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 h264bsdInterpolateChromaHor 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 320c1bc742181ded4930842b46e9507372f0b1b963James Dong;// h264bsdInterpolateChromaHor register allocation 330c1bc742181ded4930842b46e9507372f0b1b963James Dong 340c1bc742181ded4930842b46e9507372f0b1b963James Dongref RN 0 350c1bc742181ded4930842b46e9507372f0b1b963James DongptrA RN 0 360c1bc742181ded4930842b46e9507372f0b1b963James Dong 370c1bc742181ded4930842b46e9507372f0b1b963James Dongmb RN 1 380c1bc742181ded4930842b46e9507372f0b1b963James Dongblock RN 1 390c1bc742181ded4930842b46e9507372f0b1b963James Dong 400c1bc742181ded4930842b46e9507372f0b1b963James Dongx0 RN 2 410c1bc742181ded4930842b46e9507372f0b1b963James Dongcount RN 2 420c1bc742181ded4930842b46e9507372f0b1b963James Dong 430c1bc742181ded4930842b46e9507372f0b1b963James Dongy0 RN 3 440c1bc742181ded4930842b46e9507372f0b1b963James DongvalX RN 3 450c1bc742181ded4930842b46e9507372f0b1b963James Dong 460c1bc742181ded4930842b46e9507372f0b1b963James Dongwidth RN 4 470c1bc742181ded4930842b46e9507372f0b1b963James Dong 480c1bc742181ded4930842b46e9507372f0b1b963James Dongheight RN 5 490c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp7 RN 5 500c1bc742181ded4930842b46e9507372f0b1b963James Dong 510c1bc742181ded4930842b46e9507372f0b1b963James DongchrPW RN 6 520c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp8 RN 6 530c1bc742181ded4930842b46e9507372f0b1b963James Dong 540c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp1 RN 7 550c1bc742181ded4930842b46e9507372f0b1b963James DongchrPH RN 7 560c1bc742181ded4930842b46e9507372f0b1b963James Dong 570c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp2 RN 8 580c1bc742181ded4930842b46e9507372f0b1b963James Dong 590c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp3 RN 9 600c1bc742181ded4930842b46e9507372f0b1b963James Dong 610c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp4 RN 10 620c1bc742181ded4930842b46e9507372f0b1b963James Dong 630c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp5 RN 11 640c1bc742181ded4930842b46e9507372f0b1b963James Dong 650c1bc742181ded4930842b46e9507372f0b1b963James Dongtmp6 RN 12 660c1bc742181ded4930842b46e9507372f0b1b963James Dong 670c1bc742181ded4930842b46e9507372f0b1b963James Dongc32 RN 14 680c1bc742181ded4930842b46e9507372f0b1b963James DongxFrac RN 14 690c1bc742181ded4930842b46e9507372f0b1b963James Dong 700c1bc742181ded4930842b46e9507372f0b1b963James Dong;// Function exports and imports 710c1bc742181ded4930842b46e9507372f0b1b963James Dong 720c1bc742181ded4930842b46e9507372f0b1b963James Dong IMPORT h264bsdFillBlock 730c1bc742181ded4930842b46e9507372f0b1b963James Dong 740c1bc742181ded4930842b46e9507372f0b1b963James Dong EXPORT h264bsdInterpolateChromaHor 750c1bc742181ded4930842b46e9507372f0b1b963James Dong 760c1bc742181ded4930842b46e9507372f0b1b963James Dong;// Function arguments 770c1bc742181ded4930842b46e9507372f0b1b963James Dong;// 780c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u8 *ref, : 0xc4 790c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u8 *predPartChroma, : 0xc8 800c1bc742181ded4930842b46e9507372f0b1b963James Dong;// i32 x0, : 0xcc 810c1bc742181ded4930842b46e9507372f0b1b963James Dong;// i32 y0, : 0xd0 820c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u32 width, : 0xf8 830c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u32 height, : 0xfc 840c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u32 xFrac, : 0x100 850c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u32 chromaPartWidth, : 0x104 860c1bc742181ded4930842b46e9507372f0b1b963James Dong;// u32 chromaPartHeight : 0x108 870c1bc742181ded4930842b46e9507372f0b1b963James Dong 880c1bc742181ded4930842b46e9507372f0b1b963James Dongh264bsdInterpolateChromaHor 890c1bc742181ded4930842b46e9507372f0b1b963James Dong STMFD sp!, {r0-r11,lr} 900c1bc742181ded4930842b46e9507372f0b1b963James Dong SUB sp, sp, #0xc4 910c1bc742181ded4930842b46e9507372f0b1b963James Dong 920c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR chrPW, [sp, #0x104] ;// chromaPartWidth 930c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR width, [sp, #0xf8] ;// width 940c1bc742181ded4930842b46e9507372f0b1b963James Dong CMP x0, #0 950c1bc742181ded4930842b46e9507372f0b1b963James Dong BLT do_fill 960c1bc742181ded4930842b46e9507372f0b1b963James Dong 970c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD tmp6, x0, chrPW ;// tmp6 = x0+ chromaPartWidth 980c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD tmp6, tmp6, #1 ;// tmp6 = x0 + chromaPartWidth + 1 990c1bc742181ded4930842b46e9507372f0b1b963James Dong CMP tmp6, width ;// x0+chromaPartWidth+1 > width 1000c1bc742181ded4930842b46e9507372f0b1b963James Dong BHI do_fill 1010c1bc742181ded4930842b46e9507372f0b1b963James Dong 1020c1bc742181ded4930842b46e9507372f0b1b963James Dong CMP y0, #0 1030c1bc742181ded4930842b46e9507372f0b1b963James Dong BLT do_fill 1040c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR chrPH, [sp, #0x108] ;// chromaPartHeight 1050c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR height, [sp, #0xfc] ;// height 1060c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD tmp6, y0, chrPH ;// tmp6 = y0 + chromaPartHeight 1070c1bc742181ded4930842b46e9507372f0b1b963James Dong CMP tmp6, height 1080c1bc742181ded4930842b46e9507372f0b1b963James Dong BLS skip_fill 1090c1bc742181ded4930842b46e9507372f0b1b963James Dong 1100c1bc742181ded4930842b46e9507372f0b1b963James Dongdo_fill 1110c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR chrPH, [sp, #0x108] ;// chromaPartHeight 1120c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR height, [sp, #0xfc] ;// height 1130c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD tmp8, chrPW, #1 ;// tmp8 = chromaPartWidth+1 1140c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp2, tmp8 ;// tmp2 = chromaPartWidth+1 1150c1bc742181ded4930842b46e9507372f0b1b963James Dong STMIA sp,{width,height,tmp8,chrPH,tmp2} 1160c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD block, sp, #0x1c ;// block 1170c1bc742181ded4930842b46e9507372f0b1b963James Dong BL h264bsdFillBlock 1180c1bc742181ded4930842b46e9507372f0b1b963James Dong 1190c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR x0, [sp, #0xcc] 1200c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR y0, [sp, #0xd0] 1210c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR ref, [sp, #0xc4] ;// ref 1220c1bc742181ded4930842b46e9507372f0b1b963James Dong STMIA sp,{width,height,tmp8,chrPH,tmp2} 1230c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD block, sp, #0x1c ;// block 1240c1bc742181ded4930842b46e9507372f0b1b963James Dong MLA ref, height, width, ref ;// ref += width * height; 1250c1bc742181ded4930842b46e9507372f0b1b963James Dong MLA block, chrPH, tmp8, block;// block + (chromaPH)*(chromaPW+1) 1260c1bc742181ded4930842b46e9507372f0b1b963James Dong BL h264bsdFillBlock 1270c1bc742181ded4930842b46e9507372f0b1b963James Dong 1280c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV x0, #0 ;// x0 = 0 1290c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV y0, #0 ;// y0 = 0 1300c1bc742181ded4930842b46e9507372f0b1b963James Dong STR x0, [sp, #0xcc] 1310c1bc742181ded4930842b46e9507372f0b1b963James Dong STR y0, [sp, #0xd0] 1320c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD ref, sp, #0x1c ;// ref = block 1330c1bc742181ded4930842b46e9507372f0b1b963James Dong STR ref, [sp, #0xc4] ;// ref 1340c1bc742181ded4930842b46e9507372f0b1b963James Dong 1350c1bc742181ded4930842b46e9507372f0b1b963James Dong STR chrPH, [sp, #0xfc] ;// height 1360c1bc742181ded4930842b46e9507372f0b1b963James Dong STR tmp8, [sp, #0xf8] ;// width 1370c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV width, tmp8 1380c1bc742181ded4930842b46e9507372f0b1b963James Dong SUB chrPW, chrPW, #1 1390c1bc742181ded4930842b46e9507372f0b1b963James Dong 1400c1bc742181ded4930842b46e9507372f0b1b963James Dongskip_fill 1410c1bc742181ded4930842b46e9507372f0b1b963James Dong MLA tmp3, y0, width, x0 ;// tmp3 = y0*width+x0 1420c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR xFrac, [sp, #0x100] ;// xFrac 1430c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD ptrA, ref, tmp3 ;// ptrA = ref + y0*width+x0 1440c1bc742181ded4930842b46e9507372f0b1b963James Dong RSB valX, xFrac, #8 ;// valX = 8-xFrac 1450c1bc742181ded4930842b46e9507372f0b1b963James Dong 1460c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR mb, [sp, #0xc8] ;// predPartChroma 1470c1bc742181ded4930842b46e9507372f0b1b963James Dong 1480c1bc742181ded4930842b46e9507372f0b1b963James Dong 1490c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// pack values to count register 1500c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// [31:28] loop_x (chromaPartWidth-1) 1510c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// [27:24] loop_y (chromaPartHeight-1) 1520c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// [23:20] chromaPartWidth-1 1530c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// [19:16] chromaPartHeight-1 1540c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// [15:00] nothing 1550c1bc742181ded4930842b46e9507372f0b1b963James Dong 1560c1bc742181ded4930842b46e9507372f0b1b963James Dong SUB tmp2, chrPH, #1 ;// chromaPartHeight-1 1570c1bc742181ded4930842b46e9507372f0b1b963James Dong SUB tmp1, chrPW, #1 ;// chromaPartWidth-1 1580c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, tmp2, LSL #16 ;// chromaPartHeight-1 1590c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, tmp2, LSL #24 ;// loop_y 1600c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, tmp1, LSL #20 ;// chromaPartWidth-1 1610c1bc742181ded4930842b46e9507372f0b1b963James Dong AND tmp2, count, #0x00F00000 ;// loop_x 1620c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT valX, valX, xFrac, LSL #16 ;// |xFrac|valX | 1630c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV valX, valX, LSL #3 ;// multiply by 8 in advance 1640c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV c32, #32 1650c1bc742181ded4930842b46e9507372f0b1b963James Dong 1660c1bc742181ded4930842b46e9507372f0b1b963James Dong 1670c1bc742181ded4930842b46e9507372f0b1b963James Dong ;/////////////////////////////////////////////////////////////////////////// 1680c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Cb 1690c1bc742181ded4930842b46e9507372f0b1b963James Dong ;/////////////////////////////////////////////////////////////////////////// 1700c1bc742181ded4930842b46e9507372f0b1b963James Dong 1710c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// 2x2 pels per iteration 1720c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// bilinear vertical interpolation 1730c1bc742181ded4930842b46e9507372f0b1b963James Dong 1740c1bc742181ded4930842b46e9507372f0b1b963James Dongloop1_y 1750c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, tmp2, LSL #8 1760c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp1, [ptrA, width] 1770c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp2, [ptrA], #1 1780c1bc742181ded4930842b46e9507372f0b1b963James Dong 1790c1bc742181ded4930842b46e9507372f0b1b963James Dongloop1_x 1800c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp3, [ptrA, width] 1810c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp4, [ptrA], #1 1820c1bc742181ded4930842b46e9507372f0b1b963James Dong 1830c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp5, tmp1, tmp3, LSL #16 1840c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp6, tmp2, tmp4, LSL #16 1850c1bc742181ded4930842b46e9507372f0b1b963James Dong 1860c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp1, [ptrA, width] 1870c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp2, [ptrA], #1 1880c1bc742181ded4930842b46e9507372f0b1b963James Dong 1890c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp5, tmp5, valX, c32 ;// multiply 1900c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp6, tmp6, valX, c32 ;// multiply 1910c1bc742181ded4930842b46e9507372f0b1b963James Dong 1920c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp7, tmp3, tmp1, LSL #16 1930c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp8, tmp4, tmp2, LSL #16 1940c1bc742181ded4930842b46e9507372f0b1b963James Dong 1950c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp7, tmp7, valX, c32 ;// multiply 1960c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp8, tmp8, valX, c32 ;// multiply 1970c1bc742181ded4930842b46e9507372f0b1b963James Dong 1980c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp5, tmp5, LSR #6 ;// scale down 1990c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp5, [mb,#8] ;// store row 2 col 1 2000c1bc742181ded4930842b46e9507372f0b1b963James Dong 2010c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp6, tmp6, LSR #6 ;// scale down 2020c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp6, [mb],#1 ;// store row 1 col 1 2030c1bc742181ded4930842b46e9507372f0b1b963James Dong 2040c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp7, tmp7, LSR #6 ;// scale down 2050c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp7, [mb,#8] ;// store row 2 col 2 2060c1bc742181ded4930842b46e9507372f0b1b963James Dong 2070c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp8, tmp8, LSR #6 ;// scale down 2080c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp8, [mb],#1 ;// store row 1 col 2 2090c1bc742181ded4930842b46e9507372f0b1b963James Dong 2100c1bc742181ded4930842b46e9507372f0b1b963James Dong SUBS count, count, #2<<28 2110c1bc742181ded4930842b46e9507372f0b1b963James Dong BCS loop1_x 2120c1bc742181ded4930842b46e9507372f0b1b963James Dong 2130c1bc742181ded4930842b46e9507372f0b1b963James Dong AND tmp2, count, #0x00F00000 2140c1bc742181ded4930842b46e9507372f0b1b963James Dong 2150c1bc742181ded4930842b46e9507372f0b1b963James Dong ADDS mb, mb, #16 2160c1bc742181ded4930842b46e9507372f0b1b963James Dong SBC mb, mb, tmp2, LSR #20 2170c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD ptrA, ptrA, width, LSL #1 2180c1bc742181ded4930842b46e9507372f0b1b963James Dong SBC ptrA, ptrA, tmp2, LSR #20 2190c1bc742181ded4930842b46e9507372f0b1b963James Dong SUB ptrA, ptrA, #1 2200c1bc742181ded4930842b46e9507372f0b1b963James Dong 2210c1bc742181ded4930842b46e9507372f0b1b963James Dong ADDS count, count, #0xE << 24 2220c1bc742181ded4930842b46e9507372f0b1b963James Dong BGE loop1_y 2230c1bc742181ded4930842b46e9507372f0b1b963James Dong 2240c1bc742181ded4930842b46e9507372f0b1b963James Dong ;/////////////////////////////////////////////////////////////////////////// 2250c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// Cr 2260c1bc742181ded4930842b46e9507372f0b1b963James Dong ;/////////////////////////////////////////////////////////////////////////// 2270c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR height, [sp,#0xfc] ;// height 2280c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR ref, [sp, #0xc4] ;// ref 2290c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR tmp1, [sp, #0xd0] ;// y0 2300c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR tmp2, [sp, #0xcc] ;// x0 2310c1bc742181ded4930842b46e9507372f0b1b963James Dong LDR mb, [sp, #0xc8] ;// predPartChroma 2320c1bc742181ded4930842b46e9507372f0b1b963James Dong 2330c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD tmp1, height, tmp1 2340c1bc742181ded4930842b46e9507372f0b1b963James Dong MLA tmp3, tmp1, width, tmp2 2350c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD ptrA, ref, tmp3 2360c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD mb, mb, #64 2370c1bc742181ded4930842b46e9507372f0b1b963James Dong 2380c1bc742181ded4930842b46e9507372f0b1b963James Dong AND count, count, #0x00FFFFFF 2390c1bc742181ded4930842b46e9507372f0b1b963James Dong AND tmp1, count, #0x000F0000 2400c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, tmp1, LSL #8 2410c1bc742181ded4930842b46e9507372f0b1b963James Dong AND tmp2, count, #0x00F00000 2420c1bc742181ded4930842b46e9507372f0b1b963James Dong 2430c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// 2x2 pels per iteration 2440c1bc742181ded4930842b46e9507372f0b1b963James Dong ;// bilinear vertical interpolation 2450c1bc742181ded4930842b46e9507372f0b1b963James Dongloop2_y 2460c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD count, count, tmp2, LSL #8 2470c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp1, [ptrA, width] 2480c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp2, [ptrA], #1 2490c1bc742181ded4930842b46e9507372f0b1b963James Dong 2500c1bc742181ded4930842b46e9507372f0b1b963James Dongloop2_x 2510c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp3, [ptrA, width] 2520c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp4, [ptrA], #1 2530c1bc742181ded4930842b46e9507372f0b1b963James Dong 2540c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp5, tmp1, tmp3, LSL #16 2550c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp6, tmp2, tmp4, LSL #16 2560c1bc742181ded4930842b46e9507372f0b1b963James Dong 2570c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp1, [ptrA, width] 2580c1bc742181ded4930842b46e9507372f0b1b963James Dong LDRB tmp2, [ptrA], #1 2590c1bc742181ded4930842b46e9507372f0b1b963James Dong 2600c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp5, tmp5, valX, c32 ;// multiply 2610c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp6, tmp6, valX, c32 ;// multiply 2620c1bc742181ded4930842b46e9507372f0b1b963James Dong 2630c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp7, tmp3, tmp1, LSL #16 2640c1bc742181ded4930842b46e9507372f0b1b963James Dong PKHBT tmp8, tmp4, tmp2, LSL #16 2650c1bc742181ded4930842b46e9507372f0b1b963James Dong 2660c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp7, tmp7, valX, c32 ;// multiply 2670c1bc742181ded4930842b46e9507372f0b1b963James Dong SMLAD tmp8, tmp8, valX, c32 ;// multiply 2680c1bc742181ded4930842b46e9507372f0b1b963James Dong 2690c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp5, tmp5, LSR #6 ;// scale down 2700c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp5, [mb,#8] ;// store row 2 col 1 2710c1bc742181ded4930842b46e9507372f0b1b963James Dong 2720c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp6, tmp6, LSR #6 ;// scale down 2730c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp6, [mb],#1 ;// store row 1 col 1 2740c1bc742181ded4930842b46e9507372f0b1b963James Dong 2750c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp7, tmp7, LSR #6 ;// scale down 2760c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp7, [mb,#8] ;// store row 2 col 2 2770c1bc742181ded4930842b46e9507372f0b1b963James Dong 2780c1bc742181ded4930842b46e9507372f0b1b963James Dong MOV tmp8, tmp8, LSR #6 ;// scale down 2790c1bc742181ded4930842b46e9507372f0b1b963James Dong STRB tmp8, [mb],#1 ;// store row 1 col 2 2800c1bc742181ded4930842b46e9507372f0b1b963James Dong 2810c1bc742181ded4930842b46e9507372f0b1b963James Dong SUBS count, count, #2<<28 2820c1bc742181ded4930842b46e9507372f0b1b963James Dong BCS loop2_x 2830c1bc742181ded4930842b46e9507372f0b1b963James Dong 2840c1bc742181ded4930842b46e9507372f0b1b963James Dong AND tmp2, count, #0x00F00000 2850c1bc742181ded4930842b46e9507372f0b1b963James Dong 2860c1bc742181ded4930842b46e9507372f0b1b963James Dong ADDS mb, mb, #16 2870c1bc742181ded4930842b46e9507372f0b1b963James Dong SBC mb, mb, tmp2, LSR #20 2880c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD ptrA, ptrA, width, LSL #1 2890c1bc742181ded4930842b46e9507372f0b1b963James Dong SBC ptrA, ptrA, tmp2, LSR #20 2900c1bc742181ded4930842b46e9507372f0b1b963James Dong SUB ptrA, ptrA, #1 2910c1bc742181ded4930842b46e9507372f0b1b963James Dong 2920c1bc742181ded4930842b46e9507372f0b1b963James Dong ADDS count, count, #0xE << 24 2930c1bc742181ded4930842b46e9507372f0b1b963James Dong BGE loop2_y 2940c1bc742181ded4930842b46e9507372f0b1b963James Dong 2950c1bc742181ded4930842b46e9507372f0b1b963James Dong ADD sp,sp,#0xd4 2960c1bc742181ded4930842b46e9507372f0b1b963James Dong LDMFD sp!, {r4-r11,pc} 2970c1bc742181ded4930842b46e9507372f0b1b963James Dong 2980c1bc742181ded4930842b46e9507372f0b1b963James Dong END 299