10c1bc742181ded4930842b46e9507372f0b1b963James Dong;// 20c1bc742181ded4930842b46e9507372f0b1b963James Dong;// 30c1bc742181ded4930842b46e9507372f0b1b963James Dong;// File Name: omxVCM4P10_PredictIntra_16x16_s.s 40c1bc742181ded4930842b46e9507372f0b1b963James Dong;// OpenMAX DL: v1.0.2 50c1bc742181ded4930842b46e9507372f0b1b963James Dong;// Revision: 12290 60c1bc742181ded4930842b46e9507372f0b1b963James Dong;// Date: Wednesday, April 9, 2008 70c1bc742181ded4930842b46e9507372f0b1b963James Dong;// 80c1bc742181ded4930842b46e9507372f0b1b963James Dong;// (c) Copyright 2007-2008 ARM Limited. All Rights Reserved. 90c1bc742181ded4930842b46e9507372f0b1b963James Dong;// 100c1bc742181ded4930842b46e9507372f0b1b963James Dong;// 110c1bc742181ded4930842b46e9507372f0b1b963James Dong;// 120c1bc742181ded4930842b46e9507372f0b1b963James Dong 130c1bc742181ded4930842b46e9507372f0b1b963James Dong INCLUDE omxtypes_s.h 140c1bc742181ded4930842b46e9507372f0b1b963James Dong INCLUDE armCOMM_s.h 150c1bc742181ded4930842b46e9507372f0b1b963James Dong 160c1bc742181ded4930842b46e9507372f0b1b963James Dong M_VARIANTS CortexA8 170c1bc742181ded4930842b46e9507372f0b1b963James Dong 180c1bc742181ded4930842b46e9507372f0b1b963James Dong 190c1bc742181ded4930842b46e9507372f0b1b963James Dong;//------------------------------------------------------- 200c1bc742181ded4930842b46e9507372f0b1b963James Dong;// This table for implementing switch case of C in asm by 210c1bc742181ded4930842b46e9507372f0b1b963James Dong;// the mehtod of two levels of indexing. 220c1bc742181ded4930842b46e9507372f0b1b963James Dong;//------------------------------------------------------- 230c1bc742181ded4930842b46e9507372f0b1b963James Dong 240c1bc742181ded4930842b46e9507372f0b1b963James Dong M_TABLE armVCM4P10_pIndexTable16x16 250c1bc742181ded4930842b46e9507372f0b1b963James Dong DCD OMX_VC_16X16_VERT, OMX_VC_16X16_HOR 260c1bc742181ded4930842b46e9507372f0b1b963James Dong DCD OMX_VC_16X16_DC, OMX_VC_16X16_PLANE 270c1bc742181ded4930842b46e9507372f0b1b963James Dong 280c1bc742181ded4930842b46e9507372f0b1b963James Dong 290c1bc742181ded4930842b46e9507372f0b1b963James Dong IF CortexA8 300c1bc742181ded4930842b46e9507372f0b1b963James Dong 310c1bc742181ded4930842b46e9507372f0b1b963James Dong M_TABLE armVCM4P10_MultiplierTable16x16,1 320c1bc742181ded4930842b46e9507372f0b1b963James Dong DCW 7, 6, 5, 4, 3, 2, 1, 8 330c1bc742181ded4930842b46e9507372f0b1b963James Dong DCW 0, 1, 2, 3, 4, 5, 6, 7 340c1bc742181ded4930842b46e9507372f0b1b963James Dong DCW 8, 9, 10, 11, 12, 13, 14, 15 350c1bc742181ded4930842b46e9507372f0b1b963James Dong 360c1bc742181ded4930842b46e9507372f0b1b963James Dong;//-------------------------------------------- 370c1bc742181ded4930842b46e9507372f0b1b963James Dong;// Constants 380c1bc742181ded4930842b46e9507372f0b1b963James Dong;//-------------------------------------------- 390c1bc742181ded4930842b46e9507372f0b1b963James DongBLK_SIZE EQU 0x10 400c1bc742181ded4930842b46e9507372f0b1b963James DongMUL_CONST0 EQU 0x01010101 410c1bc742181ded4930842b46e9507372f0b1b963James DongMUL_CONST1 EQU 0x00060004 420c1bc742181ded4930842b46e9507372f0b1b963James DongMUL_CONST2 EQU 0x00070005 430c1bc742181ded4930842b46e9507372f0b1b963James DongMUL_CONST3 EQU 0x00030001 440c1bc742181ded4930842b46e9507372f0b1b963James DongMASK_CONST EQU 0x00FF00FF 450c1bc742181ded4930842b46e9507372f0b1b963James Dong 460c1bc742181ded4930842b46e9507372f0b1b963James Dong;//-------------------------------------------- 470c1bc742181ded4930842b46e9507372f0b1b963James Dong;// Scratch variable 480c1bc742181ded4930842b46e9507372f0b1b963James Dong;//-------------------------------------------- 490c1bc742181ded4930842b46e9507372f0b1b963James Dongy RN 12 500c1bc742181ded4930842b46e9507372f0b1b963James Dongpc RN 15 510c1bc742181ded4930842b46e9507372f0b1b963James Dong 520c1bc742181ded4930842b46e9507372f0b1b963James Dongreturn RN 0 530c1bc742181ded4930842b46e9507372f0b1b963James DongpTable RN 9 540c1bc742181ded4930842b46e9507372f0b1b963James Dongcount RN 11 550c1bc742181ded4930842b46e9507372f0b1b963James DongpMultTable RN 9 560c1bc742181ded4930842b46e9507372f0b1b963James Dong; ---------------------------------------------- 570c1bc742181ded4930842b46e9507372f0b1b963James Dong; Neon registers 580c1bc742181ded4930842b46e9507372f0b1b963James Dong; ---------------------------------------------- 590c1bc742181ded4930842b46e9507372f0b1b963James DongqAbove QN Q0.U8 600c1bc742181ded4930842b46e9507372f0b1b963James DongqLeft QN Q1.U8 610c1bc742181ded4930842b46e9507372f0b1b963James DongqSum8 QN Q0.U16 620c1bc742181ded4930842b46e9507372f0b1b963James DongdSum80 DN D0.U16 630c1bc742181ded4930842b46e9507372f0b1b963James DongdSum81 DN D1.U16 640c1bc742181ded4930842b46e9507372f0b1b963James DongdSum4 DN D0.U16 650c1bc742181ded4930842b46e9507372f0b1b963James DongdSum2 DN D0.U32 660c1bc742181ded4930842b46e9507372f0b1b963James DongdSum1 DN D0.U64 670c1bc742181ded4930842b46e9507372f0b1b963James DongqOut QN Q3.U8 680c1bc742181ded4930842b46e9507372f0b1b963James DongdSumLeft DN D6.U64 690c1bc742181ded4930842b46e9507372f0b1b963James DongdSumAbove DN D7.U64 700c1bc742181ded4930842b46e9507372f0b1b963James DongdSum DN D8.U64 710c1bc742181ded4930842b46e9507372f0b1b963James DongdSum0 DN D8.U8[0] 720c1bc742181ded4930842b46e9507372f0b1b963James Dong 730c1bc742181ded4930842b46e9507372f0b1b963James DongqH QN Q11.S32 740c1bc742181ded4930842b46e9507372f0b1b963James DongqV QN Q12.S32 750c1bc742181ded4930842b46e9507372f0b1b963James DongqA QN Q11.S16 760c1bc742181ded4930842b46e9507372f0b1b963James DongqB QN Q6.S16 770c1bc742181ded4930842b46e9507372f0b1b963James DongqC QN Q7.S16 780c1bc742181ded4930842b46e9507372f0b1b963James Dong 790c1bc742181ded4930842b46e9507372f0b1b963James DongqB0 QN Q5.S16 800c1bc742181ded4930842b46e9507372f0b1b963James DongqB1 QN Q6.S16 810c1bc742181ded4930842b46e9507372f0b1b963James DongdA1 DN D23.S16 820c1bc742181ded4930842b46e9507372f0b1b963James Dong 830c1bc742181ded4930842b46e9507372f0b1b963James DongdH0 DN D22.S32 840c1bc742181ded4930842b46e9507372f0b1b963James DongdH1 DN D23.S32 850c1bc742181ded4930842b46e9507372f0b1b963James DongdV0 DN D24.S32 860c1bc742181ded4930842b46e9507372f0b1b963James DongdV1 DN D25.S32 870c1bc742181ded4930842b46e9507372f0b1b963James Dong 880c1bc742181ded4930842b46e9507372f0b1b963James DongqHV QN Q11.S64 890c1bc742181ded4930842b46e9507372f0b1b963James DongqHV0 QN Q11.S32 900c1bc742181ded4930842b46e9507372f0b1b963James DongqHV1 QN Q12.S64 910c1bc742181ded4930842b46e9507372f0b1b963James Dong 920c1bc742181ded4930842b46e9507372f0b1b963James DongdHV00 DN D22.S32 930c1bc742181ded4930842b46e9507372f0b1b963James DongdHV01 DN D23.S32 940c1bc742181ded4930842b46e9507372f0b1b963James Dong 950c1bc742181ded4930842b46e9507372f0b1b963James DongdHV0 DN D22.S16[0] 960c1bc742181ded4930842b46e9507372f0b1b963James DongdHV1 DN D23.S16[0] 970c1bc742181ded4930842b46e9507372f0b1b963James DongdHV10 DN D24.S64 980c1bc742181ded4930842b46e9507372f0b1b963James DongdHV11 DN D25.S64 990c1bc742181ded4930842b46e9507372f0b1b963James Dong 1000c1bc742181ded4930842b46e9507372f0b1b963James DongqSum0 QN Q0.S16 1010c1bc742181ded4930842b46e9507372f0b1b963James DongqSum1 QN Q1.S16 1020c1bc742181ded4930842b46e9507372f0b1b963James Dong 1030c1bc742181ded4930842b46e9507372f0b1b963James DongdOut0 DN D6.U8 1040c1bc742181ded4930842b46e9507372f0b1b963James DongdOut1 DN D7.U8 1050c1bc742181ded4930842b46e9507372f0b1b963James Dong 1060c1bc742181ded4930842b46e9507372f0b1b963James DongdLeft0 DN D2.U8 1070c1bc742181ded4930842b46e9507372f0b1b963James DongdLeft1 DN D3.U8 1080c1bc742181ded4930842b46e9507372f0b1b963James DongqConst QN Q13.S16 1090c1bc742181ded4930842b46e9507372f0b1b963James Dong 1100c1bc742181ded4930842b46e9507372f0b1b963James DongdAbove0 DN D0.U8 1110c1bc742181ded4930842b46e9507372f0b1b963James DongdAbove1 DN D1.U8 1120c1bc742181ded4930842b46e9507372f0b1b963James Dong 1130c1bc742181ded4930842b46e9507372f0b1b963James DongdRevLeft64 DN D12.U64 1140c1bc742181ded4930842b46e9507372f0b1b963James DongdRevLeft DN D12.U8 1150c1bc742181ded4930842b46e9507372f0b1b963James DongdRevAbove64 DN D5.U64 1160c1bc742181ded4930842b46e9507372f0b1b963James DongdRevAbove DN D5.U8 1170c1bc742181ded4930842b46e9507372f0b1b963James DongqLeftDiff QN Q8.S16 1180c1bc742181ded4930842b46e9507372f0b1b963James DongdLeftDiff1 DN D17.S16 1190c1bc742181ded4930842b46e9507372f0b1b963James DongdLeftDiff64 DN D17.S64 1200c1bc742181ded4930842b46e9507372f0b1b963James DongqDiffLeft QN Q8.S16 1210c1bc742181ded4930842b46e9507372f0b1b963James DongqDiffAbove QN Q4.S16 1220c1bc742181ded4930842b46e9507372f0b1b963James DongdAboveDiff1 DN D9.S16 1230c1bc742181ded4930842b46e9507372f0b1b963James DongdAboveDiff64 DN D9.S64 1240c1bc742181ded4930842b46e9507372f0b1b963James DongqAboveDiff QN Q4.S16 1250c1bc742181ded4930842b46e9507372f0b1b963James Dong 1260c1bc742181ded4930842b46e9507372f0b1b963James DongdAboveLeft DN D4.U8 1270c1bc742181ded4930842b46e9507372f0b1b963James Dong 1280c1bc742181ded4930842b46e9507372f0b1b963James DongdDiffLeft0 DN D16.S16 1290c1bc742181ded4930842b46e9507372f0b1b963James DongdDiffLeft1 DN D17.S16 1300c1bc742181ded4930842b46e9507372f0b1b963James DongdDiffAbove0 DN D8.S16 1310c1bc742181ded4930842b46e9507372f0b1b963James DongdDiffAbove1 DN D9.S16 1320c1bc742181ded4930842b46e9507372f0b1b963James Dong 1330c1bc742181ded4930842b46e9507372f0b1b963James DongqLeft15minus0 QN Q7.S16 1340c1bc742181ded4930842b46e9507372f0b1b963James DongdLeft15minus0 DN D14.S16 1350c1bc742181ded4930842b46e9507372f0b1b963James DongqAbove15minus0 QN Q3.S16 1360c1bc742181ded4930842b46e9507372f0b1b963James DongdAbove15minus0 DN D6.S16 1370c1bc742181ded4930842b46e9507372f0b1b963James Dong 1380c1bc742181ded4930842b46e9507372f0b1b963James DongqMultiplier QN Q10.S16 1390c1bc742181ded4930842b46e9507372f0b1b963James DongqMultiplier0 QN Q10.S16 1400c1bc742181ded4930842b46e9507372f0b1b963James DongqMultiplier1 QN Q12.S16 1410c1bc742181ded4930842b46e9507372f0b1b963James DongdMultiplier0 DN D20.S16 1420c1bc742181ded4930842b46e9507372f0b1b963James DongdMultiplier1 DN D21.S16 1430c1bc742181ded4930842b46e9507372f0b1b963James Dong 1440c1bc742181ded4930842b46e9507372f0b1b963James DongdBPlusCMult7 DN D1.S64 1450c1bc742181ded4930842b46e9507372f0b1b963James DongdBPlusCMult7S16 DN D1.S16 1460c1bc742181ded4930842b46e9507372f0b1b963James Dong 1470c1bc742181ded4930842b46e9507372f0b1b963James DongqTmp QN Q0.U8 1480c1bc742181ded4930842b46e9507372f0b1b963James Dong 1490c1bc742181ded4930842b46e9507372f0b1b963James Dong;//-------------------------------------------- 1500c1bc742181ded4930842b46e9507372f0b1b963James Dong;// Declare input registers 1510c1bc742181ded4930842b46e9507372f0b1b963James Dong;//-------------------------------------------- 1520c1bc742181ded4930842b46e9507372f0b1b963James DongpSrcLeft RN 0 ;// input pointer 1530c1bc742181ded4930842b46e9507372f0b1b963James DongpSrcAbove RN 1 ;// input pointer 1540c1bc742181ded4930842b46e9507372f0b1b963James DongpSrcAboveLeft RN 2 ;// input pointer 1550c1bc742181ded4930842b46e9507372f0b1b963James DongpDst RN 3 ;// output pointer 1560c1bc742181ded4930842b46e9507372f0b1b963James DongleftStep RN 4 ;// input variable 1570c1bc742181ded4930842b46e9507372f0b1b963James DongdstStep RN 5 ;// input variable 1580c1bc742181ded4930842b46e9507372f0b1b963James DongpredMode RN 6 ;// input variable 1590c1bc742181ded4930842b46e9507372f0b1b963James Dongavailability RN 7 ;// input variable 1600c1bc742181ded4930842b46e9507372f0b1b963James Dong 1610c1bc742181ded4930842b46e9507372f0b1b963James DongpTmp RN 8 1620c1bc742181ded4930842b46e9507372f0b1b963James Dongstep RN 10 1630c1bc742181ded4930842b46e9507372f0b1b963James DongpTmp2 RN 11 1640c1bc742181ded4930842b46e9507372f0b1b963James Dong 1650c1bc742181ded4930842b46e9507372f0b1b963James Dong;//----------------------------------------------------------------------------------------------- 1660c1bc742181ded4930842b46e9507372f0b1b963James Dong;// omxVCM4P10_PredictIntra_16x16 starts 1670c1bc742181ded4930842b46e9507372f0b1b963James Dong;//----------------------------------------------------------------------------------------------- 1680c1bc742181ded4930842b46e9507372f0b1b963James Dong 1690c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Write function header 1700c1bc742181ded4930842b46e9507372f0b1b963James Dong M_START omxVCM4P10_PredictIntra_16x16, r11, d15 1710c1bc742181ded4930842b46e9507372f0b1b963James Dong 1720c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Define stack arguments 1730c1bc742181ded4930842b46e9507372f0b1b963James Dong M_ARG LeftStep, 4 1740c1bc742181ded4930842b46e9507372f0b1b963James Dong M_ARG DstStep, 4 1750c1bc742181ded4930842b46e9507372f0b1b963James Dong M_ARG PredMode, 4 1760c1bc742181ded4930842b46e9507372f0b1b963James Dong M_ARG Availability, 4 1770c1bc742181ded4930842b46e9507372f0b1b963James Dong 1780c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// M_STALL ARM1136JS=4 1790c1bc742181ded4930842b46e9507372f0b1b963James Dong 1800c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR pTable,=armVCM4P10_pIndexTable16x16 ;// Load index table for switch case 1810c1bc742181ded4930842b46e9507372f0b1b963James Dong 1820c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Load argument from the stack 1830c1bc742181ded4930842b46e9507372f0b1b963James Dong M_LDR predMode, PredMode ;// Arg predMode loaded from stack to reg 1840c1bc742181ded4930842b46e9507372f0b1b963James Dong M_LDR leftStep, LeftStep ;// Arg leftStep loaded from stack to reg 1850c1bc742181ded4930842b46e9507372f0b1b963James Dong M_LDR dstStep, DstStep ;// Arg dstStep loaded from stack to reg 1860c1bc742181ded4930842b46e9507372f0b1b963James Dong M_LDR availability, Availability ;// Arg availability loaded from stack to reg 1870c1bc742181ded4930842b46e9507372f0b1b963James Dong 1880c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV y, #BLK_SIZE ;// Outer Loop Count 1890c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR pc, [pTable, predMode, LSL #2] ;// Branch to the case based on preMode 1900c1bc742181ded4930842b46e9507372f0b1b963James Dong 1910c1bc742181ded4930842b46e9507372f0b1b963James DongOMX_VC_16X16_VERT 1920c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 qAbove, [pSrcAbove] 1930c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD pTmp, pDst, dstStep 1940c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD step, dstStep, dstStep 1950c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pDst], step 1960c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pTmp], step 1970c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pDst], step 1980c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pTmp], step 1990c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pDst], step 2000c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pTmp], step 2010c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pDst], step 2020c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pTmp], step 2030c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pDst], step 2040c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pTmp], step 2050c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pDst], step 2060c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pTmp], step 2070c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pDst], step 2080c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pTmp], step 2090c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pDst] 2100c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qAbove, [pTmp] 2110c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV return, #OMX_Sts_NoErr ;// returnNoError 2120c1bc742181ded4930842b46e9507372f0b1b963James Dong M_EXIT 2130c1bc742181ded4930842b46e9507372f0b1b963James Dong 2140c1bc742181ded4930842b46e9507372f0b1b963James DongOMX_VC_16X16_HOR 2150c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD pTmp, pSrcLeft, leftStep 2160c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD leftStep, leftStep, leftStep 2170c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD pTmp2, pDst, dstStep 2180c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD dstStep, dstStep, dstStep 2190c1bc742181ded4930842b46e9507372f0b1b963James DongLoopHor 2200c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[]}, [pSrcLeft], leftStep 2210c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qTmp[]}, [pTmp], leftStep 2220c1bc742181ded4930842b46e9507372f0b1b963James Dong SUBS y, y, #8 2230c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qLeft, [pDst], dstStep 2240c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qTmp, [pTmp2], dstStep 2250c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[]}, [pSrcLeft], leftStep 2260c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qTmp[]}, [pTmp], leftStep 2270c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qLeft, [pDst], dstStep 2280c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qTmp, [pTmp2], dstStep 2290c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[]}, [pSrcLeft], leftStep 2300c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qTmp[]}, [pTmp], leftStep 2310c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qLeft, [pDst], dstStep 2320c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qTmp, [pTmp2], dstStep 2330c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[]}, [pSrcLeft], leftStep 2340c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qTmp[]}, [pTmp], leftStep 2350c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qLeft, [pDst], dstStep 2360c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qTmp, [pTmp2], dstStep 2370c1bc742181ded4930842b46e9507372f0b1b963James Dong 2380c1bc742181ded4930842b46e9507372f0b1b963James Dong BNE LoopHor ;// Loop for 16 times 2390c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV return, #OMX_Sts_NoErr 2400c1bc742181ded4930842b46e9507372f0b1b963James Dong M_EXIT 2410c1bc742181ded4930842b46e9507372f0b1b963James Dong 2420c1bc742181ded4930842b46e9507372f0b1b963James DongOMX_VC_16X16_DC 2430c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV count, #0 ;// count = 0 2440c1bc742181ded4930842b46e9507372f0b1b963James Dong TST availability, #OMX_VC_LEFT 2450c1bc742181ded4930842b46e9507372f0b1b963James Dong BEQ UpperOrNoneAvailable ;// Jump to Upper if not left 2460c1bc742181ded4930842b46e9507372f0b1b963James Dong 2470c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD pTmp, pSrcLeft, leftStep 2480c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD step, leftStep, leftStep 2490c1bc742181ded4930842b46e9507372f0b1b963James Dong 2500c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[0]}, [pSrcLeft],step 2510c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[1]}, [pTmp],step 2520c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[2]}, [pSrcLeft],step 2530c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[3]}, [pTmp],step 2540c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[4]}, [pSrcLeft],step 2550c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[5]}, [pTmp],step 2560c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[6]}, [pSrcLeft],step 2570c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[7]}, [pTmp],step 2580c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[8]}, [pSrcLeft],step 2590c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[9]}, [pTmp],step 2600c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[10]},[pSrcLeft],step 2610c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[11]},[pTmp],step 2620c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[12]},[pSrcLeft],step 2630c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[13]},[pTmp],step 2640c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[14]},[pSrcLeft],step 2650c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[15]},[pTmp] 2660c1bc742181ded4930842b46e9507372f0b1b963James Dong 2670c1bc742181ded4930842b46e9507372f0b1b963James Dong VPADDL qSum8, qLeft 2680c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, #1 2690c1bc742181ded4930842b46e9507372f0b1b963James Dong VPADD dSum4, dSum80, dSum81 2700c1bc742181ded4930842b46e9507372f0b1b963James Dong VPADDL dSum2, dSum4 2710c1bc742181ded4930842b46e9507372f0b1b963James Dong VPADDL dSumLeft, dSum2 2720c1bc742181ded4930842b46e9507372f0b1b963James Dong VRSHR dSum, dSumLeft, #4 2730c1bc742181ded4930842b46e9507372f0b1b963James Dong 2740c1bc742181ded4930842b46e9507372f0b1b963James DongUpperOrNoneAvailable 2750c1bc742181ded4930842b46e9507372f0b1b963James Dong TST availability, #OMX_VC_UPPER ;// if(availability & #OMX_VC_UPPER) 2760c1bc742181ded4930842b46e9507372f0b1b963James Dong BEQ BothOrNoneAvailable ;// Jump to Left if not upper 2770c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 qAbove, [pSrcAbove] 2780c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, #1 ;// if upper inc count by 1 2790c1bc742181ded4930842b46e9507372f0b1b963James Dong VPADDL qSum8, qAbove 2800c1bc742181ded4930842b46e9507372f0b1b963James Dong VPADD dSum4, dSum80, dSum81 2810c1bc742181ded4930842b46e9507372f0b1b963James Dong VPADDL dSum2, dSum4 2820c1bc742181ded4930842b46e9507372f0b1b963James Dong VPADDL dSumAbove, dSum2 2830c1bc742181ded4930842b46e9507372f0b1b963James Dong VRSHR dSum, dSumAbove, #4 2840c1bc742181ded4930842b46e9507372f0b1b963James Dong 2850c1bc742181ded4930842b46e9507372f0b1b963James DongBothOrNoneAvailable 2860c1bc742181ded4930842b46e9507372f0b1b963James Dong CMP count, #2 ;// check if both available 2870c1bc742181ded4930842b46e9507372f0b1b963James Dong BNE NoneAvailable 2880c1bc742181ded4930842b46e9507372f0b1b963James Dong VADD dSum, dSumAbove, dSumLeft 2890c1bc742181ded4930842b46e9507372f0b1b963James Dong VRSHR dSum, dSum, #5 2900c1bc742181ded4930842b46e9507372f0b1b963James Dong 2910c1bc742181ded4930842b46e9507372f0b1b963James Dong 2920c1bc742181ded4930842b46e9507372f0b1b963James DongNoneAvailable 2930c1bc742181ded4930842b46e9507372f0b1b963James Dong VDUP qOut, dSum0 2940c1bc742181ded4930842b46e9507372f0b1b963James Dong CMP count, #0 ;// check if none available 2950c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD pTmp, pDst, dstStep 2960c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD step, dstStep, dstStep 2970c1bc742181ded4930842b46e9507372f0b1b963James Dong BNE LoopDC 2980c1bc742181ded4930842b46e9507372f0b1b963James Dong VMOV qOut, #128 2990c1bc742181ded4930842b46e9507372f0b1b963James DongLoopDC 3000c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pDst], step 3010c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pTmp], step 3020c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pDst], step 3030c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pTmp], step 3040c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pDst], step 3050c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pTmp], step 3060c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pDst], step 3070c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pTmp], step 3080c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pDst], step 3090c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pTmp], step 3100c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pDst], step 3110c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pTmp], step 3120c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pDst], step 3130c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pTmp], step 3140c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pDst], step 3150c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut, [pTmp], step 3160c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV return, #OMX_Sts_NoErr 3170c1bc742181ded4930842b46e9507372f0b1b963James Dong M_EXIT 3180c1bc742181ded4930842b46e9507372f0b1b963James Dong 3190c1bc742181ded4930842b46e9507372f0b1b963James DongOMX_VC_16X16_PLANE 3200c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR pMultTable, =armVCM4P10_MultiplierTable16x16 3210c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 qAbove, [pSrcAbove] ;// pSrcAbove[x] :0<= x <= 7 3220c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 dAboveLeft[0],[pSrcAboveLeft] 3230c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD pTmp, pSrcLeft, leftStep 3240c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD step, leftStep, leftStep 3250c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[0]}, [pSrcLeft],step 3260c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[1]}, [pTmp],step 3270c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[2]}, [pSrcLeft],step 3280c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[3]}, [pTmp],step 3290c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[4]}, [pSrcLeft],step 3300c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[5]}, [pTmp],step 3310c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[6]}, [pSrcLeft],step 3320c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[7]}, [pTmp],step 3330c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[8]}, [pSrcLeft],step 3340c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[9]}, [pTmp],step 3350c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[10]}, [pSrcLeft],step 3360c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[11]}, [pTmp],step 3370c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[12]}, [pSrcLeft],step 3380c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[13]}, [pTmp],step 3390c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[14]}, [pSrcLeft],step 3400c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 {qLeft[15]}, [pTmp] 3410c1bc742181ded4930842b46e9507372f0b1b963James Dong 3420c1bc742181ded4930842b46e9507372f0b1b963James Dong VREV64 dRevAbove, dAbove1 ;// pSrcAbove[15:14:13:12:11:10:9:8] 3430c1bc742181ded4930842b46e9507372f0b1b963James Dong VSUBL qAbove15minus0, dRevAbove, dAboveLeft ;// qAbove7minus0[0] = pSrcAbove[15] - pSrcAboveLeft[0] 3440c1bc742181ded4930842b46e9507372f0b1b963James Dong VSHR dRevAbove64, dRevAbove64, #8 ;// pSrcAbove[14:13:12:11:10:9:8:X] 3450c1bc742181ded4930842b46e9507372f0b1b963James Dong VSUBL qAboveDiff, dRevAbove, dAbove0 3460c1bc742181ded4930842b46e9507372f0b1b963James Dong 3470c1bc742181ded4930842b46e9507372f0b1b963James Dong VSHL dAboveDiff64, dAboveDiff64, #16 3480c1bc742181ded4930842b46e9507372f0b1b963James Dong VEXT dDiffAbove1, dAboveDiff1, dAbove15minus0, #1 3490c1bc742181ded4930842b46e9507372f0b1b963James Dong 3500c1bc742181ded4930842b46e9507372f0b1b963James Dong VREV64 dRevLeft,dLeft1 ;// pSrcLeft[15:14:13:12:11:10:9:8] 3510c1bc742181ded4930842b46e9507372f0b1b963James Dong VSUBL qLeft15minus0,dRevLeft, dAboveLeft ;// qAbove7minus0[0] = pSrcLeft[7] - pSrcAboveLeft[0] 3520c1bc742181ded4930842b46e9507372f0b1b963James Dong VSHR dRevLeft64, dRevLeft64, #8 ;// pSrcLeft[14:13:12:11:10:9:8:X] 3530c1bc742181ded4930842b46e9507372f0b1b963James Dong VSUBL qLeftDiff,dRevLeft, dLeft0 3540c1bc742181ded4930842b46e9507372f0b1b963James Dong 3550c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Multiplier = [8|1|2|...|6|7] 3560c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 qMultiplier, [pMultTable]! 3570c1bc742181ded4930842b46e9507372f0b1b963James Dong 3580c1bc742181ded4930842b46e9507372f0b1b963James Dong VSHL dLeftDiff64, dLeftDiff64, #16 3590c1bc742181ded4930842b46e9507372f0b1b963James Dong VEXT dDiffLeft1, dLeftDiff1, dLeft15minus0, #1 3600c1bc742181ded4930842b46e9507372f0b1b963James Dong 3610c1bc742181ded4930842b46e9507372f0b1b963James Dong VMULL qH,dDiffAbove0, dMultiplier0 3620c1bc742181ded4930842b46e9507372f0b1b963James Dong VMULL qV,dDiffLeft0, dMultiplier0 3630c1bc742181ded4930842b46e9507372f0b1b963James Dong VMLAL qH,dDiffAbove1, dMultiplier1 3640c1bc742181ded4930842b46e9507372f0b1b963James Dong VMLAL qV,dDiffLeft1, dMultiplier1 3650c1bc742181ded4930842b46e9507372f0b1b963James Dong 3660c1bc742181ded4930842b46e9507372f0b1b963James Dong VPADD dHV00,dH1,dH0 3670c1bc742181ded4930842b46e9507372f0b1b963James Dong VPADD dHV01,dV1,dV0 3680c1bc742181ded4930842b46e9507372f0b1b963James Dong VPADDL qHV, qHV0 3690c1bc742181ded4930842b46e9507372f0b1b963James Dong VSHL qHV1,qHV,#2 3700c1bc742181ded4930842b46e9507372f0b1b963James Dong VADD qHV,qHV,qHV1 3710c1bc742181ded4930842b46e9507372f0b1b963James Dong 3720c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// HV = [c = ((5*V+32)>>6) | b = ((5*H+32)>>6)] 3730c1bc742181ded4930842b46e9507372f0b1b963James Dong VRSHR qHV,qHV,#6 3740c1bc742181ded4930842b46e9507372f0b1b963James Dong 3750c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// HV1 = [c*7|b*7] 3760c1bc742181ded4930842b46e9507372f0b1b963James Dong VSHL qHV1,qHV,#3 3770c1bc742181ded4930842b46e9507372f0b1b963James Dong VSUB qHV1,qHV1,qHV 3780c1bc742181ded4930842b46e9507372f0b1b963James Dong 3790c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Multiplier1 = [0|1|2|...|7] 3800c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 qMultiplier0, [pMultTable]! 3810c1bc742181ded4930842b46e9507372f0b1b963James Dong VDUP qB, dHV0 3820c1bc742181ded4930842b46e9507372f0b1b963James Dong VDUP qC, dHV1 3830c1bc742181ded4930842b46e9507372f0b1b963James Dong 3840c1bc742181ded4930842b46e9507372f0b1b963James Dong VADDL qA,dAbove1,dLeft1 3850c1bc742181ded4930842b46e9507372f0b1b963James Dong VSHL qA,qA, #4 3860c1bc742181ded4930842b46e9507372f0b1b963James Dong VDUP qA,dA1[3] 3870c1bc742181ded4930842b46e9507372f0b1b963James Dong VADD dBPlusCMult7, dHV10, dHV11 3880c1bc742181ded4930842b46e9507372f0b1b963James Dong 3890c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Multiplier1 = [8|9|10|...|15] 3900c1bc742181ded4930842b46e9507372f0b1b963James Dong VLD1 qMultiplier1, [pMultTable] 3910c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Const = a - 7*(b+c) 3920c1bc742181ded4930842b46e9507372f0b1b963James Dong VDUP qConst, dBPlusCMult7S16[0] 3930c1bc742181ded4930842b46e9507372f0b1b963James Dong VSUB qConst, qA, qConst 3940c1bc742181ded4930842b46e9507372f0b1b963James Dong 3950c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// B0 = [0*b|1*b|2*b|3*b|......|7*b] 3960c1bc742181ded4930842b46e9507372f0b1b963James Dong VMUL qB0,qB,qMultiplier0 3970c1bc742181ded4930842b46e9507372f0b1b963James Dong 3980c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// B0 = [8*b|9*b|10*b|11*b|....|15*b] 3990c1bc742181ded4930842b46e9507372f0b1b963James Dong VMUL qB1,qB,qMultiplier1 4000c1bc742181ded4930842b46e9507372f0b1b963James Dong 4010c1bc742181ded4930842b46e9507372f0b1b963James Dong VADD qSum0, qB0, qConst 4020c1bc742181ded4930842b46e9507372f0b1b963James Dong VADD qSum1, qB1, qConst 4030c1bc742181ded4930842b46e9507372f0b1b963James Dong 4040c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Loops for 16 times 4050c1bc742181ded4930842b46e9507372f0b1b963James DongLoopPlane 4060c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// (b*x + c*y + C)>>5 4070c1bc742181ded4930842b46e9507372f0b1b963James Dong VQRSHRUN dOut0, qSum0,#5 4080c1bc742181ded4930842b46e9507372f0b1b963James Dong VQRSHRUN dOut1, qSum1,#5 4090c1bc742181ded4930842b46e9507372f0b1b963James Dong SUBS y, y, #1 4100c1bc742181ded4930842b46e9507372f0b1b963James Dong VST1 qOut,[pDst],dstStep 4110c1bc742181ded4930842b46e9507372f0b1b963James Dong VADD qSum0,qSum0,qC 4120c1bc742181ded4930842b46e9507372f0b1b963James Dong VADD qSum1,qSum1,qC 4130c1bc742181ded4930842b46e9507372f0b1b963James Dong BNE LoopPlane 4140c1bc742181ded4930842b46e9507372f0b1b963James Dong 4150c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV return, #OMX_Sts_NoErr 4160c1bc742181ded4930842b46e9507372f0b1b963James Dong 4170c1bc742181ded4930842b46e9507372f0b1b963James Dong M_END 4180c1bc742181ded4930842b46e9507372f0b1b963James Dong 4190c1bc742181ded4930842b46e9507372f0b1b963James Dong ENDIF ;// CortexA8 4200c1bc742181ded4930842b46e9507372f0b1b963James Dong 4210c1bc742181ded4930842b46e9507372f0b1b963James Dong END 4220c1bc742181ded4930842b46e9507372f0b1b963James Dong;----------------------------------------------------------------------------------------------- 4230c1bc742181ded4930842b46e9507372f0b1b963James Dong; omxVCM4P10_PredictIntra_16x16 ends 4240c1bc742181ded4930842b46e9507372f0b1b963James Dong;----------------------------------------------------------------------------------------------- 425