xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/up32x9-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_qs8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16(
18     size_t channels,
19     size_t output_width,
20     const int8_t** input,
21     const void* weights,
22     int8_t* output,
23     size_t input_stride,
24     size_t output_increment,
25     size_t input_offset,
26     const int8_t* zero,
27     const union xnn_qs8_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 int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
33   const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
34   const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
35   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
36   const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
37   const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
38   do {
39     const int8_t* i0 = input[0];
40     assert(i0 != NULL);
41     if XNN_UNPREDICTABLE(i0 != zero) {
42       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
43     }
44     const int8_t* i1 = input[1];
45     assert(i1 != NULL);
46     if XNN_UNPREDICTABLE(i1 != zero) {
47       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
48     }
49     const int8_t* i2 = input[2];
50     assert(i2 != NULL);
51     if XNN_UNPREDICTABLE(i2 != zero) {
52       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
53     }
54     const int8_t* i3 = input[3];
55     assert(i3 != NULL);
56     if XNN_UNPREDICTABLE(i3 != zero) {
57       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
58     }
59     const int8_t* i4 = input[4];
60     assert(i4 != NULL);
61     if XNN_UNPREDICTABLE(i4 != zero) {
62       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
63     }
64     const int8_t* i5 = input[5];
65     assert(i5 != NULL);
66     if XNN_UNPREDICTABLE(i5 != zero) {
67       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
68     }
69     const int8_t* i6 = input[6];
70     assert(i6 != NULL);
71     if XNN_UNPREDICTABLE(i6 != zero) {
72       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
73     }
74     const int8_t* i7 = input[7];
75     assert(i7 != NULL);
76     if XNN_UNPREDICTABLE(i7 != zero) {
77       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
78     }
79     const int8_t* i8 = input[8];
80     assert(i8 != NULL);
81     if XNN_UNPREDICTABLE(i8 != zero) {
82       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
83     }
84     input = (const int8_t**) ((uintptr_t) input + input_stride);
85 
86     size_t c = channels;
87     const void* w = weights;
88     for (; c >= 32; c -= 32) {
89       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
90       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
91       int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
92       int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
93       int32x4_t vaccGHIJ = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
94       int32x4_t vaccKLMN = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
95       int32x4_t vaccOPQR = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
96       int32x4_t vaccSTUV = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
97 
98 
99       const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
100       const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
101       const int16x8_t vi0x89ABCDEF = vmovl_s8(vld1_s8(i0)); i0 += 8;
102       const int16x8_t vk0x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
103       const int16x8_t vi0xGHIJKLMN = vmovl_s8(vld1_s8(i0)); i0 += 8;
104       const int16x8_t vk0xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
105       const int16x8_t vi0xOPQRSTUV = vmovl_s8(vld1_s8(i0)); i0 += 8;
106       const int16x8_t vk0xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
107 
108       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
109       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
110       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
111       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
112       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi0xGHIJKLMN), vget_low_s16(vk0xGHIJKLMN));
113       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi0xGHIJKLMN), vget_high_s16(vk0xGHIJKLMN));
114       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi0xOPQRSTUV), vget_low_s16(vk0xOPQRSTUV));
115       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi0xOPQRSTUV), vget_high_s16(vk0xOPQRSTUV));
116 
117       const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
118       const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
119       const int16x8_t vi1x89ABCDEF = vmovl_s8(vld1_s8(i1)); i1 += 8;
120       const int16x8_t vk1x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
121       const int16x8_t vi1xGHIJKLMN = vmovl_s8(vld1_s8(i1)); i1 += 8;
122       const int16x8_t vk1xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
123       const int16x8_t vi1xOPQRSTUV = vmovl_s8(vld1_s8(i1)); i1 += 8;
124       const int16x8_t vk1xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
125 
126       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
127       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
128       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
129       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
130       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi1xGHIJKLMN), vget_low_s16(vk1xGHIJKLMN));
131       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi1xGHIJKLMN), vget_high_s16(vk1xGHIJKLMN));
132       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi1xOPQRSTUV), vget_low_s16(vk1xOPQRSTUV));
133       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi1xOPQRSTUV), vget_high_s16(vk1xOPQRSTUV));
134 
135       const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
136       const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
137       const int16x8_t vi2x89ABCDEF = vmovl_s8(vld1_s8(i2)); i2 += 8;
138       const int16x8_t vk2x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
139       const int16x8_t vi2xGHIJKLMN = vmovl_s8(vld1_s8(i2)); i2 += 8;
140       const int16x8_t vk2xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
141       const int16x8_t vi2xOPQRSTUV = vmovl_s8(vld1_s8(i2)); i2 += 8;
142       const int16x8_t vk2xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
143 
144       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
145       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
146       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
147       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
148       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi2xGHIJKLMN), vget_low_s16(vk2xGHIJKLMN));
149       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi2xGHIJKLMN), vget_high_s16(vk2xGHIJKLMN));
150       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi2xOPQRSTUV), vget_low_s16(vk2xOPQRSTUV));
151       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi2xOPQRSTUV), vget_high_s16(vk2xOPQRSTUV));
152 
153       const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
154       const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
155       const int16x8_t vi3x89ABCDEF = vmovl_s8(vld1_s8(i3)); i3 += 8;
156       const int16x8_t vk3x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
157       const int16x8_t vi3xGHIJKLMN = vmovl_s8(vld1_s8(i3)); i3 += 8;
158       const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
159       const int16x8_t vi3xOPQRSTUV = vmovl_s8(vld1_s8(i3)); i3 += 8;
160       const int16x8_t vk3xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
161 
162       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
163       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
164       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
165       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
166       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN));
167       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN));
168       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi3xOPQRSTUV), vget_low_s16(vk3xOPQRSTUV));
169       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi3xOPQRSTUV), vget_high_s16(vk3xOPQRSTUV));
170 
171       const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
172       const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
173       const int16x8_t vi4x89ABCDEF = vmovl_s8(vld1_s8(i4)); i4 += 8;
174       const int16x8_t vk4x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
175       const int16x8_t vi4xGHIJKLMN = vmovl_s8(vld1_s8(i4)); i4 += 8;
176       const int16x8_t vk4xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
177       const int16x8_t vi4xOPQRSTUV = vmovl_s8(vld1_s8(i4)); i4 += 8;
178       const int16x8_t vk4xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
179 
180       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
181       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
182       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
183       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
184       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi4xGHIJKLMN), vget_low_s16(vk4xGHIJKLMN));
185       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi4xGHIJKLMN), vget_high_s16(vk4xGHIJKLMN));
186       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV));
187       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV));
188 
189       const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
190       const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
191       const int16x8_t vi5x89ABCDEF = vmovl_s8(vld1_s8(i5)); i5 += 8;
192       const int16x8_t vk5x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
193       const int16x8_t vi5xGHIJKLMN = vmovl_s8(vld1_s8(i5)); i5 += 8;
194       const int16x8_t vk5xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
195       const int16x8_t vi5xOPQRSTUV = vmovl_s8(vld1_s8(i5)); i5 += 8;
196       const int16x8_t vk5xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
197 
198       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
199       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
200       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF));
201       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF));
202       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi5xGHIJKLMN), vget_low_s16(vk5xGHIJKLMN));
203       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi5xGHIJKLMN), vget_high_s16(vk5xGHIJKLMN));
204       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi5xOPQRSTUV), vget_low_s16(vk5xOPQRSTUV));
205       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi5xOPQRSTUV), vget_high_s16(vk5xOPQRSTUV));
206 
207       const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
208       const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
209       const int16x8_t vi6x89ABCDEF = vmovl_s8(vld1_s8(i6)); i6 += 8;
210       const int16x8_t vk6x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
211       const int16x8_t vi6xGHIJKLMN = vmovl_s8(vld1_s8(i6)); i6 += 8;
212       const int16x8_t vk6xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
213       const int16x8_t vi6xOPQRSTUV = vmovl_s8(vld1_s8(i6)); i6 += 8;
214       const int16x8_t vk6xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
215 
216       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
217       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
218       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF));
219       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF));
220       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi6xGHIJKLMN), vget_low_s16(vk6xGHIJKLMN));
221       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi6xGHIJKLMN), vget_high_s16(vk6xGHIJKLMN));
222       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi6xOPQRSTUV), vget_low_s16(vk6xOPQRSTUV));
223       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi6xOPQRSTUV), vget_high_s16(vk6xOPQRSTUV));
224 
225       const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
226       const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
227       const int16x8_t vi7x89ABCDEF = vmovl_s8(vld1_s8(i7)); i7 += 8;
228       const int16x8_t vk7x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
229       const int16x8_t vi7xGHIJKLMN = vmovl_s8(vld1_s8(i7)); i7 += 8;
230       const int16x8_t vk7xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
231       const int16x8_t vi7xOPQRSTUV = vmovl_s8(vld1_s8(i7)); i7 += 8;
232       const int16x8_t vk7xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
233 
234       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
235       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
236       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi7x89ABCDEF), vget_low_s16(vk7x89ABCDEF));
237       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi7x89ABCDEF), vget_high_s16(vk7x89ABCDEF));
238       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN));
239       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN));
240       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi7xOPQRSTUV), vget_low_s16(vk7xOPQRSTUV));
241       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi7xOPQRSTUV), vget_high_s16(vk7xOPQRSTUV));
242 
243       const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
244       const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
245       const int16x8_t vi8x89ABCDEF = vmovl_s8(vld1_s8(i8)); i8 += 8;
246       const int16x8_t vk8x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
247       const int16x8_t vi8xGHIJKLMN = vmovl_s8(vld1_s8(i8)); i8 += 8;
248       const int16x8_t vk8xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
249       const int16x8_t vi8xOPQRSTUV = vmovl_s8(vld1_s8(i8)); i8 += 8;
250       const int16x8_t vk8xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
251 
252       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
253       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
254       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi8x89ABCDEF), vget_low_s16(vk8x89ABCDEF));
255       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi8x89ABCDEF), vget_high_s16(vk8x89ABCDEF));
256       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi8xGHIJKLMN), vget_low_s16(vk8xGHIJKLMN));
257       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi8xGHIJKLMN), vget_high_s16(vk8xGHIJKLMN));
258       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi8xOPQRSTUV), vget_low_s16(vk8xOPQRSTUV));
259       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi8xOPQRSTUV), vget_high_s16(vk8xOPQRSTUV));
260 
261       vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
262       vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
263       vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
264       vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
265       vaccGHIJ = vqshlq_s32(vaccGHIJ, vright_pre_shift);
266       vaccKLMN = vqshlq_s32(vaccKLMN, vright_pre_shift);
267       vaccOPQR = vqshlq_s32(vaccOPQR, vright_pre_shift);
268       vaccSTUV = vqshlq_s32(vaccSTUV, vright_pre_shift);
269 
270       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
271       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
272       vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
273       vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
274       vaccGHIJ = vqdmulhq_s32(vaccGHIJ, vmultiplier);
275       vaccKLMN = vqdmulhq_s32(vaccKLMN, vmultiplier);
276       vaccOPQR = vqdmulhq_s32(vaccOPQR, vmultiplier);
277       vaccSTUV = vqdmulhq_s32(vaccSTUV, vmultiplier);
278 
279       vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
280       vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
281       vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
282       vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
283       vaccGHIJ = vrshlq_s32(vaccGHIJ, vright_post_shift);
284       vaccKLMN = vrshlq_s32(vaccKLMN, vright_post_shift);
285       vaccOPQR = vrshlq_s32(vaccOPQR, vright_post_shift);
286       vaccSTUV = vrshlq_s32(vaccSTUV, vright_post_shift);
287 
288 #if XNN_ARCH_ARM64
289       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
290       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
291       int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);
292       int16x8_t vaccOPQRSTUV = vqmovn_high_s32(vqmovn_s32(vaccOPQR), vaccSTUV);
293 
294       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
295       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
296       vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
297       vaccOPQRSTUV = vqaddq_s16(vaccOPQRSTUV, voutput_zero_point);
298 
299       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
300       int8x16_t voutGHIJKLMNOPQRSTUV = vqmovn_high_s16(vqmovn_s16(vaccGHIJKLMN), vaccOPQRSTUV);
301 #else  // !XNN_ARCH_ARM64
302       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
303       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
304       int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));
305       int16x8_t vaccOPQRSTUV = vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV));
306 
307       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
308       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
309       vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
310       vaccOPQRSTUV = vqaddq_s16(vaccOPQRSTUV, voutput_zero_point);
311 
312       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
313       int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV));
314 #endif  // !XNN_ARCH_ARM64
315 
316       vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
317       voutGHIJKLMNOPQRSTUV = vmaxq_s8(voutGHIJKLMNOPQRSTUV, voutput_min);
318 
319       vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
320       voutGHIJKLMNOPQRSTUV = vminq_s8(voutGHIJKLMNOPQRSTUV, voutput_max);
321 
322       vst1q_s8(output, vout0123456789ABCDEF); output += 16;
323       vst1q_s8(output, voutGHIJKLMNOPQRSTUV); output += 16;
324     }
325     if XNN_UNLIKELY(c != 0) {
326       const int8_t* k = (const int8_t*) ((const int32_t*) w + 32);
327       do {
328         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
329         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
330 
331         const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
332         const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(k)); k += 8;
333 
334         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
335         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
336         const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
337         const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8((const void*) (k + 24)));
338 
339         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
340         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
341         const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
342         const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8((const void*) (k + 56)));
343 
344         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
345         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
346         const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
347         const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8((const void*) (k + 88)));
348 
349         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
350         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
351         const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
352         const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8((const void*) (k + 120)));
353 
354         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
355         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
356         const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
357         const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8((const void*) (k + 152)));
358 
359         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
360         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
361         const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
362         const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8((const void*) (k + 184)));
363 
364         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
365         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
366         const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
367         const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8((const void*) (k + 216)));
368 
369         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
370         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
371         const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
372         const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8((const void*) (k + 248)));
373 
374         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
375         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
376 
377         vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
378         vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
379 
380         vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
381         vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
382 
383         vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
384         vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
385 
386 #if XNN_ARCH_ARM64
387         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
388 #else
389         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
390 #endif
391         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
392 
393         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
394         vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
395         vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
396 
397         if XNN_LIKELY(c >= 8) {
398           vst1_s8(output, vout01234567); output += 8;
399           c -= 8;
400         } else {
401           if (c & 4) {
402             vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
403             vout01234567 = vext_s8(vout01234567, vout01234567, 4);
404           }
405           if (c & 2) {
406             vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
407             vout01234567 = vext_s8(vout01234567, vout01234567, 2);
408           }
409           if (c & 1) {
410             vst1_lane_s8(output, vout01234567, 0); output += 1;
411           }
412           c = 0;
413         }
414       } while (c != 0);
415     }
416 
417     output = (int8_t*) ((uintptr_t) output + output_increment);
418   } while (--output_width != 0);
419 }
420