xref: /aosp_15_r20/external/XNNPACK/src/f16-dwconv/gen/up32x25-minmax-neonfp16arith-acc2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-dwconv/up-neonfp16arith.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_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith_acc2(size_t channels,size_t output_width,const void ** input,const void * weights,void * output_ptr,size_t input_stride,size_t output_increment,size_t input_offset,const void * zero,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith_acc2(
18     size_t channels,
19     size_t output_width,
20     const void** input,
21     const void* weights,
22     void* output_ptr,
23     size_t input_stride,
24     size_t output_increment,
25     size_t input_offset,
26     const void* zero,
27     const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(channels != 0);
30   assert(output_width != 0);
31 
32   __fp16* output = (__fp16*) output_ptr;
33   const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.max));
34   const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.min));
35   do {
36     const __fp16* i0 = (const __fp16*) input[0];
37     assert(i0 != NULL);
38     if XNN_UNPREDICTABLE(i0 != (const __fp16*) zero) {
39       i0 = (const __fp16*) ((uintptr_t) i0 + input_offset);
40     }
41     const __fp16* i1 = (const __fp16*) input[1];
42     assert(i1 != NULL);
43     if XNN_UNPREDICTABLE(i1 != (const __fp16*) zero) {
44       i1 = (const __fp16*) ((uintptr_t) i1 + input_offset);
45     }
46     const __fp16* i2 = (const __fp16*) input[2];
47     assert(i2 != NULL);
48     if XNN_UNPREDICTABLE(i2 != (const __fp16*) zero) {
49       i2 = (const __fp16*) ((uintptr_t) i2 + input_offset);
50     }
51     const __fp16* i3 = (const __fp16*) input[3];
52     assert(i3 != NULL);
53     if XNN_UNPREDICTABLE(i3 != (const __fp16*) zero) {
54       i3 = (const __fp16*) ((uintptr_t) i3 + input_offset);
55     }
56     const __fp16* i4 = (const __fp16*) input[4];
57     assert(i4 != NULL);
58     if XNN_UNPREDICTABLE(i4 != (const __fp16*) zero) {
59       i4 = (const __fp16*) ((uintptr_t) i4 + input_offset);
60     }
61     const __fp16* i5 = (const __fp16*) input[5];
62     assert(i5 != NULL);
63     if XNN_UNPREDICTABLE(i5 != (const __fp16*) zero) {
64       i5 = (const __fp16*) ((uintptr_t) i5 + input_offset);
65     }
66     const __fp16* i6 = (const __fp16*) input[6];
67     assert(i6 != NULL);
68     if XNN_UNPREDICTABLE(i6 != (const __fp16*) zero) {
69       i6 = (const __fp16*) ((uintptr_t) i6 + input_offset);
70     }
71     const __fp16* i7 = (const __fp16*) input[7];
72     assert(i7 != NULL);
73     if XNN_UNPREDICTABLE(i7 != (const __fp16*) zero) {
74       i7 = (const __fp16*) ((uintptr_t) i7 + input_offset);
75     }
76     const __fp16* i8 = (const __fp16*) input[8];
77     assert(i8 != NULL);
78     if XNN_UNPREDICTABLE(i8 != (const __fp16*) zero) {
79       i8 = (const __fp16*) ((uintptr_t) i8 + input_offset);
80     }
81     const __fp16* i9 = (const __fp16*) input[9];
82     assert(i9 != NULL);
83     if XNN_UNPREDICTABLE(i9 != (const __fp16*) zero) {
84       i9 = (const __fp16*) ((uintptr_t) i9 + input_offset);
85     }
86     const __fp16* i10 = (const __fp16*) input[10];
87     assert(i10 != NULL);
88     if XNN_UNPREDICTABLE(i10 != (const __fp16*) zero) {
89       i10 = (const __fp16*) ((uintptr_t) i10 + input_offset);
90     }
91     const __fp16* i11 = (const __fp16*) input[11];
92     assert(i11 != NULL);
93     if XNN_UNPREDICTABLE(i11 != (const __fp16*) zero) {
94       i11 = (const __fp16*) ((uintptr_t) i11 + input_offset);
95     }
96     const __fp16* i12 = (const __fp16*) input[12];
97     assert(i12 != NULL);
98     if XNN_UNPREDICTABLE(i12 != (const __fp16*) zero) {
99       i12 = (const __fp16*) ((uintptr_t) i12 + input_offset);
100     }
101     const __fp16* i13 = (const __fp16*) input[13];
102     assert(i13 != NULL);
103     if XNN_UNPREDICTABLE(i13 != (const __fp16*) zero) {
104       i13 = (const __fp16*) ((uintptr_t) i13 + input_offset);
105     }
106     const __fp16* i14 = (const __fp16*) input[14];
107     assert(i14 != NULL);
108     if XNN_UNPREDICTABLE(i14 != (const __fp16*) zero) {
109       i14 = (const __fp16*) ((uintptr_t) i14 + input_offset);
110     }
111     const __fp16* i15 = (const __fp16*) input[15];
112     assert(i15 != NULL);
113     if XNN_UNPREDICTABLE(i15 != (const __fp16*) zero) {
114       i15 = (const __fp16*) ((uintptr_t) i15 + input_offset);
115     }
116     const __fp16* i16 = (const __fp16*) input[16];
117     assert(i16 != NULL);
118     if XNN_UNPREDICTABLE(i16 != (const __fp16*) zero) {
119       i16 = (const __fp16*) ((uintptr_t) i16 + input_offset);
120     }
121     const __fp16* i17 = (const __fp16*) input[17];
122     assert(i17 != NULL);
123     if XNN_UNPREDICTABLE(i17 != (const __fp16*) zero) {
124       i17 = (const __fp16*) ((uintptr_t) i17 + input_offset);
125     }
126     const __fp16* i18 = (const __fp16*) input[18];
127     assert(i18 != NULL);
128     if XNN_UNPREDICTABLE(i18 != (const __fp16*) zero) {
129       i18 = (const __fp16*) ((uintptr_t) i18 + input_offset);
130     }
131     const __fp16* i19 = (const __fp16*) input[19];
132     assert(i19 != NULL);
133     if XNN_UNPREDICTABLE(i19 != (const __fp16*) zero) {
134       i19 = (const __fp16*) ((uintptr_t) i19 + input_offset);
135     }
136     const __fp16* i20 = (const __fp16*) input[20];
137     assert(i20 != NULL);
138     if XNN_UNPREDICTABLE(i20 != (const __fp16*) zero) {
139       i20 = (const __fp16*) ((uintptr_t) i20 + input_offset);
140     }
141     const __fp16* i21 = (const __fp16*) input[21];
142     assert(i21 != NULL);
143     if XNN_UNPREDICTABLE(i21 != (const __fp16*) zero) {
144       i21 = (const __fp16*) ((uintptr_t) i21 + input_offset);
145     }
146     const __fp16* i22 = (const __fp16*) input[22];
147     assert(i22 != NULL);
148     if XNN_UNPREDICTABLE(i22 != (const __fp16*) zero) {
149       i22 = (const __fp16*) ((uintptr_t) i22 + input_offset);
150     }
151     const __fp16* i23 = (const __fp16*) input[23];
152     assert(i23 != NULL);
153     if XNN_UNPREDICTABLE(i23 != (const __fp16*) zero) {
154       i23 = (const __fp16*) ((uintptr_t) i23 + input_offset);
155     }
156     const __fp16* i24 = (const __fp16*) input[24];
157     assert(i24 != NULL);
158     if XNN_UNPREDICTABLE(i24 != (const __fp16*) zero) {
159       i24 = (const __fp16*) ((uintptr_t) i24 + input_offset);
160     }
161 
162     input = (const void**) ((uintptr_t) input + input_stride);
163 
164     size_t c = channels;
165     const __fp16* w = (const __fp16*) weights;
166     for (; c >= 32; c -= 32) {
167       float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
168       float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8;
169       float16x8_t vaccGHIJKLMNp0 = vld1q_f16(w); w += 8;
170       float16x8_t vaccOPQRSTUVp0 = vld1q_f16(w); w += 8;
171 
172 
173       const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
174       const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8;
175       const float16x8_t vi0xGHIJKLMN = vld1q_f16(i0); i0 += 8;
176       const float16x8_t vi0xOPQRSTUV = vld1q_f16(i0); i0 += 8;
177       const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8;
178       const float16x8_t vk0x89ABCDEF = vld1q_f16(w); w += 8;
179       const float16x8_t vk0xGHIJKLMN = vld1q_f16(w); w += 8;
180       const float16x8_t vk0xOPQRSTUV = vld1q_f16(w); w += 8;
181       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
182       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi0x89ABCDEF, vk0x89ABCDEF);
183       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi0xGHIJKLMN, vk0xGHIJKLMN);
184       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi0xOPQRSTUV, vk0xOPQRSTUV);
185 
186       const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
187       const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8;
188       const float16x8_t vi1xGHIJKLMN = vld1q_f16(i1); i1 += 8;
189       const float16x8_t vi1xOPQRSTUV = vld1q_f16(i1); i1 += 8;
190       const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8;
191       const float16x8_t vk1x89ABCDEF = vld1q_f16(w); w += 8;
192       const float16x8_t vk1xGHIJKLMN = vld1q_f16(w); w += 8;
193       const float16x8_t vk1xOPQRSTUV = vld1q_f16(w); w += 8;
194       float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
195       float16x8_t vacc89ABCDEFp1 = vmulq_f16(vi1x89ABCDEF, vk1x89ABCDEF);
196       float16x8_t vaccGHIJKLMNp1 = vmulq_f16(vi1xGHIJKLMN, vk1xGHIJKLMN);
197       float16x8_t vaccOPQRSTUVp1 = vmulq_f16(vi1xOPQRSTUV, vk1xOPQRSTUV);
198 
199       const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
200       const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8;
201       const float16x8_t vi2xGHIJKLMN = vld1q_f16(i2); i2 += 8;
202       const float16x8_t vi2xOPQRSTUV = vld1q_f16(i2); i2 += 8;
203       const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8;
204       const float16x8_t vk2x89ABCDEF = vld1q_f16(w); w += 8;
205       const float16x8_t vk2xGHIJKLMN = vld1q_f16(w); w += 8;
206       const float16x8_t vk2xOPQRSTUV = vld1q_f16(w); w += 8;
207       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
208       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF);
209       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi2xGHIJKLMN, vk2xGHIJKLMN);
210       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi2xOPQRSTUV, vk2xOPQRSTUV);
211 
212       const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
213       const float16x8_t vi3x89ABCDEF = vld1q_f16(i3); i3 += 8;
214       const float16x8_t vi3xGHIJKLMN = vld1q_f16(i3); i3 += 8;
215       const float16x8_t vi3xOPQRSTUV = vld1q_f16(i3); i3 += 8;
216       const float16x8_t vk3x01234567 = vld1q_f16(w); w += 8;
217       const float16x8_t vk3x89ABCDEF = vld1q_f16(w); w += 8;
218       const float16x8_t vk3xGHIJKLMN = vld1q_f16(w); w += 8;
219       const float16x8_t vk3xOPQRSTUV = vld1q_f16(w); w += 8;
220       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi3x01234567, vk3x01234567);
221       vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi3x89ABCDEF, vk3x89ABCDEF);
222       vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi3xGHIJKLMN, vk3xGHIJKLMN);
223       vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi3xOPQRSTUV, vk3xOPQRSTUV);
224 
225       const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
226       const float16x8_t vi4x89ABCDEF = vld1q_f16(i4); i4 += 8;
227       const float16x8_t vi4xGHIJKLMN = vld1q_f16(i4); i4 += 8;
228       const float16x8_t vi4xOPQRSTUV = vld1q_f16(i4); i4 += 8;
229       const float16x8_t vk4x01234567 = vld1q_f16(w); w += 8;
230       const float16x8_t vk4x89ABCDEF = vld1q_f16(w); w += 8;
231       const float16x8_t vk4xGHIJKLMN = vld1q_f16(w); w += 8;
232       const float16x8_t vk4xOPQRSTUV = vld1q_f16(w); w += 8;
233       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);
234       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi4x89ABCDEF, vk4x89ABCDEF);
235       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi4xGHIJKLMN, vk4xGHIJKLMN);
236       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi4xOPQRSTUV, vk4xOPQRSTUV);
237 
238       const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
239       const float16x8_t vi5x89ABCDEF = vld1q_f16(i5); i5 += 8;
240       const float16x8_t vi5xGHIJKLMN = vld1q_f16(i5); i5 += 8;
241       const float16x8_t vi5xOPQRSTUV = vld1q_f16(i5); i5 += 8;
242       const float16x8_t vk5x01234567 = vld1q_f16(w); w += 8;
243       const float16x8_t vk5x89ABCDEF = vld1q_f16(w); w += 8;
244       const float16x8_t vk5xGHIJKLMN = vld1q_f16(w); w += 8;
245       const float16x8_t vk5xOPQRSTUV = vld1q_f16(w); w += 8;
246       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi5x01234567, vk5x01234567);
247       vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi5x89ABCDEF, vk5x89ABCDEF);
248       vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi5xGHIJKLMN, vk5xGHIJKLMN);
249       vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi5xOPQRSTUV, vk5xOPQRSTUV);
250 
251       const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
252       const float16x8_t vi6x89ABCDEF = vld1q_f16(i6); i6 += 8;
253       const float16x8_t vi6xGHIJKLMN = vld1q_f16(i6); i6 += 8;
254       const float16x8_t vi6xOPQRSTUV = vld1q_f16(i6); i6 += 8;
255       const float16x8_t vk6x01234567 = vld1q_f16(w); w += 8;
256       const float16x8_t vk6x89ABCDEF = vld1q_f16(w); w += 8;
257       const float16x8_t vk6xGHIJKLMN = vld1q_f16(w); w += 8;
258       const float16x8_t vk6xOPQRSTUV = vld1q_f16(w); w += 8;
259       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);
260       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi6x89ABCDEF, vk6x89ABCDEF);
261       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi6xGHIJKLMN, vk6xGHIJKLMN);
262       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi6xOPQRSTUV, vk6xOPQRSTUV);
263 
264       const float16x8_t vi7x01234567 = vld1q_f16(i7); i7 += 8;
265       const float16x8_t vi7x89ABCDEF = vld1q_f16(i7); i7 += 8;
266       const float16x8_t vi7xGHIJKLMN = vld1q_f16(i7); i7 += 8;
267       const float16x8_t vi7xOPQRSTUV = vld1q_f16(i7); i7 += 8;
268       const float16x8_t vk7x01234567 = vld1q_f16(w); w += 8;
269       const float16x8_t vk7x89ABCDEF = vld1q_f16(w); w += 8;
270       const float16x8_t vk7xGHIJKLMN = vld1q_f16(w); w += 8;
271       const float16x8_t vk7xOPQRSTUV = vld1q_f16(w); w += 8;
272       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi7x01234567, vk7x01234567);
273       vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi7x89ABCDEF, vk7x89ABCDEF);
274       vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi7xGHIJKLMN, vk7xGHIJKLMN);
275       vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi7xOPQRSTUV, vk7xOPQRSTUV);
276 
277       const float16x8_t vi8x01234567 = vld1q_f16(i8); i8 += 8;
278       const float16x8_t vi8x89ABCDEF = vld1q_f16(i8); i8 += 8;
279       const float16x8_t vi8xGHIJKLMN = vld1q_f16(i8); i8 += 8;
280       const float16x8_t vi8xOPQRSTUV = vld1q_f16(i8); i8 += 8;
281       const float16x8_t vk8x01234567 = vld1q_f16(w); w += 8;
282       const float16x8_t vk8x89ABCDEF = vld1q_f16(w); w += 8;
283       const float16x8_t vk8xGHIJKLMN = vld1q_f16(w); w += 8;
284       const float16x8_t vk8xOPQRSTUV = vld1q_f16(w); w += 8;
285       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);
286       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi8x89ABCDEF, vk8x89ABCDEF);
287       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi8xGHIJKLMN, vk8xGHIJKLMN);
288       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi8xOPQRSTUV, vk8xOPQRSTUV);
289 
290       const float16x8_t vi9x01234567 = vld1q_f16(i9); i9 += 8;
291       const float16x8_t vi9x89ABCDEF = vld1q_f16(i9); i9 += 8;
292       const float16x8_t vi9xGHIJKLMN = vld1q_f16(i9); i9 += 8;
293       const float16x8_t vi9xOPQRSTUV = vld1q_f16(i9); i9 += 8;
294       const float16x8_t vk9x01234567 = vld1q_f16(w); w += 8;
295       const float16x8_t vk9x89ABCDEF = vld1q_f16(w); w += 8;
296       const float16x8_t vk9xGHIJKLMN = vld1q_f16(w); w += 8;
297       const float16x8_t vk9xOPQRSTUV = vld1q_f16(w); w += 8;
298       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi9x01234567, vk9x01234567);
299       vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi9x89ABCDEF, vk9x89ABCDEF);
300       vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi9xGHIJKLMN, vk9xGHIJKLMN);
301       vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi9xOPQRSTUV, vk9xOPQRSTUV);
302 
303       const float16x8_t vi10x01234567 = vld1q_f16(i10); i10 += 8;
304       const float16x8_t vi10x89ABCDEF = vld1q_f16(i10); i10 += 8;
305       const float16x8_t vi10xGHIJKLMN = vld1q_f16(i10); i10 += 8;
306       const float16x8_t vi10xOPQRSTUV = vld1q_f16(i10); i10 += 8;
307       const float16x8_t vk10x01234567 = vld1q_f16(w); w += 8;
308       const float16x8_t vk10x89ABCDEF = vld1q_f16(w); w += 8;
309       const float16x8_t vk10xGHIJKLMN = vld1q_f16(w); w += 8;
310       const float16x8_t vk10xOPQRSTUV = vld1q_f16(w); w += 8;
311       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567);
312       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi10x89ABCDEF, vk10x89ABCDEF);
313       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi10xGHIJKLMN, vk10xGHIJKLMN);
314       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi10xOPQRSTUV, vk10xOPQRSTUV);
315 
316       const float16x8_t vi11x01234567 = vld1q_f16(i11); i11 += 8;
317       const float16x8_t vi11x89ABCDEF = vld1q_f16(i11); i11 += 8;
318       const float16x8_t vi11xGHIJKLMN = vld1q_f16(i11); i11 += 8;
319       const float16x8_t vi11xOPQRSTUV = vld1q_f16(i11); i11 += 8;
320       const float16x8_t vk11x01234567 = vld1q_f16(w); w += 8;
321       const float16x8_t vk11x89ABCDEF = vld1q_f16(w); w += 8;
322       const float16x8_t vk11xGHIJKLMN = vld1q_f16(w); w += 8;
323       const float16x8_t vk11xOPQRSTUV = vld1q_f16(w); w += 8;
324       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi11x01234567, vk11x01234567);
325       vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi11x89ABCDEF, vk11x89ABCDEF);
326       vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi11xGHIJKLMN, vk11xGHIJKLMN);
327       vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi11xOPQRSTUV, vk11xOPQRSTUV);
328 
329       const float16x8_t vi12x01234567 = vld1q_f16(i12); i12 += 8;
330       const float16x8_t vi12x89ABCDEF = vld1q_f16(i12); i12 += 8;
331       const float16x8_t vi12xGHIJKLMN = vld1q_f16(i12); i12 += 8;
332       const float16x8_t vi12xOPQRSTUV = vld1q_f16(i12); i12 += 8;
333       const float16x8_t vk12x01234567 = vld1q_f16(w); w += 8;
334       const float16x8_t vk12x89ABCDEF = vld1q_f16(w); w += 8;
335       const float16x8_t vk12xGHIJKLMN = vld1q_f16(w); w += 8;
336       const float16x8_t vk12xOPQRSTUV = vld1q_f16(w); w += 8;
337       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567);
338       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi12x89ABCDEF, vk12x89ABCDEF);
339       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi12xGHIJKLMN, vk12xGHIJKLMN);
340       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi12xOPQRSTUV, vk12xOPQRSTUV);
341 
342       const float16x8_t vi13x01234567 = vld1q_f16(i13); i13 += 8;
343       const float16x8_t vi13x89ABCDEF = vld1q_f16(i13); i13 += 8;
344       const float16x8_t vi13xGHIJKLMN = vld1q_f16(i13); i13 += 8;
345       const float16x8_t vi13xOPQRSTUV = vld1q_f16(i13); i13 += 8;
346       const float16x8_t vk13x01234567 = vld1q_f16(w); w += 8;
347       const float16x8_t vk13x89ABCDEF = vld1q_f16(w); w += 8;
348       const float16x8_t vk13xGHIJKLMN = vld1q_f16(w); w += 8;
349       const float16x8_t vk13xOPQRSTUV = vld1q_f16(w); w += 8;
350       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi13x01234567, vk13x01234567);
351       vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi13x89ABCDEF, vk13x89ABCDEF);
352       vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi13xGHIJKLMN, vk13xGHIJKLMN);
353       vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi13xOPQRSTUV, vk13xOPQRSTUV);
354 
355       const float16x8_t vi14x01234567 = vld1q_f16(i14); i14 += 8;
356       const float16x8_t vi14x89ABCDEF = vld1q_f16(i14); i14 += 8;
357       const float16x8_t vi14xGHIJKLMN = vld1q_f16(i14); i14 += 8;
358       const float16x8_t vi14xOPQRSTUV = vld1q_f16(i14); i14 += 8;
359       const float16x8_t vk14x01234567 = vld1q_f16(w); w += 8;
360       const float16x8_t vk14x89ABCDEF = vld1q_f16(w); w += 8;
361       const float16x8_t vk14xGHIJKLMN = vld1q_f16(w); w += 8;
362       const float16x8_t vk14xOPQRSTUV = vld1q_f16(w); w += 8;
363       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567);
364       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi14x89ABCDEF, vk14x89ABCDEF);
365       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi14xGHIJKLMN, vk14xGHIJKLMN);
366       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi14xOPQRSTUV, vk14xOPQRSTUV);
367 
368       const float16x8_t vi15x01234567 = vld1q_f16(i15); i15 += 8;
369       const float16x8_t vi15x89ABCDEF = vld1q_f16(i15); i15 += 8;
370       const float16x8_t vi15xGHIJKLMN = vld1q_f16(i15); i15 += 8;
371       const float16x8_t vi15xOPQRSTUV = vld1q_f16(i15); i15 += 8;
372       const float16x8_t vk15x01234567 = vld1q_f16(w); w += 8;
373       const float16x8_t vk15x89ABCDEF = vld1q_f16(w); w += 8;
374       const float16x8_t vk15xGHIJKLMN = vld1q_f16(w); w += 8;
375       const float16x8_t vk15xOPQRSTUV = vld1q_f16(w); w += 8;
376       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi15x01234567, vk15x01234567);
377       vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi15x89ABCDEF, vk15x89ABCDEF);
378       vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi15xGHIJKLMN, vk15xGHIJKLMN);
379       vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi15xOPQRSTUV, vk15xOPQRSTUV);
380 
381       const float16x8_t vi16x01234567 = vld1q_f16(i16); i16 += 8;
382       const float16x8_t vi16x89ABCDEF = vld1q_f16(i16); i16 += 8;
383       const float16x8_t vi16xGHIJKLMN = vld1q_f16(i16); i16 += 8;
384       const float16x8_t vi16xOPQRSTUV = vld1q_f16(i16); i16 += 8;
385       const float16x8_t vk16x01234567 = vld1q_f16(w); w += 8;
386       const float16x8_t vk16x89ABCDEF = vld1q_f16(w); w += 8;
387       const float16x8_t vk16xGHIJKLMN = vld1q_f16(w); w += 8;
388       const float16x8_t vk16xOPQRSTUV = vld1q_f16(w); w += 8;
389       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567);
390       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi16x89ABCDEF, vk16x89ABCDEF);
391       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi16xGHIJKLMN, vk16xGHIJKLMN);
392       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi16xOPQRSTUV, vk16xOPQRSTUV);
393 
394       const float16x8_t vi17x01234567 = vld1q_f16(i17); i17 += 8;
395       const float16x8_t vi17x89ABCDEF = vld1q_f16(i17); i17 += 8;
396       const float16x8_t vi17xGHIJKLMN = vld1q_f16(i17); i17 += 8;
397       const float16x8_t vi17xOPQRSTUV = vld1q_f16(i17); i17 += 8;
398       const float16x8_t vk17x01234567 = vld1q_f16(w); w += 8;
399       const float16x8_t vk17x89ABCDEF = vld1q_f16(w); w += 8;
400       const float16x8_t vk17xGHIJKLMN = vld1q_f16(w); w += 8;
401       const float16x8_t vk17xOPQRSTUV = vld1q_f16(w); w += 8;
402       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi17x01234567, vk17x01234567);
403       vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi17x89ABCDEF, vk17x89ABCDEF);
404       vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi17xGHIJKLMN, vk17xGHIJKLMN);
405       vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi17xOPQRSTUV, vk17xOPQRSTUV);
406 
407       const float16x8_t vi18x01234567 = vld1q_f16(i18); i18 += 8;
408       const float16x8_t vi18x89ABCDEF = vld1q_f16(i18); i18 += 8;
409       const float16x8_t vi18xGHIJKLMN = vld1q_f16(i18); i18 += 8;
410       const float16x8_t vi18xOPQRSTUV = vld1q_f16(i18); i18 += 8;
411       const float16x8_t vk18x01234567 = vld1q_f16(w); w += 8;
412       const float16x8_t vk18x89ABCDEF = vld1q_f16(w); w += 8;
413       const float16x8_t vk18xGHIJKLMN = vld1q_f16(w); w += 8;
414       const float16x8_t vk18xOPQRSTUV = vld1q_f16(w); w += 8;
415       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567);
416       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi18x89ABCDEF, vk18x89ABCDEF);
417       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi18xGHIJKLMN, vk18xGHIJKLMN);
418       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi18xOPQRSTUV, vk18xOPQRSTUV);
419 
420       const float16x8_t vi19x01234567 = vld1q_f16(i19); i19 += 8;
421       const float16x8_t vi19x89ABCDEF = vld1q_f16(i19); i19 += 8;
422       const float16x8_t vi19xGHIJKLMN = vld1q_f16(i19); i19 += 8;
423       const float16x8_t vi19xOPQRSTUV = vld1q_f16(i19); i19 += 8;
424       const float16x8_t vk19x01234567 = vld1q_f16(w); w += 8;
425       const float16x8_t vk19x89ABCDEF = vld1q_f16(w); w += 8;
426       const float16x8_t vk19xGHIJKLMN = vld1q_f16(w); w += 8;
427       const float16x8_t vk19xOPQRSTUV = vld1q_f16(w); w += 8;
428       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi19x01234567, vk19x01234567);
429       vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi19x89ABCDEF, vk19x89ABCDEF);
430       vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi19xGHIJKLMN, vk19xGHIJKLMN);
431       vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi19xOPQRSTUV, vk19xOPQRSTUV);
432 
433       const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8;
434       const float16x8_t vi20x89ABCDEF = vld1q_f16(i20); i20 += 8;
435       const float16x8_t vi20xGHIJKLMN = vld1q_f16(i20); i20 += 8;
436       const float16x8_t vi20xOPQRSTUV = vld1q_f16(i20); i20 += 8;
437       const float16x8_t vk20x01234567 = vld1q_f16(w); w += 8;
438       const float16x8_t vk20x89ABCDEF = vld1q_f16(w); w += 8;
439       const float16x8_t vk20xGHIJKLMN = vld1q_f16(w); w += 8;
440       const float16x8_t vk20xOPQRSTUV = vld1q_f16(w); w += 8;
441       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567);
442       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi20x89ABCDEF, vk20x89ABCDEF);
443       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi20xGHIJKLMN, vk20xGHIJKLMN);
444       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi20xOPQRSTUV, vk20xOPQRSTUV);
445 
446       const float16x8_t vi21x01234567 = vld1q_f16(i21); i21 += 8;
447       const float16x8_t vi21x89ABCDEF = vld1q_f16(i21); i21 += 8;
448       const float16x8_t vi21xGHIJKLMN = vld1q_f16(i21); i21 += 8;
449       const float16x8_t vi21xOPQRSTUV = vld1q_f16(i21); i21 += 8;
450       const float16x8_t vk21x01234567 = vld1q_f16(w); w += 8;
451       const float16x8_t vk21x89ABCDEF = vld1q_f16(w); w += 8;
452       const float16x8_t vk21xGHIJKLMN = vld1q_f16(w); w += 8;
453       const float16x8_t vk21xOPQRSTUV = vld1q_f16(w); w += 8;
454       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi21x01234567, vk21x01234567);
455       vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi21x89ABCDEF, vk21x89ABCDEF);
456       vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi21xGHIJKLMN, vk21xGHIJKLMN);
457       vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi21xOPQRSTUV, vk21xOPQRSTUV);
458 
459       const float16x8_t vi22x01234567 = vld1q_f16(i22); i22 += 8;
460       const float16x8_t vi22x89ABCDEF = vld1q_f16(i22); i22 += 8;
461       const float16x8_t vi22xGHIJKLMN = vld1q_f16(i22); i22 += 8;
462       const float16x8_t vi22xOPQRSTUV = vld1q_f16(i22); i22 += 8;
463       const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8;
464       const float16x8_t vk22x89ABCDEF = vld1q_f16(w); w += 8;
465       const float16x8_t vk22xGHIJKLMN = vld1q_f16(w); w += 8;
466       const float16x8_t vk22xOPQRSTUV = vld1q_f16(w); w += 8;
467       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567);
468       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi22x89ABCDEF, vk22x89ABCDEF);
469       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi22xGHIJKLMN, vk22xGHIJKLMN);
470       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi22xOPQRSTUV, vk22xOPQRSTUV);
471 
472       const float16x8_t vi23x01234567 = vld1q_f16(i23); i23 += 8;
473       const float16x8_t vi23x89ABCDEF = vld1q_f16(i23); i23 += 8;
474       const float16x8_t vi23xGHIJKLMN = vld1q_f16(i23); i23 += 8;
475       const float16x8_t vi23xOPQRSTUV = vld1q_f16(i23); i23 += 8;
476       const float16x8_t vk23x01234567 = vld1q_f16(w); w += 8;
477       const float16x8_t vk23x89ABCDEF = vld1q_f16(w); w += 8;
478       const float16x8_t vk23xGHIJKLMN = vld1q_f16(w); w += 8;
479       const float16x8_t vk23xOPQRSTUV = vld1q_f16(w); w += 8;
480       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi23x01234567, vk23x01234567);
481       vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi23x89ABCDEF, vk23x89ABCDEF);
482       vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi23xGHIJKLMN, vk23xGHIJKLMN);
483       vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi23xOPQRSTUV, vk23xOPQRSTUV);
484 
485       const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8;
486       const float16x8_t vi24x89ABCDEF = vld1q_f16(i24); i24 += 8;
487       const float16x8_t vi24xGHIJKLMN = vld1q_f16(i24); i24 += 8;
488       const float16x8_t vi24xOPQRSTUV = vld1q_f16(i24); i24 += 8;
489       const float16x8_t vk24x01234567 = vld1q_f16(w); w += 8;
490       const float16x8_t vk24x89ABCDEF = vld1q_f16(w); w += 8;
491       const float16x8_t vk24xGHIJKLMN = vld1q_f16(w); w += 8;
492       const float16x8_t vk24xOPQRSTUV = vld1q_f16(w); w += 8;
493       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567);
494       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi24x89ABCDEF, vk24x89ABCDEF);
495       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi24xGHIJKLMN, vk24xGHIJKLMN);
496       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi24xOPQRSTUV, vk24xOPQRSTUV);
497 
498       // Add up all accumulators to vacc0123456789ABCDEFGHIJKLMNOPQRSTUVp0
499       vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
500       vacc89ABCDEFp0 = vaddq_f16(vacc89ABCDEFp0, vacc89ABCDEFp1);
501       vaccGHIJKLMNp0 = vaddq_f16(vaccGHIJKLMNp0, vaccGHIJKLMNp1);
502       vaccOPQRSTUVp0 = vaddq_f16(vaccOPQRSTUVp0, vaccOPQRSTUVp1);
503 
504       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
505       float16x8_t vacc89ABCDEF = vmaxq_f16(vacc89ABCDEFp0, vmin);
506       float16x8_t vaccGHIJKLMN = vmaxq_f16(vaccGHIJKLMNp0, vmin);
507       float16x8_t vaccOPQRSTUV = vmaxq_f16(vaccOPQRSTUVp0, vmin);
508       vacc01234567 = vminq_f16(vacc01234567, vmax);
509       vacc89ABCDEF = vminq_f16(vacc89ABCDEF, vmax);
510       vaccGHIJKLMN = vminq_f16(vaccGHIJKLMN, vmax);
511       vaccOPQRSTUV = vminq_f16(vaccOPQRSTUV, vmax);
512 
513       vst1q_f16(output, vacc01234567); output += 8;
514       vst1q_f16(output, vacc89ABCDEF); output += 8;
515       vst1q_f16(output, vaccGHIJKLMN); output += 8;
516       vst1q_f16(output, vaccOPQRSTUV); output += 8;
517     }
518     for (; c >= 8; c -= 8) {
519       float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
520 
521 
522       const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
523       const float16x8_t vk0x01234567 = vld1q_f16(w + 24);
524       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
525 
526       const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
527       const float16x8_t vk1x01234567 = vld1q_f16(w + 56);
528       float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
529 
530       const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
531       const float16x8_t vk2x01234567 = vld1q_f16(w + 88);
532       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
533 
534       const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
535       const float16x8_t vk3x01234567 = vld1q_f16(w + 120);
536       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi3x01234567, vk3x01234567);
537 
538       const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
539       const float16x8_t vk4x01234567 = vld1q_f16(w + 152);
540       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);
541 
542       const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
543       const float16x8_t vk5x01234567 = vld1q_f16(w + 184);
544       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi5x01234567, vk5x01234567);
545 
546       const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
547       const float16x8_t vk6x01234567 = vld1q_f16(w + 216);
548       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);
549 
550       const float16x8_t vi7x01234567 = vld1q_f16(i7); i7 += 8;
551       const float16x8_t vk7x01234567 = vld1q_f16(w + 248);
552       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi7x01234567, vk7x01234567);
553 
554       const float16x8_t vi8x01234567 = vld1q_f16(i8); i8 += 8;
555       const float16x8_t vk8x01234567 = vld1q_f16(w + 280);
556       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);
557 
558       const float16x8_t vi9x01234567 = vld1q_f16(i9); i9 += 8;
559       const float16x8_t vk9x01234567 = vld1q_f16(w + 312);
560       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi9x01234567, vk9x01234567);
561 
562       const float16x8_t vi10x01234567 = vld1q_f16(i10); i10 += 8;
563       const float16x8_t vk10x01234567 = vld1q_f16(w + 344);
564       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567);
565 
566       const float16x8_t vi11x01234567 = vld1q_f16(i11); i11 += 8;
567       const float16x8_t vk11x01234567 = vld1q_f16(w + 376);
568       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi11x01234567, vk11x01234567);
569 
570       const float16x8_t vi12x01234567 = vld1q_f16(i12); i12 += 8;
571       const float16x8_t vk12x01234567 = vld1q_f16(w + 408);
572       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567);
573 
574       const float16x8_t vi13x01234567 = vld1q_f16(i13); i13 += 8;
575       const float16x8_t vk13x01234567 = vld1q_f16(w + 440);
576       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi13x01234567, vk13x01234567);
577 
578       const float16x8_t vi14x01234567 = vld1q_f16(i14); i14 += 8;
579       const float16x8_t vk14x01234567 = vld1q_f16(w + 472);
580       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567);
581 
582       const float16x8_t vi15x01234567 = vld1q_f16(i15); i15 += 8;
583       const float16x8_t vk15x01234567 = vld1q_f16(w + 504);
584       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi15x01234567, vk15x01234567);
585 
586       const float16x8_t vi16x01234567 = vld1q_f16(i16); i16 += 8;
587       const float16x8_t vk16x01234567 = vld1q_f16(w + 536);
588       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567);
589 
590       const float16x8_t vi17x01234567 = vld1q_f16(i17); i17 += 8;
591       const float16x8_t vk17x01234567 = vld1q_f16(w + 568);
592       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi17x01234567, vk17x01234567);
593 
594       const float16x8_t vi18x01234567 = vld1q_f16(i18); i18 += 8;
595       const float16x8_t vk18x01234567 = vld1q_f16(w + 600);
596       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567);
597 
598       const float16x8_t vi19x01234567 = vld1q_f16(i19); i19 += 8;
599       const float16x8_t vk19x01234567 = vld1q_f16(w + 632);
600       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi19x01234567, vk19x01234567);
601 
602       const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8;
603       const float16x8_t vk20x01234567 = vld1q_f16(w + 664);
604       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567);
605 
606       const float16x8_t vi21x01234567 = vld1q_f16(i21); i21 += 8;
607       const float16x8_t vk21x01234567 = vld1q_f16(w + 696);
608       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi21x01234567, vk21x01234567);
609 
610       const float16x8_t vi22x01234567 = vld1q_f16(i22); i22 += 8;
611       const float16x8_t vk22x01234567 = vld1q_f16(w + 728);
612       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567);
613 
614       const float16x8_t vi23x01234567 = vld1q_f16(i23); i23 += 8;
615       const float16x8_t vk23x01234567 = vld1q_f16(w + 760);
616       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi23x01234567, vk23x01234567);
617 
618       const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8;
619       const float16x8_t vk24x01234567 = vld1q_f16(w + 792);
620       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567);
621 
622       // Add up all accumulators to vacc01234567p0
623       vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
624 
625       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
626       vacc01234567 = vminq_f16(vacc01234567, vmax);
627 
628       vst1q_f16(output, vacc01234567); output += 8;
629     }
630     if XNN_UNLIKELY(c != 0) {
631       float16x8_t vacc01234567p0 = vld1q_f16(w);
632 
633 
634       const float16x8_t vi0x01234567 = vld1q_f16(i0);
635       const float16x8_t vk0x01234567 = vld1q_f16(w + 32);
636       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
637 
638       const float16x8_t vi1x01234567 = vld1q_f16(i1);
639       const float16x8_t vk1x01234567 = vld1q_f16(w + 64);
640       float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
641 
642       const float16x8_t vi2x01234567 = vld1q_f16(i2);
643       const float16x8_t vk2x01234567 = vld1q_f16(w + 96);
644       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
645 
646       const float16x8_t vi3x01234567 = vld1q_f16(i3);
647       const float16x8_t vk3x01234567 = vld1q_f16(w + 128);
648       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi3x01234567, vk3x01234567);
649 
650       const float16x8_t vi4x01234567 = vld1q_f16(i4);
651       const float16x8_t vk4x01234567 = vld1q_f16(w + 160);
652       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);
653 
654       const float16x8_t vi5x01234567 = vld1q_f16(i5);
655       const float16x8_t vk5x01234567 = vld1q_f16(w + 192);
656       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi5x01234567, vk5x01234567);
657 
658       const float16x8_t vi6x01234567 = vld1q_f16(i6);
659       const float16x8_t vk6x01234567 = vld1q_f16(w + 224);
660       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);
661 
662       const float16x8_t vi7x01234567 = vld1q_f16(i7);
663       const float16x8_t vk7x01234567 = vld1q_f16(w + 256);
664       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi7x01234567, vk7x01234567);
665 
666       const float16x8_t vi8x01234567 = vld1q_f16(i8);
667       const float16x8_t vk8x01234567 = vld1q_f16(w + 288);
668       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);
669 
670       const float16x8_t vi9x01234567 = vld1q_f16(i9);
671       const float16x8_t vk9x01234567 = vld1q_f16(w + 320);
672       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi9x01234567, vk9x01234567);
673 
674       const float16x8_t vi10x01234567 = vld1q_f16(i10);
675       const float16x8_t vk10x01234567 = vld1q_f16(w + 352);
676       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567);
677 
678       const float16x8_t vi11x01234567 = vld1q_f16(i11);
679       const float16x8_t vk11x01234567 = vld1q_f16(w + 384);
680       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi11x01234567, vk11x01234567);
681 
682       const float16x8_t vi12x01234567 = vld1q_f16(i12);
683       const float16x8_t vk12x01234567 = vld1q_f16(w + 416);
684       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567);
685 
686       const float16x8_t vi13x01234567 = vld1q_f16(i13);
687       const float16x8_t vk13x01234567 = vld1q_f16(w + 448);
688       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi13x01234567, vk13x01234567);
689 
690       const float16x8_t vi14x01234567 = vld1q_f16(i14);
691       const float16x8_t vk14x01234567 = vld1q_f16(w + 480);
692       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567);
693 
694       const float16x8_t vi15x01234567 = vld1q_f16(i15);
695       const float16x8_t vk15x01234567 = vld1q_f16(w + 512);
696       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi15x01234567, vk15x01234567);
697 
698       const float16x8_t vi16x01234567 = vld1q_f16(i16);
699       const float16x8_t vk16x01234567 = vld1q_f16(w + 544);
700       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567);
701 
702       const float16x8_t vi17x01234567 = vld1q_f16(i17);
703       const float16x8_t vk17x01234567 = vld1q_f16(w + 576);
704       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi17x01234567, vk17x01234567);
705 
706       const float16x8_t vi18x01234567 = vld1q_f16(i18);
707       const float16x8_t vk18x01234567 = vld1q_f16(w + 608);
708       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567);
709 
710       const float16x8_t vi19x01234567 = vld1q_f16(i19);
711       const float16x8_t vk19x01234567 = vld1q_f16(w + 640);
712       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi19x01234567, vk19x01234567);
713 
714       const float16x8_t vi20x01234567 = vld1q_f16(i20);
715       const float16x8_t vk20x01234567 = vld1q_f16(w + 672);
716       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567);
717 
718       const float16x8_t vi21x01234567 = vld1q_f16(i21);
719       const float16x8_t vk21x01234567 = vld1q_f16(w + 704);
720       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi21x01234567, vk21x01234567);
721 
722       const float16x8_t vi22x01234567 = vld1q_f16(i22);
723       const float16x8_t vk22x01234567 = vld1q_f16(w + 736);
724       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567);
725 
726       const float16x8_t vi23x01234567 = vld1q_f16(i23);
727       const float16x8_t vk23x01234567 = vld1q_f16(w + 768);
728       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi23x01234567, vk23x01234567);
729 
730       const float16x8_t vi24x01234567 = vld1q_f16(i24);
731       const float16x8_t vk24x01234567 = vld1q_f16(w + 800);
732       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567);
733 
734       // Add up all accumulators to vacc01234567p0
735       vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
736 
737       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
738       vacc01234567 = vminq_f16(vacc01234567, vmax);
739 
740       float16x4_t vacc0123 = vget_low_f16(vacc01234567);
741       if (c & 4) {
742         vst1_f16(output, vacc0123); output += 4;
743         vacc0123 = vget_high_f16(vacc01234567);
744       }
745       if (c & 2) {
746         vst1_lane_u32((void*) output, vreinterpret_u32_f16(vacc0123), 0); output += 2;
747         vacc0123 = vext_f16(vacc0123, vacc0123, 2);
748       }
749       if (c & 1) {
750         vst1_lane_f16(output, vacc0123, 0); output += 1;
751       }
752     }
753 
754     output = (__fp16*) ((uintptr_t) output + output_increment);
755   } while (--output_width != 0);
756 }
757