40   const v128_t vw89 = wasm_v64x2_load_splat(weights + 8);  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()  local
114 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
115 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
116 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
117 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
165 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
166 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
167 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
168 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
218 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
219 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
220 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
221 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
262 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
263 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
264 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()
265 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))…  in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4()