yading@10: /* yading@10: * Copyright (c) 2010 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 vp6_edge_filter yading@10: vdup.16 q3, r2 @ t yading@10: vmov.i16 q13, #1 yading@10: vsubl.u8 q0, d20, d18 @ p[ 0] - p[-s] yading@10: vsubl.u8 q1, d16, d22 @ p[-2*s] - p[ s] yading@10: vsubl.u8 q14, d21, d19 yading@10: vsubl.u8 q15, d17, d23 yading@10: vadd.i16 q2, q0, q0 @ 2*(p[0]-p[-s]) yading@10: vadd.i16 d29, d28, d28 yading@10: vadd.i16 q0, q0, q1 @ p[0]-p[-s] + p[-2*s]-p[s] yading@10: vadd.i16 d28, d28, d30 yading@10: vadd.i16 q0, q0, q2 @ 3*(p[0]-p[-s]) + p[-2*s]-p[s] yading@10: vadd.i16 d28, d28, d29 yading@10: vrshr.s16 q0, q0, #3 @ v yading@10: vrshr.s16 d28, d28, #3 yading@10: vsub.i16 q8, q3, q13 @ t-1 yading@10: vabs.s16 q1, q0 @ V yading@10: vshr.s16 q2, q0, #15 @ s yading@10: vabs.s16 d30, d28 yading@10: vshr.s16 d29, d28, #15 yading@10: vsub.i16 q12, q1, q3 @ V-t yading@10: vsub.i16 d31, d30, d6 yading@10: vsub.i16 q12, q12, q13 @ V-t-1 yading@10: vsub.i16 d31, d31, d26 yading@10: vcge.u16 q12, q12, q8 @ V-t-1 >= t-1 yading@10: vcge.u16 d31, d31, d16 yading@10: vadd.i16 q13, q3, q3 @ 2*t yading@10: vadd.i16 d16, d6, d6 yading@10: vsub.i16 q13, q13, q1 @ 2*t - V yading@10: vsub.i16 d16, d16, d30 yading@10: vadd.i16 q13, q13, q2 @ += s yading@10: vadd.i16 d16, d16, d29 yading@10: veor q13, q13, q2 @ ^= s yading@10: veor d16, d16, d29 yading@10: vbif q0, q13, q12 yading@10: vbif d28, d16, d31 yading@10: vmovl.u8 q1, d20 yading@10: vmovl.u8 q15, d21 yading@10: vaddw.u8 q2, q0, d18 yading@10: vaddw.u8 q3, q14, d19 yading@10: vsub.i16 q1, q1, q0 yading@10: vsub.i16 d30, d30, d28 yading@10: vqmovun.s16 d18, q2 yading@10: vqmovun.s16 d19, q3 yading@10: vqmovun.s16 d20, q1 yading@10: vqmovun.s16 d21, q15 yading@10: .endm yading@10: yading@10: function ff_vp6_edge_filter_ver_neon, export=1 yading@10: sub r0, r0, r1, lsl #1 yading@10: vld1.8 {q8}, [r0], r1 @ p[-2*s] yading@10: vld1.8 {q9}, [r0], r1 @ p[-s] yading@10: vld1.8 {q10}, [r0], r1 @ p[0] yading@10: vld1.8 {q11}, [r0] @ p[s] yading@10: vp6_edge_filter yading@10: sub r0, r0, r1, lsl #1 yading@10: sub r1, r1, #8 yading@10: vst1.8 {d18}, [r0]! yading@10: vst1.32 {d19[0]}, [r0], r1 yading@10: vst1.8 {d20}, [r0]! yading@10: vst1.32 {d21[0]}, [r0] yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_vp6_edge_filter_hor_neon, export=1 yading@10: sub r3, r0, #1 yading@10: sub r0, r0, #2 yading@10: vld1.32 {d16[0]}, [r0], r1 yading@10: vld1.32 {d18[0]}, [r0], r1 yading@10: vld1.32 {d20[0]}, [r0], r1 yading@10: vld1.32 {d22[0]}, [r0], r1 yading@10: vld1.32 {d16[1]}, [r0], r1 yading@10: vld1.32 {d18[1]}, [r0], r1 yading@10: vld1.32 {d20[1]}, [r0], r1 yading@10: vld1.32 {d22[1]}, [r0], r1 yading@10: vld1.32 {d17[0]}, [r0], r1 yading@10: vld1.32 {d19[0]}, [r0], r1 yading@10: vld1.32 {d21[0]}, [r0], r1 yading@10: vld1.32 {d23[0]}, [r0], r1 yading@10: vtrn.8 q8, q9 yading@10: vtrn.8 q10, q11 yading@10: vtrn.16 q8, q10 yading@10: vtrn.16 q9, q11 yading@10: vp6_edge_filter yading@10: vtrn.8 q9, q10 yading@10: vst1.16 {d18[0]}, [r3], r1 yading@10: vst1.16 {d20[0]}, [r3], r1 yading@10: vst1.16 {d18[1]}, [r3], r1 yading@10: vst1.16 {d20[1]}, [r3], r1 yading@10: vst1.16 {d18[2]}, [r3], r1 yading@10: vst1.16 {d20[2]}, [r3], r1 yading@10: vst1.16 {d18[3]}, [r3], r1 yading@10: vst1.16 {d20[3]}, [r3], r1 yading@10: vst1.16 {d19[0]}, [r3], r1 yading@10: vst1.16 {d21[0]}, [r3], r1 yading@10: vst1.16 {d19[1]}, [r3], r1 yading@10: vst1.16 {d21[1]}, [r3], r1 yading@10: bx lr yading@10: endfunc