1cabdff1aSopenharmony_ci/*
2cabdff1aSopenharmony_ci * Copyright (c) 2008 Mans Rullgard <mans@mansr.com>
3cabdff1aSopenharmony_ci *
4cabdff1aSopenharmony_ci * This file is part of FFmpeg
5cabdff1aSopenharmony_ci *
6cabdff1aSopenharmony_ci * FFmpeg is free software; you can redistribute it and/or
7cabdff1aSopenharmony_ci * modify it under the terms of the GNU Lesser General Public
8cabdff1aSopenharmony_ci * License as published by the Free Software Foundation; either
9cabdff1aSopenharmony_ci * version 2.1 of the License, or (at your option) any later version.
10cabdff1aSopenharmony_ci *
11cabdff1aSopenharmony_ci * FFmpeg is distributed in the hope that it will be useful,
12cabdff1aSopenharmony_ci * but WITHOUT ANY WARRANTY; without even the implied warranty of
13cabdff1aSopenharmony_ci * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14cabdff1aSopenharmony_ci * Lesser General Public License for more details.
15cabdff1aSopenharmony_ci *
16cabdff1aSopenharmony_ci * You should have received a copy of the GNU Lesser General Public
17cabdff1aSopenharmony_ci * License along with FFmpeg; if not, write to the Free Software
18cabdff1aSopenharmony_ci * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19cabdff1aSopenharmony_ci */
20cabdff1aSopenharmony_ci
21cabdff1aSopenharmony_ci#include "config.h"
22cabdff1aSopenharmony_ci#include "libavutil/arm/asm.S"
23cabdff1aSopenharmony_ci
24cabdff1aSopenharmony_cifunction ff_conv_flt_to_s16_neon, export=1
25cabdff1aSopenharmony_ci        subs            r2,  r2,  #8
26cabdff1aSopenharmony_ci        vld1.32         {q0},     [r1,:128]!
27cabdff1aSopenharmony_ci        vcvt.s32.f32    q8,  q0,  #31
28cabdff1aSopenharmony_ci        vld1.32         {q1},     [r1,:128]!
29cabdff1aSopenharmony_ci        vcvt.s32.f32    q9,  q1,  #31
30cabdff1aSopenharmony_ci        beq             3f
31cabdff1aSopenharmony_ci        bics            r12, r2,  #15
32cabdff1aSopenharmony_ci        beq             2f
33cabdff1aSopenharmony_ci1:      subs            r12, r12, #16
34cabdff1aSopenharmony_ci        vqrshrn.s32     d4,  q8,  #16
35cabdff1aSopenharmony_ci        vld1.32         {q0},     [r1,:128]!
36cabdff1aSopenharmony_ci        vcvt.s32.f32    q0,  q0,  #31
37cabdff1aSopenharmony_ci        vqrshrn.s32     d5,  q9,  #16
38cabdff1aSopenharmony_ci        vld1.32         {q1},     [r1,:128]!
39cabdff1aSopenharmony_ci        vcvt.s32.f32    q1,  q1,  #31
40cabdff1aSopenharmony_ci        vqrshrn.s32     d6,  q0,  #16
41cabdff1aSopenharmony_ci        vst1.16         {q2},     [r0,:128]!
42cabdff1aSopenharmony_ci        vqrshrn.s32     d7,  q1,  #16
43cabdff1aSopenharmony_ci        vld1.32         {q8},     [r1,:128]!
44cabdff1aSopenharmony_ci        vcvt.s32.f32    q8,  q8,  #31
45cabdff1aSopenharmony_ci        vld1.32         {q9},     [r1,:128]!
46cabdff1aSopenharmony_ci        vcvt.s32.f32    q9,  q9,  #31
47cabdff1aSopenharmony_ci        vst1.16         {q3},     [r0,:128]!
48cabdff1aSopenharmony_ci        bne             1b
49cabdff1aSopenharmony_ci        ands            r2,  r2,  #15
50cabdff1aSopenharmony_ci        beq             3f
51cabdff1aSopenharmony_ci2:      vld1.32         {q0},     [r1,:128]!
52cabdff1aSopenharmony_ci        vqrshrn.s32     d4,  q8,  #16
53cabdff1aSopenharmony_ci        vcvt.s32.f32    q0,  q0,  #31
54cabdff1aSopenharmony_ci        vld1.32         {q1},     [r1,:128]!
55cabdff1aSopenharmony_ci        vqrshrn.s32     d5,  q9,  #16
56cabdff1aSopenharmony_ci        vcvt.s32.f32    q1,  q1,  #31
57cabdff1aSopenharmony_ci        vqrshrn.s32     d6,  q0,  #16
58cabdff1aSopenharmony_ci        vst1.16         {q2},     [r0,:128]!
59cabdff1aSopenharmony_ci        vqrshrn.s32     d7,  q1,  #16
60cabdff1aSopenharmony_ci        vst1.16         {q3},     [r0,:128]!
61cabdff1aSopenharmony_ci        bx              lr
62cabdff1aSopenharmony_ci3:      vqrshrn.s32     d4,  q8,  #16
63cabdff1aSopenharmony_ci        vqrshrn.s32     d5,  q9,  #16
64cabdff1aSopenharmony_ci        vst1.16         {q2},     [r0,:128]!
65cabdff1aSopenharmony_ci        bx              lr
66cabdff1aSopenharmony_ciendfunc
67cabdff1aSopenharmony_ci
68cabdff1aSopenharmony_cifunction ff_conv_fltp_to_s16_2ch_neon, export=1
69cabdff1aSopenharmony_ci        ldm             r1,  {r1, r3}
70cabdff1aSopenharmony_ci        subs            r2,  r2,  #8
71cabdff1aSopenharmony_ci        vld1.32         {q0},     [r1,:128]!
72cabdff1aSopenharmony_ci        vcvt.s32.f32    q8,  q0,  #31
73cabdff1aSopenharmony_ci        vld1.32         {q1},     [r1,:128]!
74cabdff1aSopenharmony_ci        vcvt.s32.f32    q9,  q1,  #31
75cabdff1aSopenharmony_ci        vld1.32         {q10},    [r3,:128]!
76cabdff1aSopenharmony_ci        vcvt.s32.f32    q10, q10, #31
77cabdff1aSopenharmony_ci        vld1.32         {q11},    [r3,:128]!
78cabdff1aSopenharmony_ci        vcvt.s32.f32    q11, q11, #31
79cabdff1aSopenharmony_ci        beq             3f
80cabdff1aSopenharmony_ci        bics            r12, r2,  #15
81cabdff1aSopenharmony_ci        beq             2f
82cabdff1aSopenharmony_ci1:      subs            r12, r12, #16
83cabdff1aSopenharmony_ci        vld1.32         {q0},     [r1,:128]!
84cabdff1aSopenharmony_ci        vcvt.s32.f32    q0,  q0,  #31
85cabdff1aSopenharmony_ci        vsri.32         q10, q8,  #16
86cabdff1aSopenharmony_ci        vld1.32         {q1},     [r1,:128]!
87cabdff1aSopenharmony_ci        vcvt.s32.f32    q1,  q1,  #31
88cabdff1aSopenharmony_ci        vld1.32         {q12},    [r3,:128]!
89cabdff1aSopenharmony_ci        vcvt.s32.f32    q12, q12, #31
90cabdff1aSopenharmony_ci        vld1.32         {q13},    [r3,:128]!
91cabdff1aSopenharmony_ci        vsri.32         q11, q9,  #16
92cabdff1aSopenharmony_ci        vst1.16         {q10},    [r0,:128]!
93cabdff1aSopenharmony_ci        vcvt.s32.f32    q13, q13, #31
94cabdff1aSopenharmony_ci        vst1.16         {q11},    [r0,:128]!
95cabdff1aSopenharmony_ci        vsri.32         q12, q0,  #16
96cabdff1aSopenharmony_ci        vld1.32         {q8},     [r1,:128]!
97cabdff1aSopenharmony_ci        vsri.32         q13, q1,  #16
98cabdff1aSopenharmony_ci        vst1.16         {q12},    [r0,:128]!
99cabdff1aSopenharmony_ci        vcvt.s32.f32    q8,  q8,  #31
100cabdff1aSopenharmony_ci        vld1.32         {q9},     [r1,:128]!
101cabdff1aSopenharmony_ci        vcvt.s32.f32    q9,  q9,  #31
102cabdff1aSopenharmony_ci        vld1.32         {q10},    [r3,:128]!
103cabdff1aSopenharmony_ci        vcvt.s32.f32    q10, q10, #31
104cabdff1aSopenharmony_ci        vld1.32         {q11},    [r3,:128]!
105cabdff1aSopenharmony_ci        vcvt.s32.f32    q11, q11, #31
106cabdff1aSopenharmony_ci        vst1.16         {q13},    [r0,:128]!
107cabdff1aSopenharmony_ci        bne             1b
108cabdff1aSopenharmony_ci        ands            r2,  r2,  #15
109cabdff1aSopenharmony_ci        beq             3f
110cabdff1aSopenharmony_ci2:      vsri.32         q10, q8,  #16
111cabdff1aSopenharmony_ci        vld1.32         {q0},     [r1,:128]!
112cabdff1aSopenharmony_ci        vcvt.s32.f32    q0,  q0,  #31
113cabdff1aSopenharmony_ci        vld1.32         {q1},     [r1,:128]!
114cabdff1aSopenharmony_ci        vcvt.s32.f32    q1,  q1,  #31
115cabdff1aSopenharmony_ci        vld1.32         {q12},    [r3,:128]!
116cabdff1aSopenharmony_ci        vcvt.s32.f32    q12, q12, #31
117cabdff1aSopenharmony_ci        vsri.32         q11, q9,  #16
118cabdff1aSopenharmony_ci        vld1.32         {q13},    [r3,:128]!
119cabdff1aSopenharmony_ci        vcvt.s32.f32    q13, q13, #31
120cabdff1aSopenharmony_ci        vst1.16         {q10},    [r0,:128]!
121cabdff1aSopenharmony_ci        vsri.32         q12, q0,  #16
122cabdff1aSopenharmony_ci        vst1.16         {q11},    [r0,:128]!
123cabdff1aSopenharmony_ci        vsri.32         q13, q1,  #16
124cabdff1aSopenharmony_ci        vst1.16         {q12-q13},[r0,:128]!
125cabdff1aSopenharmony_ci        bx              lr
126cabdff1aSopenharmony_ci3:      vsri.32         q10, q8,  #16
127cabdff1aSopenharmony_ci        vsri.32         q11, q9,  #16
128cabdff1aSopenharmony_ci        vst1.16         {q10-q11},[r0,:128]!
129cabdff1aSopenharmony_ci        bx              lr
130cabdff1aSopenharmony_ciendfunc
131cabdff1aSopenharmony_ci
132cabdff1aSopenharmony_cifunction ff_conv_fltp_to_s16_neon, export=1
133cabdff1aSopenharmony_ci        cmp             r3,  #2
134cabdff1aSopenharmony_ci        itt             lt
135cabdff1aSopenharmony_ci        ldrlt           r1,  [r1]
136cabdff1aSopenharmony_ci        blt             X(ff_conv_flt_to_s16_neon)
137cabdff1aSopenharmony_ci        beq             X(ff_conv_fltp_to_s16_2ch_neon)
138cabdff1aSopenharmony_ci
139cabdff1aSopenharmony_ci        push            {r4-r8, lr}
140cabdff1aSopenharmony_ci        cmp             r3,  #4
141cabdff1aSopenharmony_ci        lsl             r12, r3,  #1
142cabdff1aSopenharmony_ci        blt             4f
143cabdff1aSopenharmony_ci
144cabdff1aSopenharmony_ci        @ 4 channels
145cabdff1aSopenharmony_ci5:      ldm             r1!, {r4-r7}
146cabdff1aSopenharmony_ci        mov             lr,  r2
147cabdff1aSopenharmony_ci        mov             r8,  r0
148cabdff1aSopenharmony_ci        vld1.32         {q8},     [r4,:128]!
149cabdff1aSopenharmony_ci        vcvt.s32.f32    q8,  q8,  #31
150cabdff1aSopenharmony_ci        vld1.32         {q9},     [r5,:128]!
151cabdff1aSopenharmony_ci        vcvt.s32.f32    q9,  q9,  #31
152cabdff1aSopenharmony_ci        vld1.32         {q10},    [r6,:128]!
153cabdff1aSopenharmony_ci        vcvt.s32.f32    q10, q10, #31
154cabdff1aSopenharmony_ci        vld1.32         {q11},    [r7,:128]!
155cabdff1aSopenharmony_ci        vcvt.s32.f32    q11, q11, #31
156cabdff1aSopenharmony_ci6:      subs            lr,  lr,  #8
157cabdff1aSopenharmony_ci        vld1.32         {q0},     [r4,:128]!
158cabdff1aSopenharmony_ci        vcvt.s32.f32    q0,  q0,  #31
159cabdff1aSopenharmony_ci        vsri.32         q9,  q8,  #16
160cabdff1aSopenharmony_ci        vld1.32         {q1},     [r5,:128]!
161cabdff1aSopenharmony_ci        vcvt.s32.f32    q1,  q1,  #31
162cabdff1aSopenharmony_ci        vsri.32         q11, q10, #16
163cabdff1aSopenharmony_ci        vld1.32         {q2},     [r6,:128]!
164cabdff1aSopenharmony_ci        vcvt.s32.f32    q2,  q2,  #31
165cabdff1aSopenharmony_ci        vzip.32         d18, d22
166cabdff1aSopenharmony_ci        vld1.32         {q3},     [r7,:128]!
167cabdff1aSopenharmony_ci        vcvt.s32.f32    q3,  q3,  #31
168cabdff1aSopenharmony_ci        vzip.32         d19, d23
169cabdff1aSopenharmony_ci        vst1.16         {d18},    [r8], r12
170cabdff1aSopenharmony_ci        vsri.32         q1,  q0,  #16
171cabdff1aSopenharmony_ci        vst1.16         {d22},    [r8], r12
172cabdff1aSopenharmony_ci        vsri.32         q3,  q2,  #16
173cabdff1aSopenharmony_ci        vst1.16         {d19},    [r8], r12
174cabdff1aSopenharmony_ci        vzip.32         d2,  d6
175cabdff1aSopenharmony_ci        vst1.16         {d23},    [r8], r12
176cabdff1aSopenharmony_ci        vzip.32         d3,  d7
177cabdff1aSopenharmony_ci        beq             7f
178cabdff1aSopenharmony_ci        vld1.32         {q8},     [r4,:128]!
179cabdff1aSopenharmony_ci        vcvt.s32.f32    q8,  q8,  #31
180cabdff1aSopenharmony_ci        vst1.16         {d2},     [r8], r12
181cabdff1aSopenharmony_ci        vld1.32         {q9},     [r5,:128]!
182cabdff1aSopenharmony_ci        vcvt.s32.f32    q9,  q9,  #31
183cabdff1aSopenharmony_ci        vst1.16         {d6},     [r8], r12
184cabdff1aSopenharmony_ci        vld1.32         {q10},    [r6,:128]!
185cabdff1aSopenharmony_ci        vcvt.s32.f32    q10, q10, #31
186cabdff1aSopenharmony_ci        vst1.16         {d3},     [r8], r12
187cabdff1aSopenharmony_ci        vld1.32         {q11},    [r7,:128]!
188cabdff1aSopenharmony_ci        vcvt.s32.f32    q11, q11, #31
189cabdff1aSopenharmony_ci        vst1.16         {d7},     [r8], r12
190cabdff1aSopenharmony_ci        b               6b
191cabdff1aSopenharmony_ci7:      vst1.16         {d2},     [r8], r12
192cabdff1aSopenharmony_ci        vst1.16         {d6},     [r8], r12
193cabdff1aSopenharmony_ci        vst1.16         {d3},     [r8], r12
194cabdff1aSopenharmony_ci        vst1.16         {d7},     [r8], r12
195cabdff1aSopenharmony_ci        subs            r3,  r3,  #4
196cabdff1aSopenharmony_ci        it              eq
197cabdff1aSopenharmony_ci        popeq           {r4-r8, pc}
198cabdff1aSopenharmony_ci        cmp             r3,  #4
199cabdff1aSopenharmony_ci        add             r0,  r0,  #8
200cabdff1aSopenharmony_ci        bge             5b
201cabdff1aSopenharmony_ci
202cabdff1aSopenharmony_ci        @ 2 channels
203cabdff1aSopenharmony_ci4:      cmp             r3,  #2
204cabdff1aSopenharmony_ci        blt             4f
205cabdff1aSopenharmony_ci        ldm             r1!, {r4-r5}
206cabdff1aSopenharmony_ci        mov             lr,  r2
207cabdff1aSopenharmony_ci        mov             r8,  r0
208cabdff1aSopenharmony_ci        tst             lr,  #8
209cabdff1aSopenharmony_ci        vld1.32         {q8},     [r4,:128]!
210cabdff1aSopenharmony_ci        vcvt.s32.f32    q8,  q8,  #31
211cabdff1aSopenharmony_ci        vld1.32         {q9},     [r5,:128]!
212cabdff1aSopenharmony_ci        vcvt.s32.f32    q9,  q9,  #31
213cabdff1aSopenharmony_ci        vld1.32         {q10},    [r4,:128]!
214cabdff1aSopenharmony_ci        vcvt.s32.f32    q10, q10, #31
215cabdff1aSopenharmony_ci        vld1.32         {q11},    [r5,:128]!
216cabdff1aSopenharmony_ci        vcvt.s32.f32    q11, q11, #31
217cabdff1aSopenharmony_ci        beq             6f
218cabdff1aSopenharmony_ci        subs            lr,  lr,  #8
219cabdff1aSopenharmony_ci        beq             7f
220cabdff1aSopenharmony_ci        vsri.32         d18, d16, #16
221cabdff1aSopenharmony_ci        vsri.32         d19, d17, #16
222cabdff1aSopenharmony_ci        vld1.32         {q8},     [r4,:128]!
223cabdff1aSopenharmony_ci        vcvt.s32.f32    q8,  q8,  #31
224cabdff1aSopenharmony_ci        vst1.32         {d18[0]}, [r8], r12
225cabdff1aSopenharmony_ci        vsri.32         d22, d20, #16
226cabdff1aSopenharmony_ci        vst1.32         {d18[1]}, [r8], r12
227cabdff1aSopenharmony_ci        vsri.32         d23, d21, #16
228cabdff1aSopenharmony_ci        vst1.32         {d19[0]}, [r8], r12
229cabdff1aSopenharmony_ci        vst1.32         {d19[1]}, [r8], r12
230cabdff1aSopenharmony_ci        vld1.32         {q9},     [r5,:128]!
231cabdff1aSopenharmony_ci        vcvt.s32.f32    q9,  q9,  #31
232cabdff1aSopenharmony_ci        vst1.32         {d22[0]}, [r8], r12
233cabdff1aSopenharmony_ci        vst1.32         {d22[1]}, [r8], r12
234cabdff1aSopenharmony_ci        vld1.32         {q10},    [r4,:128]!
235cabdff1aSopenharmony_ci        vcvt.s32.f32    q10, q10, #31
236cabdff1aSopenharmony_ci        vst1.32         {d23[0]}, [r8], r12
237cabdff1aSopenharmony_ci        vst1.32         {d23[1]}, [r8], r12
238cabdff1aSopenharmony_ci        vld1.32         {q11},    [r5,:128]!
239cabdff1aSopenharmony_ci        vcvt.s32.f32    q11, q11, #31
240cabdff1aSopenharmony_ci6:      subs            lr,  lr,  #16
241cabdff1aSopenharmony_ci        vld1.32         {q0},     [r4,:128]!
242cabdff1aSopenharmony_ci        vcvt.s32.f32    q0,  q0,  #31
243cabdff1aSopenharmony_ci        vsri.32         d18, d16, #16
244cabdff1aSopenharmony_ci        vld1.32         {q1},     [r5,:128]!
245cabdff1aSopenharmony_ci        vcvt.s32.f32    q1,  q1,  #31
246cabdff1aSopenharmony_ci        vsri.32         d19, d17, #16
247cabdff1aSopenharmony_ci        vld1.32         {q2},     [r4,:128]!
248cabdff1aSopenharmony_ci        vcvt.s32.f32    q2,  q2,  #31
249cabdff1aSopenharmony_ci        vld1.32         {q3},     [r5,:128]!
250cabdff1aSopenharmony_ci        vcvt.s32.f32    q3,  q3,  #31
251cabdff1aSopenharmony_ci        vst1.32         {d18[0]}, [r8], r12
252cabdff1aSopenharmony_ci        vsri.32         d22, d20, #16
253cabdff1aSopenharmony_ci        vst1.32         {d18[1]}, [r8], r12
254cabdff1aSopenharmony_ci        vsri.32         d23, d21, #16
255cabdff1aSopenharmony_ci        vst1.32         {d19[0]}, [r8], r12
256cabdff1aSopenharmony_ci        vsri.32         d2,  d0,  #16
257cabdff1aSopenharmony_ci        vst1.32         {d19[1]}, [r8], r12
258cabdff1aSopenharmony_ci        vsri.32         d3,  d1,  #16
259cabdff1aSopenharmony_ci        vst1.32         {d22[0]}, [r8], r12
260cabdff1aSopenharmony_ci        vsri.32         d6,  d4,  #16
261cabdff1aSopenharmony_ci        vst1.32         {d22[1]}, [r8], r12
262cabdff1aSopenharmony_ci        vsri.32         d7,  d5,  #16
263cabdff1aSopenharmony_ci        vst1.32         {d23[0]}, [r8], r12
264cabdff1aSopenharmony_ci        vst1.32         {d23[1]}, [r8], r12
265cabdff1aSopenharmony_ci        beq             6f
266cabdff1aSopenharmony_ci        vld1.32         {q8},     [r4,:128]!
267cabdff1aSopenharmony_ci        vcvt.s32.f32    q8,  q8,  #31
268cabdff1aSopenharmony_ci        vst1.32         {d2[0]},  [r8], r12
269cabdff1aSopenharmony_ci        vst1.32         {d2[1]},  [r8], r12
270cabdff1aSopenharmony_ci        vld1.32         {q9},     [r5,:128]!
271cabdff1aSopenharmony_ci        vcvt.s32.f32    q9,  q9,  #31
272cabdff1aSopenharmony_ci        vst1.32         {d3[0]},  [r8], r12
273cabdff1aSopenharmony_ci        vst1.32         {d3[1]},  [r8], r12
274cabdff1aSopenharmony_ci        vld1.32         {q10},    [r4,:128]!
275cabdff1aSopenharmony_ci        vcvt.s32.f32    q10, q10, #31
276cabdff1aSopenharmony_ci        vst1.32         {d6[0]},  [r8], r12
277cabdff1aSopenharmony_ci        vst1.32         {d6[1]},  [r8], r12
278cabdff1aSopenharmony_ci        vld1.32         {q11},    [r5,:128]!
279cabdff1aSopenharmony_ci        vcvt.s32.f32    q11, q11, #31
280cabdff1aSopenharmony_ci        vst1.32         {d7[0]},  [r8], r12
281cabdff1aSopenharmony_ci        vst1.32         {d7[1]},  [r8], r12
282cabdff1aSopenharmony_ci        bgt             6b
283cabdff1aSopenharmony_ci6:      vst1.32         {d2[0]},  [r8], r12
284cabdff1aSopenharmony_ci        vst1.32         {d2[1]},  [r8], r12
285cabdff1aSopenharmony_ci        vst1.32         {d3[0]},  [r8], r12
286cabdff1aSopenharmony_ci        vst1.32         {d3[1]},  [r8], r12
287cabdff1aSopenharmony_ci        vst1.32         {d6[0]},  [r8], r12
288cabdff1aSopenharmony_ci        vst1.32         {d6[1]},  [r8], r12
289cabdff1aSopenharmony_ci        vst1.32         {d7[0]},  [r8], r12
290cabdff1aSopenharmony_ci        vst1.32         {d7[1]},  [r8], r12
291cabdff1aSopenharmony_ci        b               8f
292cabdff1aSopenharmony_ci7:      vsri.32         d18, d16, #16
293cabdff1aSopenharmony_ci        vsri.32         d19, d17, #16
294cabdff1aSopenharmony_ci        vst1.32         {d18[0]}, [r8], r12
295cabdff1aSopenharmony_ci        vsri.32         d22, d20, #16
296cabdff1aSopenharmony_ci        vst1.32         {d18[1]}, [r8], r12
297cabdff1aSopenharmony_ci        vsri.32         d23, d21, #16
298cabdff1aSopenharmony_ci        vst1.32         {d19[0]}, [r8], r12
299cabdff1aSopenharmony_ci        vst1.32         {d19[1]}, [r8], r12
300cabdff1aSopenharmony_ci        vst1.32         {d22[0]}, [r8], r12
301cabdff1aSopenharmony_ci        vst1.32         {d22[1]}, [r8], r12
302cabdff1aSopenharmony_ci        vst1.32         {d23[0]}, [r8], r12
303cabdff1aSopenharmony_ci        vst1.32         {d23[1]}, [r8], r12
304cabdff1aSopenharmony_ci8:      subs            r3,  r3,  #2
305cabdff1aSopenharmony_ci        add             r0,  r0,  #4
306cabdff1aSopenharmony_ci        it              eq
307cabdff1aSopenharmony_ci        popeq           {r4-r8, pc}
308cabdff1aSopenharmony_ci
309cabdff1aSopenharmony_ci        @ 1 channel
310cabdff1aSopenharmony_ci4:      ldr             r4,  [r1]
311cabdff1aSopenharmony_ci        tst             r2,  #8
312cabdff1aSopenharmony_ci        mov             lr,  r2
313cabdff1aSopenharmony_ci        mov             r5,  r0
314cabdff1aSopenharmony_ci        vld1.32         {q0},     [r4,:128]!
315cabdff1aSopenharmony_ci        vcvt.s32.f32    q0,  q0,  #31
316cabdff1aSopenharmony_ci        vld1.32         {q1},     [r4,:128]!
317cabdff1aSopenharmony_ci        vcvt.s32.f32    q1,  q1,  #31
318cabdff1aSopenharmony_ci        bne             8f
319cabdff1aSopenharmony_ci6:      subs            lr,  lr,  #16
320cabdff1aSopenharmony_ci        vld1.32         {q2},     [r4,:128]!
321cabdff1aSopenharmony_ci        vcvt.s32.f32    q2,  q2,  #31
322cabdff1aSopenharmony_ci        vld1.32         {q3},     [r4,:128]!
323cabdff1aSopenharmony_ci        vcvt.s32.f32    q3,  q3,  #31
324cabdff1aSopenharmony_ci        vst1.16         {d0[1]},  [r5,:16], r12
325cabdff1aSopenharmony_ci        vst1.16         {d0[3]},  [r5,:16], r12
326cabdff1aSopenharmony_ci        vst1.16         {d1[1]},  [r5,:16], r12
327cabdff1aSopenharmony_ci        vst1.16         {d1[3]},  [r5,:16], r12
328cabdff1aSopenharmony_ci        vst1.16         {d2[1]},  [r5,:16], r12
329cabdff1aSopenharmony_ci        vst1.16         {d2[3]},  [r5,:16], r12
330cabdff1aSopenharmony_ci        vst1.16         {d3[1]},  [r5,:16], r12
331cabdff1aSopenharmony_ci        vst1.16         {d3[3]},  [r5,:16], r12
332cabdff1aSopenharmony_ci        beq             7f
333cabdff1aSopenharmony_ci        vld1.32         {q0},     [r4,:128]!
334cabdff1aSopenharmony_ci        vcvt.s32.f32    q0,  q0,  #31
335cabdff1aSopenharmony_ci        vld1.32         {q1},     [r4,:128]!
336cabdff1aSopenharmony_ci        vcvt.s32.f32    q1,  q1,  #31
337cabdff1aSopenharmony_ci7:      vst1.16         {d4[1]},  [r5,:16], r12
338cabdff1aSopenharmony_ci        vst1.16         {d4[3]},  [r5,:16], r12
339cabdff1aSopenharmony_ci        vst1.16         {d5[1]},  [r5,:16], r12
340cabdff1aSopenharmony_ci        vst1.16         {d5[3]},  [r5,:16], r12
341cabdff1aSopenharmony_ci        vst1.16         {d6[1]},  [r5,:16], r12
342cabdff1aSopenharmony_ci        vst1.16         {d6[3]},  [r5,:16], r12
343cabdff1aSopenharmony_ci        vst1.16         {d7[1]},  [r5,:16], r12
344cabdff1aSopenharmony_ci        vst1.16         {d7[3]},  [r5,:16], r12
345cabdff1aSopenharmony_ci        bgt             6b
346cabdff1aSopenharmony_ci        pop             {r4-r8, pc}
347cabdff1aSopenharmony_ci8:      subs            lr,  lr,  #8
348cabdff1aSopenharmony_ci        vst1.16         {d0[1]},  [r5,:16], r12
349cabdff1aSopenharmony_ci        vst1.16         {d0[3]},  [r5,:16], r12
350cabdff1aSopenharmony_ci        vst1.16         {d1[1]},  [r5,:16], r12
351cabdff1aSopenharmony_ci        vst1.16         {d1[3]},  [r5,:16], r12
352cabdff1aSopenharmony_ci        vst1.16         {d2[1]},  [r5,:16], r12
353cabdff1aSopenharmony_ci        vst1.16         {d2[3]},  [r5,:16], r12
354cabdff1aSopenharmony_ci        vst1.16         {d3[1]},  [r5,:16], r12
355cabdff1aSopenharmony_ci        vst1.16         {d3[3]},  [r5,:16], r12
356cabdff1aSopenharmony_ci        it              eq
357cabdff1aSopenharmony_ci        popeq           {r4-r8, pc}
358cabdff1aSopenharmony_ci        vld1.32         {q0},     [r4,:128]!
359cabdff1aSopenharmony_ci        vcvt.s32.f32    q0,  q0,  #31
360cabdff1aSopenharmony_ci        vld1.32         {q1},     [r4,:128]!
361cabdff1aSopenharmony_ci        vcvt.s32.f32    q1,  q1,  #31
362cabdff1aSopenharmony_ci        b               6b
363cabdff1aSopenharmony_ciendfunc
364