48   const v128_t vk12 = wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()  local
215       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk12));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
216       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk12));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
217       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk12));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
218       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, vk12));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
219       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, vk12));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
220       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi6x5678, vk12));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
339       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk12));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
340       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk12));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
341       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk12));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
342       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, vk12));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
343       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, vk12));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
344       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi6x5678, vk12));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()