42   const v128_t vwGHIJ = wasm_v128_load(weights + 16);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()  local
134 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
135 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
136 …vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
137 …vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
168 …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_arm_splat_4x4_acc2()
169 …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_arm_splat_4x4_acc2()
170 …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_arm_splat_4x4_acc2()
171 …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_arm_splat_4x4_acc2()
210 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
211 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
212 …vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
213 …vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
244 …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_arm_splat_4x4_acc2()
245 …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_arm_splat_4x4_acc2()
246 …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_arm_splat_4x4_acc2()
247 …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_arm_splat_4x4_acc2()
355 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
356 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
357 …vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
358 …vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
389 …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_arm_splat_4x4_acc2()
390 …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_arm_splat_4x4_acc2()
391 …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_arm_splat_4x4_acc2()
392 …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_arm_splat_4x4_acc2()
431 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
432 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
433 …vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
434 …vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
465 …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_arm_splat_4x4_acc2()
466 …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_arm_splat_4x4_acc2()
467 …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_arm_splat_4x4_acc2()
468 …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_arm_splat_4x4_acc2()
570 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
571 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
572 …vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
573 …vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
604 …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_arm_splat_4x4_acc2()
605 …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_arm_splat_4x4_acc2()
606 …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_arm_splat_4x4_acc2()
607 …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_arm_splat_4x4_acc2()
638 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
639 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
640 …vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
641 …vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi6x2345, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
672 …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_arm_splat_4x4_acc2()
673 …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_arm_splat_4x4_acc2()
674 …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_arm_splat_4x4_acc2()
675 …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_arm_splat_4x4_acc2()