xref: /aosp_15_r20/external/XNNPACK/src/qu8-dwconv/gen/up16x9-minmax-fp32-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_fp32_ukernel_up16x9__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_fp32_ukernel_up16x9__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->fp32_neon.kernel_zero_point);
33   const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neon.scale);
34   const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
35   const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
36   const uint8x16_t voutput_min = vld1q_dup_u8(&params->fp32_neon.output_min);
37   const uint8x16_t voutput_max = vld1q_dup_u8(&params->fp32_neon.output_max);
38   do {
39     const uint8_t* i0 = input[0];
40     assert(i0 != NULL);
41     if XNN_UNPREDICTABLE(i0 != zero) {
42       i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
43     }
44     const uint8_t* i1 = input[1];
45     assert(i1 != NULL);
46     if XNN_UNPREDICTABLE(i1 != zero) {
47       i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
48     }
49     const uint8_t* i2 = input[2];
50     assert(i2 != NULL);
51     if XNN_UNPREDICTABLE(i2 != zero) {
52       i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
53     }
54     const uint8_t* i3 = input[3];
55     assert(i3 != NULL);
56     if XNN_UNPREDICTABLE(i3 != zero) {
57       i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
58     }
59     const uint8_t* i4 = input[4];
60     assert(i4 != NULL);
61     if XNN_UNPREDICTABLE(i4 != zero) {
62       i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
63     }
64     const uint8_t* i5 = input[5];
65     assert(i5 != NULL);
66     if XNN_UNPREDICTABLE(i5 != zero) {
67       i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
68     }
69     const uint8_t* i6 = input[6];
70     assert(i6 != NULL);
71     if XNN_UNPREDICTABLE(i6 != zero) {
72       i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
73     }
74     const uint8_t* i7 = input[7];
75     assert(i7 != NULL);
76     if XNN_UNPREDICTABLE(i7 != zero) {
77       i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
78     }
79     const uint8_t* i8 = input[8];
80     assert(i8 != NULL);
81     if XNN_UNPREDICTABLE(i8 != zero) {
82       i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
83     }
84     input = (const uint8_t**) ((uintptr_t) input + input_stride);
85 
86     size_t c = channels;
87     const void* w = weights;
88     for (; c >= 16; c -= 16) {
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 
94 
95       const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
96       const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
97       const int16x8_t vi0x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
98       const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
99 
100       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
101       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
102       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
103       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
104 
105       const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
106       const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
107       const int16x8_t vi1x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
108       const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
109 
110       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
111       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
112       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
113       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
114 
115       const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
116       const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
117       const int16x8_t vi2x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
118       const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
119 
120       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
121       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
122       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
123       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
124 
125       const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
126       const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
127       const int16x8_t vi3x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
128       const int16x8_t vk3x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
129 
130       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
131       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
132       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
133       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
134 
135       const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
136       const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
137       const int16x8_t vi4x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
138       const int16x8_t vk4x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
139 
140       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
141       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
142       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
143       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
144 
145       const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
146       const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
147       const int16x8_t vi5x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
148       const int16x8_t vk5x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
149 
150       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
151       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
152       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF));
153       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF));
154 
155       const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
156       const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
157       const int16x8_t vi6x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
158       const int16x8_t vk6x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
159 
160       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
161       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
162       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF));
163       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF));
164 
165       const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
166       const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
167       const int16x8_t vi7x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
168       const int16x8_t vk7x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
169 
170       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
171       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
172       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi7x89ABCDEF), vget_low_s16(vk7x89ABCDEF));
173       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi7x89ABCDEF), vget_high_s16(vk7x89ABCDEF));
174 
175       const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
176       const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
177       const int16x8_t vi8x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
178       const int16x8_t vk8x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
179 
180       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
181       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
182       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi8x89ABCDEF), vget_low_s16(vk8x89ABCDEF));
183       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi8x89ABCDEF), vget_high_s16(vk8x89ABCDEF));
184 
185       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
186       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
187       float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
188       float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
189 
190       vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
191       vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
192       vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
193       vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
194 
195       vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
196       vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
197       vacc89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc89AB, vmagic_bias));
198       vaccCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpaccCDEF, vmagic_bias));
199 
200       vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
201       vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
202       vacc89AB = vqsubq_s32(vacc89AB, vmagic_bias_less_output_zero_point);
203       vaccCDEF = vqsubq_s32(vaccCDEF, vmagic_bias_less_output_zero_point);
204 
205 #if XNN_ARCH_ARM64
206       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
207       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
208 
209 
210       uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
211 #else  // !XNN_ARCH_ARM64
212       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
213       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
214 
215 
216       uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
217 #endif  // !XNN_ARCH_ARM64
218 
219       vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
220 
221       vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
222 
223       vst1q_u8(output, vout0123456789ABCDEF); output += 16;
224     }
225     if XNN_UNLIKELY(c != 0) {
226       const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 16);
227       do {
228         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
229         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
230 
231         const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
232         const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(k), vkernel_zero_point)); k += 8;
233 
234         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
235         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
236         const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
237         const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 8)), vkernel_zero_point));
238 
239         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
240         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
241         const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
242         const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 24)), vkernel_zero_point));
243 
244         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
245         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
246         const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
247         const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 40)), vkernel_zero_point));
248 
249         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
250         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
251         const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
252         const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 56)), vkernel_zero_point));
253 
254         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
255         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
256         const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
257         const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 72)), vkernel_zero_point));
258 
259         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
260         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
261         const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
262         const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 88)), vkernel_zero_point));
263 
264         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
265         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
266         const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
267         const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 104)), vkernel_zero_point));
268 
269         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
270         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
271         const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
272         const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 120)), vkernel_zero_point));
273 
274         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
275         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
276 
277         float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
278         float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
279 
280         vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
281         vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
282 
283         vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
284         vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
285 
286         vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
287         vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
288 
289 #if XNN_ARCH_ARM64
290         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
291 #else
292         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
293 #endif
294 
295         uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
296         vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
297         vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
298 
299         if XNN_LIKELY(c >= 8) {
300           vst1_u8(output, vout01234567); output += 8;
301           c -= 8;
302         } else {
303           if (c & 4) {
304             vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
305             vout01234567 = vext_u8(vout01234567, vout01234567, 4);
306           }
307           if (c & 2) {
308             vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
309             vout01234567 = vext_u8(vout01234567, vout01234567, 2);
310           }
311           if (c & 1) {
312             vst1_lane_u8(output, vout01234567, 0); output += 1;
313           }
314           c = 0;
315         }
316       } while (c != 0);
317     }
318 
319     output = (uint8_t*) ((uintptr_t) output + output_increment);
320   } while (--output_width != 0);
321 }
322