Fixed a vme search path bug, current search area is 48x40.
authorZhou Chang <chang.zhou@intel.com>
Thu, 21 Jul 2011 03:06:47 +0000 (11:06 +0800)
committerGwenole Beauchesne <gwenole.beauchesne@intel.com>
Thu, 25 Aug 2011 12:29:01 +0000 (14:29 +0200)
src/gen6_vme.c
src/shaders/vme/gen6_vme_header.inc
src/shaders/vme/gen7_vme_header.inc
src/shaders/vme/inter_frame.asm
src/shaders/vme/inter_frame.g6b
src/shaders/vme/inter_frame.g7b

index 290982b..dd6fb65 100644 (file)
@@ -661,20 +661,20 @@ static VAStatus gen6_vme_vme_state_setup(VADriverContextP ctx,
     assert(vme_context->vme_state.bo->virtual);
     vme_state_message = (unsigned int *)vme_context->vme_state.bo->virtual;
        
-       vme_state_message[0] = 0x11010101;
+       vme_state_message[0] = 0x01010101;
        vme_state_message[1] = 0x01010110;
        vme_state_message[2] = 0x0F0F0F0F;
-       vme_state_message[3] = 0x0F0F1001;
+       vme_state_message[3] = 0x0F0F0F10;
        vme_state_message[4] = 0x01010101;
-       vme_state_message[5] = 0x01100F0F;
+       vme_state_message[5] = 0x01010110;
        vme_state_message[6] = 0x0F0F0F0F;
-       vme_state_message[7] = 0x10010101;
-       vme_state_message[8] = 0x0101100F;
-       vme_state_message[9] = 0x0F0F0F0F;
-    vme_state_message[10] = 0x0F100101;
-    vme_state_message[11] = 0x01010101;
-    vme_state_message[12] = 0x100F0F0F;
-    vme_state_message[13] = 0x0F0F0F00;
+       vme_state_message[7] = 0x0F0F0F10;
+       vme_state_message[8] = 0x01010101;
+       vme_state_message[9] = 0x01010110;
+    vme_state_message[10] = 0x0F0F0F0F;
+    vme_state_message[11] = 0x0F0F0F10;
+    vme_state_message[12] = 0x01010101;
+    vme_state_message[13] = 0x01010100;
 
     for(i = 14; i < 32; i++) {
         vme_state_message[i] = 0x00000000;
index 0d9f15b..5e27fa7 100644 (file)
@@ -48,7 +48,7 @@ define(`INTRA_SAD_HAAR',                `0x00800000')
 
 define(`INTER_PART_MASK',               `0x7E000000')
 
-define(`REF_REGION_SIZE',               `0x3030:UW')
+define(`REF_REGION_SIZE',               `0x2830:UW')
 
 define(`BI_SUB_MB_PART_MASK',           `0x0c000000')
 define(`MAX_NUM_MV',                    `0x00000020')
index 9cec738..44ab7b6 100644 (file)
@@ -48,7 +48,7 @@ define(`INTRA_SAD_HAAR',                `0x00800000')
 
 define(`INTER_PART_MASK',               `0x7E000000')
 
-define(`REF_REGION_SIZE',               `0x2020:UW')
+define(`REF_REGION_SIZE',               `0x2830:UW')
 
 define(`BI_SUB_MB_PART_MASK',           `0x0c000000')
 define(`MAX_NUM_MV',                    `0x00000020')
index afff11d..0a4db66 100644 (file)
@@ -28,7 +28,8 @@ mov  (16) tmp_reg2.0<1>:UD      0x0:UD {align1};
 /* m0 */        
 mul  (2) tmp_reg0.8<1>:UW       orig_xy_ub<2,2,1>:UB 16:UW {align1};    /* Source =  (x, y) * 16 */
 mul  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB 16:UW {align1};   
-add  (2) tmp_reg0.0<1>:W        tmp_reg0.0<2,2,1>:W -16:W {align1};            /* Reference = (x-16,y-16)-(x+32,y+32) */
+add  (1) tmp_reg0.0<1>:W        tmp_reg0.0<2,2,1>:W -16:W {align1};            /* Reference = (x-16,y-12)-(x+32,y+24) */
+add  (1) tmp_reg0.2<1>:W        tmp_reg0.2<2,2,1>:W -12:W {align1};
 mov  (1) tmp_reg0.12<1>:UD      INTER_PART_MASK + INTER_SAD_HAAR + SUB_PEL_MODE_QUARTER:UD {align1};    /* 16x16 Source, 1/4 pixel, harr */
 
 mov  (1) tmp_reg0.20<1>:UB      thread_id_ub {align1};                  /* dispatch id */
@@ -74,7 +75,9 @@ mov  (1) tmp_reg3.20<1>:UB      thread_id_ub {align1};                  /* dispa
 mov  (8) msg_reg0.0<1>:UD       tmp_reg3.0<8,8,1>:UD {align1};
 
 mov  (2) tmp_reg3.0<1>:UW       vme_wb1.0<2,2,1>:UB  {align1};
-add  (2) tmp_reg3.0<1>:W        tmp_reg3.0<2,2,1>:W -64:W {align1}; 
+add  (1) tmp_reg3.0<1>:W        tmp_reg3.0<2,2,1>:W -64:W {align1};
+add  (1) tmp_reg3.2<1>:W        tmp_reg3.2<2,2,1>:W -48:W {align1}; 
+        
         
 mov  (8) msg_reg1.0<1>:UD       tmp_reg3.0<8,8,0>:UD   {align1};
 
index 0e798e8..45fa058 100644 (file)
@@ -2,10 +2,11 @@
    { 0x00800001, 0x24400061, 0x00000000, 0x00000000 },
    { 0x00200041, 0x24082e29, 0x004500a0, 0x00100010 },
    { 0x00200041, 0x24002e29, 0x004500a0, 0x00100010 },
-   { 0x00200040, 0x24003dad, 0x00450400, 0xfff0fff0 },
+   { 0x00000040, 0x24003dad, 0x00450400, 0xfff0fff0 },
+   { 0x00000040, 0x24023dad, 0x00450402, 0xfff4fff4 },
    { 0x00000001, 0x240c0061, 0x00000000, 0x7e203000 },
    { 0x00000001, 0x24140231, 0x00000014, 0x00000000 },
-   { 0x00000001, 0x24160169, 0x00000000, 0x30303030 },
+   { 0x00000001, 0x24160169, 0x00000000, 0x28302830 },
    { 0x00600001, 0x20000022, 0x008d0400, 0x00000000 },
    { 0x00000001, 0x24240061, 0x00000000, 0x00000020 },
    { 0x00000001, 0x24280061, 0x00000000, 0x00003f3f },
@@ -20,7 +21,8 @@
    { 0x00000001, 0x24740231, 0x00000014, 0x00000000 },
    { 0x00600001, 0x20000022, 0x008d0460, 0x00000000 },
    { 0x00200001, 0x24600229, 0x004501a0, 0x00000000 },
-   { 0x00200040, 0x24603dad, 0x00450460, 0xffc0ffc0 },
+   { 0x00000040, 0x24603dad, 0x00450460, 0xffc0ffc0 },
+   { 0x00000040, 0x24623dad, 0x00450462, 0xffd0ffd0 },
    { 0x00600001, 0x20200022, 0x008c0460, 0x00000000 },
    { 0x00600001, 0x20400022, 0x008c0460, 0x00000000 },
    { 0x05800031, 0x22001cdd, 0x00000000, 0x061b0303 },
index 4c40974..68fe540 100644 (file)
@@ -2,10 +2,11 @@
    { 0x00800001, 0x24400061, 0x00000000, 0x00000000 },
    { 0x00200041, 0x24082e29, 0x004500a0, 0x00100010 },
    { 0x00200041, 0x24002e29, 0x004500a0, 0x00100010 },
-   { 0x00200040, 0x24003dad, 0x00450400, 0xfff0fff0 },
+   { 0x00000040, 0x24003dad, 0x00450400, 0xfff0fff0 },
+   { 0x00000040, 0x24023dad, 0x00450402, 0xfff4fff4 },
    { 0x00000001, 0x240c0061, 0x00000000, 0x7e203000 },
    { 0x00000001, 0x24140231, 0x00000014, 0x00000000 },
-   { 0x00000001, 0x24160169, 0x00000000, 0x20202020 },
+   { 0x00000001, 0x24160169, 0x00000000, 0x28302830 },
    { 0x00600001, 0x28000021, 0x008d0400, 0x00000000 },
    { 0x00000001, 0x24240061, 0x00000000, 0x00000020 },
    { 0x00000001, 0x24280061, 0x00000000, 0x00003f3f },
@@ -20,7 +21,8 @@
    { 0x00000001, 0x24740231, 0x00000014, 0x00000000 },
    { 0x00600001, 0x28000021, 0x008d0460, 0x00000000 },
    { 0x00200001, 0x24600229, 0x004501a0, 0x00000000 },
-   { 0x00200040, 0x24603dad, 0x00450460, 0xffc0ffc0 },
+   { 0x00000040, 0x24603dad, 0x00450460, 0xffc0ffc0 },
+   { 0x00000040, 0x24623dad, 0x00450462, 0xffd0ffd0 },
    { 0x00600001, 0x28200021, 0x008c0460, 0x00000000 },
    { 0x00600001, 0x28400021, 0x008c0460, 0x00000000 },
    { 0x0a800031, 0x20001cac, 0x00000800, 0x060a0300 },