xref: /aosp_15_r20/external/XNNPACK/src/qu8-gavgpool/gen/7p7x-minmax-fp32-neon-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/math.h>
16 
17 
xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_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)])18 void xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24(
19     size_t rows,
20     size_t channels,
21     const uint8_t* input,
22     size_t input_stride,
23     const uint8_t* zero,
24     int32_t* buffer,
25     uint8_t* output,
26     const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27 {
28   assert(rows > 7);
29   assert(channels != 0);
30 
31   const uint8_t* i0 = input;
32   const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
33   const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
34   const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
35   const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
36   const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
37   const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
38   const size_t input_increment = 7 * input_stride - round_up_po2(channels, 8) * sizeof(uint8_t);
39 
40   const int32x4_t vinit_bias = vld1q_dup_s32(&params->fp32_neon.init_bias);
41   int32_t* b = buffer;
42   size_t c = channels;
43   for (; c >= 24; c -= 24) {
44     const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
45     const uint8x8_t vi0x89ABCDEF = vld1_u8(i0); i0 += 8;
46     const uint8x8_t vi0xGHIJKLMN = vld1_u8(i0); i0 += 8;
47     const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
48     const uint8x8_t vi1x89ABCDEF = vld1_u8(i1); i1 += 8;
49     const uint8x8_t vi1xGHIJKLMN = vld1_u8(i1); i1 += 8;
50 
51     const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
52     uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
53     const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
54     uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
55     const uint8x8_t vi2xGHIJKLMN = vld1_u8(i2); i2 += 8;
56     uint16x8_t vsumGHIJKLMN = vaddl_u8(vi0xGHIJKLMN, vi1xGHIJKLMN);
57 
58     const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
59     vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
60     const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
61     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
62     const uint8x8_t vi3xGHIJKLMN = vld1_u8(i3); i3 += 8;
63     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi2xGHIJKLMN);
64     const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
65     vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
66     const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
67     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
68     const uint8x8_t vi4xGHIJKLMN = vld1_u8(i4); i4 += 8;
69     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi3xGHIJKLMN);
70     const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
71     vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
72     const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
73     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
74     const uint8x8_t vi5xGHIJKLMN = vld1_u8(i5); i5 += 8;
75     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi4xGHIJKLMN);
76     const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
77     vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
78     const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
79     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
80     const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8;
81     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi5xGHIJKLMN);
82     vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
83     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
84     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN);
85 
86     const int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
87     const int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
88     const int32x4_t vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum89ABCDEF)));
89     const int32x4_t vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum89ABCDEF)));
90     const int32x4_t vaccGHIJ = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumGHIJKLMN)));
91     const int32x4_t vaccKLMN = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumGHIJKLMN)));
92 
93     vst1q_s32(b, vacc0123); b += 4;
94     vst1q_s32(b, vacc4567); b += 4;
95     vst1q_s32(b, vacc89AB); b += 4;
96     vst1q_s32(b, vaccCDEF); b += 4;
97     vst1q_s32(b, vaccGHIJ); b += 4;
98     vst1q_s32(b, vaccKLMN); b += 4;
99   }
100   if XNN_UNLIKELY(c != 0) {
101     do {
102       const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
103       const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
104       const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
105       uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
106 
107       const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
108       vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
109       const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
110       vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
111       const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
112       vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
113       const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
114       vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
115       vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
116 
117       const int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
118       const int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
119 
120       vst1q_s32(b, vacc0123); b += 4;
121       vst1q_s32(b, vacc4567); b += 4;
122 
123       c = doz(c, 8);
124     } while (c != 0);
125   }
126 
127   for (rows -= 7; rows > 7; rows -= 7) {
128     i0 = (const uint8_t*) ((uintptr_t) i0 + input_increment);
129     i1 = (const uint8_t*) ((uintptr_t) i1 + input_increment);
130     i2 = (const uint8_t*) ((uintptr_t) i2 + input_increment);
131     i3 = (const uint8_t*) ((uintptr_t) i3 + input_increment);
132     i4 = (const uint8_t*) ((uintptr_t) i4 + input_increment);
133     i5 = (const uint8_t*) ((uintptr_t) i5 + input_increment);
134     i6 = (const uint8_t*) ((uintptr_t) i6 + input_increment);
135 
136     int32_t* b = buffer;
137     size_t c = channels;
138     for (; c >= 24; c -= 24) {
139       const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
140       const uint8x8_t vi0x89ABCDEF = vld1_u8(i0); i0 += 8;
141       const uint8x8_t vi0xGHIJKLMN = vld1_u8(i0); i0 += 8;
142       const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
143       const uint8x8_t vi1x89ABCDEF = vld1_u8(i1); i1 += 8;
144       const uint8x8_t vi1xGHIJKLMN = vld1_u8(i1); i1 += 8;
145 
146       const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
147       uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
148       const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
149       uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
150       const uint8x8_t vi2xGHIJKLMN = vld1_u8(i2); i2 += 8;
151       uint16x8_t vsumGHIJKLMN = vaddl_u8(vi0xGHIJKLMN, vi1xGHIJKLMN);
152 
153       const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
154       vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
155       const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
156       vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
157       const uint8x8_t vi3xGHIJKLMN = vld1_u8(i3); i3 += 8;
158       vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi2xGHIJKLMN);
159       const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
160       vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
161       const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
162       vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
163       const uint8x8_t vi4xGHIJKLMN = vld1_u8(i4); i4 += 8;
164       vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi3xGHIJKLMN);
165       const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
166       vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
167       const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
168       vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
169       const uint8x8_t vi5xGHIJKLMN = vld1_u8(i5); i5 += 8;
170       vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi4xGHIJKLMN);
171       const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
172       vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
173       const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
174       vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
175       const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8;
176       vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi5xGHIJKLMN);
177       int32x4_t vacc0123 = vld1q_s32(b);
178       int32x4_t vacc4567 = vld1q_s32(b + 4);
179       vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
180       int32x4_t vacc89AB = vld1q_s32(b + 8);
181       int32x4_t vaccCDEF = vld1q_s32(b + 12);
182       vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
183       int32x4_t vaccGHIJ = vld1q_s32(b + 16);
184       int32x4_t vaccKLMN = vld1q_s32(b + 20);
185       vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN);
186 
187       vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567)));
188       vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567)));
189       vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vsum89ABCDEF)));
190       vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vsum89ABCDEF)));
191       vaccGHIJ = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccGHIJ), vget_low_u16(vsumGHIJKLMN)));
192       vaccKLMN = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccKLMN), vget_high_u16(vsumGHIJKLMN)));
193 
194       vst1q_s32(b, vacc0123); b += 4;
195       vst1q_s32(b, vacc4567); b += 4;
196       vst1q_s32(b, vacc89AB); b += 4;
197       vst1q_s32(b, vaccCDEF); b += 4;
198       vst1q_s32(b, vaccGHIJ); b += 4;
199       vst1q_s32(b, vaccKLMN); b += 4;
200     }
201     if XNN_UNLIKELY(c != 0) {
202       do {
203         const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
204         const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
205         const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
206         uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
207 
208         const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
209         vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
210         const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
211         vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
212         const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
213         vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
214         const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
215         vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
216         int32x4_t vacc0123 = vld1q_s32(b);
217         int32x4_t vacc4567 = vld1q_s32(b + 4);
218         vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
219 
220         vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567)));
221         vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567)));
222 
223         vst1q_s32(b, vacc0123); b += 4;
224         vst1q_s32(b, vacc4567); b += 4;
225 
226         c = doz(c, 8);
227       } while (c != 0);
228     }
229   }
230 
231   i0 = (const uint8_t*) ((uintptr_t) i0 + input_increment);
232   i1 = (const uint8_t*) ((uintptr_t) i1 + input_increment);
233   if XNN_UNPREDICTABLE(rows < 2) {
234     i1 = zero;
235   }
236   i2 = (const uint8_t*) ((uintptr_t) i2 + input_increment);
237   if XNN_UNPREDICTABLE(rows <= 2) {
238     i2 = zero;
239   }
240   i3 = (const uint8_t*) ((uintptr_t) i3 + input_increment);
241   if XNN_UNPREDICTABLE(rows < 4) {
242     i3 = zero;
243   }
244   i4 = (const uint8_t*) ((uintptr_t) i4 + input_increment);
245   if XNN_UNPREDICTABLE(rows <= 4) {
246     i4 = zero;
247   }
248   i5 = (const uint8_t*) ((uintptr_t) i5 + input_increment);
249   if XNN_UNPREDICTABLE(rows < 6) {
250     i5 = zero;
251   }
252   i6 = (const uint8_t*) ((uintptr_t) i6 + input_increment);
253   if XNN_UNPREDICTABLE(rows <= 6) {
254     i6 = zero;
255   }
256 
257   const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neon.scale);
258   const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
259   const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
260   const uint8x16_t voutput_min = vld1q_dup_u8(&params->fp32_neon.output_min);
261   const uint8x16_t voutput_max = vld1q_dup_u8(&params->fp32_neon.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 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
333     vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
334     vacc89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc89AB, vmagic_bias));
335     vaccCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpaccCDEF, vmagic_bias));
336     vaccGHIJ = vreinterpretq_s32_f32(vaddq_f32(vfpaccGHIJ, vmagic_bias));
337     vaccKLMN = vreinterpretq_s32_f32(vaddq_f32(vfpaccKLMN, vmagic_bias));
338 
339     vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
340     vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
341     vacc89AB = vqsubq_s32(vacc89AB, vmagic_bias_less_output_zero_point);
342     vaccCDEF = vqsubq_s32(vaccCDEF, vmagic_bias_less_output_zero_point);
343     vaccGHIJ = vqsubq_s32(vaccGHIJ, vmagic_bias_less_output_zero_point);
344     vaccKLMN = vqsubq_s32(vaccKLMN, vmagic_bias_less_output_zero_point);
345 
346     #if XNN_ARCH_ARM64
347       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
348       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
349       int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);
350     #else  // !XNN_ARCH_ARM64
351       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
352       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
353       int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));
354     #endif  // !XNN_ARCH_ARM64
355 
356 
357     #if XNN_ARCH_ARM64
358       uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
359       uint8x8_t voutGHIJKLMN = vqmovun_s16(vaccGHIJKLMN);
360     #else  // !XNN_ARCH_ARM64
361       uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
362       uint8x8_t voutGHIJKLMN = vqmovun_s16(vaccGHIJKLMN);
363     #endif  // !XNN_ARCH_ARM64
364 
365     vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
366     voutGHIJKLMN = vmax_u8(voutGHIJKLMN, vget_low_u8(voutput_min));
367 
368     vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
369     voutGHIJKLMN = vmin_u8(voutGHIJKLMN, vget_low_u8(voutput_max));
370 
371     vst1q_u8(output, vout0123456789ABCDEF); output += 16;
372     vst1_u8(output, voutGHIJKLMN); output += 8;
373   }
374   if XNN_UNLIKELY(channels != 0) {
375     do {
376       const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
377       const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
378       const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
379       uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
380 
381       const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
382       vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
383       const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
384       vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
385       const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
386       vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
387       const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
388       vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
389       int32x4_t vacc0123 = vld1q_s32(buffer); buffer += 4;
390       int32x4_t vacc4567 = vld1q_s32(buffer); buffer += 4;
391       vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
392 
393       vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567)));
394       vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567)));
395 
396       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
397       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
398 
399       vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
400       vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
401 
402       vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
403       vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
404 
405       vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
406       vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
407 
408       #if XNN_ARCH_ARM64
409         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
410       #else
411         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
412       #endif
413 
414       uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
415       vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
416       vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
417 
418       if XNN_LIKELY(channels >= 8) {
419         vst1_u8(output, vout01234567); output += 8;
420         channels -= 8;
421       } else {
422         if (channels & 4) {
423           vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
424           vout01234567 = vext_u8(vout01234567, vout01234567, 4);
425         }
426         if (channels & 2) {
427           vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
428           vout01234567 = vext_u8(vout01234567, vout01234567, 2);
429         }
430         if (channels & 1) {
431           vst1_lane_u8(output, vout01234567, 0); output += 1;
432         }
433         channels = 0;
434       }
435     } while (channels != 0);
436   }
437 }
438