xref: /aosp_15_r20/external/XNNPACK/src/qu8-dwconv/gen/up24x9-minmax-fp32-neonv8-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 #include <xnnpack/intrinsics-polyfill.h>
16 
17 
xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_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)])18 void xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16(
19     size_t channels,
20     size_t output_width,
21     const uint8_t** input,
22     const void* weights,
23     uint8_t* output,
24     size_t input_stride,
25     size_t output_increment,
26     size_t input_offset,
27     const uint8_t* zero,
28     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(channels != 0);
31   assert(output_width != 0);
32 
33   const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->fp32_neonv8.kernel_zero_point);
34   const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neonv8.scale);
35   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
36   const uint8x16_t voutput_min = vld1q_dup_u8(&params->fp32_neonv8.output_min);
37   const uint8x16_t voutput_max = vld1q_dup_u8(&params->fp32_neonv8.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 >= 24; c -= 24) {
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 
96 
97       const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
98       const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
99       const int16x8_t vi0x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
100       const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
101       const int16x8_t vi0xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
102       const int16x8_t vk0xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
103 
104       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
105       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
106       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
107       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
108       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi0xGHIJKLMN), vget_low_s16(vk0xGHIJKLMN));
109       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi0xGHIJKLMN), vget_high_s16(vk0xGHIJKLMN));
110 
111       const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
112       const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
113       const int16x8_t vi1x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
114       const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
115       const int16x8_t vi1xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
116       const int16x8_t vk1xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
117 
118       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
119       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
120       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
121       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
122       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi1xGHIJKLMN), vget_low_s16(vk1xGHIJKLMN));
123       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi1xGHIJKLMN), vget_high_s16(vk1xGHIJKLMN));
124 
125       const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
126       const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
127       const int16x8_t vi2x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
128       const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
129       const int16x8_t vi2xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
130       const int16x8_t vk2xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
131 
132       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
133       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
134       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
135       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
136       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi2xGHIJKLMN), vget_low_s16(vk2xGHIJKLMN));
137       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi2xGHIJKLMN), vget_high_s16(vk2xGHIJKLMN));
138 
139       const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
140       const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
141       const int16x8_t vi3x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
142       const int16x8_t vk3x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
143       const int16x8_t vi3xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
144       const int16x8_t vk3xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
145 
146       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
147       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
148       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
149       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
150       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN));
151       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN));
152 
153       const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
154       const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
155       const int16x8_t vi4x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
156       const int16x8_t vk4x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
157       const int16x8_t vi4xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
158       const int16x8_t vk4xGHIJKLMN = 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(vi4x01234567), vget_low_s16(vk4x01234567));
161       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
162       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
163       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
164       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi4xGHIJKLMN), vget_low_s16(vk4xGHIJKLMN));
165       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi4xGHIJKLMN), vget_high_s16(vk4xGHIJKLMN));
166 
167       const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
168       const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
169       const int16x8_t vi5x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
170       const int16x8_t vk5x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
171       const int16x8_t vi5xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
172       const int16x8_t vk5xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
173 
174       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
175       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
176       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF));
177       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF));
178       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi5xGHIJKLMN), vget_low_s16(vk5xGHIJKLMN));
179       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi5xGHIJKLMN), vget_high_s16(vk5xGHIJKLMN));
180 
181       const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
182       const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
183       const int16x8_t vi6x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
184       const int16x8_t vk6x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
185       const int16x8_t vi6xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
186       const int16x8_t vk6xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
187 
188       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
189       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
190       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF));
191       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF));
192       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi6xGHIJKLMN), vget_low_s16(vk6xGHIJKLMN));
193       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi6xGHIJKLMN), vget_high_s16(vk6xGHIJKLMN));
194 
195       const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
196       const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
197       const int16x8_t vi7x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
198       const int16x8_t vk7x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
199       const int16x8_t vi7xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
200       const int16x8_t vk7xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
201 
202       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
203       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
204       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi7x89ABCDEF), vget_low_s16(vk7x89ABCDEF));
205       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi7x89ABCDEF), vget_high_s16(vk7x89ABCDEF));
206       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN));
207       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN));
208 
209       const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
210       const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
211       const int16x8_t vi8x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
212       const int16x8_t vk8x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
213       const int16x8_t vi8xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
214       const int16x8_t vk8xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
215 
216       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
217       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
218       vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi8x89ABCDEF), vget_low_s16(vk8x89ABCDEF));
219       vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi8x89ABCDEF), vget_high_s16(vk8x89ABCDEF));
220       vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi8xGHIJKLMN), vget_low_s16(vk8xGHIJKLMN));
221       vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi8xGHIJKLMN), vget_high_s16(vk8xGHIJKLMN));
222 
223       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
224       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
225       float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
226       float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
227       float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ);
228       float32x4_t vfpaccKLMN = vcvtq_f32_s32(vaccKLMN);
229 
230       vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
231       vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
232       vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
233       vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
234       vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale);
235       vfpaccKLMN = vmulq_f32(vfpaccKLMN, vscale);
236 
237       vacc0123 = vcvtnq_s32_f32(vfpacc0123);
238       vacc4567 = vcvtnq_s32_f32(vfpacc4567);
239       vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
240       vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
241       vaccGHIJ = vcvtnq_s32_f32(vfpaccGHIJ);
242       vaccKLMN = vcvtnq_s32_f32(vfpaccKLMN);
243 
244 #if XNN_ARCH_ARM64
245       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
246       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
247       int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);
248 
249       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
250       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
251       vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
252 
253       uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
254       uint8x8_t voutGHIJKLMN = vqmovun_s16(vaccGHIJKLMN);
255 #else  // !XNN_ARCH_ARM64
256       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
257       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
258       int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));
259 
260       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
261       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
262       vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
263 
264       uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
265       uint8x8_t voutGHIJKLMN = vqmovun_s16(vaccGHIJKLMN);
266 #endif  // !XNN_ARCH_ARM64
267 
268       vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
269       voutGHIJKLMN = vmax_u8(voutGHIJKLMN, vget_low_u8(voutput_min));
270 
271       vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
272       voutGHIJKLMN = vmin_u8(voutGHIJKLMN, vget_low_u8(voutput_max));
273 
274       vst1q_u8(output, vout0123456789ABCDEF); output += 16;
275       vst1_u8(output, voutGHIJKLMN); output += 8;
276     }
277     if XNN_UNLIKELY(c != 0) {
278       const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 24);
279       do {
280         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
281         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
282 
283         const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
284         const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(k), vkernel_zero_point)); k += 8;
285 
286         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
287         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
288         const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
289         const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 16)), vkernel_zero_point));
290 
291         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
292         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
293         const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
294         const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 40)), vkernel_zero_point));
295 
296         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
297         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
298         const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
299         const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 64)), vkernel_zero_point));
300 
301         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
302         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
303         const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
304         const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 88)), vkernel_zero_point));
305 
306         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
307         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
308         const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
309         const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 112)), vkernel_zero_point));
310 
311         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
312         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
313         const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
314         const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 136)), vkernel_zero_point));
315 
316         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
317         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
318         const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
319         const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 160)), vkernel_zero_point));
320 
321         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
322         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
323         const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
324         const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 184)), vkernel_zero_point));
325 
326         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
327         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
328 
329         float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
330         float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
331 
332         vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
333         vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
334 
335         vacc0123 = vcvtnq_s32_f32(vfpacc0123);
336         vacc4567 = vcvtnq_s32_f32(vfpacc4567);
337 
338 #if XNN_ARCH_ARM64
339         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
340 #else
341         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
342 #endif
343         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
344 
345         uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
346         vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
347         vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
348 
349         if XNN_LIKELY(c >= 8) {
350           vst1_u8(output, vout01234567); output += 8;
351           c -= 8;
352         } else {
353           if (c & 4) {
354             vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
355             vout01234567 = vext_u8(vout01234567, vout01234567, 4);
356           }
357           if (c & 2) {
358             vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
359             vout01234567 = vext_u8(vout01234567, vout01234567, 2);
360           }
361           if (c & 1) {
362             vst1_lane_u8(output, vout01234567, 0); output += 1;
363           }
364           c = 0;
365         }
366       } while (c != 0);
367     }
368 
369     output = (uint8_t*) ((uintptr_t) output + output_increment);
370   } while (--output_width != 0);
371 }
372