xref: /aosp_15_r20/external/XNNPACK/src/qu8-gavgpool/gen/7p7x-minmax-fp32-neonv8-c24.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gavgpool/multipass-neon.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/gavgpool.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 
18 
xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24(size_t rows,size_t channels,const uint8_t * input,size_t input_stride,const uint8_t * zero,int32_t * buffer,uint8_t * output,const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24(
20     size_t rows,
21     size_t channels,
22     const uint8_t* input,
23     size_t input_stride,
24     const uint8_t* zero,
25     int32_t* buffer,
26     uint8_t* output,
27     const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(rows > 7);
30   assert(channels != 0);
31 
32   const uint8_t* i0 = input;
33   const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
34   const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
35   const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
36   const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
37   const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
38   const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
39   const size_t input_increment = 7 * input_stride - round_up_po2(channels, 8) * sizeof(uint8_t);
40 
41   const int32x4_t vinit_bias = vld1q_dup_s32(&params->fp32_neonv8.init_bias);
42   int32_t* b = buffer;
43   size_t c = channels;
44   for (; c >= 24; c -= 24) {
45     const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
46     const uint8x8_t vi0x89ABCDEF = vld1_u8(i0); i0 += 8;
47     const uint8x8_t vi0xGHIJKLMN = vld1_u8(i0); i0 += 8;
48     const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
49     const uint8x8_t vi1x89ABCDEF = vld1_u8(i1); i1 += 8;
50     const uint8x8_t vi1xGHIJKLMN = vld1_u8(i1); i1 += 8;
51 
52     const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
53     uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
54     const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
55     uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
56     const uint8x8_t vi2xGHIJKLMN = vld1_u8(i2); i2 += 8;
57     uint16x8_t vsumGHIJKLMN = vaddl_u8(vi0xGHIJKLMN, vi1xGHIJKLMN);
58 
59     const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
60     vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
61     const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
62     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
63     const uint8x8_t vi3xGHIJKLMN = vld1_u8(i3); i3 += 8;
64     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi2xGHIJKLMN);
65     const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
66     vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
67     const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
68     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
69     const uint8x8_t vi4xGHIJKLMN = vld1_u8(i4); i4 += 8;
70     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi3xGHIJKLMN);
71     const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
72     vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
73     const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
74     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
75     const uint8x8_t vi5xGHIJKLMN = vld1_u8(i5); i5 += 8;
76     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi4xGHIJKLMN);
77     const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
78     vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
79     const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
80     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
81     const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8;
82     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi5xGHIJKLMN);
83     vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
84     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
85     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN);
86 
87     const int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
88     const int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
89     const int32x4_t vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum89ABCDEF)));
90     const int32x4_t vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum89ABCDEF)));
91     const int32x4_t vaccGHIJ = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumGHIJKLMN)));
92     const int32x4_t vaccKLMN = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumGHIJKLMN)));
93 
94     vst1q_s32(b, vacc0123); b += 4;
95     vst1q_s32(b, vacc4567); b += 4;
96     vst1q_s32(b, vacc89AB); b += 4;
97     vst1q_s32(b, vaccCDEF); b += 4;
98     vst1q_s32(b, vaccGHIJ); b += 4;
99     vst1q_s32(b, vaccKLMN); b += 4;
100   }
101   if XNN_UNLIKELY(c != 0) {
102     do {
103       const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
104       const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
105       const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
106       uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
107 
108       const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
109       vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
110       const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
111       vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
112       const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
113       vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
114       const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
115       vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
116       vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
117 
118       const int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
119       const int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
120 
121       vst1q_s32(b, vacc0123); b += 4;
122       vst1q_s32(b, vacc4567); b += 4;
123 
124       c = doz(c, 8);
125     } while (c != 0);
126   }
127 
128   for (rows -= 7; rows > 7; rows -= 7) {
129     i0 = (const uint8_t*) ((uintptr_t) i0 + input_increment);
130     i1 = (const uint8_t*) ((uintptr_t) i1 + input_increment);
131     i2 = (const uint8_t*) ((uintptr_t) i2 + input_increment);
132     i3 = (const uint8_t*) ((uintptr_t) i3 + input_increment);
133     i4 = (const uint8_t*) ((uintptr_t) i4 + input_increment);
134     i5 = (const uint8_t*) ((uintptr_t) i5 + input_increment);
135     i6 = (const uint8_t*) ((uintptr_t) i6 + input_increment);
136 
137     int32_t* b = buffer;
138     size_t c = channels;
139     for (; c >= 24; c -= 24) {
140       const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
141       const uint8x8_t vi0x89ABCDEF = vld1_u8(i0); i0 += 8;
142       const uint8x8_t vi0xGHIJKLMN = vld1_u8(i0); i0 += 8;
143       const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
144       const uint8x8_t vi1x89ABCDEF = vld1_u8(i1); i1 += 8;
145       const uint8x8_t vi1xGHIJKLMN = vld1_u8(i1); i1 += 8;
146 
147       const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
148       uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
149       const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
150       uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
151       const uint8x8_t vi2xGHIJKLMN = vld1_u8(i2); i2 += 8;
152       uint16x8_t vsumGHIJKLMN = vaddl_u8(vi0xGHIJKLMN, vi1xGHIJKLMN);
153 
154       const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
155       vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
156       const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
157       vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
158       const uint8x8_t vi3xGHIJKLMN = vld1_u8(i3); i3 += 8;
159       vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi2xGHIJKLMN);
160       const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
161       vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
162       const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
163       vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
164       const uint8x8_t vi4xGHIJKLMN = vld1_u8(i4); i4 += 8;
165       vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi3xGHIJKLMN);
166       const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
167       vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
168       const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
169       vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
170       const uint8x8_t vi5xGHIJKLMN = vld1_u8(i5); i5 += 8;
171       vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi4xGHIJKLMN);
172       const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
173       vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
174       const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
175       vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
176       const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8;
177       vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi5xGHIJKLMN);
178       int32x4_t vacc0123 = vld1q_s32(b);
179       int32x4_t vacc4567 = vld1q_s32(b + 4);
180       vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
181       int32x4_t vacc89AB = vld1q_s32(b + 8);
182       int32x4_t vaccCDEF = vld1q_s32(b + 12);
183       vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
184       int32x4_t vaccGHIJ = vld1q_s32(b + 16);
185       int32x4_t vaccKLMN = vld1q_s32(b + 20);
186       vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN);
187 
188       vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567)));
189       vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567)));
190       vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vsum89ABCDEF)));
191       vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vsum89ABCDEF)));
192       vaccGHIJ = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccGHIJ), vget_low_u16(vsumGHIJKLMN)));
193       vaccKLMN = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccKLMN), vget_high_u16(vsumGHIJKLMN)));
194 
195       vst1q_s32(b, vacc0123); b += 4;
196       vst1q_s32(b, vacc4567); b += 4;
197       vst1q_s32(b, vacc89AB); b += 4;
198       vst1q_s32(b, vaccCDEF); b += 4;
199       vst1q_s32(b, vaccGHIJ); b += 4;
200       vst1q_s32(b, vaccKLMN); b += 4;
201     }
202     if XNN_UNLIKELY(c != 0) {
203       do {
204         const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
205         const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
206         const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
207         uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
208 
209         const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
210         vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
211         const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
212         vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
213         const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
214         vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
215         const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
216         vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
217         int32x4_t vacc0123 = vld1q_s32(b);
218         int32x4_t vacc4567 = vld1q_s32(b + 4);
219         vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
220 
221         vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567)));
222         vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567)));
223 
224         vst1q_s32(b, vacc0123); b += 4;
225         vst1q_s32(b, vacc4567); b += 4;
226 
227         c = doz(c, 8);
228       } while (c != 0);
229     }
230   }
231 
232   i0 = (const uint8_t*) ((uintptr_t) i0 + input_increment);
233   i1 = (const uint8_t*) ((uintptr_t) i1 + input_increment);
234   if XNN_UNPREDICTABLE(rows < 2) {
235     i1 = zero;
236   }
237   i2 = (const uint8_t*) ((uintptr_t) i2 + input_increment);
238   if XNN_UNPREDICTABLE(rows <= 2) {
239     i2 = zero;
240   }
241   i3 = (const uint8_t*) ((uintptr_t) i3 + input_increment);
242   if XNN_UNPREDICTABLE(rows < 4) {
243     i3 = zero;
244   }
245   i4 = (const uint8_t*) ((uintptr_t) i4 + input_increment);
246   if XNN_UNPREDICTABLE(rows <= 4) {
247     i4 = zero;
248   }
249   i5 = (const uint8_t*) ((uintptr_t) i5 + input_increment);
250   if XNN_UNPREDICTABLE(rows < 6) {
251     i5 = zero;
252   }
253   i6 = (const uint8_t*) ((uintptr_t) i6 + input_increment);
254   if XNN_UNPREDICTABLE(rows <= 6) {
255     i6 = zero;
256   }
257 
258   const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neonv8.scale);
259   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
260   const uint8x16_t voutput_min = vld1q_dup_u8(&params->fp32_neonv8.output_min);
261   const uint8x16_t voutput_max = vld1q_dup_u8(&params->fp32_neonv8.output_max);
262   for (; channels >= 24; channels -= 24) {
263     const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
264     const uint8x8_t vi0x89ABCDEF = vld1_u8(i0); i0 += 8;
265     const uint8x8_t vi0xGHIJKLMN = vld1_u8(i0); i0 += 8;
266     const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
267     const uint8x8_t vi1x89ABCDEF = vld1_u8(i1); i1 += 8;
268     const uint8x8_t vi1xGHIJKLMN = vld1_u8(i1); i1 += 8;
269 
270     const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
271     uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
272     const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
273     uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
274     const uint8x8_t vi2xGHIJKLMN = vld1_u8(i2); i2 += 8;
275     uint16x8_t vsumGHIJKLMN = vaddl_u8(vi0xGHIJKLMN, vi1xGHIJKLMN);
276 
277     const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
278     vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
279     const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
280     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
281     const uint8x8_t vi3xGHIJKLMN = vld1_u8(i3); i3 += 8;
282     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi2xGHIJKLMN);
283     const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
284     vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
285     const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
286     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
287     const uint8x8_t vi4xGHIJKLMN = vld1_u8(i4); i4 += 8;
288     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi3xGHIJKLMN);
289     const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
290     vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
291     const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
292     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
293     const uint8x8_t vi5xGHIJKLMN = vld1_u8(i5); i5 += 8;
294     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi4xGHIJKLMN);
295     const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
296     vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
297     const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
298     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
299     const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8;
300     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi5xGHIJKLMN);
301     int32x4_t vacc0123 = vld1q_s32(buffer); buffer += 4;
302     int32x4_t vacc4567 = vld1q_s32(buffer); buffer += 4;
303     vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
304     int32x4_t vacc89AB = vld1q_s32(buffer); buffer += 4;
305     int32x4_t vaccCDEF = vld1q_s32(buffer); buffer += 4;
306     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
307     int32x4_t vaccGHIJ = vld1q_s32(buffer); buffer += 4;
308     int32x4_t vaccKLMN = vld1q_s32(buffer); buffer += 4;
309     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN);
310 
311     vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567)));
312     vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567)));
313     vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vsum89ABCDEF)));
314     vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vsum89ABCDEF)));
315     vaccGHIJ = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccGHIJ), vget_low_u16(vsumGHIJKLMN)));
316     vaccKLMN = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccKLMN), vget_high_u16(vsumGHIJKLMN)));
317 
318     float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
319     float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
320     float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
321     float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
322     float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ);
323     float32x4_t vfpaccKLMN = vcvtq_f32_s32(vaccKLMN);
324 
325     vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
326     vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
327     vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
328     vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
329     vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale);
330     vfpaccKLMN = vmulq_f32(vfpaccKLMN, vscale);
331 
332     vacc0123 = vcvtnq_s32_f32(vfpacc0123);
333     vacc4567 = vcvtnq_s32_f32(vfpacc4567);
334     vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
335     vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
336     vaccGHIJ = vcvtnq_s32_f32(vfpaccGHIJ);
337     vaccKLMN = vcvtnq_s32_f32(vfpaccKLMN);
338 
339     #if XNN_ARCH_ARM64
340       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
341       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
342       int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);
343     #else  // !XNN_ARCH_ARM64
344       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
345       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
346       int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));
347     #endif  // !XNN_ARCH_ARM64
348 
349     vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
350     vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
351     vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
352 
353     #if XNN_ARCH_ARM64
354       uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
355       uint8x8_t voutGHIJKLMN = vqmovun_s16(vaccGHIJKLMN);
356     #else  // !XNN_ARCH_ARM64
357       uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
358       uint8x8_t voutGHIJKLMN = vqmovun_s16(vaccGHIJKLMN);
359     #endif  // !XNN_ARCH_ARM64
360 
361     vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
362     voutGHIJKLMN = vmax_u8(voutGHIJKLMN, vget_low_u8(voutput_min));
363 
364     vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
365     voutGHIJKLMN = vmin_u8(voutGHIJKLMN, vget_low_u8(voutput_max));
366 
367     vst1q_u8(output, vout0123456789ABCDEF); output += 16;
368     vst1_u8(output, voutGHIJKLMN); output += 8;
369   }
370   if XNN_UNLIKELY(channels != 0) {
371     do {
372       const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
373       const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
374       const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
375       uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
376 
377       const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
378       vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
379       const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
380       vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
381       const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
382       vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
383       const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
384       vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
385       int32x4_t vacc0123 = vld1q_s32(buffer); buffer += 4;
386       int32x4_t vacc4567 = vld1q_s32(buffer); buffer += 4;
387       vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
388 
389       vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567)));
390       vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567)));
391 
392       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
393       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
394 
395       vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
396       vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
397 
398       vacc0123 = vcvtnq_s32_f32(vfpacc0123);
399       vacc4567 = vcvtnq_s32_f32(vfpacc4567);
400 
401       #if XNN_ARCH_ARM64
402         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
403       #else
404         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
405       #endif
406       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
407 
408       uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
409       vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
410       vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
411 
412       if XNN_LIKELY(channels >= 8) {
413         vst1_u8(output, vout01234567); output += 8;
414         channels -= 8;
415       } else {
416         if (channels & 4) {
417           vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
418           vout01234567 = vext_u8(vout01234567, vout01234567, 4);
419         }
420         if (channels & 2) {
421           vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
422           vout01234567 = vext_u8(vout01234567, vout01234567, 2);
423         }
424         if (channels & 1) {
425           vst1_lane_u8(output, vout01234567, 0); output += 1;
426         }
427         channels = 0;
428       }
429     } while (channels != 0);
430   }
431 }
432