xref: /aosp_15_r20/external/XNNPACK/src/f16-dwconv2d-chw/gen/5x5p2-minmax-neonfp16arith-4x4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f16-dwconv2d-chw/5x5p2-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 #include <xnnpack/math.h>
16*4bdc9457SAndroid Build Coastguard Worker 
17*4bdc9457SAndroid Build Coastguard Worker 
xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_4x4(size_t input_height,size_t input_width,const void * input,const void * weights,const void * zero,void * output,uint32_t padding_top,const union xnn_f16_chw_params params[restrict XNN_MIN_ELEMENTS (1)])18*4bdc9457SAndroid Build Coastguard Worker void xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_4x4(
19*4bdc9457SAndroid Build Coastguard Worker     size_t input_height,
20*4bdc9457SAndroid Build Coastguard Worker     size_t input_width,
21*4bdc9457SAndroid Build Coastguard Worker     const void* input,
22*4bdc9457SAndroid Build Coastguard Worker     const void* weights,
23*4bdc9457SAndroid Build Coastguard Worker     const void* zero,
24*4bdc9457SAndroid Build Coastguard Worker     void* output,
25*4bdc9457SAndroid Build Coastguard Worker     uint32_t padding_top,
26*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f16_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27*4bdc9457SAndroid Build Coastguard Worker {
28*4bdc9457SAndroid Build Coastguard Worker   assert(input_height != 0);
29*4bdc9457SAndroid Build Coastguard Worker   assert(input_width != 0);
30*4bdc9457SAndroid Build Coastguard Worker   assert(input_width % sizeof(__fp16) == 0);
31*4bdc9457SAndroid Build Coastguard Worker   assert(padding_top == 2);
32*4bdc9457SAndroid Build Coastguard Worker 
33*4bdc9457SAndroid Build Coastguard Worker   const uint16x4_t vmask = vld1_u16(params->neonfp16arith.mask);
34*4bdc9457SAndroid Build Coastguard Worker   const float16x4_t vmax = vld1_dup_f16(&params->neonfp16arith.max);
35*4bdc9457SAndroid Build Coastguard Worker   const float16x4_t vmin = vld1_dup_f16(&params->neonfp16arith.min);
36*4bdc9457SAndroid Build Coastguard Worker 
37*4bdc9457SAndroid Build Coastguard Worker   const __fp16* w0 = (const __fp16*)weights;
38*4bdc9457SAndroid Build Coastguard Worker   const float16x8_t vw01234567 = vld1q_f16(w0);
39*4bdc9457SAndroid Build Coastguard Worker   const float16x8_t vw89ABCDEF = vld1q_f16(w0 + 8);
40*4bdc9457SAndroid Build Coastguard Worker   const float16x8_t vwGHIJKLMN = vld1q_f16(w0 + 16);
41*4bdc9457SAndroid Build Coastguard Worker   const float16x4_t vwOP = vreinterpret_f16_u32(vld1_lane_u32((const void*)(w0 + 24), vmov_n_u32(0), 0));
42*4bdc9457SAndroid Build Coastguard Worker 
43*4bdc9457SAndroid Build Coastguard Worker   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(__fp16));
44*4bdc9457SAndroid Build Coastguard Worker 
45*4bdc9457SAndroid Build Coastguard Worker   const __fp16* i0 = zero;
46*4bdc9457SAndroid Build Coastguard Worker   const __fp16* i1 = zero;
47*4bdc9457SAndroid Build Coastguard Worker   const __fp16* i2 = input;
48*4bdc9457SAndroid Build Coastguard Worker   const __fp16* i3 = (const __fp16*) ((uintptr_t) i2 + input_width);
49*4bdc9457SAndroid Build Coastguard Worker   const __fp16* i4 = (const __fp16*) ((uintptr_t) i3 + input_width);
50*4bdc9457SAndroid Build Coastguard Worker   const __fp16* i5 = (const __fp16*) ((uintptr_t) i4 + input_width);
51*4bdc9457SAndroid Build Coastguard Worker   const __fp16* i6 = (const __fp16*) ((uintptr_t) i5 + input_width);
52*4bdc9457SAndroid Build Coastguard Worker   const __fp16* i7 = (const __fp16*) ((uintptr_t) i6 + input_width);
53*4bdc9457SAndroid Build Coastguard Worker 
54*4bdc9457SAndroid Build Coastguard Worker   __fp16* o0 = output;
55*4bdc9457SAndroid Build Coastguard Worker   __fp16* o1 = (__fp16*) ((uintptr_t) o0 + input_width);
56*4bdc9457SAndroid Build Coastguard Worker   __fp16* o2 = (__fp16*) ((uintptr_t) o1 + input_width);
57*4bdc9457SAndroid Build Coastguard Worker   __fp16* o3 = (__fp16*) ((uintptr_t) o2 + input_width);
58*4bdc9457SAndroid Build Coastguard Worker 
59*4bdc9457SAndroid Build Coastguard Worker   size_t output_height = input_height;
60*4bdc9457SAndroid Build Coastguard Worker   do {
61*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 2) {
62*4bdc9457SAndroid Build Coastguard Worker       i3 = zero;
63*4bdc9457SAndroid Build Coastguard Worker       o1 = o0;
64*4bdc9457SAndroid Build Coastguard Worker     }
65*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 3) {
66*4bdc9457SAndroid Build Coastguard Worker       i4 = zero;
67*4bdc9457SAndroid Build Coastguard Worker       o2 = o1;
68*4bdc9457SAndroid Build Coastguard Worker     }
69*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 4) {
70*4bdc9457SAndroid Build Coastguard Worker       i5 = zero;
71*4bdc9457SAndroid Build Coastguard Worker       o3 = o2;
72*4bdc9457SAndroid Build Coastguard Worker     }
73*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 5) {
74*4bdc9457SAndroid Build Coastguard Worker       i6 = zero;
75*4bdc9457SAndroid Build Coastguard Worker     }
76*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 6) {
77*4bdc9457SAndroid Build Coastguard Worker       i7 = zero;
78*4bdc9457SAndroid Build Coastguard Worker     }
79*4bdc9457SAndroid Build Coastguard Worker 
80*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi0x0123 = vmov_n_f16(0);
81*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi1x0123 = vmov_n_f16(0);
82*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi2x0123 = vmov_n_f16(0);
83*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi3x0123 = vmov_n_f16(0);
84*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi4x0123 = vmov_n_f16(0);
85*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi5x0123 = vmov_n_f16(0);
86*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi6x0123 = vmov_n_f16(0);
87*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi7x0123 = vmov_n_f16(0);
88*4bdc9457SAndroid Build Coastguard Worker 
89*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi0x4567 = vld1_f16(i0); i0 += 4;
90*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi1x4567 = vld1_f16(i1); i1 += 4;
91*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi2x4567 = vld1_f16(i2); i2 += 4;
92*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi3x4567 = vld1_f16(i3); i3 += 4;
93*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi4x4567 = vld1_f16(i4); i4 += 4;
94*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi5x4567 = vld1_f16(i5); i5 += 4;
95*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi6x4567 = vld1_f16(i6); i6 += 4;
96*4bdc9457SAndroid Build Coastguard Worker     float16x4_t vi7x4567 = vld1_f16(i7); i7 += 4;
97*4bdc9457SAndroid Build Coastguard Worker 
98*4bdc9457SAndroid Build Coastguard Worker     size_t w = input_width;
99*4bdc9457SAndroid Build Coastguard Worker     for (; w > 8 * sizeof(__fp16); w -= 4 * sizeof(__fp16)) {
100*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo0p0 = vdup_laneq_f16(vw01234567, 0);
101*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo1p0 = vdup_laneq_f16(vw01234567, 0);
102*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo2p0 = vdup_laneq_f16(vw01234567, 0);
103*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo3p0 = vdup_laneq_f16(vw01234567, 0);
104*4bdc9457SAndroid Build Coastguard Worker 
105*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi0x89AB = vld1_f16(i0); i0 += 4;
106*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi1x89AB = vld1_f16(i1); i1 += 4;
107*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi2x89AB = vld1_f16(i2); i2 += 4;
108*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi3x89AB = vld1_f16(i3); i3 += 4;
109*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi4x89AB = vld1_f16(i4); i4 += 4;
110*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi5x89AB = vld1_f16(i5); i5 += 4;
111*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi6x89AB = vld1_f16(i6); i6 += 4;
112*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi7x89AB = vld1_f16(i7); i7 += 4;
113*4bdc9457SAndroid Build Coastguard Worker 
114*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x4567, vw01234567, 3);
115*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x4567, vw01234567, 3);
116*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x4567, vw01234567, 3);
117*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x4567, vw01234567, 3);
118*4bdc9457SAndroid Build Coastguard Worker 
119*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x4567, vw89ABCDEF, 0);
120*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x4567, vw89ABCDEF, 0);
121*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x4567, vw89ABCDEF, 0);
122*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x4567, vw89ABCDEF, 0);
123*4bdc9457SAndroid Build Coastguard Worker 
124*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x4567, vw89ABCDEF, 5);
125*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x4567, vw89ABCDEF, 5);
126*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x4567, vw89ABCDEF, 5);
127*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x4567, vw89ABCDEF, 5);
128*4bdc9457SAndroid Build Coastguard Worker 
129*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x4567, vwGHIJKLMN, 2);
130*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x4567, vwGHIJKLMN, 2);
131*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x4567, vwGHIJKLMN, 2);
132*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x4567, vwGHIJKLMN, 2);
133*4bdc9457SAndroid Build Coastguard Worker 
134*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi4x4567, vwGHIJKLMN, 7);
135*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi5x4567, vwGHIJKLMN, 7);
136*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi6x4567, vwGHIJKLMN, 7);
137*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi7x4567, vwGHIJKLMN, 7);
138*4bdc9457SAndroid Build Coastguard Worker 
139*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi0x3456 = vext_f16(vi0x0123, vi0x4567, 3);
140*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi1x3456 = vext_f16(vi1x0123, vi1x4567, 3);
141*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi2x3456 = vext_f16(vi2x0123, vi2x4567, 3);
142*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi3x3456 = vext_f16(vi3x0123, vi3x4567, 3);
143*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi4x3456 = vext_f16(vi4x0123, vi4x4567, 3);
144*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi5x3456 = vext_f16(vi5x0123, vi5x4567, 3);
145*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi6x3456 = vext_f16(vi6x0123, vi6x4567, 3);
146*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi7x3456 = vext_f16(vi7x0123, vi7x4567, 3);
147*4bdc9457SAndroid Build Coastguard Worker 
148*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x3456, vw01234567, 2);
149*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x3456, vw01234567, 2);
150*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x3456, vw01234567, 2);
151*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x3456, vw01234567, 2);
152*4bdc9457SAndroid Build Coastguard Worker 
153*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x3456, vw01234567, 7);
154*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x3456, vw01234567, 7);
155*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x3456, vw01234567, 7);
156*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x3456, vw01234567, 7);
157*4bdc9457SAndroid Build Coastguard Worker 
158*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x3456, vw89ABCDEF, 4);
159*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x3456, vw89ABCDEF, 4);
160*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x3456, vw89ABCDEF, 4);
161*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x3456, vw89ABCDEF, 4);
162*4bdc9457SAndroid Build Coastguard Worker 
163*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x3456, vwGHIJKLMN, 1);
164*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x3456, vwGHIJKLMN, 1);
165*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x3456, vwGHIJKLMN, 1);
166*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x3456, vwGHIJKLMN, 1);
167*4bdc9457SAndroid Build Coastguard Worker 
168*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi4x3456, vwGHIJKLMN, 6);
169*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi5x3456, vwGHIJKLMN, 6);
170*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi6x3456, vwGHIJKLMN, 6);
171*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi7x3456, vwGHIJKLMN, 6);
172*4bdc9457SAndroid Build Coastguard Worker 
173*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi0x2345 = vext_f16(vi0x0123, vi0x4567, 2);
174*4bdc9457SAndroid Build Coastguard Worker       vi0x0123 = vi0x4567;
175*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi1x2345 = vext_f16(vi1x0123, vi1x4567, 2);
176*4bdc9457SAndroid Build Coastguard Worker       vi1x0123 = vi1x4567;
177*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2);
178*4bdc9457SAndroid Build Coastguard Worker       vi2x0123 = vi2x4567;
179*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi3x2345 = vext_f16(vi3x0123, vi3x4567, 2);
180*4bdc9457SAndroid Build Coastguard Worker       vi3x0123 = vi3x4567;
181*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi4x2345 = vext_f16(vi4x0123, vi4x4567, 2);
182*4bdc9457SAndroid Build Coastguard Worker       vi4x0123 = vi4x4567;
183*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi5x2345 = vext_f16(vi5x0123, vi5x4567, 2);
184*4bdc9457SAndroid Build Coastguard Worker       vi5x0123 = vi5x4567;
185*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi6x2345 = vext_f16(vi6x0123, vi6x4567, 2);
186*4bdc9457SAndroid Build Coastguard Worker       vi6x0123 = vi6x4567;
187*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi7x2345 = vext_f16(vi7x0123, vi7x4567, 2);
188*4bdc9457SAndroid Build Coastguard Worker       vi7x0123 = vi7x4567;
189*4bdc9457SAndroid Build Coastguard Worker 
190*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x2345, vw01234567, 1);
191*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x2345, vw01234567, 1);
192*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x2345, vw01234567, 1);
193*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x2345, vw01234567, 1);
194*4bdc9457SAndroid Build Coastguard Worker 
195*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x2345, vw01234567, 6);
196*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x2345, vw01234567, 6);
197*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x2345, vw01234567, 6);
198*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x2345, vw01234567, 6);
199*4bdc9457SAndroid Build Coastguard Worker 
200*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3);
201*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x2345, vw89ABCDEF, 3);
202*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x2345, vw89ABCDEF, 3);
203*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x2345, vw89ABCDEF, 3);
204*4bdc9457SAndroid Build Coastguard Worker 
205*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x2345, vwGHIJKLMN, 0);
206*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x2345, vwGHIJKLMN, 0);
207*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x2345, vwGHIJKLMN, 0);
208*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x2345, vwGHIJKLMN, 0);
209*4bdc9457SAndroid Build Coastguard Worker 
210*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi4x2345, vwGHIJKLMN, 5);
211*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi5x2345, vwGHIJKLMN, 5);
212*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi6x2345, vwGHIJKLMN, 5);
213*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi7x2345, vwGHIJKLMN, 5);
214*4bdc9457SAndroid Build Coastguard Worker 
215*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi0x5678 = vext_f16(vi0x4567, vi0x89AB, 1);
216*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi1x5678 = vext_f16(vi1x4567, vi1x89AB, 1);
217*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi2x5678 = vext_f16(vi2x4567, vi2x89AB, 1);
218*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi3x5678 = vext_f16(vi3x4567, vi3x89AB, 1);
219*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi4x5678 = vext_f16(vi4x4567, vi4x89AB, 1);
220*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi5x5678 = vext_f16(vi5x4567, vi5x89AB, 1);
221*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi6x5678 = vext_f16(vi6x4567, vi6x89AB, 1);
222*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi7x5678 = vext_f16(vi7x4567, vi7x89AB, 1);
223*4bdc9457SAndroid Build Coastguard Worker 
224*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x5678, vw01234567, 4);
225*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x5678, vw01234567, 4);
226*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x5678, vw01234567, 4);
227*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x5678, vw01234567, 4);
228*4bdc9457SAndroid Build Coastguard Worker 
229*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x5678, vw89ABCDEF, 1);
230*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x5678, vw89ABCDEF, 1);
231*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x5678, vw89ABCDEF, 1);
232*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x5678, vw89ABCDEF, 1);
233*4bdc9457SAndroid Build Coastguard Worker 
234*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x5678, vw89ABCDEF, 6);
235*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x5678, vw89ABCDEF, 6);
236*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x5678, vw89ABCDEF, 6);
237*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x5678, vw89ABCDEF, 6);
238*4bdc9457SAndroid Build Coastguard Worker 
239*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x5678, vwGHIJKLMN, 3);
240*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x5678, vwGHIJKLMN, 3);
241*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x5678, vwGHIJKLMN, 3);
242*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x5678, vwGHIJKLMN, 3);
243*4bdc9457SAndroid Build Coastguard Worker 
244*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_lane_f16(vo0p0, vi4x5678, vwOP, 0);
245*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_lane_f16(vo1p0, vi5x5678, vwOP, 0);
246*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_lane_f16(vo2p0, vi6x5678, vwOP, 0);
247*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_lane_f16(vo3p0, vi7x5678, vwOP, 0);
248*4bdc9457SAndroid Build Coastguard Worker 
249*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi0x6789 = vext_f16(vi0x4567, vi0x89AB, 2);
250*4bdc9457SAndroid Build Coastguard Worker       vi0x4567 = vi0x89AB;
251*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi1x6789 = vext_f16(vi1x4567, vi1x89AB, 2);
252*4bdc9457SAndroid Build Coastguard Worker       vi1x4567 = vi1x89AB;
253*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi2x6789 = vext_f16(vi2x4567, vi2x89AB, 2);
254*4bdc9457SAndroid Build Coastguard Worker       vi2x4567 = vi2x89AB;
255*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi3x6789 = vext_f16(vi3x4567, vi3x89AB, 2);
256*4bdc9457SAndroid Build Coastguard Worker       vi3x4567 = vi3x89AB;
257*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi4x6789 = vext_f16(vi4x4567, vi4x89AB, 2);
258*4bdc9457SAndroid Build Coastguard Worker       vi4x4567 = vi4x89AB;
259*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi5x6789 = vext_f16(vi5x4567, vi5x89AB, 2);
260*4bdc9457SAndroid Build Coastguard Worker       vi5x4567 = vi5x89AB;
261*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi6x6789 = vext_f16(vi6x4567, vi6x89AB, 2);
262*4bdc9457SAndroid Build Coastguard Worker       vi6x4567 = vi6x89AB;
263*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi7x6789 = vext_f16(vi7x4567, vi7x89AB, 2);
264*4bdc9457SAndroid Build Coastguard Worker       vi7x4567 = vi7x89AB;
265*4bdc9457SAndroid Build Coastguard Worker 
266*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x6789, vw01234567, 5);
267*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x6789, vw01234567, 5);
268*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x6789, vw01234567, 5);
269*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x6789, vw01234567, 5);
270*4bdc9457SAndroid Build Coastguard Worker 
271*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x6789, vw89ABCDEF, 2);
272*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x6789, vw89ABCDEF, 2);
273*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x6789, vw89ABCDEF, 2);
274*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x6789, vw89ABCDEF, 2);
275*4bdc9457SAndroid Build Coastguard Worker 
276*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x6789, vw89ABCDEF, 7);
277*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x6789, vw89ABCDEF, 7);
278*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x6789, vw89ABCDEF, 7);
279*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x6789, vw89ABCDEF, 7);
280*4bdc9457SAndroid Build Coastguard Worker 
281*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x6789, vwGHIJKLMN, 4);
282*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x6789, vwGHIJKLMN, 4);
283*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x6789, vwGHIJKLMN, 4);
284*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x6789, vwGHIJKLMN, 4);
285*4bdc9457SAndroid Build Coastguard Worker 
286*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_lane_f16(vo0p0, vi4x6789, vwOP, 1);
287*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_lane_f16(vo1p0, vi5x6789, vwOP, 1);
288*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_lane_f16(vo2p0, vi6x6789, vwOP, 1);
289*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_lane_f16(vo3p0, vi7x6789, vwOP, 1);
290*4bdc9457SAndroid Build Coastguard Worker 
291*4bdc9457SAndroid Build Coastguard Worker 
292*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo0 = vmax_f16(vo0p0, vmin);
293*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo1 = vmax_f16(vo1p0, vmin);
294*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo2 = vmax_f16(vo2p0, vmin);
295*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo3 = vmax_f16(vo3p0, vmin);
296*4bdc9457SAndroid Build Coastguard Worker 
297*4bdc9457SAndroid Build Coastguard Worker       vo0 = vmin_f16(vo0, vmax);
298*4bdc9457SAndroid Build Coastguard Worker       vo1 = vmin_f16(vo1, vmax);
299*4bdc9457SAndroid Build Coastguard Worker       vo2 = vmin_f16(vo2, vmax);
300*4bdc9457SAndroid Build Coastguard Worker       vo3 = vmin_f16(vo3, vmax);
301*4bdc9457SAndroid Build Coastguard Worker 
302*4bdc9457SAndroid Build Coastguard Worker       vst1_f16(o3, vo3); o3 += 4;
303*4bdc9457SAndroid Build Coastguard Worker       vst1_f16(o2, vo2); o2 += 4;
304*4bdc9457SAndroid Build Coastguard Worker       vst1_f16(o1, vo1); o1 += 4;
305*4bdc9457SAndroid Build Coastguard Worker       vst1_f16(o0, vo0); o0 += 4;
306*4bdc9457SAndroid Build Coastguard Worker     }
307*4bdc9457SAndroid Build Coastguard Worker     // Always process the last block of 5..8 pixels.
308*4bdc9457SAndroid Build Coastguard Worker     if XNN_LIKELY(w > 4 * sizeof(__fp16)) {
309*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo0p0 = vdup_laneq_f16(vw01234567, 0);
310*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo1p0 = vdup_laneq_f16(vw01234567, 0);
311*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo2p0 = vdup_laneq_f16(vw01234567, 0);
312*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo3p0 = vdup_laneq_f16(vw01234567, 0);
313*4bdc9457SAndroid Build Coastguard Worker 
314*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vi0x89AB = vld1_f16(i0); i0 += 4;
315*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vi1x89AB = vld1_f16(i1); i1 += 4;
316*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vi2x89AB = vld1_f16(i2); i2 += 4;
317*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vi3x89AB = vld1_f16(i3); i3 += 4;
318*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vi4x89AB = vld1_f16(i4); i4 += 4;
319*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vi5x89AB = vld1_f16(i5); i5 += 4;
320*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vi6x89AB = vld1_f16(i6); i6 += 4;
321*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vi7x89AB = vld1_f16(i7); i7 += 4;
322*4bdc9457SAndroid Build Coastguard Worker 
323*4bdc9457SAndroid Build Coastguard Worker       vi0x89AB = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi0x89AB)));
324*4bdc9457SAndroid Build Coastguard Worker       vi1x89AB = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi1x89AB)));
325*4bdc9457SAndroid Build Coastguard Worker       vi2x89AB = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi2x89AB)));
326*4bdc9457SAndroid Build Coastguard Worker       vi3x89AB = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi3x89AB)));
327*4bdc9457SAndroid Build Coastguard Worker       vi4x89AB = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi4x89AB)));
328*4bdc9457SAndroid Build Coastguard Worker       vi5x89AB = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi5x89AB)));
329*4bdc9457SAndroid Build Coastguard Worker       vi6x89AB = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi6x89AB)));
330*4bdc9457SAndroid Build Coastguard Worker       vi7x89AB = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi7x89AB)));
331*4bdc9457SAndroid Build Coastguard Worker 
332*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x4567, vw01234567, 3);
333*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x4567, vw01234567, 3);
334*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x4567, vw01234567, 3);
335*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x4567, vw01234567, 3);
336*4bdc9457SAndroid Build Coastguard Worker 
337*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x4567, vw89ABCDEF, 0);
338*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x4567, vw89ABCDEF, 0);
339*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x4567, vw89ABCDEF, 0);
340*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x4567, vw89ABCDEF, 0);
341*4bdc9457SAndroid Build Coastguard Worker 
342*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x4567, vw89ABCDEF, 5);
343*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x4567, vw89ABCDEF, 5);
344*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x4567, vw89ABCDEF, 5);
345*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x4567, vw89ABCDEF, 5);
346*4bdc9457SAndroid Build Coastguard Worker 
347*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x4567, vwGHIJKLMN, 2);
348*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x4567, vwGHIJKLMN, 2);
349*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x4567, vwGHIJKLMN, 2);
350*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x4567, vwGHIJKLMN, 2);
351*4bdc9457SAndroid Build Coastguard Worker 
352*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi4x4567, vwGHIJKLMN, 7);
353*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi5x4567, vwGHIJKLMN, 7);
354*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi6x4567, vwGHIJKLMN, 7);
355*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi7x4567, vwGHIJKLMN, 7);
356*4bdc9457SAndroid Build Coastguard Worker 
357*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi0x3456 = vext_f16(vi0x0123, vi0x4567, 3);
358*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi1x3456 = vext_f16(vi1x0123, vi1x4567, 3);
359*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi2x3456 = vext_f16(vi2x0123, vi2x4567, 3);
360*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi3x3456 = vext_f16(vi3x0123, vi3x4567, 3);
361*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi4x3456 = vext_f16(vi4x0123, vi4x4567, 3);
362*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi5x3456 = vext_f16(vi5x0123, vi5x4567, 3);
363*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi6x3456 = vext_f16(vi6x0123, vi6x4567, 3);
364*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi7x3456 = vext_f16(vi7x0123, vi7x4567, 3);
365*4bdc9457SAndroid Build Coastguard Worker 
366*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x3456, vw01234567, 2);
367*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x3456, vw01234567, 2);
368*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x3456, vw01234567, 2);
369*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x3456, vw01234567, 2);
370*4bdc9457SAndroid Build Coastguard Worker 
371*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x3456, vw01234567, 7);
372*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x3456, vw01234567, 7);
373*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x3456, vw01234567, 7);
374*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x3456, vw01234567, 7);
375*4bdc9457SAndroid Build Coastguard Worker 
376*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x3456, vw89ABCDEF, 4);
377*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x3456, vw89ABCDEF, 4);
378*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x3456, vw89ABCDEF, 4);
379*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x3456, vw89ABCDEF, 4);
380*4bdc9457SAndroid Build Coastguard Worker 
381*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x3456, vwGHIJKLMN, 1);
382*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x3456, vwGHIJKLMN, 1);
383*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x3456, vwGHIJKLMN, 1);
384*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x3456, vwGHIJKLMN, 1);
385*4bdc9457SAndroid Build Coastguard Worker 
386*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi4x3456, vwGHIJKLMN, 6);
387*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi5x3456, vwGHIJKLMN, 6);
388*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi6x3456, vwGHIJKLMN, 6);
389*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi7x3456, vwGHIJKLMN, 6);
390*4bdc9457SAndroid Build Coastguard Worker 
391*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi0x2345 = vext_f16(vi0x0123, vi0x4567, 2);
392*4bdc9457SAndroid Build Coastguard Worker       vi0x0123 = vi0x4567;
393*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi1x2345 = vext_f16(vi1x0123, vi1x4567, 2);
394*4bdc9457SAndroid Build Coastguard Worker       vi1x0123 = vi1x4567;
395*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2);
396*4bdc9457SAndroid Build Coastguard Worker       vi2x0123 = vi2x4567;
397*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi3x2345 = vext_f16(vi3x0123, vi3x4567, 2);
398*4bdc9457SAndroid Build Coastguard Worker       vi3x0123 = vi3x4567;
399*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi4x2345 = vext_f16(vi4x0123, vi4x4567, 2);
400*4bdc9457SAndroid Build Coastguard Worker       vi4x0123 = vi4x4567;
401*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi5x2345 = vext_f16(vi5x0123, vi5x4567, 2);
402*4bdc9457SAndroid Build Coastguard Worker       vi5x0123 = vi5x4567;
403*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi6x2345 = vext_f16(vi6x0123, vi6x4567, 2);
404*4bdc9457SAndroid Build Coastguard Worker       vi6x0123 = vi6x4567;
405*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi7x2345 = vext_f16(vi7x0123, vi7x4567, 2);
406*4bdc9457SAndroid Build Coastguard Worker       vi7x0123 = vi7x4567;
407*4bdc9457SAndroid Build Coastguard Worker 
408*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x2345, vw01234567, 1);
409*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x2345, vw01234567, 1);
410*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x2345, vw01234567, 1);
411*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x2345, vw01234567, 1);
412*4bdc9457SAndroid Build Coastguard Worker 
413*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x2345, vw01234567, 6);
414*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x2345, vw01234567, 6);
415*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x2345, vw01234567, 6);
416*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x2345, vw01234567, 6);
417*4bdc9457SAndroid Build Coastguard Worker 
418*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3);
419*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x2345, vw89ABCDEF, 3);
420*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x2345, vw89ABCDEF, 3);
421*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x2345, vw89ABCDEF, 3);
422*4bdc9457SAndroid Build Coastguard Worker 
423*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x2345, vwGHIJKLMN, 0);
424*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x2345, vwGHIJKLMN, 0);
425*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x2345, vwGHIJKLMN, 0);
426*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x2345, vwGHIJKLMN, 0);
427*4bdc9457SAndroid Build Coastguard Worker 
428*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi4x2345, vwGHIJKLMN, 5);
429*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi5x2345, vwGHIJKLMN, 5);
430*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi6x2345, vwGHIJKLMN, 5);
431*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi7x2345, vwGHIJKLMN, 5);
432*4bdc9457SAndroid Build Coastguard Worker 
433*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi0x5678 = vext_f16(vi0x4567, vi0x89AB, 1);
434*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi1x5678 = vext_f16(vi1x4567, vi1x89AB, 1);
435*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi2x5678 = vext_f16(vi2x4567, vi2x89AB, 1);
436*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi3x5678 = vext_f16(vi3x4567, vi3x89AB, 1);
437*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi4x5678 = vext_f16(vi4x4567, vi4x89AB, 1);
438*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi5x5678 = vext_f16(vi5x4567, vi5x89AB, 1);
439*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi6x5678 = vext_f16(vi6x4567, vi6x89AB, 1);
440*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi7x5678 = vext_f16(vi7x4567, vi7x89AB, 1);
441*4bdc9457SAndroid Build Coastguard Worker 
442*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x5678, vw01234567, 4);
443*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x5678, vw01234567, 4);
444*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x5678, vw01234567, 4);
445*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x5678, vw01234567, 4);
446*4bdc9457SAndroid Build Coastguard Worker 
447*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x5678, vw89ABCDEF, 1);
448*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x5678, vw89ABCDEF, 1);
449*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x5678, vw89ABCDEF, 1);
450*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x5678, vw89ABCDEF, 1);
451*4bdc9457SAndroid Build Coastguard Worker 
452*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x5678, vw89ABCDEF, 6);
453*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x5678, vw89ABCDEF, 6);
454*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x5678, vw89ABCDEF, 6);
455*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x5678, vw89ABCDEF, 6);
456*4bdc9457SAndroid Build Coastguard Worker 
457*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x5678, vwGHIJKLMN, 3);
458*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x5678, vwGHIJKLMN, 3);
459*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x5678, vwGHIJKLMN, 3);
460*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x5678, vwGHIJKLMN, 3);
461*4bdc9457SAndroid Build Coastguard Worker 
462*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_lane_f16(vo0p0, vi4x5678, vwOP, 0);
463*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_lane_f16(vo1p0, vi5x5678, vwOP, 0);
464*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_lane_f16(vo2p0, vi6x5678, vwOP, 0);
465*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_lane_f16(vo3p0, vi7x5678, vwOP, 0);
466*4bdc9457SAndroid Build Coastguard Worker 
467*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi0x6789 = vext_f16(vi0x4567, vi0x89AB, 2);
468*4bdc9457SAndroid Build Coastguard Worker       vi0x4567 = vi0x89AB;
469*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi1x6789 = vext_f16(vi1x4567, vi1x89AB, 2);
470*4bdc9457SAndroid Build Coastguard Worker       vi1x4567 = vi1x89AB;
471*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi2x6789 = vext_f16(vi2x4567, vi2x89AB, 2);
472*4bdc9457SAndroid Build Coastguard Worker       vi2x4567 = vi2x89AB;
473*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi3x6789 = vext_f16(vi3x4567, vi3x89AB, 2);
474*4bdc9457SAndroid Build Coastguard Worker       vi3x4567 = vi3x89AB;
475*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi4x6789 = vext_f16(vi4x4567, vi4x89AB, 2);
476*4bdc9457SAndroid Build Coastguard Worker       vi4x4567 = vi4x89AB;
477*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi5x6789 = vext_f16(vi5x4567, vi5x89AB, 2);
478*4bdc9457SAndroid Build Coastguard Worker       vi5x4567 = vi5x89AB;
479*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi6x6789 = vext_f16(vi6x4567, vi6x89AB, 2);
480*4bdc9457SAndroid Build Coastguard Worker       vi6x4567 = vi6x89AB;
481*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi7x6789 = vext_f16(vi7x4567, vi7x89AB, 2);
482*4bdc9457SAndroid Build Coastguard Worker       vi7x4567 = vi7x89AB;
483*4bdc9457SAndroid Build Coastguard Worker 
484*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x6789, vw01234567, 5);
485*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x6789, vw01234567, 5);
486*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x6789, vw01234567, 5);
487*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x6789, vw01234567, 5);
488*4bdc9457SAndroid Build Coastguard Worker 
489*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x6789, vw89ABCDEF, 2);
490*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x6789, vw89ABCDEF, 2);
491*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x6789, vw89ABCDEF, 2);
492*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x6789, vw89ABCDEF, 2);
493*4bdc9457SAndroid Build Coastguard Worker 
494*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x6789, vw89ABCDEF, 7);
495*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x6789, vw89ABCDEF, 7);
496*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x6789, vw89ABCDEF, 7);
497*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x6789, vw89ABCDEF, 7);
498*4bdc9457SAndroid Build Coastguard Worker 
499*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x6789, vwGHIJKLMN, 4);
500*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x6789, vwGHIJKLMN, 4);
501*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x6789, vwGHIJKLMN, 4);
502*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x6789, vwGHIJKLMN, 4);
503*4bdc9457SAndroid Build Coastguard Worker 
504*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_lane_f16(vo0p0, vi4x6789, vwOP, 1);
505*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_lane_f16(vo1p0, vi5x6789, vwOP, 1);
506*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_lane_f16(vo2p0, vi6x6789, vwOP, 1);
507*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_lane_f16(vo3p0, vi7x6789, vwOP, 1);
508*4bdc9457SAndroid Build Coastguard Worker 
509*4bdc9457SAndroid Build Coastguard Worker 
510*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo0 = vmax_f16(vo0p0, vmin);
511*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo1 = vmax_f16(vo1p0, vmin);
512*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo2 = vmax_f16(vo2p0, vmin);
513*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo3 = vmax_f16(vo3p0, vmin);
514*4bdc9457SAndroid Build Coastguard Worker 
515*4bdc9457SAndroid Build Coastguard Worker       vo0 = vmin_f16(vo0, vmax);
516*4bdc9457SAndroid Build Coastguard Worker       vo1 = vmin_f16(vo1, vmax);
517*4bdc9457SAndroid Build Coastguard Worker       vo2 = vmin_f16(vo2, vmax);
518*4bdc9457SAndroid Build Coastguard Worker       vo3 = vmin_f16(vo3, vmax);
519*4bdc9457SAndroid Build Coastguard Worker 
520*4bdc9457SAndroid Build Coastguard Worker       vst1_f16(o3, vo3); o3 += 4;
521*4bdc9457SAndroid Build Coastguard Worker       vst1_f16(o2, vo2); o2 += 4;
522*4bdc9457SAndroid Build Coastguard Worker       vst1_f16(o1, vo1); o1 += 4;
523*4bdc9457SAndroid Build Coastguard Worker       vst1_f16(o0, vo0); o0 += 4;
524*4bdc9457SAndroid Build Coastguard Worker 
525*4bdc9457SAndroid Build Coastguard Worker       w -= 4 * sizeof(__fp16);
526*4bdc9457SAndroid Build Coastguard Worker     }
527*4bdc9457SAndroid Build Coastguard Worker     assert(w >= 1 * sizeof(__fp16));
528*4bdc9457SAndroid Build Coastguard Worker     assert(w <= 4 * sizeof(__fp16));
529*4bdc9457SAndroid Build Coastguard Worker     {
530*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo0p0 = vdup_laneq_f16(vw01234567, 0);
531*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo1p0 = vdup_laneq_f16(vw01234567, 0);
532*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo2p0 = vdup_laneq_f16(vw01234567, 0);
533*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo3p0 = vdup_laneq_f16(vw01234567, 0);
534*4bdc9457SAndroid Build Coastguard Worker 
535*4bdc9457SAndroid Build Coastguard Worker       vi0x4567 = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi0x4567)));
536*4bdc9457SAndroid Build Coastguard Worker       vi1x4567 = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi1x4567)));
537*4bdc9457SAndroid Build Coastguard Worker       vi2x4567 = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi2x4567)));
538*4bdc9457SAndroid Build Coastguard Worker       vi3x4567 = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi3x4567)));
539*4bdc9457SAndroid Build Coastguard Worker       vi4x4567 = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi4x4567)));
540*4bdc9457SAndroid Build Coastguard Worker       vi5x4567 = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi5x4567)));
541*4bdc9457SAndroid Build Coastguard Worker       vi6x4567 = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi6x4567)));
542*4bdc9457SAndroid Build Coastguard Worker       vi7x4567 = vreinterpret_f16_u16(vand_u16(vmask, vreinterpret_u16_f16(vi7x4567)));
543*4bdc9457SAndroid Build Coastguard Worker 
544*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x4567, vw01234567, 3);
545*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x4567, vw01234567, 3);
546*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x4567, vw01234567, 3);
547*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x4567, vw01234567, 3);
548*4bdc9457SAndroid Build Coastguard Worker 
549*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x4567, vw89ABCDEF, 0);
550*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x4567, vw89ABCDEF, 0);
551*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x4567, vw89ABCDEF, 0);
552*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x4567, vw89ABCDEF, 0);
553*4bdc9457SAndroid Build Coastguard Worker 
554*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x4567, vw89ABCDEF, 5);
555*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x4567, vw89ABCDEF, 5);
556*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x4567, vw89ABCDEF, 5);
557*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x4567, vw89ABCDEF, 5);
558*4bdc9457SAndroid Build Coastguard Worker 
559*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x4567, vwGHIJKLMN, 2);
560*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x4567, vwGHIJKLMN, 2);
561*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x4567, vwGHIJKLMN, 2);
562*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x4567, vwGHIJKLMN, 2);
563*4bdc9457SAndroid Build Coastguard Worker 
564*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi4x4567, vwGHIJKLMN, 7);
565*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi5x4567, vwGHIJKLMN, 7);
566*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi6x4567, vwGHIJKLMN, 7);
567*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi7x4567, vwGHIJKLMN, 7);
568*4bdc9457SAndroid Build Coastguard Worker 
569*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi0x3456 = vext_f16(vi0x0123, vi0x4567, 3);
570*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi1x3456 = vext_f16(vi1x0123, vi1x4567, 3);
571*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi2x3456 = vext_f16(vi2x0123, vi2x4567, 3);
572*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi3x3456 = vext_f16(vi3x0123, vi3x4567, 3);
573*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi4x3456 = vext_f16(vi4x0123, vi4x4567, 3);
574*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi5x3456 = vext_f16(vi5x0123, vi5x4567, 3);
575*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi6x3456 = vext_f16(vi6x0123, vi6x4567, 3);
576*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi7x3456 = vext_f16(vi7x0123, vi7x4567, 3);
577*4bdc9457SAndroid Build Coastguard Worker 
578*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x3456, vw01234567, 2);
579*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x3456, vw01234567, 2);
580*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x3456, vw01234567, 2);
581*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x3456, vw01234567, 2);
582*4bdc9457SAndroid Build Coastguard Worker 
583*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x3456, vw01234567, 7);
584*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x3456, vw01234567, 7);
585*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x3456, vw01234567, 7);
586*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x3456, vw01234567, 7);
587*4bdc9457SAndroid Build Coastguard Worker 
588*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x3456, vw89ABCDEF, 4);
589*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x3456, vw89ABCDEF, 4);
590*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x3456, vw89ABCDEF, 4);
591*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x3456, vw89ABCDEF, 4);
592*4bdc9457SAndroid Build Coastguard Worker 
593*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x3456, vwGHIJKLMN, 1);
594*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x3456, vwGHIJKLMN, 1);
595*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x3456, vwGHIJKLMN, 1);
596*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x3456, vwGHIJKLMN, 1);
597*4bdc9457SAndroid Build Coastguard Worker 
598*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi4x3456, vwGHIJKLMN, 6);
599*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi5x3456, vwGHIJKLMN, 6);
600*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi6x3456, vwGHIJKLMN, 6);
601*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi7x3456, vwGHIJKLMN, 6);
602*4bdc9457SAndroid Build Coastguard Worker 
603*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi0x2345 = vext_f16(vi0x0123, vi0x4567, 2);
604*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi1x2345 = vext_f16(vi1x0123, vi1x4567, 2);
605*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi2x2345 = vext_f16(vi2x0123, vi2x4567, 2);
606*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi3x2345 = vext_f16(vi3x0123, vi3x4567, 2);
607*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi4x2345 = vext_f16(vi4x0123, vi4x4567, 2);
608*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi5x2345 = vext_f16(vi5x0123, vi5x4567, 2);
609*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi6x2345 = vext_f16(vi6x0123, vi6x4567, 2);
610*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi7x2345 = vext_f16(vi7x0123, vi7x4567, 2);
611*4bdc9457SAndroid Build Coastguard Worker 
612*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x2345, vw01234567, 1);
613*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x2345, vw01234567, 1);
614*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x2345, vw01234567, 1);
615*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x2345, vw01234567, 1);
616*4bdc9457SAndroid Build Coastguard Worker 
617*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x2345, vw01234567, 6);
618*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x2345, vw01234567, 6);
619*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x2345, vw01234567, 6);
620*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x2345, vw01234567, 6);
621*4bdc9457SAndroid Build Coastguard Worker 
622*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x2345, vw89ABCDEF, 3);
623*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x2345, vw89ABCDEF, 3);
624*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x2345, vw89ABCDEF, 3);
625*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x2345, vw89ABCDEF, 3);
626*4bdc9457SAndroid Build Coastguard Worker 
627*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x2345, vwGHIJKLMN, 0);
628*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x2345, vwGHIJKLMN, 0);
629*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x2345, vwGHIJKLMN, 0);
630*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x2345, vwGHIJKLMN, 0);
631*4bdc9457SAndroid Build Coastguard Worker 
632*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi4x2345, vwGHIJKLMN, 5);
633*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi5x2345, vwGHIJKLMN, 5);
634*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi6x2345, vwGHIJKLMN, 5);
635*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi7x2345, vwGHIJKLMN, 5);
636*4bdc9457SAndroid Build Coastguard Worker 
637*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vzero = vmov_n_f16(0);
638*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi0x5678 = vext_f16(vi0x4567, vzero, 1);
639*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi1x5678 = vext_f16(vi1x4567, vzero, 1);
640*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi2x5678 = vext_f16(vi2x4567, vzero, 1);
641*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi3x5678 = vext_f16(vi3x4567, vzero, 1);
642*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi4x5678 = vext_f16(vi4x4567, vzero, 1);
643*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi5x5678 = vext_f16(vi5x4567, vzero, 1);
644*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi6x5678 = vext_f16(vi6x4567, vzero, 1);
645*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi7x5678 = vext_f16(vi7x4567, vzero, 1);
646*4bdc9457SAndroid Build Coastguard Worker 
647*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x5678, vw01234567, 4);
648*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x5678, vw01234567, 4);
649*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x5678, vw01234567, 4);
650*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x5678, vw01234567, 4);
651*4bdc9457SAndroid Build Coastguard Worker 
652*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x5678, vw89ABCDEF, 1);
653*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x5678, vw89ABCDEF, 1);
654*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x5678, vw89ABCDEF, 1);
655*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x5678, vw89ABCDEF, 1);
656*4bdc9457SAndroid Build Coastguard Worker 
657*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x5678, vw89ABCDEF, 6);
658*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x5678, vw89ABCDEF, 6);
659*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x5678, vw89ABCDEF, 6);
660*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x5678, vw89ABCDEF, 6);
661*4bdc9457SAndroid Build Coastguard Worker 
662*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x5678, vwGHIJKLMN, 3);
663*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x5678, vwGHIJKLMN, 3);
664*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x5678, vwGHIJKLMN, 3);
665*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x5678, vwGHIJKLMN, 3);
666*4bdc9457SAndroid Build Coastguard Worker 
667*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_lane_f16(vo0p0, vi4x5678, vwOP, 0);
668*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_lane_f16(vo1p0, vi5x5678, vwOP, 0);
669*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_lane_f16(vo2p0, vi6x5678, vwOP, 0);
670*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_lane_f16(vo3p0, vi7x5678, vwOP, 0);
671*4bdc9457SAndroid Build Coastguard Worker 
672*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi0x6789 = vext_f16(vi0x5678, vzero, 1);
673*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi1x6789 = vext_f16(vi1x5678, vzero, 1);
674*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi2x6789 = vext_f16(vi2x5678, vzero, 1);
675*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi3x6789 = vext_f16(vi3x5678, vzero, 1);
676*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi4x6789 = vext_f16(vi4x5678, vzero, 1);
677*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi5x6789 = vext_f16(vi5x5678, vzero, 1);
678*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi6x6789 = vext_f16(vi6x5678, vzero, 1);
679*4bdc9457SAndroid Build Coastguard Worker       const float16x4_t vi7x6789 = vext_f16(vi7x5678, vzero, 1);
680*4bdc9457SAndroid Build Coastguard Worker 
681*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi0x6789, vw01234567, 5);
682*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi1x6789, vw01234567, 5);
683*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi2x6789, vw01234567, 5);
684*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi3x6789, vw01234567, 5);
685*4bdc9457SAndroid Build Coastguard Worker 
686*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi1x6789, vw89ABCDEF, 2);
687*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi2x6789, vw89ABCDEF, 2);
688*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi3x6789, vw89ABCDEF, 2);
689*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi4x6789, vw89ABCDEF, 2);
690*4bdc9457SAndroid Build Coastguard Worker 
691*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi2x6789, vw89ABCDEF, 7);
692*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi3x6789, vw89ABCDEF, 7);
693*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi4x6789, vw89ABCDEF, 7);
694*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi5x6789, vw89ABCDEF, 7);
695*4bdc9457SAndroid Build Coastguard Worker 
696*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_laneq_f16(vo0p0, vi3x6789, vwGHIJKLMN, 4);
697*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_laneq_f16(vo1p0, vi4x6789, vwGHIJKLMN, 4);
698*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_laneq_f16(vo2p0, vi5x6789, vwGHIJKLMN, 4);
699*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_laneq_f16(vo3p0, vi6x6789, vwGHIJKLMN, 4);
700*4bdc9457SAndroid Build Coastguard Worker 
701*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vfma_lane_f16(vo0p0, vi4x6789, vwOP, 1);
702*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vfma_lane_f16(vo1p0, vi5x6789, vwOP, 1);
703*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vfma_lane_f16(vo2p0, vi6x6789, vwOP, 1);
704*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vfma_lane_f16(vo3p0, vi7x6789, vwOP, 1);
705*4bdc9457SAndroid Build Coastguard Worker 
706*4bdc9457SAndroid Build Coastguard Worker 
707*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo0 = vmax_f16(vo0p0, vmin);
708*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo1 = vmax_f16(vo1p0, vmin);
709*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo2 = vmax_f16(vo2p0, vmin);
710*4bdc9457SAndroid Build Coastguard Worker       float16x4_t vo3 = vmax_f16(vo3p0, vmin);
711*4bdc9457SAndroid Build Coastguard Worker 
712*4bdc9457SAndroid Build Coastguard Worker       vo0 = vmin_f16(vo0, vmax);
713*4bdc9457SAndroid Build Coastguard Worker       vo1 = vmin_f16(vo1, vmax);
714*4bdc9457SAndroid Build Coastguard Worker       vo2 = vmin_f16(vo2, vmax);
715*4bdc9457SAndroid Build Coastguard Worker       vo3 = vmin_f16(vo3, vmax);
716*4bdc9457SAndroid Build Coastguard Worker 
717*4bdc9457SAndroid Build Coastguard Worker       if XNN_LIKELY(w & (4 * sizeof(__fp16))) {
718*4bdc9457SAndroid Build Coastguard Worker         vst1_f16(o3, vo3); o3 += 4;
719*4bdc9457SAndroid Build Coastguard Worker         vst1_f16(o2, vo2); o2 += 4;
720*4bdc9457SAndroid Build Coastguard Worker         vst1_f16(o1, vo1); o1 += 4;
721*4bdc9457SAndroid Build Coastguard Worker         vst1_f16(o0, vo0); o0 += 4;
722*4bdc9457SAndroid Build Coastguard Worker       } else {
723*4bdc9457SAndroid Build Coastguard Worker         if (w & (2 * sizeof(__fp16))) {
724*4bdc9457SAndroid Build Coastguard Worker           vst1_lane_u32((void*) o3, vreinterpret_u32_f16(vo3), 0); o3 += 2;
725*4bdc9457SAndroid Build Coastguard Worker           vst1_lane_u32((void*) o2, vreinterpret_u32_f16(vo2), 0); o2 += 2;
726*4bdc9457SAndroid Build Coastguard Worker           vst1_lane_u32((void*) o1, vreinterpret_u32_f16(vo1), 0); o1 += 2;
727*4bdc9457SAndroid Build Coastguard Worker           vst1_lane_u32((void*) o0, vreinterpret_u32_f16(vo0), 0); o0 += 2;
728*4bdc9457SAndroid Build Coastguard Worker 
729*4bdc9457SAndroid Build Coastguard Worker           vo0 = vext_f16(vo0, vo0, 2);
730*4bdc9457SAndroid Build Coastguard Worker           vo1 = vext_f16(vo1, vo1, 2);
731*4bdc9457SAndroid Build Coastguard Worker           vo2 = vext_f16(vo2, vo2, 2);
732*4bdc9457SAndroid Build Coastguard Worker           vo3 = vext_f16(vo3, vo3, 2);
733*4bdc9457SAndroid Build Coastguard Worker         }
734*4bdc9457SAndroid Build Coastguard Worker         if (w & (1 * sizeof(__fp16))) {
735*4bdc9457SAndroid Build Coastguard Worker           vst1_lane_f16(o3, vo3, 0); o3 += 1;
736*4bdc9457SAndroid Build Coastguard Worker           vst1_lane_f16(o2, vo2, 0); o2 += 1;
737*4bdc9457SAndroid Build Coastguard Worker           vst1_lane_f16(o1, vo1, 0); o1 += 1;
738*4bdc9457SAndroid Build Coastguard Worker           vst1_lane_f16(o0, vo0, 0); o0 += 1;
739*4bdc9457SAndroid Build Coastguard Worker         }
740*4bdc9457SAndroid Build Coastguard Worker       }
741*4bdc9457SAndroid Build Coastguard Worker     }
742*4bdc9457SAndroid Build Coastguard Worker 
743*4bdc9457SAndroid Build Coastguard Worker     i0 = (const __fp16*) ((uintptr_t) i4 - input_decrement);
744*4bdc9457SAndroid Build Coastguard Worker     i1 = (const __fp16*) ((uintptr_t) i5 - input_decrement);
745*4bdc9457SAndroid Build Coastguard Worker     i2 = (const __fp16*) ((uintptr_t) i1 + input_width);
746*4bdc9457SAndroid Build Coastguard Worker     i3 = (const __fp16*) ((uintptr_t) i2 + input_width);
747*4bdc9457SAndroid Build Coastguard Worker     i4 = (const __fp16*) ((uintptr_t) i3 + input_width);
748*4bdc9457SAndroid Build Coastguard Worker     i5 = (const __fp16*) ((uintptr_t) i4 + input_width);
749*4bdc9457SAndroid Build Coastguard Worker     i6 = (const __fp16*) ((uintptr_t) i5 + input_width);
750*4bdc9457SAndroid Build Coastguard Worker     i7 = (const __fp16*) ((uintptr_t) i6 + input_width);
751*4bdc9457SAndroid Build Coastguard Worker 
752*4bdc9457SAndroid Build Coastguard Worker     o0 = o3;
753*4bdc9457SAndroid Build Coastguard Worker     o1 = (__fp16*) ((uintptr_t) o0 + input_width);
754*4bdc9457SAndroid Build Coastguard Worker     o2 = (__fp16*) ((uintptr_t) o1 + input_width);
755*4bdc9457SAndroid Build Coastguard Worker     o3 = (__fp16*) ((uintptr_t) o2 + input_width);
756*4bdc9457SAndroid Build Coastguard Worker 
757*4bdc9457SAndroid Build Coastguard Worker     output_height = doz(output_height, 4);
758*4bdc9457SAndroid Build Coastguard Worker   } while (output_height != 0);
759*4bdc9457SAndroid Build Coastguard Worker }
760