yading@10: /* yading@10: * Copyright (c) 2009 Mans Rullgard yading@10: * yading@10: * This file is part of FFmpeg. yading@10: * yading@10: * FFmpeg is free software; you can redistribute it and/or yading@10: * modify it under the terms of the GNU Lesser General Public yading@10: * License as published by the Free Software Foundation; either yading@10: * version 2.1 of the License, or (at your option) any later version. yading@10: * yading@10: * FFmpeg is distributed in the hope that it will be useful, yading@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: * Lesser General Public License for more details. yading@10: * yading@10: * You should have received a copy of the GNU Lesser General Public yading@10: * License along with FFmpeg; if not, write to the Free Software yading@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: */ yading@10: yading@10: #include "libavutil/arm/asm.S" yading@10: yading@10: .macro call_2x_pixels type, subp yading@10: function ff_\type\()_pixels16\subp\()_armv6, export=1 yading@10: push {r0-r3, lr} yading@10: bl ff_\type\()_pixels8\subp\()_armv6 yading@10: pop {r0-r3, lr} yading@10: add r0, r0, #8 yading@10: add r1, r1, #8 yading@10: b ff_\type\()_pixels8\subp\()_armv6 yading@10: endfunc yading@10: .endm yading@10: yading@10: call_2x_pixels avg yading@10: call_2x_pixels put, _x2 yading@10: call_2x_pixels put, _y2 yading@10: call_2x_pixels put, _x2_no_rnd yading@10: call_2x_pixels put, _y2_no_rnd yading@10: yading@10: function ff_put_pixels16_armv6, export=1 yading@10: push {r4-r11} yading@10: 1: yading@10: ldr r5, [r1, #4] yading@10: ldr r6, [r1, #8] yading@10: ldr r7, [r1, #12] yading@10: ldr_post r4, r1, r2 yading@10: strd r6, r7, [r0, #8] yading@10: ldr r9, [r1, #4] yading@10: strd_post r4, r5, r0, r2 yading@10: ldr r10, [r1, #8] yading@10: ldr r11, [r1, #12] yading@10: ldr_post r8, r1, r2 yading@10: strd r10, r11, [r0, #8] yading@10: subs r3, r3, #2 yading@10: strd_post r8, r9, r0, r2 yading@10: bne 1b yading@10: yading@10: pop {r4-r11} yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_put_pixels8_armv6, export=1 yading@10: push {r4-r7} yading@10: 1: yading@10: ldr r5, [r1, #4] yading@10: ldr_post r4, r1, r2 yading@10: ldr r7, [r1, #4] yading@10: strd_post r4, r5, r0, r2 yading@10: ldr_post r6, r1, r2 yading@10: subs r3, r3, #2 yading@10: strd_post r6, r7, r0, r2 yading@10: bne 1b yading@10: yading@10: pop {r4-r7} yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_put_pixels8_x2_armv6, export=1 yading@10: push {r4-r11, lr} yading@10: mov r12, #1 yading@10: orr r12, r12, r12, lsl #8 yading@10: orr r12, r12, r12, lsl #16 yading@10: 1: yading@10: ldr r4, [r1] yading@10: subs r3, r3, #2 yading@10: ldr r5, [r1, #4] yading@10: ldr r7, [r1, #5] yading@10: lsr r6, r4, #8 yading@10: ldr_pre r8, r1, r2 yading@10: orr r6, r6, r5, lsl #24 yading@10: ldr r9, [r1, #4] yading@10: ldr r11, [r1, #5] yading@10: lsr r10, r8, #8 yading@10: add r1, r1, r2 yading@10: orr r10, r10, r9, lsl #24 yading@10: eor r14, r4, r6 yading@10: uhadd8 r4, r4, r6 yading@10: eor r6, r5, r7 yading@10: uhadd8 r5, r5, r7 yading@10: and r14, r14, r12 yading@10: and r6, r6, r12 yading@10: uadd8 r4, r4, r14 yading@10: eor r14, r8, r10 yading@10: uadd8 r5, r5, r6 yading@10: eor r6, r9, r11 yading@10: uhadd8 r8, r8, r10 yading@10: and r14, r14, r12 yading@10: uhadd8 r9, r9, r11 yading@10: and r6, r6, r12 yading@10: uadd8 r8, r8, r14 yading@10: strd_post r4, r5, r0, r2 yading@10: uadd8 r9, r9, r6 yading@10: strd_post r8, r9, r0, r2 yading@10: bne 1b yading@10: yading@10: pop {r4-r11, pc} yading@10: endfunc yading@10: yading@10: function ff_put_pixels8_y2_armv6, export=1 yading@10: push {r4-r11} yading@10: mov r12, #1 yading@10: orr r12, r12, r12, lsl #8 yading@10: orr r12, r12, r12, lsl #16 yading@10: ldr r4, [r1] yading@10: ldr r5, [r1, #4] yading@10: ldr_pre r6, r1, r2 yading@10: ldr r7, [r1, #4] yading@10: 1: yading@10: subs r3, r3, #2 yading@10: uhadd8 r8, r4, r6 yading@10: eor r10, r4, r6 yading@10: uhadd8 r9, r5, r7 yading@10: eor r11, r5, r7 yading@10: and r10, r10, r12 yading@10: ldr_pre r4, r1, r2 yading@10: uadd8 r8, r8, r10 yading@10: and r11, r11, r12 yading@10: uadd8 r9, r9, r11 yading@10: ldr r5, [r1, #4] yading@10: uhadd8 r10, r4, r6 yading@10: eor r6, r4, r6 yading@10: uhadd8 r11, r5, r7 yading@10: and r6, r6, r12 yading@10: eor r7, r5, r7 yading@10: uadd8 r10, r10, r6 yading@10: and r7, r7, r12 yading@10: ldr_pre r6, r1, r2 yading@10: uadd8 r11, r11, r7 yading@10: strd_post r8, r9, r0, r2 yading@10: ldr r7, [r1, #4] yading@10: strd_post r10, r11, r0, r2 yading@10: bne 1b yading@10: yading@10: pop {r4-r11} yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_put_pixels8_x2_no_rnd_armv6, export=1 yading@10: push {r4-r9, lr} yading@10: 1: yading@10: subs r3, r3, #2 yading@10: ldr r4, [r1] yading@10: ldr r5, [r1, #4] yading@10: ldr r7, [r1, #5] yading@10: ldr_pre r8, r1, r2 yading@10: ldr r9, [r1, #4] yading@10: ldr r14, [r1, #5] yading@10: add r1, r1, r2 yading@10: lsr r6, r4, #8 yading@10: orr r6, r6, r5, lsl #24 yading@10: lsr r12, r8, #8 yading@10: orr r12, r12, r9, lsl #24 yading@10: uhadd8 r4, r4, r6 yading@10: uhadd8 r5, r5, r7 yading@10: uhadd8 r8, r8, r12 yading@10: uhadd8 r9, r9, r14 yading@10: stm r0, {r4,r5} yading@10: add r0, r0, r2 yading@10: stm r0, {r8,r9} yading@10: add r0, r0, r2 yading@10: bne 1b yading@10: yading@10: pop {r4-r9, pc} yading@10: endfunc yading@10: yading@10: function ff_put_pixels8_y2_no_rnd_armv6, export=1 yading@10: push {r4-r9, lr} yading@10: ldr r4, [r1] yading@10: ldr r5, [r1, #4] yading@10: ldr_pre r6, r1, r2 yading@10: ldr r7, [r1, #4] yading@10: 1: yading@10: subs r3, r3, #2 yading@10: uhadd8 r8, r4, r6 yading@10: ldr_pre r4, r1, r2 yading@10: uhadd8 r9, r5, r7 yading@10: ldr r5, [r1, #4] yading@10: uhadd8 r12, r4, r6 yading@10: ldr_pre r6, r1, r2 yading@10: uhadd8 r14, r5, r7 yading@10: ldr r7, [r1, #4] yading@10: stm r0, {r8,r9} yading@10: add r0, r0, r2 yading@10: stm r0, {r12,r14} yading@10: add r0, r0, r2 yading@10: bne 1b yading@10: yading@10: pop {r4-r9, pc} yading@10: endfunc yading@10: yading@10: function ff_avg_pixels8_armv6, export=1 yading@10: pld [r1, r2] yading@10: push {r4-r10, lr} yading@10: mov lr, #1 yading@10: orr lr, lr, lr, lsl #8 yading@10: orr lr, lr, lr, lsl #16 yading@10: ldrd r4, r5, [r0] yading@10: ldr r10, [r1, #4] yading@10: ldr_post r9, r1, r2 yading@10: subs r3, r3, #2 yading@10: 1: yading@10: pld [r1, r2] yading@10: eor r8, r4, r9 yading@10: uhadd8 r4, r4, r9 yading@10: eor r12, r5, r10 yading@10: ldrd_reg r6, r7, r0, r2 yading@10: uhadd8 r5, r5, r10 yading@10: and r8, r8, lr yading@10: ldr r10, [r1, #4] yading@10: and r12, r12, lr yading@10: uadd8 r4, r4, r8 yading@10: ldr_post r9, r1, r2 yading@10: eor r8, r6, r9 yading@10: uadd8 r5, r5, r12 yading@10: pld [r1, r2, lsl #1] yading@10: eor r12, r7, r10 yading@10: uhadd8 r6, r6, r9 yading@10: strd_post r4, r5, r0, r2 yading@10: uhadd8 r7, r7, r10 yading@10: beq 2f yading@10: and r8, r8, lr yading@10: ldrd_reg r4, r5, r0, r2 yading@10: uadd8 r6, r6, r8 yading@10: ldr r10, [r1, #4] yading@10: and r12, r12, lr yading@10: subs r3, r3, #2 yading@10: uadd8 r7, r7, r12 yading@10: ldr_post r9, r1, r2 yading@10: strd_post r6, r7, r0, r2 yading@10: b 1b yading@10: 2: yading@10: and r8, r8, lr yading@10: and r12, r12, lr yading@10: uadd8 r6, r6, r8 yading@10: uadd8 r7, r7, r12 yading@10: strd_post r6, r7, r0, r2 yading@10: yading@10: pop {r4-r10, pc} yading@10: endfunc