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 | // Module name: Transpose_Y_16x16.asm |
||
| 11 | // |
||
| 12 | // Transpose Y 16x16 block. |
||
| 13 | // |
||
| 14 | //---------------------------------------------------------------------------------------- |
||
| 15 | // Symbols need to be defined before including this module |
||
| 16 | // |
||
| 17 | // Source region is :ub |
||
| 18 | // SRC_YB: SRC_YB Base=rxx ElementSize=1 SrcRegion=REGION(16,1) Type=ub // 8 GRFs |
||
| 19 | // |
||
| 20 | // Temp buffer: |
||
| 21 | // CUR_TEMP_B: BUF_B Base=rxx ElementSize=1 SrcRegion=REGION(16,1) Type=ub // 8 GRFs |
||
| 22 | // |
||
| 23 | ////////////////////////////////////////////////////////////////////////////////////////// |
||
| 24 | |||
| 25 | #if defined(_DEBUG) |
||
| 26 | mov (1) EntrySignatureC:w 0xDDDA:w |
||
| 27 | #endif |
||
| 28 | |||
| 29 | |||
| 30 | // Transpose Y (16x16 bytes) |
||
| 31 | |||
| 32 | // The first step |
||
| 33 | mov (16) CUR_TEMP_B(0,0)<1> SRC_YB(0,0)<16;4,1> { NoDDClr } |
||
| 34 | mov (16) CUR_TEMP_B(0,16)<1> SRC_YB(2,0)<16;4,1> { NoDDChk } |
||
| 35 | mov (16) CUR_TEMP_B(1,0)<1> SRC_YB(4,0)<16;4,1> { NoDDClr } |
||
| 36 | mov (16) CUR_TEMP_B(1,16)<1> SRC_YB(6,0)<16;4,1> { NoDDChk } |
||
| 37 | |||
| 38 | mov (16) CUR_TEMP_B(2,0)<1> SRC_YB(0,4)<16;4,1> { NoDDClr } |
||
| 39 | mov (16) CUR_TEMP_B(2,16)<1> SRC_YB(2,4)<16;4,1> { NoDDChk } |
||
| 40 | mov (16) CUR_TEMP_B(3,0)<1> SRC_YB(4,4)<16;4,1> { NoDDClr } |
||
| 41 | mov (16) CUR_TEMP_B(3,16)<1> SRC_YB(6,4)<16;4,1> { NoDDChk } |
||
| 42 | |||
| 43 | mov (16) CUR_TEMP_B(4,0)<1> SRC_YB(0,8)<16;4,1> { NoDDClr } |
||
| 44 | mov (16) CUR_TEMP_B(4,16)<1> SRC_YB(2,8)<16;4,1> { NoDDChk } |
||
| 45 | mov (16) CUR_TEMP_B(5,0)<1> SRC_YB(4,8)<16;4,1> { NoDDClr } |
||
| 46 | mov (16) CUR_TEMP_B(5,16)<1> SRC_YB(6,8)<16;4,1> { NoDDChk } |
||
| 47 | |||
| 48 | mov (16) CUR_TEMP_B(6,0)<1> SRC_YB(0,12)<16;4,1> { NoDDClr } |
||
| 49 | mov (16) CUR_TEMP_B(6,16)<1> SRC_YB(2,12)<16;4,1> { NoDDChk } |
||
| 50 | mov (16) CUR_TEMP_B(7,0)<1> SRC_YB(4,12)<16;4,1> { NoDDClr } |
||
| 51 | mov (16) CUR_TEMP_B(7,16)<1> SRC_YB(6,12)<16;4,1> { NoDDChk } |
||
| 52 | |||
| 53 | // The second step |
||
| 54 | mov (16) SRC_YB(0,0)<1> CUR_TEMP_B(0,0)<32;8,4> { NoDDClr } |
||
| 55 | mov (16) SRC_YB(0,16)<1> CUR_TEMP_B(0,1)<32;8,4> { NoDDChk } |
||
| 56 | mov (16) SRC_YB(1,0)<1> CUR_TEMP_B(0,2)<32;8,4> { NoDDClr } |
||
| 57 | mov (16) SRC_YB(1,16)<1> CUR_TEMP_B(0,3)<32;8,4> { NoDDChk } |
||
| 58 | |||
| 59 | mov (16) SRC_YB(2,0)<1> CUR_TEMP_B(2,0)<32;8,4> { NoDDClr } |
||
| 60 | mov (16) SRC_YB(2,16)<1> CUR_TEMP_B(2,1)<32;8,4> { NoDDChk } |
||
| 61 | mov (16) SRC_YB(3,0)<1> CUR_TEMP_B(2,2)<32;8,4> { NoDDClr } |
||
| 62 | mov (16) SRC_YB(3,16)<1> CUR_TEMP_B(2,3)<32;8,4> { NoDDChk } |
||
| 63 | |||
| 64 | mov (16) SRC_YB(4,0)<1> CUR_TEMP_B(4,0)<32;8,4> { NoDDClr } |
||
| 65 | mov (16) SRC_YB(4,16)<1> CUR_TEMP_B(4,1)<32;8,4> { NoDDChk } |
||
| 66 | mov (16) SRC_YB(5,0)<1> CUR_TEMP_B(4,2)<32;8,4> { NoDDClr } |
||
| 67 | mov (16) SRC_YB(5,16)<1> CUR_TEMP_B(4,3)<32;8,4> { NoDDChk } |
||
| 68 | |||
| 69 | mov (16) SRC_YB(6,0)<1> CUR_TEMP_B(6,0)<32;8,4> { NoDDClr } |
||
| 70 | mov (16) SRC_YB(6,16)<1> CUR_TEMP_B(6,1)<32;8,4> { NoDDChk } |
||
| 71 | mov (16) SRC_YB(7,0)<1> CUR_TEMP_B(6,2)<32;8,4> { NoDDClr } |
||
| 72 | mov (16) SRC_YB(7,16)<1> CUR_TEMP_B(6,3)<32;8,4> { NoDDChk } |
||
| 73 | |||
| 74 | // Y is transposed. |