xref: /aosp_15_r20/external/XNNPACK/src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f32-conv-hwc/3x3s2p1c3-neon-x2.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 
11*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
12*4bdc9457SAndroid Build Coastguard Worker 
13*4bdc9457SAndroid Build Coastguard Worker #include <arm_neon.h>
14*4bdc9457SAndroid Build Coastguard Worker 
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/conv.h>
16*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/math.h>
17*4bdc9457SAndroid Build Coastguard Worker 
18*4bdc9457SAndroid Build Coastguard Worker 
xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2(size_t input_height,size_t input_width,size_t output_y_start,size_t output_y_end,const float * input,const float * zero,const float * weights,float * output,size_t input_padding_top,size_t output_channels,size_t output_height_stride,size_t output_width_stride,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2(
20*4bdc9457SAndroid Build Coastguard Worker     size_t input_height,
21*4bdc9457SAndroid Build Coastguard Worker     size_t input_width,
22*4bdc9457SAndroid Build Coastguard Worker     size_t output_y_start,
23*4bdc9457SAndroid Build Coastguard Worker     size_t output_y_end,
24*4bdc9457SAndroid Build Coastguard Worker     const float* input,
25*4bdc9457SAndroid Build Coastguard Worker     const float* zero,
26*4bdc9457SAndroid Build Coastguard Worker     const float* weights,
27*4bdc9457SAndroid Build Coastguard Worker     float* output,
28*4bdc9457SAndroid Build Coastguard Worker     size_t input_padding_top,
29*4bdc9457SAndroid Build Coastguard Worker     size_t output_channels,
30*4bdc9457SAndroid Build Coastguard Worker     size_t output_height_stride,
31*4bdc9457SAndroid Build Coastguard Worker     size_t output_width_stride,
32*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33*4bdc9457SAndroid Build Coastguard Worker {
34*4bdc9457SAndroid Build Coastguard Worker   assert(input_width != 0);
35*4bdc9457SAndroid Build Coastguard Worker   assert(output_y_end > output_y_start);
36*4bdc9457SAndroid Build Coastguard Worker   assert(input_padding_top <= 1);
37*4bdc9457SAndroid Build Coastguard Worker   assert(output_channels != 0);
38*4bdc9457SAndroid Build Coastguard Worker 
39*4bdc9457SAndroid Build Coastguard Worker   const size_t input_height_stride = input_width * 3 /* channels */ * sizeof(float);
40*4bdc9457SAndroid Build Coastguard Worker   const size_t input_width_decrement = input_width * 3 /* channels */ * sizeof(float);
41*4bdc9457SAndroid Build Coastguard Worker   const size_t output_width = (input_width + 1) / 2;
42*4bdc9457SAndroid Build Coastguard Worker   const size_t output_channel_decrement = output_width * output_width_stride - 8 * sizeof(float);
43*4bdc9457SAndroid Build Coastguard Worker   const size_t output_height_increment = output_height_stride * 2 - round_up_po2(output_channels, 8) * sizeof(float);
44*4bdc9457SAndroid Build Coastguard Worker 
45*4bdc9457SAndroid Build Coastguard Worker   // Adjustment for padding processed below
46*4bdc9457SAndroid Build Coastguard Worker   const float* i0 = (const float*) ((uintptr_t) input +
47*4bdc9457SAndroid Build Coastguard Worker     input_height_stride * (output_y_start * 2 /* vertical stride */ - input_padding_top));
48*4bdc9457SAndroid Build Coastguard Worker   const float* i1 = (const float*) ((uintptr_t) i0 + input_height_stride);
49*4bdc9457SAndroid Build Coastguard Worker   const float* i2 = (const float*) ((uintptr_t) i1 + input_height_stride);
50*4bdc9457SAndroid Build Coastguard Worker   const float* i3 = (const float*) ((uintptr_t) i2 + input_height_stride);
51*4bdc9457SAndroid Build Coastguard Worker   const float* i4 = (const float*) ((uintptr_t) i3 + input_height_stride);
52*4bdc9457SAndroid Build Coastguard Worker   float* o0 = (float*) ((uintptr_t) output + output_height_stride * output_y_start);
53*4bdc9457SAndroid Build Coastguard Worker   float* o1 = (float*) ((uintptr_t) o0 + output_height_stride);
54*4bdc9457SAndroid Build Coastguard Worker 
55*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNPREDICTABLE(output_y_start < input_padding_top) {
56*4bdc9457SAndroid Build Coastguard Worker     i0 = zero;
57*4bdc9457SAndroid Build Coastguard Worker   }
58*4bdc9457SAndroid Build Coastguard Worker 
59*4bdc9457SAndroid Build Coastguard Worker 
60*4bdc9457SAndroid Build Coastguard Worker   for (size_t output_y = output_y_start; output_y < output_y_end; output_y += 2) {
61*4bdc9457SAndroid Build Coastguard Worker     const size_t input_y2 = output_y * 2 + 2 - input_padding_top;
62*4bdc9457SAndroid Build Coastguard Worker     const size_t input_y4 = input_y2 + 2;
63*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(input_y2 > input_height) {
64*4bdc9457SAndroid Build Coastguard Worker       i1 = zero;
65*4bdc9457SAndroid Build Coastguard Worker     }
66*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(input_y2 >= input_height) {
67*4bdc9457SAndroid Build Coastguard Worker       i2 = zero;
68*4bdc9457SAndroid Build Coastguard Worker     }
69*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(input_y4 > input_height) {
70*4bdc9457SAndroid Build Coastguard Worker       i3 = zero;
71*4bdc9457SAndroid Build Coastguard Worker     }
72*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(input_y4 >= input_height) {
73*4bdc9457SAndroid Build Coastguard Worker       i4 = zero;
74*4bdc9457SAndroid Build Coastguard Worker     }
75*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_y + 2 > output_y_end) {
76*4bdc9457SAndroid Build Coastguard Worker       o1 = o0;
77*4bdc9457SAndroid Build Coastguard Worker     }
78*4bdc9457SAndroid Build Coastguard Worker 
79*4bdc9457SAndroid Build Coastguard Worker     const float* w = weights;
80*4bdc9457SAndroid Build Coastguard Worker     size_t c = output_channels;
81*4bdc9457SAndroid Build Coastguard Worker     do {
82*4bdc9457SAndroid Build Coastguard Worker       // viMx0 = ( iM0c2, iM0c1, iM0c0, --- )
83*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vi0x0 = vmovq_n_f32(0.0f);
84*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vi1x0 = vmovq_n_f32(0.0f);
85*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vi2x0 = vmovq_n_f32(0.0f);
86*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vi3x0 = vmovq_n_f32(0.0f);
87*4bdc9457SAndroid Build Coastguard Worker       float32x4_t vi4x0 = vmovq_n_f32(0.0f);
88*4bdc9457SAndroid Build Coastguard Worker 
89*4bdc9457SAndroid Build Coastguard Worker       size_t iw = input_width;
90*4bdc9457SAndroid Build Coastguard Worker       for (; iw >= 4; iw -= 4) {
91*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo0x0c0123 = vld1q_f32(w);
92*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo0x0c4567 = vld1q_f32(w + 4);
93*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo1x0c0123 = vo0x0c0123;
94*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo1x0c4567 = vo0x0c4567;
95*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo0x1c0123 = vo0x0c0123;
96*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo0x1c4567 = vo0x0c4567;
97*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo1x1c0123 = vo0x0c0123;
98*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo1x1c4567 = vo0x0c4567;
99*4bdc9457SAndroid Build Coastguard Worker 
100*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c0x0123 = vld1q_f32(w + 8);
101*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c0x4567 = vld1q_f32(w + 12);
102*4bdc9457SAndroid Build Coastguard Worker 
103*4bdc9457SAndroid Build Coastguard Worker         // viMx1 = ( iM2c0, iM1c2, iM1c1, iM1c0 )
104*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi0x1 = vld1q_f32(i0); i0 += 4;
105*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi1x1 = vld1q_f32(i1); i1 += 4;
106*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4;
107*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi3x1 = vld1q_f32(i3); i3 += 4;
108*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi4x1 = vld1q_f32(i4); i4 += 4;
109*4bdc9457SAndroid Build Coastguard Worker 
110*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c0x0123, vget_low_f32(vi0x0), 1);
111*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);
112*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk00c0x4567, vget_low_f32(vi0x0), 1);
113*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk00c0x4567, vget_low_f32(vi2x0), 1);
114*4bdc9457SAndroid Build Coastguard Worker 
115*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk00c0x0123, vget_high_f32(vi0x1), 1);
116*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1);
117*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk00c0x4567, vget_high_f32(vi0x1), 1);
118*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk00c0x4567, vget_high_f32(vi2x1), 1);
119*4bdc9457SAndroid Build Coastguard Worker 
120*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c0x0123 = vld1q_f32(w + 16);
121*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c0x4567 = vld1q_f32(w + 20);
122*4bdc9457SAndroid Build Coastguard Worker 
123*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1);
124*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 1);
125*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk10c0x4567, vget_low_f32(vi1x0), 1);
126*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk10c0x4567, vget_low_f32(vi3x0), 1);
127*4bdc9457SAndroid Build Coastguard Worker 
128*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c0x0123, vget_high_f32(vi1x1), 1);
129*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 1);
130*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk10c0x4567, vget_high_f32(vi1x1), 1);
131*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk10c0x4567, vget_high_f32(vi3x1), 1);
132*4bdc9457SAndroid Build Coastguard Worker 
133*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c0x0123 = vld1q_f32(w + 24);
134*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c0x4567 = vld1q_f32(w + 28);
135*4bdc9457SAndroid Build Coastguard Worker 
136*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1);
137*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 1);
138*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1);
139*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk20c0x4567, vget_low_f32(vi4x0), 1);
140*4bdc9457SAndroid Build Coastguard Worker 
141*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 1);
142*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 1);
143*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c0x4567, vget_high_f32(vi2x1), 1);
144*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk20c0x4567, vget_high_f32(vi4x1), 1);
145*4bdc9457SAndroid Build Coastguard Worker 
146*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c1x0123 = vld1q_f32(w + 32);
147*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c1x4567 = vld1q_f32(w + 36);
148*4bdc9457SAndroid Build Coastguard Worker 
149*4bdc9457SAndroid Build Coastguard Worker         // viMx2 = ( iM3c1, iM3c0, iM2c2, iM2c1 )
150*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi0x2 = vld1q_f32(i0); i0 += 4;
151*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi1x2 = vld1q_f32(i1); i1 += 4;
152*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi2x2 = vld1q_f32(i2); i2 += 4;
153*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi3x2 = vld1q_f32(i3); i3 += 4;
154*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi4x2 = vld1q_f32(i4); i4 += 4;
155*4bdc9457SAndroid Build Coastguard Worker 
156*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_high_f32(vi0x0), 0);
157*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_high_f32(vi2x0), 0);
158*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk00c1x4567, vget_high_f32(vi0x0), 0);
159*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk00c1x4567, vget_high_f32(vi2x0), 0);
160*4bdc9457SAndroid Build Coastguard Worker 
161*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk00c1x0123, vget_low_f32(vi0x2), 0);
162*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_low_f32(vi2x2), 0);
163*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk00c1x4567, vget_low_f32(vi0x2), 0);
164*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk00c1x4567, vget_low_f32(vi2x2), 0);
165*4bdc9457SAndroid Build Coastguard Worker 
166*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c1x0123 = vld1q_f32(w + 40);
167*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c1x4567 = vld1q_f32(w + 44);
168*4bdc9457SAndroid Build Coastguard Worker 
169*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0);
170*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_high_f32(vi3x0), 0);
171*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk10c1x4567, vget_high_f32(vi1x0), 0);
172*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk10c1x4567, vget_high_f32(vi3x0), 0);
173*4bdc9457SAndroid Build Coastguard Worker 
174*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_low_f32(vi1x2), 0);
175*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_low_f32(vi3x2), 0);
176*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk10c1x4567, vget_low_f32(vi1x2), 0);
177*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk10c1x4567, vget_low_f32(vi3x2), 0);
178*4bdc9457SAndroid Build Coastguard Worker 
179*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c1x0123 = vld1q_f32(w + 48);
180*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c1x4567 = vld1q_f32(w + 52);
181*4bdc9457SAndroid Build Coastguard Worker 
182*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_high_f32(vi2x0), 0);
183*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c1x0123, vget_high_f32(vi4x0), 0);
184*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0);
185*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk20c1x4567, vget_high_f32(vi4x0), 0);
186*4bdc9457SAndroid Build Coastguard Worker 
187*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c1x0123, vget_low_f32(vi2x2), 0);
188*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c1x0123, vget_low_f32(vi4x2), 0);
189*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c1x4567, vget_low_f32(vi2x2), 0);
190*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk20c1x4567, vget_low_f32(vi4x2), 0);
191*4bdc9457SAndroid Build Coastguard Worker 
192*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c2x0123 = vld1q_f32(w + 56);
193*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c2x4567 = vld1q_f32(w + 60);
194*4bdc9457SAndroid Build Coastguard Worker 
195*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c2x0123, vget_high_f32(vi0x0), 1);
196*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 1);
197*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk00c2x4567, vget_high_f32(vi0x0), 1);
198*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk00c2x4567, vget_high_f32(vi2x0), 1);
199*4bdc9457SAndroid Build Coastguard Worker 
200*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk00c2x0123, vget_low_f32(vi0x2), 1);
201*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c2x0123, vget_low_f32(vi2x2), 1);
202*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk00c2x4567, vget_low_f32(vi0x2), 1);
203*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk00c2x4567, vget_low_f32(vi2x2), 1);
204*4bdc9457SAndroid Build Coastguard Worker 
205*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c2x0123 = vld1q_f32(w + 64);
206*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c2x4567 = vld1q_f32(w + 68);
207*4bdc9457SAndroid Build Coastguard Worker 
208*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c2x0123, vget_high_f32(vi1x0), 1);
209*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c2x0123, vget_high_f32(vi3x0), 1);
210*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1);
211*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk10c2x4567, vget_high_f32(vi3x0), 1);
212*4bdc9457SAndroid Build Coastguard Worker 
213*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c2x0123, vget_low_f32(vi1x2), 1);
214*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c2x0123, vget_low_f32(vi3x2), 1);
215*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk10c2x4567, vget_low_f32(vi1x2), 1);
216*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk10c2x4567, vget_low_f32(vi3x2), 1);
217*4bdc9457SAndroid Build Coastguard Worker 
218*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c2x0123 = vld1q_f32(w + 72);
219*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c2x4567 = vld1q_f32(w + 76);
220*4bdc9457SAndroid Build Coastguard Worker 
221*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c2x0123, vget_high_f32(vi2x0), 1);
222*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c2x0123, vget_high_f32(vi4x0), 1);
223*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1);
224*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk20c2x4567, vget_high_f32(vi4x0), 1);
225*4bdc9457SAndroid Build Coastguard Worker 
226*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c2x0123, vget_low_f32(vi2x2), 1);
227*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c2x0123, vget_low_f32(vi4x2), 1);
228*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c2x4567, vget_low_f32(vi2x2), 1);
229*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk20c2x4567, vget_low_f32(vi4x2), 1);
230*4bdc9457SAndroid Build Coastguard Worker 
231*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c0x0123 = vld1q_f32(w + 80);
232*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c0x4567 = vld1q_f32(w + 84);
233*4bdc9457SAndroid Build Coastguard Worker 
234*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0);
235*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);
236*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk01c0x4567, vget_low_f32(vi0x1), 0);
237*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk01c0x4567, vget_low_f32(vi2x1), 0);
238*4bdc9457SAndroid Build Coastguard Worker 
239*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c0x0123, vget_high_f32(vi0x2), 0);
240*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c0x0123, vget_high_f32(vi2x2), 0);
241*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk01c0x4567, vget_high_f32(vi0x2), 0);
242*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk01c0x4567, vget_high_f32(vi2x2), 0);
243*4bdc9457SAndroid Build Coastguard Worker 
244*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c0x0123 = vld1q_f32(w + 88);
245*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c0x4567 = vld1q_f32(w + 92);
246*4bdc9457SAndroid Build Coastguard Worker 
247*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0);
248*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);
249*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk11c0x4567, vget_low_f32(vi1x1), 0);
250*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk11c0x4567, vget_low_f32(vi3x1), 0);
251*4bdc9457SAndroid Build Coastguard Worker 
252*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c0x0123, vget_high_f32(vi1x2), 0);
253*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c0x0123, vget_high_f32(vi3x2), 0);
254*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk11c0x4567, vget_high_f32(vi1x2), 0);
255*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk11c0x4567, vget_high_f32(vi3x2), 0);
256*4bdc9457SAndroid Build Coastguard Worker 
257*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c0x0123 = vld1q_f32(w + 96);
258*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c0x4567 = vld1q_f32(w + 100);
259*4bdc9457SAndroid Build Coastguard Worker 
260*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0);
261*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c0x0123, vget_low_f32(vi4x1), 0);
262*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0);
263*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk21c0x4567, vget_low_f32(vi4x1), 0);
264*4bdc9457SAndroid Build Coastguard Worker 
265*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk21c0x0123, vget_high_f32(vi2x2), 0);
266*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk21c0x0123, vget_high_f32(vi4x2), 0);
267*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk21c0x4567, vget_high_f32(vi2x2), 0);
268*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk21c0x4567, vget_high_f32(vi4x2), 0);
269*4bdc9457SAndroid Build Coastguard Worker 
270*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c1x0123 = vld1q_f32(w + 104);
271*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c1x4567 = vld1q_f32(w + 108);
272*4bdc9457SAndroid Build Coastguard Worker 
273*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c1x0123, vget_low_f32(vi0x1), 1);
274*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);
275*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk01c1x4567, vget_low_f32(vi0x1), 1);
276*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk01c1x4567, vget_low_f32(vi2x1), 1);
277*4bdc9457SAndroid Build Coastguard Worker 
278*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c1x0123, vget_high_f32(vi0x2), 1);
279*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c1x0123, vget_high_f32(vi2x2), 1);
280*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk01c1x4567, vget_high_f32(vi0x2), 1);
281*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk01c1x4567, vget_high_f32(vi2x2), 1);
282*4bdc9457SAndroid Build Coastguard Worker 
283*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c1x0123 = vld1q_f32(w + 112);
284*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c1x4567 = vld1q_f32(w + 116);
285*4bdc9457SAndroid Build Coastguard Worker 
286*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1);
287*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c1x0123, vget_low_f32(vi3x1), 1);
288*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk11c1x4567, vget_low_f32(vi1x1), 1);
289*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk11c1x4567, vget_low_f32(vi3x1), 1);
290*4bdc9457SAndroid Build Coastguard Worker 
291*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c1x0123, vget_high_f32(vi1x2), 1);
292*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c1x0123, vget_high_f32(vi3x2), 1);
293*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk11c1x4567, vget_high_f32(vi1x2), 1);
294*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk11c1x4567, vget_high_f32(vi3x2), 1);
295*4bdc9457SAndroid Build Coastguard Worker 
296*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c1x0123 = vld1q_f32(w + 120);
297*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c1x4567 = vld1q_f32(w + 124);
298*4bdc9457SAndroid Build Coastguard Worker 
299*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1);
300*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 1);
301*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1);
302*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk21c1x4567, vget_low_f32(vi4x1), 1);
303*4bdc9457SAndroid Build Coastguard Worker 
304*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 1);
305*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk21c1x0123, vget_high_f32(vi4x2), 1);
306*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk21c1x4567, vget_high_f32(vi2x2), 1);
307*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk21c1x4567, vget_high_f32(vi4x2), 1);
308*4bdc9457SAndroid Build Coastguard Worker 
309*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c2x0123 = vld1q_f32(w + 128);
310*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c2x4567 = vld1q_f32(w + 132);
311*4bdc9457SAndroid Build Coastguard Worker 
312*4bdc9457SAndroid Build Coastguard Worker         // viMx3 = ( iM4c2, iM4c1, iM4c0, iM3c2 )
313*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi0x3 = vld1q_f32(i0); i0 += 4;
314*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi1x3 = vld1q_f32(i1); i1 += 4;
315*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi2x3 = vld1q_f32(i2); i2 += 4;
316*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi3x3 = vld1q_f32(i3); i3 += 4;
317*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi4x3 = vld1q_f32(i4); i4 += 4;
318*4bdc9457SAndroid Build Coastguard Worker 
319*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0);
320*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);
321*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk01c2x4567, vget_high_f32(vi0x1), 0);
322*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk01c2x4567, vget_high_f32(vi2x1), 0);
323*4bdc9457SAndroid Build Coastguard Worker 
324*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c2x0123, vget_low_f32(vi0x3), 0);
325*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c2x0123, vget_low_f32(vi2x3), 0);
326*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk01c2x4567, vget_low_f32(vi0x3), 0);
327*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk01c2x4567, vget_low_f32(vi2x3), 0);
328*4bdc9457SAndroid Build Coastguard Worker 
329*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c2x0123 = vld1q_f32(w + 136);
330*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c2x4567 = vld1q_f32(w + 140);
331*4bdc9457SAndroid Build Coastguard Worker 
332*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0);
333*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c2x0123, vget_high_f32(vi3x1), 0);
334*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk11c2x4567, vget_high_f32(vi1x1), 0);
335*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk11c2x4567, vget_high_f32(vi3x1), 0);
336*4bdc9457SAndroid Build Coastguard Worker 
337*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c2x0123, vget_low_f32(vi1x3), 0);
338*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c2x0123, vget_low_f32(vi3x3), 0);
339*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk11c2x4567, vget_low_f32(vi1x3), 0);
340*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk11c2x4567, vget_low_f32(vi3x3), 0);
341*4bdc9457SAndroid Build Coastguard Worker 
342*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c2x0123 = vld1q_f32(w + 144);
343*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c2x4567 = vld1q_f32(w + 148);
344*4bdc9457SAndroid Build Coastguard Worker 
345*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0);
346*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c2x0123, vget_high_f32(vi4x1), 0);
347*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0);
348*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk21c2x4567, vget_high_f32(vi4x1), 0);
349*4bdc9457SAndroid Build Coastguard Worker 
350*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk21c2x0123, vget_low_f32(vi2x3), 0);
351*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk21c2x0123, vget_low_f32(vi4x3), 0);
352*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk21c2x4567, vget_low_f32(vi2x3), 0);
353*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk21c2x4567, vget_low_f32(vi4x3), 0);
354*4bdc9457SAndroid Build Coastguard Worker 
355*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk02c0x0123 = vld1q_f32(w + 152);
356*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk02c0x4567 = vld1q_f32(w + 156);
357*4bdc9457SAndroid Build Coastguard Worker 
358*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c0x0123, vget_high_f32(vi0x1), 1);
359*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c0x0123, vget_high_f32(vi2x1), 1);
360*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk02c0x4567, vget_high_f32(vi0x1), 1);
361*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk02c0x4567, vget_high_f32(vi2x1), 1);
362*4bdc9457SAndroid Build Coastguard Worker 
363*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk02c0x0123, vget_low_f32(vi0x3), 1);
364*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk02c0x0123, vget_low_f32(vi2x3), 1);
365*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk02c0x4567, vget_low_f32(vi0x3), 1);
366*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk02c0x4567, vget_low_f32(vi2x3), 1);
367*4bdc9457SAndroid Build Coastguard Worker 
368*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk12c0x0123 = vld1q_f32(w + 160);
369*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk12c0x4567 = vld1q_f32(w + 164);
370*4bdc9457SAndroid Build Coastguard Worker 
371*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c0x0123, vget_high_f32(vi1x1), 1);
372*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c0x0123, vget_high_f32(vi3x1), 1);
373*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk12c0x4567, vget_high_f32(vi1x1), 1);
374*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk12c0x4567, vget_high_f32(vi3x1), 1);
375*4bdc9457SAndroid Build Coastguard Worker 
376*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c0x0123, vget_low_f32(vi1x3), 1);
377*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c0x0123, vget_low_f32(vi3x3), 1);
378*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk12c0x4567, vget_low_f32(vi1x3), 1);
379*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk12c0x4567, vget_low_f32(vi3x3), 1);
380*4bdc9457SAndroid Build Coastguard Worker 
381*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk22c0x0123 = vld1q_f32(w + 168);
382*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk22c0x4567 = vld1q_f32(w + 172);
383*4bdc9457SAndroid Build Coastguard Worker 
384*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1);
385*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c0x0123, vget_high_f32(vi4x1), 1);
386*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk22c0x4567, vget_high_f32(vi2x1), 1);
387*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk22c0x4567, vget_high_f32(vi4x1), 1);
388*4bdc9457SAndroid Build Coastguard Worker 
389*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c0x0123, vget_low_f32(vi2x3), 1);
390*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c0x0123, vget_low_f32(vi4x3), 1);
391*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk22c0x4567, vget_low_f32(vi2x3), 1);
392*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk22c0x4567, vget_low_f32(vi4x3), 1);
393*4bdc9457SAndroid Build Coastguard Worker 
394*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk02c1x0123 = vld1q_f32(w + 176);
395*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk02c1x4567 = vld1q_f32(w + 180);
396*4bdc9457SAndroid Build Coastguard Worker 
397*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c1x0123, vget_low_f32(vi0x2), 0);
398*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c1x0123, vget_low_f32(vi2x2), 0);
399*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk02c1x4567, vget_low_f32(vi0x2), 0);
400*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk02c1x4567, vget_low_f32(vi2x2), 0);
401*4bdc9457SAndroid Build Coastguard Worker 
402*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk02c1x0123, vget_high_f32(vi0x3), 0);
403*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk02c1x0123, vget_high_f32(vi2x3), 0);
404*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk02c1x4567, vget_high_f32(vi0x3), 0);
405*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk02c1x4567, vget_high_f32(vi2x3), 0);
406*4bdc9457SAndroid Build Coastguard Worker 
407*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk12c1x0123 = vld1q_f32(w + 184);
408*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk12c1x4567 = vld1q_f32(w + 188);
409*4bdc9457SAndroid Build Coastguard Worker 
410*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c1x0123, vget_low_f32(vi1x2), 0);
411*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c1x0123, vget_low_f32(vi3x2), 0);
412*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk12c1x4567, vget_low_f32(vi1x2), 0);
413*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk12c1x4567, vget_low_f32(vi3x2), 0);
414*4bdc9457SAndroid Build Coastguard Worker 
415*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c1x0123, vget_high_f32(vi1x3), 0);
416*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c1x0123, vget_high_f32(vi3x3), 0);
417*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk12c1x4567, vget_high_f32(vi1x3), 0);
418*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk12c1x4567, vget_high_f32(vi3x3), 0);
419*4bdc9457SAndroid Build Coastguard Worker 
420*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk22c1x0123 = vld1q_f32(w + 192);
421*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk22c1x4567 = vld1q_f32(w + 196);
422*4bdc9457SAndroid Build Coastguard Worker 
423*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_low_f32(vi2x2), 0);
424*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c1x0123, vget_low_f32(vi4x2), 0);
425*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk22c1x4567, vget_low_f32(vi2x2), 0);
426*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk22c1x4567, vget_low_f32(vi4x2), 0);
427*4bdc9457SAndroid Build Coastguard Worker 
428*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c1x0123, vget_high_f32(vi2x3), 0);
429*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c1x0123, vget_high_f32(vi4x3), 0);
430*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk22c1x4567, vget_high_f32(vi2x3), 0);
431*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk22c1x4567, vget_high_f32(vi4x3), 0);
432*4bdc9457SAndroid Build Coastguard Worker 
433*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk02c2x0123 = vld1q_f32(w + 200);
434*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk02c2x4567 = vld1q_f32(w + 204);
435*4bdc9457SAndroid Build Coastguard Worker 
436*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c2x0123, vget_low_f32(vi0x2), 1);
437*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 1);
438*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk02c2x4567, vget_low_f32(vi0x2), 1);
439*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk02c2x4567, vget_low_f32(vi2x2), 1);
440*4bdc9457SAndroid Build Coastguard Worker 
441*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk02c2x0123, vget_high_f32(vi0x3), 1);
442*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk02c2x0123, vget_high_f32(vi2x3), 1);
443*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk02c2x4567, vget_high_f32(vi0x3), 1);
444*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk02c2x4567, vget_high_f32(vi2x3), 1);
445*4bdc9457SAndroid Build Coastguard Worker 
446*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk12c2x0123 = vld1q_f32(w + 208);
447*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk12c2x4567 = vld1q_f32(w + 212);
448*4bdc9457SAndroid Build Coastguard Worker 
449*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c2x0123, vget_low_f32(vi1x2), 1);
450*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c2x0123, vget_low_f32(vi3x2), 1);
451*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk12c2x4567, vget_low_f32(vi1x2), 1);
452*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk12c2x4567, vget_low_f32(vi3x2), 1);
453*4bdc9457SAndroid Build Coastguard Worker 
454*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c2x0123, vget_high_f32(vi1x3), 1);
455*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c2x0123, vget_high_f32(vi3x3), 1);
456*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk12c2x4567, vget_high_f32(vi1x3), 1);
457*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk12c2x4567, vget_high_f32(vi3x3), 1);
458*4bdc9457SAndroid Build Coastguard Worker 
459*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk22c2x0123 = vld1q_f32(w + 216);
460*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk22c2x4567 = vld1q_f32(w + 220);
461*4bdc9457SAndroid Build Coastguard Worker 
462*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c2x0123, vget_low_f32(vi2x2), 1);
463*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c2x0123, vget_low_f32(vi4x2), 1);
464*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk22c2x4567, vget_low_f32(vi2x2), 1);
465*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk22c2x4567, vget_low_f32(vi4x2), 1);
466*4bdc9457SAndroid Build Coastguard Worker 
467*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c2x0123, vget_high_f32(vi2x3), 1);
468*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c2x0123, vget_high_f32(vi4x3), 1);
469*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk22c2x4567, vget_high_f32(vi2x3), 1);
470*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk22c2x4567, vget_high_f32(vi4x3), 1);
471*4bdc9457SAndroid Build Coastguard Worker 
472*4bdc9457SAndroid Build Coastguard Worker         vi0x0 = vi0x3;
473*4bdc9457SAndroid Build Coastguard Worker         vi1x0 = vi1x3;
474*4bdc9457SAndroid Build Coastguard Worker         vi2x0 = vi2x3;
475*4bdc9457SAndroid Build Coastguard Worker         vi3x0 = vi3x3;
476*4bdc9457SAndroid Build Coastguard Worker         vi4x0 = vi4x3;
477*4bdc9457SAndroid Build Coastguard Worker 
478*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
479*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
480*4bdc9457SAndroid Build Coastguard Worker 
481*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vmaxq_f32(vo0x0c0123, vmin);
482*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vmaxq_f32(vo1x0c0123, vmin);
483*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vmaxq_f32(vo0x0c4567, vmin);
484*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vmaxq_f32(vo1x0c4567, vmin);
485*4bdc9457SAndroid Build Coastguard Worker 
486*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vmaxq_f32(vo0x1c0123, vmin);
487*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vmaxq_f32(vo1x1c0123, vmin);
488*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vmaxq_f32(vo0x1c4567, vmin);
489*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vmaxq_f32(vo1x1c4567, vmin);
490*4bdc9457SAndroid Build Coastguard Worker 
491*4bdc9457SAndroid Build Coastguard Worker         vo0x0c0123 = vminq_f32(vo0x0c0123, vmax);
492*4bdc9457SAndroid Build Coastguard Worker         vo1x0c0123 = vminq_f32(vo1x0c0123, vmax);
493*4bdc9457SAndroid Build Coastguard Worker         vo0x0c4567 = vminq_f32(vo0x0c4567, vmax);
494*4bdc9457SAndroid Build Coastguard Worker         vo1x0c4567 = vminq_f32(vo1x0c4567, vmax);
495*4bdc9457SAndroid Build Coastguard Worker 
496*4bdc9457SAndroid Build Coastguard Worker         vo0x1c0123 = vminq_f32(vo0x1c0123, vmax);
497*4bdc9457SAndroid Build Coastguard Worker         vo1x1c0123 = vminq_f32(vo1x1c0123, vmax);
498*4bdc9457SAndroid Build Coastguard Worker         vo0x1c4567 = vminq_f32(vo0x1c4567, vmax);
499*4bdc9457SAndroid Build Coastguard Worker         vo1x1c4567 = vminq_f32(vo1x1c4567, vmax);
500*4bdc9457SAndroid Build Coastguard Worker 
501*4bdc9457SAndroid Build Coastguard Worker         if XNN_LIKELY(c >= 8) {
502*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o1, vo1x0c0123);
503*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o1 + 4, vo1x0c4567);
504*4bdc9457SAndroid Build Coastguard Worker           o1 = (float*) ((uintptr_t) o1 + output_width_stride);
505*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o0, vo0x0c0123);
506*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o0 + 4, vo0x0c4567);
507*4bdc9457SAndroid Build Coastguard Worker           o0 = (float*) ((uintptr_t) o0 + output_width_stride);
508*4bdc9457SAndroid Build Coastguard Worker 
509*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o1, vo1x1c0123);
510*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o1 + 4, vo1x1c4567);
511*4bdc9457SAndroid Build Coastguard Worker           o1 = (float*) ((uintptr_t) o1 + output_width_stride);
512*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o0, vo0x1c0123);
513*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o0 + 4, vo0x1c4567);
514*4bdc9457SAndroid Build Coastguard Worker           o0 = (float*) ((uintptr_t) o0 + output_width_stride);
515*4bdc9457SAndroid Build Coastguard Worker         } else {
516*4bdc9457SAndroid Build Coastguard Worker           float* o0_tmp = o0;
517*4bdc9457SAndroid Build Coastguard Worker           float* o1_tmp = o1;
518*4bdc9457SAndroid Build Coastguard Worker           if (c & 4) {
519*4bdc9457SAndroid Build Coastguard Worker             vst1q_f32((float*) ((uintptr_t) o1_tmp + output_width_stride), vo1x1c0123);
520*4bdc9457SAndroid Build Coastguard Worker             vo1x1c0123 = vo1x1c4567;
521*4bdc9457SAndroid Build Coastguard Worker             vst1q_f32((float*) ((uintptr_t) o0_tmp + output_width_stride), vo0x1c0123);
522*4bdc9457SAndroid Build Coastguard Worker             vo0x1c0123 = vo0x1c4567;
523*4bdc9457SAndroid Build Coastguard Worker 
524*4bdc9457SAndroid Build Coastguard Worker             vst1q_f32(o1_tmp, vo1x0c0123); o1_tmp += 4;
525*4bdc9457SAndroid Build Coastguard Worker             vo1x0c0123 = vo1x0c4567;
526*4bdc9457SAndroid Build Coastguard Worker             vst1q_f32(o0_tmp, vo0x0c0123); o0_tmp += 4;
527*4bdc9457SAndroid Build Coastguard Worker             vo0x0c0123 = vo0x0c4567;
528*4bdc9457SAndroid Build Coastguard Worker           }
529*4bdc9457SAndroid Build Coastguard Worker           float32x2_t vo0x0c01 = vget_low_f32(vo0x0c0123);
530*4bdc9457SAndroid Build Coastguard Worker           float32x2_t vo1x0c01 = vget_low_f32(vo1x0c0123);
531*4bdc9457SAndroid Build Coastguard Worker           float32x2_t vo0x1c01 = vget_low_f32(vo0x1c0123);
532*4bdc9457SAndroid Build Coastguard Worker           float32x2_t vo1x1c01 = vget_low_f32(vo1x1c0123);
533*4bdc9457SAndroid Build Coastguard Worker           if (c & 2) {
534*4bdc9457SAndroid Build Coastguard Worker             vst1_f32((float*) ((uintptr_t) o1_tmp + output_width_stride), vo1x1c01);
535*4bdc9457SAndroid Build Coastguard Worker             vo1x1c01 = vget_high_f32(vo1x1c0123);
536*4bdc9457SAndroid Build Coastguard Worker             vst1_f32((float*) ((uintptr_t) o0_tmp + output_width_stride), vo0x1c01);
537*4bdc9457SAndroid Build Coastguard Worker             vo0x1c01 = vget_high_f32(vo0x1c0123);
538*4bdc9457SAndroid Build Coastguard Worker 
539*4bdc9457SAndroid Build Coastguard Worker             vst1_f32(o1_tmp, vo1x0c01); o1_tmp += 2;
540*4bdc9457SAndroid Build Coastguard Worker             vo1x0c01 = vget_high_f32(vo1x0c0123);
541*4bdc9457SAndroid Build Coastguard Worker             vst1_f32(o0_tmp, vo0x0c01); o0_tmp += 2;
542*4bdc9457SAndroid Build Coastguard Worker             vo0x0c01 = vget_high_f32(vo0x0c0123);
543*4bdc9457SAndroid Build Coastguard Worker           }
544*4bdc9457SAndroid Build Coastguard Worker           if (c & 1) {
545*4bdc9457SAndroid Build Coastguard Worker             vst1_lane_f32(o1_tmp, vo1x0c01, 0);
546*4bdc9457SAndroid Build Coastguard Worker             vst1_lane_f32(o0_tmp, vo0x0c01, 0);
547*4bdc9457SAndroid Build Coastguard Worker 
548*4bdc9457SAndroid Build Coastguard Worker             vst1_lane_f32((float*) ((uintptr_t) o1_tmp + output_width_stride), vo1x1c01, 0);
549*4bdc9457SAndroid Build Coastguard Worker             vst1_lane_f32((float*) ((uintptr_t) o0_tmp + output_width_stride), vo0x1c01, 0);
550*4bdc9457SAndroid Build Coastguard Worker           }
551*4bdc9457SAndroid Build Coastguard Worker 
552*4bdc9457SAndroid Build Coastguard Worker           o0 = (float*) ((uintptr_t) o0 + output_width_stride * 2);
553*4bdc9457SAndroid Build Coastguard Worker           o1 = (float*) ((uintptr_t) o1 + output_width_stride * 2);
554*4bdc9457SAndroid Build Coastguard Worker         }
555*4bdc9457SAndroid Build Coastguard Worker       }
556*4bdc9457SAndroid Build Coastguard Worker       assert(iw < 4);
557*4bdc9457SAndroid Build Coastguard Worker       if XNN_UNLIKELY(iw & 2) {
558*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo0c0123 = vld1q_f32(w);
559*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo0c4567 = vld1q_f32(w + 4);
560*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo1c0123 = vo0c0123;
561*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo1c4567 = vo0c4567;
562*4bdc9457SAndroid Build Coastguard Worker 
563*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c0x0123 = vld1q_f32(w + 8);
564*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c0x4567 = vld1q_f32(w + 12);
565*4bdc9457SAndroid Build Coastguard Worker 
566*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk00c0x0123, vget_low_f32(vi0x0), 1);
567*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);
568*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 1);
569*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c0x4567, vget_low_f32(vi2x0), 1);
570*4bdc9457SAndroid Build Coastguard Worker 
571*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c0x0123 = vld1q_f32(w + 16);
572*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c0x4567 = vld1q_f32(w + 20);
573*4bdc9457SAndroid Build Coastguard Worker 
574*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1);
575*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1);
576*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 1);
577*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c0x4567, vget_low_f32(vi3x0), 1);
578*4bdc9457SAndroid Build Coastguard Worker 
579*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c0x0123 = vld1q_f32(w + 24);
580*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c0x4567 = vld1q_f32(w + 28);
581*4bdc9457SAndroid Build Coastguard Worker 
582*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1);
583*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1);
584*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1);
585*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 1);
586*4bdc9457SAndroid Build Coastguard Worker 
587*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c1x0123 = vld1q_f32(w + 32);
588*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c1x4567 = vld1q_f32(w + 36);
589*4bdc9457SAndroid Build Coastguard Worker 
590*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk00c1x0123, vget_high_f32(vi0x0), 0);
591*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c1x0123, vget_high_f32(vi2x0), 0);
592*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c1x4567, vget_high_f32(vi0x0), 0);
593*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c1x4567, vget_high_f32(vi2x0), 0);
594*4bdc9457SAndroid Build Coastguard Worker 
595*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c1x0123 = vld1q_f32(w + 40);
596*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c1x4567 = vld1q_f32(w + 44);
597*4bdc9457SAndroid Build Coastguard Worker 
598*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0);
599*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0);
600*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c1x4567, vget_high_f32(vi1x0), 0);
601*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c1x4567, vget_high_f32(vi3x0), 0);
602*4bdc9457SAndroid Build Coastguard Worker 
603*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c1x0123 = vld1q_f32(w + 48);
604*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c1x4567 = vld1q_f32(w + 52);
605*4bdc9457SAndroid Build Coastguard Worker 
606*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c1x0123, vget_high_f32(vi2x0), 0);
607*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c1x0123, vget_high_f32(vi4x0), 0);
608*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0);
609*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c1x4567, vget_high_f32(vi4x0), 0);
610*4bdc9457SAndroid Build Coastguard Worker 
611*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c2x0123 = vld1q_f32(w + 56);
612*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c2x4567 = vld1q_f32(w + 60);
613*4bdc9457SAndroid Build Coastguard Worker 
614*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk00c2x0123, vget_high_f32(vi0x0), 1);
615*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 1);
616*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 1);
617*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c2x4567, vget_high_f32(vi2x0), 1);
618*4bdc9457SAndroid Build Coastguard Worker 
619*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c2x0123 = vld1q_f32(w + 64);
620*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c2x4567 = vld1q_f32(w + 68);
621*4bdc9457SAndroid Build Coastguard Worker 
622*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c2x0123, vget_high_f32(vi1x0), 1);
623*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 1);
624*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1);
625*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 1);
626*4bdc9457SAndroid Build Coastguard Worker 
627*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c2x0123 = vld1q_f32(w + 72);
628*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c2x4567 = vld1q_f32(w + 76);
629*4bdc9457SAndroid Build Coastguard Worker 
630*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c2x0123, vget_high_f32(vi2x0), 1);
631*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 1);
632*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1);
633*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 1);
634*4bdc9457SAndroid Build Coastguard Worker 
635*4bdc9457SAndroid Build Coastguard Worker         // viMx1 = ( iM2c0, iM1c2, iM1c1, iM1c0 )
636*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi0x1 = vld1q_f32(i0); i0 += 4;
637*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi1x1 = vld1q_f32(i1); i1 += 4;
638*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4;
639*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi3x1 = vld1q_f32(i3); i3 += 4;
640*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi4x1 = vld1q_f32(i4); i4 += 4;
641*4bdc9457SAndroid Build Coastguard Worker 
642*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c0x0123 = vld1q_f32(w + 80);
643*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c0x4567 = vld1q_f32(w + 84);
644*4bdc9457SAndroid Build Coastguard Worker 
645*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0);
646*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);
647*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_low_f32(vi0x1), 0);
648*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c0x4567, vget_low_f32(vi2x1), 0);
649*4bdc9457SAndroid Build Coastguard Worker 
650*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c0x0123 = vld1q_f32(w + 88);
651*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c0x4567 = vld1q_f32(w + 92);
652*4bdc9457SAndroid Build Coastguard Worker 
653*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0);
654*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);
655*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_low_f32(vi1x1), 0);
656*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c0x4567, vget_low_f32(vi3x1), 0);
657*4bdc9457SAndroid Build Coastguard Worker 
658*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c0x0123 = vld1q_f32(w + 96);
659*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c0x4567 = vld1q_f32(w + 100);
660*4bdc9457SAndroid Build Coastguard Worker 
661*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0);
662*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c0x0123, vget_low_f32(vi4x1), 0);
663*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0);
664*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c0x4567, vget_low_f32(vi4x1), 0);
665*4bdc9457SAndroid Build Coastguard Worker 
666*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c1x0123 = vld1q_f32(w + 104);
667*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c1x4567 = vld1q_f32(w + 108);
668*4bdc9457SAndroid Build Coastguard Worker 
669*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c1x0123, vget_low_f32(vi0x1), 1);
670*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);
671*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 1);
672*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 1);
673*4bdc9457SAndroid Build Coastguard Worker 
674*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c1x0123 = vld1q_f32(w + 112);
675*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c1x4567 = vld1q_f32(w + 116);
676*4bdc9457SAndroid Build Coastguard Worker 
677*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1);
678*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1);
679*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c1x4567, vget_low_f32(vi1x1), 1);
680*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c1x4567, vget_low_f32(vi3x1), 1);
681*4bdc9457SAndroid Build Coastguard Worker 
682*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c1x0123 = vld1q_f32(w + 120);
683*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c1x4567 = vld1q_f32(w + 124);
684*4bdc9457SAndroid Build Coastguard Worker 
685*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1);
686*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1);
687*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1);
688*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c1x4567, vget_low_f32(vi4x1), 1);
689*4bdc9457SAndroid Build Coastguard Worker 
690*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c2x0123 = vld1q_f32(w + 128);
691*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c2x4567 = vld1q_f32(w + 132);
692*4bdc9457SAndroid Build Coastguard Worker 
693*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0);
694*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);
695*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c2x4567, vget_high_f32(vi0x1), 0);
696*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c2x4567, vget_high_f32(vi2x1), 0);
697*4bdc9457SAndroid Build Coastguard Worker 
698*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c2x0123 = vld1q_f32(w + 136);
699*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c2x4567 = vld1q_f32(w + 140);
700*4bdc9457SAndroid Build Coastguard Worker 
701*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0);
702*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0);
703*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vget_high_f32(vi1x1), 0);
704*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c2x4567, vget_high_f32(vi3x1), 0);
705*4bdc9457SAndroid Build Coastguard Worker 
706*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c2x0123 = vld1q_f32(w + 144);
707*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c2x4567 = vld1q_f32(w + 148);
708*4bdc9457SAndroid Build Coastguard Worker 
709*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0);
710*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c2x0123, vget_high_f32(vi4x1), 0);
711*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0);
712*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c2x4567, vget_high_f32(vi4x1), 0);
713*4bdc9457SAndroid Build Coastguard Worker 
714*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk02c0x0123 = vld1q_f32(w + 152);
715*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk02c0x4567 = vld1q_f32(w + 156);
716*4bdc9457SAndroid Build Coastguard Worker 
717*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c0x0123, vget_high_f32(vi0x1), 1);
718*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 1);
719*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c0x4567, vget_high_f32(vi0x1), 1);
720*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c0x4567, vget_high_f32(vi2x1), 1);
721*4bdc9457SAndroid Build Coastguard Worker 
722*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk12c0x0123 = vld1q_f32(w + 160);
723*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk12c0x4567 = vld1q_f32(w + 164);
724*4bdc9457SAndroid Build Coastguard Worker 
725*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c0x0123, vget_high_f32(vi1x1), 1);
726*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c0x0123, vget_high_f32(vi3x1), 1);
727*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c0x4567, vget_high_f32(vi1x1), 1);
728*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c0x4567, vget_high_f32(vi3x1), 1);
729*4bdc9457SAndroid Build Coastguard Worker 
730*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk22c0x0123 = vld1q_f32(w + 168);
731*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk22c0x4567 = vld1q_f32(w + 172);
732*4bdc9457SAndroid Build Coastguard Worker 
733*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1);
734*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 1);
735*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 1);
736*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c0x4567, vget_high_f32(vi4x1), 1);
737*4bdc9457SAndroid Build Coastguard Worker 
738*4bdc9457SAndroid Build Coastguard Worker         // viMx2 = ( iM2c2, iM2c1 )
739*4bdc9457SAndroid Build Coastguard Worker         const float32x2_t vi0x2 = vld1_f32(i0); i0 += 2;
740*4bdc9457SAndroid Build Coastguard Worker         const float32x2_t vi1x2 = vld1_f32(i1); i1 += 2;
741*4bdc9457SAndroid Build Coastguard Worker         const float32x2_t vi2x2 = vld1_f32(i2); i2 += 2;
742*4bdc9457SAndroid Build Coastguard Worker         const float32x2_t vi3x2 = vld1_f32(i3); i3 += 2;
743*4bdc9457SAndroid Build Coastguard Worker         const float32x2_t vi4x2 = vld1_f32(i4); i4 += 2;
744*4bdc9457SAndroid Build Coastguard Worker 
745*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk02c1x0123 = vld1q_f32(w + 176);
746*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk02c1x4567 = vld1q_f32(w + 180);
747*4bdc9457SAndroid Build Coastguard Worker 
748*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c1x0123, vi0x2, 0);
749*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c1x0123, vi2x2, 0);
750*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c1x4567, vi0x2, 0);
751*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c1x4567, vi2x2, 0);
752*4bdc9457SAndroid Build Coastguard Worker 
753*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk12c1x0123 = vld1q_f32(w + 184);
754*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk12c1x4567 = vld1q_f32(w + 188);
755*4bdc9457SAndroid Build Coastguard Worker 
756*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c1x0123, vi1x2, 0);
757*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c1x0123, vi3x2, 0);
758*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c1x4567, vi1x2, 0);
759*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c1x4567, vi3x2, 0);
760*4bdc9457SAndroid Build Coastguard Worker 
761*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk22c1x0123 = vld1q_f32(w + 192);
762*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk22c1x4567 = vld1q_f32(w + 196);
763*4bdc9457SAndroid Build Coastguard Worker 
764*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c1x0123, vi2x2, 0);
765*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c1x0123, vi4x2, 0);
766*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c1x4567, vi2x2, 0);
767*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c1x4567, vi4x2, 0);
768*4bdc9457SAndroid Build Coastguard Worker 
769*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk02c2x0123 = vld1q_f32(w + 200);
770*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk02c2x4567 = vld1q_f32(w + 204);
771*4bdc9457SAndroid Build Coastguard Worker 
772*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 1);
773*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1);
774*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c2x4567, vi0x2, 1);
775*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c2x4567, vi2x2, 1);
776*4bdc9457SAndroid Build Coastguard Worker 
777*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk12c2x0123 = vld1q_f32(w + 208);
778*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk12c2x4567 = vld1q_f32(w + 212);
779*4bdc9457SAndroid Build Coastguard Worker 
780*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 1);
781*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 1);
782*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c2x4567, vi1x2, 1);
783*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c2x4567, vi3x2, 1);
784*4bdc9457SAndroid Build Coastguard Worker 
785*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk22c2x0123 = vld1q_f32(w + 216);
786*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk22c2x4567 = vld1q_f32(w + 220);
787*4bdc9457SAndroid Build Coastguard Worker 
788*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c2x0123, vi2x2, 1);
789*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c2x0123, vi4x2, 1);
790*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 1);
791*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c2x4567, vi4x2, 1);
792*4bdc9457SAndroid Build Coastguard Worker 
793*4bdc9457SAndroid Build Coastguard Worker         vi0x0 = vcombine_f32(vget_high_f32(vi0x1), vi0x2);
794*4bdc9457SAndroid Build Coastguard Worker         vi1x0 = vcombine_f32(vget_high_f32(vi1x1), vi1x2);
795*4bdc9457SAndroid Build Coastguard Worker         vi2x0 = vcombine_f32(vget_high_f32(vi2x1), vi2x2);
796*4bdc9457SAndroid Build Coastguard Worker         vi3x0 = vcombine_f32(vget_high_f32(vi3x1), vi3x2);
797*4bdc9457SAndroid Build Coastguard Worker         vi4x0 = vcombine_f32(vget_high_f32(vi4x1), vi4x2);
798*4bdc9457SAndroid Build Coastguard Worker 
799*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
800*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
801*4bdc9457SAndroid Build Coastguard Worker 
802*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmaxq_f32(vo0c0123, vmin);
803*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmaxq_f32(vo1c0123, vmin);
804*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmaxq_f32(vo0c4567, vmin);
805*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmaxq_f32(vo1c4567, vmin);
806*4bdc9457SAndroid Build Coastguard Worker 
807*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vminq_f32(vo0c0123, vmax);
808*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vminq_f32(vo1c0123, vmax);
809*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vminq_f32(vo0c4567, vmax);
810*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vminq_f32(vo1c4567, vmax);
811*4bdc9457SAndroid Build Coastguard Worker 
812*4bdc9457SAndroid Build Coastguard Worker         if XNN_LIKELY(c >= 8) {
813*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o1, vo1c0123);
814*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o1 + 4, vo1c4567);
815*4bdc9457SAndroid Build Coastguard Worker           o1 = (float*) ((uintptr_t) o1 + output_width_stride);
816*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o0, vo0c0123);
817*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o0 + 4, vo0c4567);
818*4bdc9457SAndroid Build Coastguard Worker           o0 = (float*) ((uintptr_t) o0 + output_width_stride);
819*4bdc9457SAndroid Build Coastguard Worker         } else {
820*4bdc9457SAndroid Build Coastguard Worker           float* o0_tmp = o0;
821*4bdc9457SAndroid Build Coastguard Worker           float* o1_tmp = o1;
822*4bdc9457SAndroid Build Coastguard Worker           if (c & 4) {
823*4bdc9457SAndroid Build Coastguard Worker             vst1q_f32(o1_tmp, vo1c0123); o1_tmp += 4;
824*4bdc9457SAndroid Build Coastguard Worker             vo1c0123 = vo1c4567;
825*4bdc9457SAndroid Build Coastguard Worker             vst1q_f32(o0_tmp, vo0c0123); o0_tmp += 4;
826*4bdc9457SAndroid Build Coastguard Worker             vo0c0123 = vo0c4567;
827*4bdc9457SAndroid Build Coastguard Worker           }
828*4bdc9457SAndroid Build Coastguard Worker           float32x2_t vo0c01 = vget_low_f32(vo0c0123);
829*4bdc9457SAndroid Build Coastguard Worker           float32x2_t vo1c01 = vget_low_f32(vo1c0123);
830*4bdc9457SAndroid Build Coastguard Worker           if (c & 2) {
831*4bdc9457SAndroid Build Coastguard Worker             vst1_f32(o1_tmp, vo1c01); o1_tmp += 2;
832*4bdc9457SAndroid Build Coastguard Worker             vo1c01 = vget_high_f32(vo1c0123);
833*4bdc9457SAndroid Build Coastguard Worker             vst1_f32(o0_tmp, vo0c01); o0_tmp += 2;
834*4bdc9457SAndroid Build Coastguard Worker             vo0c01 = vget_high_f32(vo0c0123);
835*4bdc9457SAndroid Build Coastguard Worker           }
836*4bdc9457SAndroid Build Coastguard Worker           if (c & 1) {
837*4bdc9457SAndroid Build Coastguard Worker             vst1_lane_f32(o1_tmp, vo1c01, 0);
838*4bdc9457SAndroid Build Coastguard Worker             vst1_lane_f32(o0_tmp, vo0c01, 0);
839*4bdc9457SAndroid Build Coastguard Worker           }
840*4bdc9457SAndroid Build Coastguard Worker 
841*4bdc9457SAndroid Build Coastguard Worker           o0 = (float*) ((uintptr_t) o0 + output_width_stride);
842*4bdc9457SAndroid Build Coastguard Worker           o1 = (float*) ((uintptr_t) o1 + output_width_stride);
843*4bdc9457SAndroid Build Coastguard Worker         }
844*4bdc9457SAndroid Build Coastguard Worker       }
845*4bdc9457SAndroid Build Coastguard Worker       if XNN_UNLIKELY(iw & 1) {
846*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo0c0123 = vld1q_f32(w);
847*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo0c4567 = vld1q_f32(w + 4);
848*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo1c0123 = vo0c0123;
849*4bdc9457SAndroid Build Coastguard Worker         float32x4_t vo1c4567 = vo0c4567;
850*4bdc9457SAndroid Build Coastguard Worker 
851*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c0x0123 = vld1q_f32(w + 8);
852*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c0x4567 = vld1q_f32(w + 12);
853*4bdc9457SAndroid Build Coastguard Worker 
854*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk00c0x0123, vget_low_f32(vi0x0), 1);
855*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);
856*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 1);
857*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c0x4567, vget_low_f32(vi2x0), 1);
858*4bdc9457SAndroid Build Coastguard Worker 
859*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c0x0123 = vld1q_f32(w + 16);
860*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c0x4567 = vld1q_f32(w + 20);
861*4bdc9457SAndroid Build Coastguard Worker 
862*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1);
863*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1);
864*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 1);
865*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c0x4567, vget_low_f32(vi3x0), 1);
866*4bdc9457SAndroid Build Coastguard Worker 
867*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c0x0123 = vld1q_f32(w + 24);
868*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c0x4567 = vld1q_f32(w + 28);
869*4bdc9457SAndroid Build Coastguard Worker 
870*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1);
871*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1);
872*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1);
873*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 1);
874*4bdc9457SAndroid Build Coastguard Worker 
875*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c1x0123 = vld1q_f32(w + 32);
876*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c1x4567 = vld1q_f32(w + 36);
877*4bdc9457SAndroid Build Coastguard Worker 
878*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk00c1x0123, vget_high_f32(vi0x0), 0);
879*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c1x0123, vget_high_f32(vi2x0), 0);
880*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c1x4567, vget_high_f32(vi0x0), 0);
881*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c1x4567, vget_high_f32(vi2x0), 0);
882*4bdc9457SAndroid Build Coastguard Worker 
883*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c1x0123 = vld1q_f32(w + 40);
884*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c1x4567 = vld1q_f32(w + 44);
885*4bdc9457SAndroid Build Coastguard Worker 
886*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0);
887*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0);
888*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c1x4567, vget_high_f32(vi1x0), 0);
889*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c1x4567, vget_high_f32(vi3x0), 0);
890*4bdc9457SAndroid Build Coastguard Worker 
891*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c1x0123 = vld1q_f32(w + 48);
892*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c1x4567 = vld1q_f32(w + 52);
893*4bdc9457SAndroid Build Coastguard Worker 
894*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c1x0123, vget_high_f32(vi2x0), 0);
895*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c1x0123, vget_high_f32(vi4x0), 0);
896*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0);
897*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c1x4567, vget_high_f32(vi4x0), 0);
898*4bdc9457SAndroid Build Coastguard Worker 
899*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c2x0123 = vld1q_f32(w + 56);
900*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk00c2x4567 = vld1q_f32(w + 60);
901*4bdc9457SAndroid Build Coastguard Worker 
902*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk00c2x0123, vget_high_f32(vi0x0), 1);
903*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 1);
904*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 1);
905*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c2x4567, vget_high_f32(vi2x0), 1);
906*4bdc9457SAndroid Build Coastguard Worker 
907*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c2x0123 = vld1q_f32(w + 64);
908*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk10c2x4567 = vld1q_f32(w + 68);
909*4bdc9457SAndroid Build Coastguard Worker 
910*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c2x0123, vget_high_f32(vi1x0), 1);
911*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 1);
912*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1);
913*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 1);
914*4bdc9457SAndroid Build Coastguard Worker 
915*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c2x0123 = vld1q_f32(w + 72);
916*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk20c2x4567 = vld1q_f32(w + 76);
917*4bdc9457SAndroid Build Coastguard Worker 
918*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c2x0123, vget_high_f32(vi2x0), 1);
919*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 1);
920*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1);
921*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 1);
922*4bdc9457SAndroid Build Coastguard Worker 
923*4bdc9457SAndroid Build Coastguard Worker         // viMx1 = ( iM2c0, iM1c2, iM1c1, iM1c0 )
924*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi0x1 = vld1q_f32(i0); i0 += 3;
925*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi1x1 = vld1q_f32(i1); i1 += 3;
926*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 3;
927*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi3x1 = vld1q_f32(i3); i3 += 3;
928*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vi4x1 = vld1q_f32(i4); i4 += 3;
929*4bdc9457SAndroid Build Coastguard Worker 
930*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c0x0123 = vld1q_f32(w + 80);
931*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c0x4567 = vld1q_f32(w + 84);
932*4bdc9457SAndroid Build Coastguard Worker 
933*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0);
934*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);
935*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_low_f32(vi0x1), 0);
936*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c0x4567, vget_low_f32(vi2x1), 0);
937*4bdc9457SAndroid Build Coastguard Worker 
938*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c0x0123 = vld1q_f32(w + 88);
939*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c0x4567 = vld1q_f32(w + 92);
940*4bdc9457SAndroid Build Coastguard Worker 
941*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0);
942*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);
943*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_low_f32(vi1x1), 0);
944*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c0x4567, vget_low_f32(vi3x1), 0);
945*4bdc9457SAndroid Build Coastguard Worker 
946*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c0x0123 = vld1q_f32(w + 96);
947*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c0x4567 = vld1q_f32(w + 100);
948*4bdc9457SAndroid Build Coastguard Worker 
949*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0);
950*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c0x0123, vget_low_f32(vi4x1), 0);
951*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0);
952*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c0x4567, vget_low_f32(vi4x1), 0);
953*4bdc9457SAndroid Build Coastguard Worker 
954*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c1x0123 = vld1q_f32(w + 104);
955*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c1x4567 = vld1q_f32(w + 108);
956*4bdc9457SAndroid Build Coastguard Worker 
957*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c1x0123, vget_low_f32(vi0x1), 1);
958*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);
959*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 1);
960*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 1);
961*4bdc9457SAndroid Build Coastguard Worker 
962*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c1x0123 = vld1q_f32(w + 112);
963*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c1x4567 = vld1q_f32(w + 116);
964*4bdc9457SAndroid Build Coastguard Worker 
965*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1);
966*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1);
967*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c1x4567, vget_low_f32(vi1x1), 1);
968*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c1x4567, vget_low_f32(vi3x1), 1);
969*4bdc9457SAndroid Build Coastguard Worker 
970*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c1x0123 = vld1q_f32(w + 120);
971*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c1x4567 = vld1q_f32(w + 124);
972*4bdc9457SAndroid Build Coastguard Worker 
973*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1);
974*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1);
975*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1);
976*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c1x4567, vget_low_f32(vi4x1), 1);
977*4bdc9457SAndroid Build Coastguard Worker 
978*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c2x0123 = vld1q_f32(w + 128);
979*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk01c2x4567 = vld1q_f32(w + 132);
980*4bdc9457SAndroid Build Coastguard Worker 
981*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0);
982*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);
983*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c2x4567, vget_high_f32(vi0x1), 0);
984*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c2x4567, vget_high_f32(vi2x1), 0);
985*4bdc9457SAndroid Build Coastguard Worker 
986*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c2x0123 = vld1q_f32(w + 136);
987*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk11c2x4567 = vld1q_f32(w + 140);
988*4bdc9457SAndroid Build Coastguard Worker 
989*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0);
990*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0);
991*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vget_high_f32(vi1x1), 0);
992*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c2x4567, vget_high_f32(vi3x1), 0);
993*4bdc9457SAndroid Build Coastguard Worker 
994*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c2x0123 = vld1q_f32(w + 144);
995*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vk21c2x4567 = vld1q_f32(w + 148);
996*4bdc9457SAndroid Build Coastguard Worker 
997*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0);
998*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c2x0123, vget_high_f32(vi4x1), 0);
999*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0);
1000*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c2x4567, vget_high_f32(vi4x1), 0);
1001*4bdc9457SAndroid Build Coastguard Worker 
1002*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
1003*4bdc9457SAndroid Build Coastguard Worker         const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
1004*4bdc9457SAndroid Build Coastguard Worker 
1005*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vmaxq_f32(vo0c0123, vmin);
1006*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vmaxq_f32(vo1c0123, vmin);
1007*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vmaxq_f32(vo0c4567, vmin);
1008*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vmaxq_f32(vo1c4567, vmin);
1009*4bdc9457SAndroid Build Coastguard Worker 
1010*4bdc9457SAndroid Build Coastguard Worker         vo0c0123 = vminq_f32(vo0c0123, vmax);
1011*4bdc9457SAndroid Build Coastguard Worker         vo1c0123 = vminq_f32(vo1c0123, vmax);
1012*4bdc9457SAndroid Build Coastguard Worker         vo0c4567 = vminq_f32(vo0c4567, vmax);
1013*4bdc9457SAndroid Build Coastguard Worker         vo1c4567 = vminq_f32(vo1c4567, vmax);
1014*4bdc9457SAndroid Build Coastguard Worker 
1015*4bdc9457SAndroid Build Coastguard Worker         if XNN_LIKELY(c >= 8) {
1016*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o1, vo1c0123);
1017*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o1 + 4, vo1c4567);
1018*4bdc9457SAndroid Build Coastguard Worker           o1 = (float*) ((uintptr_t) o1 + output_width_stride);
1019*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o0, vo0c0123);
1020*4bdc9457SAndroid Build Coastguard Worker           vst1q_f32(o0 + 4, vo0c4567);
1021*4bdc9457SAndroid Build Coastguard Worker           o0 = (float*) ((uintptr_t) o0 + output_width_stride);
1022*4bdc9457SAndroid Build Coastguard Worker         } else {
1023*4bdc9457SAndroid Build Coastguard Worker           float* o0_tmp = o0;
1024*4bdc9457SAndroid Build Coastguard Worker           float* o1_tmp = o1;
1025*4bdc9457SAndroid Build Coastguard Worker           if (c & 4) {
1026*4bdc9457SAndroid Build Coastguard Worker             vst1q_f32(o1_tmp, vo1c0123); o1_tmp += 4;
1027*4bdc9457SAndroid Build Coastguard Worker             vo1c0123 = vo1c4567;
1028*4bdc9457SAndroid Build Coastguard Worker             vst1q_f32(o0_tmp, vo0c0123); o0_tmp += 4;
1029*4bdc9457SAndroid Build Coastguard Worker             vo0c0123 = vo0c4567;
1030*4bdc9457SAndroid Build Coastguard Worker           }
1031*4bdc9457SAndroid Build Coastguard Worker           float32x2_t vo0c01 = vget_low_f32(vo0c0123);
1032*4bdc9457SAndroid Build Coastguard Worker           float32x2_t vo1c01 = vget_low_f32(vo1c0123);
1033*4bdc9457SAndroid Build Coastguard Worker           if (c & 2) {
1034*4bdc9457SAndroid Build Coastguard Worker             vst1_f32(o1_tmp, vo1c01); o1_tmp += 2;
1035*4bdc9457SAndroid Build Coastguard Worker             vo1c01 = vget_high_f32(vo1c0123);
1036*4bdc9457SAndroid Build Coastguard Worker             vst1_f32(o0_tmp, vo0c01); o0_tmp += 2;
1037*4bdc9457SAndroid Build Coastguard Worker             vo0c01 = vget_high_f32(vo0c0123);
1038*4bdc9457SAndroid Build Coastguard Worker           }
1039*4bdc9457SAndroid Build Coastguard Worker           if (c & 1) {
1040*4bdc9457SAndroid Build Coastguard Worker             vst1_lane_f32(o1_tmp, vo1c01, 0);
1041*4bdc9457SAndroid Build Coastguard Worker             vst1_lane_f32(o0_tmp, vo0c01, 0);
1042*4bdc9457SAndroid Build Coastguard Worker           }
1043*4bdc9457SAndroid Build Coastguard Worker           o0 = (float*) ((uintptr_t) o0 + output_width_stride);
1044*4bdc9457SAndroid Build Coastguard Worker           o1 = (float*) ((uintptr_t) o1 + output_width_stride);
1045*4bdc9457SAndroid Build Coastguard Worker         }
1046*4bdc9457SAndroid Build Coastguard Worker       }
1047*4bdc9457SAndroid Build Coastguard Worker       // Move output pointers back to the position of the first pixel in a row,
1048*4bdc9457SAndroid Build Coastguard Worker       // and forward to the next block of output channels
1049*4bdc9457SAndroid Build Coastguard Worker       o0 = (float*) ((uintptr_t) o0 - output_channel_decrement);
1050*4bdc9457SAndroid Build Coastguard Worker       o1 = (float*) ((uintptr_t) o1 - output_channel_decrement);
1051*4bdc9457SAndroid Build Coastguard Worker       // Revert input pointers to the position of the first pixel in a row
1052*4bdc9457SAndroid Build Coastguard Worker       i0 = (const float*) ((uintptr_t) i0 - input_width_decrement);
1053*4bdc9457SAndroid Build Coastguard Worker       i1 = (const float*) ((uintptr_t) i1 - input_width_decrement);
1054*4bdc9457SAndroid Build Coastguard Worker       i2 = (const float*) ((uintptr_t) i2 - input_width_decrement);
1055*4bdc9457SAndroid Build Coastguard Worker       i3 = (const float*) ((uintptr_t) i3 - input_width_decrement);
1056*4bdc9457SAndroid Build Coastguard Worker       i4 = (const float*) ((uintptr_t) i4 - input_width_decrement);
1057*4bdc9457SAndroid Build Coastguard Worker       // Move to the block of weights for the next 8 output channels
1058*4bdc9457SAndroid Build Coastguard Worker       w += 224;
1059*4bdc9457SAndroid Build Coastguard Worker       c = doz(c, 8);
1060*4bdc9457SAndroid Build Coastguard Worker     } while (c != 0);
1061*4bdc9457SAndroid Build Coastguard Worker     // Move output pointers back to the position of the first channel, and forward to the next block of rows
1062*4bdc9457SAndroid Build Coastguard Worker     o0 = (float*) ((uintptr_t) o0 + output_height_increment);
1063*4bdc9457SAndroid Build Coastguard Worker     o1 = (float*) ((uintptr_t) o1 + output_height_increment);
1064*4bdc9457SAndroid Build Coastguard Worker     // Move input pointers forward to the next four rows
1065*4bdc9457SAndroid Build Coastguard Worker     i0 = i4;
1066*4bdc9457SAndroid Build Coastguard Worker     i1 = (const float*) ((uintptr_t) i0 + input_height_stride);
1067*4bdc9457SAndroid Build Coastguard Worker     i2 = (const float*) ((uintptr_t) i1 + input_height_stride);
1068*4bdc9457SAndroid Build Coastguard Worker     i3 = (const float*) ((uintptr_t) i2 + input_height_stride);
1069*4bdc9457SAndroid Build Coastguard Worker     i4 = (const float*) ((uintptr_t) i3 + input_height_stride);
1070*4bdc9457SAndroid Build Coastguard Worker   }
1071*4bdc9457SAndroid Build Coastguard Worker }
1072