Rewrite the GPU VME shader for MPEG2 encoding on Gen8
authorZhao Yakui <yakui.zhao@intel.com>
Thu, 15 Aug 2013 07:18:39 +0000 (15:18 +0800)
committerXiang, Haihao <haihao.xiang@intel.com>
Thu, 27 Feb 2014 02:22:18 +0000 (10:22 +0800)
This is from that on Haswell/Ivybridge. Now the MPEG2/H264 uses the same
mode/motion vector prediction shader. But the MV search region of mpeg2
is different with that on H264, which causes that the wrong mode/motion
vector prediction is used for MPEG2.

Signed-off-by: Zhao Yakui <yakui.zhao@intel.com>
src/gen8_vme.c
src/shaders/vme/Makefile.am
src/shaders/vme/mpeg2_inter_gen8.asm [new file with mode: 0644]
src/shaders/vme/mpeg2_inter_gen8.g8a [new file with mode: 0644]
src/shaders/vme/mpeg2_inter_gen8.g8b [new file with mode: 0644]

index 4a631ca..dc325df 100644 (file)
@@ -100,7 +100,7 @@ static const uint32_t gen8_vme_mpeg2_intra_frame[][4] = {
 };
 
 static const uint32_t gen8_vme_mpeg2_inter_frame[][4] = {
-#include "shaders/vme/mpeg2_inter_frame_gen8.g8b"
+#include "shaders/vme/mpeg2_inter_gen8.g8b"
 };
 
 static struct i965_kernel gen8_vme_mpeg2_kernels[] = {
@@ -827,7 +827,6 @@ gen8_vme_mpeg2_fill_vme_batchbuffer(VADriverContextP ctx,
             int slice_mb_begin = slice_param->macroblock_address;
             int slice_mb_number = slice_param->num_macroblocks;
             unsigned int mb_intra_ub;
-            int slice_mb_x = slice_param->macroblock_address % mb_width;
 
             for (i = 0; i < slice_mb_number;) {
                 int mb_count = i + slice_mb_begin;    
@@ -850,21 +849,6 @@ gen8_vme_mpeg2_fill_vme_batchbuffer(VADriverContextP ctx,
                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
                 }
 
-                if (i < mb_width) {
-                    if (i == 0)
-                        mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_AE);
-
-                    mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_BCD_MASK);
-
-                    if ((i == (mb_width - 1)) && slice_mb_x) {
-                        mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
-                    }
-                }
-               
-                if ((i == mb_width) && slice_mb_x) {
-                    mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_D);
-                }
-
                 *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
                 *command_ptr++ = kernel;
                 *command_ptr++ = 0;
@@ -928,11 +912,20 @@ gen8_vme_mpeg2_prepare(VADriverContextP ctx,
 {
     VAStatus vaStatus = VA_STATUS_SUCCESS;
     VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[0]->buffer;
+    VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
+    struct gen6_vme_context *vme_context = encoder_context->vme_context;
+
+    if ((!vme_context->mpeg2_level) ||
+               (vme_context->mpeg2_level != (seq_param->sequence_extension.bits.profile_and_level_indication & MPEG2_LEVEL_MASK))) {
+       vme_context->mpeg2_level = seq_param->sequence_extension.bits.profile_and_level_indication & MPEG2_LEVEL_MASK;
+    }
+
        
     /*Setup all the memory object*/
     gen8_vme_mpeg2_surface_setup(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
     gen8_vme_interface_setup(ctx, encode_state, encoder_context);
-    gen8_vme_vme_state_setup(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
+    //gen8_vme_vme_state_setup(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
+    intel_vme_mpeg2_state_setup(ctx, encode_state, encoder_context);
     gen8_vme_constant_setup(ctx, encode_state, encoder_context);
 
     /*Programing media pipeline*/
index d84f795..29e70b8 100644 (file)
@@ -19,8 +19,8 @@ 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_G8B      = intra_frame_gen8.g8b inter_frame_gen8.g8b mpeg2_inter_frame_gen8.g8b inter_bframe_gen8.g8b mpeg2_inter_gen8.g8b
+INTEL_G8A      = intra_frame_gen8.g8a inter_frame_gen8.g8a mpeg2_inter_frame_gen8.g8a inter_bframe_gen8.g8a mpeg2_inter_gen8.g8a
 INTEL_GEN8_INC = vme8.inc vme8_mpeg2.inc
 INTEL_GEN8_ASM = $(INTEL_G8A:%.g8a=%.gen8.asm)
 
diff --git a/src/shaders/vme/mpeg2_inter_gen8.asm b/src/shaders/vme/mpeg2_inter_gen8.asm
new file mode 100644 (file)
index 0000000..aea2cc6
--- /dev/null
@@ -0,0 +1,722 @@
+/*
+ * 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: mpeg2_inter_gen8.asm
+//
+// Make inter predition estimation for MPEG2 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) 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 */
+
+        
+shl    (2) pic_ref.0<1>:uw     r4.24<2,2,1>:uw 4:uw    {align1};
+mov    (2) pic_ref.16<1>:uw    r4.20<2,2,1>:uw {align1};
+mov  (8) mb_mvp_ref.0<1>:ud    0:ud            {align1};
+mov  (8) mb_ref_win.0<1>:ud    0:ud            {align1};
+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};
+
+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)
+*/
+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 2 oword (32 bytes), 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_2,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 1
+        {align1};
+/* TODO: RefID is required after multi-references are added */
+/* MV */
+mov       (2)          mba_result.4<1>:ud              mb_mv0.0<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 2 oword (32bytes), 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_2,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 1
+        {align1};
+/* TODO: RefID is required after multi-references are added */
+mov       (2)          mbb_result.4<1>:ud              mb_mv0.0<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 2 oword (32bytes), 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_2,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 1
+        {align1};
+/* TODO: RefID is required after multi-references are added */
+/* Forward MV */
+mov       (2)          mbc_result.4<1>:ud              mb_mv0.0<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 2 oword (32bytes), 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_2,
+                OBR_BIND_IDX,
+                OBR_WRITE_COMMIT_CATEGORY,
+                OBR_HEADER_PRESENT
+        )
+        mlen 1
+        rlen 1
+        {align1};
+
+/* TODO: RefID is required after multi-references are added */
+
+/* Forward MV */
+mov       (2)          mbc_result.4<1>:ud              mb_mv0.0<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:
+
+/* Calibrate the ref window for MPEG2 */
+mov  (1) vme_m0.0<1>:W         -16:W                   {align1};
+mov  (1) vme_m0.2<1>:W         -12:W                   {align1};
+
+mov  (1) INPUT_ARG0.0<1>:ud    vme_m0.0<0,1,0>:ud      {align1};
+mov  (1) INPUT_ARG0.8<1>:ud    vme_m0.8<0,1,0>:ud      {align1};
+mov  (8) INPUT_ARG1.0<1>:ud    pic_ref.0<8,8,1>:ud     {align1};
+
+SAVE_RET       {align1};
+jmpi   (1)     ref_boundary_check;
+mov  (2) vme_m0.0<1>:w         RET_ARG<2,2,1>:w        {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  (8) vme_msg_5<1>:UD        0x0: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};
+
+mov  (1) tmp_reg0.0<1>:UW      INTRA16_DC_PRED:UW {align1};
+mov  (1) vme_msg_5.4<1>:ub     tmp_reg0.0<0,1,0>:UB {align1};
+
+/* m6 */        
+mov  (8) vme_msg_6<1>:UD         0x0:UD {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};
+
+
+/* m7 */
+
+mov  (8) vme_msg_7.0<1>:ud       0x0:ud {align1};
+
+/*
+ * SIC VME message
+ */
+
+/* Disable Intra8x8/Intra4x4 Intra-prediction */
+/* m1 */
+mov  (8) vme_m1.0<1>:ud                0x0:UD  {align1};
+
+mov  (1) intra_flag<1>:UW       0x0:UW {align1}                     ;
+mov   (1) tmp_reg0.0<1>:uw     LUMA_INTRA_8x8_DISABLE:uw               {align1};
+add   (1) tmp_reg0.0<1>:uw     tmp_reg0.0<0,1,0>:uw    LUMA_INTRA_4x4_DISABLE:uw {align1};
+mov  (1) intra_part_mask_ub<1>:UB  tmp_reg0.0<0,1,0>:ub {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}; 
+
+/* Enable DC HAAR component when calculating HARR SATD block */
+mov  (1) tmp_reg0.0<1>:UW      DC_HARR_ENABLE: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_m0.4<1>:UD                vme_m0.0<0,1,0>:UD      {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                0x0:ud  {align1};
+mov  (4) vme_m3.4<2>:ud                0x0: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_HALF + FBR_BME_DISABLE:UD {align1};    /* 16x16 Source, 1/2 pixel, harr, BME disable */
+/* Bilinear filter */
+mov  (1) tmp_reg0.0<1>:uw      0x04:uw {align1};
+add  (1) vme_m1.30<1>:ub       vme_m1.30<0,1,0>:ub     tmp_reg0.0<0,1,0>:ub    {align1};
+
+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  2 oword (32 bytes), 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};
+
+/* 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             ;
+       
+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};
+
+nop;
+nop;
+ref_boundary_check:
+
+/* The left/up coordinate of reference window */
+add  (2) TEMP_VAR0.0<1>:w      INPUT_ARG0.8<2,2,1>:w   INPUT_ARG0.0<2,2,1>:w   {align1};
+/* The right/bottom coordinate of reference window */
+add  (1) TEMP_VAR0.16<1>:w     TEMP_VAR0.0<0,1,0>:w    48:w                    {align1};
+add  (1) TEMP_VAR0.18<1>:w     TEMP_VAR0.2<0,1,0>:w    40:w                    {align1};
+
+/* Firstly the MV range is checked */
+mul  (2) TEMP_VAR1.16<1>:w     INPUT_ARG1.16<2,2,1>:w  -1:w    {align1};
+add  (2) TEMP_VAR1.0<1>:w      INPUT_ARG0.8<2,2,1>:w   TEMP_VAR1.16<2,2,1>:w   {align1};
+add  (2) TEMP_VAR1.4<1>:w      INPUT_ARG0.8<2,2,1>:w   INPUT_ARG1.16<2,2,1>:w  {align1};
+
+cmp.l.f0.0 (1) null:w  TEMP_VAR0.0<0,1,0>:w    TEMP_VAR1.0<0,1,0>:w    {align1};
+(f0.0) mov     (1) TEMP_VAR0.0<1>:w    TEMP_VAR1.0<0,1,0>:w    {align1};
+cmp.g.f0.0 (1) null:w  TEMP_VAR0.16<0,1,0>:w   TEMP_VAR1.4<0,1,0>:w    {align1};
+(f0.0) add     (1) TEMP_VAR0.0<1>:w    TEMP_VAR1.4<0,1,0>:w    -48:w   {align1};
+cmp.l.f0.0 (1) null:w  TEMP_VAR0.2<0,1,0>:w    TEMP_VAR1.2<0,1,0>:w    {align1};
+(f0.0) mov     (1) TEMP_VAR0.2<1>:w    TEMP_VAR1.2<0,1,0>:w    {align1};
+cmp.g.f0.0 (1) null:w  TEMP_VAR0.18<0,1,0>:w   TEMP_VAR1.6<0,1,0>:w    {align1};
+(f0.0) add     (1) TEMP_VAR0.2<1>:w    TEMP_VAR1.6<0,1,0>:w    -40:w   {align1};
+
+x_left_cmp:
+       cmp.l.f0.0 (1)          null:w          TEMP_VAR0.0<0,1,0>:w    0:w     {align1};
+       (-f0.0) jmpi    (1)     x_right_cmp;
+       (f0.0)  mov     (1)     TEMP_VAR0.0<1>:w        0:w             {align1};
+       jmpi    (1)     y_top_cmp;
+x_right_cmp:
+       cmp.g.f0.0 (1)  null:w  TEMP_VAR0.16<0,1,0>:w   INPUT_ARG1.0<0,1,0>:w   {align1};
+       (-f0.0) jmpi    (1)     y_top_cmp;      
+       (f0.0)  add     (1)     TEMP_VAR0.0<1>:w        INPUT_ARG1.0<0,1,0>:w   -48:w   {align1};
+y_top_cmp:
+       cmp.l.f0.0  (1)         null:w  TEMP_VAR0.2<0,1,0>:w    0:w             {align1};
+       (-f0.0) jmpi    (1)     y_bottom_cmp;
+       (f0.0)  mov     (1)     TEMP_VAR0.2<1>:w        0:w             {align1};
+       jmpi    (1)     y_bottom_end;
+y_bottom_cmp:
+       cmp.g.f0.0  (1)         null:w  TEMP_VAR0.18<0,1,0>:w   INPUT_ARG1.2<0,1,0>:w {align1};
+       (f0.0)  add     (1)     TEMP_VAR0.2<1>:w                INPUT_ARG1.2<0,1,0>:w   -40:w   {align1};
+
+y_bottom_end:
+mul    (2)     TEMP_VAR1.0<1>:w        INPUT_ARG0.8<2,2,1>:w   -1:w    {align1};
+add    (2)     RET_ARG<1>:w            TEMP_VAR0.0<2,2,1>:w    TEMP_VAR1.0<2,2,1>:w    {align1};       
+       RETURN  {align1};
+nop;
+nop;
+
diff --git a/src/shaders/vme/mpeg2_inter_gen8.g8a b/src/shaders/vme/mpeg2_inter_gen8.g8a
new file mode 100644 (file)
index 0000000..26f94a7
--- /dev/null
@@ -0,0 +1,3 @@
+#include "vme8.inc"
+#include "vme75_mpeg2.inc"
+#include "mpeg2_inter_gen8.asm"
diff --git a/src/shaders/vme/mpeg2_inter_gen8.g8b b/src/shaders/vme/mpeg2_inter_gen8.g8b
new file mode 100644 (file)
index 0000000..8bdabf0
--- /dev/null
@@ -0,0 +1,292 @@
+   { 0x00800001, 0x24000608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24400608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24800608, 0x00000000, 0x00000000 },
+   { 0x00800001, 0x24c00608, 0x00000000, 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 },
+   { 0x00200009, 0x2a401248, 0x16450098, 0x00040004 },
+   { 0x00200001, 0x2a501248, 0x00450094, 0x00000000 },
+   { 0x00600001, 0x2ac00608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2a800608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2ae00608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2b000608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2b200608, 0x00000000, 0x00000000 },
+   { 0x01000005, 0x20001240, 0x160000a6, 0x00040004 },
+   { 0x00010020, 0x34000000, 0x0e001400, 0x00000720 },
+   { 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, 0x02180203 },
+   { 0x00200001, 0x2ae40208, 0x00450ba0, 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, 0x02180203 },
+   { 0x00200001, 0x2b040208, 0x00450ba0, 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, 0x02180203 },
+   { 0x00200001, 0x2b240208, 0x00450ba0, 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, 0x02180203 },
+   { 0x00200001, 0x2b240208, 0x00450ba0, 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, 0x00000740 },
+   { 0x00000001, 0x2ac01a68, 0x00000fe4, 0x00000000 },
+   { 0x00000001, 0x2fa01a68, 0x00000ae6, 0x00000000 },
+   { 0x00000001, 0x2fa41a68, 0x00000b06, 0x00000000 },
+   { 0x00000001, 0x2fa81a68, 0x00000b26, 0x00000000 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x000006e0 },
+   { 0x00000001, 0x2ac21a68, 0x00000fe4, 0x00000000 },
+   { 0x00000001, 0x24401e68, 0x18000000, 0xfff0fff0 },
+   { 0x00000001, 0x24421e68, 0x18000000, 0xfff4fff4 },
+   { 0x00000001, 0x2fa00208, 0x00000440, 0x00000000 },
+   { 0x00000001, 0x2fa80208, 0x00000448, 0x00000000 },
+   { 0x00600001, 0x2fc00208, 0x008d0a40, 0x00000000 },
+   { 0x00000040, 0x2fe00008, 0x06001400, 0x00000020 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000780 },
+   { 0x00200001, 0x24401a68, 0x00450fe4, 0x00000000 },
+   { 0x00600001, 0x25600208, 0x008d0020, 0x00000000 },
+   { 0x00600001, 0x28400208, 0x008d0560, 0x00000000 },
+   { 0x00600001, 0x28600608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x28800608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x28a00608, 0x00000000, 0x00000000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00010001 },
+   { 0x00000001, 0x28a52288, 0x00000400, 0x00000000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00bb00bb },
+   { 0x00000001, 0x28a42288, 0x00000400, 0x00000000 },
+   { 0x00600001, 0x28c00608, 0x00000000, 0x00000000 },
+   { 0x00000001, 0x28d00608, 0x00000000, 0x11111111 },
+   { 0x00000001, 0x28dc0608, 0x00000000, 0x00010101 },
+   { 0x00600001, 0x28e00608, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x24600608, 0x00000000, 0x00000000 },
+   { 0x00000001, 0x247c1648, 0x10000000, 0x00000000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00020002 },
+   { 0x00000040, 0x24001248, 0x16000400, 0x00040004 },
+   { 0x00000001, 0x247c2288, 0x00000400, 0x00000000 },
+   { 0x00000001, 0x247d2288, 0x000000a5, 0x00000000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00000000 },
+   { 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, 0x24440208, 0x00000440, 0x00000000 },
+   { 0x00600001, 0x28000208, 0x008d0440, 0x00000000 },
+   { 0x00000001, 0x24600608, 0x00000000, 0x00000002 },
+   { 0x00000001, 0x24642288, 0x0000009c, 0x00000000 },
+   { 0x00000001, 0x24680608, 0x00000000, 0x30003030 },
+   { 0x00600001, 0x28200208, 0x008d0460, 0x00000000 },
+   { 0x00400001, 0x45800608, 0x00000000, 0x00000000 },
+   { 0x00400001, 0x45840608, 0x00000000, 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, 0x00241000 },
+   { 0x00000001, 0x24001648, 0x10000000, 0x00040004 },
+   { 0x00000040, 0x247e2288, 0x2200047e, 0x00000400 },
+   { 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, 0x040a0203 },
+   { 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 },
+   { 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 },
+   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
+   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
+   { 0x00200040, 0x2f601a68, 0x1a450fa8, 0x00450fa0 },
+   { 0x00000040, 0x2f701a68, 0x1e000f60, 0x00300030 },
+   { 0x00000040, 0x2f721a68, 0x1e000f62, 0x00280028 },
+   { 0x00200041, 0x2f901a68, 0x1e450fd0, 0xffffffff },
+   { 0x00200040, 0x2f801a68, 0x1a450fa8, 0x00450f90 },
+   { 0x00200040, 0x2f841a68, 0x1a450fa8, 0x00450fd0 },
+   { 0x05000010, 0x20001a60, 0x1a000f60, 0x00000f80 },
+   { 0x00010001, 0x2f601a68, 0x00000f80, 0x00000000 },
+   { 0x03000010, 0x20001a60, 0x1a000f70, 0x00000f84 },
+   { 0x00010040, 0x2f601a68, 0x1e000f84, 0xffd0ffd0 },
+   { 0x05000010, 0x20001a60, 0x1a000f62, 0x00000f82 },
+   { 0x00010001, 0x2f621a68, 0x00000f82, 0x00000000 },
+   { 0x03000010, 0x20001a60, 0x1a000f72, 0x00000f86 },
+   { 0x00010040, 0x2f621a68, 0x1e000f86, 0xffd8ffd8 },
+   { 0x05000010, 0x20001a60, 0x1e000f60, 0x00000000 },
+   { 0x00110020, 0x34000000, 0x0e001400, 0x00000020 },
+   { 0x00010001, 0x2f601e68, 0x18000000, 0x00000000 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000030 },
+   { 0x03000010, 0x20001a60, 0x1a000f70, 0x00000fc0 },
+   { 0x00110020, 0x34000000, 0x0e001400, 0x00000010 },
+   { 0x00010040, 0x2f601a68, 0x1e000fc0, 0xffd0ffd0 },
+   { 0x05000010, 0x20001a60, 0x1e000f62, 0x00000000 },
+   { 0x00110020, 0x34000000, 0x0e001400, 0x00000020 },
+   { 0x00010001, 0x2f621e68, 0x18000000, 0x00000000 },
+   { 0x00000020, 0x34000000, 0x0e001400, 0x00000020 },
+   { 0x03000010, 0x20001a60, 0x1a000f72, 0x00000fc2 },
+   { 0x00010040, 0x2f621a68, 0x1e000fc2, 0xffd8ffd8 },
+   { 0x00200041, 0x2f801a68, 0x1e450fa8, 0xffffffff },
+   { 0x00200040, 0x2fe41a68, 0x1a450f60, 0x00450f80 },
+   { 0x00000001, 0x34000200, 0x00000fe0, 0x00000000 },
+   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
+   { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },