1 // Auto-generated file. Do not edit!
2 // Template: src/qu8-dwconv/unipass-neon-mul8.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/dwconv.h>
15
16
xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x9__neon_mul8(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x9__neon_mul8(
18 size_t channels,
19 size_t output_width,
20 const uint8_t** input,
21 const void* weights,
22 uint8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const uint8_t* zero,
27 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->rndnu_neon.kernel_zero_point);
33 const uint16x8_t vkernel_zero_point16 = vmovl_u8(vkernel_zero_point);
34 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
35 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
36 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
37 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
38 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
39 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
40 do {
41 const uint8_t* i0 = input[0];
42 assert(i0 != NULL);
43 if XNN_UNPREDICTABLE(i0 != zero) {
44 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
45 }
46 const uint8_t* i1 = input[1];
47 assert(i1 != NULL);
48 if XNN_UNPREDICTABLE(i1 != zero) {
49 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
50 }
51 const uint8_t* i2 = input[2];
52 assert(i2 != NULL);
53 if XNN_UNPREDICTABLE(i2 != zero) {
54 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
55 }
56 const uint8_t* i3 = input[3];
57 assert(i3 != NULL);
58 if XNN_UNPREDICTABLE(i3 != zero) {
59 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
60 }
61 const uint8_t* i4 = input[4];
62 assert(i4 != NULL);
63 if XNN_UNPREDICTABLE(i4 != zero) {
64 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
65 }
66 const uint8_t* i5 = input[5];
67 assert(i5 != NULL);
68 if XNN_UNPREDICTABLE(i5 != zero) {
69 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
70 }
71 const uint8_t* i6 = input[6];
72 assert(i6 != NULL);
73 if XNN_UNPREDICTABLE(i6 != zero) {
74 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
75 }
76 const uint8_t* i7 = input[7];
77 assert(i7 != NULL);
78 if XNN_UNPREDICTABLE(i7 != zero) {
79 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
80 }
81 const uint8_t* i8 = input[8];
82 assert(i8 != NULL);
83 if XNN_UNPREDICTABLE(i8 != zero) {
84 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
85 }
86 input = (const uint8_t**) ((uintptr_t) input + input_stride);
87
88
89 size_t c = channels;
90 const void* w = weights;
91 for (; c >= 16; c -= 16) {
92 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
93 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
94 int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
95 int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
96
97
98 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
99 const uint8x8_t vk0x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
100 const uint8x8_t vi0x89ABCDEF = vld1_u8(i0); i0 += 8;
101 const uint8x8_t vk0x89ABCDEF = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
102
103 uint16x8_t vprod01234567 = vmull_u8(vi0x01234567, vk0x01234567);
104 uint16x8_t vprod89ABCDEF = vmull_u8(vi0x89ABCDEF, vk0x89ABCDEF);
105
106 vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
107 vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
108 vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vprod89ABCDEF)));
109 vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vprod89ABCDEF)));
110 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
111 const uint8x8_t vk1x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
112 const uint8x8_t vi1x89ABCDEF = vld1_u8(i1); i1 += 8;
113 const uint8x8_t vk1x89ABCDEF = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
114
115 vprod01234567 = vmull_u8(vi1x01234567, vk1x01234567);
116 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
117 vprod89ABCDEF = vmull_u8(vi1x89ABCDEF, vk1x89ABCDEF);
118 uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
119
120 vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
121 vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
122 vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vprod89ABCDEF)));
123 vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vprod89ABCDEF)));
124 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
125 const uint8x8_t vk2x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
126 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
127 const uint8x8_t vk2x89ABCDEF = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
128
129 vprod01234567 = vmull_u8(vi2x01234567, vk2x01234567);
130 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
131 vprod89ABCDEF = vmull_u8(vi2x89ABCDEF, vk2x89ABCDEF);
132 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
133
134 vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
135 vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
136 vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vprod89ABCDEF)));
137 vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vprod89ABCDEF)));
138 const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
139 const uint8x8_t vk3x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
140 const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
141 const uint8x8_t vk3x89ABCDEF = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
142
143 vprod01234567 = vmull_u8(vi3x01234567, vk3x01234567);
144 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
145 vprod89ABCDEF = vmull_u8(vi3x89ABCDEF, vk3x89ABCDEF);
146 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
147
148 vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
149 vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
150 vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vprod89ABCDEF)));
151 vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vprod89ABCDEF)));
152 const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
153 const uint8x8_t vk4x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
154 const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
155 const uint8x8_t vk4x89ABCDEF = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
156
157 vprod01234567 = vmull_u8(vi4x01234567, vk4x01234567);
158 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
159 vprod89ABCDEF = vmull_u8(vi4x89ABCDEF, vk4x89ABCDEF);
160 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
161
162 vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
163 vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
164 vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vprod89ABCDEF)));
165 vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vprod89ABCDEF)));
166 const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
167 const uint8x8_t vk5x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
168 const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
169 const uint8x8_t vk5x89ABCDEF = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
170
171 vprod01234567 = vmull_u8(vi5x01234567, vk5x01234567);
172 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
173 vprod89ABCDEF = vmull_u8(vi5x89ABCDEF, vk5x89ABCDEF);
174 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
175
176 vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
177 vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
178 vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vprod89ABCDEF)));
179 vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vprod89ABCDEF)));
180 const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
181 const uint8x8_t vk6x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
182 const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
183 const uint8x8_t vk6x89ABCDEF = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
184
185 vprod01234567 = vmull_u8(vi6x01234567, vk6x01234567);
186 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
187 vprod89ABCDEF = vmull_u8(vi6x89ABCDEF, vk6x89ABCDEF);
188 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
189
190 vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
191 vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
192 vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vprod89ABCDEF)));
193 vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vprod89ABCDEF)));
194 const uint8x8_t vi7x01234567 = vld1_u8(i7); i7 += 8;
195 const uint8x8_t vk7x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
196 const uint8x8_t vi7x89ABCDEF = vld1_u8(i7); i7 += 8;
197 const uint8x8_t vk7x89ABCDEF = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
198
199 vprod01234567 = vmull_u8(vi7x01234567, vk7x01234567);
200 vsum01234567 = vaddw_u8(vsum01234567, vi7x01234567);
201 vprod89ABCDEF = vmull_u8(vi7x89ABCDEF, vk7x89ABCDEF);
202 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi7x89ABCDEF);
203
204 vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
205 vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
206 vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vprod89ABCDEF)));
207 vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vprod89ABCDEF)));
208 const uint8x8_t vi8x01234567 = vld1_u8(i8); i8 += 8;
209 const uint8x8_t vk8x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
210 const uint8x8_t vi8x89ABCDEF = vld1_u8(i8); i8 += 8;
211 const uint8x8_t vk8x89ABCDEF = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
212
213 vprod01234567 = vmull_u8(vi8x01234567, vk8x01234567);
214 vsum01234567 = vaddw_u8(vsum01234567, vi8x01234567);
215 vprod89ABCDEF = vmull_u8(vi8x89ABCDEF, vk8x89ABCDEF);
216 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi8x89ABCDEF);
217
218 vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
219 vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
220 vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vprod89ABCDEF)));
221 vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vprod89ABCDEF)));
222
223 vacc0123 = vreinterpretq_s32_u32(vmlsl_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567), vget_low_u16(vkernel_zero_point16)));
224 vacc4567 = vreinterpretq_s32_u32(vmlsl_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567), vget_high_u16(vkernel_zero_point16)));
225 vacc89AB = vreinterpretq_s32_u32(vmlsl_u16(vreinterpretq_u32_s32(vacc89AB), vget_low_u16(vsum89ABCDEF), vget_low_u16(vkernel_zero_point16)));
226 vaccCDEF = vreinterpretq_s32_u32(vmlsl_u16(vreinterpretq_u32_s32(vaccCDEF), vget_high_u16(vsum89ABCDEF), vget_high_u16(vkernel_zero_point16)));
227
228 vacc0123 = vshlq_s32(vacc0123, vright_pre_shift);
229 vacc4567 = vshlq_s32(vacc4567, vright_pre_shift);
230 vacc89AB = vshlq_s32(vacc89AB, vright_pre_shift);
231 vaccCDEF = vshlq_s32(vaccCDEF, vright_pre_shift);
232
233 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
234 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
235 vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
236 vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
237
238 vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
239 vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
240 vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
241 vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
242
243 #if XNN_ARCH_ARM64
244 const int16x8_t vacc01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567), voutput_zero_point);
245 const int16x8_t vacc89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF), voutput_zero_point);
246
247 uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
248 #else
249 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
250 const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF)), voutput_zero_point);
251
252 uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
253 #endif
254
255 vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
256
257 vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
258
259 vst1q_u8(output, vout0123456789ABCDEF); output += 16;
260 }
261 if XNN_UNLIKELY(c != 0) {
262 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 16);
263 do {
264 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
265 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
266
267 const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
268 const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(k), vkernel_zero_point)); k += 8;
269
270 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
271 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
272 const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
273 const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 8)), vkernel_zero_point));
274
275 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
276 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
277 const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
278 const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 24)), vkernel_zero_point));
279
280 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
281 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
282 const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
283 const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 40)), vkernel_zero_point));
284
285 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
286 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
287 const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
288 const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 56)), vkernel_zero_point));
289
290 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
291 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
292 const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
293 const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 72)), vkernel_zero_point));
294
295 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
296 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
297 const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
298 const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 88)), vkernel_zero_point));
299
300 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
301 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
302 const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
303 const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 104)), vkernel_zero_point));
304
305 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
306 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
307 const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
308 const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 120)), vkernel_zero_point));
309
310 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
311 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
312
313 vacc0123 = vrshlq_s32(vacc0123, vright_pre_shift);
314 vacc4567 = vrshlq_s32(vacc4567, vright_pre_shift);
315
316 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
317 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
318
319 vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
320 vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
321
322 #if XNN_ARCH_ARM64
323 const int16x8_t vacc01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567), voutput_zero_point);
324 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
325 #else
326 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
327 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
328 #endif
329
330 vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
331 vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
332
333 if XNN_LIKELY(c >= 8) {
334 vst1_u8(output, vout01234567); output += 8;
335 c -= 8;
336 } else {
337 if (c & 4) {
338 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
339 vout01234567 = vext_u8(vout01234567, vout01234567, 4);
340 }
341 if (c & 2) {
342 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
343 vout01234567 = vext_u8(vout01234567, vout01234567, 2);
344 }
345 if (c & 1) {
346 vst1_lane_u8(output, vout01234567, 0); output += 1;
347 }
348 c = 0;
349 }
350 } while (c != 0);
351 }
352
353 output = (uint8_t*) ((uintptr_t) output + output_increment);
354 } while (--output_width != 0);
355 }
356