VME: dual start and adaptive search
authorXiang, Haihao <haihao.xiang@intel.com>
Thu, 15 Mar 2012 06:05:48 +0000 (14:05 +0800)
committerXiang, Haihao <haihao.xiang@intel.com>
Thu, 15 Mar 2012 06:05:48 +0000 (14:05 +0800)
Signed-off-by: Xiang, Haihao <haihao.xiang@intel.com>
src/gen6_vme.c
src/shaders/vme/inter_frame.asm
src/shaders/vme/inter_frame.g6b
src/shaders/vme/inter_frame.g7b
src/shaders/vme/vme.inc

index 5145e72..46bfbfb 100644 (file)
@@ -342,19 +342,19 @@ static VAStatus gen6_vme_vme_state_setup(VADriverContextP ctx,
     vme_state_message = (unsigned int *)vme_context->vme_state.bo->virtual;
        
        vme_state_message[0] = 0x01010101;
-       vme_state_message[1] = 0x01010110;
+       vme_state_message[1] = 0x10010101;
        vme_state_message[2] = 0x0F0F0F0F;
-       vme_state_message[3] = 0x0F0F0F10;
+       vme_state_message[3] = 0x100F0F0F;
        vme_state_message[4] = 0x01010101;
-       vme_state_message[5] = 0x01010110;
-       vme_state_message[6] = 0x0F0F0F0F;
-       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;
+       vme_state_message[5] = 0x00010101;
+       vme_state_message[5] = 0x01010101;
+       vme_state_message[7] = 0x10010101;
+       vme_state_message[8] = 0x0F0F0F0F;
+       vme_state_message[9] = 0x100F0F0F;
+       vme_state_message[10] = 0x01010101;
+       vme_state_message[11] = 0x00010101;
+        vme_state_message[12] = 0x00;
+        vme_state_message[13] = 0x00;
 
     for(i = 14; i < 32; i++) {
         vme_state_message[i] = 0x00000000;
index 5bab1cc..667119e 100644 (file)
@@ -34,12 +34,13 @@ mov  (1) vme_m0.0<1>:W          -16:W {align1} ;                /* Reference = (
 mov  (1) vme_m0.2<1>:W          -12:W {align1} ;
 #endif
         
-mov  (1) vme_m0.12<1>:UD        INTER_PART_MASK + INTER_SAD_HAAR + SUB_PEL_MODE_QUARTER:UD {align1};    /* 16x16 Source, 1/4 pixel, harr */
+mov  (1) vme_m0.12<1>:UD        SEARCH_CTRL_DUAL_START + INTER_PART_MASK + INTER_SAD_HAAR + SUB_PEL_MODE_QUARTER:UD {align1};    /* 16x16 Source, 1/4 pixel, harr */
 mov  (1) vme_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
 mov  (1) vme_m0.22<1>:UW        REF_REGION_SIZE {align1};               /* Reference Width&Height, 32x32 */
 
-mov  (1) vme_m1.4<1>:UD         MAX_NUM_MV:UD {align1};                                   /* Default value MAX 32 MVs */
-mov  (1) vme_m1.8<1>:UD         SEARCH_PATH_LEN:UD {align1};
+mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
+mov  (1) vme_m1.4<1>:UD         FB_PRUNING_ENABLE + MAX_NUM_MV:UD {align1};                                   /* Default value MAX 32 MVs */
+mov  (1) vme_m1.8<1>:UD         START_CENTER + SEARCH_PATH_LEN:UD {align1};
 
 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};
index b39c58f..134f413 100644 (file)
@@ -5,11 +5,12 @@
    { 0x00200009, 0x24402e29, 0x004500a0, 0x00040004 },
    { 0x00000040, 0x24403dad, 0x00450440, 0xfff0fff0 },
    { 0x00000040, 0x24423dad, 0x00450442, 0xfff4fff4 },
-   { 0x00000001, 0x244c0061, 0x00000000, 0x7e203000 },
+   { 0x00000001, 0x244c0061, 0x00000000, 0x7e203100 },
    { 0x00000001, 0x24540231, 0x00000014, 0x00000000 },
    { 0x00000001, 0x24560169, 0x00000000, 0x28302830 },
-   { 0x00000001, 0x24640061, 0x00000000, 0x00000020 },
-   { 0x00000001, 0x24680061, 0x00000000, 0x00003f3f },
+   { 0x00000001, 0x24600061, 0x00000000, 0x00000002 },
+   { 0x00000001, 0x24640061, 0x00000000, 0x40000020 },
+   { 0x00000001, 0x24680061, 0x00000000, 0x30003f0c },
    { 0x00000041, 0x24884521, 0x000000a2, 0x000000a1 },
    { 0x00000040, 0x24884421, 0x00000488, 0x000000a0 },
    { 0x00000009, 0x24880c21, 0x00000488, 0x00000002 },
index ee9ed81..7257654 100644 (file)
@@ -4,11 +4,12 @@
    { 0x00200009, 0x24482e29, 0x004500a0, 0x00040004 },
    { 0x00000001, 0x244001ed, 0x00000000, 0xfff0fff0 },
    { 0x00000001, 0x244201ed, 0x00000000, 0xfff4fff4 },
-   { 0x00000001, 0x244c0061, 0x00000000, 0x7e203000 },
+   { 0x00000001, 0x244c0061, 0x00000000, 0x7e203100 },
    { 0x00000001, 0x24540231, 0x00000014, 0x00000000 },
    { 0x00000001, 0x24560169, 0x00000000, 0x28302830 },
-   { 0x00000001, 0x24640061, 0x00000000, 0x00000020 },
-   { 0x00000001, 0x24680061, 0x00000000, 0x00003f3f },
+   { 0x00000001, 0x24600061, 0x00000000, 0x00000002 },
+   { 0x00000001, 0x24640061, 0x00000000, 0x40000020 },
+   { 0x00000001, 0x24680061, 0x00000000, 0x30003f0c },
    { 0x00000041, 0x24884521, 0x000000a2, 0x000000a1 },
    { 0x00000040, 0x24884421, 0x00000488, 0x000000a0 },
    { 0x00000009, 0x24880c21, 0x00000488, 0x00000002 },
index 44fe320..73a554d 100644 (file)
@@ -48,12 +48,21 @@ define(`INTRA_SAD_HAAR',                `0x00800000')
 
 define(`INTER_PART_MASK',               `0x7E000000')
 
+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(`SEARCH_PATH_LEN',               `0x00003F3F')
+define(`FB_PRUNING_ENABLE',             `0x40000000')
+
+define(`SEARCH_PATH_LEN',               `0x00003F0C')
+define(`START_CENTER',                  `0x30000000')
+
+define(`ADAPTIVE_SEARCH_ENABLE',        `0x00000002') 
 define(`INTRA_PREDICTORE_MODE',         `0x11111111:UD')
 
 #ifdef DEV_SNB