50   const v128_t vk21 = wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()  local
156       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk21));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
157       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk21));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
158       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk21));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
159       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, vk21));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
160       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x4567, vk21));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
161       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi7x4567, vk21));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
289       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk21));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
290       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk21));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
291       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk21));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
292       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, vk21));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
293       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x4567, vk21));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
294       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi7x4567, vk21));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()