42   const v128_t vwGHIJ = wasm_v128_load(weights + 16);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()  local
147 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
148 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
149 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
150 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
151 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
187 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
188 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
189 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
190 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
191 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
236 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
237 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
238 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
239 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
240 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
276 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
277 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
278 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
279 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
280 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
401 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
402 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
403 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
404 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
405 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
441 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
442 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
443 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
444 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
445 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
490 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
491 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
492 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
493 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
494 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
530 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
531 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
532 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
533 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
534 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
648 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
649 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
650 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
651 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
652 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
688 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
689 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
690 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
691 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
692 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
728 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
729 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
730 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
731 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
732 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
768 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
769 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
770 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
771 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
772 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()