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: function ff_add_pixels_clamped_armv6, export=1 yading@10: push {r4-r8,lr} yading@10: mov r3, #8 yading@10: 1: yading@10: ldm r0!, {r4,r5,r12,lr} yading@10: ldrd r6, r7, [r1] yading@10: pkhbt r8, r4, r5, lsl #16 yading@10: pkhtb r5, r5, r4, asr #16 yading@10: pkhbt r4, r12, lr, lsl #16 yading@10: pkhtb lr, lr, r12, asr #16 yading@10: pld [r1, r2] yading@10: uxtab16 r8, r8, r6 yading@10: uxtab16 r5, r5, r6, ror #8 yading@10: uxtab16 r4, r4, r7 yading@10: uxtab16 lr, lr, r7, ror #8 yading@10: usat16 r8, #8, r8 yading@10: usat16 r5, #8, r5 yading@10: usat16 r4, #8, r4 yading@10: usat16 lr, #8, lr yading@10: orr r6, r8, r5, lsl #8 yading@10: orr r7, r4, lr, lsl #8 yading@10: subs r3, r3, #1 yading@10: strd_post r6, r7, r1, r2 yading@10: bgt 1b yading@10: pop {r4-r8,pc} yading@10: endfunc yading@10: yading@10: function ff_get_pixels_armv6, export=1 yading@10: pld [r1, r2] yading@10: push {r4-r8, lr} yading@10: mov lr, #8 yading@10: 1: yading@10: ldrd_post r4, r5, r1, r2 yading@10: subs lr, lr, #1 yading@10: uxtb16 r6, r4 yading@10: uxtb16 r4, r4, ror #8 yading@10: uxtb16 r12, r5 yading@10: uxtb16 r8, r5, ror #8 yading@10: pld [r1, r2] yading@10: pkhbt r5, r6, r4, lsl #16 yading@10: pkhtb r6, r4, r6, asr #16 yading@10: pkhbt r7, r12, r8, lsl #16 yading@10: pkhtb r12, r8, r12, asr #16 yading@10: stm r0!, {r5,r6,r7,r12} yading@10: bgt 1b yading@10: yading@10: pop {r4-r8, pc} yading@10: endfunc yading@10: yading@10: function ff_diff_pixels_armv6, export=1 yading@10: pld [r1, r3] yading@10: pld [r2, r3] yading@10: push {r4-r9, lr} yading@10: mov lr, #8 yading@10: 1: yading@10: ldrd_post r4, r5, r1, r3 yading@10: ldrd_post r6, r7, r2, r3 yading@10: uxtb16 r8, r4 yading@10: uxtb16 r4, r4, ror #8 yading@10: uxtb16 r9, r6 yading@10: uxtb16 r6, r6, ror #8 yading@10: pld [r1, r3] yading@10: ssub16 r9, r8, r9 yading@10: ssub16 r6, r4, r6 yading@10: uxtb16 r8, r5 yading@10: uxtb16 r5, r5, ror #8 yading@10: pld [r2, r3] yading@10: pkhbt r4, r9, r6, lsl #16 yading@10: pkhtb r6, r6, r9, asr #16 yading@10: uxtb16 r9, r7 yading@10: uxtb16 r7, r7, ror #8 yading@10: ssub16 r9, r8, r9 yading@10: ssub16 r5, r5, r7 yading@10: subs lr, lr, #1 yading@10: pkhbt r8, r9, r5, lsl #16 yading@10: pkhtb r9, r5, r9, asr #16 yading@10: stm r0!, {r4,r6,r8,r9} yading@10: bgt 1b yading@10: yading@10: pop {r4-r9, pc} yading@10: endfunc yading@10: yading@10: function ff_pix_abs16_armv6, export=1 yading@10: ldr r0, [sp] yading@10: push {r4-r9, lr} yading@10: mov r12, #0 yading@10: mov lr, #0 yading@10: ldm r1, {r4-r7} yading@10: ldr r8, [r2] yading@10: 1: yading@10: ldr r9, [r2, #4] yading@10: pld [r1, r3] yading@10: usada8 r12, r4, r8, r12 yading@10: ldr r8, [r2, #8] yading@10: pld [r2, r3] yading@10: usada8 lr, r5, r9, lr yading@10: ldr r9, [r2, #12] yading@10: usada8 r12, r6, r8, r12 yading@10: subs r0, r0, #1 yading@10: usada8 lr, r7, r9, lr yading@10: beq 2f yading@10: add r1, r1, r3 yading@10: ldm r1, {r4-r7} yading@10: add r2, r2, r3 yading@10: ldr r8, [r2] yading@10: b 1b yading@10: 2: yading@10: add r0, r12, lr yading@10: pop {r4-r9, pc} yading@10: endfunc yading@10: yading@10: function ff_pix_abs16_x2_armv6, export=1 yading@10: ldr r12, [sp] yading@10: push {r4-r11, lr} yading@10: mov r0, #0 yading@10: mov lr, #1 yading@10: orr lr, lr, lr, lsl #8 yading@10: orr lr, lr, lr, lsl #16 yading@10: 1: yading@10: ldr r8, [r2] yading@10: ldr r9, [r2, #4] yading@10: lsr r10, r8, #8 yading@10: ldr r4, [r1] yading@10: lsr r6, r9, #8 yading@10: orr r10, r10, r9, lsl #24 yading@10: ldr r5, [r2, #8] yading@10: eor r11, r8, r10 yading@10: uhadd8 r7, r8, r10 yading@10: orr r6, r6, r5, lsl #24 yading@10: and r11, r11, lr yading@10: uadd8 r7, r7, r11 yading@10: ldr r8, [r1, #4] yading@10: usada8 r0, r4, r7, r0 yading@10: eor r7, r9, r6 yading@10: lsr r10, r5, #8 yading@10: and r7, r7, lr yading@10: uhadd8 r4, r9, r6 yading@10: ldr r6, [r2, #12] yading@10: uadd8 r4, r4, r7 yading@10: pld [r1, r3] yading@10: orr r10, r10, r6, lsl #24 yading@10: usada8 r0, r8, r4, r0 yading@10: ldr r4, [r1, #8] yading@10: eor r11, r5, r10 yading@10: ldrb r7, [r2, #16] yading@10: and r11, r11, lr yading@10: uhadd8 r8, r5, r10 yading@10: ldr r5, [r1, #12] yading@10: uadd8 r8, r8, r11 yading@10: pld [r2, r3] yading@10: lsr r10, r6, #8 yading@10: usada8 r0, r4, r8, r0 yading@10: orr r10, r10, r7, lsl #24 yading@10: subs r12, r12, #1 yading@10: eor r11, r6, r10 yading@10: add r1, r1, r3 yading@10: uhadd8 r9, r6, r10 yading@10: and r11, r11, lr yading@10: uadd8 r9, r9, r11 yading@10: add r2, r2, r3 yading@10: usada8 r0, r5, r9, r0 yading@10: bgt 1b yading@10: yading@10: pop {r4-r11, pc} yading@10: endfunc yading@10: yading@10: .macro usad_y2 p0, p1, p2, p3, n0, n1, n2, n3 yading@10: ldr \n0, [r2] yading@10: eor \n1, \p0, \n0 yading@10: uhadd8 \p0, \p0, \n0 yading@10: and \n1, \n1, lr yading@10: ldr \n2, [r1] yading@10: uadd8 \p0, \p0, \n1 yading@10: ldr \n1, [r2, #4] yading@10: usada8 r0, \p0, \n2, r0 yading@10: pld [r1, r3] yading@10: eor \n3, \p1, \n1 yading@10: uhadd8 \p1, \p1, \n1 yading@10: and \n3, \n3, lr yading@10: ldr \p0, [r1, #4] yading@10: uadd8 \p1, \p1, \n3 yading@10: ldr \n2, [r2, #8] yading@10: usada8 r0, \p1, \p0, r0 yading@10: pld [r2, r3] yading@10: eor \p0, \p2, \n2 yading@10: uhadd8 \p2, \p2, \n2 yading@10: and \p0, \p0, lr yading@10: ldr \p1, [r1, #8] yading@10: uadd8 \p2, \p2, \p0 yading@10: ldr \n3, [r2, #12] yading@10: usada8 r0, \p2, \p1, r0 yading@10: eor \p1, \p3, \n3 yading@10: uhadd8 \p3, \p3, \n3 yading@10: and \p1, \p1, lr yading@10: ldr \p0, [r1, #12] yading@10: uadd8 \p3, \p3, \p1 yading@10: add r1, r1, r3 yading@10: usada8 r0, \p3, \p0, r0 yading@10: add r2, r2, r3 yading@10: .endm yading@10: yading@10: function ff_pix_abs16_y2_armv6, export=1 yading@10: pld [r1] yading@10: pld [r2] yading@10: ldr r12, [sp] yading@10: push {r4-r11, lr} yading@10: mov r0, #0 yading@10: mov lr, #1 yading@10: orr lr, lr, lr, lsl #8 yading@10: orr lr, lr, lr, lsl #16 yading@10: ldr r4, [r2] yading@10: ldr r5, [r2, #4] yading@10: ldr r6, [r2, #8] yading@10: ldr r7, [r2, #12] yading@10: add r2, r2, r3 yading@10: 1: yading@10: usad_y2 r4, r5, r6, r7, r8, r9, r10, r11 yading@10: subs r12, r12, #2 yading@10: usad_y2 r8, r9, r10, r11, r4, r5, r6, r7 yading@10: bgt 1b yading@10: yading@10: pop {r4-r11, pc} yading@10: endfunc yading@10: yading@10: function ff_pix_abs8_armv6, export=1 yading@10: pld [r2, r3] yading@10: ldr r12, [sp] yading@10: push {r4-r9, lr} yading@10: mov r0, #0 yading@10: mov lr, #0 yading@10: ldrd_post r4, r5, r1, r3 yading@10: 1: yading@10: subs r12, r12, #2 yading@10: ldr r7, [r2, #4] yading@10: ldr_post r6, r2, r3 yading@10: ldrd_post r8, r9, r1, r3 yading@10: usada8 r0, r4, r6, r0 yading@10: pld [r2, r3] yading@10: usada8 lr, r5, r7, lr yading@10: ldr r7, [r2, #4] yading@10: ldr_post r6, r2, r3 yading@10: beq 2f yading@10: ldrd_post r4, r5, r1, r3 yading@10: usada8 r0, r8, r6, r0 yading@10: pld [r2, r3] yading@10: usada8 lr, r9, r7, lr yading@10: b 1b yading@10: 2: yading@10: usada8 r0, r8, r6, r0 yading@10: usada8 lr, r9, r7, lr yading@10: add r0, r0, lr yading@10: pop {r4-r9, pc} yading@10: endfunc yading@10: yading@10: function ff_sse16_armv6, export=1 yading@10: ldr r12, [sp] yading@10: push {r4-r9, lr} yading@10: mov r0, #0 yading@10: 1: yading@10: ldrd r4, r5, [r1] yading@10: ldr r8, [r2] yading@10: uxtb16 lr, r4 yading@10: uxtb16 r4, r4, ror #8 yading@10: uxtb16 r9, r8 yading@10: uxtb16 r8, r8, ror #8 yading@10: ldr r7, [r2, #4] yading@10: usub16 lr, lr, r9 yading@10: usub16 r4, r4, r8 yading@10: smlad r0, lr, lr, r0 yading@10: uxtb16 r6, r5 yading@10: uxtb16 lr, r5, ror #8 yading@10: uxtb16 r8, r7 yading@10: uxtb16 r9, r7, ror #8 yading@10: smlad r0, r4, r4, r0 yading@10: ldrd r4, r5, [r1, #8] yading@10: usub16 r6, r6, r8 yading@10: usub16 r8, lr, r9 yading@10: ldr r7, [r2, #8] yading@10: smlad r0, r6, r6, r0 yading@10: uxtb16 lr, r4 yading@10: uxtb16 r4, r4, ror #8 yading@10: uxtb16 r9, r7 yading@10: uxtb16 r7, r7, ror #8 yading@10: smlad r0, r8, r8, r0 yading@10: ldr r8, [r2, #12] yading@10: usub16 lr, lr, r9 yading@10: usub16 r4, r4, r7 yading@10: smlad r0, lr, lr, r0 yading@10: uxtb16 r6, r5 yading@10: uxtb16 r5, r5, ror #8 yading@10: uxtb16 r9, r8 yading@10: uxtb16 r8, r8, ror #8 yading@10: smlad r0, r4, r4, r0 yading@10: usub16 r6, r6, r9 yading@10: usub16 r5, r5, r8 yading@10: smlad r0, r6, r6, r0 yading@10: add r1, r1, r3 yading@10: add r2, r2, r3 yading@10: subs r12, r12, #1 yading@10: smlad r0, r5, r5, r0 yading@10: bgt 1b yading@10: yading@10: pop {r4-r9, pc} yading@10: endfunc yading@10: yading@10: function ff_pix_norm1_armv6, export=1 yading@10: push {r4-r6, lr} yading@10: mov r12, #16 yading@10: mov lr, #0 yading@10: 1: yading@10: ldm r0, {r2-r5} yading@10: uxtb16 r6, r2 yading@10: uxtb16 r2, r2, ror #8 yading@10: smlad lr, r6, r6, lr yading@10: uxtb16 r6, r3 yading@10: smlad lr, r2, r2, lr yading@10: uxtb16 r3, r3, ror #8 yading@10: smlad lr, r6, r6, lr yading@10: uxtb16 r6, r4 yading@10: smlad lr, r3, r3, lr yading@10: uxtb16 r4, r4, ror #8 yading@10: smlad lr, r6, r6, lr yading@10: uxtb16 r6, r5 yading@10: smlad lr, r4, r4, lr yading@10: uxtb16 r5, r5, ror #8 yading@10: smlad lr, r6, r6, lr yading@10: subs r12, r12, #1 yading@10: add r0, r0, r1 yading@10: smlad lr, r5, r5, lr yading@10: bgt 1b yading@10: yading@10: mov r0, lr yading@10: pop {r4-r6, pc} yading@10: endfunc yading@10: yading@10: function ff_pix_sum_armv6, export=1 yading@10: push {r4-r7, lr} yading@10: mov r12, #16 yading@10: mov r2, #0 yading@10: mov r3, #0 yading@10: mov lr, #0 yading@10: ldr r4, [r0] yading@10: 1: yading@10: subs r12, r12, #1 yading@10: ldr r5, [r0, #4] yading@10: usada8 r2, r4, lr, r2 yading@10: ldr r6, [r0, #8] yading@10: usada8 r3, r5, lr, r3 yading@10: ldr r7, [r0, #12] yading@10: usada8 r2, r6, lr, r2 yading@10: beq 2f yading@10: ldr_pre r4, r0, r1 yading@10: usada8 r3, r7, lr, r3 yading@10: bgt 1b yading@10: 2: yading@10: usada8 r3, r7, lr, r3 yading@10: add r0, r2, r3 yading@10: pop {r4-r7, pc} yading@10: endfunc