xref: /aosp_15_r20/external/XNNPACK/src/qu8-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_qu8_dwconv_minmax_rndnu_ukernel_up32x9__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_up32x9__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     input = (const uint8_t**) ((uintptr_t) input + input_stride);
86 
87     size_t c = channels;
88     const void* w = weights;
89     for (; c >= 32; c -= 32) {
90       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
91       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
92       int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
93       int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
94       int32x4_t vaccGHIJ = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
95       int32x4_t vaccKLMN = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
96       int32x4_t vaccOPQR = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
97       int32x4_t vaccSTUV = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
98 
99 
100       const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
101       const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
102       const int16x8_t vi0x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
103       const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
104       const int16x8_t vi0xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
105       const int16x8_t vk0xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
106       const int16x8_t vi0xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
107       const int16x8_t vk0xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
108 
109       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
110       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
111       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
112       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
113       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi0xGHIJKLMN), vget_low_s16(vk0xGHIJKLMN));
114       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi0xGHIJKLMN), vget_high_s16(vk0xGHIJKLMN));
115       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi0xOPQRSTUV), vget_low_s16(vk0xOPQRSTUV));
116       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi0xOPQRSTUV), vget_high_s16(vk0xOPQRSTUV));
117 
118       const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
119       const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
120       const int16x8_t vi1x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
121       const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
122       const int16x8_t vi1xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
123       const int16x8_t vk1xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
124       const int16x8_t vi1xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
125       const int16x8_t vk1xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
126 
127       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
128       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
129       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
130       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
131       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi1xGHIJKLMN), vget_low_s16(vk1xGHIJKLMN));
132       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi1xGHIJKLMN), vget_high_s16(vk1xGHIJKLMN));
133       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi1xOPQRSTUV), vget_low_s16(vk1xOPQRSTUV));
134       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi1xOPQRSTUV), vget_high_s16(vk1xOPQRSTUV));
135 
136       const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
137       const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
138       const int16x8_t vi2x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
139       const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
140       const int16x8_t vi2xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
141       const int16x8_t vk2xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
142       const int16x8_t vi2xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
143       const int16x8_t vk2xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
144 
145       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
146       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
147       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
148       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
149       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi2xGHIJKLMN), vget_low_s16(vk2xGHIJKLMN));
150       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi2xGHIJKLMN), vget_high_s16(vk2xGHIJKLMN));
151       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi2xOPQRSTUV), vget_low_s16(vk2xOPQRSTUV));
152       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi2xOPQRSTUV), vget_high_s16(vk2xOPQRSTUV));
153 
154       const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
155       const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
156       const int16x8_t vi3x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
157       const int16x8_t vk3x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
158       const int16x8_t vi3xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
159       const int16x8_t vk3xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
160       const int16x8_t vi3xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
161       const int16x8_t vk3xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
162 
163       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
164       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
165       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
166       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
167       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN));
168       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN));
169       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi3xOPQRSTUV), vget_low_s16(vk3xOPQRSTUV));
170       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi3xOPQRSTUV), vget_high_s16(vk3xOPQRSTUV));
171 
172       const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
173       const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
174       const int16x8_t vi4x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
175       const int16x8_t vk4x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
176       const int16x8_t vi4xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
177       const int16x8_t vk4xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
178       const int16x8_t vi4xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
179       const int16x8_t vk4xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
180 
181       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
182       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
183       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
184       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
185       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi4xGHIJKLMN), vget_low_s16(vk4xGHIJKLMN));
186       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi4xGHIJKLMN), vget_high_s16(vk4xGHIJKLMN));
187       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV));
188       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV));
189 
190       const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
191       const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
192       const int16x8_t vi5x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
193       const int16x8_t vk5x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
194       const int16x8_t vi5xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
195       const int16x8_t vk5xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
196       const int16x8_t vi5xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
197       const int16x8_t vk5xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
198 
199       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
200       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
201       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF));
202       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF));
203       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi5xGHIJKLMN), vget_low_s16(vk5xGHIJKLMN));
204       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi5xGHIJKLMN), vget_high_s16(vk5xGHIJKLMN));
205       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi5xOPQRSTUV), vget_low_s16(vk5xOPQRSTUV));
206       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi5xOPQRSTUV), vget_high_s16(vk5xOPQRSTUV));
207 
208       const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
209       const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
210       const int16x8_t vi6x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
211       const int16x8_t vk6x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
212       const int16x8_t vi6xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
213       const int16x8_t vk6xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
214       const int16x8_t vi6xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
215       const int16x8_t vk6xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
216 
217       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
218       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
219       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF));
220       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF));
221       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi6xGHIJKLMN), vget_low_s16(vk6xGHIJKLMN));
222       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi6xGHIJKLMN), vget_high_s16(vk6xGHIJKLMN));
223       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi6xOPQRSTUV), vget_low_s16(vk6xOPQRSTUV));
224       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi6xOPQRSTUV), vget_high_s16(vk6xOPQRSTUV));
225 
226       const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
227       const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
228       const int16x8_t vi7x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
229       const int16x8_t vk7x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
230       const int16x8_t vi7xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
231       const int16x8_t vk7xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
232       const int16x8_t vi7xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
233       const int16x8_t vk7xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
234 
235       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
236       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
237       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi7x89ABCDEF), vget_low_s16(vk7x89ABCDEF));
238       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi7x89ABCDEF), vget_high_s16(vk7x89ABCDEF));
239       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN));
240       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN));
241       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi7xOPQRSTUV), vget_low_s16(vk7xOPQRSTUV));
242       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi7xOPQRSTUV), vget_high_s16(vk7xOPQRSTUV));
243 
244       const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
245       const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
246       const int16x8_t vi8x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
247       const int16x8_t vk8x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
248       const int16x8_t vi8xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
249       const int16x8_t vk8xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
250       const int16x8_t vi8xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
251       const int16x8_t vk8xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
252 
253       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
254       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
255       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi8x89ABCDEF), vget_low_s16(vk8x89ABCDEF));
256       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi8x89ABCDEF), vget_high_s16(vk8x89ABCDEF));
257       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi8xGHIJKLMN), vget_low_s16(vk8xGHIJKLMN));
258       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi8xGHIJKLMN), vget_high_s16(vk8xGHIJKLMN));
259       vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi8xOPQRSTUV), vget_low_s16(vk8xOPQRSTUV));
260       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi8xOPQRSTUV), vget_high_s16(vk8xOPQRSTUV));
261 
262       vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
263       vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
264       vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
265       vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
266       vaccGHIJ = vqshlq_s32(vaccGHIJ, vright_pre_shift);
267       vaccKLMN = vqshlq_s32(vaccKLMN, vright_pre_shift);
268       vaccOPQR = vqshlq_s32(vaccOPQR, vright_pre_shift);
269       vaccSTUV = vqshlq_s32(vaccSTUV, vright_pre_shift);
270 
271       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
272       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
273       vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
274       vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
275       vaccGHIJ = vqdmulhq_s32(vaccGHIJ, vmultiplier);
276       vaccKLMN = vqdmulhq_s32(vaccKLMN, vmultiplier);
277       vaccOPQR = vqdmulhq_s32(vaccOPQR, vmultiplier);
278       vaccSTUV = vqdmulhq_s32(vaccSTUV, vmultiplier);
279 
280       vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
281       vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
282       vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
283       vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
284       vaccGHIJ = vrshlq_s32(vaccGHIJ, vright_post_shift);
285       vaccKLMN = vrshlq_s32(vaccKLMN, vright_post_shift);
286       vaccOPQR = vrshlq_s32(vaccOPQR, vright_post_shift);
287       vaccSTUV = vrshlq_s32(vaccSTUV, vright_post_shift);
288 
289 #if XNN_ARCH_ARM64
290       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
291       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
292       int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);
293       int16x8_t vaccOPQRSTUV = vqmovn_high_s32(vqmovn_s32(vaccOPQR), vaccSTUV);
294 
295       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
296       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
297       vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
298       vaccOPQRSTUV = vqaddq_s16(vaccOPQRSTUV, voutput_zero_point);
299 
300       uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
301       uint8x16_t voutGHIJKLMNOPQRSTUV = vqmovun_high_s16(vqmovun_s16(vaccGHIJKLMN), vaccOPQRSTUV);
302 #else  // !XNN_ARCH_ARM64
303       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
304       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
305       int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));
306       int16x8_t vaccOPQRSTUV = vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV));
307 
308       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
309       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
310       vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
311       vaccOPQRSTUV = vqaddq_s16(vaccOPQRSTUV, voutput_zero_point);
312 
313       uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
314       uint8x16_t voutGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vaccGHIJKLMN), vqmovun_s16(vaccOPQRSTUV));
315 #endif  // !XNN_ARCH_ARM64
316 
317       vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
318       voutGHIJKLMNOPQRSTUV = vmaxq_u8(voutGHIJKLMNOPQRSTUV, voutput_min);
319 
320       vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
321       voutGHIJKLMNOPQRSTUV = vminq_u8(voutGHIJKLMNOPQRSTUV, voutput_max);
322 
323       vst1q_u8(output, vout0123456789ABCDEF); output += 16;
324       vst1q_u8(output, voutGHIJKLMNOPQRSTUV); output += 16;
325     }
326     if XNN_UNLIKELY(c != 0) {
327       const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 32);
328       do {
329         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
330         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
331 
332         const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
333         const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(k), vkernel_zero_point)); k += 8;
334 
335         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
336         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
337         const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
338         const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 24)), vkernel_zero_point));
339 
340         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
341         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
342         const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
343         const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 56)), vkernel_zero_point));
344 
345         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
346         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
347         const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
348         const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 88)), vkernel_zero_point));
349 
350         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
351         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
352         const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
353         const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 120)), vkernel_zero_point));
354 
355         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
356         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
357         const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
358         const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 152)), vkernel_zero_point));
359 
360         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
361         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
362         const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
363         const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 184)), vkernel_zero_point));
364 
365         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
366         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
367         const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
368         const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 216)), vkernel_zero_point));
369 
370         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
371         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
372         const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
373         const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 248)), vkernel_zero_point));
374 
375         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
376         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
377 
378         vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
379         vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
380 
381         vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
382         vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
383 
384         vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
385         vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
386 
387 #if XNN_ARCH_ARM64
388         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
389 #else
390         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
391 #endif
392         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
393 
394         uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
395         vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
396         vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
397 
398         if XNN_LIKELY(c >= 8) {
399           vst1_u8(output, vout01234567); output += 8;
400           c -= 8;
401         } else {
402           if (c & 4) {
403             vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
404             vout01234567 = vext_u8(vout01234567, vout01234567, 4);
405           }
406           if (c & 2) {
407             vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
408             vout01234567 = vext_u8(vout01234567, vout01234567, 2);
409           }
410           if (c & 1) {
411             vst1_lane_u8(output, vout01234567, 0); output += 1;
412           }
413           c = 0;
414         }
415       } while (c != 0);
416     }
417 
418     output = (uint8_t*) ((uintptr_t) output + output_increment);
419   } while (--output_width != 0);
420 }
421