yading@11: /* yading@11: * Copyright (c) 2008 Mans Rullgard yading@11: * yading@11: * This file is part of FFmpeg yading@11: * yading@11: * FFmpeg is free software; you can redistribute it and/or yading@11: * modify it under the terms of the GNU Lesser General Public yading@11: * License as published by the Free Software Foundation; either yading@11: * version 2.1 of the License, or (at your option) any later version. yading@11: * yading@11: * FFmpeg is distributed in the hope that it will be useful, yading@11: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@11: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@11: * Lesser General Public License for more details. yading@11: * yading@11: * You should have received a copy of the GNU Lesser General Public yading@11: * License along with FFmpeg; if not, write to the Free Software yading@11: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@11: */ yading@11: yading@11: #include "config.h" yading@11: #include "libavutil/arm/asm.S" yading@11: yading@11: function ff_conv_flt_to_s16_neon, export=1 yading@11: subs r2, r2, #8 yading@11: vld1.32 {q0}, [r1,:128]! yading@11: vcvt.s32.f32 q8, q0, #31 yading@11: vld1.32 {q1}, [r1,:128]! yading@11: vcvt.s32.f32 q9, q1, #31 yading@11: beq 3f yading@11: bics r12, r2, #15 yading@11: beq 2f yading@11: 1: subs r12, r12, #16 yading@11: vqrshrn.s32 d4, q8, #16 yading@11: vld1.32 {q0}, [r1,:128]! yading@11: vcvt.s32.f32 q0, q0, #31 yading@11: vqrshrn.s32 d5, q9, #16 yading@11: vld1.32 {q1}, [r1,:128]! yading@11: vcvt.s32.f32 q1, q1, #31 yading@11: vqrshrn.s32 d6, q0, #16 yading@11: vst1.16 {q2}, [r0,:128]! yading@11: vqrshrn.s32 d7, q1, #16 yading@11: vld1.32 {q8}, [r1,:128]! yading@11: vcvt.s32.f32 q8, q8, #31 yading@11: vld1.32 {q9}, [r1,:128]! yading@11: vcvt.s32.f32 q9, q9, #31 yading@11: vst1.16 {q3}, [r0,:128]! yading@11: bne 1b yading@11: ands r2, r2, #15 yading@11: beq 3f yading@11: 2: vld1.32 {q0}, [r1,:128]! yading@11: vqrshrn.s32 d4, q8, #16 yading@11: vcvt.s32.f32 q0, q0, #31 yading@11: vld1.32 {q1}, [r1,:128]! yading@11: vqrshrn.s32 d5, q9, #16 yading@11: vcvt.s32.f32 q1, q1, #31 yading@11: vqrshrn.s32 d6, q0, #16 yading@11: vst1.16 {q2}, [r0,:128]! yading@11: vqrshrn.s32 d7, q1, #16 yading@11: vst1.16 {q3}, [r0,:128]! yading@11: bx lr yading@11: 3: vqrshrn.s32 d4, q8, #16 yading@11: vqrshrn.s32 d5, q9, #16 yading@11: vst1.16 {q2}, [r0,:128]! yading@11: bx lr yading@11: endfunc yading@11: yading@11: function ff_conv_fltp_to_s16_2ch_neon, export=1 yading@11: ldm r1, {r1, r3} yading@11: subs r2, r2, #8 yading@11: vld1.32 {q0}, [r1,:128]! yading@11: vcvt.s32.f32 q8, q0, #31 yading@11: vld1.32 {q1}, [r1,:128]! yading@11: vcvt.s32.f32 q9, q1, #31 yading@11: vld1.32 {q10}, [r3,:128]! yading@11: vcvt.s32.f32 q10, q10, #31 yading@11: vld1.32 {q11}, [r3,:128]! yading@11: vcvt.s32.f32 q11, q11, #31 yading@11: beq 3f yading@11: bics r12, r2, #15 yading@11: beq 2f yading@11: 1: subs r12, r12, #16 yading@11: vld1.32 {q0}, [r1,:128]! yading@11: vcvt.s32.f32 q0, q0, #31 yading@11: vsri.32 q10, q8, #16 yading@11: vld1.32 {q1}, [r1,:128]! yading@11: vcvt.s32.f32 q1, q1, #31 yading@11: vld1.32 {q12}, [r3,:128]! yading@11: vcvt.s32.f32 q12, q12, #31 yading@11: vld1.32 {q13}, [r3,:128]! yading@11: vsri.32 q11, q9, #16 yading@11: vst1.16 {q10}, [r0,:128]! yading@11: vcvt.s32.f32 q13, q13, #31 yading@11: vst1.16 {q11}, [r0,:128]! yading@11: vsri.32 q12, q0, #16 yading@11: vld1.32 {q8}, [r1,:128]! yading@11: vsri.32 q13, q1, #16 yading@11: vst1.16 {q12}, [r0,:128]! yading@11: vcvt.s32.f32 q8, q8, #31 yading@11: vld1.32 {q9}, [r1,:128]! yading@11: vcvt.s32.f32 q9, q9, #31 yading@11: vld1.32 {q10}, [r3,:128]! yading@11: vcvt.s32.f32 q10, q10, #31 yading@11: vld1.32 {q11}, [r3,:128]! yading@11: vcvt.s32.f32 q11, q11, #31 yading@11: vst1.16 {q13}, [r0,:128]! yading@11: bne 1b yading@11: ands r2, r2, #15 yading@11: beq 3f yading@11: 2: vsri.32 q10, q8, #16 yading@11: vld1.32 {q0}, [r1,:128]! yading@11: vcvt.s32.f32 q0, q0, #31 yading@11: vld1.32 {q1}, [r1,:128]! yading@11: vcvt.s32.f32 q1, q1, #31 yading@11: vld1.32 {q12}, [r3,:128]! yading@11: vcvt.s32.f32 q12, q12, #31 yading@11: vsri.32 q11, q9, #16 yading@11: vld1.32 {q13}, [r3,:128]! yading@11: vcvt.s32.f32 q13, q13, #31 yading@11: vst1.16 {q10}, [r0,:128]! yading@11: vsri.32 q12, q0, #16 yading@11: vst1.16 {q11}, [r0,:128]! yading@11: vsri.32 q13, q1, #16 yading@11: vst1.16 {q12-q13},[r0,:128]! yading@11: bx lr yading@11: 3: vsri.32 q10, q8, #16 yading@11: vsri.32 q11, q9, #16 yading@11: vst1.16 {q10-q11},[r0,:128]! yading@11: bx lr yading@11: endfunc yading@11: yading@11: function ff_conv_fltp_to_s16_neon, export=1 yading@11: cmp r3, #2 yading@11: itt lt yading@11: ldrlt r1, [r1] yading@11: blt ff_conv_flt_to_s16_neon yading@11: beq ff_conv_fltp_to_s16_2ch_neon yading@11: yading@11: push {r4-r8, lr} yading@11: cmp r3, #4 yading@11: lsl r12, r3, #1 yading@11: blt 4f yading@11: yading@11: @ 4 channels yading@11: 5: ldm r1!, {r4-r7} yading@11: mov lr, r2 yading@11: mov r8, r0 yading@11: vld1.32 {q8}, [r4,:128]! yading@11: vcvt.s32.f32 q8, q8, #31 yading@11: vld1.32 {q9}, [r5,:128]! yading@11: vcvt.s32.f32 q9, q9, #31 yading@11: vld1.32 {q10}, [r6,:128]! yading@11: vcvt.s32.f32 q10, q10, #31 yading@11: vld1.32 {q11}, [r7,:128]! yading@11: vcvt.s32.f32 q11, q11, #31 yading@11: 6: subs lr, lr, #8 yading@11: vld1.32 {q0}, [r4,:128]! yading@11: vcvt.s32.f32 q0, q0, #31 yading@11: vsri.32 q9, q8, #16 yading@11: vld1.32 {q1}, [r5,:128]! yading@11: vcvt.s32.f32 q1, q1, #31 yading@11: vsri.32 q11, q10, #16 yading@11: vld1.32 {q2}, [r6,:128]! yading@11: vcvt.s32.f32 q2, q2, #31 yading@11: vzip.32 d18, d22 yading@11: vld1.32 {q3}, [r7,:128]! yading@11: vcvt.s32.f32 q3, q3, #31 yading@11: vzip.32 d19, d23 yading@11: vst1.16 {d18}, [r8], r12 yading@11: vsri.32 q1, q0, #16 yading@11: vst1.16 {d22}, [r8], r12 yading@11: vsri.32 q3, q2, #16 yading@11: vst1.16 {d19}, [r8], r12 yading@11: vzip.32 d2, d6 yading@11: vst1.16 {d23}, [r8], r12 yading@11: vzip.32 d3, d7 yading@11: beq 7f yading@11: vld1.32 {q8}, [r4,:128]! yading@11: vcvt.s32.f32 q8, q8, #31 yading@11: vst1.16 {d2}, [r8], r12 yading@11: vld1.32 {q9}, [r5,:128]! yading@11: vcvt.s32.f32 q9, q9, #31 yading@11: vst1.16 {d6}, [r8], r12 yading@11: vld1.32 {q10}, [r6,:128]! yading@11: vcvt.s32.f32 q10, q10, #31 yading@11: vst1.16 {d3}, [r8], r12 yading@11: vld1.32 {q11}, [r7,:128]! yading@11: vcvt.s32.f32 q11, q11, #31 yading@11: vst1.16 {d7}, [r8], r12 yading@11: b 6b yading@11: 7: vst1.16 {d2}, [r8], r12 yading@11: vst1.16 {d6}, [r8], r12 yading@11: vst1.16 {d3}, [r8], r12 yading@11: vst1.16 {d7}, [r8], r12 yading@11: subs r3, r3, #4 yading@11: it eq yading@11: popeq {r4-r8, pc} yading@11: cmp r3, #4 yading@11: add r0, r0, #8 yading@11: bge 5b yading@11: yading@11: @ 2 channels yading@11: 4: cmp r3, #2 yading@11: blt 4f yading@11: ldm r1!, {r4-r5} yading@11: mov lr, r2 yading@11: mov r8, r0 yading@11: tst lr, #8 yading@11: vld1.32 {q8}, [r4,:128]! yading@11: vcvt.s32.f32 q8, q8, #31 yading@11: vld1.32 {q9}, [r5,:128]! yading@11: vcvt.s32.f32 q9, q9, #31 yading@11: vld1.32 {q10}, [r4,:128]! yading@11: vcvt.s32.f32 q10, q10, #31 yading@11: vld1.32 {q11}, [r5,:128]! yading@11: vcvt.s32.f32 q11, q11, #31 yading@11: beq 6f yading@11: subs lr, lr, #8 yading@11: beq 7f yading@11: vsri.32 d18, d16, #16 yading@11: vsri.32 d19, d17, #16 yading@11: vld1.32 {q8}, [r4,:128]! yading@11: vcvt.s32.f32 q8, q8, #31 yading@11: vst1.32 {d18[0]}, [r8], r12 yading@11: vsri.32 d22, d20, #16 yading@11: vst1.32 {d18[1]}, [r8], r12 yading@11: vsri.32 d23, d21, #16 yading@11: vst1.32 {d19[0]}, [r8], r12 yading@11: vst1.32 {d19[1]}, [r8], r12 yading@11: vld1.32 {q9}, [r5,:128]! yading@11: vcvt.s32.f32 q9, q9, #31 yading@11: vst1.32 {d22[0]}, [r8], r12 yading@11: vst1.32 {d22[1]}, [r8], r12 yading@11: vld1.32 {q10}, [r4,:128]! yading@11: vcvt.s32.f32 q10, q10, #31 yading@11: vst1.32 {d23[0]}, [r8], r12 yading@11: vst1.32 {d23[1]}, [r8], r12 yading@11: vld1.32 {q11}, [r5,:128]! yading@11: vcvt.s32.f32 q11, q11, #31 yading@11: 6: subs lr, lr, #16 yading@11: vld1.32 {q0}, [r4,:128]! yading@11: vcvt.s32.f32 q0, q0, #31 yading@11: vsri.32 d18, d16, #16 yading@11: vld1.32 {q1}, [r5,:128]! yading@11: vcvt.s32.f32 q1, q1, #31 yading@11: vsri.32 d19, d17, #16 yading@11: vld1.32 {q2}, [r4,:128]! yading@11: vcvt.s32.f32 q2, q2, #31 yading@11: vld1.32 {q3}, [r5,:128]! yading@11: vcvt.s32.f32 q3, q3, #31 yading@11: vst1.32 {d18[0]}, [r8], r12 yading@11: vsri.32 d22, d20, #16 yading@11: vst1.32 {d18[1]}, [r8], r12 yading@11: vsri.32 d23, d21, #16 yading@11: vst1.32 {d19[0]}, [r8], r12 yading@11: vsri.32 d2, d0, #16 yading@11: vst1.32 {d19[1]}, [r8], r12 yading@11: vsri.32 d3, d1, #16 yading@11: vst1.32 {d22[0]}, [r8], r12 yading@11: vsri.32 d6, d4, #16 yading@11: vst1.32 {d22[1]}, [r8], r12 yading@11: vsri.32 d7, d5, #16 yading@11: vst1.32 {d23[0]}, [r8], r12 yading@11: vst1.32 {d23[1]}, [r8], r12 yading@11: beq 6f yading@11: vld1.32 {q8}, [r4,:128]! yading@11: vcvt.s32.f32 q8, q8, #31 yading@11: vst1.32 {d2[0]}, [r8], r12 yading@11: vst1.32 {d2[1]}, [r8], r12 yading@11: vld1.32 {q9}, [r5,:128]! yading@11: vcvt.s32.f32 q9, q9, #31 yading@11: vst1.32 {d3[0]}, [r8], r12 yading@11: vst1.32 {d3[1]}, [r8], r12 yading@11: vld1.32 {q10}, [r4,:128]! yading@11: vcvt.s32.f32 q10, q10, #31 yading@11: vst1.32 {d6[0]}, [r8], r12 yading@11: vst1.32 {d6[1]}, [r8], r12 yading@11: vld1.32 {q11}, [r5,:128]! yading@11: vcvt.s32.f32 q11, q11, #31 yading@11: vst1.32 {d7[0]}, [r8], r12 yading@11: vst1.32 {d7[1]}, [r8], r12 yading@11: bgt 6b yading@11: 6: vst1.32 {d2[0]}, [r8], r12 yading@11: vst1.32 {d2[1]}, [r8], r12 yading@11: vst1.32 {d3[0]}, [r8], r12 yading@11: vst1.32 {d3[1]}, [r8], r12 yading@11: vst1.32 {d6[0]}, [r8], r12 yading@11: vst1.32 {d6[1]}, [r8], r12 yading@11: vst1.32 {d7[0]}, [r8], r12 yading@11: vst1.32 {d7[1]}, [r8], r12 yading@11: b 8f yading@11: 7: vsri.32 d18, d16, #16 yading@11: vsri.32 d19, d17, #16 yading@11: vst1.32 {d18[0]}, [r8], r12 yading@11: vsri.32 d22, d20, #16 yading@11: vst1.32 {d18[1]}, [r8], r12 yading@11: vsri.32 d23, d21, #16 yading@11: vst1.32 {d19[0]}, [r8], r12 yading@11: vst1.32 {d19[1]}, [r8], r12 yading@11: vst1.32 {d22[0]}, [r8], r12 yading@11: vst1.32 {d22[1]}, [r8], r12 yading@11: vst1.32 {d23[0]}, [r8], r12 yading@11: vst1.32 {d23[1]}, [r8], r12 yading@11: 8: subs r3, r3, #2 yading@11: add r0, r0, #4 yading@11: it eq yading@11: popeq {r4-r8, pc} yading@11: yading@11: @ 1 channel yading@11: 4: ldr r4, [r1] yading@11: tst r2, #8 yading@11: mov lr, r2 yading@11: mov r5, r0 yading@11: vld1.32 {q0}, [r4,:128]! yading@11: vcvt.s32.f32 q0, q0, #31 yading@11: vld1.32 {q1}, [r4,:128]! yading@11: vcvt.s32.f32 q1, q1, #31 yading@11: bne 8f yading@11: 6: subs lr, lr, #16 yading@11: vld1.32 {q2}, [r4,:128]! yading@11: vcvt.s32.f32 q2, q2, #31 yading@11: vld1.32 {q3}, [r4,:128]! yading@11: vcvt.s32.f32 q3, q3, #31 yading@11: vst1.16 {d0[1]}, [r5,:16], r12 yading@11: vst1.16 {d0[3]}, [r5,:16], r12 yading@11: vst1.16 {d1[1]}, [r5,:16], r12 yading@11: vst1.16 {d1[3]}, [r5,:16], r12 yading@11: vst1.16 {d2[1]}, [r5,:16], r12 yading@11: vst1.16 {d2[3]}, [r5,:16], r12 yading@11: vst1.16 {d3[1]}, [r5,:16], r12 yading@11: vst1.16 {d3[3]}, [r5,:16], r12 yading@11: beq 7f yading@11: vld1.32 {q0}, [r4,:128]! yading@11: vcvt.s32.f32 q0, q0, #31 yading@11: vld1.32 {q1}, [r4,:128]! yading@11: vcvt.s32.f32 q1, q1, #31 yading@11: 7: vst1.16 {d4[1]}, [r5,:16], r12 yading@11: vst1.16 {d4[3]}, [r5,:16], r12 yading@11: vst1.16 {d5[1]}, [r5,:16], r12 yading@11: vst1.16 {d5[3]}, [r5,:16], r12 yading@11: vst1.16 {d6[1]}, [r5,:16], r12 yading@11: vst1.16 {d6[3]}, [r5,:16], r12 yading@11: vst1.16 {d7[1]}, [r5,:16], r12 yading@11: vst1.16 {d7[3]}, [r5,:16], r12 yading@11: bgt 6b yading@11: pop {r4-r8, pc} yading@11: 8: subs lr, lr, #8 yading@11: vst1.16 {d0[1]}, [r5,:16], r12 yading@11: vst1.16 {d0[3]}, [r5,:16], r12 yading@11: vst1.16 {d1[1]}, [r5,:16], r12 yading@11: vst1.16 {d1[3]}, [r5,:16], r12 yading@11: vst1.16 {d2[1]}, [r5,:16], r12 yading@11: vst1.16 {d2[3]}, [r5,:16], r12 yading@11: vst1.16 {d3[1]}, [r5,:16], r12 yading@11: vst1.16 {d3[3]}, [r5,:16], r12 yading@11: it eq yading@11: popeq {r4-r8, pc} yading@11: vld1.32 {q0}, [r4,:128]! yading@11: vcvt.s32.f32 q0, q0, #31 yading@11: vld1.32 {q1}, [r4,:128]! yading@11: vcvt.s32.f32 q1, q1, #31 yading@11: b 6b yading@11: endfunc