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_c32(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_c32(
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(¶ms->fp32_neon.init_bias);
41 int32_t* b = buffer;
42 size_t c = channels;
43 for (; c >= 32; c -= 32) {
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 vi0xOPQRSTUV = 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 const uint8x8_t vi1xOPQRSTUV = vld1_u8(i1); i1 += 8;
52
53 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
54 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
55 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
56 uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
57 const uint8x8_t vi2xGHIJKLMN = vld1_u8(i2); i2 += 8;
58 uint16x8_t vsumGHIJKLMN = vaddl_u8(vi0xGHIJKLMN, vi1xGHIJKLMN);
59 const uint8x8_t vi2xOPQRSTUV = vld1_u8(i2); i2 += 8;
60 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV);
61
62 const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
63 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
64 const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
65 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
66 const uint8x8_t vi3xGHIJKLMN = vld1_u8(i3); i3 += 8;
67 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi2xGHIJKLMN);
68 const uint8x8_t vi3xOPQRSTUV = vld1_u8(i3); i3 += 8;
69 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV);
70 const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
71 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
72 const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
73 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
74 const uint8x8_t vi4xGHIJKLMN = vld1_u8(i4); i4 += 8;
75 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi3xGHIJKLMN);
76 const uint8x8_t vi4xOPQRSTUV = vld1_u8(i4); i4 += 8;
77 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi3xOPQRSTUV);
78 const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
79 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
80 const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
81 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
82 const uint8x8_t vi5xGHIJKLMN = vld1_u8(i5); i5 += 8;
83 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi4xGHIJKLMN);
84 const uint8x8_t vi5xOPQRSTUV = vld1_u8(i5); i5 += 8;
85 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi4xOPQRSTUV);
86 const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
87 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
88 const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
89 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
90 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8;
91 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi5xGHIJKLMN);
92 const uint8x8_t vi6xOPQRSTUV = vld1_u8(i6); i6 += 8;
93 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi5xOPQRSTUV);
94 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
95 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
96 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN);
97 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi6xOPQRSTUV);
98
99 const int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
100 const int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
101 const int32x4_t vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum89ABCDEF)));
102 const int32x4_t vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum89ABCDEF)));
103 const int32x4_t vaccGHIJ = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumGHIJKLMN)));
104 const int32x4_t vaccKLMN = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumGHIJKLMN)));
105 const int32x4_t vaccOPQR = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumOPQRSTUV)));
106 const int32x4_t vaccSTUV = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumOPQRSTUV)));
107
108 vst1q_s32(b, vacc0123); b += 4;
109 vst1q_s32(b, vacc4567); b += 4;
110 vst1q_s32(b, vacc89AB); b += 4;
111 vst1q_s32(b, vaccCDEF); b += 4;
112 vst1q_s32(b, vaccGHIJ); b += 4;
113 vst1q_s32(b, vaccKLMN); b += 4;
114 vst1q_s32(b, vaccOPQR); b += 4;
115 vst1q_s32(b, vaccSTUV); b += 4;
116 }
117 if XNN_UNLIKELY(c != 0) {
118 do {
119 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
120 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
121 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
122 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
123
124 const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
125 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
126 const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
127 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
128 const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
129 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
130 const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
131 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
132 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
133
134 const int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
135 const int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
136
137 vst1q_s32(b, vacc0123); b += 4;
138 vst1q_s32(b, vacc4567); b += 4;
139
140 c = doz(c, 8);
141 } while (c != 0);
142 }
143
144 for (rows -= 7; rows > 7; rows -= 7) {
145 i0 = (const uint8_t*) ((uintptr_t) i0 + input_increment);
146 i1 = (const uint8_t*) ((uintptr_t) i1 + input_increment);
147 i2 = (const uint8_t*) ((uintptr_t) i2 + input_increment);
148 i3 = (const uint8_t*) ((uintptr_t) i3 + input_increment);
149 i4 = (const uint8_t*) ((uintptr_t) i4 + input_increment);
150 i5 = (const uint8_t*) ((uintptr_t) i5 + input_increment);
151 i6 = (const uint8_t*) ((uintptr_t) i6 + input_increment);
152
153 int32_t* b = buffer;
154 size_t c = channels;
155 for (; c >= 32; c -= 32) {
156 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
157 const uint8x8_t vi0x89ABCDEF = vld1_u8(i0); i0 += 8;
158 const uint8x8_t vi0xGHIJKLMN = vld1_u8(i0); i0 += 8;
159 const uint8x8_t vi0xOPQRSTUV = vld1_u8(i0); i0 += 8;
160 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
161 const uint8x8_t vi1x89ABCDEF = vld1_u8(i1); i1 += 8;
162 const uint8x8_t vi1xGHIJKLMN = vld1_u8(i1); i1 += 8;
163 const uint8x8_t vi1xOPQRSTUV = vld1_u8(i1); i1 += 8;
164
165 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
166 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
167 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
168 uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
169 const uint8x8_t vi2xGHIJKLMN = vld1_u8(i2); i2 += 8;
170 uint16x8_t vsumGHIJKLMN = vaddl_u8(vi0xGHIJKLMN, vi1xGHIJKLMN);
171 const uint8x8_t vi2xOPQRSTUV = vld1_u8(i2); i2 += 8;
172 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV);
173
174 const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
175 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
176 const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
177 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
178 const uint8x8_t vi3xGHIJKLMN = vld1_u8(i3); i3 += 8;
179 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi2xGHIJKLMN);
180 const uint8x8_t vi3xOPQRSTUV = vld1_u8(i3); i3 += 8;
181 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV);
182 const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
183 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
184 const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
185 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
186 const uint8x8_t vi4xGHIJKLMN = vld1_u8(i4); i4 += 8;
187 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi3xGHIJKLMN);
188 const uint8x8_t vi4xOPQRSTUV = vld1_u8(i4); i4 += 8;
189 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi3xOPQRSTUV);
190 const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
191 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
192 const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
193 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
194 const uint8x8_t vi5xGHIJKLMN = vld1_u8(i5); i5 += 8;
195 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi4xGHIJKLMN);
196 const uint8x8_t vi5xOPQRSTUV = vld1_u8(i5); i5 += 8;
197 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi4xOPQRSTUV);
198 const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
199 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
200 const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
201 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
202 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8;
203 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi5xGHIJKLMN);
204 const uint8x8_t vi6xOPQRSTUV = vld1_u8(i6); i6 += 8;
205 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi5xOPQRSTUV);
206 int32x4_t vacc0123 = vld1q_s32(b);
207 int32x4_t vacc4567 = vld1q_s32(b + 4);
208 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
209 int32x4_t vacc89AB = vld1q_s32(b + 8);
210 int32x4_t vaccCDEF = vld1q_s32(b + 12);
211 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
212 int32x4_t vaccGHIJ = vld1q_s32(b + 16);
213 int32x4_t vaccKLMN = vld1q_s32(b + 20);
214 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN);
215 int32x4_t vaccOPQR = vld1q_s32(b + 24);
216 int32x4_t vaccSTUV = vld1q_s32(b + 28);
217 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi6xOPQRSTUV);
218
219 vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567)));
220 vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567)));
221 vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vsum89ABCDEF)));
222 vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vsum89ABCDEF)));
223 vaccGHIJ = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccGHIJ), vget_low_u16(vsumGHIJKLMN)));
224 vaccKLMN = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccKLMN), vget_high_u16(vsumGHIJKLMN)));
225 vaccOPQR = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccOPQR), vget_low_u16(vsumOPQRSTUV)));
226 vaccSTUV = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccSTUV), vget_high_u16(vsumOPQRSTUV)));
227
228 vst1q_s32(b, vacc0123); b += 4;
229 vst1q_s32(b, vacc4567); b += 4;
230 vst1q_s32(b, vacc89AB); b += 4;
231 vst1q_s32(b, vaccCDEF); b += 4;
232 vst1q_s32(b, vaccGHIJ); b += 4;
233 vst1q_s32(b, vaccKLMN); b += 4;
234 vst1q_s32(b, vaccOPQR); b += 4;
235 vst1q_s32(b, vaccSTUV); b += 4;
236 }
237 if XNN_UNLIKELY(c != 0) {
238 do {
239 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
240 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
241 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
242 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
243
244 const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
245 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
246 const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
247 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
248 const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
249 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
250 const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
251 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
252 int32x4_t vacc0123 = vld1q_s32(b);
253 int32x4_t vacc4567 = vld1q_s32(b + 4);
254 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
255
256 vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567)));
257 vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567)));
258
259 vst1q_s32(b, vacc0123); b += 4;
260 vst1q_s32(b, vacc4567); b += 4;
261
262 c = doz(c, 8);
263 } while (c != 0);
264 }
265 }
266
267 i0 = (const uint8_t*) ((uintptr_t) i0 + input_increment);
268 i1 = (const uint8_t*) ((uintptr_t) i1 + input_increment);
269 if XNN_UNPREDICTABLE(rows < 2) {
270 i1 = zero;
271 }
272 i2 = (const uint8_t*) ((uintptr_t) i2 + input_increment);
273 if XNN_UNPREDICTABLE(rows <= 2) {
274 i2 = zero;
275 }
276 i3 = (const uint8_t*) ((uintptr_t) i3 + input_increment);
277 if XNN_UNPREDICTABLE(rows < 4) {
278 i3 = zero;
279 }
280 i4 = (const uint8_t*) ((uintptr_t) i4 + input_increment);
281 if XNN_UNPREDICTABLE(rows <= 4) {
282 i4 = zero;
283 }
284 i5 = (const uint8_t*) ((uintptr_t) i5 + input_increment);
285 if XNN_UNPREDICTABLE(rows < 6) {
286 i5 = zero;
287 }
288 i6 = (const uint8_t*) ((uintptr_t) i6 + input_increment);
289 if XNN_UNPREDICTABLE(rows <= 6) {
290 i6 = zero;
291 }
292
293 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neon.scale);
294 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
295 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
296 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->fp32_neon.output_min);
297 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->fp32_neon.output_max);
298 for (; channels >= 32; channels -= 32) {
299 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
300 const uint8x8_t vi0x89ABCDEF = vld1_u8(i0); i0 += 8;
301 const uint8x8_t vi0xGHIJKLMN = vld1_u8(i0); i0 += 8;
302 const uint8x8_t vi0xOPQRSTUV = vld1_u8(i0); i0 += 8;
303 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
304 const uint8x8_t vi1x89ABCDEF = vld1_u8(i1); i1 += 8;
305 const uint8x8_t vi1xGHIJKLMN = vld1_u8(i1); i1 += 8;
306 const uint8x8_t vi1xOPQRSTUV = vld1_u8(i1); i1 += 8;
307
308 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
309 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
310 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
311 uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
312 const uint8x8_t vi2xGHIJKLMN = vld1_u8(i2); i2 += 8;
313 uint16x8_t vsumGHIJKLMN = vaddl_u8(vi0xGHIJKLMN, vi1xGHIJKLMN);
314 const uint8x8_t vi2xOPQRSTUV = vld1_u8(i2); i2 += 8;
315 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV);
316
317 const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
318 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
319 const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
320 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
321 const uint8x8_t vi3xGHIJKLMN = vld1_u8(i3); i3 += 8;
322 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi2xGHIJKLMN);
323 const uint8x8_t vi3xOPQRSTUV = vld1_u8(i3); i3 += 8;
324 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV);
325 const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
326 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
327 const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
328 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
329 const uint8x8_t vi4xGHIJKLMN = vld1_u8(i4); i4 += 8;
330 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi3xGHIJKLMN);
331 const uint8x8_t vi4xOPQRSTUV = vld1_u8(i4); i4 += 8;
332 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi3xOPQRSTUV);
333 const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
334 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
335 const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
336 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
337 const uint8x8_t vi5xGHIJKLMN = vld1_u8(i5); i5 += 8;
338 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi4xGHIJKLMN);
339 const uint8x8_t vi5xOPQRSTUV = vld1_u8(i5); i5 += 8;
340 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi4xOPQRSTUV);
341 const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
342 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
343 const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
344 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
345 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8;
346 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi5xGHIJKLMN);
347 const uint8x8_t vi6xOPQRSTUV = vld1_u8(i6); i6 += 8;
348 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi5xOPQRSTUV);
349 int32x4_t vacc0123 = vld1q_s32(buffer); buffer += 4;
350 int32x4_t vacc4567 = vld1q_s32(buffer); buffer += 4;
351 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
352 int32x4_t vacc89AB = vld1q_s32(buffer); buffer += 4;
353 int32x4_t vaccCDEF = vld1q_s32(buffer); buffer += 4;
354 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
355 int32x4_t vaccGHIJ = vld1q_s32(buffer); buffer += 4;
356 int32x4_t vaccKLMN = vld1q_s32(buffer); buffer += 4;
357 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN);
358 int32x4_t vaccOPQR = vld1q_s32(buffer); buffer += 4;
359 int32x4_t vaccSTUV = vld1q_s32(buffer); buffer += 4;
360 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi6xOPQRSTUV);
361
362 vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567)));
363 vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567)));
364 vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vsum89ABCDEF)));
365 vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vsum89ABCDEF)));
366 vaccGHIJ = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccGHIJ), vget_low_u16(vsumGHIJKLMN)));
367 vaccKLMN = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccKLMN), vget_high_u16(vsumGHIJKLMN)));
368 vaccOPQR = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccOPQR), vget_low_u16(vsumOPQRSTUV)));
369 vaccSTUV = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccSTUV), vget_high_u16(vsumOPQRSTUV)));
370
371 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
372 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
373 float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
374 float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
375 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ);
376 float32x4_t vfpaccKLMN = vcvtq_f32_s32(vaccKLMN);
377 float32x4_t vfpaccOPQR = vcvtq_f32_s32(vaccOPQR);
378 float32x4_t vfpaccSTUV = vcvtq_f32_s32(vaccSTUV);
379
380 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
381 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
382 vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
383 vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
384 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale);
385 vfpaccKLMN = vmulq_f32(vfpaccKLMN, vscale);
386 vfpaccOPQR = vmulq_f32(vfpaccOPQR, vscale);
387 vfpaccSTUV = vmulq_f32(vfpaccSTUV, vscale);
388
389 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
390 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
391 vacc89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc89AB, vmagic_bias));
392 vaccCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpaccCDEF, vmagic_bias));
393 vaccGHIJ = vreinterpretq_s32_f32(vaddq_f32(vfpaccGHIJ, vmagic_bias));
394 vaccKLMN = vreinterpretq_s32_f32(vaddq_f32(vfpaccKLMN, vmagic_bias));
395 vaccOPQR = vreinterpretq_s32_f32(vaddq_f32(vfpaccOPQR, vmagic_bias));
396 vaccSTUV = vreinterpretq_s32_f32(vaddq_f32(vfpaccSTUV, vmagic_bias));
397
398 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
399 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
400 vacc89AB = vqsubq_s32(vacc89AB, vmagic_bias_less_output_zero_point);
401 vaccCDEF = vqsubq_s32(vaccCDEF, vmagic_bias_less_output_zero_point);
402 vaccGHIJ = vqsubq_s32(vaccGHIJ, vmagic_bias_less_output_zero_point);
403 vaccKLMN = vqsubq_s32(vaccKLMN, vmagic_bias_less_output_zero_point);
404 vaccOPQR = vqsubq_s32(vaccOPQR, vmagic_bias_less_output_zero_point);
405 vaccSTUV = vqsubq_s32(vaccSTUV, vmagic_bias_less_output_zero_point);
406
407 #if XNN_ARCH_ARM64
408 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
409 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
410 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);
411 int16x8_t vaccOPQRSTUV = vqmovn_high_s32(vqmovn_s32(vaccOPQR), vaccSTUV);
412 #else // !XNN_ARCH_ARM64
413 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
414 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
415 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));
416 int16x8_t vaccOPQRSTUV = vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV));
417 #endif // !XNN_ARCH_ARM64
418
419
420 #if XNN_ARCH_ARM64
421 uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
422 uint8x16_t voutGHIJKLMNOPQRSTUV = vqmovun_high_s16(vqmovun_s16(vaccGHIJKLMN), vaccOPQRSTUV);
423 #else // !XNN_ARCH_ARM64
424 uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
425 uint8x16_t voutGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vaccGHIJKLMN), vqmovun_s16(vaccOPQRSTUV));
426 #endif // !XNN_ARCH_ARM64
427
428 vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
429 voutGHIJKLMNOPQRSTUV = vmaxq_u8(voutGHIJKLMNOPQRSTUV, voutput_min);
430
431 vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
432 voutGHIJKLMNOPQRSTUV = vminq_u8(voutGHIJKLMNOPQRSTUV, voutput_max);
433
434 vst1q_u8(output, vout0123456789ABCDEF); output += 16;
435 vst1q_u8(output, voutGHIJKLMNOPQRSTUV); output += 16;
436 }
437 if XNN_UNLIKELY(channels != 0) {
438 do {
439 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
440 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
441 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
442 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
443
444 const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
445 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
446 const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
447 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
448 const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
449 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
450 const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
451 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
452 int32x4_t vacc0123 = vld1q_s32(buffer); buffer += 4;
453 int32x4_t vacc4567 = vld1q_s32(buffer); buffer += 4;
454 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
455
456 vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567)));
457 vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567)));
458
459 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
460 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
461
462 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
463 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
464
465 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
466 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
467
468 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
469 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
470
471 #if XNN_ARCH_ARM64
472 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
473 #else
474 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
475 #endif
476
477 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
478 vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
479 vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
480
481 if XNN_LIKELY(channels >= 8) {
482 vst1_u8(output, vout01234567); output += 8;
483 channels -= 8;
484 } else {
485 if (channels & 4) {
486 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
487 vout01234567 = vext_u8(vout01234567, vout01234567, 4);
488 }
489 if (channels & 2) {
490 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
491 vout01234567 = vext_u8(vout01234567, vout01234567, 2);
492 }
493 if (channels & 1) {
494 vst1_lane_u8(output, vout01234567, 0); output += 1;
495 }
496 channels = 0;
497 }
498 } while (channels != 0);
499 }
500 }
501