42   const v128_t vwGHIJ = wasm_v128_load(weights + 16);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4()  local
121 …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_arm_splat_3x4()
122 …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_arm_splat_3x4()
123 …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_arm_splat_3x4()
149 …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_3x4()
150 …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_3x4()
151 …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_3x4()
184 …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_arm_splat_3x4()
185 …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_arm_splat_3x4()
186 …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_arm_splat_3x4()
212 …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_3x4()
213 …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_3x4()
214 …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_3x4()
301 …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_arm_splat_3x4()
302 …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_arm_splat_3x4()
303 …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_arm_splat_3x4()
329 …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_3x4()
330 …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_3x4()
331 …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_3x4()
364 …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_arm_splat_3x4()
365 …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_arm_splat_3x4()
366 …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_arm_splat_3x4()
392 …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_3x4()
393 …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_3x4()
394 …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_3x4()
476 …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_arm_splat_3x4()
477 …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_arm_splat_3x4()
478 …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_arm_splat_3x4()
504 …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_3x4()
505 …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_3x4()
506 …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_3x4()
532 …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_arm_splat_3x4()
533 …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_arm_splat_3x4()
534 …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_arm_splat_3x4()
560 …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_3x4()
561 …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_3x4()
562 …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_3x4()