70   const v128_t vk44 =  wasm_v32x4_shuffle(vwOP,   vwOP,   1, 1, 1, 1);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()  local
317       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
318       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
319       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
320       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
538       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
539       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
540       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
541       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
737       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
738       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
739       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
740       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x6789, vk44));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()