xref: /aosp_15_r20/external/XNNPACK/src/f16-dwconv/gen/up32x25-minmax-neonfp16arith.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f16-dwconv/up-neonfp16arith.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2020 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker #include <arm_neon.h>
13*4bdc9457SAndroid Build Coastguard Worker 
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/dwconv.h>
15*4bdc9457SAndroid Build Coastguard Worker 
16*4bdc9457SAndroid Build Coastguard Worker 
xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith(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*4bdc9457SAndroid Build Coastguard Worker void xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith(
18*4bdc9457SAndroid Build Coastguard Worker     size_t channels,
19*4bdc9457SAndroid Build Coastguard Worker     size_t output_width,
20*4bdc9457SAndroid Build Coastguard Worker     const void** input,
21*4bdc9457SAndroid Build Coastguard Worker     const void* weights,
22*4bdc9457SAndroid Build Coastguard Worker     void* output_ptr,
23*4bdc9457SAndroid Build Coastguard Worker     size_t input_stride,
24*4bdc9457SAndroid Build Coastguard Worker     size_t output_increment,
25*4bdc9457SAndroid Build Coastguard Worker     size_t input_offset,
26*4bdc9457SAndroid Build Coastguard Worker     const void* zero,
27*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28*4bdc9457SAndroid Build Coastguard Worker {
29*4bdc9457SAndroid Build Coastguard Worker   assert(channels != 0);
30*4bdc9457SAndroid Build Coastguard Worker   assert(output_width != 0);
31*4bdc9457SAndroid Build Coastguard Worker 
32*4bdc9457SAndroid Build Coastguard Worker   __fp16* output = (__fp16*) output_ptr;
33*4bdc9457SAndroid Build Coastguard Worker   const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.max));
34*4bdc9457SAndroid Build Coastguard Worker   const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.min));
35*4bdc9457SAndroid Build Coastguard Worker   do {
36*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i0 = (const __fp16*) input[0];
37*4bdc9457SAndroid Build Coastguard Worker     assert(i0 != NULL);
38*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i0 != (const __fp16*) zero) {
39*4bdc9457SAndroid Build Coastguard Worker       i0 = (const __fp16*) ((uintptr_t) i0 + input_offset);
40*4bdc9457SAndroid Build Coastguard Worker     }
41*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i1 = (const __fp16*) input[1];
42*4bdc9457SAndroid Build Coastguard Worker     assert(i1 != NULL);
43*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i1 != (const __fp16*) zero) {
44*4bdc9457SAndroid Build Coastguard Worker       i1 = (const __fp16*) ((uintptr_t) i1 + input_offset);
45*4bdc9457SAndroid Build Coastguard Worker     }
46*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i2 = (const __fp16*) input[2];
47*4bdc9457SAndroid Build Coastguard Worker     assert(i2 != NULL);
48*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i2 != (const __fp16*) zero) {
49*4bdc9457SAndroid Build Coastguard Worker       i2 = (const __fp16*) ((uintptr_t) i2 + input_offset);
50*4bdc9457SAndroid Build Coastguard Worker     }
51*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i3 = (const __fp16*) input[3];
52*4bdc9457SAndroid Build Coastguard Worker     assert(i3 != NULL);
53*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i3 != (const __fp16*) zero) {
54*4bdc9457SAndroid Build Coastguard Worker       i3 = (const __fp16*) ((uintptr_t) i3 + input_offset);
55*4bdc9457SAndroid Build Coastguard Worker     }
56*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i4 = (const __fp16*) input[4];
57*4bdc9457SAndroid Build Coastguard Worker     assert(i4 != NULL);
58*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i4 != (const __fp16*) zero) {
59*4bdc9457SAndroid Build Coastguard Worker       i4 = (const __fp16*) ((uintptr_t) i4 + input_offset);
60*4bdc9457SAndroid Build Coastguard Worker     }
61*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i5 = (const __fp16*) input[5];
62*4bdc9457SAndroid Build Coastguard Worker     assert(i5 != NULL);
63*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i5 != (const __fp16*) zero) {
64*4bdc9457SAndroid Build Coastguard Worker       i5 = (const __fp16*) ((uintptr_t) i5 + input_offset);
65*4bdc9457SAndroid Build Coastguard Worker     }
66*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i6 = (const __fp16*) input[6];
67*4bdc9457SAndroid Build Coastguard Worker     assert(i6 != NULL);
68*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i6 != (const __fp16*) zero) {
69*4bdc9457SAndroid Build Coastguard Worker       i6 = (const __fp16*) ((uintptr_t) i6 + input_offset);
70*4bdc9457SAndroid Build Coastguard Worker     }
71*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i7 = (const __fp16*) input[7];
72*4bdc9457SAndroid Build Coastguard Worker     assert(i7 != NULL);
73*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i7 != (const __fp16*) zero) {
74*4bdc9457SAndroid Build Coastguard Worker       i7 = (const __fp16*) ((uintptr_t) i7 + input_offset);
75*4bdc9457SAndroid Build Coastguard Worker     }
76*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i8 = (const __fp16*) input[8];
77*4bdc9457SAndroid Build Coastguard Worker     assert(i8 != NULL);
78*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i8 != (const __fp16*) zero) {
79*4bdc9457SAndroid Build Coastguard Worker       i8 = (const __fp16*) ((uintptr_t) i8 + input_offset);
80*4bdc9457SAndroid Build Coastguard Worker     }
81*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i9 = (const __fp16*) input[9];
82*4bdc9457SAndroid Build Coastguard Worker     assert(i9 != NULL);
83*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i9 != (const __fp16*) zero) {
84*4bdc9457SAndroid Build Coastguard Worker       i9 = (const __fp16*) ((uintptr_t) i9 + input_offset);
85*4bdc9457SAndroid Build Coastguard Worker     }
86*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i10 = (const __fp16*) input[10];
87*4bdc9457SAndroid Build Coastguard Worker     assert(i10 != NULL);
88*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i10 != (const __fp16*) zero) {
89*4bdc9457SAndroid Build Coastguard Worker       i10 = (const __fp16*) ((uintptr_t) i10 + input_offset);
90*4bdc9457SAndroid Build Coastguard Worker     }
91*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i11 = (const __fp16*) input[11];
92*4bdc9457SAndroid Build Coastguard Worker     assert(i11 != NULL);
93*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i11 != (const __fp16*) zero) {
94*4bdc9457SAndroid Build Coastguard Worker       i11 = (const __fp16*) ((uintptr_t) i11 + input_offset);
95*4bdc9457SAndroid Build Coastguard Worker     }
96*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i12 = (const __fp16*) input[12];
97*4bdc9457SAndroid Build Coastguard Worker     assert(i12 != NULL);
98*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i12 != (const __fp16*) zero) {
99*4bdc9457SAndroid Build Coastguard Worker       i12 = (const __fp16*) ((uintptr_t) i12 + input_offset);
100*4bdc9457SAndroid Build Coastguard Worker     }
101*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i13 = (const __fp16*) input[13];
102*4bdc9457SAndroid Build Coastguard Worker     assert(i13 != NULL);
103*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i13 != (const __fp16*) zero) {
104*4bdc9457SAndroid Build Coastguard Worker       i13 = (const __fp16*) ((uintptr_t) i13 + input_offset);
105*4bdc9457SAndroid Build Coastguard Worker     }
106*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i14 = (const __fp16*) input[14];
107*4bdc9457SAndroid Build Coastguard Worker     assert(i14 != NULL);
108*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i14 != (const __fp16*) zero) {
109*4bdc9457SAndroid Build Coastguard Worker       i14 = (const __fp16*) ((uintptr_t) i14 + input_offset);
110*4bdc9457SAndroid Build Coastguard Worker     }
111*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i15 = (const __fp16*) input[15];
112*4bdc9457SAndroid Build Coastguard Worker     assert(i15 != NULL);
113*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i15 != (const __fp16*) zero) {
114*4bdc9457SAndroid Build Coastguard Worker       i15 = (const __fp16*) ((uintptr_t) i15 + input_offset);
115*4bdc9457SAndroid Build Coastguard Worker     }
116*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i16 = (const __fp16*) input[16];
117*4bdc9457SAndroid Build Coastguard Worker     assert(i16 != NULL);
118*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i16 != (const __fp16*) zero) {
119*4bdc9457SAndroid Build Coastguard Worker       i16 = (const __fp16*) ((uintptr_t) i16 + input_offset);
120*4bdc9457SAndroid Build Coastguard Worker     }
121*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i17 = (const __fp16*) input[17];
122*4bdc9457SAndroid Build Coastguard Worker     assert(i17 != NULL);
123*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i17 != (const __fp16*) zero) {
124*4bdc9457SAndroid Build Coastguard Worker       i17 = (const __fp16*) ((uintptr_t) i17 + input_offset);
125*4bdc9457SAndroid Build Coastguard Worker     }
126*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i18 = (const __fp16*) input[18];
127*4bdc9457SAndroid Build Coastguard Worker     assert(i18 != NULL);
128*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i18 != (const __fp16*) zero) {
129*4bdc9457SAndroid Build Coastguard Worker       i18 = (const __fp16*) ((uintptr_t) i18 + input_offset);
130*4bdc9457SAndroid Build Coastguard Worker     }
131*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i19 = (const __fp16*) input[19];
132*4bdc9457SAndroid Build Coastguard Worker     assert(i19 != NULL);
133*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i19 != (const __fp16*) zero) {
134*4bdc9457SAndroid Build Coastguard Worker       i19 = (const __fp16*) ((uintptr_t) i19 + input_offset);
135*4bdc9457SAndroid Build Coastguard Worker     }
136*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i20 = (const __fp16*) input[20];
137*4bdc9457SAndroid Build Coastguard Worker     assert(i20 != NULL);
138*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i20 != (const __fp16*) zero) {
139*4bdc9457SAndroid Build Coastguard Worker       i20 = (const __fp16*) ((uintptr_t) i20 + input_offset);
140*4bdc9457SAndroid Build Coastguard Worker     }
141*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i21 = (const __fp16*) input[21];
142*4bdc9457SAndroid Build Coastguard Worker     assert(i21 != NULL);
143*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i21 != (const __fp16*) zero) {
144*4bdc9457SAndroid Build Coastguard Worker       i21 = (const __fp16*) ((uintptr_t) i21 + input_offset);
145*4bdc9457SAndroid Build Coastguard Worker     }
146*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i22 = (const __fp16*) input[22];
147*4bdc9457SAndroid Build Coastguard Worker     assert(i22 != NULL);
148*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i22 != (const __fp16*) zero) {
149*4bdc9457SAndroid Build Coastguard Worker       i22 = (const __fp16*) ((uintptr_t) i22 + input_offset);
150*4bdc9457SAndroid Build Coastguard Worker     }
151*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i23 = (const __fp16*) input[23];
152*4bdc9457SAndroid Build Coastguard Worker     assert(i23 != NULL);
153*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i23 != (const __fp16*) zero) {
154*4bdc9457SAndroid Build Coastguard Worker       i23 = (const __fp16*) ((uintptr_t) i23 + input_offset);
155*4bdc9457SAndroid Build Coastguard Worker     }
156*4bdc9457SAndroid Build Coastguard Worker     const __fp16* i24 = (const __fp16*) input[24];
157*4bdc9457SAndroid Build Coastguard Worker     assert(i24 != NULL);
158*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i24 != (const __fp16*) zero) {
159*4bdc9457SAndroid Build Coastguard Worker       i24 = (const __fp16*) ((uintptr_t) i24 + input_offset);
160*4bdc9457SAndroid Build Coastguard Worker     }
161*4bdc9457SAndroid Build Coastguard Worker 
162*4bdc9457SAndroid Build Coastguard Worker     input = (const void**) ((uintptr_t) input + input_stride);
163*4bdc9457SAndroid Build Coastguard Worker 
164*4bdc9457SAndroid Build Coastguard Worker     size_t c = channels;
165*4bdc9457SAndroid Build Coastguard Worker     const __fp16* w = (const __fp16*) weights;
166*4bdc9457SAndroid Build Coastguard Worker     for (; c >= 32; c -= 32) {
167*4bdc9457SAndroid Build Coastguard Worker       float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
168*4bdc9457SAndroid Build Coastguard Worker       float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8;
169*4bdc9457SAndroid Build Coastguard Worker       float16x8_t vaccGHIJKLMNp0 = vld1q_f16(w); w += 8;
170*4bdc9457SAndroid Build Coastguard Worker       float16x8_t vaccOPQRSTUVp0 = vld1q_f16(w); w += 8;
171*4bdc9457SAndroid Build Coastguard Worker 
172*4bdc9457SAndroid Build Coastguard Worker 
173*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
174*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8;
175*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi0xGHIJKLMN = vld1q_f16(i0); i0 += 8;
176*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi0xOPQRSTUV = vld1q_f16(i0); i0 += 8;
177*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8;
178*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk0x89ABCDEF = vld1q_f16(w); w += 8;
179*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk0xGHIJKLMN = vld1q_f16(w); w += 8;
180*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk0xOPQRSTUV = vld1q_f16(w); w += 8;
181*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
182*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi0x89ABCDEF, vk0x89ABCDEF);
183*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi0xGHIJKLMN, vk0xGHIJKLMN);
184*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi0xOPQRSTUV, vk0xOPQRSTUV);
185*4bdc9457SAndroid Build Coastguard Worker 
186*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
187*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8;
188*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi1xGHIJKLMN = vld1q_f16(i1); i1 += 8;
189*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi1xOPQRSTUV = vld1q_f16(i1); i1 += 8;
190*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8;
191*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk1x89ABCDEF = vld1q_f16(w); w += 8;
192*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk1xGHIJKLMN = vld1q_f16(w); w += 8;
193*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk1xOPQRSTUV = vld1q_f16(w); w += 8;
194*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);
195*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi1x89ABCDEF, vk1x89ABCDEF);
196*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi1xGHIJKLMN, vk1xGHIJKLMN);
197*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi1xOPQRSTUV, vk1xOPQRSTUV);
198*4bdc9457SAndroid Build Coastguard Worker 
199*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
200*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8;
201*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi2xGHIJKLMN = vld1q_f16(i2); i2 += 8;
202*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi2xOPQRSTUV = vld1q_f16(i2); i2 += 8;
203*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8;
204*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk2x89ABCDEF = vld1q_f16(w); w += 8;
205*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk2xGHIJKLMN = vld1q_f16(w); w += 8;
206*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk2xOPQRSTUV = vld1q_f16(w); w += 8;
207*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
208*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF);
209*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi2xGHIJKLMN, vk2xGHIJKLMN);
210*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi2xOPQRSTUV, vk2xOPQRSTUV);
211*4bdc9457SAndroid Build Coastguard Worker 
212*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
213*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi3x89ABCDEF = vld1q_f16(i3); i3 += 8;
214*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi3xGHIJKLMN = vld1q_f16(i3); i3 += 8;
215*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi3xOPQRSTUV = vld1q_f16(i3); i3 += 8;
216*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk3x01234567 = vld1q_f16(w); w += 8;
217*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk3x89ABCDEF = vld1q_f16(w); w += 8;
218*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk3xGHIJKLMN = vld1q_f16(w); w += 8;
219*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk3xOPQRSTUV = vld1q_f16(w); w += 8;
220*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi3x01234567, vk3x01234567);
221*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi3x89ABCDEF, vk3x89ABCDEF);
222*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi3xGHIJKLMN, vk3xGHIJKLMN);
223*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi3xOPQRSTUV, vk3xOPQRSTUV);
224*4bdc9457SAndroid Build Coastguard Worker 
225*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
226*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi4x89ABCDEF = vld1q_f16(i4); i4 += 8;
227*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi4xGHIJKLMN = vld1q_f16(i4); i4 += 8;
228*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi4xOPQRSTUV = vld1q_f16(i4); i4 += 8;
229*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk4x01234567 = vld1q_f16(w); w += 8;
230*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk4x89ABCDEF = vld1q_f16(w); w += 8;
231*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk4xGHIJKLMN = vld1q_f16(w); w += 8;
232*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk4xOPQRSTUV = vld1q_f16(w); w += 8;
233*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);
234*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi4x89ABCDEF, vk4x89ABCDEF);
235*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi4xGHIJKLMN, vk4xGHIJKLMN);
236*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi4xOPQRSTUV, vk4xOPQRSTUV);
237*4bdc9457SAndroid Build Coastguard Worker 
238*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
239*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi5x89ABCDEF = vld1q_f16(i5); i5 += 8;
240*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi5xGHIJKLMN = vld1q_f16(i5); i5 += 8;
241*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi5xOPQRSTUV = vld1q_f16(i5); i5 += 8;
242*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk5x01234567 = vld1q_f16(w); w += 8;
243*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk5x89ABCDEF = vld1q_f16(w); w += 8;
244*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk5xGHIJKLMN = vld1q_f16(w); w += 8;
245*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk5xOPQRSTUV = vld1q_f16(w); w += 8;
246*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi5x01234567, vk5x01234567);
247*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi5x89ABCDEF, vk5x89ABCDEF);
248*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi5xGHIJKLMN, vk5xGHIJKLMN);
249*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi5xOPQRSTUV, vk5xOPQRSTUV);
250*4bdc9457SAndroid Build Coastguard Worker 
251*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
252*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi6x89ABCDEF = vld1q_f16(i6); i6 += 8;
253*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi6xGHIJKLMN = vld1q_f16(i6); i6 += 8;
254*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi6xOPQRSTUV = vld1q_f16(i6); i6 += 8;
255*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk6x01234567 = vld1q_f16(w); w += 8;
256*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk6x89ABCDEF = vld1q_f16(w); w += 8;
257*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk6xGHIJKLMN = vld1q_f16(w); w += 8;
258*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk6xOPQRSTUV = vld1q_f16(w); w += 8;
259*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);
260*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi6x89ABCDEF, vk6x89ABCDEF);
261*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi6xGHIJKLMN, vk6xGHIJKLMN);
262*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi6xOPQRSTUV, vk6xOPQRSTUV);
263*4bdc9457SAndroid Build Coastguard Worker 
264*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi7x01234567 = vld1q_f16(i7); i7 += 8;
265*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi7x89ABCDEF = vld1q_f16(i7); i7 += 8;
266*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi7xGHIJKLMN = vld1q_f16(i7); i7 += 8;
267*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi7xOPQRSTUV = vld1q_f16(i7); i7 += 8;
268*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk7x01234567 = vld1q_f16(w); w += 8;
269*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk7x89ABCDEF = vld1q_f16(w); w += 8;
270*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk7xGHIJKLMN = vld1q_f16(w); w += 8;
271*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk7xOPQRSTUV = vld1q_f16(w); w += 8;
272*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi7x01234567, vk7x01234567);
273*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi7x89ABCDEF, vk7x89ABCDEF);
274*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi7xGHIJKLMN, vk7xGHIJKLMN);
275*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi7xOPQRSTUV, vk7xOPQRSTUV);
276*4bdc9457SAndroid Build Coastguard Worker 
277*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi8x01234567 = vld1q_f16(i8); i8 += 8;
278*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi8x89ABCDEF = vld1q_f16(i8); i8 += 8;
279*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi8xGHIJKLMN = vld1q_f16(i8); i8 += 8;
280*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi8xOPQRSTUV = vld1q_f16(i8); i8 += 8;
281*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk8x01234567 = vld1q_f16(w); w += 8;
282*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk8x89ABCDEF = vld1q_f16(w); w += 8;
283*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk8xGHIJKLMN = vld1q_f16(w); w += 8;
284*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk8xOPQRSTUV = vld1q_f16(w); w += 8;
285*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);
286*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi8x89ABCDEF, vk8x89ABCDEF);
287*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi8xGHIJKLMN, vk8xGHIJKLMN);
288*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi8xOPQRSTUV, vk8xOPQRSTUV);
289*4bdc9457SAndroid Build Coastguard Worker 
290*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi9x01234567 = vld1q_f16(i9); i9 += 8;
291*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi9x89ABCDEF = vld1q_f16(i9); i9 += 8;
292*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi9xGHIJKLMN = vld1q_f16(i9); i9 += 8;
293*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi9xOPQRSTUV = vld1q_f16(i9); i9 += 8;
294*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk9x01234567 = vld1q_f16(w); w += 8;
295*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk9x89ABCDEF = vld1q_f16(w); w += 8;
296*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk9xGHIJKLMN = vld1q_f16(w); w += 8;
297*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk9xOPQRSTUV = vld1q_f16(w); w += 8;
298*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi9x01234567, vk9x01234567);
299*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi9x89ABCDEF, vk9x89ABCDEF);
300*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi9xGHIJKLMN, vk9xGHIJKLMN);
301*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi9xOPQRSTUV, vk9xOPQRSTUV);
302*4bdc9457SAndroid Build Coastguard Worker 
303*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi10x01234567 = vld1q_f16(i10); i10 += 8;
304*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi10x89ABCDEF = vld1q_f16(i10); i10 += 8;
305*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi10xGHIJKLMN = vld1q_f16(i10); i10 += 8;
306*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi10xOPQRSTUV = vld1q_f16(i10); i10 += 8;
307*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk10x01234567 = vld1q_f16(w); w += 8;
308*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk10x89ABCDEF = vld1q_f16(w); w += 8;
309*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk10xGHIJKLMN = vld1q_f16(w); w += 8;
310*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk10xOPQRSTUV = vld1q_f16(w); w += 8;
311*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567);
312*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi10x89ABCDEF, vk10x89ABCDEF);
313*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi10xGHIJKLMN, vk10xGHIJKLMN);
314*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi10xOPQRSTUV, vk10xOPQRSTUV);
315*4bdc9457SAndroid Build Coastguard Worker 
316*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi11x01234567 = vld1q_f16(i11); i11 += 8;
317*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi11x89ABCDEF = vld1q_f16(i11); i11 += 8;
318*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi11xGHIJKLMN = vld1q_f16(i11); i11 += 8;
319*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi11xOPQRSTUV = vld1q_f16(i11); i11 += 8;
320*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk11x01234567 = vld1q_f16(w); w += 8;
321*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk11x89ABCDEF = vld1q_f16(w); w += 8;
322*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk11xGHIJKLMN = vld1q_f16(w); w += 8;
323*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk11xOPQRSTUV = vld1q_f16(w); w += 8;
324*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi11x01234567, vk11x01234567);
325*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi11x89ABCDEF, vk11x89ABCDEF);
326*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi11xGHIJKLMN, vk11xGHIJKLMN);
327*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi11xOPQRSTUV, vk11xOPQRSTUV);
328*4bdc9457SAndroid Build Coastguard Worker 
329*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi12x01234567 = vld1q_f16(i12); i12 += 8;
330*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi12x89ABCDEF = vld1q_f16(i12); i12 += 8;
331*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi12xGHIJKLMN = vld1q_f16(i12); i12 += 8;
332*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi12xOPQRSTUV = vld1q_f16(i12); i12 += 8;
333*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk12x01234567 = vld1q_f16(w); w += 8;
334*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk12x89ABCDEF = vld1q_f16(w); w += 8;
335*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk12xGHIJKLMN = vld1q_f16(w); w += 8;
336*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk12xOPQRSTUV = vld1q_f16(w); w += 8;
337*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567);
338*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi12x89ABCDEF, vk12x89ABCDEF);
339*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi12xGHIJKLMN, vk12xGHIJKLMN);
340*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi12xOPQRSTUV, vk12xOPQRSTUV);
341*4bdc9457SAndroid Build Coastguard Worker 
342*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi13x01234567 = vld1q_f16(i13); i13 += 8;
343*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi13x89ABCDEF = vld1q_f16(i13); i13 += 8;
344*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi13xGHIJKLMN = vld1q_f16(i13); i13 += 8;
345*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi13xOPQRSTUV = vld1q_f16(i13); i13 += 8;
346*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk13x01234567 = vld1q_f16(w); w += 8;
347*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk13x89ABCDEF = vld1q_f16(w); w += 8;
348*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk13xGHIJKLMN = vld1q_f16(w); w += 8;
349*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk13xOPQRSTUV = vld1q_f16(w); w += 8;
350*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi13x01234567, vk13x01234567);
351*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi13x89ABCDEF, vk13x89ABCDEF);
352*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi13xGHIJKLMN, vk13xGHIJKLMN);
353*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi13xOPQRSTUV, vk13xOPQRSTUV);
354*4bdc9457SAndroid Build Coastguard Worker 
355*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi14x01234567 = vld1q_f16(i14); i14 += 8;
356*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi14x89ABCDEF = vld1q_f16(i14); i14 += 8;
357*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi14xGHIJKLMN = vld1q_f16(i14); i14 += 8;
358*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi14xOPQRSTUV = vld1q_f16(i14); i14 += 8;
359*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk14x01234567 = vld1q_f16(w); w += 8;
360*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk14x89ABCDEF = vld1q_f16(w); w += 8;
361*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk14xGHIJKLMN = vld1q_f16(w); w += 8;
362*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk14xOPQRSTUV = vld1q_f16(w); w += 8;
363*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567);
364*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi14x89ABCDEF, vk14x89ABCDEF);
365*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi14xGHIJKLMN, vk14xGHIJKLMN);
366*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi14xOPQRSTUV, vk14xOPQRSTUV);
367*4bdc9457SAndroid Build Coastguard Worker 
368*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi15x01234567 = vld1q_f16(i15); i15 += 8;
369*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi15x89ABCDEF = vld1q_f16(i15); i15 += 8;
370*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi15xGHIJKLMN = vld1q_f16(i15); i15 += 8;
371*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi15xOPQRSTUV = vld1q_f16(i15); i15 += 8;
372*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk15x01234567 = vld1q_f16(w); w += 8;
373*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk15x89ABCDEF = vld1q_f16(w); w += 8;
374*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk15xGHIJKLMN = vld1q_f16(w); w += 8;
375*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk15xOPQRSTUV = vld1q_f16(w); w += 8;
376*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi15x01234567, vk15x01234567);
377*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi15x89ABCDEF, vk15x89ABCDEF);
378*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi15xGHIJKLMN, vk15xGHIJKLMN);
379*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi15xOPQRSTUV, vk15xOPQRSTUV);
380*4bdc9457SAndroid Build Coastguard Worker 
381*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi16x01234567 = vld1q_f16(i16); i16 += 8;
382*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi16x89ABCDEF = vld1q_f16(i16); i16 += 8;
383*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi16xGHIJKLMN = vld1q_f16(i16); i16 += 8;
384*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi16xOPQRSTUV = vld1q_f16(i16); i16 += 8;
385*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk16x01234567 = vld1q_f16(w); w += 8;
386*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk16x89ABCDEF = vld1q_f16(w); w += 8;
387*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk16xGHIJKLMN = vld1q_f16(w); w += 8;
388*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk16xOPQRSTUV = vld1q_f16(w); w += 8;
389*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567);
390*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi16x89ABCDEF, vk16x89ABCDEF);
391*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi16xGHIJKLMN, vk16xGHIJKLMN);
392*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi16xOPQRSTUV, vk16xOPQRSTUV);
393*4bdc9457SAndroid Build Coastguard Worker 
394*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi17x01234567 = vld1q_f16(i17); i17 += 8;
395*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi17x89ABCDEF = vld1q_f16(i17); i17 += 8;
396*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi17xGHIJKLMN = vld1q_f16(i17); i17 += 8;
397*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi17xOPQRSTUV = vld1q_f16(i17); i17 += 8;
398*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk17x01234567 = vld1q_f16(w); w += 8;
399*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk17x89ABCDEF = vld1q_f16(w); w += 8;
400*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk17xGHIJKLMN = vld1q_f16(w); w += 8;
401*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk17xOPQRSTUV = vld1q_f16(w); w += 8;
402*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi17x01234567, vk17x01234567);
403*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi17x89ABCDEF, vk17x89ABCDEF);
404*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi17xGHIJKLMN, vk17xGHIJKLMN);
405*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi17xOPQRSTUV, vk17xOPQRSTUV);
406*4bdc9457SAndroid Build Coastguard Worker 
407*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi18x01234567 = vld1q_f16(i18); i18 += 8;
408*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi18x89ABCDEF = vld1q_f16(i18); i18 += 8;
409*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi18xGHIJKLMN = vld1q_f16(i18); i18 += 8;
410*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi18xOPQRSTUV = vld1q_f16(i18); i18 += 8;
411*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk18x01234567 = vld1q_f16(w); w += 8;
412*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk18x89ABCDEF = vld1q_f16(w); w += 8;
413*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk18xGHIJKLMN = vld1q_f16(w); w += 8;
414*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk18xOPQRSTUV = vld1q_f16(w); w += 8;
415*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567);
416*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi18x89ABCDEF, vk18x89ABCDEF);
417*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi18xGHIJKLMN, vk18xGHIJKLMN);
418*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi18xOPQRSTUV, vk18xOPQRSTUV);
419*4bdc9457SAndroid Build Coastguard Worker 
420*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi19x01234567 = vld1q_f16(i19); i19 += 8;
421*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi19x89ABCDEF = vld1q_f16(i19); i19 += 8;
422*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi19xGHIJKLMN = vld1q_f16(i19); i19 += 8;
423*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi19xOPQRSTUV = vld1q_f16(i19); i19 += 8;
424*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk19x01234567 = vld1q_f16(w); w += 8;
425*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk19x89ABCDEF = vld1q_f16(w); w += 8;
426*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk19xGHIJKLMN = vld1q_f16(w); w += 8;
427*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk19xOPQRSTUV = vld1q_f16(w); w += 8;
428*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi19x01234567, vk19x01234567);
429*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi19x89ABCDEF, vk19x89ABCDEF);
430*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi19xGHIJKLMN, vk19xGHIJKLMN);
431*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi19xOPQRSTUV, vk19xOPQRSTUV);
432*4bdc9457SAndroid Build Coastguard Worker 
433*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8;
434*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi20x89ABCDEF = vld1q_f16(i20); i20 += 8;
435*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi20xGHIJKLMN = vld1q_f16(i20); i20 += 8;
436*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi20xOPQRSTUV = vld1q_f16(i20); i20 += 8;
437*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk20x01234567 = vld1q_f16(w); w += 8;
438*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk20x89ABCDEF = vld1q_f16(w); w += 8;
439*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk20xGHIJKLMN = vld1q_f16(w); w += 8;
440*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk20xOPQRSTUV = vld1q_f16(w); w += 8;
441*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567);
442*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi20x89ABCDEF, vk20x89ABCDEF);
443*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi20xGHIJKLMN, vk20xGHIJKLMN);
444*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi20xOPQRSTUV, vk20xOPQRSTUV);
445*4bdc9457SAndroid Build Coastguard Worker 
446*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi21x01234567 = vld1q_f16(i21); i21 += 8;
447*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi21x89ABCDEF = vld1q_f16(i21); i21 += 8;
448*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi21xGHIJKLMN = vld1q_f16(i21); i21 += 8;
449*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi21xOPQRSTUV = vld1q_f16(i21); i21 += 8;
450*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk21x01234567 = vld1q_f16(w); w += 8;
451*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk21x89ABCDEF = vld1q_f16(w); w += 8;
452*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk21xGHIJKLMN = vld1q_f16(w); w += 8;
453*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk21xOPQRSTUV = vld1q_f16(w); w += 8;
454*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi21x01234567, vk21x01234567);
455*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi21x89ABCDEF, vk21x89ABCDEF);
456*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi21xGHIJKLMN, vk21xGHIJKLMN);
457*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi21xOPQRSTUV, vk21xOPQRSTUV);
458*4bdc9457SAndroid Build Coastguard Worker 
459*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi22x01234567 = vld1q_f16(i22); i22 += 8;
460*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi22x89ABCDEF = vld1q_f16(i22); i22 += 8;
461*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi22xGHIJKLMN = vld1q_f16(i22); i22 += 8;
462*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi22xOPQRSTUV = vld1q_f16(i22); i22 += 8;
463*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8;
464*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk22x89ABCDEF = vld1q_f16(w); w += 8;
465*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk22xGHIJKLMN = vld1q_f16(w); w += 8;
466*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk22xOPQRSTUV = vld1q_f16(w); w += 8;
467*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567);
468*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi22x89ABCDEF, vk22x89ABCDEF);
469*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi22xGHIJKLMN, vk22xGHIJKLMN);
470*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi22xOPQRSTUV, vk22xOPQRSTUV);
471*4bdc9457SAndroid Build Coastguard Worker 
472*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi23x01234567 = vld1q_f16(i23); i23 += 8;
473*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi23x89ABCDEF = vld1q_f16(i23); i23 += 8;
474*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi23xGHIJKLMN = vld1q_f16(i23); i23 += 8;
475*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi23xOPQRSTUV = vld1q_f16(i23); i23 += 8;
476*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk23x01234567 = vld1q_f16(w); w += 8;
477*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk23x89ABCDEF = vld1q_f16(w); w += 8;
478*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk23xGHIJKLMN = vld1q_f16(w); w += 8;
479*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk23xOPQRSTUV = vld1q_f16(w); w += 8;
480*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi23x01234567, vk23x01234567);
481*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi23x89ABCDEF, vk23x89ABCDEF);
482*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi23xGHIJKLMN, vk23xGHIJKLMN);
483*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi23xOPQRSTUV, vk23xOPQRSTUV);
484*4bdc9457SAndroid Build Coastguard Worker 
485*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8;
486*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi24x89ABCDEF = vld1q_f16(i24); i24 += 8;
487*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi24xGHIJKLMN = vld1q_f16(i24); i24 += 8;
488*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi24xOPQRSTUV = vld1q_f16(i24); i24 += 8;
489*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk24x01234567 = vld1q_f16(w); w += 8;
490*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk24x89ABCDEF = vld1q_f16(w); w += 8;
491*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk24xGHIJKLMN = vld1q_f16(w); w += 8;
492*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk24xOPQRSTUV = vld1q_f16(w); w += 8;
493*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567);
494*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi24x89ABCDEF, vk24x89ABCDEF);
495*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi24xGHIJKLMN, vk24xGHIJKLMN);
496*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi24xOPQRSTUV, vk24xOPQRSTUV);
497*4bdc9457SAndroid Build Coastguard Worker 
498*4bdc9457SAndroid Build Coastguard Worker 
499*4bdc9457SAndroid Build Coastguard Worker       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
500*4bdc9457SAndroid Build Coastguard Worker       float16x8_t vacc89ABCDEF = vmaxq_f16(vacc89ABCDEFp0, vmin);
501*4bdc9457SAndroid Build Coastguard Worker       float16x8_t vaccGHIJKLMN = vmaxq_f16(vaccGHIJKLMNp0, vmin);
502*4bdc9457SAndroid Build Coastguard Worker       float16x8_t vaccOPQRSTUV = vmaxq_f16(vaccOPQRSTUVp0, vmin);
503*4bdc9457SAndroid Build Coastguard Worker       vacc01234567 = vminq_f16(vacc01234567, vmax);
504*4bdc9457SAndroid Build Coastguard Worker       vacc89ABCDEF = vminq_f16(vacc89ABCDEF, vmax);
505*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMN = vminq_f16(vaccGHIJKLMN, vmax);
506*4bdc9457SAndroid Build Coastguard Worker       vaccOPQRSTUV = vminq_f16(vaccOPQRSTUV, vmax);
507*4bdc9457SAndroid Build Coastguard Worker 
508*4bdc9457SAndroid Build Coastguard Worker       vst1q_f16(output, vacc01234567); output += 8;
509*4bdc9457SAndroid Build Coastguard Worker       vst1q_f16(output, vacc89ABCDEF); output += 8;
510*4bdc9457SAndroid Build Coastguard Worker       vst1q_f16(output, vaccGHIJKLMN); output += 8;
511*4bdc9457SAndroid Build Coastguard Worker       vst1q_f16(output, vaccOPQRSTUV); output += 8;
512*4bdc9457SAndroid Build Coastguard Worker     }
513*4bdc9457SAndroid Build Coastguard Worker     for (; c >= 8; c -= 8) {
514*4bdc9457SAndroid Build Coastguard Worker       float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
515*4bdc9457SAndroid Build Coastguard Worker 
516*4bdc9457SAndroid Build Coastguard Worker 
517*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
518*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk0x01234567 = vld1q_f16(w + 24);
519*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
520*4bdc9457SAndroid Build Coastguard Worker 
521*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
522*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk1x01234567 = vld1q_f16(w + 56);
523*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);
524*4bdc9457SAndroid Build Coastguard Worker 
525*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
526*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk2x01234567 = vld1q_f16(w + 88);
527*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
528*4bdc9457SAndroid Build Coastguard Worker 
529*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
530*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk3x01234567 = vld1q_f16(w + 120);
531*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi3x01234567, vk3x01234567);
532*4bdc9457SAndroid Build Coastguard Worker 
533*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
534*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk4x01234567 = vld1q_f16(w + 152);
535*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);
536*4bdc9457SAndroid Build Coastguard Worker 
537*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
538*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk5x01234567 = vld1q_f16(w + 184);
539*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi5x01234567, vk5x01234567);
540*4bdc9457SAndroid Build Coastguard Worker 
541*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
542*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk6x01234567 = vld1q_f16(w + 216);
543*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);
544*4bdc9457SAndroid Build Coastguard Worker 
545*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi7x01234567 = vld1q_f16(i7); i7 += 8;
546*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk7x01234567 = vld1q_f16(w + 248);
547*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi7x01234567, vk7x01234567);
548*4bdc9457SAndroid Build Coastguard Worker 
549*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi8x01234567 = vld1q_f16(i8); i8 += 8;
550*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk8x01234567 = vld1q_f16(w + 280);
551*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);
552*4bdc9457SAndroid Build Coastguard Worker 
553*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi9x01234567 = vld1q_f16(i9); i9 += 8;
554*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk9x01234567 = vld1q_f16(w + 312);
555*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi9x01234567, vk9x01234567);
556*4bdc9457SAndroid Build Coastguard Worker 
557*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi10x01234567 = vld1q_f16(i10); i10 += 8;
558*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk10x01234567 = vld1q_f16(w + 344);
559*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567);
560*4bdc9457SAndroid Build Coastguard Worker 
561*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi11x01234567 = vld1q_f16(i11); i11 += 8;
562*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk11x01234567 = vld1q_f16(w + 376);
563*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi11x01234567, vk11x01234567);
564*4bdc9457SAndroid Build Coastguard Worker 
565*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi12x01234567 = vld1q_f16(i12); i12 += 8;
566*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk12x01234567 = vld1q_f16(w + 408);
567*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567);
568*4bdc9457SAndroid Build Coastguard Worker 
569*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi13x01234567 = vld1q_f16(i13); i13 += 8;
570*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk13x01234567 = vld1q_f16(w + 440);
571*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi13x01234567, vk13x01234567);
572*4bdc9457SAndroid Build Coastguard Worker 
573*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi14x01234567 = vld1q_f16(i14); i14 += 8;
574*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk14x01234567 = vld1q_f16(w + 472);
575*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567);
576*4bdc9457SAndroid Build Coastguard Worker 
577*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi15x01234567 = vld1q_f16(i15); i15 += 8;
578*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk15x01234567 = vld1q_f16(w + 504);
579*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi15x01234567, vk15x01234567);
580*4bdc9457SAndroid Build Coastguard Worker 
581*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi16x01234567 = vld1q_f16(i16); i16 += 8;
582*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk16x01234567 = vld1q_f16(w + 536);
583*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567);
584*4bdc9457SAndroid Build Coastguard Worker 
585*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi17x01234567 = vld1q_f16(i17); i17 += 8;
586*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk17x01234567 = vld1q_f16(w + 568);
587*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi17x01234567, vk17x01234567);
588*4bdc9457SAndroid Build Coastguard Worker 
589*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi18x01234567 = vld1q_f16(i18); i18 += 8;
590*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk18x01234567 = vld1q_f16(w + 600);
591*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567);
592*4bdc9457SAndroid Build Coastguard Worker 
593*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi19x01234567 = vld1q_f16(i19); i19 += 8;
594*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk19x01234567 = vld1q_f16(w + 632);
595*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi19x01234567, vk19x01234567);
596*4bdc9457SAndroid Build Coastguard Worker 
597*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8;
598*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk20x01234567 = vld1q_f16(w + 664);
599*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567);
600*4bdc9457SAndroid Build Coastguard Worker 
601*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi21x01234567 = vld1q_f16(i21); i21 += 8;
602*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk21x01234567 = vld1q_f16(w + 696);
603*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi21x01234567, vk21x01234567);
604*4bdc9457SAndroid Build Coastguard Worker 
605*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi22x01234567 = vld1q_f16(i22); i22 += 8;
606*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk22x01234567 = vld1q_f16(w + 728);
607*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567);
608*4bdc9457SAndroid Build Coastguard Worker 
609*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi23x01234567 = vld1q_f16(i23); i23 += 8;
610*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk23x01234567 = vld1q_f16(w + 760);
611*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi23x01234567, vk23x01234567);
612*4bdc9457SAndroid Build Coastguard Worker 
613*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8;
614*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk24x01234567 = vld1q_f16(w + 792);
615*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567);
616*4bdc9457SAndroid Build Coastguard Worker 
617*4bdc9457SAndroid Build Coastguard Worker 
618*4bdc9457SAndroid Build Coastguard Worker       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
619*4bdc9457SAndroid Build Coastguard Worker       vacc01234567 = vminq_f16(vacc01234567, vmax);
620*4bdc9457SAndroid Build Coastguard Worker 
621*4bdc9457SAndroid Build Coastguard Worker       vst1q_f16(output, vacc01234567); output += 8;
622*4bdc9457SAndroid Build Coastguard Worker     }
623*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNLIKELY(c != 0) {
624*4bdc9457SAndroid Build Coastguard Worker       float16x8_t vacc01234567p0 = vld1q_f16(w);
625*4bdc9457SAndroid Build Coastguard Worker 
626*4bdc9457SAndroid Build Coastguard Worker 
627*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi0x01234567 = vld1q_f16(i0);
628*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk0x01234567 = vld1q_f16(w + 32);
629*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
630*4bdc9457SAndroid Build Coastguard Worker 
631*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi1x01234567 = vld1q_f16(i1);
632*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk1x01234567 = vld1q_f16(w + 64);
633*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);
634*4bdc9457SAndroid Build Coastguard Worker 
635*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi2x01234567 = vld1q_f16(i2);
636*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk2x01234567 = vld1q_f16(w + 96);
637*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
638*4bdc9457SAndroid Build Coastguard Worker 
639*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi3x01234567 = vld1q_f16(i3);
640*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk3x01234567 = vld1q_f16(w + 128);
641*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi3x01234567, vk3x01234567);
642*4bdc9457SAndroid Build Coastguard Worker 
643*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi4x01234567 = vld1q_f16(i4);
644*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk4x01234567 = vld1q_f16(w + 160);
645*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);
646*4bdc9457SAndroid Build Coastguard Worker 
647*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi5x01234567 = vld1q_f16(i5);
648*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk5x01234567 = vld1q_f16(w + 192);
649*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi5x01234567, vk5x01234567);
650*4bdc9457SAndroid Build Coastguard Worker 
651*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi6x01234567 = vld1q_f16(i6);
652*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk6x01234567 = vld1q_f16(w + 224);
653*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);
654*4bdc9457SAndroid Build Coastguard Worker 
655*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi7x01234567 = vld1q_f16(i7);
656*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk7x01234567 = vld1q_f16(w + 256);
657*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi7x01234567, vk7x01234567);
658*4bdc9457SAndroid Build Coastguard Worker 
659*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi8x01234567 = vld1q_f16(i8);
660*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk8x01234567 = vld1q_f16(w + 288);
661*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);
662*4bdc9457SAndroid Build Coastguard Worker 
663*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi9x01234567 = vld1q_f16(i9);
664*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk9x01234567 = vld1q_f16(w + 320);
665*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi9x01234567, vk9x01234567);
666*4bdc9457SAndroid Build Coastguard Worker 
667*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi10x01234567 = vld1q_f16(i10);
668*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk10x01234567 = vld1q_f16(w + 352);
669*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567);
670*4bdc9457SAndroid Build Coastguard Worker 
671*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi11x01234567 = vld1q_f16(i11);
672*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk11x01234567 = vld1q_f16(w + 384);
673*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi11x01234567, vk11x01234567);
674*4bdc9457SAndroid Build Coastguard Worker 
675*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi12x01234567 = vld1q_f16(i12);
676*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk12x01234567 = vld1q_f16(w + 416);
677*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567);
678*4bdc9457SAndroid Build Coastguard Worker 
679*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi13x01234567 = vld1q_f16(i13);
680*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk13x01234567 = vld1q_f16(w + 448);
681*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi13x01234567, vk13x01234567);
682*4bdc9457SAndroid Build Coastguard Worker 
683*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi14x01234567 = vld1q_f16(i14);
684*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk14x01234567 = vld1q_f16(w + 480);
685*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567);
686*4bdc9457SAndroid Build Coastguard Worker 
687*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi15x01234567 = vld1q_f16(i15);
688*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk15x01234567 = vld1q_f16(w + 512);
689*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi15x01234567, vk15x01234567);
690*4bdc9457SAndroid Build Coastguard Worker 
691*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi16x01234567 = vld1q_f16(i16);
692*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk16x01234567 = vld1q_f16(w + 544);
693*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567);
694*4bdc9457SAndroid Build Coastguard Worker 
695*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi17x01234567 = vld1q_f16(i17);
696*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk17x01234567 = vld1q_f16(w + 576);
697*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi17x01234567, vk17x01234567);
698*4bdc9457SAndroid Build Coastguard Worker 
699*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi18x01234567 = vld1q_f16(i18);
700*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk18x01234567 = vld1q_f16(w + 608);
701*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567);
702*4bdc9457SAndroid Build Coastguard Worker 
703*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi19x01234567 = vld1q_f16(i19);
704*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk19x01234567 = vld1q_f16(w + 640);
705*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi19x01234567, vk19x01234567);
706*4bdc9457SAndroid Build Coastguard Worker 
707*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi20x01234567 = vld1q_f16(i20);
708*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk20x01234567 = vld1q_f16(w + 672);
709*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567);
710*4bdc9457SAndroid Build Coastguard Worker 
711*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi21x01234567 = vld1q_f16(i21);
712*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk21x01234567 = vld1q_f16(w + 704);
713*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi21x01234567, vk21x01234567);
714*4bdc9457SAndroid Build Coastguard Worker 
715*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi22x01234567 = vld1q_f16(i22);
716*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk22x01234567 = vld1q_f16(w + 736);
717*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567);
718*4bdc9457SAndroid Build Coastguard Worker 
719*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi23x01234567 = vld1q_f16(i23);
720*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk23x01234567 = vld1q_f16(w + 768);
721*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi23x01234567, vk23x01234567);
722*4bdc9457SAndroid Build Coastguard Worker 
723*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vi24x01234567 = vld1q_f16(i24);
724*4bdc9457SAndroid Build Coastguard Worker       const float16x8_t vk24x01234567 = vld1q_f16(w + 800);
725*4bdc9457SAndroid Build Coastguard Worker       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567);
726*4bdc9457SAndroid Build Coastguard Worker 
727*4bdc9457SAndroid Build Coastguard Worker 
728*4bdc9457SAndroid Build Coastguard Worker       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
729*4bdc9457SAndroid Build Coastguard Worker       vacc01234567 = vminq_f16(vacc01234567, vmax);
730*4bdc9457SAndroid Build Coastguard Worker 
731*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vacc0123 = vget_low_f16(vacc01234567);
732*4bdc9457SAndroid Build Coastguard Worker       if (c & 4) {
733*4bdc9457SAndroid Build Coastguard Worker         vst1_f16(output, vacc0123); output += 4;
734*4bdc9457SAndroid Build Coastguard Worker         vacc0123 = vget_high_f16(vacc01234567);
735*4bdc9457SAndroid Build Coastguard Worker       }
736*4bdc9457SAndroid Build Coastguard Worker       if (c & 2) {
737*4bdc9457SAndroid Build Coastguard Worker         vst1_lane_u32((void*) output, vreinterpret_u32_f16(vacc0123), 0); output += 2;
738*4bdc9457SAndroid Build Coastguard Worker         vacc0123 = vext_f16(vacc0123, vacc0123, 2);
739*4bdc9457SAndroid Build Coastguard Worker       }
740*4bdc9457SAndroid Build Coastguard Worker       if (c & 1) {
741*4bdc9457SAndroid Build Coastguard Worker         vst1_lane_f16(output, vacc0123, 0); output += 1;
742*4bdc9457SAndroid Build Coastguard Worker       }
743*4bdc9457SAndroid Build Coastguard Worker     }
744*4bdc9457SAndroid Build Coastguard Worker 
745*4bdc9457SAndroid Build Coastguard Worker     output = (__fp16*) ((uintptr_t) output + output_increment);
746*4bdc9457SAndroid Build Coastguard Worker   } while (--output_width != 0);
747*4bdc9457SAndroid Build Coastguard Worker }
748