Follow the HW spec to configure the buffer cache on Gen9+

The MOCS field is used to define the cache type for the given buffer. From the
SKL+, the MOCS field is interpreted as the index that is used to find the corresponding
cache type in kernel driver. The current MOCS setting causes that buffer uses
the wrong cache type.

Signed-off-by: Zhao Yakui <yakui.zhao@intel.com>
Reviewed-by: Sean V Kelley <sean.v.kelley@intel.com>

10 files changed, 104 insertions, 65 deletions

diff --git a/src/gen75_vpp_vebox.c b/src/gen75_vpp_vebox.c
index 0c52765..eee8e76 100644
--- a/src/gen75_vpp_vebox.c
+++ b/src/gen75_vpp_vebox.c
@@ -2292,12 +2292,13 @@ void skl_veb_state_table_setup(VADriverContextP ctx, struct intel_vebox_context
 void
 skl_veb_state_command(VADriverContextP ctx, struct intel_vebox_context *proc_ctx)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = proc_ctx->batch;
 
     BEGIN_VEB_BATCH(batch, 0x10);
     OUT_VEB_BATCH(batch, VEB_STATE | (0x10 - 2));
     OUT_VEB_BATCH(batch,
-                  0 << 25 |       // state surface control bits
+                  ((i965->intel.mocs_state) << 25) |       // state surface control bits
                   0 << 23 |       // reserved.
                   0 << 22 |       // gamut expansion position
                   0 << 15 |       // reserved.
diff --git a/src/gen8_mfc.c b/src/gen8_mfc.c
index 9ea7423..9b50f9a 100644
--- a/src/gen8_mfc.c
+++ b/src/gen8_mfc.c
@@ -232,6 +232,7 @@ static void
 gen8_mfc_ind_obj_base_addr_state(VADriverContextP ctx,
                                  struct intel_encoder_context *encoder_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = encoder_context->base.batch;
     struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
     struct gen6_vme_context *vme_context = encoder_context->vme_context;
@@ -263,7 +264,7 @@ gen8_mfc_ind_obj_base_addr_state(VADriverContextP ctx,
         /* the DW6-10 is for MFX Indirect MV Object Base Address */
         OUT_BCS_RELOC(batch, vme_context->vme_output.bo, I915_GEM_DOMAIN_INSTRUCTION, 0, 0);
         OUT_BCS_BATCH(batch, 0);
-        OUT_BCS_BATCH(batch, 0);
+        OUT_BCS_BATCH(batch, i965->intel.mocs_state);
         OUT_BCS_RELOC(batch, vme_context->vme_output.bo, I915_GEM_DOMAIN_INSTRUCTION, 0, vme_size);
         OUT_BCS_BATCH(batch, 0);
     } else {
@@ -296,7 +297,7 @@ gen8_mfc_ind_obj_base_addr_state(VADriverContextP ctx,
                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                   bse_offset);
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 	
     OUT_BCS_RELOC(batch,
                   mfc_context->mfc_indirect_pak_bse_object.bo,
@@ -642,6 +643,7 @@ static void
 gen8_mfc_pipe_buf_addr_state(VADriverContextP ctx,
                              struct intel_encoder_context *encoder_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = encoder_context->base.batch;
     struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
     int i;
@@ -659,7 +661,7 @@ gen8_mfc_pipe_buf_addr_state(VADriverContextP ctx,
         OUT_BCS_BATCH(batch, 0);											/* pre output addr   */
 
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
     /* the DW4-6 is for the post_deblocking */
 
     if (mfc_context->post_deblocking_output.bo)
@@ -670,7 +672,7 @@ gen8_mfc_pipe_buf_addr_state(VADriverContextP ctx,
         OUT_BCS_BATCH(batch, 0);
     
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* the DW7-9 is for the uncompressed_picture */
     OUT_BCS_RELOC(batch, mfc_context->uncompressed_picture_source.bo,
@@ -678,7 +680,7 @@ gen8_mfc_pipe_buf_addr_state(VADriverContextP ctx,
                   0); /* uncompressed data */
 
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* the DW10-12 is for the mb status */
     OUT_BCS_RELOC(batch, mfc_context->macroblock_status_buffer.bo,
@@ -686,7 +688,7 @@ gen8_mfc_pipe_buf_addr_state(VADriverContextP ctx,
                   0); /* StreamOut data*/
     
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* the DW13-15 is for the intra_row_store_scratch */
     OUT_BCS_RELOC(batch, mfc_context->intra_row_store_scratch_buffer.bo,
@@ -694,7 +696,7 @@ gen8_mfc_pipe_buf_addr_state(VADriverContextP ctx,
                   0);	
 
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* the DW16-18 is for the deblocking filter */
     OUT_BCS_RELOC(batch, mfc_context->deblocking_filter_row_store_scratch_buffer.bo,
@@ -702,7 +704,7 @@ gen8_mfc_pipe_buf_addr_state(VADriverContextP ctx,
                   0);
 
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* the DW 19-50 is for Reference pictures*/
     for (i = 0; i < ARRAY_ELEMS(mfc_context->reference_surfaces); i++) {
@@ -717,7 +719,7 @@ gen8_mfc_pipe_buf_addr_state(VADriverContextP ctx,
 	OUT_BCS_BATCH(batch, 0);
     }
 
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* The DW 52-54 is for the MB status buffer */
     OUT_BCS_RELOC(batch, mfc_context->macroblock_status_buffer.bo,
@@ -725,7 +727,7 @@ gen8_mfc_pipe_buf_addr_state(VADriverContextP ctx,
                   0);											/* Macroblock status buffer*/
 	
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* the DW 55-57 is the ILDB buffer */
     OUT_BCS_BATCH(batch, 0);
@@ -744,6 +746,7 @@ static void
 gen8_mfc_avc_directmode_state(VADriverContextP ctx,
                               struct intel_encoder_context *encoder_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = encoder_context->base.batch;
     struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
 
@@ -767,7 +770,7 @@ gen8_mfc_avc_directmode_state(VADriverContextP ctx,
         }
     }
     
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* the DW34-36 is the MV for the current reference */
     OUT_BCS_RELOC(batch, mfc_context->direct_mv_buffers[NUM_MFC_DMV_BUFFERS - 2].bo,
@@ -775,7 +778,7 @@ gen8_mfc_avc_directmode_state(VADriverContextP ctx,
                   0);
 
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* POL list */
     for(i = 0; i < 32; i++) {
@@ -792,6 +795,7 @@ static void
 gen8_mfc_bsp_buf_base_addr_state(VADriverContextP ctx,
                                  struct intel_encoder_context *encoder_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = encoder_context->base.batch;
     struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
 
@@ -802,7 +806,7 @@ gen8_mfc_bsp_buf_base_addr_state(VADriverContextP ctx,
                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                   0);
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 	
     /* the DW4-6 is for MPR Row Store Scratch Buffer Base Address */
     OUT_BCS_BATCH(batch, 0);
@@ -4063,13 +4067,14 @@ gen8_mfc_vp8_pic_state(VADriverContextP ctx,
     else                                                                \
         OUT_BCS_BATCH(batch, 0);                                        \
     OUT_BCS_BATCH(batch, 0);                                            \
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
 static void 
 gen8_mfc_vp8_bsp_buf_base_addr_state(VADriverContextP ctx, 
                                      struct encode_state *encode_state,
                                      struct intel_encoder_context *encoder_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = encoder_context->base.batch;
     struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
 
diff --git a/src/gen8_mfd.c b/src/gen8_mfd.c
index 9852664..61999b3 100644
--- a/src/gen8_mfd.c
+++ b/src/gen8_mfd.c
@@ -179,6 +179,7 @@ gen8_mfd_pipe_buf_addr_state(VADriverContextP ctx,
                              int standard_select,
                              struct gen7_mfd_context *gen7_mfd_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = gen7_mfd_context->base.batch;
     int i;
 
@@ -193,7 +194,8 @@ gen8_mfd_pipe_buf_addr_state(VADriverContextP ctx,
         OUT_BCS_BATCH(batch, 0);
 
 	OUT_BCS_BATCH(batch, 0);
-	OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
+
 	/* Post-debloing 4-6 */
     if (gen7_mfd_context->post_deblocking_output.valid)
         OUT_BCS_RELOC(batch, gen7_mfd_context->post_deblocking_output.bo,
@@ -203,7 +205,7 @@ gen8_mfd_pipe_buf_addr_state(VADriverContextP ctx,
         OUT_BCS_BATCH(batch, 0);
 
 	OUT_BCS_BATCH(batch, 0);
-	OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
 	/* uncompressed-video & stream out 7-12 */
     OUT_BCS_BATCH(batch, 0); /* ignore for decoding */
@@ -222,7 +224,8 @@ gen8_mfd_pipe_buf_addr_state(VADriverContextP ctx,
         OUT_BCS_BATCH(batch, 0);
 
 	OUT_BCS_BATCH(batch, 0);
-	OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
+
 	/* deblocking-filter-row-store 16-18 */
     if (gen7_mfd_context->deblocking_filter_row_store_scratch_buffer.valid)
         OUT_BCS_RELOC(batch, gen7_mfd_context->deblocking_filter_row_store_scratch_buffer.bo,
@@ -231,7 +234,8 @@ gen8_mfd_pipe_buf_addr_state(VADriverContextP ctx,
     else
         OUT_BCS_BATCH(batch, 0);
 	OUT_BCS_BATCH(batch, 0);
-	OUT_BCS_BATCH(batch, 0);
+
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 19..50 */
     for (i = 0; i < ARRAY_ELEMS(gen7_mfd_context->reference_surface); i++) {
@@ -253,7 +257,7 @@ gen8_mfd_pipe_buf_addr_state(VADriverContextP ctx,
     }
     
     /* reference property 51 */
-    OUT_BCS_BATCH(batch, 0);  
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 	
     /* Macroblock status & ILDB 52-57 */
     OUT_BCS_BATCH(batch, 0);
@@ -278,13 +282,14 @@ gen8_mfd_ind_obj_base_addr_state(VADriverContextP ctx,
                                  struct gen7_mfd_context *gen7_mfd_context)
 {
     struct intel_batchbuffer *batch = gen7_mfd_context->base.batch;
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
 
     BEGIN_BCS_BATCH(batch, 26);
     OUT_BCS_BATCH(batch, MFX_IND_OBJ_BASE_ADDR_STATE | (26 - 2));
 	/* MFX In BS 1-5 */
     OUT_BCS_RELOC(batch, slice_data_bo, I915_GEM_DOMAIN_INSTRUCTION, 0, 0); /* MFX Indirect Bitstream Object Base Address */
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 	/* Upper bound 4-5 */	
     OUT_BCS_BATCH(batch, 0);
     OUT_BCS_BATCH(batch, 0);
@@ -327,6 +332,7 @@ gen8_mfd_bsp_buf_base_addr_state(VADriverContextP ctx,
                                  struct gen7_mfd_context *gen7_mfd_context)
 {
     struct intel_batchbuffer *batch = gen7_mfd_context->base.batch;
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
 
     BEGIN_BCS_BATCH(batch, 10);
     OUT_BCS_BATCH(batch, MFX_BSP_BUF_BASE_ADDR_STATE | (10 - 2));
@@ -339,7 +345,7 @@ gen8_mfd_bsp_buf_base_addr_state(VADriverContextP ctx,
     		OUT_BCS_BATCH(batch, 0);
 		
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 	/* MPR Row Store Scratch buffer 4-6 */
     if (gen7_mfd_context->mpr_row_store_scratch_buffer.valid)
         OUT_BCS_RELOC(batch, gen7_mfd_context->mpr_row_store_scratch_buffer.bo,
@@ -349,7 +355,7 @@ gen8_mfd_bsp_buf_base_addr_state(VADriverContextP ctx,
         OUT_BCS_BATCH(batch, 0);
 
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
 	/* Bitplane 7-9 */ 
     if (gen7_mfd_context->bitplane_read_buffer.valid)
@@ -359,7 +365,7 @@ gen8_mfd_bsp_buf_base_addr_state(VADriverContextP ctx,
     else
     	OUT_BCS_BATCH(batch, 0);
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
     ADVANCE_BCS_BATCH(batch);
 }
 
@@ -510,6 +516,7 @@ gen8_mfd_avc_directmode_state(VADriverContextP ctx,
                               VASliceParameterBufferH264 *slice_param,
                               struct gen7_mfd_context *gen7_mfd_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = gen7_mfd_context->base.batch;
     struct object_surface *obj_surface;
     GenAvcSurface *gen7_avc_surface;
@@ -538,7 +545,7 @@ gen8_mfd_avc_directmode_state(VADriverContextP ctx,
         }
     }
     
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* the current decoding frame/field */
     va_pic = &pic_param->CurrPic;
@@ -551,7 +558,7 @@ gen8_mfd_avc_directmode_state(VADriverContextP ctx,
                   0);
 
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* POC List */
     for (i = 0; i < ARRAY_ELEMS(gen7_mfd_context->reference_surface); i++) {
@@ -1712,6 +1719,7 @@ gen8_mfd_vc1_directmode_state(VADriverContextP ctx,
                               struct decode_state *decode_state,
                               struct gen7_mfd_context *gen7_mfd_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = gen7_mfd_context->base.batch;
     struct object_surface *obj_surface;
     dri_bo *dmv_read_buffer = NULL, *dmv_write_buffer = NULL;
@@ -1739,7 +1747,7 @@ gen8_mfd_vc1_directmode_state(VADriverContextP ctx,
         OUT_BCS_BATCH(batch, 0);
 
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     if (dmv_read_buffer)
         OUT_BCS_RELOC(batch, dmv_read_buffer,
@@ -1749,7 +1757,7 @@ gen8_mfd_vc1_directmode_state(VADriverContextP ctx,
         OUT_BCS_BATCH(batch, 0);
     
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
                   
     ADVANCE_BCS_BATCH(batch);
 }
@@ -2308,7 +2316,7 @@ gen8_jpeg_wa_pipe_buf_addr_state(VADriverContextP ctx,
                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                   0);
 	OUT_BCS_BATCH(batch, 0);
-	OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
     
 
     OUT_BCS_BATCH(batch, 0); /* post deblocking */
@@ -2329,7 +2337,8 @@ gen8_jpeg_wa_pipe_buf_addr_state(VADriverContextP ctx,
                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                   0);
 	OUT_BCS_BATCH(batch, 0);
-	OUT_BCS_BATCH(batch, 0);
+
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
 	/* the DW 16-18 is for deblocking filter */ 
     OUT_BCS_BATCH(batch, 0);
@@ -2387,14 +2396,14 @@ gen8_jpeg_wa_bsp_buf_base_addr_state(VADriverContextP ctx,
                   0);
 
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     OUT_BCS_RELOC(batch,
                   mpr_bo,
                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                   0);
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     OUT_BCS_BATCH(batch, 0);
     OUT_BCS_BATCH(batch, 0);
@@ -2509,7 +2518,7 @@ gen8_jpeg_wa_ind_obj_base_addr_state(VADriverContextP ctx,
                   I915_GEM_DOMAIN_INSTRUCTION, 0,
                   0);
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0); /* ignore for VLD mode */
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
     OUT_BCS_BATCH(batch, 0);
     OUT_BCS_BATCH(batch, 0); /* ignore for VLD mode */
     OUT_BCS_BATCH(batch, 0);
@@ -2816,6 +2825,7 @@ gen8_mfd_vp8_pic_state(VADriverContextP ctx,
                        struct decode_state *decode_state,
                        struct gen7_mfd_context *gen7_mfd_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = gen7_mfd_context->base.batch;
     VAPictureParameterBufferVP8 *pic_param = (VAPictureParameterBufferVP8 *)decode_state->pic_param->buffer;
     VAIQMatrixBufferVP8 *iq_matrix = (VAIQMatrixBufferVP8 *)decode_state->iq_matrix->buffer;
@@ -2893,7 +2903,7 @@ gen8_mfd_vp8_pic_state(VADriverContextP ctx,
                       0, I915_GEM_DOMAIN_INSTRUCTION,
                       0);
         OUT_BCS_BATCH(batch, 0);
-        OUT_BCS_BATCH(batch, 0);
+        OUT_BCS_BATCH(batch, i965->intel.mocs_state);
     } else {
         OUT_BCS_BATCH(batch, 0);
         OUT_BCS_BATCH(batch, 0);
@@ -2951,7 +2961,7 @@ gen8_mfd_vp8_pic_state(VADriverContextP ctx,
                       0, I915_GEM_DOMAIN_INSTRUCTION,
                       0);
         OUT_BCS_BATCH(batch, 0);
-        OUT_BCS_BATCH(batch, 0);
+        OUT_BCS_BATCH(batch, i965->intel.mocs_state);
     }
     else {
         OUT_BCS_BATCH(batch, 0);
diff --git a/src/gen9_mfc_hevc.c b/src/gen9_mfc_hevc.c
index 8a84c1c..4234cf7 100644
--- a/src/gen9_mfc_hevc.c
+++ b/src/gen9_mfc_hevc.c
@@ -93,7 +93,7 @@ typedef enum _gen6_brc_status {
         }                                                               \
         OUT_BCS_BATCH(batch, 0);                                        \
         if (ma)                                                         \
-            OUT_BCS_BATCH(batch, 0);                                    \
+            OUT_BCS_BATCH(batch, i965->intel.mocs_state);                                    \
     } while (0)
 
 #define OUT_BUFFER_MA_TARGET(buf_bo)       OUT_BUFFER_X(buf_bo, 1, 1)
@@ -318,6 +318,7 @@ static void
 gen9_hcpe_ind_obj_base_addr_state(VADriverContextP ctx,
                                   struct intel_encoder_context *encoder_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = encoder_context->base.batch;
     struct gen9_hcpe_context *mfc_context = encoder_context->mfc_context;
 
@@ -334,7 +335,7 @@ gen9_hcpe_ind_obj_base_addr_state(VADriverContextP ctx,
                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                   mfc_context->hcp_indirect_pak_bse_object.offset);
     OUT_BCS_BATCH(batch, 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
     OUT_BCS_RELOC(batch,
                   mfc_context->hcp_indirect_pak_bse_object.bo,
                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
diff --git a/src/gen9_mfd.c b/src/gen9_mfd.c
index 5f42514..6c4435d 100644
--- a/src/gen9_mfd.c
+++ b/src/gen9_mfd.c
@@ -53,7 +53,7 @@
         }                                                               \
         OUT_BCS_BATCH(batch, 0);                                        \
         if (ma)                                                         \
-            OUT_BCS_BATCH(batch, 0);                                    \
+            OUT_BCS_BATCH(batch, i965->intel.mocs_state);                                    \
     } while (0)
 
 #define OUT_BUFFER_MA_TARGET(buf_bo)       OUT_BUFFER(buf_bo, 1, 1)
@@ -269,6 +269,7 @@ gen9_hcpd_pipe_buf_addr_state(VADriverContextP ctx,
                               struct decode_state *decode_state,
                               struct gen9_hcpd_context *gen9_hcpd_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = gen9_hcpd_context->base.batch;
     struct object_surface *obj_surface;
     GenHevcSurface *gen9_hevc_surface;
@@ -338,6 +339,7 @@ gen9_hcpd_ind_obj_base_addr_state(VADriverContextP ctx,
                                   dri_bo *slice_data_bo,
                                   struct gen9_hcpd_context *gen9_hcpd_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = gen9_hcpd_context->base.batch;
 
     BEGIN_BCS_BATCH(batch, 14);
@@ -1396,6 +1398,7 @@ gen9_hcpd_vp9_pipe_buf_addr_state(VADriverContextP ctx,
                               struct decode_state *decode_state,
                               struct gen9_hcpd_context *gen9_hcpd_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = gen9_hcpd_context->base.batch;
     struct object_surface *obj_surface;
     int i=0;
diff --git a/src/gen9_vdenc.c b/src/gen9_vdenc.c
index 1913a67..caaa433 100644
--- a/src/gen9_vdenc.c
+++ b/src/gen9_vdenc.c
@@ -778,7 +778,7 @@ const int vdenc_hme_cost[8][52] = {
 
 #define OUT_BUFFER_3DW(batch, bo, is_target, delta, attr)  do { \
         OUT_BUFFER_2DW(batch, bo, is_target, delta);            \
-        OUT_BCS_BATCH(batch, attr);                             \
+        OUT_BCS_BATCH(batch, i965->intel.mocs_state);                             \
     } while (0)
 
 #define ALLOC_VDENC_BUFFER_RESOURCE(buffer, bfsize, des) do {   \
@@ -1312,6 +1312,7 @@ gen9_vdenc_huc_dmem_state(VADriverContextP ctx,
                           struct intel_encoder_context *encoder_context,
                           struct huc_dmem_state_parameter *params)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = encoder_context->base.batch;
 
     BEGIN_BCS_BATCH(batch, 6);
@@ -1345,6 +1346,7 @@ gen9_vdenc_huc_virtual_addr_state(VADriverContextP ctx,
                                   struct intel_encoder_context *encoder_context,
                                   struct huc_virtual_addr_parameter *params)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = encoder_context->base.batch;
     int i;
 
@@ -1369,6 +1371,7 @@ gen9_vdenc_huc_ind_obj_base_addr_state(VADriverContextP ctx,
                                        struct intel_encoder_context *encoder_context,
                                        struct huc_ind_obj_base_addr_parameter *params)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = encoder_context->base.batch;
 
     BEGIN_BCS_BATCH(batch, 11);
@@ -2344,6 +2347,7 @@ gen9_vdenc_mfx_surface_state(VADriverContextP ctx,
 static void
 gen9_vdenc_mfx_pipe_buf_addr_state(VADriverContextP ctx, struct intel_encoder_context *encoder_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct gen9_vdenc_context *vdenc_context = encoder_context->mfc_context;
     struct intel_batchbuffer *batch = encoder_context->base.batch;
     int i;
@@ -2399,6 +2403,7 @@ gen9_vdenc_mfx_pipe_buf_addr_state(VADriverContextP ctx, struct intel_encoder_co
 static void
 gen9_vdenc_mfx_ind_obj_base_addr_state(VADriverContextP ctx, struct intel_encoder_context *encoder_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct gen9_vdenc_context *vdenc_context = encoder_context->mfc_context;
     struct intel_batchbuffer *batch = encoder_context->base.batch;
 
@@ -2440,6 +2445,7 @@ gen9_vdenc_mfx_ind_obj_base_addr_state(VADriverContextP ctx, struct intel_encode
 static void
 gen9_vdenc_mfx_bsp_buf_base_addr_state(VADriverContextP ctx, struct intel_encoder_context *encoder_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct gen9_vdenc_context *vdenc_context = encoder_context->mfc_context;
     struct intel_batchbuffer *batch = encoder_context->base.batch;
 
@@ -2636,6 +2642,7 @@ gen9_vdenc_vdenc_pipe_buf_addr_state(VADriverContextP ctx,
                                      struct encode_state *encode_state,
                                      struct intel_encoder_context *encoder_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct gen9_vdenc_context *vdenc_context = encoder_context->mfc_context;
     struct intel_batchbuffer *batch = encoder_context->base.batch;
 
diff --git a/src/gen9_vp9_encoder.c b/src/gen9_vp9_encoder.c
index 05d86da..0de2cd7 100644
--- a/src/gen9_vp9_encoder.c
+++ b/src/gen9_vp9_encoder.c
@@ -4944,6 +4944,7 @@ gen9_pak_vp9_pipe_buf_addr_state(VADriverContextP ctx,
                                  struct encode_state *encode_state,
                                  struct intel_encoder_context *encoder_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = encoder_context->base.batch;
     struct gen9_encoder_context_vp9 *pak_context = encoder_context->mfc_context;
     struct gen9_vp9_state *vp9_state;
@@ -4968,49 +4969,49 @@ gen9_pak_vp9_pipe_buf_addr_state(VADriverContextP ctx,
                 obj_surface->bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 4..6 deblocking line */
     OUT_RELOC64(batch,
                 pak_context->res_deblocking_filter_line_buffer.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 7..9 deblocking tile line */
     OUT_RELOC64(batch,
                 pak_context->res_deblocking_filter_tile_line_buffer.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 10..12 deblocking tile col */
     OUT_RELOC64(batch,
                 pak_context->res_deblocking_filter_tile_col_buffer.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 13..15 metadata line */
     OUT_RELOC64(batch,
                 pak_context->res_metadata_line_buffer.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 16..18 metadata tile line */
     OUT_RELOC64(batch,
                 pak_context->res_metadata_tile_line_buffer.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 19..21 metadata tile col */
     OUT_RELOC64(batch,
                 pak_context->res_metadata_tile_col_buffer.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 22..30 SAO is not used for VP9 */
     OUT_BCS_BATCH(batch, 0);
@@ -5028,7 +5029,7 @@ gen9_pak_vp9_pipe_buf_addr_state(VADriverContextP ctx,
                 pak_context->res_mv_temporal_buffer[vp9_state->curr_mv_temporal_index].bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 34..36 Not used */
     OUT_BCS_BATCH(batch, 0);
@@ -5058,14 +5059,14 @@ gen9_pak_vp9_pipe_buf_addr_state(VADriverContextP ctx,
         OUT_BCS_BATCH(batch, 0);
     }
 
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 54..56 for source input */
     OUT_RELOC64(batch,
                 pak_context->uncompressed_picture_source.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, 0,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 57..59 StreamOut is not used */
     OUT_BCS_BATCH(batch, 0);
@@ -5098,7 +5099,7 @@ gen9_pak_vp9_pipe_buf_addr_state(VADriverContextP ctx,
         OUT_BCS_BATCH(batch, 0);
         OUT_BCS_BATCH(batch, 0);
     }
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 83..85 VP9 prob buffer */
     OUT_RELOC64(batch,
@@ -5106,7 +5107,7 @@ gen9_pak_vp9_pipe_buf_addr_state(VADriverContextP ctx,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
 
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 86..88 Segment id buffer */
     if (pak_context->res_segmentid_buffer.bo) {
@@ -5118,21 +5119,21 @@ gen9_pak_vp9_pipe_buf_addr_state(VADriverContextP ctx,
         OUT_BCS_BATCH(batch, 0);
         OUT_BCS_BATCH(batch, 0);
     }
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 89..91 HVD line rowstore buffer */
     OUT_RELOC64(batch,
                 pak_context->res_hvd_line_buffer.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 92..94 HVD tile line rowstore buffer */
     OUT_RELOC64(batch,
                 pak_context->res_hvd_tile_line_buffer.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 95..97 SAO streamout. Not used for VP9 */
     OUT_BCS_BATCH(batch, 0);
@@ -5157,6 +5158,7 @@ gen9_pak_vp9_ind_obj_base_addr_state(VADriverContextP ctx,
                                      struct encode_state *encode_state,
                                      struct intel_encoder_context *encoder_context)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct intel_batchbuffer *batch = encoder_context->base.batch;
     struct gen9_encoder_context_vp9 *pak_context = encoder_context->mfc_context;
     struct gen9_vp9_state *vp9_state;
@@ -5182,14 +5184,14 @@ gen9_pak_vp9_ind_obj_base_addr_state(VADriverContextP ctx,
                 I915_GEM_DOMAIN_INSTRUCTION, 0,   /* No write domain */
                 vp9_state->mb_data_offset);
     /* default attribute */
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 9..11, PAK-BSE */
     OUT_RELOC64(batch,
                   pak_context->indirect_pak_bse_object.bo,
                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                   pak_context->indirect_pak_bse_object.offset);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 12..13 upper bound */
     OUT_RELOC64(batch,
@@ -5202,35 +5204,35 @@ gen9_pak_vp9_ind_obj_base_addr_state(VADriverContextP ctx,
                 pak_context->res_compressed_input_buffer.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, 0,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 17..19 prob counter streamout */
     OUT_RELOC64(batch,
                 pak_context->res_prob_counter_buffer.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 20..22 prob delta streamin */
     OUT_RELOC64(batch,
                 pak_context->res_prob_delta_buffer.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 23..25 Tile record streamout */
     OUT_RELOC64(batch,
                 pak_context->res_tile_record_streamout_buffer.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     /* DW 26..28 CU record streamout */
     OUT_RELOC64(batch,
                 pak_context->res_cu_stat_streamout_buffer.bo,
                 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
                 0);
-    OUT_BCS_BATCH(batch, 0);
+    OUT_BCS_BATCH(batch, i965->intel.mocs_state);
 
     ADVANCE_BCS_BATCH(batch);
 }
diff --git a/src/i965_gpe_utils.c b/src/i965_gpe_utils.c
index 548cbf4..31976a2 100644
--- a/src/i965_gpe_utils.c
+++ b/src/i965_gpe_utils.c
@@ -1264,6 +1264,7 @@ gen9_gpe_state_base_address(VADriverContextP ctx,
                             struct i965_gpe_context *gpe_context,
                             struct intel_batchbuffer *batch)
 {
+    struct i965_driver_data *i965 = i965_driver_data(ctx);
     BEGIN_BATCH(batch, 19);
 
     OUT_BATCH(batch, CMD_STATE_BASE_ADDRESS | (19 - 2));
@@ -1273,13 +1274,14 @@ gen9_gpe_state_base_address(VADriverContextP ctx,
     OUT_BATCH(batch, 0);
 
 	/*DW4 Surface state base address */
-    OUT_RELOC64(batch, gpe_context->surface_state_binding_table.bo, I915_GEM_DOMAIN_INSTRUCTION, 0, BASE_ADDRESS_MODIFY); /* Surface state base address */
+    OUT_RELOC64(batch, gpe_context->surface_state_binding_table.bo, I915_GEM_DOMAIN_INSTRUCTION, 0, BASE_ADDRESS_MODIFY | (i965->intel.mocs_state << 4)); /* Surface state base address */
 
 	/*DW6. Dynamic state base address */
     if (gpe_context->dynamic_state.bo)
         OUT_RELOC64(batch, gpe_context->dynamic_state.bo,
                   I915_GEM_DOMAIN_RENDER | I915_GEM_DOMAIN_SAMPLER,
-                  I915_GEM_DOMAIN_RENDER, BASE_ADDRESS_MODIFY);
+                  I915_GEM_DOMAIN_RENDER,
+                  BASE_ADDRESS_MODIFY | (i965->intel.mocs_state << 4));
     else {
         OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY);
         OUT_BATCH(batch, 0);
@@ -1290,7 +1292,7 @@ gen9_gpe_state_base_address(VADriverContextP ctx,
     if (gpe_context->indirect_state.bo)
         OUT_RELOC64(batch, gpe_context->indirect_state.bo,
                   I915_GEM_DOMAIN_SAMPLER,
-                  0, BASE_ADDRESS_MODIFY);
+                  0, BASE_ADDRESS_MODIFY | (i965->intel.mocs_state << 4));
     else {
         OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY);
         OUT_BATCH(batch, 0);
@@ -1301,7 +1303,7 @@ gen9_gpe_state_base_address(VADriverContextP ctx,
     if (gpe_context->instruction_state.bo)
         OUT_RELOC64(batch, gpe_context->instruction_state.bo,
                   I915_GEM_DOMAIN_INSTRUCTION,
-                  0, BASE_ADDRESS_MODIFY);
+                  0, BASE_ADDRESS_MODIFY | (i965->intel.mocs_state << 4));
     else {
         OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY);
         OUT_BATCH(batch, 0);
diff --git a/src/intel_driver.c b/src/intel_driver.c
index 67d7de1..bcc635b 100644
--- a/src/intel_driver.c
+++ b/src/intel_driver.c
@@ -153,6 +153,13 @@ intel_driver_init(VADriverContextP ctx)
         intel->eu_total = ret_value;
     }
 
+    intel->mocs_state = 0;
+
+#define GEN9_PTE_CACHE    2
+
+    if (IS_GEN9(intel->device_info))
+        intel->mocs_state = GEN9_PTE_CACHE;
+
     intel_driver_get_revid(intel, &intel->revision);
     return true;
 }
diff --git a/src/intel_driver.h b/src/intel_driver.h
index 036e150..ad3c04b 100644
--- a/src/intel_driver.h
+++ b/src/intel_driver.h
@@ -192,6 +192,7 @@ struct intel_driver_data
     int eu_total;
 
     const struct intel_device_info *device_info;
+    unsigned int mocs_state;
 };
 
 bool intel_driver_init(VADriverContextP ctx);