49   const v128_t vk20 = wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()  local
184       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
185       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk20));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
186       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk20));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
187       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk20));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
188       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x3456, vk20));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
189       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi7x3456, vk20));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
317       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
318       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk20));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
319       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk20));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
320       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk20));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
321       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x3456, vk20));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
322       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi7x3456, vk20));  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()