xref: /aosp_15_r20/external/XNNPACK/src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f32-dwconv2d-chw/5x5p2-neon.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_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])18*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_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 float* input,
22*4bdc9457SAndroid Build Coastguard Worker     const float* weights,
23*4bdc9457SAndroid Build Coastguard Worker     const float* zero,
24*4bdc9457SAndroid Build Coastguard Worker     float* output,
25*4bdc9457SAndroid Build Coastguard Worker     uint32_t padding_top,
26*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f32_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(float) == 0);
31*4bdc9457SAndroid Build Coastguard Worker   assert(padding_top == 2);
32*4bdc9457SAndroid Build Coastguard Worker 
33*4bdc9457SAndroid Build Coastguard Worker   const uint32x4_t vmask = vld1q_u32(params->neon.mask);
34*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vmax = vld1q_dup_f32(&params->neon.max);
35*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vmin = vld1q_dup_f32(&params->neon.min);
36*4bdc9457SAndroid Build Coastguard Worker 
37*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vw0123 = vld1q_f32(weights);
38*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vw4567 = vld1q_f32(weights + 4);
39*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vw89AB = vld1q_f32(weights + 8);
40*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vwCDEF = vld1q_f32(weights + 12);
41*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vwGHIJ = vld1q_f32(weights + 16);
42*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vwKLMN = vld1q_f32(weights + 20);
43*4bdc9457SAndroid Build Coastguard Worker   const float32x2_t vwOP = vld1_f32(weights + 24);
44*4bdc9457SAndroid Build Coastguard Worker 
45*4bdc9457SAndroid Build Coastguard Worker   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
46*4bdc9457SAndroid Build Coastguard Worker 
47*4bdc9457SAndroid Build Coastguard Worker   const float* i0 = zero;
48*4bdc9457SAndroid Build Coastguard Worker   const float* i1 = zero;
49*4bdc9457SAndroid Build Coastguard Worker   const float* i2 = input;
50*4bdc9457SAndroid Build Coastguard Worker   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
51*4bdc9457SAndroid Build Coastguard Worker   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
52*4bdc9457SAndroid Build Coastguard Worker   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
53*4bdc9457SAndroid Build Coastguard Worker   const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
54*4bdc9457SAndroid Build Coastguard Worker   const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
55*4bdc9457SAndroid Build Coastguard Worker 
56*4bdc9457SAndroid Build Coastguard Worker   float* o0 = output;
57*4bdc9457SAndroid Build Coastguard Worker   float* o1 = (float*) ((uintptr_t) o0 + input_width);
58*4bdc9457SAndroid Build Coastguard Worker   float* o2 = (float*) ((uintptr_t) o1 + input_width);
59*4bdc9457SAndroid Build Coastguard Worker   float* o3 = (float*) ((uintptr_t) o2 + input_width);
60*4bdc9457SAndroid Build Coastguard Worker 
61*4bdc9457SAndroid Build Coastguard Worker   size_t output_height = input_height;
62*4bdc9457SAndroid Build Coastguard Worker   do {
63*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 2) {
64*4bdc9457SAndroid Build Coastguard Worker       i3 = zero;
65*4bdc9457SAndroid Build Coastguard Worker       o1 = o0;
66*4bdc9457SAndroid Build Coastguard Worker     }
67*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 3) {
68*4bdc9457SAndroid Build Coastguard Worker       i4 = zero;
69*4bdc9457SAndroid Build Coastguard Worker       o2 = o1;
70*4bdc9457SAndroid Build Coastguard Worker     }
71*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 4) {
72*4bdc9457SAndroid Build Coastguard Worker       i5 = zero;
73*4bdc9457SAndroid Build Coastguard Worker       o3 = o2;
74*4bdc9457SAndroid Build Coastguard Worker     }
75*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 5) {
76*4bdc9457SAndroid Build Coastguard Worker       i6 = zero;
77*4bdc9457SAndroid Build Coastguard Worker     }
78*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 6) {
79*4bdc9457SAndroid Build Coastguard Worker       i7 = zero;
80*4bdc9457SAndroid Build Coastguard Worker     }
81*4bdc9457SAndroid Build Coastguard Worker 
82*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi0x0123 = vmovq_n_f32(0.0f);
83*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi1x0123 = vmovq_n_f32(0.0f);
84*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi2x0123 = vmovq_n_f32(0.0f);
85*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi3x0123 = vmovq_n_f32(0.0f);
86*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi4x0123 = vmovq_n_f32(0.0f);
87*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi5x0123 = vmovq_n_f32(0.0f);
88*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi6x0123 = vmovq_n_f32(0.0f);
89*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi7x0123 = vmovq_n_f32(0.0f);
90*4bdc9457SAndroid Build Coastguard Worker 
91*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
92*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
93*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
94*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;
95*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4;
96*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi5x4567 = vld1q_f32(i5); i5 += 4;
97*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi6x4567 = vld1q_f32(i6); i6 += 4;
98*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vi7x4567 = vld1q_f32(i7); i7 += 4;
99*4bdc9457SAndroid Build Coastguard Worker 
100*4bdc9457SAndroid Build Coastguard Worker     size_t w = input_width;
101*4bdc9457SAndroid Build Coastguard Worker     for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
102*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
103*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
104*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
105*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo3p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
106*4bdc9457SAndroid Build Coastguard Worker 
107*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
108*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4;
109*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi2x89AB = vld1q_f32(i2); i2 += 4;
110*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4;
111*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4;
112*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi5x89AB = vld1q_f32(i5); i5 += 4;
113*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi6x89AB = vld1q_f32(i6); i6 += 4;
114*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi7x89AB = vld1q_f32(i7); i7 += 4;
115*4bdc9457SAndroid Build Coastguard Worker 
116*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1);
117*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1);
118*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x4567, vget_high_f32(vw0123), 1);
119*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x4567, vget_high_f32(vw0123), 1);
120*4bdc9457SAndroid Build Coastguard Worker 
121*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0);
122*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw89AB), 0);
123*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x4567, vget_low_f32(vw89AB), 0);
124*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x4567, vget_low_f32(vw89AB), 0);
125*4bdc9457SAndroid Build Coastguard Worker 
126*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1);
127*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vget_low_f32(vwCDEF), 1);
128*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x4567, vget_low_f32(vwCDEF), 1);
129*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x4567, vget_low_f32(vwCDEF), 1);
130*4bdc9457SAndroid Build Coastguard Worker 
131*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0);
132*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0);
133*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x4567, vget_high_f32(vwGHIJ), 0);
134*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x4567, vget_high_f32(vwGHIJ), 0);
135*4bdc9457SAndroid Build Coastguard Worker 
136*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1);
137*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1);
138*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1);
139*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x4567, vget_high_f32(vwKLMN), 1);
140*4bdc9457SAndroid Build Coastguard Worker 
141*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
142*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
143*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
144*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
145*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);
146*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);
147*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi6x3456 = vextq_f32(vi6x0123, vi6x4567, 3);
148*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3);
149*4bdc9457SAndroid Build Coastguard Worker 
150*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0);
151*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x3456, vget_high_f32(vw0123), 0);
152*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x3456, vget_high_f32(vw0123), 0);
153*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x3456, vget_high_f32(vw0123), 0);
154*4bdc9457SAndroid Build Coastguard Worker 
155*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1);
156*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_high_f32(vw4567), 1);
157*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x3456, vget_high_f32(vw4567), 1);
158*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x3456, vget_high_f32(vw4567), 1);
159*4bdc9457SAndroid Build Coastguard Worker 
160*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0);
161*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x3456, vget_low_f32(vwCDEF), 0);
162*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x3456, vget_low_f32(vwCDEF), 0);
163*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0);
164*4bdc9457SAndroid Build Coastguard Worker 
165*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1);
166*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1);
167*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1);
168*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x3456, vget_low_f32(vwGHIJ), 1);
169*4bdc9457SAndroid Build Coastguard Worker 
170*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0);
171*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0);
172*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x3456, vget_high_f32(vwKLMN), 0);
173*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x3456, vget_high_f32(vwKLMN), 0);
174*4bdc9457SAndroid Build Coastguard Worker 
175*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2);
176*4bdc9457SAndroid Build Coastguard Worker       vi0x0123 = vi0x4567;
177*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi1x2345 = vextq_f32(vi1x0123, vi1x4567, 2);
178*4bdc9457SAndroid Build Coastguard Worker       vi1x0123 = vi1x4567;
179*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2);
180*4bdc9457SAndroid Build Coastguard Worker       vi2x0123 = vi2x4567;
181*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);
182*4bdc9457SAndroid Build Coastguard Worker       vi3x0123 = vi3x4567;
183*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);
184*4bdc9457SAndroid Build Coastguard Worker       vi4x0123 = vi4x4567;
185*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);
186*4bdc9457SAndroid Build Coastguard Worker       vi5x0123 = vi5x4567;
187*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi6x2345 = vextq_f32(vi6x0123, vi6x4567, 2);
188*4bdc9457SAndroid Build Coastguard Worker       vi6x0123 = vi6x4567;
189*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi7x2345 = vextq_f32(vi7x0123, vi7x4567, 2);
190*4bdc9457SAndroid Build Coastguard Worker       vi7x0123 = vi7x4567;
191*4bdc9457SAndroid Build Coastguard Worker 
192*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1);
193*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x2345, vget_low_f32(vw0123), 1);
194*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1);
195*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x2345, vget_low_f32(vw0123), 1);
196*4bdc9457SAndroid Build Coastguard Worker 
197*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x2345, vget_high_f32(vw4567), 0);
198*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0);
199*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x2345, vget_high_f32(vw4567), 0);
200*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x2345, vget_high_f32(vw4567), 0);
201*4bdc9457SAndroid Build Coastguard Worker 
202*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1);
203*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x2345, vget_high_f32(vw89AB), 1);
204*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x2345, vget_high_f32(vw89AB), 1);
205*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x2345, vget_high_f32(vw89AB), 1);
206*4bdc9457SAndroid Build Coastguard Worker 
207*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0);
208*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x2345, vget_low_f32(vwGHIJ), 0);
209*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x2345, vget_low_f32(vwGHIJ), 0);
210*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x2345, vget_low_f32(vwGHIJ), 0);
211*4bdc9457SAndroid Build Coastguard Worker 
212*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1);
213*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x2345, vget_low_f32(vwKLMN), 1);
214*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x2345, vget_low_f32(vwKLMN), 1);
215*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x2345, vget_low_f32(vwKLMN), 1);
216*4bdc9457SAndroid Build Coastguard Worker 
217*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1);
218*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vi1x89AB, 1);
219*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vi2x89AB, 1);
220*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);
221*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1);
222*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1);
223*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1);
224*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1);
225*4bdc9457SAndroid Build Coastguard Worker 
226*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0);
227*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x5678, vget_low_f32(vw4567), 0);
228*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x5678, vget_low_f32(vw4567), 0);
229*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x5678, vget_low_f32(vw4567), 0);
230*4bdc9457SAndroid Build Coastguard Worker 
231*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1);
232*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x5678, vget_low_f32(vw89AB), 1);
233*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x5678, vget_low_f32(vw89AB), 1);
234*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x5678, vget_low_f32(vw89AB), 1);
235*4bdc9457SAndroid Build Coastguard Worker 
236*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vget_high_f32(vwCDEF), 0);
237*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x5678, vget_high_f32(vwCDEF), 0);
238*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x5678, vget_high_f32(vwCDEF), 0);
239*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vget_high_f32(vwCDEF), 0);
240*4bdc9457SAndroid Build Coastguard Worker 
241*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1);
242*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x5678, vget_high_f32(vwGHIJ), 1);
243*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1);
244*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x5678, vget_high_f32(vwGHIJ), 1);
245*4bdc9457SAndroid Build Coastguard Worker 
246*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x5678, vwOP, 0);
247*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x5678, vwOP, 0);
248*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x5678, vwOP, 0);
249*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x5678, vwOP, 0);
250*4bdc9457SAndroid Build Coastguard Worker 
251*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2);
252*4bdc9457SAndroid Build Coastguard Worker       vi0x4567 = vi0x89AB;
253*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi1x6789 = vextq_f32(vi1x4567, vi1x89AB, 2);
254*4bdc9457SAndroid Build Coastguard Worker       vi1x4567 = vi1x89AB;
255*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2);
256*4bdc9457SAndroid Build Coastguard Worker       vi2x4567 = vi2x89AB;
257*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);
258*4bdc9457SAndroid Build Coastguard Worker       vi3x4567 = vi3x89AB;
259*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2);
260*4bdc9457SAndroid Build Coastguard Worker       vi4x4567 = vi4x89AB;
261*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2);
262*4bdc9457SAndroid Build Coastguard Worker       vi5x4567 = vi5x89AB;
263*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2);
264*4bdc9457SAndroid Build Coastguard Worker       vi6x4567 = vi6x89AB;
265*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi7x6789 = vextq_f32(vi7x4567, vi7x89AB, 2);
266*4bdc9457SAndroid Build Coastguard Worker       vi7x4567 = vi7x89AB;
267*4bdc9457SAndroid Build Coastguard Worker 
268*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1);
269*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x6789, vget_low_f32(vw4567), 1);
270*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1);
271*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x6789, vget_low_f32(vw4567), 1);
272*4bdc9457SAndroid Build Coastguard Worker 
273*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x6789, vget_high_f32(vw89AB), 0);
274*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0);
275*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x6789, vget_high_f32(vw89AB), 0);
276*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x6789, vget_high_f32(vw89AB), 0);
277*4bdc9457SAndroid Build Coastguard Worker 
278*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1);
279*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x6789, vget_high_f32(vwCDEF), 1);
280*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1);
281*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1);
282*4bdc9457SAndroid Build Coastguard Worker 
283*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0);
284*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0);
285*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0);
286*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x6789, vget_low_f32(vwKLMN), 0);
287*4bdc9457SAndroid Build Coastguard Worker 
288*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1);
289*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1);
290*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x6789, vwOP, 1);
291*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x6789, vwOP, 1);
292*4bdc9457SAndroid Build Coastguard Worker 
293*4bdc9457SAndroid Build Coastguard Worker 
294*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
295*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
296*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo2 = vmaxq_f32(vo2p0, vmin);
297*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo3 = vmaxq_f32(vo3p0, vmin);
298*4bdc9457SAndroid Build Coastguard Worker 
299*4bdc9457SAndroid Build Coastguard Worker       vo0 = vminq_f32(vo0, vmax);
300*4bdc9457SAndroid Build Coastguard Worker       vo1 = vminq_f32(vo1, vmax);
301*4bdc9457SAndroid Build Coastguard Worker       vo2 = vminq_f32(vo2, vmax);
302*4bdc9457SAndroid Build Coastguard Worker       vo3 = vminq_f32(vo3, vmax);
303*4bdc9457SAndroid Build Coastguard Worker 
304*4bdc9457SAndroid Build Coastguard Worker       vst1q_f32(o3, vo3); o3 += 4;
305*4bdc9457SAndroid Build Coastguard Worker       vst1q_f32(o2, vo2); o2 += 4;
306*4bdc9457SAndroid Build Coastguard Worker       vst1q_f32(o1, vo1); o1 += 4;
307*4bdc9457SAndroid Build Coastguard Worker       vst1q_f32(o0, vo0); o0 += 4;
308*4bdc9457SAndroid Build Coastguard Worker     }
309*4bdc9457SAndroid Build Coastguard Worker     // Always process the last block of 5..8 pixels.
310*4bdc9457SAndroid Build Coastguard Worker     if XNN_LIKELY(w > 4 * sizeof(float)) {
311*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
312*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
313*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
314*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo3p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
315*4bdc9457SAndroid Build Coastguard Worker 
316*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
317*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4;
318*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vi2x89AB = vld1q_f32(i2); i2 += 4;
319*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4;
320*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4;
321*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vi5x89AB = vld1q_f32(i5); i5 += 4;
322*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vi6x89AB = vld1q_f32(i6); i6 += 4;
323*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vi7x89AB = vld1q_f32(i7); i7 += 4;
324*4bdc9457SAndroid Build Coastguard Worker 
325*4bdc9457SAndroid Build Coastguard Worker       vi0x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x89AB)));
326*4bdc9457SAndroid Build Coastguard Worker       vi1x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x89AB)));
327*4bdc9457SAndroid Build Coastguard Worker       vi2x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x89AB)));
328*4bdc9457SAndroid Build Coastguard Worker       vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB)));
329*4bdc9457SAndroid Build Coastguard Worker       vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB)));
330*4bdc9457SAndroid Build Coastguard Worker       vi5x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi5x89AB)));
331*4bdc9457SAndroid Build Coastguard Worker       vi6x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi6x89AB)));
332*4bdc9457SAndroid Build Coastguard Worker       vi7x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi7x89AB)));
333*4bdc9457SAndroid Build Coastguard Worker 
334*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1);
335*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1);
336*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x4567, vget_high_f32(vw0123), 1);
337*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x4567, vget_high_f32(vw0123), 1);
338*4bdc9457SAndroid Build Coastguard Worker 
339*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0);
340*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw89AB), 0);
341*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x4567, vget_low_f32(vw89AB), 0);
342*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x4567, vget_low_f32(vw89AB), 0);
343*4bdc9457SAndroid Build Coastguard Worker 
344*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1);
345*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vget_low_f32(vwCDEF), 1);
346*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x4567, vget_low_f32(vwCDEF), 1);
347*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x4567, vget_low_f32(vwCDEF), 1);
348*4bdc9457SAndroid Build Coastguard Worker 
349*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0);
350*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0);
351*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x4567, vget_high_f32(vwGHIJ), 0);
352*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x4567, vget_high_f32(vwGHIJ), 0);
353*4bdc9457SAndroid Build Coastguard Worker 
354*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1);
355*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1);
356*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1);
357*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x4567, vget_high_f32(vwKLMN), 1);
358*4bdc9457SAndroid Build Coastguard Worker 
359*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
360*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
361*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
362*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
363*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);
364*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);
365*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi6x3456 = vextq_f32(vi6x0123, vi6x4567, 3);
366*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3);
367*4bdc9457SAndroid Build Coastguard Worker 
368*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0);
369*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x3456, vget_high_f32(vw0123), 0);
370*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x3456, vget_high_f32(vw0123), 0);
371*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x3456, vget_high_f32(vw0123), 0);
372*4bdc9457SAndroid Build Coastguard Worker 
373*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1);
374*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_high_f32(vw4567), 1);
375*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x3456, vget_high_f32(vw4567), 1);
376*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x3456, vget_high_f32(vw4567), 1);
377*4bdc9457SAndroid Build Coastguard Worker 
378*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0);
379*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x3456, vget_low_f32(vwCDEF), 0);
380*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x3456, vget_low_f32(vwCDEF), 0);
381*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0);
382*4bdc9457SAndroid Build Coastguard Worker 
383*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1);
384*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1);
385*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1);
386*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x3456, vget_low_f32(vwGHIJ), 1);
387*4bdc9457SAndroid Build Coastguard Worker 
388*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0);
389*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0);
390*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x3456, vget_high_f32(vwKLMN), 0);
391*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x3456, vget_high_f32(vwKLMN), 0);
392*4bdc9457SAndroid Build Coastguard Worker 
393*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2);
394*4bdc9457SAndroid Build Coastguard Worker       vi0x0123 = vi0x4567;
395*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi1x2345 = vextq_f32(vi1x0123, vi1x4567, 2);
396*4bdc9457SAndroid Build Coastguard Worker       vi1x0123 = vi1x4567;
397*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2);
398*4bdc9457SAndroid Build Coastguard Worker       vi2x0123 = vi2x4567;
399*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);
400*4bdc9457SAndroid Build Coastguard Worker       vi3x0123 = vi3x4567;
401*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);
402*4bdc9457SAndroid Build Coastguard Worker       vi4x0123 = vi4x4567;
403*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);
404*4bdc9457SAndroid Build Coastguard Worker       vi5x0123 = vi5x4567;
405*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi6x2345 = vextq_f32(vi6x0123, vi6x4567, 2);
406*4bdc9457SAndroid Build Coastguard Worker       vi6x0123 = vi6x4567;
407*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi7x2345 = vextq_f32(vi7x0123, vi7x4567, 2);
408*4bdc9457SAndroid Build Coastguard Worker       vi7x0123 = vi7x4567;
409*4bdc9457SAndroid Build Coastguard Worker 
410*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1);
411*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x2345, vget_low_f32(vw0123), 1);
412*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1);
413*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x2345, vget_low_f32(vw0123), 1);
414*4bdc9457SAndroid Build Coastguard Worker 
415*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x2345, vget_high_f32(vw4567), 0);
416*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0);
417*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x2345, vget_high_f32(vw4567), 0);
418*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x2345, vget_high_f32(vw4567), 0);
419*4bdc9457SAndroid Build Coastguard Worker 
420*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1);
421*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x2345, vget_high_f32(vw89AB), 1);
422*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x2345, vget_high_f32(vw89AB), 1);
423*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x2345, vget_high_f32(vw89AB), 1);
424*4bdc9457SAndroid Build Coastguard Worker 
425*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0);
426*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x2345, vget_low_f32(vwGHIJ), 0);
427*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x2345, vget_low_f32(vwGHIJ), 0);
428*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x2345, vget_low_f32(vwGHIJ), 0);
429*4bdc9457SAndroid Build Coastguard Worker 
430*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1);
431*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x2345, vget_low_f32(vwKLMN), 1);
432*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x2345, vget_low_f32(vwKLMN), 1);
433*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x2345, vget_low_f32(vwKLMN), 1);
434*4bdc9457SAndroid Build Coastguard Worker 
435*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1);
436*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vi1x89AB, 1);
437*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vi2x89AB, 1);
438*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);
439*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1);
440*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1);
441*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vi6x89AB, 1);
442*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vi7x89AB, 1);
443*4bdc9457SAndroid Build Coastguard Worker 
444*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0);
445*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x5678, vget_low_f32(vw4567), 0);
446*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x5678, vget_low_f32(vw4567), 0);
447*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x5678, vget_low_f32(vw4567), 0);
448*4bdc9457SAndroid Build Coastguard Worker 
449*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1);
450*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x5678, vget_low_f32(vw89AB), 1);
451*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x5678, vget_low_f32(vw89AB), 1);
452*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x5678, vget_low_f32(vw89AB), 1);
453*4bdc9457SAndroid Build Coastguard Worker 
454*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vget_high_f32(vwCDEF), 0);
455*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x5678, vget_high_f32(vwCDEF), 0);
456*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x5678, vget_high_f32(vwCDEF), 0);
457*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vget_high_f32(vwCDEF), 0);
458*4bdc9457SAndroid Build Coastguard Worker 
459*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1);
460*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x5678, vget_high_f32(vwGHIJ), 1);
461*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1);
462*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x5678, vget_high_f32(vwGHIJ), 1);
463*4bdc9457SAndroid Build Coastguard Worker 
464*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x5678, vwOP, 0);
465*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x5678, vwOP, 0);
466*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x5678, vwOP, 0);
467*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x5678, vwOP, 0);
468*4bdc9457SAndroid Build Coastguard Worker 
469*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2);
470*4bdc9457SAndroid Build Coastguard Worker       vi0x4567 = vi0x89AB;
471*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi1x6789 = vextq_f32(vi1x4567, vi1x89AB, 2);
472*4bdc9457SAndroid Build Coastguard Worker       vi1x4567 = vi1x89AB;
473*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2);
474*4bdc9457SAndroid Build Coastguard Worker       vi2x4567 = vi2x89AB;
475*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);
476*4bdc9457SAndroid Build Coastguard Worker       vi3x4567 = vi3x89AB;
477*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2);
478*4bdc9457SAndroid Build Coastguard Worker       vi4x4567 = vi4x89AB;
479*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2);
480*4bdc9457SAndroid Build Coastguard Worker       vi5x4567 = vi5x89AB;
481*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi6x6789 = vextq_f32(vi6x4567, vi6x89AB, 2);
482*4bdc9457SAndroid Build Coastguard Worker       vi6x4567 = vi6x89AB;
483*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi7x6789 = vextq_f32(vi7x4567, vi7x89AB, 2);
484*4bdc9457SAndroid Build Coastguard Worker       vi7x4567 = vi7x89AB;
485*4bdc9457SAndroid Build Coastguard Worker 
486*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1);
487*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x6789, vget_low_f32(vw4567), 1);
488*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1);
489*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x6789, vget_low_f32(vw4567), 1);
490*4bdc9457SAndroid Build Coastguard Worker 
491*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x6789, vget_high_f32(vw89AB), 0);
492*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0);
493*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x6789, vget_high_f32(vw89AB), 0);
494*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x6789, vget_high_f32(vw89AB), 0);
495*4bdc9457SAndroid Build Coastguard Worker 
496*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1);
497*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x6789, vget_high_f32(vwCDEF), 1);
498*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1);
499*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1);
500*4bdc9457SAndroid Build Coastguard Worker 
501*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0);
502*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0);
503*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0);
504*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x6789, vget_low_f32(vwKLMN), 0);
505*4bdc9457SAndroid Build Coastguard Worker 
506*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1);
507*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1);
508*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x6789, vwOP, 1);
509*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x6789, vwOP, 1);
510*4bdc9457SAndroid Build Coastguard Worker 
511*4bdc9457SAndroid Build Coastguard Worker 
512*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
513*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
514*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo2 = vmaxq_f32(vo2p0, vmin);
515*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo3 = vmaxq_f32(vo3p0, vmin);
516*4bdc9457SAndroid Build Coastguard Worker 
517*4bdc9457SAndroid Build Coastguard Worker       vo0 = vminq_f32(vo0, vmax);
518*4bdc9457SAndroid Build Coastguard Worker       vo1 = vminq_f32(vo1, vmax);
519*4bdc9457SAndroid Build Coastguard Worker       vo2 = vminq_f32(vo2, vmax);
520*4bdc9457SAndroid Build Coastguard Worker       vo3 = vminq_f32(vo3, vmax);
521*4bdc9457SAndroid Build Coastguard Worker 
522*4bdc9457SAndroid Build Coastguard Worker       vst1q_f32(o3, vo3); o3 += 4;
523*4bdc9457SAndroid Build Coastguard Worker       vst1q_f32(o2, vo2); o2 += 4;
524*4bdc9457SAndroid Build Coastguard Worker       vst1q_f32(o1, vo1); o1 += 4;
525*4bdc9457SAndroid Build Coastguard Worker       vst1q_f32(o0, vo0); o0 += 4;
526*4bdc9457SAndroid Build Coastguard Worker 
527*4bdc9457SAndroid Build Coastguard Worker       w -= 4 * sizeof(float);
528*4bdc9457SAndroid Build Coastguard Worker     }
529*4bdc9457SAndroid Build Coastguard Worker     assert(w >= 1 * sizeof(float));
530*4bdc9457SAndroid Build Coastguard Worker     assert(w <= 4 * sizeof(float));
531*4bdc9457SAndroid Build Coastguard Worker     {
532*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
533*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
534*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
535*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo3p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
536*4bdc9457SAndroid Build Coastguard Worker 
537*4bdc9457SAndroid Build Coastguard Worker       vi0x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x4567)));
538*4bdc9457SAndroid Build Coastguard Worker       vi1x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x4567)));
539*4bdc9457SAndroid Build Coastguard Worker       vi2x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x4567)));
540*4bdc9457SAndroid Build Coastguard Worker       vi3x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x4567)));
541*4bdc9457SAndroid Build Coastguard Worker       vi4x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x4567)));
542*4bdc9457SAndroid Build Coastguard Worker       vi5x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi5x4567)));
543*4bdc9457SAndroid Build Coastguard Worker       vi6x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi6x4567)));
544*4bdc9457SAndroid Build Coastguard Worker       vi7x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi7x4567)));
545*4bdc9457SAndroid Build Coastguard Worker 
546*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1);
547*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1);
548*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x4567, vget_high_f32(vw0123), 1);
549*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x4567, vget_high_f32(vw0123), 1);
550*4bdc9457SAndroid Build Coastguard Worker 
551*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0);
552*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw89AB), 0);
553*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x4567, vget_low_f32(vw89AB), 0);
554*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x4567, vget_low_f32(vw89AB), 0);
555*4bdc9457SAndroid Build Coastguard Worker 
556*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1);
557*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vget_low_f32(vwCDEF), 1);
558*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x4567, vget_low_f32(vwCDEF), 1);
559*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x4567, vget_low_f32(vwCDEF), 1);
560*4bdc9457SAndroid Build Coastguard Worker 
561*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0);
562*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0);
563*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x4567, vget_high_f32(vwGHIJ), 0);
564*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x4567, vget_high_f32(vwGHIJ), 0);
565*4bdc9457SAndroid Build Coastguard Worker 
566*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1);
567*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1);
568*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1);
569*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x4567, vget_high_f32(vwKLMN), 1);
570*4bdc9457SAndroid Build Coastguard Worker 
571*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
572*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
573*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
574*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
575*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);
576*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);
577*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi6x3456 = vextq_f32(vi6x0123, vi6x4567, 3);
578*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3);
579*4bdc9457SAndroid Build Coastguard Worker 
580*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0);
581*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x3456, vget_high_f32(vw0123), 0);
582*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x3456, vget_high_f32(vw0123), 0);
583*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x3456, vget_high_f32(vw0123), 0);
584*4bdc9457SAndroid Build Coastguard Worker 
585*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1);
586*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_high_f32(vw4567), 1);
587*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x3456, vget_high_f32(vw4567), 1);
588*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x3456, vget_high_f32(vw4567), 1);
589*4bdc9457SAndroid Build Coastguard Worker 
590*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0);
591*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x3456, vget_low_f32(vwCDEF), 0);
592*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x3456, vget_low_f32(vwCDEF), 0);
593*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0);
594*4bdc9457SAndroid Build Coastguard Worker 
595*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1);
596*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1);
597*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1);
598*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x3456, vget_low_f32(vwGHIJ), 1);
599*4bdc9457SAndroid Build Coastguard Worker 
600*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0);
601*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0);
602*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x3456, vget_high_f32(vwKLMN), 0);
603*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x3456, vget_high_f32(vwKLMN), 0);
604*4bdc9457SAndroid Build Coastguard Worker 
605*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2);
606*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi1x2345 = vextq_f32(vi1x0123, vi1x4567, 2);
607*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2);
608*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);
609*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);
610*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);
611*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi6x2345 = vextq_f32(vi6x0123, vi6x4567, 2);
612*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi7x2345 = vextq_f32(vi7x0123, vi7x4567, 2);
613*4bdc9457SAndroid Build Coastguard Worker 
614*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1);
615*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x2345, vget_low_f32(vw0123), 1);
616*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x2345, vget_low_f32(vw0123), 1);
617*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x2345, vget_low_f32(vw0123), 1);
618*4bdc9457SAndroid Build Coastguard Worker 
619*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x2345, vget_high_f32(vw4567), 0);
620*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0);
621*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x2345, vget_high_f32(vw4567), 0);
622*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x2345, vget_high_f32(vw4567), 0);
623*4bdc9457SAndroid Build Coastguard Worker 
624*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1);
625*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x2345, vget_high_f32(vw89AB), 1);
626*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x2345, vget_high_f32(vw89AB), 1);
627*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x2345, vget_high_f32(vw89AB), 1);
628*4bdc9457SAndroid Build Coastguard Worker 
629*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0);
630*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x2345, vget_low_f32(vwGHIJ), 0);
631*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x2345, vget_low_f32(vwGHIJ), 0);
632*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x2345, vget_low_f32(vwGHIJ), 0);
633*4bdc9457SAndroid Build Coastguard Worker 
634*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1);
635*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x2345, vget_low_f32(vwKLMN), 1);
636*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x2345, vget_low_f32(vwKLMN), 1);
637*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x2345, vget_low_f32(vwKLMN), 1);
638*4bdc9457SAndroid Build Coastguard Worker 
639*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vzero = vmovq_n_f32(0.0f);
640*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1);
641*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vzero, 1);
642*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vzero, 1);
643*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1);
644*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1);
645*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1);
646*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi6x5678 = vextq_f32(vi6x4567, vzero, 1);
647*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi7x5678 = vextq_f32(vi7x4567, vzero, 1);
648*4bdc9457SAndroid Build Coastguard Worker 
649*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0);
650*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x5678, vget_low_f32(vw4567), 0);
651*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x5678, vget_low_f32(vw4567), 0);
652*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x5678, vget_low_f32(vw4567), 0);
653*4bdc9457SAndroid Build Coastguard Worker 
654*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1);
655*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x5678, vget_low_f32(vw89AB), 1);
656*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x5678, vget_low_f32(vw89AB), 1);
657*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x5678, vget_low_f32(vw89AB), 1);
658*4bdc9457SAndroid Build Coastguard Worker 
659*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vget_high_f32(vwCDEF), 0);
660*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x5678, vget_high_f32(vwCDEF), 0);
661*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x5678, vget_high_f32(vwCDEF), 0);
662*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vget_high_f32(vwCDEF), 0);
663*4bdc9457SAndroid Build Coastguard Worker 
664*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1);
665*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x5678, vget_high_f32(vwGHIJ), 1);
666*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1);
667*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x5678, vget_high_f32(vwGHIJ), 1);
668*4bdc9457SAndroid Build Coastguard Worker 
669*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x5678, vwOP, 0);
670*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x5678, vwOP, 0);
671*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x5678, vwOP, 0);
672*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x5678, vwOP, 0);
673*4bdc9457SAndroid Build Coastguard Worker 
674*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1);
675*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi1x6789 = vextq_f32(vi1x5678, vzero, 1);
676*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1);
677*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi3x6789 = vextq_f32(vi3x5678, vzero, 1);
678*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1);
679*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi5x6789 = vextq_f32(vi5x5678, vzero, 1);
680*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi6x6789 = vextq_f32(vi6x5678, vzero, 1);
681*4bdc9457SAndroid Build Coastguard Worker       const float32x4_t vi7x6789 = vextq_f32(vi7x5678, vzero, 1);
682*4bdc9457SAndroid Build Coastguard Worker 
683*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1);
684*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x6789, vget_low_f32(vw4567), 1);
685*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi2x6789, vget_low_f32(vw4567), 1);
686*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi3x6789, vget_low_f32(vw4567), 1);
687*4bdc9457SAndroid Build Coastguard Worker 
688*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x6789, vget_high_f32(vw89AB), 0);
689*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0);
690*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi3x6789, vget_high_f32(vw89AB), 0);
691*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi4x6789, vget_high_f32(vw89AB), 0);
692*4bdc9457SAndroid Build Coastguard Worker 
693*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1);
694*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x6789, vget_high_f32(vwCDEF), 1);
695*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1);
696*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1);
697*4bdc9457SAndroid Build Coastguard Worker 
698*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0);
699*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0);
700*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0);
701*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x6789, vget_low_f32(vwKLMN), 0);
702*4bdc9457SAndroid Build Coastguard Worker 
703*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1);
704*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1);
705*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x6789, vwOP, 1);
706*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x6789, vwOP, 1);
707*4bdc9457SAndroid Build Coastguard Worker 
708*4bdc9457SAndroid Build Coastguard Worker 
709*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
710*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
711*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo2 = vmaxq_f32(vo2p0, vmin);
712*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vo3 = vmaxq_f32(vo3p0, vmin);
713*4bdc9457SAndroid Build Coastguard Worker 
714*4bdc9457SAndroid Build Coastguard Worker       vo0 = vminq_f32(vo0, vmax);
715*4bdc9457SAndroid Build Coastguard Worker       vo1 = vminq_f32(vo1, vmax);
716*4bdc9457SAndroid Build Coastguard Worker       vo2 = vminq_f32(vo2, vmax);
717*4bdc9457SAndroid Build Coastguard Worker       vo3 = vminq_f32(vo3, vmax);
718*4bdc9457SAndroid Build Coastguard Worker 
719*4bdc9457SAndroid Build Coastguard Worker       if XNN_LIKELY(w & (4 * sizeof(float))) {
720*4bdc9457SAndroid Build Coastguard Worker         vst1q_f32(o3, vo3); o3 += 4;
721*4bdc9457SAndroid Build Coastguard Worker         vst1q_f32(o2, vo2); o2 += 4;
722*4bdc9457SAndroid Build Coastguard Worker         vst1q_f32(o1, vo1); o1 += 4;
723*4bdc9457SAndroid Build Coastguard Worker         vst1q_f32(o0, vo0); o0 += 4;
724*4bdc9457SAndroid Build Coastguard Worker       } else {
725*4bdc9457SAndroid Build Coastguard Worker         float32x2_t vo0_lo = vget_low_f32(vo0);
726*4bdc9457SAndroid Build Coastguard Worker         float32x2_t vo1_lo = vget_low_f32(vo1);
727*4bdc9457SAndroid Build Coastguard Worker         float32x2_t vo2_lo = vget_low_f32(vo2);
728*4bdc9457SAndroid Build Coastguard Worker         float32x2_t vo3_lo = vget_low_f32(vo3);
729*4bdc9457SAndroid Build Coastguard Worker         if (w & (2 * sizeof(float))) {
730*4bdc9457SAndroid Build Coastguard Worker           vst1_f32(o3, vo3_lo); o3 += 2;
731*4bdc9457SAndroid Build Coastguard Worker           vst1_f32(o2, vo2_lo); o2 += 2;
732*4bdc9457SAndroid Build Coastguard Worker           vst1_f32(o1, vo1_lo); o1 += 2;
733*4bdc9457SAndroid Build Coastguard Worker           vst1_f32(o0, vo0_lo); o0 += 2;
734*4bdc9457SAndroid Build Coastguard Worker 
735*4bdc9457SAndroid Build Coastguard Worker           vo0_lo = vget_high_f32(vo0);
736*4bdc9457SAndroid Build Coastguard Worker           vo1_lo = vget_high_f32(vo1);
737*4bdc9457SAndroid Build Coastguard Worker           vo2_lo = vget_high_f32(vo2);
738*4bdc9457SAndroid Build Coastguard Worker           vo3_lo = vget_high_f32(vo3);
739*4bdc9457SAndroid Build Coastguard Worker         }
740*4bdc9457SAndroid Build Coastguard Worker         if (w & (1 * sizeof(float))) {
741*4bdc9457SAndroid Build Coastguard Worker           vst1_lane_f32(o3, vo3_lo, 0); o3 += 1;
742*4bdc9457SAndroid Build Coastguard Worker           vst1_lane_f32(o2, vo2_lo, 0); o2 += 1;
743*4bdc9457SAndroid Build Coastguard Worker           vst1_lane_f32(o1, vo1_lo, 0); o1 += 1;
744*4bdc9457SAndroid Build Coastguard Worker           vst1_lane_f32(o0, vo0_lo, 0); o0 += 1;
745*4bdc9457SAndroid Build Coastguard Worker         }
746*4bdc9457SAndroid Build Coastguard Worker       }
747*4bdc9457SAndroid Build Coastguard Worker     }
748*4bdc9457SAndroid Build Coastguard Worker 
749*4bdc9457SAndroid Build Coastguard Worker     i0 = (const float*) ((uintptr_t) i4 - input_decrement);
750*4bdc9457SAndroid Build Coastguard Worker     i1 = (const float*) ((uintptr_t) i5 - input_decrement);
751*4bdc9457SAndroid Build Coastguard Worker     i2 = (const float*) ((uintptr_t) i1 + input_width);
752*4bdc9457SAndroid Build Coastguard Worker     i3 = (const float*) ((uintptr_t) i2 + input_width);
753*4bdc9457SAndroid Build Coastguard Worker     i4 = (const float*) ((uintptr_t) i3 + input_width);
754*4bdc9457SAndroid Build Coastguard Worker     i5 = (const float*) ((uintptr_t) i4 + input_width);
755*4bdc9457SAndroid Build Coastguard Worker     i6 = (const float*) ((uintptr_t) i5 + input_width);
756*4bdc9457SAndroid Build Coastguard Worker     i7 = (const float*) ((uintptr_t) i6 + input_width);
757*4bdc9457SAndroid Build Coastguard Worker 
758*4bdc9457SAndroid Build Coastguard Worker     o0 = o3;
759*4bdc9457SAndroid Build Coastguard Worker     o1 = (float*) ((uintptr_t) o0 + input_width);
760*4bdc9457SAndroid Build Coastguard Worker     o2 = (float*) ((uintptr_t) o1 + input_width);
761*4bdc9457SAndroid Build Coastguard Worker     o3 = (float*) ((uintptr_t) o2 + input_width);
762*4bdc9457SAndroid Build Coastguard Worker 
763*4bdc9457SAndroid Build Coastguard Worker     output_height = doz(output_height, 4);
764*4bdc9457SAndroid Build Coastguard Worker   } while (output_height != 0);
765*4bdc9457SAndroid Build Coastguard Worker }
766