xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/up8x25-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_qs8_dwconv_minmax_fp32_ukernel_up8x25__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_fp32_ukernel_up8x25__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 float32x4_t vscale = vld1q_dup_f32(&params->fp32_neon.scale);
33   const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
34   const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
35   const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neon.output_min);
36   const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neon.output_max);
37   do {
38     const int8_t* i0 = input[0];
39     assert(i0 != NULL);
40     if XNN_UNPREDICTABLE(i0 != zero) {
41       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
42     }
43     const int8_t* i1 = input[1];
44     assert(i1 != NULL);
45     if XNN_UNPREDICTABLE(i1 != zero) {
46       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
47     }
48     const int8_t* i2 = input[2];
49     assert(i2 != NULL);
50     if XNN_UNPREDICTABLE(i2 != zero) {
51       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
52     }
53     const int8_t* i3 = input[3];
54     assert(i3 != NULL);
55     if XNN_UNPREDICTABLE(i3 != zero) {
56       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
57     }
58     const int8_t* i4 = input[4];
59     assert(i4 != NULL);
60     if XNN_UNPREDICTABLE(i4 != zero) {
61       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
62     }
63     const int8_t* i5 = input[5];
64     assert(i5 != NULL);
65     if XNN_UNPREDICTABLE(i5 != zero) {
66       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
67     }
68     const int8_t* i6 = input[6];
69     assert(i6 != NULL);
70     if XNN_UNPREDICTABLE(i6 != zero) {
71       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
72     }
73     const int8_t* i7 = input[7];
74     assert(i7 != NULL);
75     if XNN_UNPREDICTABLE(i7 != zero) {
76       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
77     }
78     const int8_t* i8 = input[8];
79     assert(i8 != NULL);
80     if XNN_UNPREDICTABLE(i8 != zero) {
81       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
82     }
83     const int8_t* i9 = input[9];
84     assert(i9 != NULL);
85     if XNN_UNPREDICTABLE(i9 != zero) {
86       i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
87     }
88     const int8_t* i10 = input[10];
89     assert(i10 != NULL);
90     if XNN_UNPREDICTABLE(i10 != zero) {
91       i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
92     }
93     const int8_t* i11 = input[11];
94     assert(i11 != NULL);
95     if XNN_UNPREDICTABLE(i11 != zero) {
96       i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
97     }
98     const int8_t* i12 = input[12];
99     assert(i12 != NULL);
100     if XNN_UNPREDICTABLE(i12 != zero) {
101       i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
102     }
103     const int8_t* i13 = input[13];
104     assert(i13 != NULL);
105     if XNN_UNPREDICTABLE(i13 != zero) {
106       i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
107     }
108     const int8_t* i14 = input[14];
109     assert(i14 != NULL);
110     if XNN_UNPREDICTABLE(i14 != zero) {
111       i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
112     }
113     const int8_t* i15 = input[15];
114     assert(i15 != NULL);
115     if XNN_UNPREDICTABLE(i15 != zero) {
116       i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
117     }
118     const int8_t* i16 = input[16];
119     assert(i16 != NULL);
120     if XNN_UNPREDICTABLE(i16 != zero) {
121       i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
122     }
123     const int8_t* i17 = input[17];
124     assert(i17 != NULL);
125     if XNN_UNPREDICTABLE(i17 != zero) {
126       i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
127     }
128     const int8_t* i18 = input[18];
129     assert(i18 != NULL);
130     if XNN_UNPREDICTABLE(i18 != zero) {
131       i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
132     }
133     const int8_t* i19 = input[19];
134     assert(i19 != NULL);
135     if XNN_UNPREDICTABLE(i19 != zero) {
136       i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
137     }
138     const int8_t* i20 = input[20];
139     assert(i20 != NULL);
140     if XNN_UNPREDICTABLE(i20 != zero) {
141       i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
142     }
143     const int8_t* i21 = input[21];
144     assert(i21 != NULL);
145     if XNN_UNPREDICTABLE(i21 != zero) {
146       i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
147     }
148     const int8_t* i22 = input[22];
149     assert(i22 != NULL);
150     if XNN_UNPREDICTABLE(i22 != zero) {
151       i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
152     }
153     const int8_t* i23 = input[23];
154     assert(i23 != NULL);
155     if XNN_UNPREDICTABLE(i23 != zero) {
156       i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
157     }
158     const int8_t* i24 = input[24];
159     assert(i24 != NULL);
160     if XNN_UNPREDICTABLE(i24 != zero) {
161       i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
162     }
163     input = (const int8_t**) ((uintptr_t) input + input_stride);
164 
165     size_t c = channels;
166     const void* w = weights;
167     for (; c >= 8; c -= 8) {
168       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
169       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
170 
171 
172       const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
173       const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
174 
175       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
176       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
177 
178       const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
179       const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
180 
181       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
182       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
183 
184       const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
185       const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
186 
187       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
188       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
189 
190       const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
191       const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
192 
193       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
194       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
195 
196       const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
197       const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
198 
199       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
200       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
201 
202       const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
203       const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
204 
205       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
206       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
207 
208       const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
209       const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
210 
211       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
212       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
213 
214       const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
215       const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
216 
217       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
218       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
219 
220       const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
221       const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
222 
223       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
224       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
225 
226       const int16x8_t vi9x01234567 = vmovl_s8(vld1_s8(i9)); i9 += 8;
227       const int16x8_t vk9x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
228 
229       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi9x01234567), vget_low_s16(vk9x01234567));
230       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi9x01234567), vget_high_s16(vk9x01234567));
231 
232       const int16x8_t vi10x01234567 = vmovl_s8(vld1_s8(i10)); i10 += 8;
233       const int16x8_t vk10x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
234 
235       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi10x01234567), vget_low_s16(vk10x01234567));
236       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi10x01234567), vget_high_s16(vk10x01234567));
237 
238       const int16x8_t vi11x01234567 = vmovl_s8(vld1_s8(i11)); i11 += 8;
239       const int16x8_t vk11x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
240 
241       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi11x01234567), vget_low_s16(vk11x01234567));
242       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi11x01234567), vget_high_s16(vk11x01234567));
243 
244       const int16x8_t vi12x01234567 = vmovl_s8(vld1_s8(i12)); i12 += 8;
245       const int16x8_t vk12x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
246 
247       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi12x01234567), vget_low_s16(vk12x01234567));
248       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi12x01234567), vget_high_s16(vk12x01234567));
249 
250       const int16x8_t vi13x01234567 = vmovl_s8(vld1_s8(i13)); i13 += 8;
251       const int16x8_t vk13x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
252 
253       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi13x01234567), vget_low_s16(vk13x01234567));
254       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi13x01234567), vget_high_s16(vk13x01234567));
255 
256       const int16x8_t vi14x01234567 = vmovl_s8(vld1_s8(i14)); i14 += 8;
257       const int16x8_t vk14x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
258 
259       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi14x01234567), vget_low_s16(vk14x01234567));
260       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi14x01234567), vget_high_s16(vk14x01234567));
261 
262       const int16x8_t vi15x01234567 = vmovl_s8(vld1_s8(i15)); i15 += 8;
263       const int16x8_t vk15x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
264 
265       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi15x01234567), vget_low_s16(vk15x01234567));
266       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi15x01234567), vget_high_s16(vk15x01234567));
267 
268       const int16x8_t vi16x01234567 = vmovl_s8(vld1_s8(i16)); i16 += 8;
269       const int16x8_t vk16x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
270 
271       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi16x01234567), vget_low_s16(vk16x01234567));
272       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi16x01234567), vget_high_s16(vk16x01234567));
273 
274       const int16x8_t vi17x01234567 = vmovl_s8(vld1_s8(i17)); i17 += 8;
275       const int16x8_t vk17x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
276 
277       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi17x01234567), vget_low_s16(vk17x01234567));
278       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi17x01234567), vget_high_s16(vk17x01234567));
279 
280       const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); i18 += 8;
281       const int16x8_t vk18x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
282 
283       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi18x01234567), vget_low_s16(vk18x01234567));
284       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi18x01234567), vget_high_s16(vk18x01234567));
285 
286       const int16x8_t vi19x01234567 = vmovl_s8(vld1_s8(i19)); i19 += 8;
287       const int16x8_t vk19x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
288 
289       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi19x01234567), vget_low_s16(vk19x01234567));
290       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi19x01234567), vget_high_s16(vk19x01234567));
291 
292       const int16x8_t vi20x01234567 = vmovl_s8(vld1_s8(i20)); i20 += 8;
293       const int16x8_t vk20x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
294 
295       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi20x01234567), vget_low_s16(vk20x01234567));
296       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi20x01234567), vget_high_s16(vk20x01234567));
297 
298       const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); i21 += 8;
299       const int16x8_t vk21x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
300 
301       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi21x01234567), vget_low_s16(vk21x01234567));
302       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi21x01234567), vget_high_s16(vk21x01234567));
303 
304       const int16x8_t vi22x01234567 = vmovl_s8(vld1_s8(i22)); i22 += 8;
305       const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
306 
307       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567));
308       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567));
309 
310       const int16x8_t vi23x01234567 = vmovl_s8(vld1_s8(i23)); i23 += 8;
311       const int16x8_t vk23x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
312 
313       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi23x01234567), vget_low_s16(vk23x01234567));
314       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi23x01234567), vget_high_s16(vk23x01234567));
315 
316       const int16x8_t vi24x01234567 = vmovl_s8(vld1_s8(i24)); i24 += 8;
317       const int16x8_t vk24x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
318 
319       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi24x01234567), vget_low_s16(vk24x01234567));
320       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi24x01234567), vget_high_s16(vk24x01234567));
321 
322       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
323       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
324 
325       vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
326       vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
327 
328       vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
329       vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
330 
331       vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
332       vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
333 
334 #if XNN_ARCH_ARM64
335       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
336 
337 
338       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
339 #else  // !XNN_ARCH_ARM64
340       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
341 
342 
343       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
344 #endif  // !XNN_ARCH_ARM64
345 
346       vout01234567 = vmax_s8(vout01234567, voutput_min);
347 
348       vout01234567 = vmin_s8(vout01234567, voutput_max);
349 
350       vst1_s8(output, vout01234567); output += 8;
351     }
352     if XNN_UNLIKELY(c != 0) {
353       {
354         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
355         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
356 
357         const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0));
358         const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w));
359 
360         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
361         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
362         const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1));
363         const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 8)));
364 
365         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
366         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
367         const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2));
368         const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 16)));
369 
370         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
371         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
372         const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3));
373         const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 24)));
374 
375         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
376         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
377         const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4));
378         const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 32)));
379 
380         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
381         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
382         const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5));
383         const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 40)));
384 
385         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
386         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
387         const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6));
388         const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 48)));
389 
390         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
391         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
392         const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7));
393         const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 56)));
394 
395         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
396         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
397         const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8));
398         const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 64)));
399 
400         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
401         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
402         const int16x8_t vi9x01234567 = vmovl_s8(vld1_s8(i9));
403         const int16x8_t vk9x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 72)));
404 
405         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi9x01234567), vget_low_s16(vk9x01234567));
406         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi9x01234567), vget_high_s16(vk9x01234567));
407         const int16x8_t vi10x01234567 = vmovl_s8(vld1_s8(i10));
408         const int16x8_t vk10x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 80)));
409 
410         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi10x01234567), vget_low_s16(vk10x01234567));
411         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi10x01234567), vget_high_s16(vk10x01234567));
412         const int16x8_t vi11x01234567 = vmovl_s8(vld1_s8(i11));
413         const int16x8_t vk11x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 88)));
414 
415         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi11x01234567), vget_low_s16(vk11x01234567));
416         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi11x01234567), vget_high_s16(vk11x01234567));
417         const int16x8_t vi12x01234567 = vmovl_s8(vld1_s8(i12));
418         const int16x8_t vk12x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 96)));
419 
420         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi12x01234567), vget_low_s16(vk12x01234567));
421         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi12x01234567), vget_high_s16(vk12x01234567));
422         const int16x8_t vi13x01234567 = vmovl_s8(vld1_s8(i13));
423         const int16x8_t vk13x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 104)));
424 
425         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi13x01234567), vget_low_s16(vk13x01234567));
426         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi13x01234567), vget_high_s16(vk13x01234567));
427         const int16x8_t vi14x01234567 = vmovl_s8(vld1_s8(i14));
428         const int16x8_t vk14x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 112)));
429 
430         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi14x01234567), vget_low_s16(vk14x01234567));
431         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi14x01234567), vget_high_s16(vk14x01234567));
432         const int16x8_t vi15x01234567 = vmovl_s8(vld1_s8(i15));
433         const int16x8_t vk15x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 120)));
434 
435         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi15x01234567), vget_low_s16(vk15x01234567));
436         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi15x01234567), vget_high_s16(vk15x01234567));
437         const int16x8_t vi16x01234567 = vmovl_s8(vld1_s8(i16));
438         const int16x8_t vk16x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 128)));
439 
440         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi16x01234567), vget_low_s16(vk16x01234567));
441         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi16x01234567), vget_high_s16(vk16x01234567));
442         const int16x8_t vi17x01234567 = vmovl_s8(vld1_s8(i17));
443         const int16x8_t vk17x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 136)));
444 
445         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi17x01234567), vget_low_s16(vk17x01234567));
446         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi17x01234567), vget_high_s16(vk17x01234567));
447         const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18));
448         const int16x8_t vk18x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 144)));
449 
450         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi18x01234567), vget_low_s16(vk18x01234567));
451         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi18x01234567), vget_high_s16(vk18x01234567));
452         const int16x8_t vi19x01234567 = vmovl_s8(vld1_s8(i19));
453         const int16x8_t vk19x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 152)));
454 
455         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi19x01234567), vget_low_s16(vk19x01234567));
456         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi19x01234567), vget_high_s16(vk19x01234567));
457         const int16x8_t vi20x01234567 = vmovl_s8(vld1_s8(i20));
458         const int16x8_t vk20x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 160)));
459 
460         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi20x01234567), vget_low_s16(vk20x01234567));
461         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi20x01234567), vget_high_s16(vk20x01234567));
462         const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21));
463         const int16x8_t vk21x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 168)));
464 
465         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi21x01234567), vget_low_s16(vk21x01234567));
466         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi21x01234567), vget_high_s16(vk21x01234567));
467         const int16x8_t vi22x01234567 = vmovl_s8(vld1_s8(i22));
468         const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 176)));
469 
470         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567));
471         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567));
472         const int16x8_t vi23x01234567 = vmovl_s8(vld1_s8(i23));
473         const int16x8_t vk23x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 184)));
474 
475         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi23x01234567), vget_low_s16(vk23x01234567));
476         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi23x01234567), vget_high_s16(vk23x01234567));
477         const int16x8_t vi24x01234567 = vmovl_s8(vld1_s8(i24));
478         const int16x8_t vk24x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 192)));
479 
480         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi24x01234567), vget_low_s16(vk24x01234567));
481         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi24x01234567), vget_high_s16(vk24x01234567));
482 
483         float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
484         float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
485 
486         vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
487         vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
488 
489         vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
490         vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
491 
492         vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
493         vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
494 
495 #if XNN_ARCH_ARM64
496         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
497 #else
498         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
499 #endif
500 
501         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
502         vout01234567 = vmax_s8(vout01234567, voutput_min);
503         vout01234567 = vmin_s8(vout01234567, voutput_max);
504 
505         if (c & 4) {
506           vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
507           vout01234567 = vext_s8(vout01234567, vout01234567, 4);
508         }
509         if (c & 2) {
510           vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
511           vout01234567 = vext_s8(vout01234567, vout01234567, 2);
512         }
513         if (c & 1) {
514           vst1_lane_s8(output, vout01234567, 0); output += 1;
515         }
516       }
517     }
518 
519     output = (int8_t*) ((uintptr_t) output + output_increment);
520   } while (--output_width != 0);
521 }
522