xref: /aosp_15_r20/external/XNNPACK/src/qu8-dwconv/gen/up16x25-minmax-rndnu-neon-mul16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-neon-mul16.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_up16x25__neon_mul16(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_up16x25__neon_mul16(
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 int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
34   const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
35   const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
36   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
37   const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
38   const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
39   do {
40     const uint8_t* i0 = input[0];
41     assert(i0 != NULL);
42     if XNN_UNPREDICTABLE(i0 != zero) {
43       i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
44     }
45     const uint8_t* i1 = input[1];
46     assert(i1 != NULL);
47     if XNN_UNPREDICTABLE(i1 != zero) {
48       i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
49     }
50     const uint8_t* i2 = input[2];
51     assert(i2 != NULL);
52     if XNN_UNPREDICTABLE(i2 != zero) {
53       i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
54     }
55     const uint8_t* i3 = input[3];
56     assert(i3 != NULL);
57     if XNN_UNPREDICTABLE(i3 != zero) {
58       i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
59     }
60     const uint8_t* i4 = input[4];
61     assert(i4 != NULL);
62     if XNN_UNPREDICTABLE(i4 != zero) {
63       i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
64     }
65     const uint8_t* i5 = input[5];
66     assert(i5 != NULL);
67     if XNN_UNPREDICTABLE(i5 != zero) {
68       i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
69     }
70     const uint8_t* i6 = input[6];
71     assert(i6 != NULL);
72     if XNN_UNPREDICTABLE(i6 != zero) {
73       i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
74     }
75     const uint8_t* i7 = input[7];
76     assert(i7 != NULL);
77     if XNN_UNPREDICTABLE(i7 != zero) {
78       i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
79     }
80     const uint8_t* i8 = input[8];
81     assert(i8 != NULL);
82     if XNN_UNPREDICTABLE(i8 != zero) {
83       i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
84     }
85     const uint8_t* i9 = input[9];
86     assert(i9 != NULL);
87     if XNN_UNPREDICTABLE(i9 != zero) {
88       i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
89     }
90     const uint8_t* i10 = input[10];
91     assert(i10 != NULL);
92     if XNN_UNPREDICTABLE(i10 != zero) {
93       i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
94     }
95     const uint8_t* i11 = input[11];
96     assert(i11 != NULL);
97     if XNN_UNPREDICTABLE(i11 != zero) {
98       i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
99     }
100     const uint8_t* i12 = input[12];
101     assert(i12 != NULL);
102     if XNN_UNPREDICTABLE(i12 != zero) {
103       i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
104     }
105     const uint8_t* i13 = input[13];
106     assert(i13 != NULL);
107     if XNN_UNPREDICTABLE(i13 != zero) {
108       i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
109     }
110     const uint8_t* i14 = input[14];
111     assert(i14 != NULL);
112     if XNN_UNPREDICTABLE(i14 != zero) {
113       i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
114     }
115     const uint8_t* i15 = input[15];
116     assert(i15 != NULL);
117     if XNN_UNPREDICTABLE(i15 != zero) {
118       i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
119     }
120     const uint8_t* i16 = input[16];
121     assert(i16 != NULL);
122     if XNN_UNPREDICTABLE(i16 != zero) {
123       i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
124     }
125     const uint8_t* i17 = input[17];
126     assert(i17 != NULL);
127     if XNN_UNPREDICTABLE(i17 != zero) {
128       i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
129     }
130     const uint8_t* i18 = input[18];
131     assert(i18 != NULL);
132     if XNN_UNPREDICTABLE(i18 != zero) {
133       i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
134     }
135     const uint8_t* i19 = input[19];
136     assert(i19 != NULL);
137     if XNN_UNPREDICTABLE(i19 != zero) {
138       i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
139     }
140     const uint8_t* i20 = input[20];
141     assert(i20 != NULL);
142     if XNN_UNPREDICTABLE(i20 != zero) {
143       i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
144     }
145     const uint8_t* i21 = input[21];
146     assert(i21 != NULL);
147     if XNN_UNPREDICTABLE(i21 != zero) {
148       i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
149     }
150     const uint8_t* i22 = input[22];
151     assert(i22 != NULL);
152     if XNN_UNPREDICTABLE(i22 != zero) {
153       i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
154     }
155     const uint8_t* i23 = input[23];
156     assert(i23 != NULL);
157     if XNN_UNPREDICTABLE(i23 != zero) {
158       i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
159     }
160     const uint8_t* i24 = input[24];
161     assert(i24 != NULL);
162     if XNN_UNPREDICTABLE(i24 != zero) {
163       i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
164     }
165     input = (const uint8_t**) ((uintptr_t) input + input_stride);
166 
167     size_t c = channels;
168     const void* w = weights;
169     for (; c >= 16; c -= 16) {
170       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
171       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
172       int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
173       int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
174 
175 
176       const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
177       const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
178       const int16x8_t vi0x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
179       const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
180 
181       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
182       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
183       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
184       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
185 
186       const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
187       const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
188       const int16x8_t vi1x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
189       const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
190 
191       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
192       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
193       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
194       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
195 
196       const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
197       const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
198       const int16x8_t vi2x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
199       const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
200 
201       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
202       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
203       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
204       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
205 
206       const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
207       const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
208       const int16x8_t vi3x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
209       const int16x8_t vk3x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
210 
211       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
212       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
213       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
214       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
215 
216       const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
217       const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
218       const int16x8_t vi4x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
219       const int16x8_t vk4x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
220 
221       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
222       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
223       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
224       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
225 
226       const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
227       const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
228       const int16x8_t vi5x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
229       const int16x8_t vk5x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
230 
231       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
232       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
233       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF));
234       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF));
235 
236       const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
237       const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
238       const int16x8_t vi6x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
239       const int16x8_t vk6x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
240 
241       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
242       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
243       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF));
244       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF));
245 
246       const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
247       const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
248       const int16x8_t vi7x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
249       const int16x8_t vk7x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
250 
251       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
252       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
253       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi7x89ABCDEF), vget_low_s16(vk7x89ABCDEF));
254       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi7x89ABCDEF), vget_high_s16(vk7x89ABCDEF));
255 
256       const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
257       const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
258       const int16x8_t vi8x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
259       const int16x8_t vk8x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
260 
261       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
262       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
263       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi8x89ABCDEF), vget_low_s16(vk8x89ABCDEF));
264       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi8x89ABCDEF), vget_high_s16(vk8x89ABCDEF));
265 
266       const int16x8_t vi9x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i9))); i9 += 8;
267       const int16x8_t vk9x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
268       const int16x8_t vi9x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i9))); i9 += 8;
269       const int16x8_t vk9x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
270 
271       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi9x01234567), vget_low_s16(vk9x01234567));
272       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi9x01234567), vget_high_s16(vk9x01234567));
273       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi9x89ABCDEF), vget_low_s16(vk9x89ABCDEF));
274       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi9x89ABCDEF), vget_high_s16(vk9x89ABCDEF));
275 
276       const int16x8_t vi10x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i10))); i10 += 8;
277       const int16x8_t vk10x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
278       const int16x8_t vi10x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i10))); i10 += 8;
279       const int16x8_t vk10x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
280 
281       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi10x01234567), vget_low_s16(vk10x01234567));
282       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi10x01234567), vget_high_s16(vk10x01234567));
283       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi10x89ABCDEF), vget_low_s16(vk10x89ABCDEF));
284       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi10x89ABCDEF), vget_high_s16(vk10x89ABCDEF));
285 
286       const int16x8_t vi11x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i11))); i11 += 8;
287       const int16x8_t vk11x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
288       const int16x8_t vi11x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i11))); i11 += 8;
289       const int16x8_t vk11x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
290 
291       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi11x01234567), vget_low_s16(vk11x01234567));
292       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi11x01234567), vget_high_s16(vk11x01234567));
293       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi11x89ABCDEF), vget_low_s16(vk11x89ABCDEF));
294       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi11x89ABCDEF), vget_high_s16(vk11x89ABCDEF));
295 
296       const int16x8_t vi12x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i12))); i12 += 8;
297       const int16x8_t vk12x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
298       const int16x8_t vi12x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i12))); i12 += 8;
299       const int16x8_t vk12x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
300 
301       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi12x01234567), vget_low_s16(vk12x01234567));
302       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi12x01234567), vget_high_s16(vk12x01234567));
303       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi12x89ABCDEF), vget_low_s16(vk12x89ABCDEF));
304       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi12x89ABCDEF), vget_high_s16(vk12x89ABCDEF));
305 
306       const int16x8_t vi13x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i13))); i13 += 8;
307       const int16x8_t vk13x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
308       const int16x8_t vi13x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i13))); i13 += 8;
309       const int16x8_t vk13x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
310 
311       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi13x01234567), vget_low_s16(vk13x01234567));
312       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi13x01234567), vget_high_s16(vk13x01234567));
313       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi13x89ABCDEF), vget_low_s16(vk13x89ABCDEF));
314       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi13x89ABCDEF), vget_high_s16(vk13x89ABCDEF));
315 
316       const int16x8_t vi14x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i14))); i14 += 8;
317       const int16x8_t vk14x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
318       const int16x8_t vi14x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i14))); i14 += 8;
319       const int16x8_t vk14x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
320 
321       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi14x01234567), vget_low_s16(vk14x01234567));
322       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi14x01234567), vget_high_s16(vk14x01234567));
323       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi14x89ABCDEF), vget_low_s16(vk14x89ABCDEF));
324       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi14x89ABCDEF), vget_high_s16(vk14x89ABCDEF));
325 
326       const int16x8_t vi15x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i15))); i15 += 8;
327       const int16x8_t vk15x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
328       const int16x8_t vi15x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i15))); i15 += 8;
329       const int16x8_t vk15x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
330 
331       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi15x01234567), vget_low_s16(vk15x01234567));
332       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi15x01234567), vget_high_s16(vk15x01234567));
333       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi15x89ABCDEF), vget_low_s16(vk15x89ABCDEF));
334       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi15x89ABCDEF), vget_high_s16(vk15x89ABCDEF));
335 
336       const int16x8_t vi16x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i16))); i16 += 8;
337       const int16x8_t vk16x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
338       const int16x8_t vi16x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i16))); i16 += 8;
339       const int16x8_t vk16x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
340 
341       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi16x01234567), vget_low_s16(vk16x01234567));
342       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi16x01234567), vget_high_s16(vk16x01234567));
343       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi16x89ABCDEF), vget_low_s16(vk16x89ABCDEF));
344       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi16x89ABCDEF), vget_high_s16(vk16x89ABCDEF));
345 
346       const int16x8_t vi17x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i17))); i17 += 8;
347       const int16x8_t vk17x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
348       const int16x8_t vi17x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i17))); i17 += 8;
349       const int16x8_t vk17x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
350 
351       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi17x01234567), vget_low_s16(vk17x01234567));
352       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi17x01234567), vget_high_s16(vk17x01234567));
353       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi17x89ABCDEF), vget_low_s16(vk17x89ABCDEF));
354       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi17x89ABCDEF), vget_high_s16(vk17x89ABCDEF));
355 
356       const int16x8_t vi18x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); i18 += 8;
357       const int16x8_t vk18x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
358       const int16x8_t vi18x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); i18 += 8;
359       const int16x8_t vk18x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
360 
361       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi18x01234567), vget_low_s16(vk18x01234567));
362       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi18x01234567), vget_high_s16(vk18x01234567));
363       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi18x89ABCDEF), vget_low_s16(vk18x89ABCDEF));
364       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi18x89ABCDEF), vget_high_s16(vk18x89ABCDEF));
365 
366       const int16x8_t vi19x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i19))); i19 += 8;
367       const int16x8_t vk19x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
368       const int16x8_t vi19x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i19))); i19 += 8;
369       const int16x8_t vk19x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
370 
371       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi19x01234567), vget_low_s16(vk19x01234567));
372       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi19x01234567), vget_high_s16(vk19x01234567));
373       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi19x89ABCDEF), vget_low_s16(vk19x89ABCDEF));
374       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi19x89ABCDEF), vget_high_s16(vk19x89ABCDEF));
375 
376       const int16x8_t vi20x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i20))); i20 += 8;
377       const int16x8_t vk20x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
378       const int16x8_t vi20x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i20))); i20 += 8;
379       const int16x8_t vk20x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
380 
381       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi20x01234567), vget_low_s16(vk20x01234567));
382       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi20x01234567), vget_high_s16(vk20x01234567));
383       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi20x89ABCDEF), vget_low_s16(vk20x89ABCDEF));
384       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi20x89ABCDEF), vget_high_s16(vk20x89ABCDEF));
385 
386       const int16x8_t vi21x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); i21 += 8;
387       const int16x8_t vk21x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
388       const int16x8_t vi21x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); i21 += 8;
389       const int16x8_t vk21x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
390 
391       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi21x01234567), vget_low_s16(vk21x01234567));
392       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi21x01234567), vget_high_s16(vk21x01234567));
393       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi21x89ABCDEF), vget_low_s16(vk21x89ABCDEF));
394       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi21x89ABCDEF), vget_high_s16(vk21x89ABCDEF));
395 
396       const int16x8_t vi22x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i22))); i22 += 8;
397       const int16x8_t vk22x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
398       const int16x8_t vi22x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i22))); i22 += 8;
399       const int16x8_t vk22x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
400 
401       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567));
402       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567));
403       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi22x89ABCDEF), vget_low_s16(vk22x89ABCDEF));
404       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi22x89ABCDEF), vget_high_s16(vk22x89ABCDEF));
405 
406       const int16x8_t vi23x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i23))); i23 += 8;
407       const int16x8_t vk23x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
408       const int16x8_t vi23x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i23))); i23 += 8;
409       const int16x8_t vk23x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
410 
411       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi23x01234567), vget_low_s16(vk23x01234567));
412       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi23x01234567), vget_high_s16(vk23x01234567));
413       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi23x89ABCDEF), vget_low_s16(vk23x89ABCDEF));
414       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi23x89ABCDEF), vget_high_s16(vk23x89ABCDEF));
415 
416       const int16x8_t vi24x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i24))); i24 += 8;
417       const int16x8_t vk24x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
418       const int16x8_t vi24x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i24))); i24 += 8;
419       const int16x8_t vk24x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
420 
421       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi24x01234567), vget_low_s16(vk24x01234567));
422       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi24x01234567), vget_high_s16(vk24x01234567));
423       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi24x89ABCDEF), vget_low_s16(vk24x89ABCDEF));
424       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi24x89ABCDEF), vget_high_s16(vk24x89ABCDEF));
425 
426       vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
427       vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
428       vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
429       vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
430 
431       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
432       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
433       vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
434       vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
435 
436       vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
437       vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
438       vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
439       vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
440 
441 #if XNN_ARCH_ARM64
442       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
443       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
444 
445       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
446       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
447 
448       uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
449 #else  // !XNN_ARCH_ARM64
450       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
451       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
452 
453       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
454       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
455 
456       uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
457 #endif  // !XNN_ARCH_ARM64
458 
459       vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
460 
461       vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
462 
463       vst1q_u8(output, vout0123456789ABCDEF); output += 16;
464     }
465     if XNN_UNLIKELY(c != 0) {
466       const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 16);
467       do {
468         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
469         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
470 
471         const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
472         const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(k), vkernel_zero_point)); k += 8;
473 
474         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
475         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
476         const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
477         const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 8)), vkernel_zero_point));
478 
479         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
480         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
481         const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
482         const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 24)), vkernel_zero_point));
483 
484         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
485         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
486         const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
487         const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 40)), vkernel_zero_point));
488 
489         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
490         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
491         const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
492         const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 56)), vkernel_zero_point));
493 
494         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
495         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
496         const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
497         const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 72)), vkernel_zero_point));
498 
499         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
500         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
501         const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
502         const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 88)), vkernel_zero_point));
503 
504         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
505         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
506         const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
507         const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 104)), vkernel_zero_point));
508 
509         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
510         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
511         const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
512         const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 120)), vkernel_zero_point));
513 
514         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
515         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
516         const int16x8_t vi9x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i9))); i9 += 8;
517         const int16x8_t vk9x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 136)), vkernel_zero_point));
518 
519         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi9x01234567), vget_low_s16(vk9x01234567));
520         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi9x01234567), vget_high_s16(vk9x01234567));
521         const int16x8_t vi10x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i10))); i10 += 8;
522         const int16x8_t vk10x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 152)), vkernel_zero_point));
523 
524         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi10x01234567), vget_low_s16(vk10x01234567));
525         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi10x01234567), vget_high_s16(vk10x01234567));
526         const int16x8_t vi11x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i11))); i11 += 8;
527         const int16x8_t vk11x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 168)), vkernel_zero_point));
528 
529         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi11x01234567), vget_low_s16(vk11x01234567));
530         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi11x01234567), vget_high_s16(vk11x01234567));
531         const int16x8_t vi12x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i12))); i12 += 8;
532         const int16x8_t vk12x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 184)), vkernel_zero_point));
533 
534         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi12x01234567), vget_low_s16(vk12x01234567));
535         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi12x01234567), vget_high_s16(vk12x01234567));
536         const int16x8_t vi13x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i13))); i13 += 8;
537         const int16x8_t vk13x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 200)), vkernel_zero_point));
538 
539         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi13x01234567), vget_low_s16(vk13x01234567));
540         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi13x01234567), vget_high_s16(vk13x01234567));
541         const int16x8_t vi14x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i14))); i14 += 8;
542         const int16x8_t vk14x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 216)), vkernel_zero_point));
543 
544         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi14x01234567), vget_low_s16(vk14x01234567));
545         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi14x01234567), vget_high_s16(vk14x01234567));
546         const int16x8_t vi15x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i15))); i15 += 8;
547         const int16x8_t vk15x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 232)), vkernel_zero_point));
548 
549         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi15x01234567), vget_low_s16(vk15x01234567));
550         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi15x01234567), vget_high_s16(vk15x01234567));
551         const int16x8_t vi16x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i16))); i16 += 8;
552         const int16x8_t vk16x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 248)), vkernel_zero_point));
553 
554         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi16x01234567), vget_low_s16(vk16x01234567));
555         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi16x01234567), vget_high_s16(vk16x01234567));
556         const int16x8_t vi17x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i17))); i17 += 8;
557         const int16x8_t vk17x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 264)), vkernel_zero_point));
558 
559         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi17x01234567), vget_low_s16(vk17x01234567));
560         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi17x01234567), vget_high_s16(vk17x01234567));
561         const int16x8_t vi18x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); i18 += 8;
562         const int16x8_t vk18x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 280)), vkernel_zero_point));
563 
564         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi18x01234567), vget_low_s16(vk18x01234567));
565         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi18x01234567), vget_high_s16(vk18x01234567));
566         const int16x8_t vi19x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i19))); i19 += 8;
567         const int16x8_t vk19x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 296)), vkernel_zero_point));
568 
569         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi19x01234567), vget_low_s16(vk19x01234567));
570         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi19x01234567), vget_high_s16(vk19x01234567));
571         const int16x8_t vi20x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i20))); i20 += 8;
572         const int16x8_t vk20x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 312)), vkernel_zero_point));
573 
574         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi20x01234567), vget_low_s16(vk20x01234567));
575         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi20x01234567), vget_high_s16(vk20x01234567));
576         const int16x8_t vi21x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); i21 += 8;
577         const int16x8_t vk21x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 328)), vkernel_zero_point));
578 
579         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi21x01234567), vget_low_s16(vk21x01234567));
580         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi21x01234567), vget_high_s16(vk21x01234567));
581         const int16x8_t vi22x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i22))); i22 += 8;
582         const int16x8_t vk22x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 344)), vkernel_zero_point));
583 
584         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567));
585         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567));
586         const int16x8_t vi23x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i23))); i23 += 8;
587         const int16x8_t vk23x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 360)), vkernel_zero_point));
588 
589         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi23x01234567), vget_low_s16(vk23x01234567));
590         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi23x01234567), vget_high_s16(vk23x01234567));
591         const int16x8_t vi24x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i24))); i24 += 8;
592         const int16x8_t vk24x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 376)), vkernel_zero_point));
593 
594         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi24x01234567), vget_low_s16(vk24x01234567));
595         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi24x01234567), vget_high_s16(vk24x01234567));
596 
597         vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
598         vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
599 
600         vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
601         vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
602 
603         vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
604         vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
605 
606 #if XNN_ARCH_ARM64
607         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
608 #else
609         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
610 #endif
611         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
612 
613         uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
614         vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
615         vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
616 
617         if XNN_LIKELY(c >= 8) {
618           vst1_u8(output, vout01234567); output += 8;
619           c -= 8;
620         } else {
621           if (c & 4) {
622             vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
623             vout01234567 = vext_u8(vout01234567, vout01234567, 4);
624           }
625           if (c & 2) {
626             vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
627             vout01234567 = vext_u8(vout01234567, vout01234567, 2);
628           }
629           if (c & 1) {
630             vst1_lane_u8(output, vout01234567, 0); output += 1;
631           }
632           c = 0;
633         }
634       } while (c != 0);
635     }
636 
637     output = (uint8_t*) ((uintptr_t) output + output_increment);
638   } while (--output_width != 0);
639 }
640