yading@10: /* yading@10: * Copyright (c) 2011 Mans Rullgard yading@10: * yading@10: * This file is part of Libav. yading@10: * yading@10: * Libav 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: * Libav 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 Libav; 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_ac3_max_msb_abs_int16_neon, export=1 yading@10: vmov.i16 q0, #0 yading@10: vmov.i16 q2, #0 yading@10: 1: vld1.16 {q1}, [r0,:128]! yading@10: vabs.s16 q1, q1 yading@10: vld1.16 {q3}, [r0,:128]! yading@10: vabs.s16 q3, q3 yading@10: vorr q0, q0, q1 yading@10: vorr q2, q2, q3 yading@10: subs r1, r1, #16 yading@10: bgt 1b yading@10: vorr q0, q0, q2 yading@10: vorr d0, d0, d1 yading@10: vpmax.u16 d0, d0, d0 yading@10: vpmax.u16 d0, d0, d0 yading@10: vmov.u16 r0, d0[0] yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_ac3_exponent_min_neon, export=1 yading@10: cmp r1, #0 yading@10: it eq yading@10: bxeq lr yading@10: push {lr} yading@10: mov r12, #256 yading@10: 1: yading@10: vld1.8 {q0}, [r0,:128] yading@10: mov lr, r1 yading@10: add r3, r0, #256 yading@10: 2: vld1.8 {q1}, [r3,:128], r12 yading@10: subs lr, lr, #1 yading@10: vmin.u8 q0, q0, q1 yading@10: bgt 2b yading@10: subs r2, r2, #16 yading@10: vst1.8 {q0}, [r0,:128]! yading@10: bgt 1b yading@10: pop {pc} yading@10: endfunc yading@10: yading@10: function ff_ac3_lshift_int16_neon, export=1 yading@10: vdup.16 q0, r2 yading@10: 1: vld1.16 {q1}, [r0,:128] yading@10: vshl.s16 q1, q1, q0 yading@10: vst1.16 {q1}, [r0,:128]! yading@10: subs r1, r1, #8 yading@10: bgt 1b yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_ac3_rshift_int32_neon, export=1 yading@10: rsb r2, r2, #0 yading@10: vdup.32 q0, r2 yading@10: 1: vld1.32 {q1}, [r0,:128] yading@10: vshl.s32 q1, q1, q0 yading@10: vst1.32 {q1}, [r0,:128]! yading@10: subs r1, r1, #4 yading@10: bgt 1b yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_float_to_fixed24_neon, export=1 yading@10: 1: vld1.32 {q0-q1}, [r1,:128]! yading@10: vcvt.s32.f32 q0, q0, #24 yading@10: vld1.32 {q2-q3}, [r1,:128]! yading@10: vcvt.s32.f32 q1, q1, #24 yading@10: vcvt.s32.f32 q2, q2, #24 yading@10: vst1.32 {q0-q1}, [r0,:128]! yading@10: vcvt.s32.f32 q3, q3, #24 yading@10: vst1.32 {q2-q3}, [r0,:128]! yading@10: subs r2, r2, #16 yading@10: bgt 1b yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_ac3_extract_exponents_neon, export=1 yading@10: vmov.i32 q15, #8 yading@10: 1: yading@10: vld1.32 {q0}, [r1,:128]! yading@10: vabs.s32 q1, q0 yading@10: vclz.i32 q3, q1 yading@10: vsub.i32 q3, q3, q15 yading@10: vmovn.i32 d6, q3 yading@10: vmovn.i16 d6, q3 yading@10: vst1.32 {d6[0]}, [r0,:32]! yading@10: subs r2, r2, #4 yading@10: bgt 1b yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_ac3_sum_square_butterfly_int32_neon, export=1 yading@10: vmov.i64 q0, #0 yading@10: vmov.i64 q1, #0 yading@10: vmov.i64 q2, #0 yading@10: vmov.i64 q3, #0 yading@10: 1: yading@10: vld1.32 {d16}, [r1]! yading@10: vld1.32 {d17}, [r2]! yading@10: vadd.s32 d18, d16, d17 yading@10: vsub.s32 d19, d16, d17 yading@10: vmlal.s32 q0, d16, d16 yading@10: vmlal.s32 q1, d17, d17 yading@10: vmlal.s32 q2, d18, d18 yading@10: vmlal.s32 q3, d19, d19 yading@10: subs r3, r3, #2 yading@10: bgt 1b yading@10: vadd.s64 d0, d0, d1 yading@10: vadd.s64 d1, d2, d3 yading@10: vadd.s64 d2, d4, d5 yading@10: vadd.s64 d3, d6, d7 yading@10: vst1.64 {q0-q1}, [r0] yading@10: bx lr yading@10: endfunc yading@10: yading@10: function ff_ac3_sum_square_butterfly_float_neon, export=1 yading@10: vmov.f32 q0, #0.0 yading@10: vmov.f32 q1, #0.0 yading@10: 1: yading@10: vld1.32 {d16}, [r1]! yading@10: vld1.32 {d17}, [r2]! yading@10: vadd.f32 d18, d16, d17 yading@10: vsub.f32 d19, d16, d17 yading@10: vmla.f32 d0, d16, d16 yading@10: vmla.f32 d1, d17, d17 yading@10: vmla.f32 d2, d18, d18 yading@10: vmla.f32 d3, d19, d19 yading@10: subs r3, r3, #2 yading@10: bgt 1b yading@10: vpadd.f32 d0, d0, d1 yading@10: vpadd.f32 d1, d2, d3 yading@10: vst1.32 {q0}, [r0] yading@10: bx lr yading@10: endfunc