Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 5361 | serge | 1 | /* |
| 2 | * Copyright © <2010>, Intel Corporation. |
||
| 3 | * |
||
| 4 | * This program is licensed under the terms and conditions of the |
||
| 5 | * Eclipse Public License (EPL), version 1.0. The full text of the EPL is at |
||
| 6 | * http://www.opensource.org/licenses/eclipse-1.0.php. |
||
| 7 | * |
||
| 8 | */ |
||
| 9 | ////////////////////////////////////////////////////////////////////////////////////////////////////////////// |
||
| 10 | // AVC Child Kernel (Vertical and horizontal de-block a 4:2:0 MB Y comp) |
||
| 11 | // |
||
| 12 | // First, de-block vertical edges from left to right. |
||
| 13 | // Second, de-block horizontal edge from top to bottom. |
||
| 14 | // |
||
| 15 | // ***** MBAFF Mode ***** |
||
| 16 | // This version deblocks top MB first, followed by bottom MB. |
||
| 17 | // |
||
| 18 | // Need variable CurMB to indicate top MB or bottom MB (CurMB = 0 or 1). |
||
| 19 | // We can use BotFieldFlag in BitFields to represent it. |
||
| 20 | // |
||
| 21 | // Usage: |
||
| 22 | // 1) Access control data for top |
||
| 23 | // CntrlDataOffsetY + CurMB * Control data block size (64 DWs for CL, 16 DWs for BLC) |
||
| 24 | // |
||
| 25 | // 2) Load frame/field video data based on flags: FieldModeCurrentMbFlag, FieldModeLeftMbFlag, FieldModeaboveMbFlag, |
||
| 26 | // |
||
| 27 | // E.g. |
||
| 28 | // if (pCntlData->BitField & FieldModeCurrentMbFlag) |
||
| 29 | // cur_y = ORIX_CUR.y + CurMB * 1; // Add field vertical offset for bot field MB . |
||
| 30 | // else |
||
| 31 | // cur_y = ORIX_CUR.y + CurMB * MB_Rows_Y; // Add bottom MB vertical offset for bot MB |
||
| 32 | // |
||
| 33 | // |
||
| 34 | ////////////////////////////////////////////////////////////////////////////////////////////////////////////// |
||
| 35 | #define AVC_ILDB |
||
| 36 | |||
| 37 | .kernel AVC_ILDB_CHILD_MBAFF_Y |
||
| 38 | #if defined(COMBINED_KERNEL) |
||
| 39 | ILDB_LABEL(AVC_ILDB_CHILD_Y): |
||
| 40 | #endif |
||
| 41 | |||
| 42 | #include "SetupVPKernel.asm" |
||
| 43 | #include "AVC_ILDB.inc" |
||
| 44 | |||
| 45 | #if defined(_DEBUG) |
||
| 46 | mov (1) EntrySignatureC:w 0xE998:w |
||
| 47 | #endif |
||
| 48 | |||
| 49 | // Setup temp buf used by load and save code |
||
| 50 | #define BUF_B RTempB |
||
| 51 | #define BUF_D RTempD |
||
| 52 | |||
| 53 | // Init local variables |
||
| 54 | // These coordinates are in progressive fashion |
||
| 55 | mul (4) ORIX_CUR<2>:w ORIX<0;1,0>:w 16:w { NoDDClr } // Expand X addr to bytes, repeat 4 times |
||
| 56 | mul (4) ORIY_CUR<2>:w ORIY<0;1,0>:w 32:w { NoDDChk } // Expand Y addr to bytes, repeat 4 times |
||
| 57 | |||
| 58 | mov (2) f0.0<1>:w 0:w |
||
| 59 | |||
| 60 | mov (1) GateWayOffsetC:uw ORIY:uw // Use row # as Gateway offset |
||
| 61 | |||
| 62 | //=== Null Kernel =============================================================== |
||
| 63 | // jmpi POST_ILDB |
||
| 64 | //=============================================================================== |
||
| 65 | |||
| 66 | //==================================================================================== |
||
| 67 | // Assuming the MB control data is laid out in scan line order in a rectangle with width = 16 bytes. |
||
| 68 | // Control data has dimension of X x Y = 16 x N bytes, where N = W x H / 16 |
||
| 69 | // Each MB has 256 bytes of control data |
||
| 70 | |||
| 71 | // For CRESTLINE, 256 bytes are stored in memory and fetched into GRF. |
||
| 72 | // MB_offset = MBsCntX * CurRow + CurCol |
||
| 73 | // Byte_offset = MB_offset * (256 << Mbaff_flag), Mbaff_flag = 0 or 1. |
||
| 74 | // Base address of a control data block = (x, y) = (0, y'=y/x), region width is 16 bytes |
||
| 75 | // where y' = Byte_offset / 16 = MB_offset * (16 << Mbaff_flag) |
||
| 76 | // MBCntrlDataOffsetY holds y'. |
||
| 77 | |||
| 78 | // For BearLake-C, 64 bytes are stored in memory and dataport expands to 256 bytes. Need to use a special read command on BL-C. |
||
| 79 | // MB_offset = MBsCntX * CurRow + CurCol |
||
| 80 | // Byte_offset = MB_offset * (64 << Mbaff_flag), Mbaff_flag = 0 or 1. |
||
| 81 | // MBCntrlDataOffsetY holds globel byte offset. |
||
| 82 | |||
| 83 | #if !defined(DEV_CL) |
||
| 84 | mul (1) CntrlDataOffsetY:ud MBsCntX:w ORIY:w |
||
| 85 | add (1) CntrlDataOffsetY:ud CntrlDataOffsetY:ud ORIX:w |
||
| 86 | mul (1) CntrlDataOffsetY:ud CntrlDataOffsetY:ud 128:uw |
||
| 87 | #endif |
||
| 88 | |||
| 89 | //==================================================================================== |
||
| 90 | |||
| 91 | add (1) ORIX_LEFT:w ORIX_LEFT:w -4:w |
||
| 92 | add (1) ORIY_TOP:w ORIY_TOP:w -4:w |
||
| 93 | |||
| 94 | |||
| 95 | //=========== Process Top MB ============ |
||
| 96 | and (1) BitFields:w BitFields:w TopFieldFlag:w // Reset BotFieldFlag |
||
| 97 | |||
| 98 | RE_ENTRY: // for bootom field |
||
| 99 | |||
| 100 | // Load current MB control data |
||
| 101 | #if defined(DEV_CL) |
||
| 102 | #include "Load_ILDB_Cntrl_Data_64DW.asm" // Crestline |
||
| 103 | #else |
||
| 104 | #include "Load_ILDB_Cntrl_Data_16DW.asm" // Cantiga and beyond |
||
| 105 | #endif |
||
| 106 | |||
| 107 | // Init addr register for vertical control data |
||
| 108 | mov (1) ECM_AddrReg<1>:w CNTRL_DATA_BASE:w // Init edge control map AddrReg |
||
| 109 | |||
| 110 | // Check loaded control data |
||
| 111 | and.z.f0.1 (16) null<1>:uw r[ECM_AddrReg, wEdgeCntlMap_IntLeftVert]<16;16,1>:uw 0xFFFF:uw // Skip ILDB? |
||
| 112 | and.nz.f0.0 (1) null:w r[ECM_AddrReg, ExtBitFlags]:ub DISABLE_ILDB_FLAG:w // Skip ILDB? |
||
| 113 | |||
| 114 | // Use free cycles here |
||
| 115 | // Set DualFieldMode for all data read, write and deblocking |
||
| 116 | and (1) CTemp1_W:uw r[ECM_AddrReg, BitFlags]:ub FieldModeAboveMbFlag+FieldModeCurrentMbFlag:uw |
||
| 117 | |||
| 118 | // Get Vert Edge Pattern (frame vs. field MBs) |
||
| 119 | and (1) VertEdgePattern:uw r[ECM_AddrReg, BitFlags]:ub FieldModeLeftMbFlag+FieldModeCurrentMbFlag:uw |
||
| 120 | |||
| 121 | (f0.1.all16h) jmpi SKIP_ILDB // Skip ILDB |
||
| 122 | (f0.0) jmpi SKIP_ILDB // Skip ILDB |
||
| 123 | |||
| 124 | // Set DualFieldMode for all data read, write and deblocking |
||
| 125 | // and (1) CTemp1_W:uw r[ECM_AddrReg, BitFlags]:ub FieldModeAboveMbFlag+FieldModeCurrentMbFlag:uw |
||
| 126 | cmp.z.f0.0 (1) null:w CTemp1_W:uw ABOVE_FIELD_CUR_FRAME:w |
||
| 127 | and (1) DualFieldMode:w f0.0:w 0x0001:w |
||
| 128 | |||
| 129 | // Load current MB // DDD1 |
||
| 130 | #include "load_Cur_Y_16x16T_Mbaff.asm" // Load cur Y, 16x16, transpose |
||
| 131 | #include "load_Left_Y_4x16T_Mbaff.asm" // Load left MB (4x16) Y data from memory if exists |
||
| 132 | |||
| 133 | #include "Transpose_Cur_Y_16x16.asm" |
||
| 134 | #include "Transpose_Left_Y_4x16.asm" |
||
| 135 | |||
| 136 | //---------- Perform vertical ILDB filting on Y---------- |
||
| 137 | #include "AVC_ILDB_Filter_Mbaff_Y_v.asm" |
||
| 138 | //------------------------------------------------------- |
||
| 139 | |||
| 140 | #include "save_Left_Y_16x4T_Mbaff.asm" // Write left MB (4x16) Y data to memory if exists |
||
| 141 | #include "load_Top_Y_16x4_Mbaff.asm" // Load top MB (16x4) Y data from memory if exists |
||
| 142 | #include "Transpose_Cur_Y_16x16.asm" // Transpose a MB for horizontal edge de-blocking |
||
| 143 | |||
| 144 | //---------- Perform horizontal ILDB filting on Y ---------- |
||
| 145 | #include "AVC_ILDB_Filter_Mbaff_Y_h.asm" |
||
| 146 | //---------------------------------------------------------- |
||
| 147 | |||
| 148 | #include "save_Cur_Y_16x16_Mbaff.asm" // Write cur MB (16x16) |
||
| 149 | #include "save_Top_Y_16x4_Mbaff.asm" // Write top MB (16x4) if not the top row |
||
| 150 | |||
| 151 | SKIP_ILDB: |
||
| 152 | //---------------------------------------------------------- |
||
| 153 | and.z.f0.0 (1) null:w BitFields:w BotFieldFlag:w |
||
| 154 | |||
| 155 | //=========== Process Bottom MB ============ |
||
| 156 | or (1) BitFields:w BitFields:w BotFieldFlag:w // Set BotFieldFlag to 1 |
||
| 157 | (f0.0) jmpi RE_ENTRY // Loop back for bottom deblocking |
||
| 158 | |||
| 159 | // Fall through to finish |
||
| 160 | |||
| 161 | //=========== Check write commit of the last write ============ |
||
| 162 | mov (8) WritebackResponse(0)<1> WritebackResponse(0) |
||
| 163 | |||
| 164 | POST_ILDB: |
||
| 165 | |||
| 166 | //--------------------------------------------------------------------------- |
||
| 167 | // Send notification thru Gateway to root thread, update luma Status[CurRow] |
||
| 168 | |||
| 169 | #include "AVC_ILDB_ForwardMsg.asm" |
||
| 170 | |||
| 171 | #if !defined(GW_DCN) // For non-ILK chipsets |
||
| 172 | //child send EOT : Request type = 1 |
||
| 173 | END_CHILD_THREAD |
||
| 174 | #endif // !defined(DEV_ILK) |
||
| 175 | |||
| 176 | // The thread finishs here |
||
| 177 | //------------------------------------------------------------------------------ |
||
| 178 | |||
| 179 | |||
| 180 | //////////////////////////////////////////////////////////////////////////////// |
||
| 181 | // Include other subrutines being called |
||
| 182 | #include "AVC_ILDB_Luma_Core_Mbaff.asm" |
||
| 183 | |||
| 184 | #if !defined(COMBINED_KERNEL) // For standalone kernel only |
||
| 185 | .end_code |
||
| 186 | |||
| 187 | .end_kernel |
||
| 188 | #endif |