114         const __m128 vi2x1 = _mm_loadu_ps(i2); i2 += 4;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()  local
121 …vo1x1 = _mm_add_ps(vo1x1, _mm_mul_ps(vk00c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
134 …vo0x1 = _mm_add_ps(vo0x1, _mm_mul_ps(vk20c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
189 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk01c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
202 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk21c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
210 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk01c1, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(1, 1, 1, 1))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
223 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk21c1, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(1, 1, 1, 1))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
238 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk01c2, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(2, 2, 2, 2))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
251 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk21c2, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(2, 2, 2, 2))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
259 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk02c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
272 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk22c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
363         __m128 vi2x1 = _mm_loadu_ps(i2);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()  local
371 …vo1x1 = _mm_add_ps(vo1x1, _mm_mul_ps(vk00c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
388 …vo0x1 = _mm_add_ps(vo0x1, _mm_mul_ps(vk20c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
451 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk01c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
468 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk21c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
478 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk01c1, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(1, 1, 1, 1))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
495 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk21c1, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(1, 1, 1, 1))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
519 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk01c2, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(2, 2, 2, 2))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
532 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk21c2, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(2, 2, 2, 2))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
541 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk02c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
550 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk22c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()