,
diff --git a/i965_drv_video/i965_avc_hw_scoreboard.c b/i965_drv_video/i965_avc_hw_scoreboard.c
index 633e4b9..9e2263f 100644
--- a/i965_drv_video/i965_avc_hw_scoreboard.c
+++ b/i965_drv_video/i965_avc_hw_scoreboard.c
@@ -41,8 +41,6 @@
 #include "i965_media_h264.h"
 #include "i965_media.h"
 
-extern struct i965_kernel *h264_avc_kernels;
-
 /* On Ironlake */
 #include "shaders/h264/mc/export.inc.gen5"
 
@@ -335,7 +333,7 @@ i965_avc_hw_scoreboard(VADriverContextP ctx, struct decode_state *decode_state,
         avc_hw_scoreboard_context->surface.total_mbs = i965_h264_context->avc_it_command_mb_info.mbs * 2;
         
         dri_bo_unreference(avc_hw_scoreboard_context->hw_kernel.bo);
-        avc_hw_scoreboard_context->hw_kernel.bo = h264_avc_kernels[H264_AVC_COMBINED].bo;
+        avc_hw_scoreboard_context->hw_kernel.bo = i965_h264_context->avc_kernels[H264_AVC_COMBINED].bo;
         assert(avc_hw_scoreboard_context->hw_kernel.bo != NULL);
         dri_bo_reference(avc_hw_scoreboard_context->hw_kernel.bo);
 
diff --git a/i965_drv_video/i965_avc_ildb.c b/i965_drv_video/i965_avc_ildb.c
index 19146ac..739e0cf 100644
--- a/i965_drv_video/i965_avc_ildb.c
+++ b/i965_drv_video/i965_avc_ildb.c
@@ -123,8 +123,6 @@ struct avc_ildb_root_input
     unsigned int pad3;
 };
 
-extern struct i965_kernel *h264_avc_kernels;
-
 #define NUM_AVC_ILDB_INTERFACES ARRAY_ELEMS(avc_ildb_kernel_offset_gen4)
 static unsigned long *avc_ildb_kernel_offset = NULL;
 
@@ -278,7 +276,7 @@ i965_avc_ildb_interface_descriptor_table(VADriverContextP ctx, struct i965_h264_
         int kernel_offset = avc_ildb_kernel_offset[i];
         memset(desc, 0, sizeof(*desc));
         desc->desc0.grf_reg_blocks = 7; 
-        desc->desc0.kernel_start_pointer = (h264_avc_kernels[H264_AVC_COMBINED].bo->offset + kernel_offset) >> 6; /* reloc */
+        desc->desc0.kernel_start_pointer = (i965_h264_context->avc_kernels[H264_AVC_COMBINED].bo->offset + kernel_offset) >> 6; /* reloc */
         desc->desc1.const_urb_entry_read_offset = 0;
         desc->desc1.const_urb_entry_read_len = ((i == AVC_ILDB_ROOT_Y_ILDB_FRAME ||
                                                  i == AVC_ILDB_ROOT_Y_ILDB_FIELD ||
@@ -291,7 +289,7 @@ i965_avc_ildb_interface_descriptor_table(VADriverContextP ctx, struct i965_h264_
                           I915_GEM_DOMAIN_INSTRUCTION, 0,
                           desc->desc0.grf_reg_blocks + kernel_offset,
                           i * sizeof(*desc) + offsetof(struct i965_interface_descriptor, desc0),
-                          h264_avc_kernels[H264_AVC_COMBINED].bo);
+                          i965_h264_context->avc_kernels[H264_AVC_COMBINED].bo);
 
         dri_bo_emit_reloc(bo,
                           I915_GEM_DOMAIN_INSTRUCTION, 0,
diff --git a/i965_drv_video/i965_drv_video.h b/i965_drv_video/i965_drv_video.h
index 2db1b3a..abf16dc 100644
--- a/i965_drv_video/i965_drv_video.h
+++ b/i965_drv_video/i965_drv_video.h
@@ -37,8 +37,6 @@
 
 #include "intel_driver.h"
 
-#include "i965_render.h"
-
 #define I965_MAX_PROFILES                       11
 #define I965_MAX_ENTRYPOINTS                    5
 #define I965_MAX_CONFIG_ATTRIBUTES              10
@@ -204,6 +202,9 @@ struct hw_codec_info
     struct hw_context *(*enc_hw_context_init)(VADriverContextP, VAProfile);
 };
 
+
+#include "i965_render.h"
+
 struct i965_driver_data 
 {
     struct intel_driver_data intel;
diff --git a/i965_drv_video/i965_media_h264.c b/i965_drv_video/i965_media_h264.c
index 1620877..2298361 100644
--- a/i965_drv_video/i965_media_h264.c
+++ b/i965_drv_video/i965_media_h264.c
@@ -275,9 +275,6 @@ static struct i965_kernel h264_avc_kernels_gen5[] = {
     }
 };
 
-#define NUM_H264_AVC_KERNELS (sizeof(h264_avc_kernels_gen4) / sizeof(h264_avc_kernels_gen4[0]))
-struct i965_kernel *h264_avc_kernels = NULL;
-
 #define NUM_AVC_MC_INTERFACES (sizeof(avc_mc_kernel_offset_gen4) / sizeof(avc_mc_kernel_offset_gen4[0]))
 static unsigned long *avc_mc_kernel_offset = NULL;
 
@@ -448,6 +445,7 @@ i965_media_h264_binding_table(VADriverContextP ctx, struct i965_media_context *m
 static void 
 i965_media_h264_interface_descriptor_remap_table(VADriverContextP ctx, struct i965_media_context *media_context)
 {
+    struct i965_h264_context *i965_h264_context = (struct i965_h264_context *)media_context->private_context;
     struct i965_interface_descriptor *desc;
     int i;
     dri_bo *bo;
@@ -461,7 +459,7 @@ i965_media_h264_interface_descriptor_remap_table(VADriverContextP ctx, struct i9
         int kernel_offset = avc_mc_kernel_offset[i];
         memset(desc, 0, sizeof(*desc));
         desc->desc0.grf_reg_blocks = 7; 
-        desc->desc0.kernel_start_pointer = (h264_avc_kernels[H264_AVC_COMBINED].bo->offset + kernel_offset) >> 6; /* reloc */
+        desc->desc0.kernel_start_pointer = (i965_h264_context->avc_kernels[H264_AVC_COMBINED].bo->offset + kernel_offset) >> 6; /* reloc */
         desc->desc1.const_urb_entry_read_offset = 0;
         desc->desc1.const_urb_entry_read_len = 2;
         desc->desc3.binding_table_entry_count = 0;
@@ -472,7 +470,7 @@ i965_media_h264_interface_descriptor_remap_table(VADriverContextP ctx, struct i9
                           I915_GEM_DOMAIN_INSTRUCTION, 0,
                           desc->desc0.grf_reg_blocks + kernel_offset,
                           i * sizeof(*desc) + offsetof(struct i965_interface_descriptor, desc0),
-                          h264_avc_kernels[H264_AVC_COMBINED].bo);
+                          i965_h264_context->avc_kernels[H264_AVC_COMBINED].bo);
 
         dri_bo_emit_reloc(bo,
                           I915_GEM_DOMAIN_INSTRUCTION, 0,
@@ -758,15 +756,16 @@ i965_media_h264_free_private_context(void **data)
     dri_bo_unreference(i965_h264_context->avc_it_command_mb_info.bo);
     dri_bo_unreference(i965_h264_context->avc_it_data.bo);
     dri_bo_unreference(i965_h264_context->avc_ildb_data.bo);
-    free(i965_h264_context);
-    *data = NULL;
 
     for (i = 0; i < NUM_H264_AVC_KERNELS; i++) {
-        struct i965_kernel *kernel = &h264_avc_kernels[i];
+        struct i965_kernel *kernel = &i965_h264_context->avc_kernels[i];
 
         dri_bo_unreference(kernel->bo);
         kernel->bo = NULL;
     }
+
+    free(i965_h264_context);
+    *data = NULL;
 }
 
 void
@@ -850,34 +849,31 @@ i965_media_h264_dec_context_init(VADriverContextP ctx, struct i965_media_context
     i965_h264_context = calloc(1, sizeof(struct i965_h264_context));
 
     /* kernel */
-    if (h264_avc_kernels == NULL) {
-        assert(NUM_H264_AVC_KERNELS == (sizeof(h264_avc_kernels_gen5) / 
-                                        sizeof(h264_avc_kernels_gen5[0])));
-        assert(NUM_AVC_MC_INTERFACES == (sizeof(avc_mc_kernel_offset_gen5) /
-                                         sizeof(avc_mc_kernel_offset_gen5[0])));
-
-        if (IS_IRONLAKE(i965->intel.device_id)) {
-            h264_avc_kernels = h264_avc_kernels_gen5;
-            avc_mc_kernel_offset = avc_mc_kernel_offset_gen5;
-            intra_kernel_header = &intra_kernel_header_gen5;
-            i965_h264_context->use_avc_hw_scoreboard = 1;
-            i965_h264_context->use_hw_w128 = 1;
-        } else {
-            h264_avc_kernels = h264_avc_kernels_gen4;
-            avc_mc_kernel_offset = avc_mc_kernel_offset_gen4;
-            intra_kernel_header = &intra_kernel_header_gen4;
-            i965_h264_context->use_avc_hw_scoreboard = 0;
-            i965_h264_context->use_hw_w128 = 0;
-        }
+    assert(NUM_H264_AVC_KERNELS == (sizeof(h264_avc_kernels_gen5) / 
+                                    sizeof(h264_avc_kernels_gen5[0])));
+    assert(NUM_AVC_MC_INTERFACES == (sizeof(avc_mc_kernel_offset_gen5) /
+                                     sizeof(avc_mc_kernel_offset_gen5[0])));
+    if (IS_IRONLAKE(i965->intel.device_id)) {
+        memcpy(i965_h264_context->avc_kernels, h264_avc_kernels_gen5, sizeof(i965_h264_context->avc_kernels));
+        avc_mc_kernel_offset = avc_mc_kernel_offset_gen5;
+        intra_kernel_header = &intra_kernel_header_gen5;
+        i965_h264_context->use_avc_hw_scoreboard = 1;
+        i965_h264_context->use_hw_w128 = 1;
+    } else {
+        memcpy(i965_h264_context->avc_kernels, h264_avc_kernels_gen4, sizeof(i965_h264_context->avc_kernels));
+        avc_mc_kernel_offset = avc_mc_kernel_offset_gen4;
+        intra_kernel_header = &intra_kernel_header_gen4;
+        i965_h264_context->use_avc_hw_scoreboard = 0;
+        i965_h264_context->use_hw_w128 = 0;
+    }
 
-        for (i = 0; i < NUM_H264_AVC_KERNELS; i++) {
-            struct i965_kernel *kernel = &h264_avc_kernels[i];
-            kernel->bo = dri_bo_alloc(i965->intel.bufmgr, 
-                                      kernel->name, 
-                                      kernel->size, 0x1000);
-            assert(kernel->bo);
-            dri_bo_subdata(kernel->bo, 0, kernel->size, kernel->bin);
-        }
+    for (i = 0; i < NUM_H264_AVC_KERNELS; i++) {
+        struct i965_kernel *kernel = &i965_h264_context->avc_kernels[i];
+        kernel->bo = dri_bo_alloc(i965->intel.bufmgr, 
+                                  kernel->name, 
+                                  kernel->size, 0x1000);
+        assert(kernel->bo);
+        dri_bo_subdata(kernel->bo, 0, kernel->size, kernel->bin);
     }
 
     for (i = 0; i < 16; i++) {
diff --git a/i965_drv_video/i965_media_h264.h b/i965_drv_video/i965_media_h264.h
index 3af7e8d..ba5b649 100644
--- a/i965_drv_video/i965_media_h264.h
+++ b/i965_drv_video/i965_media_h264.h
@@ -20,6 +20,8 @@ enum {
     H264_AVC_NULL
 };
 
+#define NUM_H264_AVC_KERNELS    2
+
 struct i965_h264_context
 {
     struct {
@@ -62,6 +64,8 @@ struct i965_h264_context
         VASurfaceID surface_id;
         int frame_store_id;
     } fsid_list[16];
+
+    struct i965_kernel avc_kernels[NUM_H264_AVC_KERNELS];
 };
 
 void i965_media_h264_decode_init(VADriverContextP ctx, struct decode_state *decode_state, struct i965_media_context *media_context);
diff --git a/i965_drv_video/i965_media_mpeg2.c b/i965_drv_video/i965_media_mpeg2.c
index b44e27e..406ace4 100644
--- a/i965_drv_video/i965_media_mpeg2.c
+++ b/i965_drv_video/i965_media_mpeg2.c
@@ -27,6 +27,7 @@
  *
  */
 
+#include <stdlib.h>
 #include <stdio.h>
 #include <string.h>
 #include <assert.h>
@@ -450,10 +451,6 @@ static struct i965_kernel  mpeg2_vld_kernels_gen5[] = {
     }
 };
 
-static struct i965_kernel  *mpeg2_vld_kernels = NULL;
-
-#define NUM_MPEG2_VLD_KERNELS (sizeof(mpeg2_vld_kernels_gen4)/sizeof(mpeg2_vld_kernels_gen4[0]))
-
 static void
 i965_media_mpeg2_surface_state(VADriverContextP ctx, 
                                int index,
@@ -702,6 +699,7 @@ i965_media_mpeg2_vfe_state(VADriverContextP ctx, struct i965_media_context *medi
 static void 
 i965_media_mpeg2_interface_descriptor_remap_table(VADriverContextP ctx, struct i965_media_context *media_context)
 {
+    struct i965_mpeg2_context *i965_mpeg2_context = (struct i965_mpeg2_context *)media_context->private_context;
     struct i965_interface_descriptor *desc;
     int i;
     dri_bo *bo;
@@ -714,7 +712,7 @@ i965_media_mpeg2_interface_descriptor_remap_table(VADriverContextP ctx, struct i
     for (i = 0; i < NUM_MPEG2_VLD_KERNELS; i++) {
         memset(desc, 0, sizeof(*desc));
         desc->desc0.grf_reg_blocks = 15;
-        desc->desc0.kernel_start_pointer = mpeg2_vld_kernels[i].bo->offset >> 6; /* reloc */
+        desc->desc0.kernel_start_pointer = i965_mpeg2_context->vld_kernels[i].bo->offset >> 6; /* reloc */
         desc->desc1.const_urb_entry_read_offset = 0;
         desc->desc1.const_urb_entry_read_len = 30;
         desc->desc3.binding_table_entry_count = 0;
@@ -725,7 +723,7 @@ i965_media_mpeg2_interface_descriptor_remap_table(VADriverContextP ctx, struct i
                           I915_GEM_DOMAIN_INSTRUCTION, 0,
                           desc->desc0.grf_reg_blocks,
                           i * sizeof(*desc) + offsetof(struct i965_interface_descriptor, desc0),
-                          mpeg2_vld_kernels[i].bo);
+                          i965_mpeg2_context->vld_kernels[i].bo);
 
         dri_bo_emit_reloc(bo,
                           I915_GEM_DOMAIN_INSTRUCTION, 0,
@@ -809,6 +807,7 @@ i965_media_mpeg2_upload_constants(VADriverContextP ctx,
                                   struct decode_state *decode_state,
                                   struct i965_media_context *media_context)
 {
+    struct i965_mpeg2_context *i965_mpeg2_context = (struct i965_mpeg2_context *)media_context->private_context;
     int i, j;
     unsigned char *constant_buffer;
     unsigned char *qmx;
@@ -854,12 +853,12 @@ i965_media_mpeg2_upload_constants(VADriverContextP ctx,
     lib_reloc_offset = 128 + sizeof(idct_table);
     lib_reloc = (unsigned int *)(constant_buffer + lib_reloc_offset);
     for (i = 0; i < 8; i++) {
-        lib_reloc[i] = mpeg2_vld_kernels[LIB_INTERFACE].bo->offset;
+        lib_reloc[i] = i965_mpeg2_context->vld_kernels[LIB_INTERFACE].bo->offset;
         dri_bo_emit_reloc(media_context->curbe.bo,
                           I915_GEM_DOMAIN_INSTRUCTION, 0,
                           0,
                           lib_reloc_offset + i * sizeof(unsigned int),
-                          mpeg2_vld_kernels[LIB_INTERFACE].bo);
+                          i965_mpeg2_context->vld_kernels[LIB_INTERFACE].bo);
     }
 
     dri_bo_unmap(media_context->curbe.bo);
@@ -916,16 +915,21 @@ i965_media_mpeg2_objects(VADriverContextP ctx,
 static void
 i965_media_mpeg2_free_private_context(void **data)
 {
+    struct i965_mpeg2_context *i965_mpeg2_context = *data;
     int i;
 
+    if (i965_mpeg2_context == NULL)
+        return;
+
     for (i = 0; i < NUM_MPEG2_VLD_KERNELS; i++) {
-        struct i965_kernel *kernel = &mpeg2_vld_kernels[i];
+        struct i965_kernel *kernel = &i965_mpeg2_context->vld_kernels[i];
 
         dri_bo_unreference(kernel->bo);
         kernel->bo = NULL;
     }
 
-    mpeg2_vld_kernels = NULL;
+    free(i965_mpeg2_context);
+    *data = NULL;
 }
 
 void 
@@ -952,27 +956,30 @@ void
 i965_media_mpeg2_dec_context_init(VADriverContextP ctx, struct i965_media_context *media_context)
 {
     struct i965_driver_data *i965 = i965_driver_data(ctx);
+    struct i965_mpeg2_context *i965_mpeg2_context;
     int i;
 
+    i965_mpeg2_context = calloc(1, sizeof(struct i965_mpeg2_context));
+
     /* kernel */
-    if (mpeg2_vld_kernels == NULL) {
-        assert(NUM_MPEG2_VLD_KERNELS == (sizeof(mpeg2_vld_kernels_gen5) / 
-                                         sizeof(mpeg2_vld_kernels_gen5[0])));
-        assert(NUM_MPEG2_VLD_KERNELS <= MAX_INTERFACE_DESC);
-
-        if (IS_IRONLAKE(i965->intel.device_id))
-            mpeg2_vld_kernels = mpeg2_vld_kernels_gen5;
-        else
-            mpeg2_vld_kernels = mpeg2_vld_kernels_gen4;
-
-        for (i = 0; i < NUM_MPEG2_VLD_KERNELS; i++) {
-            struct i965_kernel *kernel = &mpeg2_vld_kernels[i];
-            kernel->bo = dri_bo_alloc(i965->intel.bufmgr, 
-                                      kernel->name, 
-                                      kernel->size, 64);
-            assert(kernel->bo);
-            dri_bo_subdata(kernel->bo, 0, kernel->size, kernel->bin);
-        }
+    assert(NUM_MPEG2_VLD_KERNELS == (sizeof(mpeg2_vld_kernels_gen4) / 
+                                     sizeof(mpeg2_vld_kernels_gen4[0])));
+    assert(NUM_MPEG2_VLD_KERNELS == (sizeof(mpeg2_vld_kernels_gen5) / 
+                                     sizeof(mpeg2_vld_kernels_gen5[0])));
+    assert(NUM_MPEG2_VLD_KERNELS <= MAX_INTERFACE_DESC);
+
+    if (IS_IRONLAKE(i965->intel.device_id))
+        memcpy(i965_mpeg2_context->vld_kernels, mpeg2_vld_kernels_gen5, sizeof(i965_mpeg2_context->vld_kernels));
+    else
+        memcpy(i965_mpeg2_context->vld_kernels, mpeg2_vld_kernels_gen4, sizeof(i965_mpeg2_context->vld_kernels));
+
+    for (i = 0; i < NUM_MPEG2_VLD_KERNELS; i++) {
+        struct i965_kernel *kernel = &i965_mpeg2_context->vld_kernels[i];
+        kernel->bo = dri_bo_alloc(i965->intel.bufmgr, 
+                                  kernel->name, 
+                                  kernel->size, 64);
+        assert(kernel->bo);
+        dri_bo_subdata(kernel->bo, 0, kernel->size, kernel->bin);
     }
 
     /* URB */
diff --git a/i965_drv_video/i965_media_mpeg2.h b/i965_drv_video/i965_media_mpeg2.h
index 7da9b78..40d91f1 100644
--- a/i965_drv_video/i965_media_mpeg2.h
+++ b/i965_drv_video/i965_media_mpeg2.h
@@ -38,6 +38,13 @@
 struct decode_state;
 struct i965_media_context;
 
+#define NUM_MPEG2_VLD_KERNELS           15
+
+struct i965_mpeg2_context
+{
+    struct i965_kernel vld_kernels[NUM_MPEG2_VLD_KERNELS];
+};
+
 void i965_media_mpeg2_decode_init(VADriverContextP ctx, struct decode_state * decode_state, struct i965_media_context *media_context);
 void i965_media_mpeg2_dec_context_init(VADriverContextP ctx, struct i965_media_context *media_context);
 
diff --git a/i965_drv_video/i965_post_processing.c b/i965_drv_video/i965_post_processing.c
index 4dba0c6..25c27ce 100644
--- a/i965_drv_video/i965_post_processing.c
+++ b/i965_drv_video/i965_post_processing.c
@@ -37,24 +37,13 @@
 #include "intel_driver.h"
 #include "i965_defines.h"
 #include "i965_structs.h"
-
+#include "i965_drv_video.h"
 #include "i965_post_processing.h"
 #include "i965_render.h"
-#include "i965_drv_video.h"
 
 #define HAS_PP(ctx) (IS_IRONLAKE((ctx)->intel.device_id) ||     \
                      IS_GEN6((ctx)->intel.device_id))
 
-struct pp_module
-{
-    struct i965_kernel kernel;
-    
-    /* others */
-    void (*initialize)(VADriverContextP ctx, VASurfaceID surface, int input,
-                       unsigned short srcw, unsigned short srch,
-                       unsigned short destw, unsigned short desth);
-};
-
 static const uint32_t pp_null_gen5[][4] = {
 #include "shaders/post_processing/null.g4b.gen5"
 };
@@ -235,10 +224,6 @@ static struct pp_module pp_modules_gen6[] = {
     },
 };
 
-#define NUM_PP_MODULES ARRAY_ELEMS(pp_modules_gen5)
-
-static struct pp_module *pp_modules = NULL;
-
 struct pp_static_parameter
 {
     struct {
@@ -503,7 +488,7 @@ ironlake_pp_interface_descriptor_table(struct i965_post_processing_context *pp_c
     desc = bo->virtual;
     memset(desc, 0, sizeof(*desc));
     desc->desc0.grf_reg_blocks = 10;
-    desc->desc0.kernel_start_pointer = pp_modules[pp_index].kernel.bo->offset >> 6; /* reloc */
+    desc->desc0.kernel_start_pointer = pp_context->pp_modules[pp_index].kernel.bo->offset >> 6; /* reloc */
     desc->desc1.const_urb_entry_read_offset = 0;
     desc->desc1.const_urb_entry_read_len = 4; /* grf 1-4 */
     desc->desc2.sampler_state_pointer = pp_context->sampler_state_table.bo->offset >> 5;
@@ -516,7 +501,7 @@ ironlake_pp_interface_descriptor_table(struct i965_post_processing_context *pp_c
                       I915_GEM_DOMAIN_INSTRUCTION, 0,
                       desc->desc0.grf_reg_blocks,
                       offsetof(struct i965_interface_descriptor, desc0),
-                      pp_modules[pp_index].kernel.bo);
+                      pp_context->pp_modules[pp_index].kernel.bo);
 
     dri_bo_emit_reloc(bo,
                       I915_GEM_DOMAIN_INSTRUCTION, 0,
@@ -2036,9 +2021,8 @@ ironlake_pp_initialize(VADriverContextP ctx,
     memset(&pp_static_parameter, 0, sizeof(pp_static_parameter));
     memset(&pp_inline_parameter, 0, sizeof(pp_inline_parameter));
     assert(pp_index >= PP_NULL && pp_index < NUM_PP_MODULES);
-    assert(pp_modules);
     pp_context->current_pp = pp_index;
-    pp_module = &pp_modules[pp_index];
+    pp_module = &pp_context->pp_modules[pp_index];
     
     if (pp_module->initialize)
         pp_module->initialize(ctx, surface, input, srcw, srch, destw, desth);
@@ -2157,9 +2141,8 @@ gen6_pp_initialize(VADriverContextP ctx,
     memset(&pp_static_parameter, 0, sizeof(pp_static_parameter));
     memset(&pp_inline_parameter, 0, sizeof(pp_inline_parameter));
     assert(pp_index >= PP_NULL && pp_index < NUM_PP_MODULES);
-    assert(pp_modules);
     pp_context->current_pp = pp_index;
-    pp_module = &pp_modules[pp_index];
+    pp_module = &pp_context->pp_modules[pp_index];
     
     if (pp_module->initialize)
         pp_module->initialize(ctx, surface, input, srcw, srch, destw, desth);
@@ -2207,7 +2190,7 @@ gen6_pp_interface_descriptor_table(struct i965_post_processing_context *pp_conte
     desc = bo->virtual;
     memset(desc, 0, sizeof(*desc));
     desc->desc0.kernel_start_pointer = 
-        pp_modules[pp_index].kernel.bo->offset >> 6; /* reloc */
+        pp_context->pp_modules[pp_index].kernel.bo->offset >> 6; /* reloc */
     desc->desc1.single_program_flow = 1;
     desc->desc1.floating_point_mode = FLOATING_POINT_IEEE_754;
     desc->desc2.sampler_count = 1;      /* 1 - 4 samplers used */
@@ -2223,7 +2206,7 @@ gen6_pp_interface_descriptor_table(struct i965_post_processing_context *pp_conte
                       I915_GEM_DOMAIN_INSTRUCTION, 0,
                       0,
                       offsetof(struct gen6_interface_descriptor_data, desc0),
-                      pp_modules[pp_index].kernel.bo);
+                      pp_context->pp_modules[pp_index].kernel.bo);
 
     dri_bo_emit_reloc(bo,
                       I915_GEM_DOMAIN_INSTRUCTION, 0,
@@ -2517,17 +2500,17 @@ i965_post_processing_terminate(VADriverContextP ctx)
             dri_bo_unreference(pp_context->stmm.bo);
             pp_context->stmm.bo = NULL;
 
+            for (i = 0; i < NUM_PP_MODULES; i++) {
+                struct pp_module *pp_module = &pp_context->pp_modules[i];
+
+                dri_bo_unreference(pp_module->kernel.bo);
+                pp_module->kernel.bo = NULL;
+            }
+
             free(pp_context);
         }
 
         i965->pp_context = NULL;
-
-        for (i = 0; i < NUM_PP_MODULES && pp_modules; i++) {
-            struct pp_module *pp_module = &pp_modules[i];
-
-            dri_bo_unreference(pp_module->kernel.bo);
-            pp_module->kernel.bo = NULL;
-        }
     }
 
     return True;
@@ -2544,36 +2527,36 @@ i965_post_processing_init(VADriverContextP ctx)
         if (pp_context == NULL) {
             pp_context = calloc(1, sizeof(*pp_context));
             i965->pp_context = pp_context;
-        }
 
-        pp_context->urb.size = URB_SIZE((&i965->intel));
-        pp_context->urb.num_vfe_entries = 32;
-        pp_context->urb.size_vfe_entry = 1;     /* in 512 bits unit */
-        pp_context->urb.num_cs_entries = 1;
-        pp_context->urb.size_cs_entry = 2;      /* in 512 bits unit */
-        pp_context->urb.vfe_start = 0;
-        pp_context->urb.cs_start = pp_context->urb.vfe_start + 
-            pp_context->urb.num_vfe_entries * pp_context->urb.size_vfe_entry;
-        assert(pp_context->urb.cs_start + 
-               pp_context->urb.num_cs_entries * pp_context->urb.size_cs_entry <= URB_SIZE((&i965->intel)));
-
-        assert(NUM_PP_MODULES == ARRAY_ELEMS(pp_modules_gen6));
-
-        if (IS_GEN6(i965->intel.device_id))
-            pp_modules = pp_modules_gen6;
-        else if (IS_IRONLAKE(i965->intel.device_id)) {
-            pp_modules = pp_modules_gen5;
-        }
-
-        for (i = 0; i < NUM_PP_MODULES && pp_modules; i++) {
-            struct pp_module *pp_module = &pp_modules[i];
-            dri_bo_unreference(pp_module->kernel.bo);
-            pp_module->kernel.bo = dri_bo_alloc(i965->intel.bufmgr,
-                                                pp_module->kernel.name,
-                                                pp_module->kernel.size,
-                                                4096);
-            assert(pp_module->kernel.bo);
-            dri_bo_subdata(pp_module->kernel.bo, 0, pp_module->kernel.size, pp_module->kernel.bin);
+            pp_context->urb.size = URB_SIZE((&i965->intel));
+            pp_context->urb.num_vfe_entries = 32;
+            pp_context->urb.size_vfe_entry = 1;     /* in 512 bits unit */
+            pp_context->urb.num_cs_entries = 1;
+            pp_context->urb.size_cs_entry = 2;      /* in 512 bits unit */
+            pp_context->urb.vfe_start = 0;
+            pp_context->urb.cs_start = pp_context->urb.vfe_start + 
+                pp_context->urb.num_vfe_entries * pp_context->urb.size_vfe_entry;
+            assert(pp_context->urb.cs_start + 
+                   pp_context->urb.num_cs_entries * pp_context->urb.size_cs_entry <= URB_SIZE((&i965->intel)));
+
+            assert(NUM_PP_MODULES == ARRAY_ELEMS(pp_modules_gen5));
+            assert(NUM_PP_MODULES == ARRAY_ELEMS(pp_modules_gen6));
+
+            if (IS_GEN6(i965->intel.device_id))
+                memcpy(pp_context->pp_modules, pp_modules_gen6, sizeof(pp_context->pp_modules));
+            else if (IS_IRONLAKE(i965->intel.device_id))
+                memcpy(pp_context->pp_modules, pp_modules_gen5, sizeof(pp_context->pp_modules));
+
+            for (i = 0; i < NUM_PP_MODULES; i++) {
+                struct pp_module *pp_module = &pp_context->pp_modules[i];
+                dri_bo_unreference(pp_module->kernel.bo);
+                pp_module->kernel.bo = dri_bo_alloc(i965->intel.bufmgr,
+                                                    pp_module->kernel.name,
+                                                    pp_module->kernel.size,
+                                                    4096);
+                assert(pp_module->kernel.bo);
+                dri_bo_subdata(pp_module->kernel.bo, 0, pp_module->kernel.size, pp_module->kernel.bin);
+            }
         }
     }
 
diff --git a/i965_drv_video/i965_post_processing.h b/i965_drv_video/i965_post_processing.h
index fa2d701..bfbe448 100644
--- a/i965_drv_video/i965_post_processing.h
+++ b/i965_drv_video/i965_post_processing.h
@@ -43,6 +43,8 @@ enum
     PP_NV12_DNDI,
 };
 
+#define NUM_PP_MODULES                  5
+
 struct pp_load_save_context
 {
     int dest_w;
@@ -70,9 +72,20 @@ struct pp_dndi_context
 
 };
 
+struct pp_module
+{
+    struct i965_kernel kernel;
+    
+    /* others */
+    void (*initialize)(VADriverContextP ctx, VASurfaceID surface, int input,
+                       unsigned short srcw, unsigned short srch,
+                       unsigned short destw, unsigned short desth);
+};
+
 struct i965_post_processing_context
 {
     int current_pp;
+    struct pp_module pp_modules[NUM_PP_MODULES];
 
     struct {
         dri_bo *bo;
diff --git a/i965_drv_video/i965_render.c b/i965_drv_video/i965_render.c
index 7794303..c04b72c 100644
--- a/i965_drv_video/i965_render.c
+++ b/i965_drv_video/i965_render.c
@@ -213,10 +213,6 @@ static struct i965_kernel render_kernels_gen6[] = {
     }
 };
 
-static struct i965_kernel *render_kernels = NULL;
-
-#define NUM_RENDER_KERNEL (sizeof(render_kernels_gen4)/sizeof(render_kernels_gen4[0]))
-
 #define URB_VS_ENTRIES	      8
 #define URB_VS_ENTRY_SIZE     1
 
@@ -269,7 +265,7 @@ i965_render_sf_unit(VADriverContextP ctx)
     memset(sf_state, 0, sizeof(*sf_state));
 
     sf_state->thread0.grf_reg_count = I965_GRF_BLOCKS(SF_KERNEL_NUM_GRF);
-    sf_state->thread0.kernel_start_pointer = render_kernels[SF_KERNEL].bo->offset >> 6;
+    sf_state->thread0.kernel_start_pointer = render_state->render_kernels[SF_KERNEL].bo->offset >> 6;
 
     sf_state->sf1.single_program_flow = 1; /* XXX */
     sf_state->sf1.binding_table_entry_count = 0;
@@ -308,7 +304,7 @@ i965_render_sf_unit(VADriverContextP ctx)
                       I915_GEM_DOMAIN_INSTRUCTION, 0,
                       sf_state->thread0.grf_reg_count << 1,
                       offsetof(struct i965_sf_unit_state, thread0),
-                      render_kernels[SF_KERNEL].bo);
+                      render_state->render_kernels[SF_KERNEL].bo);
 
     dri_bo_unmap(render_state->sf.state);
 }
@@ -354,7 +350,7 @@ i965_subpic_render_wm_unit(VADriverContextP ctx)
     memset(wm_state, 0, sizeof(*wm_state));
 
     wm_state->thread0.grf_reg_count = I965_GRF_BLOCKS(PS_KERNEL_NUM_GRF);
-    wm_state->thread0.kernel_start_pointer = render_kernels[PS_SUBPIC_KERNEL].bo->offset >> 6;
+    wm_state->thread0.kernel_start_pointer = render_state->render_kernels[PS_SUBPIC_KERNEL].bo->offset >> 6;
 
     wm_state->thread1.single_program_flow = 1; /* XXX */
 
@@ -392,7 +388,7 @@ i965_subpic_render_wm_unit(VADriverContextP ctx)
                       I915_GEM_DOMAIN_INSTRUCTION, 0,
                       wm_state->thread0.grf_reg_count << 1,
                       offsetof(struct i965_wm_unit_state, thread0),
-                      render_kernels[PS_SUBPIC_KERNEL].bo);
+                      render_state->render_kernels[PS_SUBPIC_KERNEL].bo);
 
     dri_bo_emit_reloc(render_state->wm.state,
                       I915_GEM_DOMAIN_INSTRUCTION, 0,
@@ -419,7 +415,7 @@ i965_render_wm_unit(VADriverContextP ctx)
     memset(wm_state, 0, sizeof(*wm_state));
 
     wm_state->thread0.grf_reg_count = I965_GRF_BLOCKS(PS_KERNEL_NUM_GRF);
-    wm_state->thread0.kernel_start_pointer = render_kernels[PS_KERNEL].bo->offset >> 6;
+    wm_state->thread0.kernel_start_pointer = render_state->render_kernels[PS_KERNEL].bo->offset >> 6;
 
     wm_state->thread1.single_program_flow = 1; /* XXX */
 
@@ -457,7 +453,7 @@ i965_render_wm_unit(VADriverContextP ctx)
                       I915_GEM_DOMAIN_INSTRUCTION, 0,
                       wm_state->thread0.grf_reg_count << 1,
                       offsetof(struct i965_wm_unit_state, thread0),
-                      render_kernels[PS_KERNEL].bo);
+                      render_state->render_kernels[PS_KERNEL].bo);
 
     dri_bo_emit_reloc(render_state->wm.state,
                       I915_GEM_DOMAIN_INSTRUCTION, 0,
@@ -1821,7 +1817,7 @@ gen6_emit_wm_state(VADriverContextP ctx, int kernel)
     OUT_BATCH(ctx, 0);
 
     OUT_BATCH(ctx, GEN6_3DSTATE_WM | (9 - 2));
-    OUT_RELOC(ctx, render_kernels[kernel].bo,
+    OUT_RELOC(ctx, render_state->render_kernels[kernel].bo,
               I915_GEM_DOMAIN_INSTRUCTION, 0,
               0);
     OUT_BATCH(ctx, (1 << GEN6_3DSTATE_WM_SAMPLER_COUNT_SHITF) |
@@ -2088,14 +2084,14 @@ i965_render_init(VADriverContextP ctx)
                                  sizeof(render_kernels_gen6[0])));
 
     if (IS_GEN6(i965->intel.device_id))
-        render_kernels = render_kernels_gen6;
+        memcpy(render_state->render_kernels, render_kernels_gen6, sizeof(render_state->render_kernels));
     else if (IS_IRONLAKE(i965->intel.device_id))
-        render_kernels = render_kernels_gen5;
+        memcpy(render_state->render_kernels, render_kernels_gen5, sizeof(render_state->render_kernels));
     else
-        render_kernels = render_kernels_gen4;
+        memcpy(render_state->render_kernels, render_kernels_gen4, sizeof(render_state->render_kernels));
 
     for (i = 0; i < NUM_RENDER_KERNEL; i++) {
-        struct i965_kernel *kernel = &render_kernels[i];
+        struct i965_kernel *kernel = &render_state->render_kernels[i];
 
         if (!kernel->size)
             continue;
@@ -2128,7 +2124,7 @@ i965_render_terminate(VADriverContextP ctx)
     render_state->curbe.bo = NULL;
 
     for (i = 0; i < NUM_RENDER_KERNEL; i++) {
-        struct i965_kernel *kernel = &render_kernels[i];
+        struct i965_kernel *kernel = &render_state->render_kernels[i];
         
         dri_bo_unreference(kernel->bo);
         kernel->bo = NULL;
diff --git a/i965_drv_video/i965_render.h b/i965_drv_video/i965_render.h
index dc7bcc2..fd4be8f 100644
--- a/i965_drv_video/i965_render.h
+++ b/i965_drv_video/i965_render.h
@@ -31,8 +31,12 @@
 #define MAX_SAMPLERS            16
 #define MAX_RENDER_SURFACES     (MAX_SAMPLERS + 1)
 
+#define NUM_RENDER_KERNEL       3
+
 #include "i965_post_processing.h"
 
+struct i965_kernel;
+
 struct i965_render_state
 {
     struct {
@@ -71,7 +75,8 @@ struct i965_render_state
     struct intel_region *draw_region;
 
     int pp_flag; /* 0: disable, 1: enable */
-    struct i965_post_processing_context pp_context;
+
+    struct i965_kernel render_kernels[3];
 };
 
 Bool i965_render_init(VADriverContextP ctx);