44   const v128_t vwGHIJ = wasm_v128_load(weights + 16);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()  local
181 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
182 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
183 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi7x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
201 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
202 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
203 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi7x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
240 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
241 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
242 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi7x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
326 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
327 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
328 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi7x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
434 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
435 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
436 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi7x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
454 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
455 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
456 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi7x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
484 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
485 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
486 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi7x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
514 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
515 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
516 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi7x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()