Rewrite the VME shader for encoding on BDW
authorZhao Yakui <yakui.zhao@intel.com>
Wed, 27 Mar 2013 01:24:15 +0000 (09:24 +0800)
committerXiang, Haihao <haihao.xiang@intel.com>
Thu, 27 Feb 2014 02:04:12 +0000 (10:04 +0800)
Signed-off-by: Zhao Yakui <yakui.zhao@intel.com>
16 files changed:
src/gen8_mfc.c
src/gen8_vme.c
src/shaders/vme/Makefile.am
src/shaders/vme/inter_bframe_gen8.asm [new file with mode: 0644]
src/shaders/vme/inter_bframe_gen8.g8a [new file with mode: 0644]
src/shaders/vme/inter_bframe_gen8.g8b [new file with mode: 0644]
src/shaders/vme/inter_frame_gen8.asm [new file with mode: 0644]
src/shaders/vme/inter_frame_gen8.g8a [new file with mode: 0644]
src/shaders/vme/inter_frame_gen8.g8b [new file with mode: 0644]
src/shaders/vme/intra_frame_gen8.asm [new file with mode: 0644]
src/shaders/vme/intra_frame_gen8.g8a [new file with mode: 0644]
src/shaders/vme/intra_frame_gen8.g8b [new file with mode: 0644]
src/shaders/vme/mpeg2_inter_frame_gen8.g8a [new file with mode: 0644]
src/shaders/vme/mpeg2_inter_frame_gen8.g8b [new file with mode: 0644]
src/shaders/vme/vme8.inc [new file with mode: 0644]
src/shaders/vme/vme8_mpeg2.inc [new file with mode: 0644]

index fb8b0d7..1deaae3 100644 (file)
@@ -2463,7 +2463,7 @@ Bool gen8_mfc_context_init(VADriverContextP ctx, struct intel_encoder_context *e
     mfc_context->avc_qm_state = gen8_mfc_avc_qm_state;
     mfc_context->avc_fqm_state = gen8_mfc_avc_fqm_state;
     mfc_context->insert_object = gen8_mfc_avc_insert_object;
-    mfc_context->buffer_suface_setup = gen7_gpe_buffer_suface_setup;
+    mfc_context->buffer_suface_setup = gen8_gpe_buffer_suface_setup;
 
     encoder_context->mfc_context = mfc_context;
     encoder_context->mfc_context_destroy = gen8_mfc_context_destroy;
index 389ca8e..65d27c1 100644 (file)
@@ -51,8 +51,7 @@
 
 #define VME_INTRA_SHADER        0
 #define VME_INTER_SHADER        1
-#define VME_BINTER_SHADER      3
-#define VME_BATCHBUFFER         2
+#define VME_BINTER_SHADER      2
 
 #define CURBE_ALLOCATION_SIZE   37              /* in 256-bit */
 #define CURBE_TOTAL_DATA_LENGTH (4 * 32)        /* in byte, it should be less than or equal to CURBE_ALLOCATION_SIZE * 32 */
 #define VME_MSG_LENGTH         32
   
 static const uint32_t gen8_vme_intra_frame[][4] = {
-#include "shaders/vme/intra_frame_haswell.g75b"
+#include "shaders/vme/intra_frame_gen8.g8b"
 };
 
 static const uint32_t gen8_vme_inter_frame[][4] = {
-#include "shaders/vme/inter_frame_haswell.g75b"
+#include "shaders/vme/inter_frame_gen8.g8b"
 };
 
 static const uint32_t gen8_vme_inter_bframe[][4] = {
-#include "shaders/vme/inter_bframe_haswell.g75b"
-};
-
-static const uint32_t gen8_vme_batchbuffer[][4] = {
-#include "shaders/vme/batchbuffer.g75b"
+#include "shaders/vme/inter_bframe_gen8.g8b"
 };
 
 static struct i965_kernel gen8_vme_kernels[] = {
@@ -92,13 +87,6 @@ static struct i965_kernel gen8_vme_kernels[] = {
         NULL
     },
     {
-        "VME BATCHBUFFER",
-        VME_BATCHBUFFER,
-        gen8_vme_batchbuffer,
-        sizeof(gen8_vme_batchbuffer),
-        NULL
-    },
-    {
         "VME inter BFrame",
         VME_BINTER_SHADER,
         gen8_vme_inter_bframe,
@@ -108,15 +96,11 @@ static struct i965_kernel gen8_vme_kernels[] = {
 };
 
 static const uint32_t gen8_vme_mpeg2_intra_frame[][4] = {
-#include "shaders/vme/intra_frame_haswell.g75b"
+#include "shaders/vme/intra_frame_gen8.g8b"
 };
 
 static const uint32_t gen8_vme_mpeg2_inter_frame[][4] = {
-#include "shaders/vme/mpeg2_inter_haswell.g75b"
-};
-
-static const uint32_t gen8_vme_mpeg2_batchbuffer[][4] = {
-#include "shaders/vme/batchbuffer.g75b"
+#include "shaders/vme/mpeg2_inter_frame_gen8.g8b"
 };
 
 static struct i965_kernel gen8_vme_mpeg2_kernels[] = {
@@ -134,13 +118,6 @@ static struct i965_kernel gen8_vme_mpeg2_kernels[] = {
         sizeof(gen8_vme_mpeg2_inter_frame),
         NULL
     },
-    {
-        "VME BATCHBUFFER",
-        VME_BATCHBUFFER,
-        gen8_vme_mpeg2_batchbuffer,
-        sizeof(gen8_vme_mpeg2_batchbuffer),
-        NULL
-    },
 };
 
 /* only used for VME source surface state */
index 634e6d4..d84f795 100644 (file)
@@ -1,6 +1,7 @@
 VME_CORE       = batchbuffer.asm intra_frame.asm inter_frame.asm
 VME7_CORE      = batchbuffer.asm intra_frame_ivb.asm inter_frame_ivb.asm inter_bframe_ivb.asm mpeg2_inter_ivb.asm
 VME75_CORE     = batchbuffer.asm intra_frame_haswell.asm inter_frame_haswell.asm inter_bframe_haswell.asm mpeg2_inter_haswell.asm
+VME8_CORE      = intra_frame_gen8.asm  inter_frame_gen8.asm inter_bframe_gen8.asm 
 
 INTEL_G6B      = batchbuffer.g6b intra_frame.g6b inter_frame.g6b
 INTEL_G6A      = batchbuffer.g6a intra_frame.g6a inter_frame.g6a
@@ -17,16 +18,24 @@ INTEL_G75A  = batchbuffer.g75a intra_frame_haswell.g75a inter_frame_haswell.g75a
 INTEL_GEN75_INC        = batchbuffer.inc vme75.inc vme75_mpeg2.inc
 INTEL_GEN75_ASM        = $(INTEL_G75A:%.g75a=%.gen75.asm)
 
+
+INTEL_G8B      = intra_frame_gen8.g8b inter_frame_gen8.g8b mpeg2_inter_frame_gen8.g8b inter_bframe_gen8.g8b
+INTEL_G8A      = intra_frame_gen8.g8a inter_frame_gen8.g8a mpeg2_inter_frame_gen8.g8a inter_bframe_gen8.g8a
+INTEL_GEN8_INC = vme8.inc vme8_mpeg2.inc
+INTEL_GEN8_ASM = $(INTEL_G8A:%.g8a=%.gen8.asm)
+
+
 TARGETS  =
 if HAVE_GEN4ASM
 TARGETS += $(INTEL_G6B)
 TARGETS += $(INTEL_G7B)
 TARGETS += $(INTEL_G75B)
+TARGETS += $(INTEL_G8B)
 endif
 
 all-local: $(TARGETS)
 
-SUFFIXES = .g6a .g6b .g7a .g7b .gen6.asm .gen7.asm .g75a .g75b .gen75.asm
+SUFFIXES = .g6a .g6b .g7a .g7b .gen6.asm .gen7.asm .g75a .g75b .gen75.asm .g8a .g8b .gen8.asm
 
 if HAVE_GEN4ASM
 $(INTEL_GEN6_ASM): $(VME_CORE) $(INTEL_GEN6_INC)
@@ -53,9 +62,18 @@ $(INTEL_GEN75_ASM): $(VME75_CORE) $(INTEL_GEN75_INC)
        rm _vme0.$@
 .gen75.asm.g75b:
        $(AM_V_GEN)$(GEN4ASM) -g 7.5 -o $@ $<
+
+$(INTEL_GEN8_ASM): $(VME8_CORE) $(INTEL_GEN8_INC)
+.g8a.gen8.asm:
+       $(AM_V_GEN)cpp -P $< > _vme0.$@         && \
+       m4 _vme0.$@ > $@                                && \
+       rm _vme0.$@
+.gen8.asm.g8b:
+       $(AM_V_GEN)$(GEN4ASM) -g 8 -o $@ $<
+
 endif
 
-CLEANFILES = $(INTEL_GEN6_ASM) $(INTEL_GEN7_ASM) $(INTEL_GEN75_ASM)
+CLEANFILES = $(INTEL_GEN6_ASM) $(INTEL_GEN7_ASM) $(INTEL_GEN75_ASM) $(INTEL_GEN8_ASM)
 
 EXTRA_DIST = \
        $(INTEL_G6A)            \
@@ -64,11 +82,15 @@ EXTRA_DIST = \
        $(INTEL_G75B)           \
        $(INTEL_G7A)            \
        $(INTEL_G7B)            \
+       $(INTEL_G8A)            \
+       $(INTEL_G8B)            \
        $(INTEL_GEN6_INC)       \
        $(INTEL_GEN75_INC)      \
        $(INTEL_GEN7_INC)       \
+       $(INTEL_GEN8_INC)       \
        $(VME75_CORE)           \
        $(VME7_CORE)            \
+       $(VME8_CORE)            \
        $(VME_CORE)             \
        $(NULL)
 
diff --git a/src/shaders/vme/inter_bframe_gen8.asm b/src/shaders/vme/inter_bframe_gen8.asm
new file mode 100644 (file)
index 0000000..d8de588
--- /dev/null
@@ -0,0 +1,875 @@
+/*
+ * Copyright © <2010>, Intel Corporation.
+ *
+ * This program is licensed under the terms and conditions of the
+ * Eclipse Public License (EPL), version 1.0.  The full text of the EPL is at
+ * http://www.opensource.org/licenses/eclipse-1.0.php.
+ * Authors: Zhao Yakui <yakui.zhao@intel.com>
+ */
+// Modual name: Inter_bframe_haswell.asm
+//
+// Make inter predition estimation for Inter frame for B-frame
+//
+
+//
+//  Now, begin source code....
+//
+
+#define SAVE_RET       add (1) RETURN_REG<1>:ud   ip:ud        32:ud
+#define        RETURN          mov (1) ip:ud   RETURN_REG<0,1,0>:ud
+
+/*
+ * __START
+ */
+__INTER_START:
+mov  (16) tmp_reg0.0<1>:UD      0x0:UD {align1};
+mov  (16) tmp_reg2.0<1>:UD      0x0:UD {align1};
+mov  (16) tmp_reg4.0<1>:UD      0x0:UD {align1} ;
+mov  (16) tmp_reg6.0<1>:UD      0x0:UD {align1} ;
+
+shl  (2) read0_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
+add  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D -8:W {align1};     /* X offset */
+add  (1) read0_header.4<1>:D    read0_header.4<0,1,0>:D -1:W {align1};     /* Y offset */ 
+mov  (1) read0_header.8<1>:UD   BLOCK_32X1 {align1};
+mov  (1) read0_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
+
+shl  (2) read1_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
+add  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D -4:W {align1};     /* X offset */
+mov  (1) read1_header.8<1>:UD   BLOCK_4X16 {align1};
+mov  (1) read1_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
+        
+shl  (2) vme_m0.8<1>:UW         orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
+mov  (1) vme_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+
+mul  (1) obw_m0.8<1>:UD         w_in_mb_uw<0,1,0>:UW orig_y_ub<0,1,0>:UB {align1};
+add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD orig_x_ub<0,1,0>:UB {align1};
+mul  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 24:UD {align1};
+mov  (1) obw_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+        
+/*
+ * Media Read Message -- fetch Luma neighbor edge pixels 
+ */
+/* ROW */
+mov  (8) msg_reg0.0<1>:UD       read0_header.0<8,8,1>:UD {align1};        
+send (8) msg_ind INEP_ROW<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 1 {align1};
+
+/* COL */
+mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};                
+send (8) msg_ind INEP_COL0<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 2 {align1};
+        
+/*
+ * Media Read Message -- fetch Chroma neighbor edge pixels 
+ */
+/* ROW */
+shl  (2) read0_header.0<1>:D    orig_xy_ub<2,2,1>:UB 3:UW {align1};    /* x * 16 , y * 8 */
+mul  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D  2:W {align1};
+add  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D -8:W {align1};     /* X offset */
+add  (1) read0_header.4<1>:D    read0_header.4<0,1,0>:D -1:W {align1};     /* Y offset */ 
+mov  (8) msg_reg0.0<1>:UD       read0_header.0<8,8,1>:UD {align1};        
+send (8) msg_ind CHROMA_ROW<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen 1 {align1};
+
+/* COL */
+shl  (2) read1_header.0<1>:D    orig_xy_ub<2,2,1>:UB 3:UW {align1};    /* x * 16, y * 8 */
+mul  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D  2:W {align1};
+add  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D -4:W {align1};     /* X offset */
+mov  (1) read1_header.8<1>:UD   BLOCK_8X4 {align1};
+mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};                
+send (8) msg_ind CHROMA_COL<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen 1 {align1};
+
+mov  (8) vme_m1.0<1>:ud                0:ud            {align1};
+mov  (8) mb_mvp_ref.0<1>:ud    0:ud            {align1};
+mov  (8) mb_ref_win.0<1>:ud    0:ud            {align1};
+and.z.f0.0 (1)         null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
+(f0.0) jmpi (1) __mb_hwdep_end;
+
+/* read back the data for MB A */
+/* the layout of MB result is: rx.0(Available). rx.4(MVa), rX.8(MVb), rX.16(Pred_L0 flag),
+*  rX.18 (Pred_L1 flag), rX.20(Forward reference ID), rX.22(Backwared reference ID)
+*/
+mov  (8) mba_result.0<1>:ud    0x0:ud          {align1};
+mov  (8) mbb_result.0<1>:ud    0x0:ud          {align1};
+mov  (8) mbc_result.0<1>:ud    0x0:ud          {align1};
+mba_start:
+mov  (8) mb_msg0.0<1>:ud       0:ud            {align1};
+and.z.f0.0 (1)         null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
+/* MB A doesn't exist. Zero MV. mba_flag is zero and ref ID = -1 */
+(f0.0)  mov  (2)       mba_result.20<1>:w      -1:w    {align1};
+(f0.0)  jmpi (1)       mbb_start;
+mov  (1) mba_result.0<1>:d     MB_AVAIL                {align1};       
+mov  (2) tmp_reg0.0<1>:UW      orig_xy_ub<2,2,1>:UB    {align1};
+add  (1) tmp_reg0.0<1>:w       tmp_reg0.0<0,1,0>:w     -1:w    {align1};
+mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
+add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
+mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
+mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+
+/* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_wb.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_4,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 2
+        {align1};
+
+/* TODO: RefID is required after multi-references are added */
+cmp.l.f0.0 (1)         null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
+(f0.0)   mov (2)       mba_result.20<1>:w                      -1:w    {align1};
+(f0.0)   jmpi  (1)     mbb_start;
+
+add   (1) mb_msg0.8<1>:UD      mb_msg0.8<0,1,0>:ud     3:ud {align1};
+/* Read MV for MB A */
+/* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_mv0.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_8,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 4
+        {align1};
+/* TODO: RefID is required after multi-references are added */
+/* MV */
+mov    (2)     mba_result.20<1>:w              -1:w            {align1};
+mov    (1)     INPUT_ARG0.0<1>:ud      mb_inter_wb.4<0,1,0>:ud {align1};
+mov    (1)     INPUT_ARG0.4<1>:ud      mb_inter_wb.0<0,1,0>:ud {align1};
+mov    (1)     INPUT_ARG0.8<1>:ud      INTER_BLOCK1:ud {align1};
+SAVE_RET       {align1};
+jmpi   (1)     mb_pred_func;
+mov    (1)     mb_pred_mode.0<1>:uw    RET_ARG<0,1,0>:uw       {align1};
+cmp.e.f0.0 (1) null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L0 {align1};
+(f0.0) mov   (1)       mba_result.16<1>:uw             MB_PRED_FLAG            {align1};
+(f0.0) mov   (1)       mba_result.20<1>:w              0:w             {align1};
+(f0.0)  mov   (1)      mba_result.4<1>:ud              mb_mv1.8<0,1,0>:ud      {align1};
+(f0.0)  jmpi   (1) mbb_start;
+cmp.e.f0.0 (1) null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L1 {align1};
+(f0.0) mov   (1)       mba_result.18<1>:uw             MB_PRED_FLAG            {align1};
+(f0.0) mov   (1)       mba_result.22<1>:w              0:w             {align1};
+(f0.0)  mov   (1)      mba_result.8<1>:ud              mb_mv1.12<0,1,0>:ud     {align1};
+(f0.0)  jmpi   (1) mbb_start;
+mov   (2)      mba_result.4<1>:ud              mb_mv1.8<2,2,1>:ud      {align1};
+mov   (2)      mba_result.16<1>:uw             MB_PRED_FLAG            {align1};
+mov   (2)      mba_result.20<1>:w              0:w             {align1};
+
+mbb_start:
+mov  (8) mb_msg0.0<1>:ud       0:ud            {align1};
+and.z.f0.0 (1)         null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
+/* MB B doesn't exist. Zero MV. mba_flag is zero */
+/* If MB B doesn't exist, neither MB C nor D exists */
+(f0.0)  mov  (2)       mbb_result.20<1>:w      -1:w            {align1};
+(f0.0)  mov  (2)       mbc_result.20<1>:w      -1:w            {align1};
+(f0.0)  jmpi (1)       mb_mvp_start;
+mov  (1) mbb_result.0<1>:d     MB_AVAIL                {align1};       
+mov  (2) tmp_reg0.0<1>:UW      orig_xy_ub<2,2,1>:UB    {align1};
+add  (1) tmp_reg0.2<1>:w       tmp_reg0.2<0,1,0>:w     -1:w    {align1};
+mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
+add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
+mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
+mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+
+/* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_wb.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_4,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 2
+        {align1};
+
+/* TODO: RefID is required after multi-references are added */
+cmp.l.f0.0 (1)         null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
+(f0.0)   mov (2)       mbb_result.20<1>:w                      -1:w    {align1};
+(f0.0)   jmpi  (1)     mbc_start;
+add   (1) mb_msg0.8<1>:UD      mb_msg0.8<0,1,0>:ud     3:ud {align1};
+/* Read MV for MB B */
+/* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_mv0.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_8,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 4
+        {align1};
+/* TODO: RefID is required after multi-references are added */
+mov       (2)          mbb_result.20<1>:w              -1:w    {align1};
+mov    (1)     INPUT_ARG0.0<1>:ud      mb_inter_wb.4<0,1,0>:ud {align1};
+mov    (1)     INPUT_ARG0.4<1>:ud      mb_inter_wb.0<0,1,0>:ud {align1};
+mov    (1)     INPUT_ARG0.8<1>:ud      INTER_BLOCK2:ud {align1};
+SAVE_RET       {align1};
+jmpi   (1)     mb_pred_func;
+mov    (1)     mb_pred_mode.0<1>:uw    RET_ARG<0,1,0>:uw       {align1};
+cmp.e.f0.0 (1) null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L0 {align1};
+(f0.0) mov   (1)       mbb_result.16<1>:uw             MB_PRED_FLAG            {align1};
+(f0.0) mov   (1)       mbb_result.20<1>:w              0:w             {align1};
+(f0.0)  mov   (1)      mbb_result.4<1>:ud              mb_mv2.16<0,1,0>:ud     {align1};
+(f0.0)  jmpi   (1) mbc_start;
+cmp.e.f0.0 (1) null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L1 {align1};
+(f0.0) mov   (1)       mbb_result.18<1>:uw             MB_PRED_FLAG            {align1};
+(f0.0) mov   (1)       mbb_result.22<1>:w              0:w             {align1};
+(f0.0)  mov   (1)      mbb_result.8<1>:ud              mb_mv2.20<0,1,0>:ud     {align1};
+(f0.0)  jmpi   (1) mbc_start;
+mov   (2)      mbb_result.16<1>:uw             MB_PRED_FLAG            {align1};
+mov   (2)      mbb_result.20<1>:w              0:w             {align1};
+mov       (2)          mbb_result.4<1>:ud              mb_mv2.16<2,2,1>:ud     {align1};
+
+mbc_start:
+mov  (8) mb_msg0.0<1>:ud       0:ud            {align1};
+and.z.f0.0 (1)         null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_C:uw   {align1};
+/* MB C doesn't exist. Zero MV. mba_flag is zero */
+/* Based on h264 spec the MB D will be replaced if MB C doesn't exist */
+(f0.0)  jmpi (1)       mbd_start;
+mov  (1) mbc_result.0<1>:d     MB_AVAIL                {align1};       
+mov  (2) tmp_reg0.0<1>:UW      orig_xy_ub<2,2,1>:UB    {align1};
+add  (1) tmp_reg0.2<1>:w       tmp_reg0.2<0,1,0>:w     -1:w    {align1};
+add  (1) tmp_reg0.0<1>:w       tmp_reg0.0<0,1,0>:w     1:w     {align1};
+mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
+add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
+mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
+mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+
+/* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_wb.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_4,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 2
+        {align1};
+
+/* TODO: RefID is required after multi-references are added */
+cmp.l.f0.0 (1)         null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
+(f0.0)   mov (2)       mbc_result.20<1>:w                      -1:w    {align1};
+(f0.0)   jmpi  (1)     mb_mvp_start;
+add   (1) mb_msg0.8<1>:UD      mb_msg0.8<0,1,0>:ud     3:ud {align1};
+/* Read MV for MB C */
+/* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_mv0.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_8,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 4
+        {align1};
+/* TODO: RefID is required after multi-references are added */
+/* Forward MV */
+mov       (2)          mbc_result.20<1>:w              -1:w    {align1};
+mov    (1)     INPUT_ARG0.0<1>:ud      mb_inter_wb.4<0,1,0>:ud {align1};
+mov    (1)     INPUT_ARG0.4<1>:ud      mb_inter_wb.0<0,1,0>:ud {align1};
+mov    (1)     INPUT_ARG0.8<1>:ud      INTER_BLOCK2:ud {align1};
+SAVE_RET       {align1};
+jmpi   (1)     mb_pred_func;
+mov    (1)     mb_pred_mode.0<1>:uw    RET_ARG<0,1,0>:uw       {align1};
+cmp.e.f0.0 (1) null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L0 {align1};
+(f0.0) mov    (1) mbc_result.16<1>:uw          MB_PRED_FLAG            {align1};
+(f0.0) mov    (1) mbc_result.20<1>:w           0:w             {align1};
+(f0.0)  mov    (1)  mbc_result.4<1>:ud         mb_mv2.16<0,1,0>:ud     {align1};
+(f0.0)  jmpi   (1) mb_mvp_start;
+cmp.e.f0.0 (1) null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L1 {align1};
+(f0.0) mov   (1)       mbc_result.18<1>:uw             MB_PRED_FLAG            {align1};
+(f0.0) mov    (1) mbc_result.22<1>:w           0:w             {align1};
+(f0.0)  mov    (1)  mbc_result.8<1>:ud         mb_mv2.20<0,1,0>:ud     {align1};
+(f0.0)  jmpi  (1) mb_mvp_start;
+mov  (2)  mbc_result.16<1>:uw          MB_PRED_FLAG            {align1};
+mov  (2)  mbc_result.20<1>:w           0:w     {align1};
+mov       (2)          mbc_result.4<1>:ud              mb_mv2.16<2,2,1>:ud     {align1};
+
+jmpi   (1)    mb_mvp_start;
+mbd_start:
+mov  (8) mb_msg0.0<1>:ud       0:ud            {align1};
+and.z.f0.0 (1)         null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_D:uw   {align1};
+(f0.0)  mov  (2)       mbc_result.20<1>:w              -1:w    {align1};
+(f0.0)  jmpi (1)       mb_mvp_start;
+mov  (1) mbc_result.0<1>:d     MB_AVAIL                {align1};       
+mov  (2) tmp_reg0.0<1>:UW      orig_xy_ub<2,2,1>:UB    {align1};
+add  (2) tmp_reg0.0<1>:w       tmp_reg0.0<2,2,1>:w     -1:w    {align1};
+mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
+add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
+mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
+mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+
+/* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_wb.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_4,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 2
+        {align1};
+
+cmp.l.f0.0 (1)         null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
+(f0.0)   mov (2)       mbc_result.20<1>:w                      -1:w    {align1};
+(f0.0)   jmpi  (1)     mb_mvp_start;
+
+add   (1) mb_msg0.8<1>:UD      mb_msg0.8<0,1,0>:ud     3:ud {align1};
+/* Read MV for MB D */
+/* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_mv0.0<1>:ub
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_8,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 4
+        {align1};
+
+/* TODO: RefID is required after multi-references are added */
+
+/* Forward MV */
+mov       (2)          mbc_result.20<1>:w              -1:w    {align1};
+mov    (1)     INPUT_ARG0.0<1>:ud      mb_inter_wb.4<0,1,0>:ud {align1};
+mov    (1)     INPUT_ARG0.4<1>:ud      mb_inter_wb.0<0,1,0>:ud {align1};
+mov    (1)     INPUT_ARG0.8<1>:ud      INTER_BLOCK3:ud {align1};
+SAVE_RET       {align1};
+jmpi   (1)     mb_pred_func;
+mov    (1)     mb_pred_mode.0<1>:uw    RET_ARG<0,1,0>:uw       {align1};
+cmp.e.f0.0 (1) null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L0 {align1};
+(f0.0) mov    (1) mbc_result.16<1>:uw          MB_PRED_FLAG            {align1};
+(f0.0) mov    (1) mbc_result.20<1>:w           0:w             {align1};
+(f0.0)  mov    (1) mbc_result.4<1>:ud          mb_mv3.24<0,1,0>:ud     {align1};
+(f0.0)  jmpi   (1) mb_mvp_start;
+cmp.e.f0.0 (1) null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L1 {align1};
+(f0.0) mov   (1)       mbc_result.18<1>:uw             MB_PRED_FLAG            {align1};
+(f0.0) mov    (1) mbc_result.22<1>:w           0:w             {align1};
+(f0.0)  mov    (1) mbc_result.8<1>:ud          mb_mv3.28<0,1,0>:ud     {align1};
+(f0.0)  jmpi  (1) mb_mvp_start;
+mov  (2)  mbc_result.16<1>:uw          MB_PRED_FLAG            {align1};
+mov  (2)  mbc_result.20<1>:w           0:w     {align1};
+mov       (2)          mbc_result.4<1>:ud              mb_mv3.24<2,2,1>:ud     {align1};
+       
+mb_mvp_start:
+/*TODO: Add the skip prediction */
+/* Check whether both MB B and C are inavailable */
+add    (1)     tmp_reg0.0<1>:d         mbb_result.0<0,1,0>:d   mbc_result.0<0,1,0>:d   {align1};
+cmp.z.f0.0 (1) null:d                  tmp_reg0.0<0,1,0>:d     0:d     {align1};
+(-f0.0)        jmpi (1)        mb_median_start;
+cmp.nz.f0.0 (1)        null:d  mba_result.0<0,1,0>:d           1:d             {align1};
+(f0.0) mov     (2)     mbb_result.4<1>:ud              mba_result.4<2,2,1>:ud  {align1};       
+(f0.0) mov     (2)     mbc_result.4<1>:ud              mba_result.4<2,2,1>:ud  {align1};       
+(f0.0) mov     (2)     mbb_result.20<1>:uw             mba_result.20<2,2,1>:uw {align1};       
+(f0.0) mov     (2)     mbc_result.20<1>:uw             mba_result.20<2,2,1>:uw {align1};       
+(f0.0)  mov     (2)    mb_mvp_ref.0<1>:ud              mba_result.4<2,2,1>:ud  {align1};
+(-f0.0) mov    (2)     mb_mvp_ref.0<1>:ud              0:ud                    {align1};
+jmpi   (1)     __mb_hwdep_end;
+       
+mb_median_start:
+/* forward_MVP */
+/* check whether only one neighbour MB has the same ref ID with the current MB */
+mov (8)        tmp_reg0.0<1>:ud                0:ud            {align1};
+cmp.z.f0.0     (1)     null:d  mba_result.20<0,1,0>:w  0:w     {align1};
+(f0.0) add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
+(f0.0) mov     (1)     tmp_reg0.4<1>:ud        mba_result.4<0,1,0>:ud  {align1};
+cmp.z.f0.0     (1)     null:d  mbb_result.20<0,1,0>:w  0:w     {align1};
+(f0.0) add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
+(f0.0) mov     (1)     tmp_reg0.4<1>:ud        mbb_result.4<0,1,0>:ud  {align1};
+cmp.z.f0.0     (1)     null:d  mbc_result.20<0,1,0>:w  0:w     {align1};
+(f0.0) add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
+(f0.0) mov     (1)     tmp_reg0.4<1>:ud        mbc_result.4<0,1,0>:ud  {align1};
+cmp.e.f0.0     (1)     null:d  tmp_reg0.0<0,1,0>:w      1:w    {align1};
+(f0.0) mov     (1)     mb_mvp_ref.0<1>:ud      tmp_reg0.4<0,1,0>:ud    {align1};
+(f0.0) jmpi (1)  mvp_backward;
+
+mov    (1)     INPUT_ARG0.0<1>:w       mba_result.4<0,1,0>:w   {align1};
+mov    (1)     INPUT_ARG0.4<1>:w       mbb_result.4<0,1,0>:w   {align1};
+mov    (1)     INPUT_ARG0.8<1>:w       mbc_result.4<0,1,0>:w   {align1};
+SAVE_RET       {align1};
+ jmpi  (1)     word_imedian;
+mov    (1)     mb_mvp_ref.0<1>:w               RET_ARG<0,1,0>:w        {align1};
+mov    (1)     INPUT_ARG0.0<1>:w       mba_result.6<0,1,0>:w   {align1};
+mov    (1)     INPUT_ARG0.4<1>:w       mbb_result.6<0,1,0>:w   {align1};
+mov    (1)     INPUT_ARG0.8<1>:w       mbc_result.6<0,1,0>:w   {align1};
+SAVE_RET       {align1};
+jmpi   (1)     word_imedian; 
+mov    (1)     mb_mvp_ref.2<1>:w               RET_ARG<0,1,0>:w        {align1};
+
+
+mvp_backward:
+/* check whether only one neighbour MB has the same ref ID with the current MB */
+mov (8)        tmp_reg0.0<1>:ud                0:ud            {align1};
+cmp.z.f0.0     (1)     null:d  mba_result.22<0,1,0>:w  0:w     {align1};
+(f0.0) add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
+(f0.0) mov     (1)     tmp_reg0.4<1>:ud        mba_result.8<0,1,0>:ud  {align1};
+cmp.z.f0.0     (1)     null:d  mbb_result.22<0,1,0>:w  0:w     {align1};
+(f0.0) add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
+(f0.0) mov     (1)     tmp_reg0.4<1>:ud        mbb_result.8<0,1,0>:ud  {align1};
+cmp.z.f0.0     (1)     null:d  mbc_result.22<0,1,0>:w  0:w     {align1};
+(f0.0) add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
+(f0.0) mov     (1)     tmp_reg0.4<1>:ud        mbc_result.8<0,1,0>:ud  {align1};
+cmp.e.f0.0     (1)     null:d  tmp_reg0.0<0,1,0>:w      1:w    {align1};
+(f0.0) mov     (1)     mb_mvp_ref.4<1>:ud      tmp_reg0.4<0,1,0>:ud    {align1};
+(f0.0) jmpi (1) __mb_hwdep_end;
+
+mov    (1)     INPUT_ARG0.0<1>:w       mba_result.8<0,1,0>:w   {align1};
+mov    (1)     INPUT_ARG0.4<1>:w       mbb_result.8<0,1,0>:w   {align1};
+mov    (1)     INPUT_ARG0.8<1>:w       mbc_result.8<0,1,0>:w   {align1};
+SAVE_RET       {align1};
+ jmpi  (1)     word_imedian;
+mov    (1)     mb_mvp_ref.4<1>:w               RET_ARG<0,1,0>:w        {align1};
+mov    (1)     INPUT_ARG0.0<1>:w       mba_result.10<0,1,0>:w  {align1};
+mov    (1)     INPUT_ARG0.4<1>:w       mbb_result.10<0,1,0>:w  {align1};
+mov    (1)     INPUT_ARG0.8<1>:w       mbc_result.10<0,1,0>:w  {align1};
+SAVE_RET       {align1};
+jmpi   (1)     word_imedian; 
+mov    (1)     mb_mvp_ref.6<1>:w               RET_ARG<0,1,0>:w        {align1};
+
+__mb_hwdep_end:
+asr    (4)     mb_ref_win.0<1>:w       mb_mvp_ref.0<4,4,1>:w   2:w     {align1};
+add    (4)     mb_ref_win.8<1>:w       mb_ref_win.0<4,4,1>:w   3:w     {align1};
+and    (4)     mb_ref_win.16<1>:uw     mb_ref_win.8<4,4,1>:uw  0xFFFC:uw {align1};
+/* m2, get the MV/Mb cost passed from constant buffer when
+spawning thread by MEDIA_OBJECT */       
+mov (8) vme_m2<1>:UD            r1.0<8,8,1>:UD {align1};
+
+mov (8) vme_msg_2<1>:UD                vme_m2.0<8,8,1>:UD {align1};
+/* m3 cost center */
+mov (8) vme_m3.0<1>:ud         0x0:ud  {align1};
+mov (8) vme_msg_3<1>:UD                vme_m3.0<8,8,1>:UD {align1};            
+
+/* m4. skip center */
+mov (8) vme_msg_4<1>:ud                0x0:ud  {align1};
+
+/* m5 */
+mov  (1) INEP_ROW.0<1>:UD       0x0:UD {align1};
+and  (1) INEP_ROW.4<1>:UD       INEP_ROW.4<0,1,0>:UD            0xFF000000:UD {align1};
+mov  (8) vme_msg_5<1>:UD         INEP_ROW.0<8,8,1>:UD {align1};
+/* Use the Luma mode */
+mov  (1) tmp_reg0.0<1>:UW      LUMA_INTRA_MODE:UW {align1};
+mov  (1) vme_msg_5.5<1>:UB     tmp_reg0.0<0,1,0>:UB {align1};
+
+/* m6 */        
+mov  (8) vme_msg_6<1>:UD         0x0:UD {align1};
+mov (16) vme_msg_6.0<1>:UB       INEP_COL0.3<32,8,4>:UB {align1};
+mov  (1) vme_msg_6.16<1>:UD      INTRA_PREDICTORE_MODE {align1};
+
+/* the penalty for Intra mode */
+mov  (1) vme_msg_6.28<1>:UD    0x010101:UD {align1};
+mov  (1) vme_msg_6.20<1>:UW      CHROMA_ROW.6<0,1,0>:UW {align1};
+
+
+/* m7 */
+
+mov  (4) vme_msg_7.16<1>:UD      CHROMA_ROW.8<4,4,1>:UD {align1};
+mov  (8) vme_msg_7.0<1>:UW       CHROMA_COL.2<16,8,2>:UW {align1};
+
+/*
+ * SIC VME message
+ */
+/* m1 */
+mov  (1) intra_flag<1>:UW       0x0:UW {align1}                     ;
+and.z.f0.0 (1) null<1>:UW transform_8x8_ub<0,1,0>:UB 1:UW {align1};
+(f0.0) mov  (1) intra_part_mask_ub<1>:UB  LUMA_INTRA_8x8_DISABLE {align1};
+
+/* assign MB intra struct from the thread payload*/
+mov (1) mb_intra_struct_ub<1>:UB input_mb_intra_ub<0,1,0>:UB {align1}; 
+
+/* Disable DC HAAR component when calculating HARR SATD block */
+mov  (1) tmp_reg0.0<1>:UW      DC_HARR_DISABLE:UW              {align1};
+mov  (1) vme_m1.30<1>:UB       tmp_reg0.0<0,1,0>:UB  {align1};
+mov  (8) vme_msg_1<1>:UD        vme_m1.0<8,8,1>:UD {align1};
+
+/* m0 */        
+mov  (1) vme_m0.12<1>:UD        INTRA_SAD_HAAR:UD {align1};    /* 16x16 Source, Intra_harr */
+mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
+
+/* after verification it will be passed by using payload */
+send (8)
+        vme_msg_ind
+        vme_wb<1>:UD
+        null
+        cre(
+                BIND_IDX_VME,
+                VME_SIC_MESSAGE_TYPE
+        )
+        mlen sic_vme_msg_length
+        rlen vme_wb_length
+        {align1};
+/*
+ * Oword Block Write message
+ */
+mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
+        
+mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
+mov  (1) msg_reg1.4<1>:UD       vme_wb.16<0,1,0>:UD     {align1};
+mov  (1) msg_reg1.8<1>:UD       vme_wb.20<0,1,0>:UD     {align1};
+mov  (1) msg_reg1.12<1>:UD      vme_wb.24<0,1,0>:UD     {align1};
+
+/* Distortion, Intra (17-16), */
+mov  (1) msg_reg1.16<1>:UW      vme_wb.12<0,1,0>:UW     {align1};
+
+mov  (1) msg_reg1.20<1>:UD      vme_wb.8<0,1,0>:UD     {align1};
+/* VME clock counts */
+mov  (1) msg_reg1.24<1>:UD      vme_wb.28<0,1,0>:UD     {align1};
+
+mov  (1) msg_reg1.28<1>:UD      obw_m0.8<0,1,0>:UD     {align1};
+
+/* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
+send (16)
+        msg_ind
+        obw_wb
+        null
+        data_port(
+                OBW_CACHE_TYPE,
+                OBW_MESSAGE_TYPE,
+                OBW_CONTROL_2,
+                OBW_BIND_IDX,
+                OBW_WRITE_COMMIT_CATEGORY,
+                OBW_HEADER_PRESENT
+        )
+        mlen 2
+        rlen obw_wb_length
+        {align1};
+
+/* IME search */
+mov  (1) vme_m0.12<1>:UD        SEARCH_CTRL_DUAL_REFERENCE + INTER_PART_MASK + INTER_SAD_HAAR:UD {align1};    /* 16x16 Source, harr */
+mov  (1) vme_m0.22<1>:UW        DREF_REGION_SIZE {align1};        
+/* Dual Reference Width&Height,32x32 */
+
+mov  (1) vme_m0.0<1>:UD                vme_m0.8<0,1,0>:UD      {align1};
+
+/* Reference = (x-8,y-8)-(x+8,y+8) */
+add  (1) vme_m0.0<1>:W          vme_m0.0<0,1,0>:W -8:W {align1};               
+add  (1) vme_m0.2<1>:W          vme_m0.2<0,1,0>:W -8:W {align1};
+
+mov  (1) vme_m0.0<1>:W         -8:W                    {align1};
+mov  (1) vme_m0.2<1>:W         -8:W                    {align1};
+
+mov  (1) vme_m0.4<1>:UD                vme_m0.0<0,1,0>:UD      {align1};
+and.z.f0.0 (1)         null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
+(f0.0) add     (1)     vme_m0.0<1>:w   vme_m0.0<0,1,0>:w       4:w     {align1};
+(f0.0) add     (1)     vme_m0.4<1>:w   vme_m0.4<0,1,0>:w       4:w     {align1};
+and.z.f0.0 (1)         null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
+(f0.0) add     (1)     vme_m0.2<1>:w   vme_m0.2<0,1,0>:w       4:w     {align1};
+(f0.0) add     (1)     vme_m0.6<1>:w   vme_m0.6<0,1,0>:w       4:w     {align1};
+
+add  (2) vme_m0.0<1>:w         vme_m0.0<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
+add  (2) vme_m0.4<1>:w         vme_m0.4<2,2,1>:w       mb_ref_win.20<2,2,1>:w  {align1};
+mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
+
+mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
+/* the Max MV number is passed by constant buffer */
+mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};          
+mov  (1) vme_m1.8<1>:UD         DSTART_CENTER + DSEARCH_PATH_LEN:UD {align1};
+mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
+
+mov (8) vme_msg_2<1>:UD                vme_m2.0<8,8,1>:UD {align1};
+
+
+/* Setup the Cost center */
+/* currently four 8x8 share the same cost center */
+mov  (4) vme_m3.0<2>:ud                mb_mvp_ref.0<0,1,0>:ud  {align1};
+mov  (4) vme_m3.4<2>:ud                mb_mvp_ref.4<0,1,0>:ud  {align1};
+
+/* M4/M5 search path */
+
+mov  (1) vme_msg_4.0<1>:UD     0x10010101:UD {align1};
+mov  (1) vme_msg_4.4<1>:UD     0x100F0F0F:UD {align1};
+mov  (1) vme_msg_4.8<1>:UD     0x10010101:UD {align1};
+mov  (1) vme_msg_4.12<1>:UD    0x000F0F0F:UD {align1};
+
+mov  (4) vme_msg_4.16<1>:UD    0x0:UD {align1};
+mov  (8) vme_msg_5.16<1>:UD    0x0:UD {align1};
+
+send (8)
+        vme_msg_ind
+        vme_wb<1>:UD
+        null
+        vme(
+                BIND_IDX_VME,
+                0,
+                0,
+                VME_IME_MESSAGE_TYPE
+        )
+        mlen ime_vme_msg_length
+        rlen vme_wb_length {align1};
+
+/* Set Macroblock-shape/mode for FBR */
+
+mov  (1) vme_m2.20<1>:UD       0x0:UD {align1};
+mov  (1) vme_m2.21<1>:UB       vme_wb.25<0,1,0>:UB     {align1};
+mov  (1) vme_m2.22<1>:UB       vme_wb.26<0,1,0>:UB     {align1};
+
+and  (1) tmp_reg0.0<1>:UW      vme_wb.0<0,1,0>:UW      0x03:UW {align1};
+mov  (1) vme_m2.20<1>:UB       tmp_reg0.0<0,1,0>:UB    {align1};
+
+/* Send FBR message into CRE */
+
+mov  (8) vme_msg_4.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
+mov  (8) vme_msg_5.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
+mov  (8) vme_msg_6.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
+mov  (8) vme_msg_7.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
+
+ /* 16x16 Source, 1/4 pixel, harr, BME ENABLE */
+mov  (1) vme_m0.12<1>:UD       INTER_SAD_HAAR + SUB_PEL_MODE_QUARTER + FBR_BME_ENABLE:UD {align1};
+   
+mov  (8) vme_msg_0.0<1>:UD     vme_m0.0<8,8,1>:UD  {align1};
+
+mov  (1) tmp_reg0.0<1>:uw      BI_WEIGHT       {align1};
+mov  (1) vme_m1.6<1>:UB                tmp_reg0.0<0,1,0>:ub    {align1};
+mov  (8) vme_msg_1.0<1>:UD     vme_m1.0<8,8,1>:UD  {align1};
+
+mov  (8) vme_msg_2.0<1>:UD             vme_m2.0<8,8,1>:UD      {align1};
+mov  (8) vme_msg_3.0<1>:UD             vme_m3.0<8,8,1>:UD      {align1};
+
+/* after verification it will be passed by using payload */
+send (8)
+        vme_msg_ind
+        vme_wb<1>:UD
+        null
+        cre(
+                BIND_IDX_VME,
+                VME_FBR_MESSAGE_TYPE
+        )
+        mlen fbr_vme_msg_length
+        rlen vme_wb_length
+        {align1};
+
+add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x02:UD {align1};
+mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
+/* write FME info */
+mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
+
+mov  (1) msg_reg1.4<1>:UD       vme_wb.24<0,1,0>:UD     {align1};
+/* Inter distortion of FME */
+mov  (1) msg_reg1.8<1>:UD       vme_wb.8<0,1,0>:UD     {align1};
+
+mov  (1) msg_reg1.12<1>:UD     vme_m2.20<0,1,0>:UD {align1};
+
+/* bind index 3, write  oword (16bytes), msg type: 8(OWord Block Write) */
+send (16)
+        msg_ind
+        obw_wb
+        null
+        data_port(
+                OBW_CACHE_TYPE,
+                OBW_MESSAGE_TYPE,
+                OBW_CONTROL_0,
+                OBW_BIND_IDX,
+                OBW_WRITE_COMMIT_CATEGORY,
+                OBW_HEADER_PRESENT
+        )
+        mlen 2
+        rlen obw_wb_length
+        {align1};
+
+/* Write FME/BME MV */
+add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x01:UD {align1};
+mov  (8) msg_reg0.0<1>:UD       obw_m0.0<8,8,1>:UD {align1};
+
+
+mov  (8) msg_reg1.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
+mov  (8) msg_reg2.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
+mov  (8) msg_reg3.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
+mov  (8) msg_reg4.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
+/* bind index 3, write  8 oword (128 bytes), msg type: 8(OWord Block Write) */
+send (16)
+        msg_ind
+        obw_wb
+        null
+        data_port(
+                OBW_CACHE_TYPE,
+                OBW_MESSAGE_TYPE,
+                OBW_CONTROL_8,
+                OBW_BIND_IDX,
+                OBW_WRITE_COMMIT_CATEGORY,
+                OBW_HEADER_PRESENT
+        )
+        mlen 5
+        rlen obw_wb_length
+        {align1};
+
+/* Write FME/BME RefID */
+add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x08:UD {align1};
+mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
+
+mov  (8) msg_reg1.0<1>:UD      vme_wb6.0<8,8,1>:UD {align1};
+
+/* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
+send (16)
+        msg_ind
+        obw_wb
+        null
+        data_port(
+                OBW_CACHE_TYPE,
+                OBW_MESSAGE_TYPE,
+                OBW_CONTROL_2,
+                OBW_BIND_IDX,
+                OBW_WRITE_COMMIT_CATEGORY,
+                OBW_HEADER_PRESENT
+        )
+        mlen 2
+        rlen obw_wb_length
+        {align1};
+
+
+/* Issue message fence so that the previous write message is committed */
+send (16)
+        mb_ind
+        obw_wb
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_FENCE,
+                OBR_MF_COMMIT,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 1
+        {align1};
+
+__EXIT: 
+/*
+ * kill thread
+ */        
+mov  (8) ts_msg_reg0<1>:UD         r0<8,8,1>:UD {align1};
+send (16) ts_msg_ind acc0<1>UW null thread_spawner(0, 0, 1) mlen 1 rlen 0 {align1 EOT};
+
+
+       nop             ;
+       nop             ;
+/* Compare three word data to get the min value */
+word_imin:
+       cmp.le.f0.0 (1)         null:w          INPUT_ARG0.0<0,1,0>:w   INPUT_ARG0.4<0,1,0>:w {align1};
+       (f0.0) mov  (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.0<0,1,0>:w                    {align1};
+       (-f0.0) mov (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.4<0,1,0>:w                    {align1};
+       cmp.le.f0.0 (1)         null:w          TEMP_VAR0.0<0,1,0>:w    INPUT_ARG0.8<0,1,0>:w {align1};
+       (f0.0) mov  (1)         RET_ARG<1>:w TEMP_VAR0.0<0,1,0>:w                         {align1};
+       (-f0.0) mov (1)         RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w                        {align1};
+       RETURN          {align1};       
+       
+/* Compare three word data to get the max value */
+word_imax:
+       cmp.ge.f0.0 (1)         null:w          INPUT_ARG0.0<0,1,0>:w   INPUT_ARG0.4<0,1,0>:w {align1};
+       (f0.0) mov  (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.0<0,1,0>:w                    {align1};
+       (-f0.0) mov (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.4<0,1,0>:w                    {align1};
+       cmp.ge.f0.0 (1)         null:w          TEMP_VAR0.0<0,1,0>:w    INPUT_ARG0.8<0,1,0>:w {align1};
+       (f0.0) mov  (1)         RET_ARG<1>:w TEMP_VAR0.0<0,1,0>:w                         {align1};
+       (-f0.0) mov (1)         RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w                        {align1};
+       RETURN          {align1};       
+       
+word_imedian:
+       cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.4<0,1,0>:w {align1};
+       (f0.0)  jmpi (1) cmp_a_ge_b;
+       cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
+       (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
+       (f0.0) jmpi (1) cmp_end;
+       cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
+       (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
+       (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
+       jmpi (1) cmp_end;
+cmp_a_ge_b:
+       cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
+       (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
+       (f0.0) jmpi (1) cmp_end;
+       cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
+       (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
+       (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
+cmp_end:
+       RETURN  {align1};
+
+mb_pred_func:
+        mov    (8)     TEMP_VAR0.0<1>:ud       0:ud    {align1};
+       mov     (1)     TEMP_VAR0.0<1>:ub       INPUT_ARG0.2<0,1,0>:ub  {align1};
+       and     (1)   TEMP_VAR0.4<1>:uw   INPUT_ARG0.4<0,1,0>:uw        INTER_MASK:uw   {align1};
+       /* INTER16x16 mode. The bit1-0 is the prediction mode */
+       cmp.e.f0.0 (1) null:uw  TEMP_VAR0.4<0,1,0>:uw   INTER_16X16MODE:uw      {align1};
+       (f0.0)  and (1) RET_ARG<1>:uw   TEMP_VAR0.0<0,1,0>:uw   PRED_MASK {align1};
+       (f0.0)  jmpi (1) end_mb_pred;
+       /* Check whether it is INTER8x8 mode. */
+       cmp.e.f0.0 (1) null:uw  TEMP_VAR0.4<0,1,0>:uw   INTER_8X8MODE:uw        {align1};
+       (f0.0)  jmpi (1) mb_pred_func_8;
+       
+       /* Check whether it is INTER16x8 mode. */
+       cmp.e.f0.0 (1) null:uw  TEMP_VAR0.4<0,1,0>:uw   INTER_16X8MODE:uw       {align1};
+       (f0.0)  jmpi (1) mb_pred_func_168;
+mb_pred_func_816:      
+       /* Block 0/2 uses the bit1-0. Block 1/3 uses the bit3-2 */
+       mov     (1) TEMP_VAR0.8<1>:uw INPUT_ARG0.8<0,1,0>:uw    {align1};
+       and.z.f0.0 (1)  null:uw TEMP_VAR0.8<0,1,0>:uw   INTER_BLOCK1:uw {align1};
+       (f0.0)  and     (1) RET_ARG<1>:uw       TEMP_VAR0.0<0,1,0>:uw   PRED_MASK {align1};
+       (f0.0)  jmpi (1) end_mb_pred;
+       shr     (1)  TEMP_VAR0.16<1>:uw TEMP_VAR0.0<0,1,0>:uw   2:uw {align1};
+       and     (1)  RET_ARG<1>:uw      TEMP_VAR0.16<0,1,0>:uw  PRED_MASK  {align1};    
+       jmpi    (1) end_mb_pred;
+
+mb_pred_func_168:
+       /* Block 0/1 uses the bit1-0. Block 2/3 uses the bit3-2 */
+       mov     (1) TEMP_VAR0.8<1>:uw INPUT_ARG0.8<0,1,0>:uw    {align1};
+       cmp.l.f0.0 (1)  null:uw TEMP_VAR0.8<0,1,0>:uw   INTER_BLOCK2:uw {align1};
+       (f0.0)  and  (1) RET_ARG<1>:uw  TEMP_VAR0.0<0,1,0>:uw   PRED_MASK {align1};
+       (f0.0)  jmpi (1) end_mb_pred;
+       shr     (1)  TEMP_VAR0.16<1>:uw TEMP_VAR0.0<0,1,0>:uw   2:uw {align1};
+       and     (1)  RET_ARG<1>:uw      TEMP_VAR0.16<0,1,0>:uw  PRED_MASK  {align1};    
+       jmpi    (1) end_mb_pred;
+
+mb_pred_func_8:
+       /* 8X8 mode. Every block uses two bits as the prediction mode. */
+       mul     (1)  TEMP_VAR0.8<1>:uw  INPUT_ARG0.8<0,1,0>:uw  2:uw {align1};
+       shr     (1)  TEMP_VAR0.16<1>:uw TEMP_VAR0.0<0,1,0>:uw   TEMP_VAR0.8<0,1,0>:uw {align1};
+       and     (1)  RET_ARG<1>:uw      TEMP_VAR0.16<0,1,0>:uw  PRED_MASK  {align1};    
+end_mb_pred:
+       RETURN  {align1};
+       
diff --git a/src/shaders/vme/inter_bframe_gen8.g8a b/src/shaders/vme/inter_bframe_gen8.g8a
new file mode 100644 (file)
index 0000000..8aff32e
--- /dev/null
@@ -0,0 +1,2 @@
+#include "vme8.inc"
+#include "inter_bframe_gen8.asm"
diff --git a/src/shaders/vme/inter_bframe_gen8.g8b b/src/shaders/vme/inter_bframe_gen8.g8b
new file mode 100644 (file)
index 0000000..b3d74cc
--- /dev/null
@@ -0,0 +1,423 @@
+   { 0x00800001, 0x24000608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24400608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24800608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24c00608, 0x00000000, 0x00000000 },
+   { 0x00200009, 0x24002228, 0x164500a0, 0x00040004 },
+   { 0x00000040, 0x24000a28, 0x1e000400, 0xfff8fff8 },
+   { 0x00000040, 0x24040a28, 0x1e000404, 0xffffffff },
+   { 0x00000001, 0x24080e08, 0x08000000, 0x0000001f },
+   { 0x00000001, 0x24142288, 0x00000014, 0x00000000 },
+   { 0x00200009, 0x24202228, 0x164500a0, 0x00040004 },
+   { 0x00000040, 0x24200a28, 0x1e000420, 0xfffcfffc },
+   { 0x00000001, 0x24280e08, 0x08000000, 0x000f0003 },
+   { 0x00000001, 0x24342288, 0x00000014, 0x00000000 },
+   { 0x00200009, 0x24482248, 0x164500a0, 0x00040004 },
+   { 0x00000001, 0x24542288, 0x00000014, 0x00000000 },
+   { 0x00000041, 0x24881208, 0x220000a2, 0x000000a1 },
+   { 0x00000040, 0x24880208, 0x22000488, 0x000000a0 },
+   { 0x00000041, 0x24880208, 0x06000488, 0x00000018 },
+   { 0x00000001, 0x24942288, 0x00000014, 0x00000000 },
+   { 0x00600001, 0x28000208, 0x008d0400, 0x00000000 },
+   { 0x04600031, 0x23800a88, 0x0e000800, 0x02190004 },
+   { 0x00600001, 0x28000208, 0x008d0420, 0x00000000 },
+   { 0x04600031, 0x23a00a88, 0x0e000800, 0x02290004 },
+   { 0x00200009, 0x24002228, 0x164500a0, 0x00030003 },
+   { 0x00000041, 0x24000a28, 0x1e000400, 0x00020002 },
+   { 0x00000040, 0x24000a28, 0x1e000400, 0xfff8fff8 },
+   { 0x00000040, 0x24040a28, 0x1e000404, 0xffffffff },
+   { 0x00600001, 0x28000208, 0x008d0400, 0x00000000 },
+   { 0x04600031, 0x26000a88, 0x0e000800, 0x02190006 },
+   { 0x00200009, 0x24202228, 0x164500a0, 0x00030003 },
+   { 0x00000041, 0x24200a28, 0x1e000420, 0x00020002 },
+   { 0x00000040, 0x24200a28, 0x1e000420, 0xfffcfffc },
+   { 0x00000001, 0x24280e08, 0x08000000, 0x00070003 },
+   { 0x00600001, 0x28000208, 0x008d0420, 0x00000000 },
+   { 0x04600031, 0x26200a88, 0x0e000800, 0x02190006 },
+   { 0x00600001, 0x24600608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2ac00608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2a800608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20001240, 0x160000a6, 0x00040004 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000d60 },
+   { 0x00600001, 0x2ae00608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2b000608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2b200608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2b400608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00600060 },
+   { 0x00210001, 0x2af41e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000210 },
+   { 0x00000001, 0x2ae00e28, 0x08000000, 0x00000001 },
+   { 0x00200001, 0x24002248, 0x004500a0, 0x00000000 },
+   { 0x00000040, 0x24001a68, 0x1e000400, 0xffffffff },
+   { 0x00000041, 0x2b481208, 0x120000a2, 0x00000402 },
+   { 0x00000040, 0x2b480208, 0x12000b48, 0x00000400 },
+   { 0x00000041, 0x2b480208, 0x06000b48, 0x00000018 },
+   { 0x00000001, 0x2b542288, 0x00000014, 0x00000000 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x02280303 },
+   { 0x05000010, 0x20001260, 0x12000b70, 0x00000b88 },
+   { 0x00210001, 0x2af41e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000160 },
+   { 0x00000040, 0x2b480208, 0x06000b48, 0x00000003 },
+   { 0x0a800031, 0x2ba00a08, 0x0e000b40, 0x02480403 },
+   { 0x00200001, 0x2af41e68, 0x18000000, 0xffffffff },
+   { 0x00000001, 0x2fa00208, 0x00000b84, 0x00000000 },
+   { 0x00000001, 0x2fa40208, 0x00000b80, 0x00000000 },
+   { 0x00000001, 0x2fa80608, 0x00000000, 0x00000001 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00001490 },
+   { 0x00000001, 0x2aa01248, 0x00000fe4, 0x00000000 },
+   { 0x01000010, 0x20001240, 0x16000aa0, 0x00000000 },
+   { 0x00010001, 0x2af01e48, 0x18000000, 0x00010001 },
+   { 0x00010001, 0x2af41e68, 0x18000000, 0x00000000 },
+   { 0x00010001, 0x2ae40208, 0x00000bc8, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000080 },
+   { 0x01000010, 0x20001240, 0x16000aa0, 0x00010001 },
+   { 0x00010001, 0x2af21e48, 0x18000000, 0x00010001 },
+   { 0x00010001, 0x2af61e68, 0x18000000, 0x00000000 },
+   { 0x00010001, 0x2ae80208, 0x00000bcc, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000030 },
+   { 0x00200001, 0x2ae40208, 0x00450bc8, 0x00000000 },
+   { 0x00200001, 0x2af01e48, 0x18000000, 0x00010001 },
+   { 0x00200001, 0x2af41e68, 0x18000000, 0x00000000 },
+   { 0x00600001, 0x2b400608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00100010 },
+   { 0x00210001, 0x2b141e68, 0x18000000, 0xffffffff },
+   { 0x00210001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000006c0 },
+   { 0x00000001, 0x2b000e28, 0x08000000, 0x00000001 },
+   { 0x00200001, 0x24002248, 0x004500a0, 0x00000000 },
+   { 0x00000040, 0x24021a68, 0x1e000402, 0xffffffff },
+   { 0x00000041, 0x2b481208, 0x120000a2, 0x00000402 },
+   { 0x00000040, 0x2b480208, 0x12000b48, 0x00000400 },
+   { 0x00000041, 0x2b480208, 0x06000b48, 0x00000018 },
+   { 0x00000001, 0x2b542288, 0x00000014, 0x00000000 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x02280303 },
+   { 0x05000010, 0x20001260, 0x12000b70, 0x00000b88 },
+   { 0x00210001, 0x2b141e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000160 },
+   { 0x00000040, 0x2b480208, 0x06000b48, 0x00000003 },
+   { 0x0a800031, 0x2ba00a08, 0x0e000b40, 0x02480403 },
+   { 0x00200001, 0x2b141e68, 0x18000000, 0xffffffff },
+   { 0x00000001, 0x2fa00208, 0x00000b84, 0x00000000 },
+   { 0x00000001, 0x2fa40208, 0x00000b80, 0x00000000 },
+   { 0x00000001, 0x2fa80608, 0x00000000, 0x00000002 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00001230 },
+   { 0x00000001, 0x2aa01248, 0x00000fe4, 0x00000000 },
+   { 0x01000010, 0x20001240, 0x16000aa0, 0x00000000 },
+   { 0x00010001, 0x2b101e48, 0x18000000, 0x00010001 },
+   { 0x00010001, 0x2b141e68, 0x18000000, 0x00000000 },
+   { 0x00010001, 0x2b040208, 0x00000bf0, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000080 },
+   { 0x01000010, 0x20001240, 0x16000aa0, 0x00010001 },
+   { 0x00010001, 0x2b121e48, 0x18000000, 0x00010001 },
+   { 0x00010001, 0x2b161e68, 0x18000000, 0x00000000 },
+   { 0x00010001, 0x2b080208, 0x00000bf4, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000030 },
+   { 0x00200001, 0x2b101e48, 0x18000000, 0x00010001 },
+   { 0x00200001, 0x2b141e68, 0x18000000, 0x00000000 },
+   { 0x00200001, 0x2b040208, 0x00450bf0, 0x00000000 },
+   { 0x00600001, 0x2b400608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00080008 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000230 },
+   { 0x00000001, 0x2b200e28, 0x08000000, 0x00000001 },
+   { 0x00200001, 0x24002248, 0x004500a0, 0x00000000 },
+   { 0x00000040, 0x24021a68, 0x1e000402, 0xffffffff },
+   { 0x00000040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00000041, 0x2b481208, 0x120000a2, 0x00000402 },
+   { 0x00000040, 0x2b480208, 0x12000b48, 0x00000400 },
+   { 0x00000041, 0x2b480208, 0x06000b48, 0x00000018 },
+   { 0x00000001, 0x2b542288, 0x00000014, 0x00000000 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x02280303 },
+   { 0x05000010, 0x20001260, 0x12000b70, 0x00000b88 },
+   { 0x00210001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000003c0 },
+   { 0x00000040, 0x2b480208, 0x06000b48, 0x00000003 },
+   { 0x0a800031, 0x2ba00a08, 0x0e000b40, 0x02480403 },
+   { 0x00200001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00000001, 0x2fa00208, 0x00000b84, 0x00000000 },
+   { 0x00000001, 0x2fa40208, 0x00000b80, 0x00000000 },
+   { 0x00000001, 0x2fa80608, 0x00000000, 0x00000002 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000fe0 },
+   { 0x00000001, 0x2aa01248, 0x00000fe4, 0x00000000 },
+   { 0x01000010, 0x20001240, 0x16000aa0, 0x00000000 },
+   { 0x00010001, 0x2b301e48, 0x18000000, 0x00010001 },
+   { 0x00010001, 0x2b341e68, 0x18000000, 0x00000000 },
+   { 0x00010001, 0x2b240208, 0x00000bf0, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000002e0 },
+   { 0x01000010, 0x20001240, 0x16000aa0, 0x00010001 },
+   { 0x00010001, 0x2b321e48, 0x18000000, 0x00010001 },
+   { 0x00010001, 0x2b361e68, 0x18000000, 0x00000000 },
+   { 0x00010001, 0x2b280208, 0x00000bf4, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000290 },
+   { 0x00200001, 0x2b301e48, 0x18000000, 0x00010001 },
+   { 0x00200001, 0x2b341e68, 0x18000000, 0x00000000 },
+   { 0x00200001, 0x2b240208, 0x00450bf0, 0x00000000 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000250 },
+   { 0x00600001, 0x2b400608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00040004 },
+   { 0x00210001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000210 },
+   { 0x00000001, 0x2b200e28, 0x08000000, 0x00000001 },
+   { 0x00200001, 0x24002248, 0x004500a0, 0x00000000 },
+   { 0x00200040, 0x24001a68, 0x1e450400, 0xffffffff },
+   { 0x00000041, 0x2b481208, 0x120000a2, 0x00000402 },
+   { 0x00000040, 0x2b480208, 0x12000b48, 0x00000400 },
+   { 0x00000041, 0x2b480208, 0x06000b48, 0x00000018 },
+   { 0x00000001, 0x2b542288, 0x00000014, 0x00000000 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x02280303 },
+   { 0x05000010, 0x20001260, 0x12000b70, 0x00000b88 },
+   { 0x00210001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000160 },
+   { 0x00000040, 0x2b480208, 0x06000b48, 0x00000003 },
+   { 0x0a800031, 0x2ba00a88, 0x0e000b40, 0x02480403 },
+   { 0x00200001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00000001, 0x2fa00208, 0x00000b84, 0x00000000 },
+   { 0x00000001, 0x2fa40208, 0x00000b80, 0x00000000 },
+   { 0x00000001, 0x2fa80608, 0x00000000, 0x00000003 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000d80 },
+   { 0x00000001, 0x2aa01248, 0x00000fe4, 0x00000000 },
+   { 0x01000010, 0x20001240, 0x16000aa0, 0x00000000 },
+   { 0x00010001, 0x2b301e48, 0x18000000, 0x00010001 },
+   { 0x00010001, 0x2b341e68, 0x18000000, 0x00000000 },
+   { 0x00010001, 0x2b240208, 0x00000c18, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000080 },
+   { 0x01000010, 0x20001240, 0x16000aa0, 0x00010001 },
+   { 0x00010001, 0x2b321e48, 0x18000000, 0x00010001 },
+   { 0x00010001, 0x2b361e68, 0x18000000, 0x00000000 },
+   { 0x00010001, 0x2b280208, 0x00000c1c, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000030 },
+   { 0x00200001, 0x2b301e48, 0x18000000, 0x00010001 },
+   { 0x00200001, 0x2b341e68, 0x18000000, 0x00000000 },
+   { 0x00200001, 0x2b240208, 0x00450c18, 0x00000000 },
+   { 0x00000040, 0x24000a28, 0x0a000b00, 0x00000b20 },
+   { 0x01000010, 0x20000a20, 0x0e000400, 0x00000000 },
+   { 0x00110020, 0x34000000, 0x0e001400, 0x00000080 },
+   { 0x02000010, 0x20000a20, 0x0e000ae0, 0x00000001 },
+   { 0x00210001, 0x2b040208, 0x00450ae4, 0x00000000 },
+   { 0x00210001, 0x2b240208, 0x00450ae4, 0x00000000 },
+   { 0x00210001, 0x2b141248, 0x00450af4, 0x00000000 },
+   { 0x00210001, 0x2b341248, 0x00450af4, 0x00000000 },
+   { 0x00210001, 0x2ac00208, 0x00450ae4, 0x00000000 },
+   { 0x00310001, 0x2ac00608, 0x00000000, 0x00000000 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000320 },
+   { 0x00600001, 0x24000608, 0x00000000, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000af4, 0x00000000 },
+   { 0x00010040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x24040208, 0x00000ae4, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000b14, 0x00000000 },
+   { 0x00010040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x24040208, 0x00000b04, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000b34, 0x00000000 },
+   { 0x00010040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x24040208, 0x00000b24, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x2ac00208, 0x00000404, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000000c0 },
+   { 0x00000001, 0x2fa01a68, 0x00000ae4, 0x00000000 },
+   { 0x00000001, 0x2fa41a68, 0x00000b04, 0x00000000 },
+   { 0x00000001, 0x2fa81a68, 0x00000b24, 0x00000000 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x000009d0 },
+   { 0x00000001, 0x2ac01a68, 0x00000fe4, 0x00000000 },
+   { 0x00000001, 0x2fa01a68, 0x00000ae6, 0x00000000 },
+   { 0x00000001, 0x2fa41a68, 0x00000b06, 0x00000000 },
+   { 0x00000001, 0x2fa81a68, 0x00000b26, 0x00000000 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000970 },
+   { 0x00000001, 0x2ac21a68, 0x00000fe4, 0x00000000 },
+   { 0x00600001, 0x24000608, 0x00000000, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000af6, 0x00000000 },
+   { 0x00010040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x24040208, 0x00000ae8, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000b16, 0x00000000 },
+   { 0x00010040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x24040208, 0x00000b08, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000b36, 0x00000000 },
+   { 0x00010040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x24040208, 0x00000b28, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x2ac40208, 0x00000404, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000000c0 },
+   { 0x00000001, 0x2fa01a68, 0x00000ae8, 0x00000000 },
+   { 0x00000001, 0x2fa41a68, 0x00000b08, 0x00000000 },
+   { 0x00000001, 0x2fa81a68, 0x00000b28, 0x00000000 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000840 },
+   { 0x00000001, 0x2ac41a68, 0x00000fe4, 0x00000000 },
+   { 0x00000001, 0x2fa01a68, 0x00000aea, 0x00000000 },
+   { 0x00000001, 0x2fa41a68, 0x00000b0a, 0x00000000 },
+   { 0x00000001, 0x2fa81a68, 0x00000b2a, 0x00000000 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x000007e0 },
+   { 0x00000001, 0x2ac61a68, 0x00000fe4, 0x00000000 },
+   { 0x0040000c, 0x2a801a68, 0x1e690ac0, 0x00020002 },
+   { 0x00400040, 0x2a881a68, 0x1e690a80, 0x00030003 },
+   { 0x00400005, 0x2a901248, 0x16690a88, 0xfffcfffc },
+   { 0x00600001, 0x25600208, 0x008d0020, 0x00000000 },
+   { 0x00600001, 0x28400208, 0x008d0560, 0x00000000 },
+   { 0x00600001, 0x25800608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x28600208, 0x008d0580, 0x00000000 },
+   { 0x00600001, 0x28800608, 0x00000000, 0x00000000 },
+   { 0x00000001, 0x23800608, 0x00000000, 0x00000000 },
+   { 0x00000005, 0x23840208, 0x06000384, 0xff000000 },
+   { 0x00600001, 0x28a00208, 0x008d0380, 0x00000000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00010001 },
+   { 0x00000001, 0x28a52288, 0x00000400, 0x00000000 },
+   { 0x00600001, 0x28c00608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x28c02288, 0x00cf03a3, 0x00000000 },
+   { 0x00000001, 0x28d00608, 0x00000000, 0x11111111 },
+   { 0x00000001, 0x28dc0608, 0x00000000, 0x00010101 },
+   { 0x00000001, 0x28d41248, 0x00000606, 0x00000000 },
+   { 0x00400001, 0x28f00208, 0x00690608, 0x00000000 },
+   { 0x00600001, 0x28e01248, 0x00ae0622, 0x00000000 },
+   { 0x00000001, 0x247c1648, 0x10000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a4, 0x00010001 },
+   { 0x00010001, 0x247c0e88, 0x08000000, 0x00000002 },
+   { 0x00000001, 0x247d2288, 0x000000a5, 0x00000000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00200020 },
+   { 0x00000001, 0x247e2288, 0x00000400, 0x00000000 },
+   { 0x00600001, 0x28200208, 0x008d0460, 0x00000000 },
+   { 0x00000001, 0x244c0608, 0x00000000, 0x00800000 },
+   { 0x00600001, 0x28000208, 0x008d0440, 0x00000000 },
+   { 0x0d600031, 0x21800a08, 0x0e000800, 0x10782000 },
+   { 0x00600001, 0x28000208, 0x008d0480, 0x00000000 },
+   { 0x00000001, 0x28200208, 0x00000180, 0x00000000 },
+   { 0x00000001, 0x28240208, 0x00000190, 0x00000000 },
+   { 0x00000001, 0x28280208, 0x00000194, 0x00000000 },
+   { 0x00000001, 0x282c0208, 0x00000198, 0x00000000 },
+   { 0x00000001, 0x28301248, 0x0000018c, 0x00000000 },
+   { 0x00000001, 0x28340208, 0x00000188, 0x00000000 },
+   { 0x00000001, 0x28380208, 0x0000019c, 0x00000000 },
+   { 0x00000001, 0x283c0208, 0x00000488, 0x00000000 },
+   { 0x0a800031, 0x20000a60, 0x0e000800, 0x040a0203 },
+   { 0x00000001, 0x244c0608, 0x00000000, 0x00200700 },
+   { 0x00000001, 0x24561648, 0x10000000, 0x20202020 },
+   { 0x00000001, 0x24400208, 0x00000448, 0x00000000 },
+   { 0x00000040, 0x24401a68, 0x1e000440, 0xfff8fff8 },
+   { 0x00000040, 0x24421a68, 0x1e000442, 0xfff8fff8 },
+   { 0x00000001, 0x24401e68, 0x18000000, 0xfff8fff8 },
+   { 0x00000001, 0x24421e68, 0x18000000, 0xfff8fff8 },
+   { 0x00000001, 0x24440208, 0x00000440, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00600060 },
+   { 0x00010040, 0x24401a68, 0x1e000440, 0x00040004 },
+   { 0x00010040, 0x24441a68, 0x1e000444, 0x00040004 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00100010 },
+   { 0x00010040, 0x24421a68, 0x1e000442, 0x00040004 },
+   { 0x00010040, 0x24461a68, 0x1e000446, 0x00040004 },
+   { 0x00200040, 0x24401a68, 0x1a450440, 0x00450a90 },
+   { 0x00200040, 0x24441a68, 0x1a450444, 0x00450a94 },
+   { 0x00600001, 0x28000208, 0x008d0440, 0x00000000 },
+   { 0x00000001, 0x24600608, 0x00000000, 0x00000002 },
+   { 0x00000001, 0x24642288, 0x0000009c, 0x00000000 },
+   { 0x00000001, 0x24680608, 0x00000000, 0x00001212 },
+   { 0x00600001, 0x28200208, 0x008d0460, 0x00000000 },
+   { 0x00600001, 0x28400208, 0x008d0560, 0x00000000 },
+   { 0x00400001, 0x45800208, 0x00000ac0, 0x00000000 },
+   { 0x00400001, 0x45840208, 0x00000ac4, 0x00000000 },
+   { 0x00000001, 0x28800608, 0x00000000, 0x10010101 },
+   { 0x00000001, 0x28840608, 0x00000000, 0x100f0f0f },
+   { 0x00000001, 0x28880608, 0x00000000, 0x10010101 },
+   { 0x00000001, 0x288c0608, 0x00000000, 0x000f0f0f },
+   { 0x00400001, 0x28900608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x28b00608, 0x00000000, 0x00000000 },
+   { 0x08600031, 0x21800a08, 0x0e000800, 0x0c784000 },
+   { 0x00000001, 0x25740608, 0x00000000, 0x00000000 },
+   { 0x00000001, 0x25752288, 0x00000199, 0x00000000 },
+   { 0x00000001, 0x25762288, 0x0000019a, 0x00000000 },
+   { 0x00000005, 0x24001248, 0x16000180, 0x00030003 },
+   { 0x00000001, 0x25742288, 0x00000400, 0x00000000 },
+   { 0x00600001, 0x28800208, 0x008d01a0, 0x00000000 },
+   { 0x00600001, 0x28a00208, 0x008d01c0, 0x00000000 },
+   { 0x00600001, 0x28c00208, 0x008d01e0, 0x00000000 },
+   { 0x00600001, 0x28e00208, 0x008d0200, 0x00000000 },
+   { 0x00000001, 0x244c0608, 0x00000000, 0x00203000 },
+   { 0x00600001, 0x28000208, 0x008d0440, 0x00000000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00200020 },
+   { 0x00000001, 0x24662288, 0x00000400, 0x00000000 },
+   { 0x00600001, 0x28200208, 0x008d0460, 0x00000000 },
+   { 0x00600001, 0x28400208, 0x008d0560, 0x00000000 },
+   { 0x00600001, 0x28600208, 0x008d0580, 0x00000000 },
+   { 0x0d600031, 0x21800a08, 0x0e000800, 0x10786000 },
+   { 0x00000040, 0x24880208, 0x06000488, 0x00000002 },
+   { 0x00600001, 0x28000208, 0x008d0480, 0x00000000 },
+   { 0x00000001, 0x28200208, 0x00000180, 0x00000000 },
+   { 0x00000001, 0x28240208, 0x00000198, 0x00000000 },
+   { 0x00000001, 0x28280208, 0x00000188, 0x00000000 },
+   { 0x00000001, 0x282c0208, 0x00000574, 0x00000000 },
+   { 0x0a800031, 0x20000a60, 0x0e000800, 0x040a0003 },
+   { 0x00000040, 0x24880208, 0x06000488, 0x00000001 },
+   { 0x00600001, 0x28000208, 0x008d0480, 0x00000000 },
+   { 0x00600001, 0x28200208, 0x008d01a0, 0x00000000 },
+   { 0x00600001, 0x28400208, 0x008d01c0, 0x00000000 },
+   { 0x00600001, 0x28600208, 0x008d01e0, 0x00000000 },
+   { 0x00600001, 0x28800208, 0x008d0200, 0x00000000 },
+   { 0x0a800031, 0x20000a60, 0x0e000800, 0x0a0a0403 },
+   { 0x00000040, 0x24880208, 0x06000488, 0x00000008 },
+   { 0x00600001, 0x28000208, 0x008d0480, 0x00000000 },
+   { 0x00600001, 0x28200208, 0x008d0240, 0x00000000 },
+   { 0x0a800031, 0x20000a60, 0x0e000800, 0x040a0203 },
+   { 0x0a800031, 0x20000a60, 0x0e000b40, 0x0219e003 },
+   { 0x00600001, 0x2e000208, 0x008d0000, 0x00000000 },
+   { 0x07800031, 0x24000a40, 0x0e000e00, 0x82000010 },
+   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
+   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
+   { 0x06000010, 0x20001a60, 0x1a000fa0, 0x00000fa4 },
+   { 0x00010001, 0x2f601a68, 0x00000fa0, 0x00000000 },
+   { 0x00110001, 0x2f601a68, 0x00000fa4, 0x00000000 },
+   { 0x06000010, 0x20001a60, 0x1a000f60, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000f60, 0x00000000 },
+   { 0x00110001, 0x2fe41a68, 0x00000fa8, 0x00000000 },
+   { 0x00000001, 0x34000200, 0x00000fe0, 0x00000000 },
+   { 0x04000010, 0x20001a60, 0x1a000fa0, 0x00000fa4 },
+   { 0x00010001, 0x2f601a68, 0x00000fa0, 0x00000000 },
+   { 0x00110001, 0x2f601a68, 0x00000fa4, 0x00000000 },
+   { 0x04000010, 0x20001a60, 0x1a000f60, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000f60, 0x00000000 },
+   { 0x00110001, 0x2fe41a68, 0x00000fa8, 0x00000000 },
+   { 0x00000001, 0x34000200, 0x00000fe0, 0x00000000 },
+   { 0x04000010, 0x20001a60, 0x1a000fa0, 0x00000fa4 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000070 },
+   { 0x04000010, 0x20001a60, 0x1a000fa0, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000fa0, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000000a0 },
+   { 0x04000010, 0x20001a60, 0x1a000fa4, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000fa8, 0x00000000 },
+   { 0x00110001, 0x2fe41a68, 0x00000fa4, 0x00000000 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000060 },
+   { 0x04000010, 0x20001a60, 0x1a000fa4, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000fa4, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000030 },
+   { 0x04000010, 0x20001a60, 0x1a000fa0, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000fa8, 0x00000000 },
+   { 0x00110001, 0x2fe41a68, 0x00000fa0, 0x00000000 },
+   { 0x00000001, 0x34000200, 0x00000fe0, 0x00000000 },
+   { 0x00600001, 0x2f600608, 0x00000000, 0x00000000 },
+   { 0x00000001, 0x2f602288, 0x00000fa2, 0x00000000 },
+   { 0x00000005, 0x2f641248, 0x16000fa4, 0x00030003 },
+   { 0x01000010, 0x20001240, 0x16000f64, 0x00000000 },
+   { 0x00010005, 0x2fe41248, 0x16000f60, 0x00030003 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000150 },
+   { 0x01000010, 0x20001240, 0x16000f64, 0x00030003 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000100 },
+   { 0x01000010, 0x20001240, 0x16000f64, 0x00010001 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000070 },
+   { 0x00000001, 0x2f681248, 0x00000fa8, 0x00000000 },
+   { 0x01000005, 0x20001240, 0x16000f68, 0x00010001 },
+   { 0x00010005, 0x2fe41248, 0x16000f60, 0x00030003 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000000d0 },
+   { 0x00000008, 0x2f701248, 0x16000f60, 0x00020002 },
+   { 0x00000005, 0x2fe41248, 0x16000f70, 0x00030003 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x000000a0 },
+   { 0x00000001, 0x2f681248, 0x00000fa8, 0x00000000 },
+   { 0x05000010, 0x20001240, 0x16000f68, 0x00020002 },
+   { 0x00010005, 0x2fe41248, 0x16000f60, 0x00030003 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000060 },
+   { 0x00000008, 0x2f701248, 0x16000f60, 0x00020002 },
+   { 0x00000005, 0x2fe41248, 0x16000f70, 0x00030003 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000030 },
+   { 0x00000041, 0x2f681248, 0x16000fa8, 0x00020002 },
+   { 0x00000008, 0x2f701248, 0x12000f60, 0x00000f68 },
+   { 0x00000005, 0x2fe41248, 0x16000f70, 0x00030003 },
+   { 0x00000001, 0x34000200, 0x00000fe0, 0x00000000 },
diff --git a/src/shaders/vme/inter_frame_gen8.asm b/src/shaders/vme/inter_frame_gen8.asm
new file mode 100644 (file)
index 0000000..aa9fb80
--- /dev/null
@@ -0,0 +1,727 @@
+/*
+ * Copyright © <2013>, Intel Corporation.
+ *
+ * This program is licensed under the terms and conditions of the
+ * Eclipse Public License (EPL), version 1.0.  The full text of the EPL is at
+ * http://www.opensource.org/licenses/eclipse-1.0.php.
+ *
+ */
+// Modual name: Inter_frame_gen8.asm
+//
+// Make inter predition estimation for Inter-frame on gen8
+//
+
+//
+//  Now, begin source code....
+//
+
+#define SAVE_RET       add (1) RETURN_REG<1>:ud   ip:ud        32:ud
+#define        RETURN          mov (1) ip:ud   RETURN_REG<0,1,0>:ud
+
+/*
+ * __START
+ */
+__INTER_START:
+mov  (16) tmp_reg0.0<1>:UD      0x0:UD {align1};
+mov  (16) tmp_reg2.0<1>:UD      0x0:UD {align1};
+mov  (16) tmp_reg4.0<1>:UD      0x0:UD {align1} ;
+mov  (16) tmp_reg6.0<1>:UD      0x0:UD {align1} ;
+
+shl  (2) read0_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
+add  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D -8:W {align1};     /* X offset */
+add  (1) read0_header.4<1>:D    read0_header.4<0,1,0>:D -1:W {align1};     /* Y offset */ 
+mov  (1) read0_header.8<1>:UD   BLOCK_32X1 {align1};
+mov  (1) read0_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
+
+shl  (2) read1_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
+add  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D -4:W {align1};     /* X offset */
+mov  (1) read1_header.8<1>:UD   BLOCK_4X16 {align1};
+mov  (1) read1_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
+        
+shl  (2) vme_m0.8<1>:UW         orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
+mov  (1) vme_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+
+mul  (1) obw_m0.8<1>:UD         w_in_mb_uw<0,1,0>:UW orig_y_ub<0,1,0>:UB {align1};
+add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD orig_x_ub<0,1,0>:UB {align1};
+mul  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 24:UD {align1};
+mov  (1) obw_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+        
+/*
+ * Media Read Message -- fetch Luma neighbor edge pixels 
+ */
+/* ROW */
+mov  (8) msg_reg0.0<1>:UD       read0_header.0<8,8,1>:UD {align1};        
+send (8) msg_ind INEP_ROW<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 1 {align1};
+
+/* COL */
+mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};                
+send (8) msg_ind INEP_COL0<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 2 {align1};
+        
+/*
+ * Media Read Message -- fetch Chroma neighbor edge pixels 
+ */
+/* ROW */
+shl  (2) read0_header.0<1>:D    orig_xy_ub<2,2,1>:UB 3:UW {align1};    /* x * 16 , y * 8 */
+mul  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D  2:W {align1};
+add  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D -8:W {align1};     /* X offset */
+add  (1) read0_header.4<1>:D    read0_header.4<0,1,0>:D -1:W {align1};     /* Y offset */ 
+mov  (8) msg_reg0.0<1>:UD       read0_header.0<8,8,1>:UD {align1};        
+send (8) msg_ind CHROMA_ROW<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen 1 {align1};
+
+/* COL */
+shl  (2) read1_header.0<1>:D    orig_xy_ub<2,2,1>:UB 3:UW {align1};    /* x * 16, y * 8 */
+mul  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D  2:W {align1};
+add  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D -4:W {align1};     /* X offset */
+mov  (1) read1_header.8<1>:UD   BLOCK_8X4 {align1};
+mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};                
+send (8) msg_ind CHROMA_COL<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen 1 {align1};
+
+mov  (8) mb_mvp_ref.0<1>:ud    0:ud            {align1};
+mov  (8) mb_ref_win.0<1>:ud    0:ud            {align1};
+and.z.f0.0 (1)         null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
+(f0.0) jmpi (1) __mb_hwdep_end;
+/* read back the data for MB A */
+/* the layout of MB result is: rx.0(Available). rx.4(MVa), rX.8(MVb), rX.16(Pred_L0 flag),
+*  rX.18 (Pred_L1 flag), rX.20(Forward reference ID), rX.22(Backwared reference ID)
+*/
+mov  (8) mba_result.0<1>:ud    0x0:ud          {align1};
+mov  (8) mbb_result.0<1>:ud    0x0:ud          {align1};
+mov  (8) mbc_result.0<1>:ud    0x0:ud          {align1};
+mba_start:
+mov  (8) mb_msg0.0<1>:ud       0:ud            {align1};
+and.z.f0.0 (1)         null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
+/* MB A doesn't exist. Zero MV. mba_flag is zero and ref ID = -1 */
+(f0.0)  mov  (2)       mba_result.20<1>:w      -1:w    {align1};
+(f0.0)  jmpi (1)       mbb_start;
+mov  (1) mba_result.0<1>:d     MB_AVAIL                {align1};       
+mov  (2) tmp_reg0.0<1>:UW      orig_xy_ub<2,2,1>:UB    {align1};
+add  (1) tmp_reg0.0<1>:w       tmp_reg0.0<0,1,0>:w     -1:w    {align1};
+mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
+add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
+mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
+mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+
+/* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_wb.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_4,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 2
+        {align1};
+
+/* TODO: RefID is required after multi-references are added */
+cmp.l.f0.0 (1)         null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
+(f0.0)   mov (2)       mba_result.20<1>:w                      -1:w    {align1};
+(f0.0)   jmpi  (1)     mbb_start;
+
+add   (1) mb_msg0.8<1>:UD      mb_msg0.8<0,1,0>:ud     3:ud {align1};
+/* Read MV for MB A */
+/* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_mv0.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_8,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 4
+        {align1};
+/* TODO: RefID is required after multi-references are added */
+/* MV */
+mov       (2)          mba_result.4<1>:ud              mb_mv1.8<2,2,1>:ud      {align1};
+mov       (1)          mba_result.16<1>:w              MB_PRED_FLAG            {align1};
+
+mbb_start:
+mov  (8) mb_msg0.0<1>:ud       0:ud            {align1};
+and.z.f0.0 (1)         null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
+/* MB B doesn't exist. Zero MV. mba_flag is zero */
+/* If MB B doesn't exist, neither MB C nor D exists */
+(f0.0)  mov  (2)       mbb_result.20<1>:w      -1:w            {align1};
+(f0.0)  mov  (2)       mbc_result.20<1>:w      -1:w            {align1};
+(f0.0)  jmpi (1)       mb_mvp_start;
+mov  (1) mbb_result.0<1>:d     MB_AVAIL                {align1};       
+mov  (2) tmp_reg0.0<1>:UW      orig_xy_ub<2,2,1>:UB    {align1};
+add  (1) tmp_reg0.2<1>:w       tmp_reg0.2<0,1,0>:w     -1:w    {align1};
+mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
+add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
+mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
+mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+
+/* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_wb.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_4,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 2
+        {align1};
+
+/* TODO: RefID is required after multi-references are added */
+cmp.l.f0.0 (1)         null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
+(f0.0)   mov (2)       mbb_result.20<1>:w                      -1:w    {align1};
+(f0.0)   jmpi  (1)     mbc_start;
+add   (1) mb_msg0.8<1>:UD      mb_msg0.8<0,1,0>:ud     3:ud {align1};
+/* Read MV for MB B */
+/* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_mv0.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_8,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 4
+        {align1};
+/* TODO: RefID is required after multi-references are added */
+mov       (2)          mbb_result.4<1>:ud              mb_mv2.16<2,2,1>:ud     {align1};
+mov       (1)          mbb_result.16<1>:w              MB_PRED_FLAG            {align1};
+
+mbc_start:
+mov  (8) mb_msg0.0<1>:ud       0:ud            {align1};
+and.z.f0.0 (1)         null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_C:uw   {align1};
+/* MB C doesn't exist. Zero MV. mba_flag is zero */
+/* Based on h264 spec the MB D will be replaced if MB C doesn't exist */
+(f0.0)  jmpi (1)       mbd_start;
+mov  (1) mbc_result.0<1>:d     MB_AVAIL                {align1};       
+mov  (2) tmp_reg0.0<1>:UW      orig_xy_ub<2,2,1>:UB    {align1};
+add  (1) tmp_reg0.2<1>:w       tmp_reg0.2<0,1,0>:w     -1:w    {align1};
+add  (1) tmp_reg0.0<1>:w       tmp_reg0.0<0,1,0>:w     1:w     {align1};
+mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
+add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
+mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
+mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+
+/* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_wb.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_4,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 2
+        {align1};
+
+/* TODO: RefID is required after multi-references are added */
+cmp.l.f0.0 (1)         null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
+(f0.0)   mov (2)       mbc_result.20<1>:w                      -1:w    {align1};
+(f0.0)   jmpi  (1)     mb_mvp_start;
+add   (1) mb_msg0.8<1>:UD      mb_msg0.8<0,1,0>:ud     3:ud {align1};
+/* Read MV for MB C */
+/* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_mv0.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_8,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 4
+        {align1};
+/* TODO: RefID is required after multi-references are added */
+/* Forward MV */
+mov       (2)          mbc_result.4<1>:ud              mb_mv2.16<2,2,1>:ud     {align1};
+mov       (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
+
+jmpi   (1)    mb_mvp_start;
+mbd_start:
+mov  (8) mb_msg0.0<1>:ud       0:ud            {align1};
+and.z.f0.0 (1)         null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_D:uw   {align1};
+(f0.0)  mov (2)        mbc_result.20<1>:w      -1:w    {align1};
+(f0.0)  jmpi (1)       mb_mvp_start;
+mov  (1) mbc_result.0<1>:d     MB_AVAIL                {align1};       
+mov  (2) tmp_reg0.0<1>:UW      orig_xy_ub<2,2,1>:UB    {align1};
+add  (2) tmp_reg0.0<1>:w       tmp_reg0.0<2,2,1>:w     -1:w    {align1};
+mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
+add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
+mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
+mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+
+/* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_wb.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_4,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 2
+        {align1};
+
+cmp.l.f0.0 (1)         null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
+(f0.0)   mov (2)       mbc_result.20<1>:w                      -1:w    {align1};
+(f0.0)   jmpi  (1)     mb_mvp_start;
+
+add   (1) mb_msg0.8<1>:UD      mb_msg0.8<0,1,0>:ud     3:ud {align1};
+/* Read MV for MB D */
+/* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
+send (16)
+        mb_ind
+        mb_mv0.0<1>:ub
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_TYPE,
+                OBR_CONTROL_8,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 4
+        {align1};
+
+/* TODO: RefID is required after multi-references are added */
+
+/* Forward MV */
+mov       (2)          mbc_result.4<1>:ud              mb_mv3.24<2,2,1>:ud     {align1};
+mov       (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
+       
+mb_mvp_start:
+/*TODO: Add the skip prediction */
+/* Check whether both MB B and C are inavailable */
+add    (1)     tmp_reg0.0<1>:d         mbb_result.0<0,1,0>:d   mbc_result.0<0,1,0>:d   {align1};
+cmp.z.f0.0 (1) null:d                  tmp_reg0.0<0,1,0>:d     0:d     {align1};
+(-f0.0)        jmpi (1)        mb_median_start;
+cmp.nz.f0.0 (1)        null:d  mba_result.0<0,1,0>:d           1:d             {align1};
+(f0.0) mov     (1)     mbb_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};       
+(f0.0) mov     (1)     mbc_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};       
+(f0.0) mov     (1)     mbb_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};       
+(f0.0) mov     (1)     mbc_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};       
+(f0.0)  mov     (1)    mb_mvp_ref.0<1>:ud              mba_result.4<0,1,0>:ud  {align1};
+(-f0.0) mov    (1)     mb_mvp_ref.0<1>:ud              0:ud                    {align1};
+jmpi   (1)     __mb_hwdep_end;
+       
+mb_median_start:
+/* check whether only one neighbour MB has the same ref ID with the current MB */
+mov (8)        tmp_reg0.0<1>:ud                0:ud            {align1};
+cmp.z.f0.0     (1)     null:d  mba_result.20<0,1,0>:w  0:w     {align1};
+(f0.0) add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
+(f0.0) mov     (1)     tmp_reg0.4<1>:ud        mba_result.4<0,1,0>:ud  {align1};
+cmp.z.f0.0     (1)     null:d  mbb_result.20<0,1,0>:w  0:w     {align1};
+(f0.0) add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
+(f0.0) mov     (1)     tmp_reg0.4<1>:ud        mbb_result.4<0,1,0>:ud  {align1};
+cmp.z.f0.0     (1)     null:d  mbc_result.20<0,1,0>:w  0:w     {align1};
+(f0.0) add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
+(f0.0) mov     (1)     tmp_reg0.4<1>:ud        mbc_result.4<0,1,0>:ud  {align1};
+cmp.e.f0.0     (1)     null:d  tmp_reg0.0<0,1,0>:w      1:w    {align1};
+(f0.0) mov     (1)     mb_mvp_ref.0<1>:ud      tmp_reg0.4<0,1,0>:ud    {align1};
+(f0.0) jmpi (1)  __mb_hwdep_end;
+
+mov    (1)     INPUT_ARG0.0<1>:w       mba_result.4<0,1,0>:w   {align1};
+mov    (1)     INPUT_ARG0.4<1>:w       mbb_result.4<0,1,0>:w   {align1};
+mov    (1)     INPUT_ARG0.8<1>:w       mbc_result.4<0,1,0>:w   {align1};
+SAVE_RET       {align1};
+ jmpi  (1)     word_imedian;
+mov    (1)     mb_mvp_ref.0<1>:w               RET_ARG<0,1,0>:w        {align1};
+mov    (1)     INPUT_ARG0.0<1>:w       mba_result.6<0,1,0>:w   {align1};
+mov    (1)     INPUT_ARG0.4<1>:w       mbb_result.6<0,1,0>:w   {align1};
+mov    (1)     INPUT_ARG0.8<1>:w       mbc_result.6<0,1,0>:w   {align1};
+SAVE_RET       {align1};
+jmpi   (1)     word_imedian; 
+mov    (1)     mb_mvp_ref.2<1>:w               RET_ARG<0,1,0>:w        {align1};
+
+__mb_hwdep_end:
+asr    (2)     mb_ref_win.0<1>:w       mb_mvp_ref.0<2,2,1>:w   2:w     {align1};
+add    (2)     mb_ref_win.8<1>:w       mb_ref_win.0<2,2,1>:w   3:w     {align1};
+and    (2)     mb_ref_win.16<1>:uw     mb_ref_win.8<2,2,1>:uw  0xFFFC:uw {align1};
+/* m2, get the MV/Mb cost passed from constant buffer when
+spawning thread by MEDIA_OBJECT */       
+mov (8) vme_m2<1>:UD            r1.0<8,8,1>:UD {align1};
+
+mov (8) vme_msg_2<1>:UD                vme_m2.0<8,8,1>:UD {align1};
+
+/* m3 FWD/BWD cost center*/
+mov (8) vme_msg_3<1>:UD                0x0:UD {align1};                
+
+/* m4 skip center*/
+mov (8) vme_msg_4<1>:UD                0x0:UD {align1};               
+
+/* m5 */ 
+mov  (1) INEP_ROW.0<1>:UD       0x0:UD {align1};
+and  (1) INEP_ROW.4<1>:UD       INEP_ROW.4<0,1,0>:UD            0xFF000000:UD {align1};
+mov  (8) vme_msg_5<1>:UD         INEP_ROW.0<8,8,1>:UD {align1};
+
+
+/* Use the Luma mode */
+mov  (1) tmp_reg0.0<1>:UW      LUMA_INTRA_MODE:UW {align1};
+mov  (1) vme_msg_5.5<1>:UB     tmp_reg0.0<0,1,0>:UB {align1};
+
+/* m6 */        
+mov  (8) vme_msg_6<1>:UD         0x0:UD {align1};
+mov (16) vme_msg_6.0<1>:UB       INEP_COL0.3<32,8,4>:UB {align1};
+mov  (1) vme_msg_6.16<1>:UD      INTRA_PREDICTORE_MODE {align1};
+
+/* the penalty for Intra mode */
+mov  (1) vme_msg_6.28<1>:UD    0x010101:UD {align1};
+mov  (1) vme_msg_6.20<1>:UW      CHROMA_ROW.6<0,1,0>:UW {align1};
+
+
+/* m7 */
+
+mov  (4) vme_msg_7.16<1>:UD      CHROMA_ROW.8<4,4,1>:UD {align1};
+mov  (8) vme_msg_7.0<1>:UW       CHROMA_COL.2<16,8,2>:UW {align1};
+
+/*
+ * SIC VME message
+ */
+
+/* m1 */
+mov  (1) intra_flag<1>:UW       0x0:UW {align1}                     ;
+and.z.f0.0 (1) null<1>:UW transform_8x8_ub<0,1,0>:UB 1:UW {align1};
+(f0.0) mov  (1) intra_part_mask_ub<1>:UB  LUMA_INTRA_8x8_DISABLE {align1};
+
+/* assign MB intra struct from the thread payload*/
+mov (1) mb_intra_struct_ub<1>:UB input_mb_intra_ub<0,1,0>:UB {align1}; 
+
+/* Disable DC HAAR component when calculating HARR SATD block */
+mov  (1) tmp_reg0.0<1>:UW      DC_HARR_DISABLE:UW              {align1};
+mov  (1) vme_m1.30<1>:UB       tmp_reg0.0<0,1,0>:UB  {align1};
+mov  (8) vme_msg_1<1>:UD        vme_m1.0<8,8,1>:UD {align1};
+
+/* m0 */        
+mov  (1) vme_m0.12<1>:UD        INTRA_SAD_HAAR:UD {align1};    /* 16x16 Source, Intra_harr */
+mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
+
+/* after verification it will be passed by using payload */
+send (8)
+        vme_msg_ind
+        vme_wb<1>:UD
+        null
+        cre(
+                BIND_IDX_VME,
+                VME_SIC_MESSAGE_TYPE
+        )
+        mlen sic_vme_msg_length
+        rlen vme_wb_length
+        {align1};
+/*
+ * Oword Block Write message
+ */
+mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
+        
+mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
+mov  (1) msg_reg1.4<1>:UD       vme_wb.16<0,1,0>:UD     {align1};
+mov  (1) msg_reg1.8<1>:UD       vme_wb.20<0,1,0>:UD     {align1};
+mov  (1) msg_reg1.12<1>:UD      vme_wb.24<0,1,0>:UD     {align1};
+
+/* Distortion, Intra (17-16), */
+mov  (1) msg_reg1.16<1>:UW      vme_wb.12<0,1,0>:UW     {align1};
+
+mov  (1) msg_reg1.20<1>:UD      vme_wb.8<0,1,0>:UD     {align1};
+/* VME clock counts */
+mov  (1) msg_reg1.24<1>:UD      vme_wb.28<0,1,0>:UD     {align1};
+
+mov  (1) msg_reg1.28<1>:UD      obw_m0.8<0,1,0>:UD     {align1};
+
+/* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
+send (16)
+        msg_ind
+        obw_wb
+        null
+        data_port(
+                OBW_CACHE_TYPE,
+                OBW_MESSAGE_TYPE,
+                OBW_CONTROL_2,
+                OBW_BIND_IDX,
+                OBW_WRITE_COMMIT_CATEGORY,
+                OBW_HEADER_PRESENT
+        )
+        mlen 2
+        rlen obw_wb_length
+        {align1};
+
+/* IME search */
+mov  (1) vme_m0.12<1>:UD        SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR:UD {align1};    /* 16x16 Source, harr */
+mov  (1) vme_m0.22<1>:UW        REF_REGION_SIZE {align1};         /* Reference Width&Height, 48x40 */
+
+mov  (1) vme_m0.0<1>:UD                vme_m0.8<0,1,0>:UD      {align1};
+
+add  (1) vme_m0.0<1>:W          vme_m0.0<0,1,0>:W -16:W {align1};              /* Reference = (x-16,y-12)-(x+32,y+28) */
+add  (1) vme_m0.2<1>:W          vme_m0.2<0,1,0>:W -12:W {align1};
+
+mov  (1) vme_m0.0<1>:W         -16:W                   {align1};
+mov  (1) vme_m0.2<1>:W         -12:W                   {align1};
+
+mov  (1) vme_m0.4<1>:UD                vme_m0.0<0,1,0>:UD      {align1};
+
+and.z.f0.0 (1)         null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
+(f0.0) add     (1)     vme_m0.0<1>:w   vme_m0.0<0,1,0>:w       12:w    {align1};
+and.z.f0.0 (1)         null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
+(f0.0) add     (1)     vme_m0.2<1>:w   vme_m0.2<0,1,0>:w       8:w     {align1};
+add  (2) vme_m0.0<1>:w         vme_m0.0<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
+add  (2) vme_m0.4<1>:w         vme_m0.4<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
+mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
+
+mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
+/* the Max MV number is passed by constant buffer */
+mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};          
+mov  (1) vme_m1.8<1>:UD         START_CENTER + SEARCH_PATH_LEN:UD {align1};
+mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
+
+/* Setup the Cost center */
+/* currently four 8x8 share the same cost center */
+mov  (4) vme_m3.0<2>:ud                mb_mvp_ref.0<0,1,0>:ud  {align1};
+mov  (4) vme_m3.4<2>:ud                mb_mvp_ref.0<0,1,0>:ud  {align1};
+
+mov (8) vme_msg_3<1>:UD                vme_m3.0<8,8,1>:UD {align1};
+mov (8) vme_msg_2<1>:UD                vme_m2.0<8,8,1>:UD {align1};
+
+/* M4/M5 search path */
+mov  (1) vme_msg_4.0<1>:UD     0x01010101:UD {align1};
+mov  (1) vme_msg_4.4<1>:UD     0x10010101:UD {align1};
+mov  (1) vme_msg_4.8<1>:UD     0x0F0F0F0F:UD {align1};
+mov  (1) vme_msg_4.12<1>:UD    0x100F0F0F:UD {align1};
+mov  (1) vme_msg_4.16<1>:UD    0x01010101:UD {align1};
+mov  (1) vme_msg_4.20<1>:UD    0x10010101:UD {align1};
+mov  (1) vme_msg_4.24<1>:UD    0x0F0F0F0F:UD {align1};
+mov  (1) vme_msg_4.28<1>:UD    0x100F0F0F:UD {align1};
+
+mov  (1) vme_msg_5.0<1>:UD     0x01010101:UD {align1};
+mov  (1) vme_msg_5.4<1>:UD     0x10010101:UD {align1};
+mov  (1) vme_msg_5.8<1>:UD     0x0F0F0F0F:UD {align1};
+mov  (1) vme_msg_5.12<1>:UD    0x000F0F0F:UD {align1};
+
+mov  (4) vme_msg_5.16<1>:UD    0x0:UD {align1};
+
+send (8)
+        vme_msg_ind
+        vme_wb<1>:UD
+        null
+        vme(
+                BIND_IDX_VME,
+                0,
+                0,
+                VME_IME_MESSAGE_TYPE
+        )
+        mlen ime_vme_msg_length
+        rlen vme_wb_length {align1};
+
+/* Set Macroblock-shape/mode for FBR */
+
+mov  (1) vme_m2.20<1>:UD       0x0:UD {align1};
+mov  (1) vme_m2.21<1>:UB       vme_wb.25<0,1,0>:UB     {align1};
+mov  (1) vme_m2.22<1>:UB       vme_wb.26<0,1,0>:UB     {align1};
+
+and  (1) tmp_reg0.0<1>:UW      vme_wb.0<0,1,0>:UW      0x03:UW {align1};
+mov  (1) vme_m2.20<1>:UB       tmp_reg0.0<0,1,0>:UB    {align1};
+
+/* Send FBR message into CRE */
+
+mov  (8) vme_msg_4.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
+mov  (8) vme_msg_5.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
+mov  (8) vme_msg_6.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
+mov  (8) vme_msg_7.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
+
+mov  (1) vme_m0.12<1>:UD       INTER_SAD_HAAR + SUB_PEL_MODE_QUARTER + FBR_BME_DISABLE:UD {align1};    /* 16x16 Source, 1/4 pixel, harr, BME disable */
+mov  (8) vme_msg_0.0<1>:UD     vme_m0.0<8,8,1>:UD  {align1};
+mov  (8) vme_msg_1.0<1>:UD     vme_m1.0<8,8,1>:UD  {align1};
+
+mov  (8) vme_msg_2.0<1>:UD             vme_m2.0<8,8,1>:UD      {align1};
+mov  (8) vme_msg_3.0<1>:UD             vme_m3.0<8,8,1>:UD      {align1};
+
+/* after verification it will be passed by using payload */
+send (8)
+        vme_msg_ind
+        vme_wb<1>:UD
+        null
+        cre(
+                BIND_IDX_VME,
+                VME_FBR_MESSAGE_TYPE
+        )
+        mlen fbr_vme_msg_length
+        rlen vme_wb_length
+        {align1};
+
+add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x02:UD {align1};
+mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
+/* write FME info */
+mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
+
+mov  (1) msg_reg1.4<1>:UD       vme_wb.24<0,1,0>:UD     {align1};
+/* Inter distortion of FME */
+mov  (1) msg_reg1.8<1>:UD       vme_wb.8<0,1,0>:UD     {align1};
+
+mov  (1) msg_reg1.12<1>:UD     vme_m2.20<0,1,0>:UD {align1};
+
+/* bind index 3, write  oword (16bytes), msg type: 8(OWord Block Write) */
+send (16)
+        msg_ind
+        obw_wb
+        null
+        data_port(
+                OBW_CACHE_TYPE,
+                OBW_MESSAGE_TYPE,
+                OBW_CONTROL_0,
+                OBW_BIND_IDX,
+                OBW_WRITE_COMMIT_CATEGORY,
+                OBW_HEADER_PRESENT
+        )
+        mlen 2
+        rlen obw_wb_length
+        {align1};
+
+/* Write FME/BME MV */
+add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x01:UD {align1};
+mov  (8) msg_reg0.0<1>:UD       obw_m0.0<8,8,1>:UD {align1};
+
+
+mov  (8) msg_reg1.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
+mov  (8) msg_reg2.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
+mov  (8) msg_reg3.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
+mov  (8) msg_reg4.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
+/* bind index 3, write  8 oword (128 bytes), msg type: 8(OWord Block Write) */
+send (16)
+        msg_ind
+        obw_wb
+        null
+        data_port(
+                OBW_CACHE_TYPE,
+                OBW_MESSAGE_TYPE,
+                OBW_CONTROL_8,
+                OBW_BIND_IDX,
+                OBW_WRITE_COMMIT_CATEGORY,
+                OBW_HEADER_PRESENT
+        )
+        mlen 5
+        rlen obw_wb_length
+        {align1};
+
+/* Write FME/BME RefID */
+add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x08:UD {align1};
+mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
+
+mov  (8) msg_reg1.0<1>:UD      vme_wb6.0<8,8,1>:UD {align1};
+
+/* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
+send (16)
+        msg_ind
+        obw_wb
+        null
+        data_port(
+                OBW_CACHE_TYPE,
+                OBW_MESSAGE_TYPE,
+                OBW_CONTROL_2,
+                OBW_BIND_IDX,
+                OBW_WRITE_COMMIT_CATEGORY,
+                OBW_HEADER_PRESENT
+        )
+        mlen 2
+        rlen obw_wb_length
+        {align1};
+
+/* Issue message fence so that the previous write message is committed */
+send (16)
+        mb_ind
+        mb_wb.0<1>:ud
+       NULL
+        data_port(
+                OBR_CACHE_TYPE,
+                OBR_MESSAGE_FENCE,
+                OBR_MF_COMMIT,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 1
+        {align1};
+
+__EXIT: 
+/*
+ * kill thread
+ */        
+mov  (8) ts_msg_reg0<1>:UD         r0<8,8,1>:UD {align1};
+send (16) ts_msg_ind acc0<1>UW null thread_spawner(0, 0, 1) mlen 1 rlen 0 {align1 EOT};
+
+
+       nop             ;
+       nop             ;
+/* Compare three word data to get the min value */
+word_imin:
+       cmp.le.f0.0 (1)         null:w          INPUT_ARG0.0<0,1,0>:w   INPUT_ARG0.4<0,1,0>:w {align1};
+       (f0.0) mov  (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.0<0,1,0>:w                    {align1};
+       (-f0.0) mov (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.4<0,1,0>:w                    {align1};
+       cmp.le.f0.0 (1)         null:w          TEMP_VAR0.0<0,1,0>:w    INPUT_ARG0.8<0,1,0>:w {align1};
+       (f0.0) mov  (1)         RET_ARG<1>:w TEMP_VAR0.0<0,1,0>:w                         {align1};
+       (-f0.0) mov (1)         RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w                        {align1};
+       RETURN          {align1};       
+       
+/* Compare three word data to get the max value */
+word_imax:
+       cmp.ge.f0.0 (1)         null:w          INPUT_ARG0.0<0,1,0>:w   INPUT_ARG0.4<0,1,0>:w {align1};
+       (f0.0) mov  (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.0<0,1,0>:w                    {align1};
+       (-f0.0) mov (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.4<0,1,0>:w                    {align1};
+       cmp.ge.f0.0 (1)         null:w          TEMP_VAR0.0<0,1,0>:w    INPUT_ARG0.8<0,1,0>:w {align1};
+       (f0.0) mov  (1)         RET_ARG<1>:w TEMP_VAR0.0<0,1,0>:w                         {align1};
+       (-f0.0) mov (1)         RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w                        {align1};
+       RETURN          {align1};       
+       
+word_imedian:
+       cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.4<0,1,0>:w {align1};
+       (f0.0)  jmpi (1) cmp_a_ge_b;
+       cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
+       (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
+       (f0.0) jmpi (1) cmp_end;
+       cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
+       (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
+       (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
+       jmpi (1) cmp_end;
+cmp_a_ge_b:
+       cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
+       (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
+       (f0.0) jmpi (1) cmp_end;
+       cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
+       (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
+       (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
+cmp_end:
+       RETURN  {align1};
+
diff --git a/src/shaders/vme/inter_frame_gen8.g8a b/src/shaders/vme/inter_frame_gen8.g8a
new file mode 100644 (file)
index 0000000..f514dd3
--- /dev/null
@@ -0,0 +1,2 @@
+#include "vme8.inc"
+#include "inter_frame_gen8.asm"
diff --git a/src/shaders/vme/inter_frame_gen8.g8b b/src/shaders/vme/inter_frame_gen8.g8b
new file mode 100644 (file)
index 0000000..c4e2c97
--- /dev/null
@@ -0,0 +1,300 @@
+   { 0x00800001, 0x24000608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24400608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24800608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24c00608, 0x00000000, 0x00000000 },
+   { 0x00200009, 0x24002228, 0x164500a0, 0x00040004 },
+   { 0x00000040, 0x24000a28, 0x1e000400, 0xfff8fff8 },
+   { 0x00000040, 0x24040a28, 0x1e000404, 0xffffffff },
+   { 0x00000001, 0x24080e08, 0x08000000, 0x0000001f },
+   { 0x00000001, 0x24142288, 0x00000014, 0x00000000 },
+   { 0x00200009, 0x24202228, 0x164500a0, 0x00040004 },
+   { 0x00000040, 0x24200a28, 0x1e000420, 0xfffcfffc },
+   { 0x00000001, 0x24280e08, 0x08000000, 0x000f0003 },
+   { 0x00000001, 0x24342288, 0x00000014, 0x00000000 },
+   { 0x00200009, 0x24482248, 0x164500a0, 0x00040004 },
+   { 0x00000001, 0x24542288, 0x00000014, 0x00000000 },
+   { 0x00000041, 0x24881208, 0x220000a2, 0x000000a1 },
+   { 0x00000040, 0x24880208, 0x22000488, 0x000000a0 },
+   { 0x00000041, 0x24880208, 0x06000488, 0x00000018 },
+   { 0x00000001, 0x24942288, 0x00000014, 0x00000000 },
+   { 0x00600001, 0x28000208, 0x008d0400, 0x00000000 },
+   { 0x04600031, 0x23800a88, 0x0e000800, 0x02190004 },
+   { 0x00600001, 0x28000208, 0x008d0420, 0x00000000 },
+   { 0x04600031, 0x23a00a88, 0x0e000800, 0x02290004 },
+   { 0x00200009, 0x24002228, 0x164500a0, 0x00030003 },
+   { 0x00000041, 0x24000a28, 0x1e000400, 0x00020002 },
+   { 0x00000040, 0x24000a28, 0x1e000400, 0xfff8fff8 },
+   { 0x00000040, 0x24040a28, 0x1e000404, 0xffffffff },
+   { 0x00600001, 0x28000208, 0x008d0400, 0x00000000 },
+   { 0x04600031, 0x26000a88, 0x0e000800, 0x02190006 },
+   { 0x00200009, 0x24202228, 0x164500a0, 0x00030003 },
+   { 0x00000041, 0x24200a28, 0x1e000420, 0x00020002 },
+   { 0x00000040, 0x24200a28, 0x1e000420, 0xfffcfffc },
+   { 0x00000001, 0x24280e08, 0x08000000, 0x00070003 },
+   { 0x00600001, 0x28000208, 0x008d0420, 0x00000000 },
+   { 0x04600031, 0x26200a88, 0x0e000800, 0x02190006 },
+   { 0x00600001, 0x2ac00608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2a800608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20001240, 0x160000a6, 0x00040004 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000750 },
+   { 0x00600001, 0x2ae00608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2b000608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2b200608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2b400608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00600060 },
+   { 0x00210001, 0x2af41e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000000f0 },
+   { 0x00000001, 0x2ae00e28, 0x08000000, 0x00000001 },
+   { 0x00200001, 0x24002248, 0x004500a0, 0x00000000 },
+   { 0x00000040, 0x24001a68, 0x1e000400, 0xffffffff },
+   { 0x00000041, 0x2b481208, 0x120000a2, 0x00000402 },
+   { 0x00000040, 0x2b480208, 0x12000b48, 0x00000400 },
+   { 0x00000041, 0x2b480208, 0x06000b48, 0x00000018 },
+   { 0x00000001, 0x2b542288, 0x00000014, 0x00000000 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x02280303 },
+   { 0x05000010, 0x20001260, 0x12000b70, 0x00000b88 },
+   { 0x00210001, 0x2af41e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000040 },
+   { 0x00000040, 0x2b480208, 0x06000b48, 0x00000003 },
+   { 0x0a800031, 0x2ba00a08, 0x0e000b40, 0x02480403 },
+   { 0x00200001, 0x2ae40208, 0x00450bc8, 0x00000000 },
+   { 0x00000001, 0x2af01e68, 0x18000000, 0x00010001 },
+   { 0x00600001, 0x2b400608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00100010 },
+   { 0x00210001, 0x2b141e68, 0x18000000, 0xffffffff },
+   { 0x00210001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000360 },
+   { 0x00000001, 0x2b000e28, 0x08000000, 0x00000001 },
+   { 0x00200001, 0x24002248, 0x004500a0, 0x00000000 },
+   { 0x00000040, 0x24021a68, 0x1e000402, 0xffffffff },
+   { 0x00000041, 0x2b481208, 0x120000a2, 0x00000402 },
+   { 0x00000040, 0x2b480208, 0x12000b48, 0x00000400 },
+   { 0x00000041, 0x2b480208, 0x06000b48, 0x00000018 },
+   { 0x00000001, 0x2b542288, 0x00000014, 0x00000000 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x02280303 },
+   { 0x05000010, 0x20001260, 0x12000b70, 0x00000b88 },
+   { 0x00210001, 0x2b141e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000040 },
+   { 0x00000040, 0x2b480208, 0x06000b48, 0x00000003 },
+   { 0x0a800031, 0x2ba00a08, 0x0e000b40, 0x02480403 },
+   { 0x00200001, 0x2b040208, 0x00450bf0, 0x00000000 },
+   { 0x00000001, 0x2b101e68, 0x18000000, 0x00010001 },
+   { 0x00600001, 0x2b400608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00080008 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000110 },
+   { 0x00000001, 0x2b200e28, 0x08000000, 0x00000001 },
+   { 0x00200001, 0x24002248, 0x004500a0, 0x00000000 },
+   { 0x00000040, 0x24021a68, 0x1e000402, 0xffffffff },
+   { 0x00000040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00000041, 0x2b481208, 0x120000a2, 0x00000402 },
+   { 0x00000040, 0x2b480208, 0x12000b48, 0x00000400 },
+   { 0x00000041, 0x2b480208, 0x06000b48, 0x00000018 },
+   { 0x00000001, 0x2b542288, 0x00000014, 0x00000000 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x02280303 },
+   { 0x05000010, 0x20001260, 0x12000b70, 0x00000b88 },
+   { 0x00210001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000180 },
+   { 0x00000040, 0x2b480208, 0x06000b48, 0x00000003 },
+   { 0x0a800031, 0x2ba00a08, 0x0e000b40, 0x02480403 },
+   { 0x00200001, 0x2b240208, 0x00450bf0, 0x00000000 },
+   { 0x00000001, 0x2b301e68, 0x18000000, 0x00010001 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000130 },
+   { 0x00600001, 0x2b400608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00040004 },
+   { 0x00210001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000000f0 },
+   { 0x00000001, 0x2b200e28, 0x08000000, 0x00000001 },
+   { 0x00200001, 0x24002248, 0x004500a0, 0x00000000 },
+   { 0x00200040, 0x24001a68, 0x1e450400, 0xffffffff },
+   { 0x00000041, 0x2b481208, 0x120000a2, 0x00000402 },
+   { 0x00000040, 0x2b480208, 0x12000b48, 0x00000400 },
+   { 0x00000041, 0x2b480208, 0x06000b48, 0x00000018 },
+   { 0x00000001, 0x2b542288, 0x00000014, 0x00000000 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x02280303 },
+   { 0x05000010, 0x20001260, 0x12000b70, 0x00000b88 },
+   { 0x00210001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000040 },
+   { 0x00000040, 0x2b480208, 0x06000b48, 0x00000003 },
+   { 0x0a800031, 0x2ba00a88, 0x0e000b40, 0x02480403 },
+   { 0x00200001, 0x2b240208, 0x00450c18, 0x00000000 },
+   { 0x00000001, 0x2b301e68, 0x18000000, 0x00010001 },
+   { 0x00000040, 0x24000a28, 0x0a000b00, 0x00000b20 },
+   { 0x01000010, 0x20000a20, 0x0e000400, 0x00000000 },
+   { 0x00110020, 0x34000000, 0x0e001400, 0x00000080 },
+   { 0x02000010, 0x20000a20, 0x0e000ae0, 0x00000001 },
+   { 0x00010001, 0x2b040208, 0x00000ae4, 0x00000000 },
+   { 0x00010001, 0x2b240208, 0x00000ae4, 0x00000000 },
+   { 0x00010001, 0x2b141248, 0x00000af4, 0x00000000 },
+   { 0x00010001, 0x2b341248, 0x00000af4, 0x00000000 },
+   { 0x00010001, 0x2ac00208, 0x00000ae4, 0x00000000 },
+   { 0x00110001, 0x2ac00608, 0x00000000, 0x00000000 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000190 },
+   { 0x00600001, 0x24000608, 0x00000000, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000af4, 0x00000000 },
+   { 0x00010040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x24040208, 0x00000ae4, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000b14, 0x00000000 },
+   { 0x00010040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x24040208, 0x00000b04, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000b34, 0x00000000 },
+   { 0x00010040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x24040208, 0x00000b24, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x2ac00208, 0x00000404, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000000c0 },
+   { 0x00000001, 0x2fa01a68, 0x00000ae4, 0x00000000 },
+   { 0x00000001, 0x2fa41a68, 0x00000b04, 0x00000000 },
+   { 0x00000001, 0x2fa81a68, 0x00000b24, 0x00000000 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000870 },
+   { 0x00000001, 0x2ac01a68, 0x00000fe4, 0x00000000 },
+   { 0x00000001, 0x2fa01a68, 0x00000ae6, 0x00000000 },
+   { 0x00000001, 0x2fa41a68, 0x00000b06, 0x00000000 },
+   { 0x00000001, 0x2fa81a68, 0x00000b26, 0x00000000 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000810 },
+   { 0x00000001, 0x2ac21a68, 0x00000fe4, 0x00000000 },
+   { 0x0020000c, 0x2a801a68, 0x1e450ac0, 0x00020002 },
+   { 0x00200040, 0x2a881a68, 0x1e450a80, 0x00030003 },
+   { 0x00200005, 0x2a901248, 0x16450a88, 0xfffcfffc },
+   { 0x00600001, 0x25600208, 0x008d0020, 0x00000000 },
+   { 0x00600001, 0x28400208, 0x008d0560, 0x00000000 },
+   { 0x00600001, 0x28600608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x28800608, 0x00000000, 0x00000000 },
+   { 0x00000001, 0x23800608, 0x00000000, 0x00000000 },
+   { 0x00000005, 0x23840208, 0x06000384, 0xff000000 },
+   { 0x00600001, 0x28a00208, 0x008d0380, 0x00000000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00010001 },
+   { 0x00000001, 0x28a52288, 0x00000400, 0x00000000 },
+   { 0x00600001, 0x28c00608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x28c02288, 0x00cf03a3, 0x00000000 },
+   { 0x00000001, 0x28d00608, 0x00000000, 0x11111111 },
+   { 0x00000001, 0x28dc0608, 0x00000000, 0x00010101 },
+   { 0x00000001, 0x28d41248, 0x00000606, 0x00000000 },
+   { 0x00400001, 0x28f00208, 0x00690608, 0x00000000 },
+   { 0x00600001, 0x28e01248, 0x00ae0622, 0x00000000 },
+   { 0x00000001, 0x247c1648, 0x10000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a4, 0x00010001 },
+   { 0x00010001, 0x247c0e88, 0x08000000, 0x00000002 },
+   { 0x00000001, 0x247d2288, 0x000000a5, 0x00000000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00200020 },
+   { 0x00000001, 0x247e2288, 0x00000400, 0x00000000 },
+   { 0x00600001, 0x28200208, 0x008d0460, 0x00000000 },
+   { 0x00000001, 0x244c0608, 0x00000000, 0x00800000 },
+   { 0x00600001, 0x28000208, 0x008d0440, 0x00000000 },
+   { 0x0d600031, 0x21800a08, 0x0e000800, 0x10782000 },
+   { 0x00600001, 0x28000208, 0x008d0480, 0x00000000 },
+   { 0x00000001, 0x28200208, 0x00000180, 0x00000000 },
+   { 0x00000001, 0x28240208, 0x00000190, 0x00000000 },
+   { 0x00000001, 0x28280208, 0x00000194, 0x00000000 },
+   { 0x00000001, 0x282c0208, 0x00000198, 0x00000000 },
+   { 0x00000001, 0x28301248, 0x0000018c, 0x00000000 },
+   { 0x00000001, 0x28340208, 0x00000188, 0x00000000 },
+   { 0x00000001, 0x28380208, 0x0000019c, 0x00000000 },
+   { 0x00000001, 0x283c0208, 0x00000488, 0x00000000 },
+   { 0x0a800031, 0x20000a60, 0x0e000800, 0x040a0203 },
+   { 0x00000001, 0x244c0608, 0x00000000, 0x00200000 },
+   { 0x00000001, 0x24561648, 0x10000000, 0x28302830 },
+   { 0x00000001, 0x24400208, 0x00000448, 0x00000000 },
+   { 0x00000040, 0x24401a68, 0x1e000440, 0xfff0fff0 },
+   { 0x00000040, 0x24421a68, 0x1e000442, 0xfff4fff4 },
+   { 0x00000001, 0x24401e68, 0x18000000, 0xfff0fff0 },
+   { 0x00000001, 0x24421e68, 0x18000000, 0xfff4fff4 },
+   { 0x00000001, 0x24440208, 0x00000440, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00600060 },
+   { 0x00010040, 0x24401a68, 0x1e000440, 0x000c000c },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00100010 },
+   { 0x00010040, 0x24421a68, 0x1e000442, 0x00080008 },
+   { 0x00200040, 0x24401a68, 0x1a450440, 0x00450a90 },
+   { 0x00200040, 0x24441a68, 0x1a450444, 0x00450a90 },
+   { 0x00600001, 0x28000208, 0x008d0440, 0x00000000 },
+   { 0x00000001, 0x24600608, 0x00000000, 0x00000002 },
+   { 0x00000001, 0x24642288, 0x0000009c, 0x00000000 },
+   { 0x00000001, 0x24680608, 0x00000000, 0x30003030 },
+   { 0x00600001, 0x28200208, 0x008d0460, 0x00000000 },
+   { 0x00400001, 0x45800208, 0x00000ac0, 0x00000000 },
+   { 0x00400001, 0x45840208, 0x00000ac0, 0x00000000 },
+   { 0x00600001, 0x28600208, 0x008d0580, 0x00000000 },
+   { 0x00600001, 0x28400208, 0x008d0560, 0x00000000 },
+   { 0x00000001, 0x28800608, 0x00000000, 0x01010101 },
+   { 0x00000001, 0x28840608, 0x00000000, 0x10010101 },
+   { 0x00000001, 0x28880608, 0x00000000, 0x0f0f0f0f },
+   { 0x00000001, 0x288c0608, 0x00000000, 0x100f0f0f },
+   { 0x00000001, 0x28900608, 0x00000000, 0x01010101 },
+   { 0x00000001, 0x28940608, 0x00000000, 0x10010101 },
+   { 0x00000001, 0x28980608, 0x00000000, 0x0f0f0f0f },
+   { 0x00000001, 0x289c0608, 0x00000000, 0x100f0f0f },
+   { 0x00000001, 0x28a00608, 0x00000000, 0x01010101 },
+   { 0x00000001, 0x28a40608, 0x00000000, 0x10010101 },
+   { 0x00000001, 0x28a80608, 0x00000000, 0x0f0f0f0f },
+   { 0x00000001, 0x28ac0608, 0x00000000, 0x000f0f0f },
+   { 0x00400001, 0x28b00608, 0x00000000, 0x00000000 },
+   { 0x08600031, 0x21800a08, 0x0e000800, 0x0c784000 },
+   { 0x00000001, 0x25740608, 0x00000000, 0x00000000 },
+   { 0x00000001, 0x25752288, 0x00000199, 0x00000000 },
+   { 0x00000001, 0x25762288, 0x0000019a, 0x00000000 },
+   { 0x00000005, 0x24001248, 0x16000180, 0x00030003 },
+   { 0x00000001, 0x25742288, 0x00000400, 0x00000000 },
+   { 0x00600001, 0x28800208, 0x008d01a0, 0x00000000 },
+   { 0x00600001, 0x28a00208, 0x008d01c0, 0x00000000 },
+   { 0x00600001, 0x28c00208, 0x008d01e0, 0x00000000 },
+   { 0x00600001, 0x28e00208, 0x008d0200, 0x00000000 },
+   { 0x00000001, 0x244c0608, 0x00000000, 0x00243000 },
+   { 0x00600001, 0x28000208, 0x008d0440, 0x00000000 },
+   { 0x00600001, 0x28200208, 0x008d0460, 0x00000000 },
+   { 0x00600001, 0x28400208, 0x008d0560, 0x00000000 },
+   { 0x00600001, 0x28600208, 0x008d0580, 0x00000000 },
+   { 0x0d600031, 0x21800a08, 0x0e000800, 0x10786000 },
+   { 0x00000040, 0x24880208, 0x06000488, 0x00000002 },
+   { 0x00600001, 0x28000208, 0x008d0480, 0x00000000 },
+   { 0x00000001, 0x28200208, 0x00000180, 0x00000000 },
+   { 0x00000001, 0x28240208, 0x00000198, 0x00000000 },
+   { 0x00000001, 0x28280208, 0x00000188, 0x00000000 },
+   { 0x00000001, 0x282c0208, 0x00000574, 0x00000000 },
+   { 0x0a800031, 0x20000a60, 0x0e000800, 0x040a0003 },
+   { 0x00000040, 0x24880208, 0x06000488, 0x00000001 },
+   { 0x00600001, 0x28000208, 0x008d0480, 0x00000000 },
+   { 0x00600001, 0x28200208, 0x008d01a0, 0x00000000 },
+   { 0x00600001, 0x28400208, 0x008d01c0, 0x00000000 },
+   { 0x00600001, 0x28600208, 0x008d01e0, 0x00000000 },
+   { 0x00600001, 0x28800208, 0x008d0200, 0x00000000 },
+   { 0x0a800031, 0x20000a60, 0x0e000800, 0x0a0a0403 },
+   { 0x00000040, 0x24880208, 0x06000488, 0x00000008 },
+   { 0x00600001, 0x28000208, 0x008d0480, 0x00000000 },
+   { 0x00600001, 0x28200208, 0x008d0240, 0x00000000 },
+   { 0x0a800031, 0x20000a60, 0x0e000800, 0x040a0203 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x0219e003 },
+   { 0x00600001, 0x2e000208, 0x008d0000, 0x00000000 },
+   { 0x07800031, 0x24000a40, 0x0e000e00, 0x82000010 },
+   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
+   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
+   { 0x06000010, 0x20001a60, 0x1a000fa0, 0x00000fa4 },
+   { 0x00010001, 0x2f601a68, 0x00000fa0, 0x00000000 },
+   { 0x00110001, 0x2f601a68, 0x00000fa4, 0x00000000 },
+   { 0x06000010, 0x20001a60, 0x1a000f60, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000f60, 0x00000000 },
+   { 0x00110001, 0x2fe41a68, 0x00000fa8, 0x00000000 },
+   { 0x00000001, 0x34000200, 0x00000fe0, 0x00000000 },
+   { 0x04000010, 0x20001a60, 0x1a000fa0, 0x00000fa4 },
+   { 0x00010001, 0x2f601a68, 0x00000fa0, 0x00000000 },
+   { 0x00110001, 0x2f601a68, 0x00000fa4, 0x00000000 },
+   { 0x04000010, 0x20001a60, 0x1a000f60, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000f60, 0x00000000 },
+   { 0x00110001, 0x2fe41a68, 0x00000fa8, 0x00000000 },
+   { 0x00000001, 0x34000200, 0x00000fe0, 0x00000000 },
+   { 0x04000010, 0x20001a60, 0x1a000fa0, 0x00000fa4 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000070 },
+   { 0x04000010, 0x20001a60, 0x1a000fa0, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000fa0, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000000a0 },
+   { 0x04000010, 0x20001a60, 0x1a000fa4, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000fa8, 0x00000000 },
+   { 0x00110001, 0x2fe41a68, 0x00000fa4, 0x00000000 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000060 },
+   { 0x04000010, 0x20001a60, 0x1a000fa4, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000fa4, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000030 },
+   { 0x04000010, 0x20001a60, 0x1a000fa0, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000fa8, 0x00000000 },
+   { 0x00110001, 0x2fe41a68, 0x00000fa0, 0x00000000 },
+   { 0x00000001, 0x34000200, 0x00000fe0, 0x00000000 },
diff --git a/src/shaders/vme/intra_frame_gen8.asm b/src/shaders/vme/intra_frame_gen8.asm
new file mode 100644 (file)
index 0000000..41cdb3a
--- /dev/null
@@ -0,0 +1,185 @@
+/*
+ * Copyright © <2010>, Intel Corporation.
+ *
+ * This program is licensed under the terms and conditions of the
+ * Eclipse Public License (EPL), version 1.0.  The full text of the EPL is at
+ * http://www.opensource.org/licenses/eclipse-1.0.php.
+ *
+ */
+// Modual name: IntraFrame_gen8.asm
+//
+// Make intra predition estimation for Intra frame on Gen8
+//
+
+//
+//  Now, begin source code....
+//
+
+/*
+ * __START
+ */
+__INTRA_START:
+mov  (16) tmp_reg0.0<1>:UD      0x0:UD {align1};
+mov  (16) tmp_reg2.0<1>:UD      0x0:UD {align1};
+mov  (16) tmp_reg4.0<1>:UD      0x0:UD {align1} ;
+mov  (16) tmp_reg6.0<1>:UD      0x0:UD {align1} ;
+
+shl  (2) read0_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
+add  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D -8:W {align1};     /* X offset */
+add  (1) read0_header.4<1>:D    read0_header.4<0,1,0>:D -1:W {align1};     /* Y offset */ 
+mov  (1) read0_header.8<1>:UD   BLOCK_32X1 {align1};
+mov  (1) read0_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
+
+shl  (2) read1_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
+add  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D -4:W {align1};     /* X offset */
+mov  (1) read1_header.8<1>:UD   BLOCK_4X16 {align1};
+mov  (1) read1_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
+        
+shl  (2) vme_m0.8<1>:UW         orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
+mov  (1) vme_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+
+mul  (1) obw_m0.8<1>:UD         w_in_mb_uw<0,1,0>:UW orig_y_ub<0,1,0>:UB {align1};
+add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD orig_x_ub<0,1,0>:UB {align1};
+mul  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x02:UD {align1};
+mov  (1) obw_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
+        
+/*
+ * Media Read Message -- fetch Luma neighbor edge pixels 
+ */
+/* ROW */
+mov  (8) msg_reg0.0<1>:UD       read0_header.0<8,8,1>:UD {align1};        
+send (8) msg_ind INEP_ROW<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 1 {align1};
+
+/* COL */
+mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};                
+send (8) msg_ind INEP_COL0<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 2 {align1};
+        
+/*
+ * Media Read Message -- fetch Chroma neighbor edge pixels 
+ */
+/* ROW */
+shl  (2) read0_header.0<1>:D    orig_xy_ub<2,2,1>:UB 3:UW {align1};    /* x * 16 , y * 8 */
+mul  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D  2:W {align1};
+add  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D -8:W {align1};     /* X offset */
+add  (1) read0_header.4<1>:D    read0_header.4<0,1,0>:D -1:W {align1};     /* Y offset */ 
+mov  (8) msg_reg0.0<1>:UD       read0_header.0<8,8,1>:UD {align1};        
+send (8) msg_ind CHROMA_ROW<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen 1 {align1};
+
+/* COL */
+shl  (2) read1_header.0<1>:D    orig_xy_ub<2,2,1>:UB 3:UW {align1};    /* x * 16, y * 8 */
+mul  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D  2:W {align1};
+add  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D -4:W {align1};     /* X offset */
+mov  (1) read1_header.8<1>:UD   BLOCK_8X4 {align1};
+mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};                
+send (8) msg_ind CHROMA_COL<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen 1 {align1};
+
+/* m2, get the MV/Mb cost passed by constant buffer 
+when creating EU thread by MEDIA_OBJECT */       
+mov (8) vme_msg_2<1>:UD         r1.0<8,8,1>:UD {align1};
+
+/* m3. This is changed for FWD/BWD cost center */
+mov (8) vme_msg_3<1>:UD                0x0:UD {align1};                
+
+/* m4.*/
+mov (8) vme_msg_4<1>:ud                0x0:ud  {align1};
+
+/* m5 */
+mov  (1) INEP_ROW.0<1>:UD       0x0:UD {align1};
+and  (1) INEP_ROW.4<1>:UD       INEP_ROW.4<0,1,0>:UD            0xFF000000:UD {align1};
+mov  (8) vme_msg_5<1>:UD         INEP_ROW.0<8,8,1>:UD {align1};
+
+mov  (1) tmp_reg0.0<1>:UW      LUMA_CHROMA_MODE:UW {align1};
+/* Use the Luma mode */
+mov  (1) vme_msg_5.5<1>:UB     tmp_reg0.0<0,1,0>:UB {align1};
+
+/* m6 */        
+mov  (8) vme_msg_6<1>:UD         0x0:UD {align1};
+mov (16) vme_msg_6.0<1>:UB       INEP_COL0.3<32,8,4>:UB {align1};
+mov  (1) vme_msg_6.16<1>:UD      INTRA_PREDICTORE_MODE {align1};
+
+/* the penalty for Intra mode */
+mov  (1) vme_msg_6.28<1>:UD    0x010101:UD {align1};
+mov  (1) vme_msg_6.20<1>:UW      CHROMA_ROW.6<0,1,0>:UW {align1};
+
+
+/* m7 */
+
+mov  (4) vme_msg_7.16<1>:UD      CHROMA_ROW.8<4,4,1>:UD {align1};
+mov  (8) vme_msg_7.0<1>:UW       CHROMA_COL.2<16,8,2>:UW {align1};
+
+/*
+ * VME message
+ */
+
+/* m1 */
+mov  (1) intra_flag<1>:UW       0x0:UW {align1}                     ;
+and.z.f0.0 (1) null<1>:UW transform_8x8_ub<0,1,0>:UB 1:UW {align1};
+(f0.0) mov  (1) intra_part_mask_ub<1>:UB  LUMA_INTRA_8x8_DISABLE {align1};
+
+/* assign MB intra struct from the thread payload*/
+mov (1) mb_intra_struct_ub<1>:UB input_mb_intra_ub<0,1,0>:UB {align1}; 
+                           
+/* Disable DC HAAR component when calculating HARR SATD block */
+mov  (1) tmp_reg0.0<1>:UW      DC_HARR_DISABLE:UW              {align1};
+mov  (1) vme_m1.30<1>:UB       tmp_reg0.0<0,1,0>:UB  {align1};
+
+mov  (8) vme_msg_1<1>:UD        vme_m1.0<8,8,1>:UD {align1};
+/* m0 */        
+/* 16x16 Source, Intra_harr */
+add  (1) vme_m0.12<1>:UD        vme_m0.12<0,1,0>:ud    INTRA_SAD_HAAR:UD {align1};
+mov  (8) vme_msg_1<1>:UD        vme_m1.0<8,8,1>:UD {align1};
+
+/* after verification it will be passed by using payload */
+send (8)
+        vme_msg_ind
+        vme_wb<1>:UD
+        null
+        cre(
+                BIND_IDX_VME,
+                VME_SIC_MESSAGE_TYPE
+        )
+        mlen sic_vme_msg_length
+        rlen vme_wb_length
+        {align1};
+/*
+ * Oword Block Write message
+ */
+mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
+        
+mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
+mov  (1) msg_reg1.4<1>:UD       vme_wb.16<0,1,0>:UD     {align1};
+mov  (1) msg_reg1.8<1>:UD       vme_wb.20<0,1,0>:UD     {align1};
+mov  (1) msg_reg1.12<1>:UD      vme_wb.24<0,1,0>:UD     {align1};
+
+/* Distortion, Intra (17-16), */
+mov  (1) msg_reg1.16<1>:UW      vme_wb.12<0,1,0>:UW     {align1};
+
+mov  (1) msg_reg1.20<1>:UD      vme_wb.8<0,1,0>:UD     {align1};
+/* VME clock counts */
+mov  (1) msg_reg1.24<1>:UD      vme_wb.28<0,1,0>:UD     {align1};
+
+mov  (1) msg_reg1.28<1>:UD      obw_m0.8<0,1,0>:UD     {align1};
+
+/* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
+send (16)
+        msg_ind
+        obw_wb
+        null
+        data_port(
+                OBW_CACHE_TYPE,
+                OBW_MESSAGE_TYPE,
+                OBW_CONTROL_2,
+                OBW_BIND_IDX,
+                OBW_WRITE_COMMIT_CATEGORY,
+                OBW_HEADER_PRESENT
+        )
+        mlen 2
+        rlen obw_wb_length
+        {align1};
+
+__EXIT: 
+/*
+ * kill thread
+ */        
+mov  (8) ts_msg_reg0<1>:UD         r0<8,8,1>:UD {align1};
+send (16) ts_msg_ind acc0<1>UW null thread_spawner(0, 0, 1) mlen 1 rlen 0 {align1 EOT};
diff --git a/src/shaders/vme/intra_frame_gen8.g8a b/src/shaders/vme/intra_frame_gen8.g8a
new file mode 100644 (file)
index 0000000..859c72c
--- /dev/null
@@ -0,0 +1,2 @@
+#include "vme8.inc"
+#include "intra_frame_gen8.asm"
diff --git a/src/shaders/vme/intra_frame_gen8.g8b b/src/shaders/vme/intra_frame_gen8.g8b
new file mode 100644 (file)
index 0000000..43c904a
--- /dev/null
@@ -0,0 +1,72 @@
+   { 0x00800001, 0x24000608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24400608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24800608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24c00608, 0x00000000, 0x00000000 },
+   { 0x00200009, 0x24002228, 0x164500a0, 0x00040004 },
+   { 0x00000040, 0x24000a28, 0x1e000400, 0xfff8fff8 },
+   { 0x00000040, 0x24040a28, 0x1e000404, 0xffffffff },
+   { 0x00000001, 0x24080e08, 0x08000000, 0x0000001f },
+   { 0x00000001, 0x24142288, 0x00000014, 0x00000000 },
+   { 0x00200009, 0x24202228, 0x164500a0, 0x00040004 },
+   { 0x00000040, 0x24200a28, 0x1e000420, 0xfffcfffc },
+   { 0x00000001, 0x24280e08, 0x08000000, 0x000f0003 },
+   { 0x00000001, 0x24342288, 0x00000014, 0x00000000 },
+   { 0x00200009, 0x24482248, 0x164500a0, 0x00040004 },
+   { 0x00000001, 0x24542288, 0x00000014, 0x00000000 },
+   { 0x00000041, 0x24881208, 0x220000a2, 0x000000a1 },
+   { 0x00000040, 0x24880208, 0x22000488, 0x000000a0 },
+   { 0x00000041, 0x24880208, 0x06000488, 0x00000002 },
+   { 0x00000001, 0x24942288, 0x00000014, 0x00000000 },
+   { 0x00600001, 0x28000208, 0x008d0400, 0x00000000 },
+   { 0x04600031, 0x23800a88, 0x0e000800, 0x02190004 },
+   { 0x00600001, 0x28000208, 0x008d0420, 0x00000000 },
+   { 0x04600031, 0x23a00a88, 0x0e000800, 0x02290004 },
+   { 0x00200009, 0x24002228, 0x164500a0, 0x00030003 },
+   { 0x00000041, 0x24000a28, 0x1e000400, 0x00020002 },
+   { 0x00000040, 0x24000a28, 0x1e000400, 0xfff8fff8 },
+   { 0x00000040, 0x24040a28, 0x1e000404, 0xffffffff },
+   { 0x00600001, 0x28000208, 0x008d0400, 0x00000000 },
+   { 0x04600031, 0x26000a88, 0x0e000800, 0x02190006 },
+   { 0x00200009, 0x24202228, 0x164500a0, 0x00030003 },
+   { 0x00000041, 0x24200a28, 0x1e000420, 0x00020002 },
+   { 0x00000040, 0x24200a28, 0x1e000420, 0xfffcfffc },
+   { 0x00000001, 0x24280e08, 0x08000000, 0x00070003 },
+   { 0x00600001, 0x28000208, 0x008d0420, 0x00000000 },
+   { 0x04600031, 0x26200a88, 0x0e000800, 0x02190006 },
+   { 0x00600001, 0x28400208, 0x008d0020, 0x00000000 },
+   { 0x00600001, 0x28600608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x28800608, 0x00000000, 0x00000000 },
+   { 0x00000001, 0x23800608, 0x00000000, 0x00000000 },
+   { 0x00000005, 0x23840208, 0x06000384, 0xff000000 },
+   { 0x00600001, 0x28a00208, 0x008d0380, 0x00000000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00000000 },
+   { 0x00000001, 0x28a52288, 0x00000400, 0x00000000 },
+   { 0x00600001, 0x28c00608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x28c02288, 0x00cf03a3, 0x00000000 },
+   { 0x00000001, 0x28d00608, 0x00000000, 0x11111111 },
+   { 0x00000001, 0x28dc0608, 0x00000000, 0x00010101 },
+   { 0x00000001, 0x28d41248, 0x00000606, 0x00000000 },
+   { 0x00400001, 0x28f00208, 0x00690608, 0x00000000 },
+   { 0x00600001, 0x28e01248, 0x00ae0622, 0x00000000 },
+   { 0x00000001, 0x247c1648, 0x10000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a4, 0x00010001 },
+   { 0x00010001, 0x247c0e88, 0x08000000, 0x00000002 },
+   { 0x00000001, 0x247d2288, 0x000000a5, 0x00000000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00200020 },
+   { 0x00000001, 0x247e2288, 0x00000400, 0x00000000 },
+   { 0x00600001, 0x28200208, 0x008d0460, 0x00000000 },
+   { 0x00000040, 0x244c0208, 0x0600044c, 0x00800000 },
+   { 0x00600001, 0x28200208, 0x008d0460, 0x00000000 },
+   { 0x0d600031, 0x21800a08, 0x0e000800, 0x10782000 },
+   { 0x00600001, 0x28000208, 0x008d0480, 0x00000000 },
+   { 0x00000001, 0x28200208, 0x00000180, 0x00000000 },
+   { 0x00000001, 0x28240208, 0x00000190, 0x00000000 },
+   { 0x00000001, 0x28280208, 0x00000194, 0x00000000 },
+   { 0x00000001, 0x282c0208, 0x00000198, 0x00000000 },
+   { 0x00000001, 0x28301248, 0x0000018c, 0x00000000 },
+   { 0x00000001, 0x28340208, 0x00000188, 0x00000000 },
+   { 0x00000001, 0x28380208, 0x0000019c, 0x00000000 },
+   { 0x00000001, 0x283c0208, 0x00000488, 0x00000000 },
+   { 0x0a800031, 0x20000a60, 0x0e000800, 0x040a0203 },
+   { 0x00600001, 0x2e000208, 0x008d0000, 0x00000000 },
+   { 0x07800031, 0x24000a40, 0x0e000e00, 0x82000010 },
diff --git a/src/shaders/vme/mpeg2_inter_frame_gen8.g8a b/src/shaders/vme/mpeg2_inter_frame_gen8.g8a
new file mode 100644 (file)
index 0000000..9925c25
--- /dev/null
@@ -0,0 +1,3 @@
+#include "vme8.inc"
+#include "vme75_mpeg2.inc"
+#include "inter_frame_gen8.asm"
diff --git a/src/shaders/vme/mpeg2_inter_frame_gen8.g8b b/src/shaders/vme/mpeg2_inter_frame_gen8.g8b
new file mode 100644 (file)
index 0000000..7e52539
--- /dev/null
@@ -0,0 +1,300 @@
+   { 0x00800001, 0x24000608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24400608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24800608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24c00608, 0x00000000, 0x00000000 },
+   { 0x00200009, 0x24002228, 0x164500a0, 0x00040004 },
+   { 0x00000040, 0x24000a28, 0x1e000400, 0xfff8fff8 },
+   { 0x00000040, 0x24040a28, 0x1e000404, 0xffffffff },
+   { 0x00000001, 0x24080e08, 0x08000000, 0x0000001f },
+   { 0x00000001, 0x24142288, 0x00000014, 0x00000000 },
+   { 0x00200009, 0x24202228, 0x164500a0, 0x00040004 },
+   { 0x00000040, 0x24200a28, 0x1e000420, 0xfffcfffc },
+   { 0x00000001, 0x24280e08, 0x08000000, 0x000f0003 },
+   { 0x00000001, 0x24342288, 0x00000014, 0x00000000 },
+   { 0x00200009, 0x24482248, 0x164500a0, 0x00040004 },
+   { 0x00000001, 0x24542288, 0x00000014, 0x00000000 },
+   { 0x00000041, 0x24881208, 0x220000a2, 0x000000a1 },
+   { 0x00000040, 0x24880208, 0x22000488, 0x000000a0 },
+   { 0x00000041, 0x24880208, 0x06000488, 0x00000018 },
+   { 0x00000001, 0x24942288, 0x00000014, 0x00000000 },
+   { 0x00600001, 0x28000208, 0x008d0400, 0x00000000 },
+   { 0x04600031, 0x23800a88, 0x0e000800, 0x02190004 },
+   { 0x00600001, 0x28000208, 0x008d0420, 0x00000000 },
+   { 0x04600031, 0x23a00a88, 0x0e000800, 0x02290004 },
+   { 0x00200009, 0x24002228, 0x164500a0, 0x00030003 },
+   { 0x00000041, 0x24000a28, 0x1e000400, 0x00020002 },
+   { 0x00000040, 0x24000a28, 0x1e000400, 0xfff8fff8 },
+   { 0x00000040, 0x24040a28, 0x1e000404, 0xffffffff },
+   { 0x00600001, 0x28000208, 0x008d0400, 0x00000000 },
+   { 0x04600031, 0x26000a88, 0x0e000800, 0x02190006 },
+   { 0x00200009, 0x24202228, 0x164500a0, 0x00030003 },
+   { 0x00000041, 0x24200a28, 0x1e000420, 0x00020002 },
+   { 0x00000040, 0x24200a28, 0x1e000420, 0xfffcfffc },
+   { 0x00000001, 0x24280e08, 0x08000000, 0x00070003 },
+   { 0x00600001, 0x28000208, 0x008d0420, 0x00000000 },
+   { 0x04600031, 0x26200a88, 0x0e000800, 0x02190006 },
+   { 0x00600001, 0x2ac00608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2a800608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20001240, 0x160000a6, 0x00040004 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000750 },
+   { 0x00600001, 0x2ae00608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2b000608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2b200608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2b400608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00600060 },
+   { 0x00210001, 0x2af41e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000000f0 },
+   { 0x00000001, 0x2ae00e28, 0x08000000, 0x00000001 },
+   { 0x00200001, 0x24002248, 0x004500a0, 0x00000000 },
+   { 0x00000040, 0x24001a68, 0x1e000400, 0xffffffff },
+   { 0x00000041, 0x2b481208, 0x120000a2, 0x00000402 },
+   { 0x00000040, 0x2b480208, 0x12000b48, 0x00000400 },
+   { 0x00000041, 0x2b480208, 0x06000b48, 0x00000018 },
+   { 0x00000001, 0x2b542288, 0x00000014, 0x00000000 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x02280303 },
+   { 0x05000010, 0x20001260, 0x12000b70, 0x00000b88 },
+   { 0x00210001, 0x2af41e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000040 },
+   { 0x00000040, 0x2b480208, 0x06000b48, 0x00000003 },
+   { 0x0a800031, 0x2ba00a08, 0x0e000b40, 0x02480403 },
+   { 0x00200001, 0x2ae40208, 0x00450bc8, 0x00000000 },
+   { 0x00000001, 0x2af01e68, 0x18000000, 0x00010001 },
+   { 0x00600001, 0x2b400608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00100010 },
+   { 0x00210001, 0x2b141e68, 0x18000000, 0xffffffff },
+   { 0x00210001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000360 },
+   { 0x00000001, 0x2b000e28, 0x08000000, 0x00000001 },
+   { 0x00200001, 0x24002248, 0x004500a0, 0x00000000 },
+   { 0x00000040, 0x24021a68, 0x1e000402, 0xffffffff },
+   { 0x00000041, 0x2b481208, 0x120000a2, 0x00000402 },
+   { 0x00000040, 0x2b480208, 0x12000b48, 0x00000400 },
+   { 0x00000041, 0x2b480208, 0x06000b48, 0x00000018 },
+   { 0x00000001, 0x2b542288, 0x00000014, 0x00000000 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x02280303 },
+   { 0x05000010, 0x20001260, 0x12000b70, 0x00000b88 },
+   { 0x00210001, 0x2b141e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000040 },
+   { 0x00000040, 0x2b480208, 0x06000b48, 0x00000003 },
+   { 0x0a800031, 0x2ba00a08, 0x0e000b40, 0x02480403 },
+   { 0x00200001, 0x2b040208, 0x00450bf0, 0x00000000 },
+   { 0x00000001, 0x2b101e68, 0x18000000, 0x00010001 },
+   { 0x00600001, 0x2b400608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00080008 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000110 },
+   { 0x00000001, 0x2b200e28, 0x08000000, 0x00000001 },
+   { 0x00200001, 0x24002248, 0x004500a0, 0x00000000 },
+   { 0x00000040, 0x24021a68, 0x1e000402, 0xffffffff },
+   { 0x00000040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00000041, 0x2b481208, 0x120000a2, 0x00000402 },
+   { 0x00000040, 0x2b480208, 0x12000b48, 0x00000400 },
+   { 0x00000041, 0x2b480208, 0x06000b48, 0x00000018 },
+   { 0x00000001, 0x2b542288, 0x00000014, 0x00000000 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x02280303 },
+   { 0x05000010, 0x20001260, 0x12000b70, 0x00000b88 },
+   { 0x00210001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000180 },
+   { 0x00000040, 0x2b480208, 0x06000b48, 0x00000003 },
+   { 0x0a800031, 0x2ba00a08, 0x0e000b40, 0x02480403 },
+   { 0x00200001, 0x2b240208, 0x00450bf0, 0x00000000 },
+   { 0x00000001, 0x2b301e68, 0x18000000, 0x00010001 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000130 },
+   { 0x00600001, 0x2b400608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00040004 },
+   { 0x00210001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000000f0 },
+   { 0x00000001, 0x2b200e28, 0x08000000, 0x00000001 },
+   { 0x00200001, 0x24002248, 0x004500a0, 0x00000000 },
+   { 0x00200040, 0x24001a68, 0x1e450400, 0xffffffff },
+   { 0x00000041, 0x2b481208, 0x120000a2, 0x00000402 },
+   { 0x00000040, 0x2b480208, 0x12000b48, 0x00000400 },
+   { 0x00000041, 0x2b480208, 0x06000b48, 0x00000018 },
+   { 0x00000001, 0x2b542288, 0x00000014, 0x00000000 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x02280303 },
+   { 0x05000010, 0x20001260, 0x12000b70, 0x00000b88 },
+   { 0x00210001, 0x2b341e68, 0x18000000, 0xffffffff },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000040 },
+   { 0x00000040, 0x2b480208, 0x06000b48, 0x00000003 },
+   { 0x0a800031, 0x2ba00a88, 0x0e000b40, 0x02480403 },
+   { 0x00200001, 0x2b240208, 0x00450c18, 0x00000000 },
+   { 0x00000001, 0x2b301e68, 0x18000000, 0x00010001 },
+   { 0x00000040, 0x24000a28, 0x0a000b00, 0x00000b20 },
+   { 0x01000010, 0x20000a20, 0x0e000400, 0x00000000 },
+   { 0x00110020, 0x34000000, 0x0e001400, 0x00000080 },
+   { 0x02000010, 0x20000a20, 0x0e000ae0, 0x00000001 },
+   { 0x00010001, 0x2b040208, 0x00000ae4, 0x00000000 },
+   { 0x00010001, 0x2b240208, 0x00000ae4, 0x00000000 },
+   { 0x00010001, 0x2b141248, 0x00000af4, 0x00000000 },
+   { 0x00010001, 0x2b341248, 0x00000af4, 0x00000000 },
+   { 0x00010001, 0x2ac00208, 0x00000ae4, 0x00000000 },
+   { 0x00110001, 0x2ac00608, 0x00000000, 0x00000000 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000190 },
+   { 0x00600001, 0x24000608, 0x00000000, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000af4, 0x00000000 },
+   { 0x00010040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x24040208, 0x00000ae4, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000b14, 0x00000000 },
+   { 0x00010040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x24040208, 0x00000b04, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000b34, 0x00000000 },
+   { 0x00010040, 0x24001a68, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x24040208, 0x00000b24, 0x00000000 },
+   { 0x01000010, 0x20001a20, 0x1e000400, 0x00010001 },
+   { 0x00010001, 0x2ac00208, 0x00000404, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000000c0 },
+   { 0x00000001, 0x2fa01a68, 0x00000ae4, 0x00000000 },
+   { 0x00000001, 0x2fa41a68, 0x00000b04, 0x00000000 },
+   { 0x00000001, 0x2fa81a68, 0x00000b24, 0x00000000 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000870 },
+   { 0x00000001, 0x2ac01a68, 0x00000fe4, 0x00000000 },
+   { 0x00000001, 0x2fa01a68, 0x00000ae6, 0x00000000 },
+   { 0x00000001, 0x2fa41a68, 0x00000b06, 0x00000000 },
+   { 0x00000001, 0x2fa81a68, 0x00000b26, 0x00000000 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000810 },
+   { 0x00000001, 0x2ac21a68, 0x00000fe4, 0x00000000 },
+   { 0x0020000c, 0x2a801a68, 0x1e450ac0, 0x00020002 },
+   { 0x00200040, 0x2a881a68, 0x1e450a80, 0x00030003 },
+   { 0x00200005, 0x2a901248, 0x16450a88, 0xfffcfffc },
+   { 0x00600001, 0x25600208, 0x008d0020, 0x00000000 },
+   { 0x00600001, 0x28400208, 0x008d0560, 0x00000000 },
+   { 0x00600001, 0x28600608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x28800608, 0x00000000, 0x00000000 },
+   { 0x00000001, 0x23800608, 0x00000000, 0x00000000 },
+   { 0x00000005, 0x23840208, 0x06000384, 0xff000000 },
+   { 0x00600001, 0x28a00208, 0x008d0380, 0x00000000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00010001 },
+   { 0x00000001, 0x28a52288, 0x00000400, 0x00000000 },
+   { 0x00600001, 0x28c00608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x28c02288, 0x00cf03a3, 0x00000000 },
+   { 0x00000001, 0x28d00608, 0x00000000, 0x11111111 },
+   { 0x00000001, 0x28dc0608, 0x00000000, 0x00010101 },
+   { 0x00000001, 0x28d41248, 0x00000606, 0x00000000 },
+   { 0x00400001, 0x28f00208, 0x00690608, 0x00000000 },
+   { 0x00600001, 0x28e01248, 0x00ae0622, 0x00000000 },
+   { 0x00000001, 0x247c1648, 0x10000000, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a4, 0x00010001 },
+   { 0x00010001, 0x247c0e88, 0x08000000, 0x00000002 },
+   { 0x00000001, 0x247d2288, 0x000000a5, 0x00000000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00200020 },
+   { 0x00000001, 0x247e2288, 0x00000400, 0x00000000 },
+   { 0x00600001, 0x28200208, 0x008d0460, 0x00000000 },
+   { 0x00000001, 0x244c0608, 0x00000000, 0x00800000 },
+   { 0x00600001, 0x28000208, 0x008d0440, 0x00000000 },
+   { 0x0d600031, 0x21800a08, 0x0e000800, 0x10782000 },
+   { 0x00600001, 0x28000208, 0x008d0480, 0x00000000 },
+   { 0x00000001, 0x28200208, 0x00000180, 0x00000000 },
+   { 0x00000001, 0x28240208, 0x00000190, 0x00000000 },
+   { 0x00000001, 0x28280208, 0x00000194, 0x00000000 },
+   { 0x00000001, 0x282c0208, 0x00000198, 0x00000000 },
+   { 0x00000001, 0x28301248, 0x0000018c, 0x00000000 },
+   { 0x00000001, 0x28340208, 0x00000188, 0x00000000 },
+   { 0x00000001, 0x28380208, 0x0000019c, 0x00000000 },
+   { 0x00000001, 0x283c0208, 0x00000488, 0x00000000 },
+   { 0x0a800031, 0x20000a60, 0x0e000800, 0x040a0203 },
+   { 0x00000001, 0x244c0608, 0x00000000, 0x7e200000 },
+   { 0x00000001, 0x24561648, 0x10000000, 0x28302830 },
+   { 0x00000001, 0x24400208, 0x00000448, 0x00000000 },
+   { 0x00000040, 0x24401a68, 0x1e000440, 0xfff0fff0 },
+   { 0x00000040, 0x24421a68, 0x1e000442, 0xfff4fff4 },
+   { 0x00000001, 0x24401e68, 0x18000000, 0xfff0fff0 },
+   { 0x00000001, 0x24421e68, 0x18000000, 0xfff4fff4 },
+   { 0x00000001, 0x24440208, 0x00000440, 0x00000000 },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00600060 },
+   { 0x00010040, 0x24401a68, 0x1e000440, 0x000c000c },
+   { 0x01000005, 0x20002240, 0x160000a5, 0x00100010 },
+   { 0x00010040, 0x24421a68, 0x1e000442, 0x00080008 },
+   { 0x00200040, 0x24401a68, 0x1a450440, 0x00450a90 },
+   { 0x00200040, 0x24441a68, 0x1a450444, 0x00450a90 },
+   { 0x00600001, 0x28000208, 0x008d0440, 0x00000000 },
+   { 0x00000001, 0x24600608, 0x00000000, 0x00000002 },
+   { 0x00000001, 0x24642288, 0x0000009c, 0x00000000 },
+   { 0x00000001, 0x24680608, 0x00000000, 0x30003030 },
+   { 0x00600001, 0x28200208, 0x008d0460, 0x00000000 },
+   { 0x00400001, 0x45800208, 0x00000ac0, 0x00000000 },
+   { 0x00400001, 0x45840208, 0x00000ac0, 0x00000000 },
+   { 0x00600001, 0x28600208, 0x008d0580, 0x00000000 },
+   { 0x00600001, 0x28400208, 0x008d0560, 0x00000000 },
+   { 0x00000001, 0x28800608, 0x00000000, 0x01010101 },
+   { 0x00000001, 0x28840608, 0x00000000, 0x10010101 },
+   { 0x00000001, 0x28880608, 0x00000000, 0x0f0f0f0f },
+   { 0x00000001, 0x288c0608, 0x00000000, 0x100f0f0f },
+   { 0x00000001, 0x28900608, 0x00000000, 0x01010101 },
+   { 0x00000001, 0x28940608, 0x00000000, 0x10010101 },
+   { 0x00000001, 0x28980608, 0x00000000, 0x0f0f0f0f },
+   { 0x00000001, 0x289c0608, 0x00000000, 0x100f0f0f },
+   { 0x00000001, 0x28a00608, 0x00000000, 0x01010101 },
+   { 0x00000001, 0x28a40608, 0x00000000, 0x10010101 },
+   { 0x00000001, 0x28a80608, 0x00000000, 0x0f0f0f0f },
+   { 0x00000001, 0x28ac0608, 0x00000000, 0x000f0f0f },
+   { 0x00400001, 0x28b00608, 0x00000000, 0x00000000 },
+   { 0x08600031, 0x21800a08, 0x0e000800, 0x0c784000 },
+   { 0x00000001, 0x25740608, 0x00000000, 0x00000000 },
+   { 0x00000001, 0x25752288, 0x00000199, 0x00000000 },
+   { 0x00000001, 0x25762288, 0x0000019a, 0x00000000 },
+   { 0x00000005, 0x24001248, 0x16000180, 0x00030003 },
+   { 0x00000001, 0x25742288, 0x00000400, 0x00000000 },
+   { 0x00600001, 0x28800208, 0x008d01a0, 0x00000000 },
+   { 0x00600001, 0x28a00208, 0x008d01c0, 0x00000000 },
+   { 0x00600001, 0x28c00208, 0x008d01e0, 0x00000000 },
+   { 0x00600001, 0x28e00208, 0x008d0200, 0x00000000 },
+   { 0x00000001, 0x244c0608, 0x00000000, 0x00243000 },
+   { 0x00600001, 0x28000208, 0x008d0440, 0x00000000 },
+   { 0x00600001, 0x28200208, 0x008d0460, 0x00000000 },
+   { 0x00600001, 0x28400208, 0x008d0560, 0x00000000 },
+   { 0x00600001, 0x28600208, 0x008d0580, 0x00000000 },
+   { 0x0d600031, 0x21800a08, 0x0e000800, 0x10786000 },
+   { 0x00000040, 0x24880208, 0x06000488, 0x00000002 },
+   { 0x00600001, 0x28000208, 0x008d0480, 0x00000000 },
+   { 0x00000001, 0x28200208, 0x00000180, 0x00000000 },
+   { 0x00000001, 0x28240208, 0x00000198, 0x00000000 },
+   { 0x00000001, 0x28280208, 0x00000188, 0x00000000 },
+   { 0x00000001, 0x282c0208, 0x00000574, 0x00000000 },
+   { 0x0a800031, 0x20000a60, 0x0e000800, 0x040a0003 },
+   { 0x00000040, 0x24880208, 0x06000488, 0x00000001 },
+   { 0x00600001, 0x28000208, 0x008d0480, 0x00000000 },
+   { 0x00600001, 0x28200208, 0x008d01a0, 0x00000000 },
+   { 0x00600001, 0x28400208, 0x008d01c0, 0x00000000 },
+   { 0x00600001, 0x28600208, 0x008d01e0, 0x00000000 },
+   { 0x00600001, 0x28800208, 0x008d0200, 0x00000000 },
+   { 0x0a800031, 0x20000a60, 0x0e000800, 0x0a0a0403 },
+   { 0x00000040, 0x24880208, 0x06000488, 0x00000008 },
+   { 0x00600001, 0x28000208, 0x008d0480, 0x00000000 },
+   { 0x00600001, 0x28200208, 0x008d0240, 0x00000000 },
+   { 0x0a800031, 0x20000a60, 0x0e000800, 0x040a0203 },
+   { 0x0a800031, 0x2b600a08, 0x0e000b40, 0x0219e003 },
+   { 0x00600001, 0x2e000208, 0x008d0000, 0x00000000 },
+   { 0x07800031, 0x24000a40, 0x0e000e00, 0x82000010 },
+   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
+   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
+   { 0x06000010, 0x20001a60, 0x1a000fa0, 0x00000fa4 },
+   { 0x00010001, 0x2f601a68, 0x00000fa0, 0x00000000 },
+   { 0x00110001, 0x2f601a68, 0x00000fa4, 0x00000000 },
+   { 0x06000010, 0x20001a60, 0x1a000f60, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000f60, 0x00000000 },
+   { 0x00110001, 0x2fe41a68, 0x00000fa8, 0x00000000 },
+   { 0x00000001, 0x34000200, 0x00000fe0, 0x00000000 },
+   { 0x04000010, 0x20001a60, 0x1a000fa0, 0x00000fa4 },
+   { 0x00010001, 0x2f601a68, 0x00000fa0, 0x00000000 },
+   { 0x00110001, 0x2f601a68, 0x00000fa4, 0x00000000 },
+   { 0x04000010, 0x20001a60, 0x1a000f60, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000f60, 0x00000000 },
+   { 0x00110001, 0x2fe41a68, 0x00000fa8, 0x00000000 },
+   { 0x00000001, 0x34000200, 0x00000fe0, 0x00000000 },
+   { 0x04000010, 0x20001a60, 0x1a000fa0, 0x00000fa4 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000070 },
+   { 0x04000010, 0x20001a60, 0x1a000fa0, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000fa0, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x000000a0 },
+   { 0x04000010, 0x20001a60, 0x1a000fa4, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000fa8, 0x00000000 },
+   { 0x00110001, 0x2fe41a68, 0x00000fa4, 0x00000000 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000060 },
+   { 0x04000010, 0x20001a60, 0x1a000fa4, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000fa4, 0x00000000 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000030 },
+   { 0x04000010, 0x20001a60, 0x1a000fa0, 0x00000fa8 },
+   { 0x00010001, 0x2fe41a68, 0x00000fa8, 0x00000000 },
+   { 0x00110001, 0x2fe41a68, 0x00000fa0, 0x00000000 },
+   { 0x00000001, 0x34000200, 0x00000fe0, 0x00000000 },
diff --git a/src/shaders/vme/vme8.inc b/src/shaders/vme/vme8.inc
new file mode 100644 (file)
index 0000000..bf60ad5
--- /dev/null
@@ -0,0 +1,341 @@
+/*
+ * Copyright © <2010>, Intel Corporation.
+ *
+ * This program is licensed under the terms and conditions of the
+ * Eclipse Public License (EPL), version 1.0.  The full text of the EPL is at
+ * http://www.opensource.org/licenses/eclipse-1.0.php.
+ *
+ */
+// Modual name: ME_header.inc for Gen8
+//
+// Global symbols define
+//
+
+/*
+ * Constant
+ */
+define(`VME_MESSAGE_TYPE_INTER',        `1')
+define(`VME_MESSAGE_TYPE_INTRA',        `2')
+define(`VME_MESSAGE_TYPE_MIXED',        `3')
+        
+define(`VME_SIC_MESSAGE_TYPE',        `1')
+define(`VME_IME_MESSAGE_TYPE',        `2')
+define(`VME_FBR_MESSAGE_TYPE',        `3')
+
+define(`BLOCK_32X1',                    `0x0000001F')
+define(`BLOCK_4X16',                    `0x000F0003')
+define(`BLOCK_8X4',                     `0x00070003')
+        
+define(`LUMA_INTRA_16x16_DISABLE',      `0x1')
+define(`LUMA_INTRA_8x8_DISABLE',        `0x2')
+define(`LUMA_INTRA_4x4_DISABLE',        `0x4')
+
+define(`INTRA_PRED_AVAIL_FLAG_AE',      `0x60')
+define(`INTRA_PRED_AVAIL_FLAG_B',       `0x10')
+define(`INTRA_PRED_AVAIL_FLAG_C',       `0x8')
+define(`INTRA_PRED_AVAIL_FLAG_D',       `0x4')
+
+define(`BIND_IDX_VME',                  `0')
+define(`BIND_IDX_VME_REF0',             `1')
+define(`BIND_IDX_VME_REF1',             `2')
+define(`BIND_IDX_OUTPUT',               `3')
+define(`BIND_IDX_INEP',                 `4')
+
+define(`SUB_PEL_MODE_INTEGER',          `0x00000000')
+define(`SUB_PEL_MODE_HALF',             `0x00001000')
+define(`SUB_PEL_MODE_QUARTER',          `0x00003000')
+
+define(`INTER_SAD_NONE',                `0x00000000')
+define(`INTER_SAD_HAAR',                `0x00200000')
+
+define(`INTRA_SAD_NONE',                `0x00000000')
+define(`INTRA_SAD_HAAR',                `0x00800000')
+
+define(`INTER_PART_MASK',               `0x00000000')
+
+define(`SEARCH_CTRL_SINGLE',            `0x00000000')
+define(`SEARCH_CTRL_DUAL_START',        `0x00000100')
+define(`SEARCH_CTRL_DUAL_RECORD',       `0x00000300')
+define(`SEARCH_CTRL_DUAL_REFERENCE',    `0x00000700')
+
+define(`REF_REGION_SIZE',               `0x2830:UW')
+
+define(`BI_SUB_MB_PART_MASK',           `0x0c000000')
+define(`MAX_NUM_MV',                    `0x00000020')
+define(`FB_PRUNING_ENABLE',             `0x40000000')
+
+define(`SEARCH_PATH_LEN',               `0x00003030')
+define(`START_CENTER',                  `0x30000000')
+
+define(`ADAPTIVE_SEARCH_ENABLE',        `0x00000002') 
+define(`INTRA_PREDICTORE_MODE',         `0x11111111:UD')
+
+define(`INTER_VME_OUTPUT_IN_OWS',       `10')
+define(`INTER_VME_OUTPUT_MV_IN_OWS',    `8')
+
+define(`INTRAMBFLAG_MASK',              `0x00002000')
+define(`MVSIZE_UW_BASE',                `0x0040')
+define(`MFC_MV32_BIT_SHIFT',            `5')
+define(`CBP_DC_YUV_UW',                 `0x000E')
+
+define(`DC_HARR_ENABLE',                `0x0000')
+define(`DC_HARR_DISABLE',              `0x0020')
+
+define(`MV32_BIT_MASK',                 `0x0020')
+define(`MV32_BIT_SHIFT',                `5')
+
+define(`OBW_CACHE_TYPE',                `10')
+
+
+define(`OBW_MESSAGE_TYPE',              `8')
+
+define(`OBW_BIND_IDX',                  `BIND_IDX_OUTPUT')
+
+define(`OBW_CONTROL_0',                 `0')    /* 1 OWord, low 128 bits */
+define(`OBW_CONTROL_1',                 `1')    /* 1 OWord, high 128 bits */
+define(`OBW_CONTROL_2',                 `2')    /* 2 OWords */
+define(`OBW_CONTROL_3',                 `3')    /* 4 OWords */
+define(`OBW_CONTROL_8',                 `4')    /* 8 OWords */
+
+define(`FBR_BME_ENABLE',                 `0x00000000')
+define(`FBR_BME_DISABLE',                `0x00040000')
+
+define(`OBW_WRITE_COMMIT_CATEGORY',     `0')    /* category on Ivybridge */
+
+
+define(`OBW_HEADER_PRESENT',            `1')
+
+/* GRF registers
+ * r0 header
+ * r1~r4 constant buffer (reserved)
+ * r5 inline data
+ * r6~r11 reserved        
+ * r12 write back of VME message
+ * r13 write back of Oword Block Write        
+ */
+/*
+ * GRF 0 -- header       
+ */        
+define(`thread_id_ub',          `r0.20<0,1,0>:UB')  /* thread id in payload */
+
+/*
+ * GRF 1~4 -- Constant Buffer (reserved)
+ */
+        
+/*
+ * GRF 5 -- inline data
+ */        
+define(`inline_reg0',           `r5')
+define(`w_in_mb_uw',            `inline_reg0.2')
+define(`orig_xy_ub',            `inline_reg0.0')
+define(`orig_x_ub',             `inline_reg0.0')    /* in macroblock */    
+define(`orig_y_ub',             `inline_reg0.1')
+define(`transform_8x8_ub',      `inline_reg0.4')
+define(`input_mb_intra_ub',     `inline_reg0.5')
+define(`num_macroblocks',       `inline_reg0.6')
+
+/*
+ * GRF 6~11 -- reserved
+ */
+
+/*
+ * GRF 12~15 -- write back for VME message 
+ */
+define(`vme_wb',                `r12')
+define(`vme_wb0',               `r12')
+define(`vme_wb1',               `r13')
+define(`vme_wb2',               `r14')
+define(`vme_wb3',               `r15')
+define(`vme_wb4',               `r16')
+define(`vme_wb5',               `r17')
+define(`vme_wb6',               `r18')
+define(`vme_ime_wb7',          `r19')
+define(`vme_ime_wb8',          `r20')
+define(`vme_ime_wb9',          `r21')
+define(`vme_ime_wb10',         `r22')
+
+
+/*
+ * GRF 24 -- write for VME output message
+ */
+define(`obw_wb',                `null<1>:W')
+define(`obw_wb_length',         `0')
+
+
+/*
+ * GRF 28~30 -- Intra Neighbor Edge Pixels
+ */
+define(`INEP_ROW',              `r28')
+define(`INEP_COL0',             `r29')
+define(`INEP_COL1',             `r30')
+        
+/*
+ * GRF 48~50 -- Chroma Neighbor Edge Pixels
+ */
+define(`CHROMA_ROW',              `r48')
+define(`CHROMA_COL',              `r49')
+
+/*
+ * temporary registers
+ */
+define(`tmp_reg0',              `r32')
+define(`read0_header',          `tmp_reg0')
+define(`tmp_reg1',              `r33')
+define(`read1_header',          `tmp_reg1')
+define(`tmp_reg2',              `r34')
+define(`vme_m0',                `tmp_reg2')
+define(`tmp_reg3',              `r35')                                
+define(`vme_m1',                `tmp_reg3')
+define(`intra_flag',            `vme_m1.28')
+define(`intra_part_mask_ub',    `vme_m1.28')        
+define(`mb_intra_struct_ub',    `vme_m1.29')
+define(`tmp_reg4',              `r36')
+define(`obw_m0',                `tmp_reg4')
+define(`tmp_reg5',              `r37')
+define(`obw_m1',                `tmp_reg5')
+define(`tmp_reg6',              `r38')
+define(`obw_m2',                `tmp_reg6')
+define(`tmp_reg7',              `r39')
+define(`obw_m3',                `tmp_reg7')
+define(`tmp_reg8',              `r40')
+define(`obw_m4',                `tmp_reg8')
+define(`tmp_reg9',              `r41')
+define(`tmp_x_w',               `tmp_reg9.0')
+define(`tmp_rega',              `r42')
+define(`tmp_ud0',               `tmp_rega.0')
+define(`tmp_ud1',               `tmp_rega.4')
+define(`tmp_ud2',               `tmp_rega.8')
+define(`tmp_ud3',               `tmp_rega.12')
+define(`tmp_uw0',               `tmp_rega.0')
+define(`tmp_uw1',               `tmp_rega.2')
+define(`tmp_uw2',               `tmp_rega.4')
+define(`tmp_uw3',               `tmp_rega.6')
+define(`tmp_uw4',               `tmp_rega.8')
+define(`tmp_uw5',               `tmp_rega.10')
+define(`tmp_uw6',               `tmp_rega.12')
+define(`tmp_uw7',               `tmp_rega.14')
+
+define(`vme_m2',                `r43')
+define(`vme_m3',                `r44')
+/*
+ * MRF registers
+ */        
+
+define(`msg_ind',               `64')
+define(`msg_reg0',              `r64')
+define(`msg_reg1',              `r65')
+define(`msg_reg2',              `r66')
+define(`msg_reg3',              `r67')
+define(`msg_reg4',              `r68')
+define(`msg_reg5',              `r69')
+define(`msg_reg6',              `r70')
+define(`msg_reg7',              `r71')
+define(`msg_reg8',              `r72')
+define(`msg_reg9',              `r73')
+
+define(`ts_msg_ind',               `112')
+define(`ts_msg_reg0',               `r112')
+/*
+ * VME message payload
+ */
+
+define(`vme_intra_wb_length',   `1')
+define(`vme_wb_length',                `7')
+define(`sic_vme_msg_length',   `8')
+define(`fbr_vme_msg_length',   `8')
+define(`ime_vme_msg_length',   `6')
+
+define(`vme_msg_ind',           `msg_ind')
+define(`vme_msg_0',             `msg_reg0')
+define(`vme_msg_1',             `msg_reg1')
+define(`vme_msg_2',             `msg_reg2')
+
+define(`vme_msg_3',             `msg_reg3')
+define(`vme_msg_4',             `msg_reg4')
+
+
+define(`vme_msg_5',             `msg_reg5')
+define(`vme_msg_6',             `msg_reg6')
+define(`vme_msg_7',             `msg_reg7')
+define(`vme_msg_8',             `msg_reg8')
+define(`vme_msg_9',             `msg_reg9')
+
+define(`BIND_IDX_CBCR',                        `6')
+
+
+define(`LUMA_CHROMA_MODE',      `0x0')
+define(`LUMA_INTRA_MODE',      `0x1')
+define(`LUMA_INTRA_DISABLE',   `0x2')
+
+define(`RETURN_REG',              `r127.0')
+define(`RET_ARG',              `r127.4')
+
+/* Now at most two registers are used for input parameter */
+define(`INPUT_ARG0',              `r125')
+define(`INPUT_ARG1',              `r126')
+
+/* Two temporal registers are used in the function */
+define(`TEMP_VAR0',              `r123')
+define(`TEMP_VAR1',              `r124')
+
+
+define(`OBR_MESSAGE_TYPE',              `0')
+define(`OBR_CACHE_TYPE',                `10')
+define(`OBR_BIND_IDX',                  `BIND_IDX_OUTPUT')
+
+define(`OBR_CONTROL_0',                 `0')    /* 1 OWord, low 128 bits */
+define(`OBR_CONTROL_1',                 `1')    /* 1 OWord, high 128 bits */
+define(`OBR_CONTROL_2',                 `2')    /* 2 OWords */
+define(`OBR_CONTROL_4',                 `3')    /* 4 OWords */
+define(`OBR_CONTROL_8',                 `4')    /* 8 OWords */
+define(`OBR_WRITE_COMMIT_CATEGORY',     `0')    /* category on SNB+ for Data port */
+define(`OBR_HEADER_PRESENT',            `1')
+
+define(`mb_hwdep',           `r5.6')
+define(`MB_AVAIL',              `1:d')
+define(`MB_PRED_FLAG',          `1:w')
+
+define(`mb_pred_mode',         `r85')
+define(`mb_mvp_ref',           `r86')
+define(`mba_result',           `r87')
+define(`mbb_result',           `r88')
+define(`mbc_result',           `r89')
+define(`mb_ind',                `90')
+define(`mb_msg0',              `r90')
+define(`mb_wb',                        `r91')
+define(`mb_intra_wb',          `r91')
+define(`mb_inter_wb',          `r92')
+define(`mb_mv0',               `r93')
+define(`mb_mv1',               `r94')
+define(`mb_mv2',               `r95')
+define(`mb_mv3',               `r96')
+define(`mb_ref',               `r97')
+define(`mb_ref_win',           `r84')
+
+define(`DREF_REGION_SIZE',               `0x2020:UW')
+define(`PRED_L0',               `0x0':uw)
+define(`PRED_L1',               `0x1':uw)
+define(`PRED_BI',               `0x2':uw)
+define(`PRED_DIRECT',           `0x3':uw)
+define(`PRED_MASK',             `0x3':uw)
+
+/* The MAX search len per reference is 16 */
+define(`DSEARCH_PATH_LEN',               `0x00001212')
+define(`BI_WEIGHT',             `0x20':uw)
+define(`DSTART_CENTER',                  `0x00000000')
+define(`INTER_MASK',                   `0x03')
+define(`INTER_16X16MODE',              `0x0')
+define(`INTER_16X8MODE',               `0x01')
+define(`INTER_8X16MODE',               `0x02')
+define(`INTER_8X8MODE',                        `0x03')
+define(`INTER_BLOCK0',                 `0x0')
+define(`INTER_BLOCK1',                 `0x1')
+define(`INTER_BLOCK2',                 `0x2')
+define(`INTER_BLOCK3',                 `0x3')
+define(`INTER_16X8MODE',               `0x01')
+define(`INTER_8X16MODE',               `0x02')
+
+
+define(`OBR_MESSAGE_FENCE',              `7')
+define(`OBR_MF_NOCOMMIT',               `0')
+define(`OBR_MF_COMMIT',                         `0x20')
diff --git a/src/shaders/vme/vme8_mpeg2.inc b/src/shaders/vme/vme8_mpeg2.inc
new file mode 100644 (file)
index 0000000..9b877ac
--- /dev/null
@@ -0,0 +1,18 @@
+/*
+ * Copyright © <2010>, Intel Corporation.
+ *
+ * This program is licensed under the terms and conditions of the
+ * Eclipse Public License (EPL), version 1.0.  The full text of the EPL is at
+ * http://www.opensource.org/licenses/eclipse-1.0.php.
+ *
+ */
+// Modual name: ME_header.inc
+//
+// Global symbols define
+//
+
+/*
+ * Constant
+ */
+
+define(`INTER_PART_MASK',               `0x7e000000')