xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up8x25-minmax-fp32-neon-mul8-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-neon-mul8.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_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul8_ld64(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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul8_ld64(
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_qc8_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 vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
33   const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
34   const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neon.output_min);
35   const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neon.output_max);
36   do {
37     const int8_t* i0 = input[0];
38     assert(i0 != NULL);
39     if XNN_UNPREDICTABLE(i0 != zero) {
40       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
41     }
42     const int8_t* i1 = input[1];
43     assert(i1 != NULL);
44     if XNN_UNPREDICTABLE(i1 != zero) {
45       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
46     }
47     const int8_t* i2 = input[2];
48     assert(i2 != NULL);
49     if XNN_UNPREDICTABLE(i2 != zero) {
50       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
51     }
52     const int8_t* i3 = input[3];
53     assert(i3 != NULL);
54     if XNN_UNPREDICTABLE(i3 != zero) {
55       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
56     }
57     const int8_t* i4 = input[4];
58     assert(i4 != NULL);
59     if XNN_UNPREDICTABLE(i4 != zero) {
60       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
61     }
62     const int8_t* i5 = input[5];
63     assert(i5 != NULL);
64     if XNN_UNPREDICTABLE(i5 != zero) {
65       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
66     }
67     const int8_t* i6 = input[6];
68     assert(i6 != NULL);
69     if XNN_UNPREDICTABLE(i6 != zero) {
70       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
71     }
72     const int8_t* i7 = input[7];
73     assert(i7 != NULL);
74     if XNN_UNPREDICTABLE(i7 != zero) {
75       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
76     }
77     const int8_t* i8 = input[8];
78     assert(i8 != NULL);
79     if XNN_UNPREDICTABLE(i8 != zero) {
80       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
81     }
82     const int8_t* i9 = input[9];
83     assert(i9 != NULL);
84     if XNN_UNPREDICTABLE(i9 != zero) {
85       i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
86     }
87     const int8_t* i10 = input[10];
88     assert(i10 != NULL);
89     if XNN_UNPREDICTABLE(i10 != zero) {
90       i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
91     }
92     const int8_t* i11 = input[11];
93     assert(i11 != NULL);
94     if XNN_UNPREDICTABLE(i11 != zero) {
95       i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
96     }
97     const int8_t* i12 = input[12];
98     assert(i12 != NULL);
99     if XNN_UNPREDICTABLE(i12 != zero) {
100       i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
101     }
102     const int8_t* i13 = input[13];
103     assert(i13 != NULL);
104     if XNN_UNPREDICTABLE(i13 != zero) {
105       i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
106     }
107     const int8_t* i14 = input[14];
108     assert(i14 != NULL);
109     if XNN_UNPREDICTABLE(i14 != zero) {
110       i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
111     }
112     const int8_t* i15 = input[15];
113     assert(i15 != NULL);
114     if XNN_UNPREDICTABLE(i15 != zero) {
115       i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
116     }
117     const int8_t* i16 = input[16];
118     assert(i16 != NULL);
119     if XNN_UNPREDICTABLE(i16 != zero) {
120       i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
121     }
122     const int8_t* i17 = input[17];
123     assert(i17 != NULL);
124     if XNN_UNPREDICTABLE(i17 != zero) {
125       i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
126     }
127     const int8_t* i18 = input[18];
128     assert(i18 != NULL);
129     if XNN_UNPREDICTABLE(i18 != zero) {
130       i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
131     }
132     const int8_t* i19 = input[19];
133     assert(i19 != NULL);
134     if XNN_UNPREDICTABLE(i19 != zero) {
135       i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
136     }
137     const int8_t* i20 = input[20];
138     assert(i20 != NULL);
139     if XNN_UNPREDICTABLE(i20 != zero) {
140       i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
141     }
142     const int8_t* i21 = input[21];
143     assert(i21 != NULL);
144     if XNN_UNPREDICTABLE(i21 != zero) {
145       i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
146     }
147     const int8_t* i22 = input[22];
148     assert(i22 != NULL);
149     if XNN_UNPREDICTABLE(i22 != zero) {
150       i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
151     }
152     const int8_t* i23 = input[23];
153     assert(i23 != NULL);
154     if XNN_UNPREDICTABLE(i23 != zero) {
155       i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
156     }
157     const int8_t* i24 = input[24];
158     assert(i24 != NULL);
159     if XNN_UNPREDICTABLE(i24 != zero) {
160       i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
161     }
162     input = (const int8_t**) ((uintptr_t) input + input_stride);
163 
164     size_t c = channels;
165     const void* w = weights;
166     for (; c >= 8; c -= 8) {
167       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
168       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
169 
170       const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
171       const int8x8_t vk0x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
172 
173       int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
174 
175       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
176       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
177       const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
178       const int8x8_t vk1x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
179 
180       vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
181 
182       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
183       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
184       const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
185       const int8x8_t vk2x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
186 
187       vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
188 
189       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
190       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
191       const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
192       const int8x8_t vk3x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
193 
194       vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
195 
196       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
197       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
198       const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
199       const int8x8_t vk4x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
200 
201       vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
202 
203       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
204       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
205       const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
206       const int8x8_t vk5x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
207 
208       vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
209 
210       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
211       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
212       const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
213       const int8x8_t vk6x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
214 
215       vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
216 
217       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
218       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
219       const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
220       const int8x8_t vk7x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
221 
222       vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
223 
224       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
225       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
226       const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
227       const int8x8_t vk8x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
228 
229       vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
230 
231       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
232       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
233       const int8x8_t vi9x01234567 = vld1_s8(i9); i9 += 8;
234       const int8x8_t vk9x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
235 
236       vprod01234567 = vmull_s8(vi9x01234567, vk9x01234567);
237 
238       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
239       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
240       const int8x8_t vi10x01234567 = vld1_s8(i10); i10 += 8;
241       const int8x8_t vk10x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
242 
243       vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
244 
245       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
246       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
247       const int8x8_t vi11x01234567 = vld1_s8(i11); i11 += 8;
248       const int8x8_t vk11x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
249 
250       vprod01234567 = vmull_s8(vi11x01234567, vk11x01234567);
251 
252       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
253       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
254       const int8x8_t vi12x01234567 = vld1_s8(i12); i12 += 8;
255       const int8x8_t vk12x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
256 
257       vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
258 
259       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
260       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
261       const int8x8_t vi13x01234567 = vld1_s8(i13); i13 += 8;
262       const int8x8_t vk13x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
263 
264       vprod01234567 = vmull_s8(vi13x01234567, vk13x01234567);
265 
266       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
267       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
268       const int8x8_t vi14x01234567 = vld1_s8(i14); i14 += 8;
269       const int8x8_t vk14x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
270 
271       vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
272 
273       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
274       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
275       const int8x8_t vi15x01234567 = vld1_s8(i15); i15 += 8;
276       const int8x8_t vk15x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
277 
278       vprod01234567 = vmull_s8(vi15x01234567, vk15x01234567);
279 
280       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
281       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
282       const int8x8_t vi16x01234567 = vld1_s8(i16); i16 += 8;
283       const int8x8_t vk16x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
284 
285       vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
286 
287       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
288       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
289       const int8x8_t vi17x01234567 = vld1_s8(i17); i17 += 8;
290       const int8x8_t vk17x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
291 
292       vprod01234567 = vmull_s8(vi17x01234567, vk17x01234567);
293 
294       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
295       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
296       const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8;
297       const int8x8_t vk18x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
298 
299       vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
300 
301       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
302       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
303       const int8x8_t vi19x01234567 = vld1_s8(i19); i19 += 8;
304       const int8x8_t vk19x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
305 
306       vprod01234567 = vmull_s8(vi19x01234567, vk19x01234567);
307 
308       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
309       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
310       const int8x8_t vi20x01234567 = vld1_s8(i20); i20 += 8;
311       const int8x8_t vk20x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
312 
313       vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
314 
315       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
316       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
317       const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8;
318       const int8x8_t vk21x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
319 
320       vprod01234567 = vmull_s8(vi21x01234567, vk21x01234567);
321 
322       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
323       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
324       const int8x8_t vi22x01234567 = vld1_s8(i22); i22 += 8;
325       const int8x8_t vk22x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
326 
327       vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
328 
329       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
330       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
331       const int8x8_t vi23x01234567 = vld1_s8(i23); i23 += 8;
332       const int8x8_t vk23x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
333 
334       vprod01234567 = vmull_s8(vi23x01234567, vk23x01234567);
335 
336       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
337       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
338       const int8x8_t vi24x01234567 = vld1_s8(i24); i24 += 8;
339       const int8x8_t vk24x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
340 
341       vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
342 
343       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
344       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
345 
346       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
347       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
348 
349       const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
350       const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
351 
352       vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
353       vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
354 
355       vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
356       vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
357 
358       vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
359       vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
360 
361 #if XNN_ARCH_ARM64
362       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
363 
364 
365       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
366 #else  // !XNN_ARCH_ARM64
367       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
368 
369 
370       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
371 #endif  // !XNN_ARCH_ARM64
372 
373       vout01234567 = vmax_s8(vout01234567, voutput_min);
374 
375       vout01234567 = vmin_s8(vout01234567, voutput_max);
376 
377       vst1_s8(output, vout01234567); output += 8;
378     }
379     if XNN_UNLIKELY(c != 0) {
380       {
381         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
382         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
383 
384         const int8x8_t vi0x01234567 = vld1_s8(i0);
385         const int8x8_t vk0x01234567 = vld1_s8(w);
386 
387         int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
388 
389         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
390         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
391         const int8x8_t vi1x01234567 = vld1_s8(i1);
392         const int8x8_t vk1x01234567 = vld1_s8((const void*) ((const int8_t*) w + 8));
393 
394         vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
395 
396         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
397         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
398         const int8x8_t vi2x01234567 = vld1_s8(i2);
399         const int8x8_t vk2x01234567 = vld1_s8((const void*) ((const int8_t*) w + 16));
400 
401         vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
402 
403         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
404         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
405         const int8x8_t vi3x01234567 = vld1_s8(i3);
406         const int8x8_t vk3x01234567 = vld1_s8((const void*) ((const int8_t*) w + 24));
407 
408         vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
409 
410         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
411         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
412         const int8x8_t vi4x01234567 = vld1_s8(i4);
413         const int8x8_t vk4x01234567 = vld1_s8((const void*) ((const int8_t*) w + 32));
414 
415         vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
416 
417         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
418         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
419         const int8x8_t vi5x01234567 = vld1_s8(i5);
420         const int8x8_t vk5x01234567 = vld1_s8((const void*) ((const int8_t*) w + 40));
421 
422         vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
423 
424         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
425         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
426         const int8x8_t vi6x01234567 = vld1_s8(i6);
427         const int8x8_t vk6x01234567 = vld1_s8((const void*) ((const int8_t*) w + 48));
428 
429         vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
430 
431         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
432         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
433         const int8x8_t vi7x01234567 = vld1_s8(i7);
434         const int8x8_t vk7x01234567 = vld1_s8((const void*) ((const int8_t*) w + 56));
435 
436         vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
437 
438         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
439         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
440         const int8x8_t vi8x01234567 = vld1_s8(i8);
441         const int8x8_t vk8x01234567 = vld1_s8((const void*) ((const int8_t*) w + 64));
442 
443         vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
444 
445         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
446         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
447         const int8x8_t vi9x01234567 = vld1_s8(i9);
448         const int8x8_t vk9x01234567 = vld1_s8((const void*) ((const int8_t*) w + 72));
449 
450         vprod01234567 = vmull_s8(vi9x01234567, vk9x01234567);
451 
452         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
453         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
454         const int8x8_t vi10x01234567 = vld1_s8(i10);
455         const int8x8_t vk10x01234567 = vld1_s8((const void*) ((const int8_t*) w + 80));
456 
457         vprod01234567 = vmull_s8(vi10x01234567, vk10x01234567);
458 
459         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
460         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
461         const int8x8_t vi11x01234567 = vld1_s8(i11);
462         const int8x8_t vk11x01234567 = vld1_s8((const void*) ((const int8_t*) w + 88));
463 
464         vprod01234567 = vmull_s8(vi11x01234567, vk11x01234567);
465 
466         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
467         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
468         const int8x8_t vi12x01234567 = vld1_s8(i12);
469         const int8x8_t vk12x01234567 = vld1_s8((const void*) ((const int8_t*) w + 96));
470 
471         vprod01234567 = vmull_s8(vi12x01234567, vk12x01234567);
472 
473         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
474         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
475         const int8x8_t vi13x01234567 = vld1_s8(i13);
476         const int8x8_t vk13x01234567 = vld1_s8((const void*) ((const int8_t*) w + 104));
477 
478         vprod01234567 = vmull_s8(vi13x01234567, vk13x01234567);
479 
480         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
481         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
482         const int8x8_t vi14x01234567 = vld1_s8(i14);
483         const int8x8_t vk14x01234567 = vld1_s8((const void*) ((const int8_t*) w + 112));
484 
485         vprod01234567 = vmull_s8(vi14x01234567, vk14x01234567);
486 
487         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
488         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
489         const int8x8_t vi15x01234567 = vld1_s8(i15);
490         const int8x8_t vk15x01234567 = vld1_s8((const void*) ((const int8_t*) w + 120));
491 
492         vprod01234567 = vmull_s8(vi15x01234567, vk15x01234567);
493 
494         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
495         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
496         const int8x8_t vi16x01234567 = vld1_s8(i16);
497         const int8x8_t vk16x01234567 = vld1_s8((const void*) ((const int8_t*) w + 128));
498 
499         vprod01234567 = vmull_s8(vi16x01234567, vk16x01234567);
500 
501         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
502         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
503         const int8x8_t vi17x01234567 = vld1_s8(i17);
504         const int8x8_t vk17x01234567 = vld1_s8((const void*) ((const int8_t*) w + 136));
505 
506         vprod01234567 = vmull_s8(vi17x01234567, vk17x01234567);
507 
508         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
509         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
510         const int8x8_t vi18x01234567 = vld1_s8(i18);
511         const int8x8_t vk18x01234567 = vld1_s8((const void*) ((const int8_t*) w + 144));
512 
513         vprod01234567 = vmull_s8(vi18x01234567, vk18x01234567);
514 
515         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
516         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
517         const int8x8_t vi19x01234567 = vld1_s8(i19);
518         const int8x8_t vk19x01234567 = vld1_s8((const void*) ((const int8_t*) w + 152));
519 
520         vprod01234567 = vmull_s8(vi19x01234567, vk19x01234567);
521 
522         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
523         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
524         const int8x8_t vi20x01234567 = vld1_s8(i20);
525         const int8x8_t vk20x01234567 = vld1_s8((const void*) ((const int8_t*) w + 160));
526 
527         vprod01234567 = vmull_s8(vi20x01234567, vk20x01234567);
528 
529         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
530         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
531         const int8x8_t vi21x01234567 = vld1_s8(i21);
532         const int8x8_t vk21x01234567 = vld1_s8((const void*) ((const int8_t*) w + 168));
533 
534         vprod01234567 = vmull_s8(vi21x01234567, vk21x01234567);
535 
536         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
537         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
538         const int8x8_t vi22x01234567 = vld1_s8(i22);
539         const int8x8_t vk22x01234567 = vld1_s8((const void*) ((const int8_t*) w + 176));
540 
541         vprod01234567 = vmull_s8(vi22x01234567, vk22x01234567);
542 
543         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
544         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
545         const int8x8_t vi23x01234567 = vld1_s8(i23);
546         const int8x8_t vk23x01234567 = vld1_s8((const void*) ((const int8_t*) w + 184));
547 
548         vprod01234567 = vmull_s8(vi23x01234567, vk23x01234567);
549 
550         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
551         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
552         const int8x8_t vi24x01234567 = vld1_s8(i24);
553         const int8x8_t vk24x01234567 = vld1_s8((const void*) ((const int8_t*) w + 192));
554 
555         vprod01234567 = vmull_s8(vi24x01234567, vk24x01234567);
556 
557         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
558         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
559 
560         float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
561         float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
562 
563         const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 0 * sizeof(int32_t) + 200 * sizeof(int8_t)));
564         const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 0 * sizeof(int32_t) + 200 * sizeof(int8_t) + 4 * sizeof(float)));
565         vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
566         vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
567 
568         vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
569         vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
570 
571         vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
572         vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
573 
574 #if XNN_ARCH_ARM64
575         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
576 #else
577         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
578 #endif
579 
580         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
581         vout01234567 = vmax_s8(vout01234567, voutput_min);
582         vout01234567 = vmin_s8(vout01234567, voutput_max);
583 
584         if (c & 4) {
585           vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
586           vout01234567 = vext_s8(vout01234567, vout01234567, 4);
587         }
588         if (c & 2) {
589           vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
590           vout01234567 = vext_s8(vout01234567, vout01234567, 2);
591         }
592         if (c & 1) {
593           vst1_lane_s8(output, vout01234567, 0); output += 1;
594         }
595       }
596     }
597 
598     output = (int8_t*) ((uintptr_t) output + output_increment);
599   } while (--output_width != 0);
600 }
601