blob: 4850206b28f28f1d2ff9b84713e5498546ebf530 [file] [log] [blame]
Marat Dukhance7a3f82020-05-17 21:46:44 -07001// Auto-generated file. Do not edit!
2// Template: src/f32-conv-hwc/3x3s2p0p1c3-neon-x2.c.in
3// Generator: tools/xngen
4//
5// Copyright 2020 Google LLC
6//
7// This source code is licensed under the BSD-style license found in the
8// LICENSE file in the root directory of this source tree.
9
10
11#include <assert.h>
12
13#include <arm_neon.h>
14
15#include <xnnpack/conv.h>
16#include <xnnpack/math.h>
17
18
19void xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2(
20 size_t input_height,
21 size_t input_width,
22 size_t output_y_start,
23 size_t output_y_end,
24 const float* input,
25 const float* zero,
26 const float* weights,
27 float* output,
28 size_t input_padding_top,
29 size_t output_channels,
30 size_t output_height_stride,
31 size_t output_width_stride,
32 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
33{
34 assert(input_width != 0);
35 assert(output_y_end > output_y_start);
36 assert(input_padding_top <= 1);
37 assert(output_channels != 0);
38
39 const size_t input_height_stride = input_width * 3 /* channels */ * sizeof(float);
40 const size_t input_width_decrement = (4 + ((input_width - 1) & 1) * 2 + (round_down_po2(input_width - 1, 2) * 3 /* channels */)) * sizeof(float);
41 const size_t output_width = input_width / 2;
42 const size_t output_channel_decrement = output_width * output_width_stride - 8 * sizeof(float);
43 const size_t output_height_increment = output_height_stride * 2 - round_up_po2(output_channels, 8) * sizeof(float);
44
45 // Adjustment for padding processed below
46 const float* i0 = (const float*) ((uintptr_t) input +
47 input_height_stride * (output_y_start * 2 /* vertical stride */ - input_padding_top));
48 const float* i1 = (const float*) ((uintptr_t) i0 + input_height_stride);
49 const float* i2 = (const float*) ((uintptr_t) i1 + input_height_stride);
50 const float* i3 = (const float*) ((uintptr_t) i2 + input_height_stride);
51 const float* i4 = (const float*) ((uintptr_t) i3 + input_height_stride);
52 float* o0 = (float*) ((uintptr_t) output + output_height_stride * output_y_start);
53 float* o1 = (float*) ((uintptr_t) o0 + output_height_stride);
54
55 if XNN_UNPREDICTABLE(output_y_start < input_padding_top) {
56 i0 = zero;
57 }
58
Marat Dukhance7a3f82020-05-17 21:46:44 -070059
60 for (size_t output_y = output_y_start; output_y < output_y_end; output_y += 2) {
61 const size_t input_y2 = output_y * 2 + 2 - input_padding_top;
62 const size_t input_y4 = input_y2 + 2;
63 if XNN_UNPREDICTABLE(input_y2 > input_height) {
64 i1 = zero;
65 }
66 if XNN_UNPREDICTABLE(input_y2 >= input_height) {
67 i2 = zero;
68 }
69 if XNN_UNPREDICTABLE(input_y4 > input_height) {
70 i3 = zero;
71 }
72 if XNN_UNPREDICTABLE(input_y4 >= input_height) {
73 i4 = zero;
74 }
75 if XNN_UNPREDICTABLE(output_y + 2 > output_y_end) {
76 o1 = o0;
77 }
78
79 const float* w = weights;
80 size_t c = output_channels;
81 do {
82 // viMx0 = ( iM1c0, iM0c2, iM0c1, iM0c0 )
83 float32x4_t vi0x0 = vld1q_f32(i0); i0 += 4;
84 float32x4_t vi1x0 = vld1q_f32(i1); i1 += 4;
85 float32x4_t vi2x0 = vld1q_f32(i2); i2 += 4;
86 float32x4_t vi3x0 = vld1q_f32(i3); i3 += 4;
87 float32x4_t vi4x0 = vld1q_f32(i4); i4 += 4;
88
89 size_t iw = input_width - 1;
90 for (; iw >= 4; iw -= 4) {
91 float32x4_t vo0x0c0123 = vld1q_f32(w);
92 float32x4_t vo0x0c4567 = vld1q_f32(w + 4);
93 float32x4_t vo1x0c0123 = vo0x0c0123;
94 float32x4_t vo1x0c4567 = vo0x0c4567;
95 float32x4_t vo0x1c0123 = vo0x0c0123;
96 float32x4_t vo0x1c4567 = vo0x0c4567;
97 float32x4_t vo1x1c0123 = vo0x0c0123;
98 float32x4_t vo1x1c4567 = vo0x0c4567;
99
100 const float32x4_t vk00c0x0123 = vld1q_f32(w + 8);
101 const float32x4_t vk00c0x4567 = vld1q_f32(w + 12);
102
103 // viMx1 = ( iM2c1, iM2c0, iM1c2, iM1c1 )
104 const float32x4_t vi0x1 = vld1q_f32(i0); i0 += 4;
105 const float32x4_t vi1x1 = vld1q_f32(i1); i1 += 4;
106 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4;
107 const float32x4_t vi3x1 = vld1q_f32(i3); i3 += 4;
108 const float32x4_t vi4x1 = vld1q_f32(i4); i4 += 4;
109
110 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c0x0123, vget_low_f32(vi0x0), 0);
111 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 0);
112 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk00c0x4567, vget_low_f32(vi0x0), 0);
113 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk00c0x4567, vget_low_f32(vi2x0), 0);
114
115 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk00c0x0123, vget_high_f32(vi0x1), 0);
116 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 0);
117 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk00c0x4567, vget_high_f32(vi0x1), 0);
118 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk00c0x4567, vget_high_f32(vi2x1), 0);
119
120 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16);
121 const float32x4_t vk10c0x4567 = vld1q_f32(w + 20);
122
123 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0);
124 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 0);
125 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0);
126 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk10c0x4567, vget_low_f32(vi3x0), 0);
127
128 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c0x0123, vget_high_f32(vi1x1), 0);
129 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 0);
130 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk10c0x4567, vget_high_f32(vi1x1), 0);
131 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk10c0x4567, vget_high_f32(vi3x1), 0);
132
133 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24);
134 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28);
135
136 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0);
137 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 0);
138 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0);
139 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk20c0x4567, vget_low_f32(vi4x0), 0);
140
141 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 0);
142 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 0);
143 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c0x4567, vget_high_f32(vi2x1), 0);
144 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk20c0x4567, vget_high_f32(vi4x1), 0);
145
146 const float32x4_t vk00c1x0123 = vld1q_f32(w + 32);
147 const float32x4_t vk00c1x4567 = vld1q_f32(w + 36);
148
149 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_low_f32(vi0x0), 1);
150 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_low_f32(vi2x0), 1);
151 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk00c1x4567, vget_low_f32(vi0x0), 1);
152 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk00c1x4567, vget_low_f32(vi2x0), 1);
153
154 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk00c1x0123, vget_high_f32(vi0x1), 1);
155 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_high_f32(vi2x1), 1);
156 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk00c1x4567, vget_high_f32(vi0x1), 1);
157 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk00c1x4567, vget_high_f32(vi2x1), 1);
158
159 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40);
160 const float32x4_t vk10c1x4567 = vld1q_f32(w + 44);
161
162 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1);
163 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_low_f32(vi3x0), 1);
164 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1);
165 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk10c1x4567, vget_low_f32(vi3x0), 1);
166
167 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_high_f32(vi1x1), 1);
168 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_high_f32(vi3x1), 1);
169 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk10c1x4567, vget_high_f32(vi1x1), 1);
170 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk10c1x4567, vget_high_f32(vi3x1), 1);
171
172 const float32x4_t vk20c1x0123 = vld1q_f32(w + 48);
173 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52);
174
175 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1);
176 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c1x0123, vget_low_f32(vi4x0), 1);
177 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1);
178 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk20c1x4567, vget_low_f32(vi4x0), 1);
179
180 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c1x0123, vget_high_f32(vi2x1), 1);
181 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c1x0123, vget_high_f32(vi4x1), 1);
182 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c1x4567, vget_high_f32(vi2x1), 1);
183 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk20c1x4567, vget_high_f32(vi4x1), 1);
184
185 const float32x4_t vk00c2x0123 = vld1q_f32(w + 56);
186 const float32x4_t vk00c2x4567 = vld1q_f32(w + 60);
187
188 // viMx2 = ( iM3c2, iM3c1, iM3c0, iM2c2 )
189 const float32x4_t vi0x2 = vld1q_f32(i0); i0 += 4;
190 const float32x4_t vi1x2 = vld1q_f32(i1); i1 += 4;
191 const float32x4_t vi2x2 = vld1q_f32(i2); i2 += 4;
192 const float32x4_t vi3x2 = vld1q_f32(i3); i3 += 4;
193 const float32x4_t vi4x2 = vld1q_f32(i4); i4 += 4;
194
195 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c2x0123, vget_high_f32(vi0x0), 0);
196 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 0);
197 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk00c2x4567, vget_high_f32(vi0x0), 0);
198 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk00c2x4567, vget_high_f32(vi2x0), 0);
199
200 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk00c2x0123, vget_low_f32(vi0x2), 0);
201 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c2x0123, vget_low_f32(vi2x2), 0);
202 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk00c2x4567, vget_low_f32(vi0x2), 0);
203 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk00c2x4567, vget_low_f32(vi2x2), 0);
204
205 const float32x4_t vk10c2x0123 = vld1q_f32(w + 64);
206 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68);
207
208 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c2x0123, vget_high_f32(vi1x0), 0);
209 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c2x0123, vget_high_f32(vi3x0), 0);
210 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0);
211 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk10c2x4567, vget_high_f32(vi3x0), 0);
212
213 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c2x0123, vget_low_f32(vi1x2), 0);
214 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c2x0123, vget_low_f32(vi3x2), 0);
215 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk10c2x4567, vget_low_f32(vi1x2), 0);
216 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk10c2x4567, vget_low_f32(vi3x2), 0);
217
218 const float32x4_t vk20c2x0123 = vld1q_f32(w + 72);
219 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76);
220
221 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c2x0123, vget_high_f32(vi2x0), 0);
222 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c2x0123, vget_high_f32(vi4x0), 0);
223 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0);
224 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk20c2x4567, vget_high_f32(vi4x0), 0);
225
226 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c2x0123, vget_low_f32(vi2x2), 0);
227 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c2x0123, vget_low_f32(vi4x2), 0);
228 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c2x4567, vget_low_f32(vi2x2), 0);
229 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk20c2x4567, vget_low_f32(vi4x2), 0);
230
231 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80);
232 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84);
233
234 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1);
235 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);
236 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1);
237 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk01c0x4567, vget_high_f32(vi2x0), 1);
238
239 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c0x0123, vget_low_f32(vi0x2), 1);
240 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c0x0123, vget_low_f32(vi2x2), 1);
241 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk01c0x4567, vget_low_f32(vi0x2), 1);
242 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk01c0x4567, vget_low_f32(vi2x2), 1);
243
244 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88);
245 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92);
246
247 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1);
248 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_high_f32(vi3x0), 1);
249 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1);
250 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk11c0x4567, vget_high_f32(vi3x0), 1);
251
252 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c0x0123, vget_low_f32(vi1x2), 1);
253 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c0x0123, vget_low_f32(vi3x2), 1);
254 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk11c0x4567, vget_low_f32(vi1x2), 1);
255 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk11c0x4567, vget_low_f32(vi3x2), 1);
256
257 const float32x4_t vk21c0x0123 = vld1q_f32(w + 96);
258 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100);
259
260 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c0x0123, vget_high_f32(vi2x0), 1);
261 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c0x0123, vget_high_f32(vi4x0), 1);
262 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1);
263 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk21c0x4567, vget_high_f32(vi4x0), 1);
264
265 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk21c0x0123, vget_low_f32(vi2x2), 1);
266 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk21c0x0123, vget_low_f32(vi4x2), 1);
267 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk21c0x4567, vget_low_f32(vi2x2), 1);
268 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk21c0x4567, vget_low_f32(vi4x2), 1);
269
270 const float32x4_t vk01c1x0123 = vld1q_f32(w + 104);
271 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108);
272
273 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c1x0123, vget_low_f32(vi0x1), 0);
274 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 0);
275 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk01c1x4567, vget_low_f32(vi0x1), 0);
276 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk01c1x4567, vget_low_f32(vi2x1), 0);
277
278 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c1x0123, vget_high_f32(vi0x2), 0);
279 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c1x0123, vget_high_f32(vi2x2), 0);
280 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk01c1x4567, vget_high_f32(vi0x2), 0);
281 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk01c1x4567, vget_high_f32(vi2x2), 0);
282
283 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112);
284 const float32x4_t vk11c1x4567 = vld1q_f32(w + 116);
285
286 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0);
287 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c1x0123, vget_low_f32(vi3x1), 0);
288 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk11c1x4567, vget_low_f32(vi1x1), 0);
289 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk11c1x4567, vget_low_f32(vi3x1), 0);
290
291 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c1x0123, vget_high_f32(vi1x2), 0);
292 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c1x0123, vget_high_f32(vi3x2), 0);
293 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk11c1x4567, vget_high_f32(vi1x2), 0);
294 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk11c1x4567, vget_high_f32(vi3x2), 0);
295
296 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120);
297 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124);
298
299 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0);
300 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 0);
301 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0);
302 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk21c1x4567, vget_low_f32(vi4x1), 0);
303
304 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 0);
305 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk21c1x0123, vget_high_f32(vi4x2), 0);
306 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk21c1x4567, vget_high_f32(vi2x2), 0);
307 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk21c1x4567, vget_high_f32(vi4x2), 0);
308
309 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128);
310 const float32x4_t vk01c2x4567 = vld1q_f32(w + 132);
311
312 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1);
313 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_low_f32(vi2x1), 1);
314 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk01c2x4567, vget_low_f32(vi0x1), 1);
315 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk01c2x4567, vget_low_f32(vi2x1), 1);
316
317 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c2x0123, vget_high_f32(vi0x2), 1);
318 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c2x0123, vget_high_f32(vi2x2), 1);
319 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk01c2x4567, vget_high_f32(vi0x2), 1);
320 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk01c2x4567, vget_high_f32(vi2x2), 1);
321
322 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136);
323 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140);
324
325 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1);
326 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c2x0123, vget_low_f32(vi3x1), 1);
327 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk11c2x4567, vget_low_f32(vi1x1), 1);
328 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk11c2x4567, vget_low_f32(vi3x1), 1);
329
330 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c2x0123, vget_high_f32(vi1x2), 1);
331 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c2x0123, vget_high_f32(vi3x2), 1);
332 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk11c2x4567, vget_high_f32(vi1x2), 1);
333 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk11c2x4567, vget_high_f32(vi3x2), 1);
334
335 const float32x4_t vk21c2x0123 = vld1q_f32(w + 144);
336 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148);
337
338 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c2x0123, vget_low_f32(vi2x1), 1);
339 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c2x0123, vget_low_f32(vi4x1), 1);
340 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1);
341 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk21c2x4567, vget_low_f32(vi4x1), 1);
342
343 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk21c2x0123, vget_high_f32(vi2x2), 1);
344 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk21c2x0123, vget_high_f32(vi4x2), 1);
345 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk21c2x4567, vget_high_f32(vi2x2), 1);
346 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk21c2x4567, vget_high_f32(vi4x2), 1);
347
348 const float32x4_t vk02c0x0123 = vld1q_f32(w + 152);
349 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156);
350
351 // viMx3 = ( iM5c0, iM4c2, iM4c1, iM4c0 )
352 const float32x4_t vi0x3 = vld1q_f32(i0); i0 += 4;
353 const float32x4_t vi1x3 = vld1q_f32(i1); i1 += 4;
354 const float32x4_t vi2x3 = vld1q_f32(i2); i2 += 4;
355 const float32x4_t vi3x3 = vld1q_f32(i3); i3 += 4;
356 const float32x4_t vi4x3 = vld1q_f32(i4); i4 += 4;
357
358 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c0x0123, vget_high_f32(vi0x1), 0);
359 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c0x0123, vget_high_f32(vi2x1), 0);
360 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk02c0x4567, vget_high_f32(vi0x1), 0);
361 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk02c0x4567, vget_high_f32(vi2x1), 0);
362
363 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk02c0x0123, vget_low_f32(vi0x3), 0);
364 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk02c0x0123, vget_low_f32(vi2x3), 0);
365 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk02c0x4567, vget_low_f32(vi0x3), 0);
366 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk02c0x4567, vget_low_f32(vi2x3), 0);
367
368 const float32x4_t vk12c0x0123 = vld1q_f32(w + 160);
369 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164);
370
371 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c0x0123, vget_high_f32(vi1x1), 0);
372 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c0x0123, vget_high_f32(vi3x1), 0);
373 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk12c0x4567, vget_high_f32(vi1x1), 0);
374 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk12c0x4567, vget_high_f32(vi3x1), 0);
375
376 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c0x0123, vget_low_f32(vi1x3), 0);
377 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c0x0123, vget_low_f32(vi3x3), 0);
378 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk12c0x4567, vget_low_f32(vi1x3), 0);
379 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk12c0x4567, vget_low_f32(vi3x3), 0);
380
381 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168);
382 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172);
383
384 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0);
385 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c0x0123, vget_high_f32(vi4x1), 0);
386 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk22c0x4567, vget_high_f32(vi2x1), 0);
387 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk22c0x4567, vget_high_f32(vi4x1), 0);
388
389 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c0x0123, vget_low_f32(vi2x3), 0);
390 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c0x0123, vget_low_f32(vi4x3), 0);
391 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk22c0x4567, vget_low_f32(vi2x3), 0);
392 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk22c0x4567, vget_low_f32(vi4x3), 0);
393
394 const float32x4_t vk02c1x0123 = vld1q_f32(w + 176);
395 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180);
396
397 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c1x0123, vget_high_f32(vi0x1), 1);
398 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c1x0123, vget_high_f32(vi2x1), 1);
399 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk02c1x4567, vget_high_f32(vi0x1), 1);
400 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk02c1x4567, vget_high_f32(vi2x1), 1);
401
402 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk02c1x0123, vget_low_f32(vi0x3), 1);
403 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk02c1x0123, vget_low_f32(vi2x3), 1);
404 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk02c1x4567, vget_low_f32(vi0x3), 1);
405 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk02c1x4567, vget_low_f32(vi2x3), 1);
406
407 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184);
408 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188);
409
410 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1);
411 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c1x0123, vget_high_f32(vi3x1), 1);
412 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk12c1x4567, vget_high_f32(vi1x1), 1);
413 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk12c1x4567, vget_high_f32(vi3x1), 1);
414
415 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c1x0123, vget_low_f32(vi1x3), 1);
416 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c1x0123, vget_low_f32(vi3x3), 1);
417 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk12c1x4567, vget_low_f32(vi1x3), 1);
418 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk12c1x4567, vget_low_f32(vi3x3), 1);
419
420 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192);
421 const float32x4_t vk22c1x4567 = vld1q_f32(w + 196);
422
423 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1);
424 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c1x0123, vget_high_f32(vi4x1), 1);
425 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk22c1x4567, vget_high_f32(vi2x1), 1);
426 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk22c1x4567, vget_high_f32(vi4x1), 1);
427
428 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c1x0123, vget_low_f32(vi2x3), 1);
429 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c1x0123, vget_low_f32(vi4x3), 1);
430 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk22c1x4567, vget_low_f32(vi2x3), 1);
431 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk22c1x4567, vget_low_f32(vi4x3), 1);
432
433 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200);
434 const float32x4_t vk02c2x4567 = vld1q_f32(w + 204);
435
436 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c2x0123, vget_low_f32(vi0x2), 0);
437 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 0);
438 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk02c2x4567, vget_low_f32(vi0x2), 0);
439 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk02c2x4567, vget_low_f32(vi2x2), 0);
440
441 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk02c2x0123, vget_high_f32(vi0x3), 0);
442 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk02c2x0123, vget_high_f32(vi2x3), 0);
443 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk02c2x4567, vget_high_f32(vi0x3), 0);
444 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk02c2x4567, vget_high_f32(vi2x3), 0);
445
446 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208);
447 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212);
448
449 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c2x0123, vget_low_f32(vi1x2), 0);
450 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c2x0123, vget_low_f32(vi3x2), 0);
451 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk12c2x4567, vget_low_f32(vi1x2), 0);
452 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk12c2x4567, vget_low_f32(vi3x2), 0);
453
454 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c2x0123, vget_high_f32(vi1x3), 0);
455 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c2x0123, vget_high_f32(vi3x3), 0);
456 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk12c2x4567, vget_high_f32(vi1x3), 0);
457 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk12c2x4567, vget_high_f32(vi3x3), 0);
458
459 const float32x4_t vk22c2x0123 = vld1q_f32(w + 216);
460 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220);
461
462 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c2x0123, vget_low_f32(vi2x2), 0);
463 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c2x0123, vget_low_f32(vi4x2), 0);
464 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk22c2x4567, vget_low_f32(vi2x2), 0);
465 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk22c2x4567, vget_low_f32(vi4x2), 0);
466
467 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c2x0123, vget_high_f32(vi2x3), 0);
468 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c2x0123, vget_high_f32(vi4x3), 0);
469 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk22c2x4567, vget_high_f32(vi2x3), 0);
470 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk22c2x4567, vget_high_f32(vi4x3), 0);
471
472 vi0x0 = vi0x3;
473 vi1x0 = vi1x3;
474 vi2x0 = vi2x3;
475 vi3x0 = vi3x3;
476 vi4x0 = vi4x3;
477
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700478 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
479 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
480
Marat Dukhance7a3f82020-05-17 21:46:44 -0700481 vo0x0c0123 = vmaxq_f32(vo0x0c0123, vmin);
482 vo1x0c0123 = vmaxq_f32(vo1x0c0123, vmin);
483 vo0x0c4567 = vmaxq_f32(vo0x0c4567, vmin);
484 vo1x0c4567 = vmaxq_f32(vo1x0c4567, vmin);
485
486 vo0x1c0123 = vmaxq_f32(vo0x1c0123, vmin);
487 vo1x1c0123 = vmaxq_f32(vo1x1c0123, vmin);
488 vo0x1c4567 = vmaxq_f32(vo0x1c4567, vmin);
489 vo1x1c4567 = vmaxq_f32(vo1x1c4567, vmin);
490
491 vo0x0c0123 = vminq_f32(vo0x0c0123, vmax);
492 vo1x0c0123 = vminq_f32(vo1x0c0123, vmax);
493 vo0x0c4567 = vminq_f32(vo0x0c4567, vmax);
494 vo1x0c4567 = vminq_f32(vo1x0c4567, vmax);
495
496 vo0x1c0123 = vminq_f32(vo0x1c0123, vmax);
497 vo1x1c0123 = vminq_f32(vo1x1c0123, vmax);
498 vo0x1c4567 = vminq_f32(vo0x1c4567, vmax);
499 vo1x1c4567 = vminq_f32(vo1x1c4567, vmax);
500
501 if XNN_LIKELY(c >= 8) {
502 vst1q_f32(o1, vo1x0c0123);
503 vst1q_f32(o1 + 4, vo1x0c4567);
504 o1 = (float*) ((uintptr_t) o1 + output_width_stride);
505 vst1q_f32(o0, vo0x0c0123);
506 vst1q_f32(o0 + 4, vo0x0c4567);
507 o0 = (float*) ((uintptr_t) o0 + output_width_stride);
508
509 vst1q_f32(o1, vo1x1c0123);
510 vst1q_f32(o1 + 4, vo1x1c4567);
511 o1 = (float*) ((uintptr_t) o1 + output_width_stride);
512 vst1q_f32(o0, vo0x1c0123);
513 vst1q_f32(o0 + 4, vo0x1c4567);
514 o0 = (float*) ((uintptr_t) o0 + output_width_stride);
515 } else {
516 float* o0_tmp = o0;
517 float* o1_tmp = o1;
518 if (c & 4) {
519 vst1q_f32((float*) ((uintptr_t) o1_tmp + output_width_stride), vo1x1c0123);
520 vo1x1c0123 = vo1x1c4567;
521 vst1q_f32((float*) ((uintptr_t) o0_tmp + output_width_stride), vo0x1c0123);
522 vo0x1c0123 = vo0x1c4567;
523
524 vst1q_f32(o1_tmp, vo1x0c0123); o1_tmp += 4;
525 vo1x0c0123 = vo1x0c4567;
526 vst1q_f32(o0_tmp, vo0x0c0123); o0_tmp += 4;
527 vo0x0c0123 = vo0x0c4567;
528 }
529 float32x2_t vo0x0c01 = vget_low_f32(vo0x0c0123);
530 float32x2_t vo1x0c01 = vget_low_f32(vo1x0c0123);
531 float32x2_t vo0x1c01 = vget_low_f32(vo0x1c0123);
532 float32x2_t vo1x1c01 = vget_low_f32(vo1x1c0123);
533 if (c & 2) {
534 vst1_f32((float*) ((uintptr_t) o1_tmp + output_width_stride), vo1x1c01);
535 vo1x1c01 = vget_high_f32(vo1x1c0123);
536 vst1_f32((float*) ((uintptr_t) o0_tmp + output_width_stride), vo0x1c01);
537 vo0x1c01 = vget_high_f32(vo0x1c0123);
538
539 vst1_f32(o1_tmp, vo1x0c01); o1_tmp += 2;
540 vo1x0c01 = vget_high_f32(vo1x0c0123);
541 vst1_f32(o0_tmp, vo0x0c01); o0_tmp += 2;
542 vo0x0c01 = vget_high_f32(vo0x0c0123);
543 }
544 if (c & 1) {
545 vst1_lane_f32(o1_tmp, vo1x0c01, 0);
546 vst1_lane_f32(o0_tmp, vo0x0c01, 0);
547
548 vst1_lane_f32((float*) ((uintptr_t) o1_tmp + output_width_stride), vo1x1c01, 0);
549 vst1_lane_f32((float*) ((uintptr_t) o0_tmp + output_width_stride), vo0x1c01, 0);
550 }
551
552 o0 = (float*) ((uintptr_t) o0 + output_width_stride * 2);
553 o1 = (float*) ((uintptr_t) o1 + output_width_stride * 2);
554 }
555 }
556 assert(iw < 4);
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700557 if XNN_LIKELY(iw & 2) {
Marat Dukhance7a3f82020-05-17 21:46:44 -0700558 float32x4_t vo0c0123 = vld1q_f32(w);
559 float32x4_t vo0c4567 = vld1q_f32(w + 4);
560 float32x4_t vo1c0123 = vo0c0123;
561 float32x4_t vo1c4567 = vo0c4567;
562
563 const float32x4_t vk00c0x0123 = vld1q_f32(w + 8);
564 const float32x4_t vk00c0x4567 = vld1q_f32(w + 12);
565
566 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk00c0x0123, vget_low_f32(vi0x0), 0);
567 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 0);
568 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 0);
569 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c0x4567, vget_low_f32(vi2x0), 0);
570
571 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16);
572 const float32x4_t vk10c0x4567 = vld1q_f32(w + 20);
573
574 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0);
575 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0);
576 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0);
577 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c0x4567, vget_low_f32(vi3x0), 0);
578
579 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24);
580 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28);
581
582 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0);
583 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0);
584 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0);
585 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 0);
586
587 const float32x4_t vk00c1x0123 = vld1q_f32(w + 32);
588 const float32x4_t vk00c1x4567 = vld1q_f32(w + 36);
589
590 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk00c1x0123, vget_low_f32(vi0x0), 1);
591 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c1x0123, vget_low_f32(vi2x0), 1);
592 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c1x4567, vget_low_f32(vi0x0), 1);
593 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c1x4567, vget_low_f32(vi2x0), 1);
594
595 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40);
596 const float32x4_t vk10c1x4567 = vld1q_f32(w + 44);
597
598 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1);
599 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1);
600 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1);
601 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c1x4567, vget_low_f32(vi3x0), 1);
602
603 const float32x4_t vk20c1x0123 = vld1q_f32(w + 48);
604 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52);
605
606 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1);
607 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c1x0123, vget_low_f32(vi4x0), 1);
608 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1);
609 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c1x4567, vget_low_f32(vi4x0), 1);
610
611 const float32x4_t vk00c2x0123 = vld1q_f32(w + 56);
612 const float32x4_t vk00c2x4567 = vld1q_f32(w + 60);
613
614 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk00c2x0123, vget_high_f32(vi0x0), 0);
615 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 0);
616 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 0);
617 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c2x4567, vget_high_f32(vi2x0), 0);
618
619 const float32x4_t vk10c2x0123 = vld1q_f32(w + 64);
620 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68);
621
622 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c2x0123, vget_high_f32(vi1x0), 0);
623 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 0);
624 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0);
625 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 0);
626
627 const float32x4_t vk20c2x0123 = vld1q_f32(w + 72);
628 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76);
629
630 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c2x0123, vget_high_f32(vi2x0), 0);
631 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 0);
632 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0);
633 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 0);
634
635 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80);
636 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84);
637
638 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1);
639 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);
640 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1);
641 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c0x4567, vget_high_f32(vi2x0), 1);
642
643 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88);
644 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92);
645
646 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1);
647 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1);
648 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1);
649 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c0x4567, vget_high_f32(vi3x0), 1);
650
651 const float32x4_t vk21c0x0123 = vld1q_f32(w + 96);
652 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100);
653
654 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c0x0123, vget_high_f32(vi2x0), 1);
655 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c0x0123, vget_high_f32(vi4x0), 1);
656 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1);
657 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c0x4567, vget_high_f32(vi4x0), 1);
658
659 const float32x4_t vk01c1x0123 = vld1q_f32(w + 104);
660 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108);
661
662 // viMx1 = ( iM2c0, iM1c2, iM1c1, iM1c0 )
663 const float32x4_t vi0x1 = vld1q_f32(i0); i0 += 4;
664 const float32x4_t vi1x1 = vld1q_f32(i1); i1 += 4;
665 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4;
666 const float32x4_t vi3x1 = vld1q_f32(i3); i3 += 4;
667 const float32x4_t vi4x1 = vld1q_f32(i4); i4 += 4;
668
669 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c1x0123, vget_low_f32(vi0x1), 0);
670 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 0);
671 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 0);
672 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 0);
673
674 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112);
675 const float32x4_t vk11c1x4567 = vld1q_f32(w + 116);
676
677 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0);
678 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 0);
679 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c1x4567, vget_low_f32(vi1x1), 0);
680 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c1x4567, vget_low_f32(vi3x1), 0);
681
682 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120);
683 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124);
684
685 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0);
686 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 0);
687 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0);
688 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c1x4567, vget_low_f32(vi4x1), 0);
689
690 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128);
691 const float32x4_t vk01c2x4567 = vld1q_f32(w + 132);
692
693 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1);
694 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1);
695 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c2x4567, vget_low_f32(vi0x1), 1);
696 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c2x4567, vget_low_f32(vi2x1), 1);
697
698 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136);
699 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140);
700
701 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1);
702 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_low_f32(vi3x1), 1);
703 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vget_low_f32(vi1x1), 1);
704 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c2x4567, vget_low_f32(vi3x1), 1);
705
706 const float32x4_t vk21c2x0123 = vld1q_f32(w + 144);
707 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148);
708
709 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c2x0123, vget_low_f32(vi2x1), 1);
710 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c2x0123, vget_low_f32(vi4x1), 1);
711 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1);
712 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c2x4567, vget_low_f32(vi4x1), 1);
713
714 const float32x4_t vk02c0x0123 = vld1q_f32(w + 152);
715 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156);
716
717 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c0x0123, vget_high_f32(vi0x1), 0);
718 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 0);
719 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c0x4567, vget_high_f32(vi0x1), 0);
720 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c0x4567, vget_high_f32(vi2x1), 0);
721
722 const float32x4_t vk12c0x0123 = vld1q_f32(w + 160);
723 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164);
724
725 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c0x0123, vget_high_f32(vi1x1), 0);
726 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c0x0123, vget_high_f32(vi3x1), 0);
727 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c0x4567, vget_high_f32(vi1x1), 0);
728 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c0x4567, vget_high_f32(vi3x1), 0);
729
730 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168);
731 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172);
732
733 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0);
734 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 0);
735 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 0);
736 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c0x4567, vget_high_f32(vi4x1), 0);
737
738 const float32x4_t vk02c1x0123 = vld1q_f32(w + 176);
739 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180);
740
741 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c1x0123, vget_high_f32(vi0x1), 1);
742 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c1x0123, vget_high_f32(vi2x1), 1);
743 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c1x4567, vget_high_f32(vi0x1), 1);
744 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c1x4567, vget_high_f32(vi2x1), 1);
745
746 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184);
747 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188);
748
749 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1);
750 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c1x0123, vget_high_f32(vi3x1), 1);
751 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c1x4567, vget_high_f32(vi1x1), 1);
752 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c1x4567, vget_high_f32(vi3x1), 1);
753
754 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192);
755 const float32x4_t vk22c1x4567 = vld1q_f32(w + 196);
756
757 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1);
758 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c1x0123, vget_high_f32(vi4x1), 1);
759 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c1x4567, vget_high_f32(vi2x1), 1);
760 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c1x4567, vget_high_f32(vi4x1), 1);
761
762 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200);
763 const float32x4_t vk02c2x4567 = vld1q_f32(w + 204);
764
765 // viMx2 = ( iM2c2, iM2c1 )
766 const float32x2_t vi0x2 = vld1_f32(i0); i0 += 2;
767 const float32x2_t vi1x2 = vld1_f32(i1); i1 += 2;
768 const float32x2_t vi2x2 = vld1_f32(i2); i2 += 2;
769 const float32x2_t vi3x2 = vld1_f32(i3); i3 += 2;
770 const float32x2_t vi4x2 = vld1_f32(i4); i4 += 2;
771
772 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 0);
773 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 0);
774 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c2x4567, vi0x2, 0);
775 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c2x4567, vi2x2, 0);
776
777 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208);
778 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212);
779
780 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 0);
781 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 0);
782 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c2x4567, vi1x2, 0);
783 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c2x4567, vi3x2, 0);
784
785 const float32x4_t vk22c2x0123 = vld1q_f32(w + 216);
786 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220);
787
788 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c2x0123, vi2x2, 0);
789 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c2x0123, vi4x2, 0);
790 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 0);
791 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c2x4567, vi4x2, 0);
792
793 vi0x0 = vcombine_f32(vget_high_f32(vi0x1), vi0x2);
794 vi1x0 = vcombine_f32(vget_high_f32(vi1x1), vi1x2);
795 vi2x0 = vcombine_f32(vget_high_f32(vi2x1), vi2x2);
796 vi3x0 = vcombine_f32(vget_high_f32(vi3x1), vi3x2);
797 vi4x0 = vcombine_f32(vget_high_f32(vi4x1), vi4x2);
798
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700799 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
800 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
801
Marat Dukhance7a3f82020-05-17 21:46:44 -0700802 vo0c0123 = vmaxq_f32(vo0c0123, vmin);
803 vo1c0123 = vmaxq_f32(vo1c0123, vmin);
804 vo0c4567 = vmaxq_f32(vo0c4567, vmin);
805 vo1c4567 = vmaxq_f32(vo1c4567, vmin);
806
807 vo0c0123 = vminq_f32(vo0c0123, vmax);
808 vo1c0123 = vminq_f32(vo1c0123, vmax);
809 vo0c4567 = vminq_f32(vo0c4567, vmax);
810 vo1c4567 = vminq_f32(vo1c4567, vmax);
811
812 if XNN_LIKELY(c >= 8) {
813 vst1q_f32(o1, vo1c0123);
814 vst1q_f32(o1 + 4, vo1c4567);
815 o1 = (float*) ((uintptr_t) o1 + output_width_stride);
816 vst1q_f32(o0, vo0c0123);
817 vst1q_f32(o0 + 4, vo0c4567);
818 o0 = (float*) ((uintptr_t) o0 + output_width_stride);
819 } else {
820 float* o0_tmp = o0;
821 float* o1_tmp = o1;
822 if (c & 4) {
823 vst1q_f32(o1_tmp, vo1c0123); o1_tmp += 4;
824 vo1c0123 = vo1c4567;
825 vst1q_f32(o0_tmp, vo0c0123); o0_tmp += 4;
826 vo0c0123 = vo0c4567;
827 }
828 float32x2_t vo0c01 = vget_low_f32(vo0c0123);
829 float32x2_t vo1c01 = vget_low_f32(vo1c0123);
830 if (c & 2) {
831 vst1_f32(o1_tmp, vo1c01); o1_tmp += 2;
832 vo1c01 = vget_high_f32(vo1c0123);
833 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2;
834 vo0c01 = vget_high_f32(vo0c0123);
835 }
836 if (c & 1) {
837 vst1_lane_f32(o1_tmp, vo1c01, 0);
838 vst1_lane_f32(o0_tmp, vo0c01, 0);
839 }
840
841 o0 = (float*) ((uintptr_t) o0 + output_width_stride);
842 o1 = (float*) ((uintptr_t) o1 + output_width_stride);
843 }
844 }
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700845 if XNN_LIKELY(iw & 1) {
Marat Dukhance7a3f82020-05-17 21:46:44 -0700846 float32x4_t vo0c0123 = vld1q_f32(w);
847 float32x4_t vo0c4567 = vld1q_f32(w + 4);
848 float32x4_t vo1c0123 = vo0c0123;
849 float32x4_t vo1c4567 = vo0c4567;
850
851 const float32x4_t vk00c0x0123 = vld1q_f32(w + 8);
852 const float32x4_t vk00c0x4567 = vld1q_f32(w + 12);
853
854 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk00c0x0123, vget_low_f32(vi0x0), 0);
855 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 0);
856 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 0);
857 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c0x4567, vget_low_f32(vi2x0), 0);
858
859 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16);
860 const float32x4_t vk10c0x4567 = vld1q_f32(w + 20);
861
862 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0);
863 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0);
864 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0);
865 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c0x4567, vget_low_f32(vi3x0), 0);
866
867 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24);
868 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28);
869
870 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0);
871 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0);
872 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0);
873 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 0);
874
875 const float32x4_t vk00c1x0123 = vld1q_f32(w + 32);
876 const float32x4_t vk00c1x4567 = vld1q_f32(w + 36);
877
878 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk00c1x0123, vget_low_f32(vi0x0), 1);
879 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c1x0123, vget_low_f32(vi2x0), 1);
880 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c1x4567, vget_low_f32(vi0x0), 1);
881 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c1x4567, vget_low_f32(vi2x0), 1);
882
883 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40);
884 const float32x4_t vk10c1x4567 = vld1q_f32(w + 44);
885
886 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1);
887 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1);
888 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1);
889 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c1x4567, vget_low_f32(vi3x0), 1);
890
891 const float32x4_t vk20c1x0123 = vld1q_f32(w + 48);
892 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52);
893
894 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1);
895 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c1x0123, vget_low_f32(vi4x0), 1);
896 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1);
897 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c1x4567, vget_low_f32(vi4x0), 1);
898
899 const float32x4_t vk00c2x0123 = vld1q_f32(w + 56);
900 const float32x4_t vk00c2x4567 = vld1q_f32(w + 60);
901
902 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk00c2x0123, vget_high_f32(vi0x0), 0);
903 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 0);
904 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 0);
905 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c2x4567, vget_high_f32(vi2x0), 0);
906
907 const float32x4_t vk10c2x0123 = vld1q_f32(w + 64);
908 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68);
909
910 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c2x0123, vget_high_f32(vi1x0), 0);
911 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 0);
912 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0);
913 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 0);
914
915 const float32x4_t vk20c2x0123 = vld1q_f32(w + 72);
916 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76);
917
918 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c2x0123, vget_high_f32(vi2x0), 0);
919 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 0);
920 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0);
921 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 0);
922
923 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80);
924 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84);
925
926 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1);
927 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);
928 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1);
929 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c0x4567, vget_high_f32(vi2x0), 1);
930
931 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88);
932 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92);
933
934 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1);
935 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1);
936 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1);
937 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c0x4567, vget_high_f32(vi3x0), 1);
938
939 const float32x4_t vk21c0x0123 = vld1q_f32(w + 96);
940 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100);
941
942 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c0x0123, vget_high_f32(vi2x0), 1);
943 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c0x0123, vget_high_f32(vi4x0), 1);
944 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1);
945 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c0x4567, vget_high_f32(vi4x0), 1);
946
947 const float32x4_t vk01c1x0123 = vld1q_f32(w + 104);
948 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108);
949
950 // viMx1 = ( iM1c2, iM1c1 )
951 const float32x2_t vi0x1 = vld1_f32(i0); i0 += 2;
952 const float32x2_t vi1x1 = vld1_f32(i1); i1 += 2;
953 const float32x2_t vi2x1 = vld1_f32(i2); i2 += 2;
954 const float32x2_t vi3x1 = vld1_f32(i3); i3 += 2;
955 const float32x2_t vi4x1 = vld1_f32(i4); i4 += 2;
956
957 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c1x0123, vi0x1, 0);
958 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c1x0123, vi2x1, 0);
959 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vi0x1, 0);
960 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c1x4567, vi2x1, 0);
961
962 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112);
963 const float32x4_t vk11c1x4567 = vld1q_f32(w + 116);
964
965 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vi1x1, 0);
966 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vi3x1, 0);
967 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c1x4567, vi1x1, 0);
968 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c1x4567, vi3x1, 0);
969
970 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120);
971 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124);
972
973 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vi2x1, 0);
974 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vi4x1, 0);
975 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vi2x1, 0);
976 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c1x4567, vi4x1, 0);
977
978 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128);
979 const float32x4_t vk01c2x4567 = vld1q_f32(w + 132);
980
981 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vi0x1, 1);
982 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vi2x1, 1);
983 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c2x4567, vi0x1, 1);
984 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c2x4567, vi2x1, 1);
985
986 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136);
987 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140);
988
989 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vi1x1, 1);
990 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vi3x1, 1);
991 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vi1x1, 1);
992 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c2x4567, vi3x1, 1);
993
994 const float32x4_t vk21c2x0123 = vld1q_f32(w + 144);
995 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148);
996
997 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c2x0123, vi2x1, 1);
998 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c2x0123, vi4x1, 1);
999 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vi2x1, 1);
1000 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c2x4567, vi4x1, 1);
1001
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001002 const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
1003 const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
1004
Marat Dukhance7a3f82020-05-17 21:46:44 -07001005 vo0c0123 = vmaxq_f32(vo0c0123, vmin);
1006 vo1c0123 = vmaxq_f32(vo1c0123, vmin);
1007 vo0c4567 = vmaxq_f32(vo0c4567, vmin);
1008 vo1c4567 = vmaxq_f32(vo1c4567, vmin);
1009
1010 vo0c0123 = vminq_f32(vo0c0123, vmax);
1011 vo1c0123 = vminq_f32(vo1c0123, vmax);
1012 vo0c4567 = vminq_f32(vo0c4567, vmax);
1013 vo1c4567 = vminq_f32(vo1c4567, vmax);
1014
1015 if XNN_LIKELY(c >= 8) {
1016 vst1q_f32(o1, vo1c0123);
1017 vst1q_f32(o1 + 4, vo1c4567);
1018 o1 = (float*) ((uintptr_t) o1 + output_width_stride);
1019 vst1q_f32(o0, vo0c0123);
1020 vst1q_f32(o0 + 4, vo0c4567);
1021 o0 = (float*) ((uintptr_t) o0 + output_width_stride);
1022 } else {
1023 float* o0_tmp = o0;
1024 float* o1_tmp = o1;
1025 if (c & 4) {
1026 vst1q_f32(o1_tmp, vo1c0123); o1_tmp += 4;
1027 vo1c0123 = vo1c4567;
1028 vst1q_f32(o0_tmp, vo0c0123); o0_tmp += 4;
1029 vo0c0123 = vo0c4567;
1030 }
1031 float32x2_t vo0c01 = vget_low_f32(vo0c0123);
1032 float32x2_t vo1c01 = vget_low_f32(vo1c0123);
1033 if (c & 2) {
1034 vst1_f32(o1_tmp, vo1c01); o1_tmp += 2;
1035 vo1c01 = vget_high_f32(vo1c0123);
1036 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2;
1037 vo0c01 = vget_high_f32(vo0c0123);
1038 }
1039 if (c & 1) {
1040 vst1_lane_f32(o1_tmp, vo1c01, 0);
1041 vst1_lane_f32(o0_tmp, vo0c01, 0);
1042 }
1043 o0 = (float*) ((uintptr_t) o0 + output_width_stride);
1044 o1 = (float*) ((uintptr_t) o1 + output_width_stride);
1045 }
1046 }
1047 // Move output pointers back to the position of the first pixel in a row,
1048 // and forward to the next block of output channels
1049 o0 = (float*) ((uintptr_t) o0 - output_channel_decrement);
1050 o1 = (float*) ((uintptr_t) o1 - output_channel_decrement);
1051 // Revert input pointers to the position of the first pixel in a row
1052 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement);
1053 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement);
1054 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement);
1055 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement);
1056 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement);
1057 // Move to the block of weights for the next 8 output channels
1058 w += 224;
1059 c = doz(c, 8);
1060 } while (c != 0);
1061 // Move output pointers back to the position of the first channel, and forward to the next block of rows
1062 o0 = (float*) ((uintptr_t) o0 + output_height_increment);
1063 o1 = (float*) ((uintptr_t) o1 + output_height_increment);
1064 // Move input pointers forward to the next four rows
1065 i0 = i4;
1066 i1 = (const float*) ((uintptr_t) i0 + input_height_stride);
1067 i2 = (const float*) ((uintptr_t) i1 + input_height_stride);
1068 i3 = (const float*) ((uintptr_t) i2 + input_height_stride);
1069 i4 = (const float*) ((uintptr_t) i3 + input_height_stride);
1070 }
1071}