111         float vo0p0 = vbias + vi0x0 * vk00;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()  local
112         vo0p0 += vi1x0 * vk10;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
113         vo0p0 += vi2x0 * vk20;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
114         vo0p0 += vi3x0 * vk30;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
115         vo0p0 += vi4x0 * vk40;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
123         vo0p0 += vi0x1 * vk01;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
124         vo0p0 += vi1x1 * vk11;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
125         vo0p0 += vi2x1 * vk21;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
126         vo0p0 += vi3x1 * vk31;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
127         vo0p0 += vi4x1 * vk41;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
135         vo0p0 += vi0x2 * vk02;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
136         vo0p0 += vi1x2 * vk12;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
137         vo0p0 += vi2x2 * vk22;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
138         vo0p0 += vi3x2 * vk32;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
139         vo0p0 += vi4x2 * vk42;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
147         vo0p0 += vi0x3 * vk03;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
148         vo0p0 += vi1x3 * vk13;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
149         vo0p0 += vi2x3 * vk23;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
150         vo0p0 += vi3x3 * vk33;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
151         vo0p0 += vi4x3 * vk43;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
159         vo0p0 += vi0x4 * vk04;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
160         vo0p0 += vi1x4 * vk14;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
161         vo0p0 += vi2x4 * vk24;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
162         vo0p0 += vi3x4 * vk34;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
163         vo0p0 += vi4x4 * vk44;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
166         float vo0 = math_max_f32(vo0p0, vmin);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
174         float vo0p0 = vbias + vi0x0 * vk00;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()  local
175         vo0p0 += vi1x0 * vk10;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
176         vo0p0 += vi2x0 * vk20;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
177         vo0p0 += vi3x0 * vk30;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
178         vo0p0 += vi4x0 * vk40;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
186         vo0p0 += vi0x1 * vk01;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
187         vo0p0 += vi1x1 * vk11;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
188         vo0p0 += vi2x1 * vk21;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
189         vo0p0 += vi3x1 * vk31;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
190         vo0p0 += vi4x1 * vk41;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
198         vo0p0 += vi0x2 * vk02;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
199         vo0p0 += vi1x2 * vk12;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
200         vo0p0 += vi2x2 * vk22;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
201         vo0p0 += vi3x2 * vk32;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
202         vo0p0 += vi4x2 * vk42;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
210         vo0p0 += vi0x3 * vk03;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
211         vo0p0 += vi1x3 * vk13;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
212         vo0p0 += vi2x3 * vk23;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
213         vo0p0 += vi3x3 * vk33;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
214         vo0p0 += vi4x3 * vk43;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
217         float vo0 = math_max_f32(vo0p0, vmin);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
227       float vo0p0 = vbias + vi0x0 * vk00;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()  local
228       vo0p0 += vi1x0 * vk10;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
229       vo0p0 += vi2x0 * vk20;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
230       vo0p0 += vi3x0 * vk30;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
231       vo0p0 += vi4x0 * vk40;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
233       vo0p0 += vi0x1 * vk01;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
234       vo0p0 += vi1x1 * vk11;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
235       vo0p0 += vi2x1 * vk21;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
236       vo0p0 += vi3x1 * vk31;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
237       vo0p0 += vi4x1 * vk41;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
239       vo0p0 += vi0x2 * vk02;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
240       vo0p0 += vi1x2 * vk12;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
241       vo0p0 += vi2x2 * vk22;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
242       vo0p0 += vi3x2 * vk32;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
243       vo0p0 += vi4x2 * vk42;  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
246       float vo0 = math_max_f32(vo0p0, vmin);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()