322       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()  local
337       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x6789, vk04));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
342       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, vk14));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
347       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
352       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, vk34));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
357       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
576       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()  local
591       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x6789, vk04));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
596       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, vk14));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
601       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
606       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, vk34));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
611       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
810       const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x5678, vzero, 1, 2, 3, 4);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()  local
820       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x6789, vk04));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
825       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, vk14));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
830       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
835       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, vk34));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
840       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()