10c1bc742181ded4930842b46e9507372f0b1b963James Dong; Copyright (C) 2009 The Android Open Source Project 20c1bc742181ded4930842b46e9507372f0b1b963James Dong; 30c1bc742181ded4930842b46e9507372f0b1b963James Dong; Licensed under the Apache License, Version 2.0 (the "License"); 40c1bc742181ded4930842b46e9507372f0b1b963James Dong; you may not use this file except in compliance with the License. 50c1bc742181ded4930842b46e9507372f0b1b963James Dong; You may obtain a copy of the License at 60c1bc742181ded4930842b46e9507372f0b1b963James Dong; 70c1bc742181ded4930842b46e9507372f0b1b963James Dong; http://www.apache.org/licenses/LICENSE-2.0 80c1bc742181ded4930842b46e9507372f0b1b963James Dong; 90c1bc742181ded4930842b46e9507372f0b1b963James Dong; Unless required by applicable law or agreed to in writing, software 100c1bc742181ded4930842b46e9507372f0b1b963James Dong; distributed under the License is distributed on an "AS IS" BASIS, 110c1bc742181ded4930842b46e9507372f0b1b963James Dong; WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 120c1bc742181ded4930842b46e9507372f0b1b963James Dong; See the License for the specific language governing permissions and 130c1bc742181ded4930842b46e9507372f0b1b963James Dong; limitations under the License. 140c1bc742181ded4930842b46e9507372f0b1b963James Dong 150c1bc742181ded4930842b46e9507372f0b1b963James Dong;------------------------------------------------------------------------------- 160c1bc742181ded4930842b46e9507372f0b1b963James Dong;-- 170c1bc742181ded4930842b46e9507372f0b1b963James Dong;-- Abstract : ARMv6 optimized version of h264bsdInterpolateChromaVer function 180c1bc742181ded4930842b46e9507372f0b1b963James Dong;-- 190c1bc742181ded4930842b46e9507372f0b1b963James Dong;------------------------------------------------------------------------------- 200c1bc742181ded4930842b46e9507372f0b1b963James Dong 210c1bc742181ded4930842b46e9507372f0b1b963James Dong 220c1bc742181ded4930842b46e9507372f0b1b963James Dong IF :DEF: H264DEC_WINASM 230c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// We dont use REQUIRE8 and PRESERVE8 for winasm 240c1bc742181ded4930842b46e9507372f0b1b963James Dong ELSE 250c1bc742181ded4930842b46e9507372f0b1b963James Dong REQUIRE8 260c1bc742181ded4930842b46e9507372f0b1b963James Dong PRESERVE8 270c1bc742181ded4930842b46e9507372f0b1b963James Dong ENDIF 280c1bc742181ded4930842b46e9507372f0b1b963James Dong 290c1bc742181ded4930842b46e9507372f0b1b963James Dong AREA |.text|, CODE 300c1bc742181ded4930842b46e9507372f0b1b963James Dong 310c1bc742181ded4930842b46e9507372f0b1b963James Dong;// h264bsdInterpolateChromaVer register allocation 320c1bc742181ded4930842b46e9507372f0b1b963James Dong 330c1bc742181ded4930842b46e9507372f0b1b963James Dongref RN 0 340c1bc742181ded4930842b46e9507372f0b1b963James DongptrA RN 0 350c1bc742181ded4930842b46e9507372f0b1b963James Dong 360c1bc742181ded4930842b46e9507372f0b1b963James Dongmb RN 1 370c1bc742181ded4930842b46e9507372f0b1b963James Dongblock RN 1 380c1bc742181ded4930842b46e9507372f0b1b963James Dong 390c1bc742181ded4930842b46e9507372f0b1b963James Dongx0 RN 2 400c1bc742181ded4930842b46e9507372f0b1b963James Dongcount RN 2 410c1bc742181ded4930842b46e9507372f0b1b963James Dong 420c1bc742181ded4930842b46e9507372f0b1b963James Dongy0 RN 3 430c1bc742181ded4930842b46e9507372f0b1b963James DongvalY RN 3 440c1bc742181ded4930842b46e9507372f0b1b963James Dong 450c1bc742181ded4930842b46e9507372f0b1b963James Dongwidth RN 4 460c1bc742181ded4930842b46e9507372f0b1b963James Dong 470c1bc742181ded4930842b46e9507372f0b1b963James Dongheight RN 5 480c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp7 RN 5 490c1bc742181ded4930842b46e9507372f0b1b963James Dong 500c1bc742181ded4930842b46e9507372f0b1b963James DongchrPW RN 6 510c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp8 RN 6 520c1bc742181ded4930842b46e9507372f0b1b963James Dong 530c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp1 RN 7 540c1bc742181ded4930842b46e9507372f0b1b963James Dong 550c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp2 RN 8 560c1bc742181ded4930842b46e9507372f0b1b963James Dong 570c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp3 RN 9 580c1bc742181ded4930842b46e9507372f0b1b963James Dong 590c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp4 RN 10 600c1bc742181ded4930842b46e9507372f0b1b963James Dong 610c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp5 RN 11 620c1bc742181ded4930842b46e9507372f0b1b963James DongchrPH RN 11 630c1bc742181ded4930842b46e9507372f0b1b963James Dong 640c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp6 RN 12 650c1bc742181ded4930842b46e9507372f0b1b963James Dong 660c1bc742181ded4930842b46e9507372f0b1b963James Dongc32 RN 14 670c1bc742181ded4930842b46e9507372f0b1b963James DongyFrac RN 14 680c1bc742181ded4930842b46e9507372f0b1b963James Dong 690c1bc742181ded4930842b46e9507372f0b1b963James Dong;// Function exports and imports 700c1bc742181ded4930842b46e9507372f0b1b963James Dong 710c1bc742181ded4930842b46e9507372f0b1b963James Dong IMPORT h264bsdFillBlock 720c1bc742181ded4930842b46e9507372f0b1b963James Dong 730c1bc742181ded4930842b46e9507372f0b1b963James Dong EXPORT h264bsdInterpolateChromaVer 740c1bc742181ded4930842b46e9507372f0b1b963James Dong 750c1bc742181ded4930842b46e9507372f0b1b963James Dong;// Function arguments 760c1bc742181ded4930842b46e9507372f0b1b963James Dong;// 770c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u8 *ref, : 0xc4 780c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u8 *predPartChroma, : 0xc8 790c1bc742181ded4930842b46e9507372f0b1b963James Dong;// i32 x0, : 0xcc 800c1bc742181ded4930842b46e9507372f0b1b963James Dong;// i32 y0, : 0xd0 810c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u32 width, : 0xf8 820c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u32 height, : 0xfc 830c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u32 yFrac, : 0x100 840c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u32 chromaPartWidth, : 0x104 850c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u32 chromaPartHeight : 0x108 860c1bc742181ded4930842b46e9507372f0b1b963James Dong 870c1bc742181ded4930842b46e9507372f0b1b963James Dongh264bsdInterpolateChromaVer 880c1bc742181ded4930842b46e9507372f0b1b963James Dong STMFD sp!, {r0-r11,lr} 890c1bc742181ded4930842b46e9507372f0b1b963James Dong SUB sp, sp, #0xc4 900c1bc742181ded4930842b46e9507372f0b1b963James Dong 910c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR chrPW, [sp, #0x104] ;// chromaPartWidth 920c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR width, [sp, #0xf8] ;// width 930c1bc742181ded4930842b46e9507372f0b1b963James Dong CMP x0, #0 940c1bc742181ded4930842b46e9507372f0b1b963James Dong BLT do_fill 950c1bc742181ded4930842b46e9507372f0b1b963James Dong 960c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD tmp1, x0, chrPW ;// tmp1 = x0+ chromaPartWidth 970c1bc742181ded4930842b46e9507372f0b1b963James Dong CMP tmp1, width ;// x0+chromaPartWidth > width 980c1bc742181ded4930842b46e9507372f0b1b963James Dong BHI do_fill 990c1bc742181ded4930842b46e9507372f0b1b963James Dong 1000c1bc742181ded4930842b46e9507372f0b1b963James Dong CMP y0, #0 1010c1bc742181ded4930842b46e9507372f0b1b963James Dong BLT do_fill 1020c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR chrPH, [sp, #0x108] ;// chromaPartHeight 1030c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR height, [sp, #0xfc] ;// height 1040c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD tmp1, y0, chrPH ;// tmp1 = y0 + chromaPartHeight 1050c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD tmp1, tmp1, #1 ;// tmp1 = y0 + chromaPartHeight + 1 1060c1bc742181ded4930842b46e9507372f0b1b963James Dong CMP tmp1, height 1070c1bc742181ded4930842b46e9507372f0b1b963James Dong BLS skip_fill 1080c1bc742181ded4930842b46e9507372f0b1b963James Dong 1090c1bc742181ded4930842b46e9507372f0b1b963James Dongdo_fill 1100c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR chrPH, [sp, #0x108] ;// chromaPartHeight 1110c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR height, [sp, #0xfc] ;// height 1120c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD tmp1, chrPH, #1 ;// tmp1 = chromaPartHeight+1 1130c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp2, chrPW ;// tmp2 = chromaPartWidth 1140c1bc742181ded4930842b46e9507372f0b1b963James Dong STMIA sp,{width,height,chrPW,tmp1,tmp2} 1150c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD block, sp, #0x1c ;// block 1160c1bc742181ded4930842b46e9507372f0b1b963James Dong BL h264bsdFillBlock 1170c1bc742181ded4930842b46e9507372f0b1b963James Dong 1180c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR x0, [sp, #0xcc] 1190c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR y0, [sp, #0xd0] 1200c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR ref, [sp, #0xc4] ;// ref 1210c1bc742181ded4930842b46e9507372f0b1b963James Dong STMIA sp,{width,height,chrPW,tmp1,tmp2} 1220c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD block, sp, #0x1c ;// block 1230c1bc742181ded4930842b46e9507372f0b1b963James Dong MLA ref, height, width, ref ;// ref += width * height; 1240c1bc742181ded4930842b46e9507372f0b1b963James Dong MLA block, chrPW, tmp1, block;// block + (chromaPW)*(chromaPH+1) 1250c1bc742181ded4930842b46e9507372f0b1b963James Dong BL h264bsdFillBlock 1260c1bc742181ded4930842b46e9507372f0b1b963James Dong 1270c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV x0, #0 ;// x0 = 0 1280c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV y0, #0 ;// y0 = 0 1290c1bc742181ded4930842b46e9507372f0b1b963James Dong STR x0, [sp, #0xcc] 1300c1bc742181ded4930842b46e9507372f0b1b963James Dong STR y0, [sp, #0xd0] 1310c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD ref, sp, #0x1c ;// ref = block 1320c1bc742181ded4930842b46e9507372f0b1b963James Dong STR ref, [sp, #0xc4] ;// ref 1330c1bc742181ded4930842b46e9507372f0b1b963James Dong 1340c1bc742181ded4930842b46e9507372f0b1b963James Dong STR tmp1, [sp, #0xfc] ;// height 1350c1bc742181ded4930842b46e9507372f0b1b963James Dong STR chrPW, [sp, #0xf8] ;// width 1360c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV width, chrPW 1370c1bc742181ded4930842b46e9507372f0b1b963James Dong 1380c1bc742181ded4930842b46e9507372f0b1b963James Dongskip_fill 1390c1bc742181ded4930842b46e9507372f0b1b963James Dong MLA tmp3, y0, width, x0 ;// tmp3 = y0*width+x0 1400c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR yFrac, [sp, #0x100] ;// yFrac 1410c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD ptrA, ref, tmp3 ;// ptrA = ref + y0*width+x0 1420c1bc742181ded4930842b46e9507372f0b1b963James Dong RSB valY, yFrac, #8 ;// valY = 8-yFrac 1430c1bc742181ded4930842b46e9507372f0b1b963James Dong 1440c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR mb, [sp, #0xc8] ;// predPartChroma 1450c1bc742181ded4930842b46e9507372f0b1b963James Dong 1460c1bc742181ded4930842b46e9507372f0b1b963James Dong 1470c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// pack values to count register 1480c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// [31:28] loop_x (chromaPartWidth-1) 1490c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// [27:24] loop_y (chromaPartHeight-1) 1500c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// [23:20] chromaPartWidth-1 1510c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// [19:16] chromaPartHeight-1 1520c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// [15:00] nothing 1530c1bc742181ded4930842b46e9507372f0b1b963James Dong 1540c1bc742181ded4930842b46e9507372f0b1b963James Dong SUB tmp2, chrPH, #1 ;// chromaPartHeight-1 1550c1bc742181ded4930842b46e9507372f0b1b963James Dong SUB tmp1, chrPW, #1 ;// chromaPartWidth-1 1560c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, tmp2, LSL #16 ;// chromaPartHeight-1 1570c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, tmp2, LSL #24 ;// loop_y 1580c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, tmp1, LSL #20 ;// chromaPartWidth-1 1590c1bc742181ded4930842b46e9507372f0b1b963James Dong AND tmp2, count, #0x00F00000 ;// loop_x 1600c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT valY, valY, yFrac, LSL #16 ;// |yFrac|valY | 1610c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV valY, valY, LSL #3 ;// multiply by 8 in advance 1620c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV c32, #32 1630c1bc742181ded4930842b46e9507372f0b1b963James Dong 1640c1bc742181ded4930842b46e9507372f0b1b963James Dong 1650c1bc742181ded4930842b46e9507372f0b1b963James Dong ;/////////////////////////////////////////////////////////////////////////// 1660c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Cb 1670c1bc742181ded4930842b46e9507372f0b1b963James Dong ;/////////////////////////////////////////////////////////////////////////// 1680c1bc742181ded4930842b46e9507372f0b1b963James Dong 1690c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// 2x2 pels per iteration 1700c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// bilinear vertical interpolation 1710c1bc742181ded4930842b46e9507372f0b1b963James Dong 1720c1bc742181ded4930842b46e9507372f0b1b963James Dongloop1_y 1730c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, tmp2, LSL #8 1740c1bc742181ded4930842b46e9507372f0b1b963James Dongloop1_x 1750c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Process 2x2 block 1760c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp2, [ptrA,width] ;// 2 row, 1 col 1770c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp3, [ptrA,width, LSL #1] ;// 3 row, 1 col 1780c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp1, [ptrA],#1 ;// 1 row, 1 col 1790c1bc742181ded4930842b46e9507372f0b1b963James Dong 1800c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp5, [ptrA,width] ;// 2 row, 2 col 1810c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp6, [ptrA,width, LSL #1] ;// 3 row, 2 col 1820c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp4, [ptrA],#1 ;// 1 row, 2 col 1830c1bc742181ded4930842b46e9507372f0b1b963James Dong 1840c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp1, tmp1, tmp2, LSL #16 ;// |B|A| 1850c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp2, tmp2, tmp3, LSL #16 ;// |C|B| 1860c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp4, tmp4, tmp5, LSL #16 ;// |B|A| 1870c1bc742181ded4930842b46e9507372f0b1b963James Dong 1880c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp7, tmp2, valY, c32 ;// multiply 1890c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp5, tmp5, tmp6, LSL #16 ;// |C|B| 1900c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp2, tmp1, valY, c32 ;// multiply 1910c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp8, tmp5, valY, c32 ;// multiply 1920c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp5, tmp4, valY, c32 ;// multiply 1930c1bc742181ded4930842b46e9507372f0b1b963James Dong 1940c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp7, tmp7, LSR #6 ;// scale down 1950c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp7, [mb,#8] ;// store row 2 col 1 1960c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp2, tmp2, LSR #6 ;// scale down 1970c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp2, [mb],#1 ;// store row 1 col 1 1980c1bc742181ded4930842b46e9507372f0b1b963James Dong 1990c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp8, tmp8, LSR #6 ;// scale down 2000c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp8, [mb,#8] ;// store row 2 col 2 2010c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp5, tmp5, LSR #6 ;// scale down 2020c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp5, [mb],#1 ;// store row 1 col 2 2030c1bc742181ded4930842b46e9507372f0b1b963James Dong 2040c1bc742181ded4930842b46e9507372f0b1b963James Dong 2050c1bc742181ded4930842b46e9507372f0b1b963James Dong SUBS count, count, #2<<28 2060c1bc742181ded4930842b46e9507372f0b1b963James Dong BCS loop1_x 2070c1bc742181ded4930842b46e9507372f0b1b963James Dong 2080c1bc742181ded4930842b46e9507372f0b1b963James Dong AND tmp2, count, #0x00F00000 2090c1bc742181ded4930842b46e9507372f0b1b963James Dong 2100c1bc742181ded4930842b46e9507372f0b1b963James Dong ADDS mb, mb, #16 2110c1bc742181ded4930842b46e9507372f0b1b963James Dong SBC mb, mb, tmp2, LSR #20 2120c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD ptrA, ptrA, width, LSL #1 2130c1bc742181ded4930842b46e9507372f0b1b963James Dong SBC ptrA, ptrA, tmp2, LSR #20 2140c1bc742181ded4930842b46e9507372f0b1b963James Dong 2150c1bc742181ded4930842b46e9507372f0b1b963James Dong ADDS count, count, #0xE << 24 2160c1bc742181ded4930842b46e9507372f0b1b963James Dong BGE loop1_y 2170c1bc742181ded4930842b46e9507372f0b1b963James Dong 2180c1bc742181ded4930842b46e9507372f0b1b963James Dong ;/////////////////////////////////////////////////////////////////////////// 2190c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Cr 2200c1bc742181ded4930842b46e9507372f0b1b963James Dong ;/////////////////////////////////////////////////////////////////////////// 2210c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR height, [sp,#0xfc] ;// height 2220c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR ref, [sp, #0xc4] ;// ref 2230c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR tmp1, [sp, #0xd0] ;// y0 2240c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR tmp2, [sp, #0xcc] ;// x0 2250c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR mb, [sp, #0xc8] ;// predPartChroma 2260c1bc742181ded4930842b46e9507372f0b1b963James Dong 2270c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD tmp1, height, tmp1 2280c1bc742181ded4930842b46e9507372f0b1b963James Dong MLA tmp3, tmp1, width, tmp2 2290c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD ptrA, ref, tmp3 2300c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD mb, mb, #64 2310c1bc742181ded4930842b46e9507372f0b1b963James Dong 2320c1bc742181ded4930842b46e9507372f0b1b963James Dong AND count, count, #0x00FFFFFF 2330c1bc742181ded4930842b46e9507372f0b1b963James Dong AND tmp1, count, #0x000F0000 2340c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, tmp1, LSL #8 2350c1bc742181ded4930842b46e9507372f0b1b963James Dong AND tmp2, count, #0x00F00000 2360c1bc742181ded4930842b46e9507372f0b1b963James Dong 2370c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// 2x2 pels per iteration 2380c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// bilinear vertical interpolation 2390c1bc742181ded4930842b46e9507372f0b1b963James Dongloop2_y 2400c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, tmp2, LSL #8 2410c1bc742181ded4930842b46e9507372f0b1b963James Dongloop2_x 2420c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Process 2x2 block 2430c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp2, [ptrA,width] ;// 2 row, 1 col 2440c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp3, [ptrA,width, LSL #1] ;// 3 row, 1 col 2450c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp1, [ptrA],#1 ;// 1 row, 1 col 2460c1bc742181ded4930842b46e9507372f0b1b963James Dong 2470c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp5, [ptrA,width] ;// 2 row, 2 col 2480c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp6, [ptrA,width, LSL #1] ;// 3 row, 2 col 2490c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp4, [ptrA],#1 ;// 1 row, 2 col 2500c1bc742181ded4930842b46e9507372f0b1b963James Dong 2510c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp1, tmp1, tmp2, LSL #16 ;// |B|A| 2520c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp2, tmp2, tmp3, LSL #16 ;// |C|B| 2530c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp4, tmp4, tmp5, LSL #16 ;// |B|A| 2540c1bc742181ded4930842b46e9507372f0b1b963James Dong 2550c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp7, tmp2, valY, c32 ;// multiply 2560c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp5, tmp5, tmp6, LSL #16 ;// |C|B| 2570c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp2, tmp1, valY, c32 ;// multiply 2580c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp8, tmp5, valY, c32 ;// multiply 2590c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp5, tmp4, valY, c32 ;// multiply 2600c1bc742181ded4930842b46e9507372f0b1b963James Dong 2610c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp7, tmp7, LSR #6 ;// scale down 2620c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp7, [mb,#8] ;// store row 2 col 1 2630c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp2, tmp2, LSR #6 ;// scale down 2640c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp2, [mb],#1 ;// store row 1 col 1 2650c1bc742181ded4930842b46e9507372f0b1b963James Dong 2660c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp8, tmp8, LSR #6 ;// scale down 2670c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp8, [mb,#8] ;// store row 2 col 2 2680c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp5, tmp5, LSR #6 ;// scale down 2690c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp5, [mb],#1 ;// store row 1 col 2 2700c1bc742181ded4930842b46e9507372f0b1b963James Dong 2710c1bc742181ded4930842b46e9507372f0b1b963James Dong 2720c1bc742181ded4930842b46e9507372f0b1b963James Dong SUBS count, count, #2<<28 2730c1bc742181ded4930842b46e9507372f0b1b963James Dong BCS loop2_x 2740c1bc742181ded4930842b46e9507372f0b1b963James Dong 2750c1bc742181ded4930842b46e9507372f0b1b963James Dong AND tmp2, count, #0x00F00000 2760c1bc742181ded4930842b46e9507372f0b1b963James Dong 2770c1bc742181ded4930842b46e9507372f0b1b963James Dong ADDS mb, mb, #16 2780c1bc742181ded4930842b46e9507372f0b1b963James Dong SBC mb, mb, tmp2, LSR #20 2790c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD ptrA, ptrA, width, LSL #1 2800c1bc742181ded4930842b46e9507372f0b1b963James Dong SBC ptrA, ptrA, tmp2, LSR #20 2810c1bc742181ded4930842b46e9507372f0b1b963James Dong 2820c1bc742181ded4930842b46e9507372f0b1b963James Dong ADDS count, count, #0xE << 24 2830c1bc742181ded4930842b46e9507372f0b1b963James Dong BGE loop2_y 2840c1bc742181ded4930842b46e9507372f0b1b963James Dong 2850c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD sp,sp,#0xd4 2860c1bc742181ded4930842b46e9507372f0b1b963James Dong LDMFD sp!, {r4-r11,pc} 2870c1bc742181ded4930842b46e9507372f0b1b963James Dong 2880c1bc742181ded4930842b46e9507372f0b1b963James Dong END 289