62   const __m128 vk44 = _mm_load1_ps(weights + 25);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()  local
359       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
360       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
361       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
362       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
363       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
612       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
613       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
614       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
615       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
616       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
842       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
843       vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
844       vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
845       vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
846       vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()