xref: /aosp_15_r20/external/XNNPACK/src/qu8-dwconv/gen/up32x25-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_up32x25__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_up32x25__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 >= 32; c -= 32) {
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       int32x4_t vaccGHIJ = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
175       int32x4_t vaccKLMN = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
176       int32x4_t vaccOPQR = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
177       int32x4_t vaccSTUV = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
178 
179 
180       const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
181       const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
182       const int16x8_t vi0x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
183       const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
184       const int16x8_t vi0xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
185       const int16x8_t vk0xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
186       const int16x8_t vi0xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
187       const int16x8_t vk0xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
188 
189       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
190       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
191       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
192       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
193       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi0xGHIJKLMN), vget_low_s16(vk0xGHIJKLMN));
194       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi0xGHIJKLMN), vget_high_s16(vk0xGHIJKLMN));
195       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi0xOPQRSTUV), vget_low_s16(vk0xOPQRSTUV));
196       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi0xOPQRSTUV), vget_high_s16(vk0xOPQRSTUV));
197 
198       const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
199       const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
200       const int16x8_t vi1x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
201       const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
202       const int16x8_t vi1xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
203       const int16x8_t vk1xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
204       const int16x8_t vi1xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
205       const int16x8_t vk1xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
206 
207       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
208       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
209       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
210       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
211       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi1xGHIJKLMN), vget_low_s16(vk1xGHIJKLMN));
212       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi1xGHIJKLMN), vget_high_s16(vk1xGHIJKLMN));
213       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi1xOPQRSTUV), vget_low_s16(vk1xOPQRSTUV));
214       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi1xOPQRSTUV), vget_high_s16(vk1xOPQRSTUV));
215 
216       const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
217       const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
218       const int16x8_t vi2x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
219       const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
220       const int16x8_t vi2xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
221       const int16x8_t vk2xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
222       const int16x8_t vi2xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
223       const int16x8_t vk2xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
224 
225       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
226       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
227       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
228       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
229       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi2xGHIJKLMN), vget_low_s16(vk2xGHIJKLMN));
230       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi2xGHIJKLMN), vget_high_s16(vk2xGHIJKLMN));
231       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi2xOPQRSTUV), vget_low_s16(vk2xOPQRSTUV));
232       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi2xOPQRSTUV), vget_high_s16(vk2xOPQRSTUV));
233 
234       const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
235       const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
236       const int16x8_t vi3x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
237       const int16x8_t vk3x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
238       const int16x8_t vi3xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
239       const int16x8_t vk3xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
240       const int16x8_t vi3xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
241       const int16x8_t vk3xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
242 
243       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
244       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
245       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
246       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
247       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN));
248       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN));
249       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi3xOPQRSTUV), vget_low_s16(vk3xOPQRSTUV));
250       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi3xOPQRSTUV), vget_high_s16(vk3xOPQRSTUV));
251 
252       const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
253       const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
254       const int16x8_t vi4x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
255       const int16x8_t vk4x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
256       const int16x8_t vi4xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
257       const int16x8_t vk4xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
258       const int16x8_t vi4xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
259       const int16x8_t vk4xOPQRSTUV = 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(vi4x01234567), vget_low_s16(vk4x01234567));
262       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
263       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
264       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
265       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi4xGHIJKLMN), vget_low_s16(vk4xGHIJKLMN));
266       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi4xGHIJKLMN), vget_high_s16(vk4xGHIJKLMN));
267       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV));
268       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV));
269 
270       const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
271       const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
272       const int16x8_t vi5x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
273       const int16x8_t vk5x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
274       const int16x8_t vi5xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
275       const int16x8_t vk5xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
276       const int16x8_t vi5xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
277       const int16x8_t vk5xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
278 
279       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
280       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
281       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF));
282       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF));
283       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi5xGHIJKLMN), vget_low_s16(vk5xGHIJKLMN));
284       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi5xGHIJKLMN), vget_high_s16(vk5xGHIJKLMN));
285       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi5xOPQRSTUV), vget_low_s16(vk5xOPQRSTUV));
286       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi5xOPQRSTUV), vget_high_s16(vk5xOPQRSTUV));
287 
288       const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
289       const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
290       const int16x8_t vi6x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
291       const int16x8_t vk6x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
292       const int16x8_t vi6xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
293       const int16x8_t vk6xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
294       const int16x8_t vi6xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
295       const int16x8_t vk6xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
296 
297       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
298       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
299       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF));
300       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF));
301       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi6xGHIJKLMN), vget_low_s16(vk6xGHIJKLMN));
302       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi6xGHIJKLMN), vget_high_s16(vk6xGHIJKLMN));
303       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi6xOPQRSTUV), vget_low_s16(vk6xOPQRSTUV));
304       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi6xOPQRSTUV), vget_high_s16(vk6xOPQRSTUV));
305 
306       const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
307       const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
308       const int16x8_t vi7x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
309       const int16x8_t vk7x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
310       const int16x8_t vi7xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
311       const int16x8_t vk7xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
312       const int16x8_t vi7xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
313       const int16x8_t vk7xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
314 
315       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
316       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
317       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi7x89ABCDEF), vget_low_s16(vk7x89ABCDEF));
318       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi7x89ABCDEF), vget_high_s16(vk7x89ABCDEF));
319       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN));
320       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN));
321       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi7xOPQRSTUV), vget_low_s16(vk7xOPQRSTUV));
322       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi7xOPQRSTUV), vget_high_s16(vk7xOPQRSTUV));
323 
324       const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
325       const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
326       const int16x8_t vi8x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
327       const int16x8_t vk8x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
328       const int16x8_t vi8xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
329       const int16x8_t vk8xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
330       const int16x8_t vi8xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
331       const int16x8_t vk8xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
332 
333       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
334       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
335       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi8x89ABCDEF), vget_low_s16(vk8x89ABCDEF));
336       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi8x89ABCDEF), vget_high_s16(vk8x89ABCDEF));
337       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi8xGHIJKLMN), vget_low_s16(vk8xGHIJKLMN));
338       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi8xGHIJKLMN), vget_high_s16(vk8xGHIJKLMN));
339       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi8xOPQRSTUV), vget_low_s16(vk8xOPQRSTUV));
340       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi8xOPQRSTUV), vget_high_s16(vk8xOPQRSTUV));
341 
342       const int16x8_t vi9x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i9))); i9 += 8;
343       const int16x8_t vk9x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
344       const int16x8_t vi9x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i9))); i9 += 8;
345       const int16x8_t vk9x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
346       const int16x8_t vi9xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i9))); i9 += 8;
347       const int16x8_t vk9xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
348       const int16x8_t vi9xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i9))); i9 += 8;
349       const int16x8_t vk9xOPQRSTUV = 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(vi9x01234567), vget_low_s16(vk9x01234567));
352       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi9x01234567), vget_high_s16(vk9x01234567));
353       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi9x89ABCDEF), vget_low_s16(vk9x89ABCDEF));
354       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi9x89ABCDEF), vget_high_s16(vk9x89ABCDEF));
355       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi9xGHIJKLMN), vget_low_s16(vk9xGHIJKLMN));
356       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi9xGHIJKLMN), vget_high_s16(vk9xGHIJKLMN));
357       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi9xOPQRSTUV), vget_low_s16(vk9xOPQRSTUV));
358       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi9xOPQRSTUV), vget_high_s16(vk9xOPQRSTUV));
359 
360       const int16x8_t vi10x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i10))); i10 += 8;
361       const int16x8_t vk10x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
362       const int16x8_t vi10x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i10))); i10 += 8;
363       const int16x8_t vk10x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
364       const int16x8_t vi10xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i10))); i10 += 8;
365       const int16x8_t vk10xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
366       const int16x8_t vi10xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i10))); i10 += 8;
367       const int16x8_t vk10xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
368 
369       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi10x01234567), vget_low_s16(vk10x01234567));
370       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi10x01234567), vget_high_s16(vk10x01234567));
371       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi10x89ABCDEF), vget_low_s16(vk10x89ABCDEF));
372       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi10x89ABCDEF), vget_high_s16(vk10x89ABCDEF));
373       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi10xGHIJKLMN), vget_low_s16(vk10xGHIJKLMN));
374       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi10xGHIJKLMN), vget_high_s16(vk10xGHIJKLMN));
375       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi10xOPQRSTUV), vget_low_s16(vk10xOPQRSTUV));
376       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi10xOPQRSTUV), vget_high_s16(vk10xOPQRSTUV));
377 
378       const int16x8_t vi11x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i11))); i11 += 8;
379       const int16x8_t vk11x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
380       const int16x8_t vi11x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i11))); i11 += 8;
381       const int16x8_t vk11x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
382       const int16x8_t vi11xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i11))); i11 += 8;
383       const int16x8_t vk11xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
384       const int16x8_t vi11xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i11))); i11 += 8;
385       const int16x8_t vk11xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
386 
387       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi11x01234567), vget_low_s16(vk11x01234567));
388       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi11x01234567), vget_high_s16(vk11x01234567));
389       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi11x89ABCDEF), vget_low_s16(vk11x89ABCDEF));
390       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi11x89ABCDEF), vget_high_s16(vk11x89ABCDEF));
391       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi11xGHIJKLMN), vget_low_s16(vk11xGHIJKLMN));
392       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi11xGHIJKLMN), vget_high_s16(vk11xGHIJKLMN));
393       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi11xOPQRSTUV), vget_low_s16(vk11xOPQRSTUV));
394       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi11xOPQRSTUV), vget_high_s16(vk11xOPQRSTUV));
395 
396       const int16x8_t vi12x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i12))); i12 += 8;
397       const int16x8_t vk12x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
398       const int16x8_t vi12x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i12))); i12 += 8;
399       const int16x8_t vk12x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
400       const int16x8_t vi12xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i12))); i12 += 8;
401       const int16x8_t vk12xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
402       const int16x8_t vi12xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i12))); i12 += 8;
403       const int16x8_t vk12xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
404 
405       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi12x01234567), vget_low_s16(vk12x01234567));
406       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi12x01234567), vget_high_s16(vk12x01234567));
407       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi12x89ABCDEF), vget_low_s16(vk12x89ABCDEF));
408       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi12x89ABCDEF), vget_high_s16(vk12x89ABCDEF));
409       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi12xGHIJKLMN), vget_low_s16(vk12xGHIJKLMN));
410       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi12xGHIJKLMN), vget_high_s16(vk12xGHIJKLMN));
411       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi12xOPQRSTUV), vget_low_s16(vk12xOPQRSTUV));
412       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi12xOPQRSTUV), vget_high_s16(vk12xOPQRSTUV));
413 
414       const int16x8_t vi13x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i13))); i13 += 8;
415       const int16x8_t vk13x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
416       const int16x8_t vi13x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i13))); i13 += 8;
417       const int16x8_t vk13x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
418       const int16x8_t vi13xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i13))); i13 += 8;
419       const int16x8_t vk13xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
420       const int16x8_t vi13xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i13))); i13 += 8;
421       const int16x8_t vk13xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
422 
423       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi13x01234567), vget_low_s16(vk13x01234567));
424       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi13x01234567), vget_high_s16(vk13x01234567));
425       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi13x89ABCDEF), vget_low_s16(vk13x89ABCDEF));
426       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi13x89ABCDEF), vget_high_s16(vk13x89ABCDEF));
427       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi13xGHIJKLMN), vget_low_s16(vk13xGHIJKLMN));
428       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi13xGHIJKLMN), vget_high_s16(vk13xGHIJKLMN));
429       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi13xOPQRSTUV), vget_low_s16(vk13xOPQRSTUV));
430       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi13xOPQRSTUV), vget_high_s16(vk13xOPQRSTUV));
431 
432       const int16x8_t vi14x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i14))); i14 += 8;
433       const int16x8_t vk14x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
434       const int16x8_t vi14x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i14))); i14 += 8;
435       const int16x8_t vk14x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
436       const int16x8_t vi14xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i14))); i14 += 8;
437       const int16x8_t vk14xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
438       const int16x8_t vi14xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i14))); i14 += 8;
439       const int16x8_t vk14xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
440 
441       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi14x01234567), vget_low_s16(vk14x01234567));
442       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi14x01234567), vget_high_s16(vk14x01234567));
443       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi14x89ABCDEF), vget_low_s16(vk14x89ABCDEF));
444       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi14x89ABCDEF), vget_high_s16(vk14x89ABCDEF));
445       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi14xGHIJKLMN), vget_low_s16(vk14xGHIJKLMN));
446       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi14xGHIJKLMN), vget_high_s16(vk14xGHIJKLMN));
447       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi14xOPQRSTUV), vget_low_s16(vk14xOPQRSTUV));
448       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi14xOPQRSTUV), vget_high_s16(vk14xOPQRSTUV));
449 
450       const int16x8_t vi15x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i15))); i15 += 8;
451       const int16x8_t vk15x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
452       const int16x8_t vi15x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i15))); i15 += 8;
453       const int16x8_t vk15x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
454       const int16x8_t vi15xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i15))); i15 += 8;
455       const int16x8_t vk15xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
456       const int16x8_t vi15xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i15))); i15 += 8;
457       const int16x8_t vk15xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
458 
459       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi15x01234567), vget_low_s16(vk15x01234567));
460       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi15x01234567), vget_high_s16(vk15x01234567));
461       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi15x89ABCDEF), vget_low_s16(vk15x89ABCDEF));
462       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi15x89ABCDEF), vget_high_s16(vk15x89ABCDEF));
463       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi15xGHIJKLMN), vget_low_s16(vk15xGHIJKLMN));
464       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi15xGHIJKLMN), vget_high_s16(vk15xGHIJKLMN));
465       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi15xOPQRSTUV), vget_low_s16(vk15xOPQRSTUV));
466       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi15xOPQRSTUV), vget_high_s16(vk15xOPQRSTUV));
467 
468       const int16x8_t vi16x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i16))); i16 += 8;
469       const int16x8_t vk16x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
470       const int16x8_t vi16x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i16))); i16 += 8;
471       const int16x8_t vk16x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
472       const int16x8_t vi16xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i16))); i16 += 8;
473       const int16x8_t vk16xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
474       const int16x8_t vi16xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i16))); i16 += 8;
475       const int16x8_t vk16xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
476 
477       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi16x01234567), vget_low_s16(vk16x01234567));
478       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi16x01234567), vget_high_s16(vk16x01234567));
479       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi16x89ABCDEF), vget_low_s16(vk16x89ABCDEF));
480       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi16x89ABCDEF), vget_high_s16(vk16x89ABCDEF));
481       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi16xGHIJKLMN), vget_low_s16(vk16xGHIJKLMN));
482       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi16xGHIJKLMN), vget_high_s16(vk16xGHIJKLMN));
483       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi16xOPQRSTUV), vget_low_s16(vk16xOPQRSTUV));
484       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi16xOPQRSTUV), vget_high_s16(vk16xOPQRSTUV));
485 
486       const int16x8_t vi17x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i17))); i17 += 8;
487       const int16x8_t vk17x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
488       const int16x8_t vi17x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i17))); i17 += 8;
489       const int16x8_t vk17x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
490       const int16x8_t vi17xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i17))); i17 += 8;
491       const int16x8_t vk17xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
492       const int16x8_t vi17xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i17))); i17 += 8;
493       const int16x8_t vk17xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
494 
495       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi17x01234567), vget_low_s16(vk17x01234567));
496       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi17x01234567), vget_high_s16(vk17x01234567));
497       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi17x89ABCDEF), vget_low_s16(vk17x89ABCDEF));
498       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi17x89ABCDEF), vget_high_s16(vk17x89ABCDEF));
499       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi17xGHIJKLMN), vget_low_s16(vk17xGHIJKLMN));
500       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi17xGHIJKLMN), vget_high_s16(vk17xGHIJKLMN));
501       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi17xOPQRSTUV), vget_low_s16(vk17xOPQRSTUV));
502       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi17xOPQRSTUV), vget_high_s16(vk17xOPQRSTUV));
503 
504       const int16x8_t vi18x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); i18 += 8;
505       const int16x8_t vk18x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
506       const int16x8_t vi18x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); i18 += 8;
507       const int16x8_t vk18x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
508       const int16x8_t vi18xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); i18 += 8;
509       const int16x8_t vk18xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
510       const int16x8_t vi18xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); i18 += 8;
511       const int16x8_t vk18xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
512 
513       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi18x01234567), vget_low_s16(vk18x01234567));
514       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi18x01234567), vget_high_s16(vk18x01234567));
515       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi18x89ABCDEF), vget_low_s16(vk18x89ABCDEF));
516       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi18x89ABCDEF), vget_high_s16(vk18x89ABCDEF));
517       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi18xGHIJKLMN), vget_low_s16(vk18xGHIJKLMN));
518       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi18xGHIJKLMN), vget_high_s16(vk18xGHIJKLMN));
519       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi18xOPQRSTUV), vget_low_s16(vk18xOPQRSTUV));
520       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi18xOPQRSTUV), vget_high_s16(vk18xOPQRSTUV));
521 
522       const int16x8_t vi19x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i19))); i19 += 8;
523       const int16x8_t vk19x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
524       const int16x8_t vi19x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i19))); i19 += 8;
525       const int16x8_t vk19x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
526       const int16x8_t vi19xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i19))); i19 += 8;
527       const int16x8_t vk19xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
528       const int16x8_t vi19xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i19))); i19 += 8;
529       const int16x8_t vk19xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
530 
531       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi19x01234567), vget_low_s16(vk19x01234567));
532       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi19x01234567), vget_high_s16(vk19x01234567));
533       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi19x89ABCDEF), vget_low_s16(vk19x89ABCDEF));
534       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi19x89ABCDEF), vget_high_s16(vk19x89ABCDEF));
535       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi19xGHIJKLMN), vget_low_s16(vk19xGHIJKLMN));
536       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi19xGHIJKLMN), vget_high_s16(vk19xGHIJKLMN));
537       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi19xOPQRSTUV), vget_low_s16(vk19xOPQRSTUV));
538       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi19xOPQRSTUV), vget_high_s16(vk19xOPQRSTUV));
539 
540       const int16x8_t vi20x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i20))); i20 += 8;
541       const int16x8_t vk20x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
542       const int16x8_t vi20x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i20))); i20 += 8;
543       const int16x8_t vk20x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
544       const int16x8_t vi20xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i20))); i20 += 8;
545       const int16x8_t vk20xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
546       const int16x8_t vi20xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i20))); i20 += 8;
547       const int16x8_t vk20xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
548 
549       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi20x01234567), vget_low_s16(vk20x01234567));
550       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi20x01234567), vget_high_s16(vk20x01234567));
551       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi20x89ABCDEF), vget_low_s16(vk20x89ABCDEF));
552       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi20x89ABCDEF), vget_high_s16(vk20x89ABCDEF));
553       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi20xGHIJKLMN), vget_low_s16(vk20xGHIJKLMN));
554       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi20xGHIJKLMN), vget_high_s16(vk20xGHIJKLMN));
555       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi20xOPQRSTUV), vget_low_s16(vk20xOPQRSTUV));
556       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi20xOPQRSTUV), vget_high_s16(vk20xOPQRSTUV));
557 
558       const int16x8_t vi21x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); i21 += 8;
559       const int16x8_t vk21x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
560       const int16x8_t vi21x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); i21 += 8;
561       const int16x8_t vk21x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
562       const int16x8_t vi21xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); i21 += 8;
563       const int16x8_t vk21xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
564       const int16x8_t vi21xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); i21 += 8;
565       const int16x8_t vk21xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
566 
567       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi21x01234567), vget_low_s16(vk21x01234567));
568       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi21x01234567), vget_high_s16(vk21x01234567));
569       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi21x89ABCDEF), vget_low_s16(vk21x89ABCDEF));
570       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi21x89ABCDEF), vget_high_s16(vk21x89ABCDEF));
571       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi21xGHIJKLMN), vget_low_s16(vk21xGHIJKLMN));
572       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi21xGHIJKLMN), vget_high_s16(vk21xGHIJKLMN));
573       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi21xOPQRSTUV), vget_low_s16(vk21xOPQRSTUV));
574       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi21xOPQRSTUV), vget_high_s16(vk21xOPQRSTUV));
575 
576       const int16x8_t vi22x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i22))); i22 += 8;
577       const int16x8_t vk22x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
578       const int16x8_t vi22x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i22))); i22 += 8;
579       const int16x8_t vk22x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
580       const int16x8_t vi22xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i22))); i22 += 8;
581       const int16x8_t vk22xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
582       const int16x8_t vi22xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i22))); i22 += 8;
583       const int16x8_t vk22xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
584 
585       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567));
586       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567));
587       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi22x89ABCDEF), vget_low_s16(vk22x89ABCDEF));
588       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi22x89ABCDEF), vget_high_s16(vk22x89ABCDEF));
589       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi22xGHIJKLMN), vget_low_s16(vk22xGHIJKLMN));
590       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi22xGHIJKLMN), vget_high_s16(vk22xGHIJKLMN));
591       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi22xOPQRSTUV), vget_low_s16(vk22xOPQRSTUV));
592       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi22xOPQRSTUV), vget_high_s16(vk22xOPQRSTUV));
593 
594       const int16x8_t vi23x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i23))); i23 += 8;
595       const int16x8_t vk23x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
596       const int16x8_t vi23x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i23))); i23 += 8;
597       const int16x8_t vk23x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
598       const int16x8_t vi23xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i23))); i23 += 8;
599       const int16x8_t vk23xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
600       const int16x8_t vi23xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i23))); i23 += 8;
601       const int16x8_t vk23xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
602 
603       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi23x01234567), vget_low_s16(vk23x01234567));
604       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi23x01234567), vget_high_s16(vk23x01234567));
605       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi23x89ABCDEF), vget_low_s16(vk23x89ABCDEF));
606       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi23x89ABCDEF), vget_high_s16(vk23x89ABCDEF));
607       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi23xGHIJKLMN), vget_low_s16(vk23xGHIJKLMN));
608       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi23xGHIJKLMN), vget_high_s16(vk23xGHIJKLMN));
609       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi23xOPQRSTUV), vget_low_s16(vk23xOPQRSTUV));
610       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi23xOPQRSTUV), vget_high_s16(vk23xOPQRSTUV));
611 
612       const int16x8_t vi24x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i24))); i24 += 8;
613       const int16x8_t vk24x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
614       const int16x8_t vi24x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i24))); i24 += 8;
615       const int16x8_t vk24x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
616       const int16x8_t vi24xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i24))); i24 += 8;
617       const int16x8_t vk24xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
618       const int16x8_t vi24xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i24))); i24 += 8;
619       const int16x8_t vk24xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
620 
621       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi24x01234567), vget_low_s16(vk24x01234567));
622       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi24x01234567), vget_high_s16(vk24x01234567));
623       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi24x89ABCDEF), vget_low_s16(vk24x89ABCDEF));
624       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi24x89ABCDEF), vget_high_s16(vk24x89ABCDEF));
625       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi24xGHIJKLMN), vget_low_s16(vk24xGHIJKLMN));
626       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi24xGHIJKLMN), vget_high_s16(vk24xGHIJKLMN));
627       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi24xOPQRSTUV), vget_low_s16(vk24xOPQRSTUV));
628       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi24xOPQRSTUV), vget_high_s16(vk24xOPQRSTUV));
629 
630       vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
631       vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
632       vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
633       vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
634       vaccGHIJ = vqshlq_s32(vaccGHIJ, vright_pre_shift);
635       vaccKLMN = vqshlq_s32(vaccKLMN, vright_pre_shift);
636       vaccOPQR = vqshlq_s32(vaccOPQR, vright_pre_shift);
637       vaccSTUV = vqshlq_s32(vaccSTUV, vright_pre_shift);
638 
639       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
640       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
641       vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
642       vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
643       vaccGHIJ = vqdmulhq_s32(vaccGHIJ, vmultiplier);
644       vaccKLMN = vqdmulhq_s32(vaccKLMN, vmultiplier);
645       vaccOPQR = vqdmulhq_s32(vaccOPQR, vmultiplier);
646       vaccSTUV = vqdmulhq_s32(vaccSTUV, vmultiplier);
647 
648       vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
649       vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
650       vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
651       vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
652       vaccGHIJ = vrshlq_s32(vaccGHIJ, vright_post_shift);
653       vaccKLMN = vrshlq_s32(vaccKLMN, vright_post_shift);
654       vaccOPQR = vrshlq_s32(vaccOPQR, vright_post_shift);
655       vaccSTUV = vrshlq_s32(vaccSTUV, vright_post_shift);
656 
657 #if XNN_ARCH_ARM64
658       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
659       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
660       int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);
661       int16x8_t vaccOPQRSTUV = vqmovn_high_s32(vqmovn_s32(vaccOPQR), vaccSTUV);
662 
663       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
664       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
665       vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
666       vaccOPQRSTUV = vqaddq_s16(vaccOPQRSTUV, voutput_zero_point);
667 
668       uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
669       uint8x16_t voutGHIJKLMNOPQRSTUV = vqmovun_high_s16(vqmovun_s16(vaccGHIJKLMN), vaccOPQRSTUV);
670 #else  // !XNN_ARCH_ARM64
671       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
672       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
673       int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));
674       int16x8_t vaccOPQRSTUV = vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV));
675 
676       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
677       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
678       vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
679       vaccOPQRSTUV = vqaddq_s16(vaccOPQRSTUV, voutput_zero_point);
680 
681       uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
682       uint8x16_t voutGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vaccGHIJKLMN), vqmovun_s16(vaccOPQRSTUV));
683 #endif  // !XNN_ARCH_ARM64
684 
685       vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
686       voutGHIJKLMNOPQRSTUV = vmaxq_u8(voutGHIJKLMNOPQRSTUV, voutput_min);
687 
688       vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
689       voutGHIJKLMNOPQRSTUV = vminq_u8(voutGHIJKLMNOPQRSTUV, voutput_max);
690 
691       vst1q_u8(output, vout0123456789ABCDEF); output += 16;
692       vst1q_u8(output, voutGHIJKLMNOPQRSTUV); output += 16;
693     }
694     if XNN_UNLIKELY(c != 0) {
695       const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 32);
696       do {
697         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
698         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
699 
700         const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
701         const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(k), vkernel_zero_point)); k += 8;
702 
703         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
704         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
705         const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
706         const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 24)), vkernel_zero_point));
707 
708         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
709         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
710         const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
711         const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 56)), vkernel_zero_point));
712 
713         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
714         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
715         const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
716         const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 88)), vkernel_zero_point));
717 
718         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
719         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
720         const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
721         const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 120)), vkernel_zero_point));
722 
723         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
724         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
725         const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
726         const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 152)), vkernel_zero_point));
727 
728         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
729         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
730         const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
731         const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 184)), vkernel_zero_point));
732 
733         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
734         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
735         const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
736         const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 216)), vkernel_zero_point));
737 
738         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
739         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
740         const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
741         const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 248)), vkernel_zero_point));
742 
743         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
744         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
745         const int16x8_t vi9x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i9))); i9 += 8;
746         const int16x8_t vk9x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 280)), vkernel_zero_point));
747 
748         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi9x01234567), vget_low_s16(vk9x01234567));
749         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi9x01234567), vget_high_s16(vk9x01234567));
750         const int16x8_t vi10x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i10))); i10 += 8;
751         const int16x8_t vk10x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 312)), vkernel_zero_point));
752 
753         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi10x01234567), vget_low_s16(vk10x01234567));
754         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi10x01234567), vget_high_s16(vk10x01234567));
755         const int16x8_t vi11x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i11))); i11 += 8;
756         const int16x8_t vk11x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 344)), vkernel_zero_point));
757 
758         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi11x01234567), vget_low_s16(vk11x01234567));
759         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi11x01234567), vget_high_s16(vk11x01234567));
760         const int16x8_t vi12x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i12))); i12 += 8;
761         const int16x8_t vk12x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 376)), vkernel_zero_point));
762 
763         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi12x01234567), vget_low_s16(vk12x01234567));
764         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi12x01234567), vget_high_s16(vk12x01234567));
765         const int16x8_t vi13x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i13))); i13 += 8;
766         const int16x8_t vk13x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 408)), vkernel_zero_point));
767 
768         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi13x01234567), vget_low_s16(vk13x01234567));
769         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi13x01234567), vget_high_s16(vk13x01234567));
770         const int16x8_t vi14x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i14))); i14 += 8;
771         const int16x8_t vk14x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 440)), vkernel_zero_point));
772 
773         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi14x01234567), vget_low_s16(vk14x01234567));
774         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi14x01234567), vget_high_s16(vk14x01234567));
775         const int16x8_t vi15x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i15))); i15 += 8;
776         const int16x8_t vk15x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 472)), vkernel_zero_point));
777 
778         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi15x01234567), vget_low_s16(vk15x01234567));
779         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi15x01234567), vget_high_s16(vk15x01234567));
780         const int16x8_t vi16x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i16))); i16 += 8;
781         const int16x8_t vk16x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 504)), vkernel_zero_point));
782 
783         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi16x01234567), vget_low_s16(vk16x01234567));
784         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi16x01234567), vget_high_s16(vk16x01234567));
785         const int16x8_t vi17x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i17))); i17 += 8;
786         const int16x8_t vk17x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 536)), vkernel_zero_point));
787 
788         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi17x01234567), vget_low_s16(vk17x01234567));
789         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi17x01234567), vget_high_s16(vk17x01234567));
790         const int16x8_t vi18x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); i18 += 8;
791         const int16x8_t vk18x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 568)), vkernel_zero_point));
792 
793         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi18x01234567), vget_low_s16(vk18x01234567));
794         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi18x01234567), vget_high_s16(vk18x01234567));
795         const int16x8_t vi19x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i19))); i19 += 8;
796         const int16x8_t vk19x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 600)), vkernel_zero_point));
797 
798         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi19x01234567), vget_low_s16(vk19x01234567));
799         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi19x01234567), vget_high_s16(vk19x01234567));
800         const int16x8_t vi20x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i20))); i20 += 8;
801         const int16x8_t vk20x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 632)), vkernel_zero_point));
802 
803         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi20x01234567), vget_low_s16(vk20x01234567));
804         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi20x01234567), vget_high_s16(vk20x01234567));
805         const int16x8_t vi21x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); i21 += 8;
806         const int16x8_t vk21x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 664)), vkernel_zero_point));
807 
808         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi21x01234567), vget_low_s16(vk21x01234567));
809         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi21x01234567), vget_high_s16(vk21x01234567));
810         const int16x8_t vi22x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i22))); i22 += 8;
811         const int16x8_t vk22x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 696)), vkernel_zero_point));
812 
813         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567));
814         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567));
815         const int16x8_t vi23x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i23))); i23 += 8;
816         const int16x8_t vk23x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 728)), vkernel_zero_point));
817 
818         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi23x01234567), vget_low_s16(vk23x01234567));
819         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi23x01234567), vget_high_s16(vk23x01234567));
820         const int16x8_t vi24x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i24))); i24 += 8;
821         const int16x8_t vk24x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 760)), vkernel_zero_point));
822 
823         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi24x01234567), vget_low_s16(vk24x01234567));
824         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi24x01234567), vget_high_s16(vk24x01234567));
825 
826         vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
827         vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
828 
829         vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
830         vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
831 
832         vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
833         vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
834 
835 #if XNN_ARCH_ARM64
836         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
837 #else
838         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
839 #endif
840         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
841 
842         uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
843         vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
844         vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
845 
846         if XNN_LIKELY(c >= 8) {
847           vst1_u8(output, vout01234567); output += 8;
848           c -= 8;
849         } else {
850           if (c & 4) {
851             vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
852             vout01234567 = vext_u8(vout01234567, vout01234567, 4);
853           }
854           if (c & 2) {
855             vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
856             vout01234567 = vext_u8(vout01234567, vout01234567, 2);
857           }
858           if (c & 1) {
859             vst1_lane_u8(output, vout01234567, 0); output += 1;
860           }
861           c = 0;
862         }
863       } while (c != 0);
864     }
865 
866     output = (uint8_t*) ((uintptr_t) output + output_increment);
867   } while (--output_width != 0);
868 }
869