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 ldcol.8 rd, rs, rt, n=8, hi=0 yading@10: .if \n == 8 || \hi == 0 yading@10: vld1.8 {\rd[0]}, [\rs], \rt yading@10: vld1.8 {\rd[1]}, [\rs], \rt yading@10: vld1.8 {\rd[2]}, [\rs], \rt yading@10: vld1.8 {\rd[3]}, [\rs], \rt yading@10: .endif yading@10: .if \n == 8 || \hi == 1 yading@10: vld1.8 {\rd[4]}, [\rs], \rt yading@10: vld1.8 {\rd[5]}, [\rs], \rt yading@10: vld1.8 {\rd[6]}, [\rs], \rt yading@10: vld1.8 {\rd[7]}, [\rs], \rt yading@10: .endif yading@10: .endm yading@10: yading@10: .macro add16x8 dq, dl, dh, rl, rh yading@10: vaddl.u8 \dq, \rl, \rh yading@10: vadd.u16 \dl, \dl, \dh yading@10: vpadd.u16 \dl, \dl, \dl yading@10: vpadd.u16 \dl, \dl, \dl yading@10: .endm yading@10: yading@10: function ff_pred16x16_128_dc_neon, export=1 yading@10: vmov.i8 q0, #128 yading@10: b .L_pred16x16_dc_end yading@10: endfunc yading@10: yading@10: function ff_pred16x16_top_dc_neon, export=1 yading@10: sub r2, r0, r1 yading@10: vld1.8 {q0}, [r2,:128] yading@10: add16x8 q0, d0, d1, d0, d1 yading@10: vrshrn.u16 d0, q0, #4 yading@10: vdup.8 q0, d0[0] yading@10: b .L_pred16x16_dc_end yading@10: endfunc yading@10: yading@10: function ff_pred16x16_left_dc_neon, export=1 yading@10: sub r2, r0, #1 yading@10: ldcol.8 d0, r2, r1 yading@10: ldcol.8 d1, r2, r1 yading@10: add16x8 q0, d0, d1, d0, d1 yading@10: vrshrn.u16 d0, q0, #4 yading@10: vdup.8 q0, d0[0] yading@10: b .L_pred16x16_dc_end yading@10: endfunc yading@10: yading@10: function ff_pred16x16_dc_neon, export=1 yading@10: sub r2, r0, r1 yading@10: vld1.8 {q0}, [r2,:128] yading@10: sub r2, r0, #1 yading@10: ldcol.8 d2, r2, r1 yading@10: ldcol.8 d3, r2, r1 yading@10: vaddl.u8 q0, d0, d1 yading@10: vaddl.u8 q1, d2, d3 yading@10: vadd.u16 q0, q0, q1 yading@10: vadd.u16 d0, d0, d1 yading@10: vpadd.u16 d0, d0, d0 yading@10: vpadd.u16 d0, d0, d0 yading@10: vrshrn.u16 d0, q0, #5 yading@10: vdup.8 q0, d0[0] yading@10: .L_pred16x16_dc_end: yading@10: mov r3, #8 yading@10: 6: vst1.8 {q0}, [r0,:128], r1 yading@10: vst1.8 {q0}, [r0,:128], r1 yading@10: subs r3, r3, #1 yading@10: bne 6b yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_pred16x16_hor_neon, export=1 yading@10: sub r2, r0, #1 yading@10: mov r3, #16 yading@10: 1: vld1.8 {d0[],d1[]},[r2], r1 yading@10: vst1.8 {q0}, [r0,:128], r1 yading@10: subs r3, r3, #1 yading@10: bne 1b yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_pred16x16_vert_neon, export=1 yading@10: sub r0, r0, r1 yading@10: vld1.8 {q0}, [r0,:128], r1 yading@10: mov r3, #8 yading@10: 1: vst1.8 {q0}, [r0,:128], r1 yading@10: vst1.8 {q0}, [r0,:128], r1 yading@10: subs r3, r3, #1 yading@10: bne 1b yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_pred16x16_plane_neon, export=1 yading@10: sub r3, r0, r1 yading@10: add r2, r3, #8 yading@10: sub r3, r3, #1 yading@10: vld1.8 {d0}, [r3] yading@10: vld1.8 {d2}, [r2,:64], r1 yading@10: ldcol.8 d1, r3, r1 yading@10: add r3, r3, r1 yading@10: ldcol.8 d3, r3, r1 yading@10: vrev64.8 q0, q0 yading@10: vaddl.u8 q8, d2, d3 yading@10: vsubl.u8 q2, d2, d0 yading@10: vsubl.u8 q3, d3, d1 yading@10: movrel r3, p16weight yading@10: vld1.8 {q0}, [r3,:128] yading@10: vmul.s16 q2, q2, q0 yading@10: vmul.s16 q3, q3, q0 yading@10: vadd.i16 d4, d4, d5 yading@10: vadd.i16 d5, d6, d7 yading@10: vpadd.i16 d4, d4, d5 yading@10: vpadd.i16 d4, d4, d4 yading@10: vshll.s16 q3, d4, #2 yading@10: vaddw.s16 q2, q3, d4 yading@10: vrshrn.s32 d4, q2, #6 yading@10: mov r3, #0 yading@10: vtrn.16 d4, d5 yading@10: vadd.i16 d2, d4, d5 yading@10: vshl.i16 d3, d2, #3 yading@10: vrev64.16 d16, d17 yading@10: vsub.i16 d3, d3, d2 yading@10: vadd.i16 d16, d16, d0 yading@10: vshl.i16 d2, d16, #4 yading@10: vsub.i16 d2, d2, d3 yading@10: vshl.i16 d3, d4, #4 yading@10: vext.16 q0, q0, q0, #7 yading@10: vsub.i16 d6, d5, d3 yading@10: vmov.16 d0[0], r3 yading@10: vmul.i16 q0, q0, d4[0] yading@10: vdup.16 q1, d2[0] yading@10: vdup.16 q2, d4[0] yading@10: vdup.16 q3, d6[0] yading@10: vshl.i16 q2, q2, #3 yading@10: vadd.i16 q1, q1, q0 yading@10: vadd.i16 q3, q3, q2 yading@10: mov r3, #16 yading@10: 1: yading@10: vqshrun.s16 d0, q1, #5 yading@10: vadd.i16 q1, q1, q2 yading@10: vqshrun.s16 d1, q1, #5 yading@10: vadd.i16 q1, q1, q3 yading@10: vst1.8 {q0}, [r0,:128], r1 yading@10: subs r3, r3, #1 yading@10: bne 1b yading@10: bx lr yading@10: endfunc yading@10: yading@10: const p16weight, align=4 yading@10: .short 1,2,3,4,5,6,7,8 yading@10: endconst yading@10: yading@10: function ff_pred8x8_hor_neon, export=1 yading@10: sub r2, r0, #1 yading@10: mov r3, #8 yading@10: 1: vld1.8 {d0[]}, [r2], r1 yading@10: vst1.8 {d0}, [r0,:64], r1 yading@10: subs r3, r3, #1 yading@10: bne 1b yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_pred8x8_vert_neon, export=1 yading@10: sub r0, r0, r1 yading@10: vld1.8 {d0}, [r0,:64], r1 yading@10: mov r3, #4 yading@10: 1: vst1.8 {d0}, [r0,:64], r1 yading@10: vst1.8 {d0}, [r0,:64], r1 yading@10: subs r3, r3, #1 yading@10: bne 1b yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_pred8x8_plane_neon, export=1 yading@10: sub r3, r0, r1 yading@10: add r2, r3, #4 yading@10: sub r3, r3, #1 yading@10: vld1.32 {d0[0]}, [r3] yading@10: vld1.32 {d2[0]}, [r2,:32], r1 yading@10: ldcol.8 d0, r3, r1, 4, hi=1 yading@10: add r3, r3, r1 yading@10: ldcol.8 d3, r3, r1, 4 yading@10: vaddl.u8 q8, d2, d3 yading@10: vrev32.8 d0, d0 yading@10: vtrn.32 d2, d3 yading@10: vsubl.u8 q2, d2, d0 yading@10: movrel r3, p16weight yading@10: vld1.16 {q0}, [r3,:128] yading@10: vmul.s16 d4, d4, d0 yading@10: vmul.s16 d5, d5, d0 yading@10: vpadd.i16 d4, d4, d5 yading@10: vpaddl.s16 d4, d4 yading@10: vshl.i32 d5, d4, #4 yading@10: vadd.s32 d4, d4, d5 yading@10: vrshrn.s32 d4, q2, #5 yading@10: mov r3, #0 yading@10: vtrn.16 d4, d5 yading@10: vadd.i16 d2, d4, d5 yading@10: vshl.i16 d3, d2, #2 yading@10: vrev64.16 d16, d16 yading@10: vsub.i16 d3, d3, d2 yading@10: vadd.i16 d16, d16, d0 yading@10: vshl.i16 d2, d16, #4 yading@10: vsub.i16 d2, d2, d3 yading@10: vshl.i16 d3, d4, #3 yading@10: vext.16 q0, q0, q0, #7 yading@10: vsub.i16 d6, d5, d3 yading@10: vmov.16 d0[0], r3 yading@10: vmul.i16 q0, q0, d4[0] yading@10: vdup.16 q1, d2[0] yading@10: vdup.16 q2, d4[0] yading@10: vdup.16 q3, d6[0] yading@10: vshl.i16 q2, q2, #3 yading@10: vadd.i16 q1, q1, q0 yading@10: vadd.i16 q3, q3, q2 yading@10: mov r3, #8 yading@10: 1: yading@10: vqshrun.s16 d0, q1, #5 yading@10: vadd.i16 q1, q1, q3 yading@10: vst1.8 {d0}, [r0,:64], r1 yading@10: subs r3, r3, #1 yading@10: bne 1b yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_pred8x8_128_dc_neon, export=1 yading@10: vmov.i8 q0, #128 yading@10: b .L_pred8x8_dc_end yading@10: endfunc yading@10: yading@10: function ff_pred8x8_top_dc_neon, export=1 yading@10: sub r2, r0, r1 yading@10: vld1.8 {d0}, [r2,:64] yading@10: vpaddl.u8 d0, d0 yading@10: vpadd.u16 d0, d0, d0 yading@10: vrshrn.u16 d0, q0, #2 yading@10: vdup.8 d1, d0[1] yading@10: vdup.8 d0, d0[0] yading@10: vtrn.32 d0, d1 yading@10: b .L_pred8x8_dc_end yading@10: endfunc yading@10: yading@10: function ff_pred8x8_left_dc_neon, export=1 yading@10: sub r2, r0, #1 yading@10: ldcol.8 d0, r2, r1 yading@10: vpaddl.u8 d0, d0 yading@10: vpadd.u16 d0, d0, d0 yading@10: vrshrn.u16 d0, q0, #2 yading@10: vdup.8 d1, d0[1] yading@10: vdup.8 d0, d0[0] yading@10: b .L_pred8x8_dc_end yading@10: endfunc yading@10: yading@10: function ff_pred8x8_dc_neon, export=1 yading@10: sub r2, r0, r1 yading@10: vld1.8 {d0}, [r2,:64] yading@10: sub r2, r0, #1 yading@10: ldcol.8 d1, r2, r1 yading@10: vtrn.32 d0, d1 yading@10: vpaddl.u8 q0, q0 yading@10: vpadd.u16 d0, d0, d1 yading@10: vpadd.u16 d1, d0, d0 yading@10: vrshrn.u16 d2, q0, #3 yading@10: vrshrn.u16 d3, q0, #2 yading@10: vdup.8 d0, d2[4] yading@10: vdup.8 d1, d3[3] yading@10: vdup.8 d4, d3[2] yading@10: vdup.8 d5, d2[5] yading@10: vtrn.32 q0, q2 yading@10: .L_pred8x8_dc_end: yading@10: mov r3, #4 yading@10: add r2, r0, r1, lsl #2 yading@10: 6: vst1.8 {d0}, [r0,:64], r1 yading@10: vst1.8 {d1}, [r2,:64], r1 yading@10: subs r3, r3, #1 yading@10: bne 6b yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_pred8x8_l0t_dc_neon, export=1 yading@10: sub r2, r0, r1 yading@10: vld1.8 {d0}, [r2,:64] yading@10: sub r2, r0, #1 yading@10: ldcol.8 d1, r2, r1, 4 yading@10: vtrn.32 d0, d1 yading@10: vpaddl.u8 q0, q0 yading@10: vpadd.u16 d0, d0, d1 yading@10: vpadd.u16 d1, d0, d0 yading@10: vrshrn.u16 d2, q0, #3 yading@10: vrshrn.u16 d3, q0, #2 yading@10: vdup.8 d0, d2[4] yading@10: vdup.8 d1, d3[0] yading@10: vdup.8 q2, d3[2] yading@10: vtrn.32 q0, q2 yading@10: b .L_pred8x8_dc_end yading@10: endfunc yading@10: yading@10: function ff_pred8x8_l00_dc_neon, export=1 yading@10: sub r2, r0, #1 yading@10: ldcol.8 d0, r2, r1, 4 yading@10: vpaddl.u8 d0, d0 yading@10: vpadd.u16 d0, d0, d0 yading@10: vrshrn.u16 d0, q0, #2 yading@10: vmov.i8 d1, #128 yading@10: vdup.8 d0, d0[0] yading@10: b .L_pred8x8_dc_end yading@10: endfunc yading@10: yading@10: function ff_pred8x8_0lt_dc_neon, export=1 yading@10: sub r2, r0, r1 yading@10: vld1.8 {d0}, [r2,:64] yading@10: add r2, r0, r1, lsl #2 yading@10: sub r2, r2, #1 yading@10: ldcol.8 d1, r2, r1, 4, hi=1 yading@10: vtrn.32 d0, d1 yading@10: vpaddl.u8 q0, q0 yading@10: vpadd.u16 d0, d0, d1 yading@10: vpadd.u16 d1, d0, d0 yading@10: vrshrn.u16 d3, q0, #2 yading@10: vrshrn.u16 d2, q0, #3 yading@10: vdup.8 d0, d3[0] yading@10: vdup.8 d1, d3[3] yading@10: vdup.8 d4, d3[2] yading@10: vdup.8 d5, d2[5] yading@10: vtrn.32 q0, q2 yading@10: b .L_pred8x8_dc_end yading@10: endfunc yading@10: yading@10: function ff_pred8x8_0l0_dc_neon, export=1 yading@10: add r2, r0, r1, lsl #2 yading@10: sub r2, r2, #1 yading@10: ldcol.8 d1, r2, r1, 4 yading@10: vpaddl.u8 d2, d1 yading@10: vpadd.u16 d2, d2, d2 yading@10: vrshrn.u16 d1, q1, #2 yading@10: vmov.i8 d0, #128 yading@10: vdup.8 d1, d1[0] yading@10: b .L_pred8x8_dc_end yading@10: endfunc