170       float vo2p0 = vbias + vi4x0 * vk00;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()  local
173       vo2p0 += vi5x0 * vk10;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
176       vo2p0 += vi6x0 * vk20;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
179       vo2p0 += vi7x0 * vk30;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
182       vo2p0 += vi8x0 * vk40;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
196       vo2p0 += vi4x1 * vk01;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
199       vo2p0 += vi5x1 * vk11;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
202       vo2p0 += vi6x1 * vk21;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
205       vo2p0 += vi7x1 * vk31;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
208       vo2p0 += vi8x1 * vk41;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
222       vo2p0 += vi4x2 * vk02;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
225       vo2p0 += vi5x2 * vk12;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
228       vo2p0 += vi6x2 * vk22;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
231       vo2p0 += vi7x2 * vk32;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
234       vo2p0 += vi8x2 * vk42;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
248       vo2p0 += vi4x3 * vk03;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
251       vo2p0 += vi5x3 * vk13;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
254       vo2p0 += vi6x3 * vk23;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
257       vo2p0 += vi7x3 * vk33;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
260       vo2p0 += vi8x3 * vk43;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
264       vo2p0 += vi4x4 * vk04;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
267       vo2p0 += vi5x4 * vk14;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
270       vo2p0 += vi6x4 * vk24;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
273       vo2p0 += vi7x4 * vk34;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
276       vo2p0 += vi8x4 * vk44;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
281       float vo2 = math_max_f32(vo2p0, vmin);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
304       float vo2p0 = vbias + vi4x0 * vk00;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()  local
307       vo2p0 += vi5x0 * vk10;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
310       vo2p0 += vi6x0 * vk20;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
313       vo2p0 += vi7x0 * vk30;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
316       vo2p0 += vi8x0 * vk40;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
320       vo2p0 += vi4x1 * vk01;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
323       vo2p0 += vi5x1 * vk11;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
326       vo2p0 += vi6x1 * vk21;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
329       vo2p0 += vi7x1 * vk31;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
332       vo2p0 += vi8x1 * vk41;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
336       vo2p0 += vi4x2 * vk02;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
339       vo2p0 += vi5x2 * vk12;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
342       vo2p0 += vi6x2 * vk22;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
345       vo2p0 += vi7x2 * vk32;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
348       vo2p0 += vi8x2 * vk42;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
352       vo2p0 += vi4x3 * vk03;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
355       vo2p0 += vi5x3 * vk13;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
358       vo2p0 += vi6x3 * vk23;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
361       vo2p0 += vi7x3 * vk33;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
364       vo2p0 += vi8x3 * vk43;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
369       float vo2 = math_max_f32(vo2p0, vmin);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
381       float vo2p0 = vbias + vi4x0 * vk00;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()  local
384       vo2p0 += vi5x0 * vk10;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
387       vo2p0 += vi6x0 * vk20;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
390       vo2p0 += vi7x0 * vk30;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
393       vo2p0 += vi8x0 * vk40;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
397       vo2p0 += vi4x1 * vk01;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
400       vo2p0 += vi5x1 * vk11;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
403       vo2p0 += vi6x1 * vk21;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
406       vo2p0 += vi7x1 * vk31;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
409       vo2p0 += vi8x1 * vk41;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
413       vo2p0 += vi4x2 * vk02;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
416       vo2p0 += vi5x2 * vk12;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
419       vo2p0 += vi6x2 * vk22;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
422       vo2p0 += vi7x2 * vk32;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
425       vo2p0 += vi8x2 * vk42;  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
430       float vo2 = math_max_f32(vo2p0, vmin);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()