116       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()  local
118 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
120 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
122 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x8ACE, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
124 …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_1x4()
126 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
128 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
130 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
132 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x9BDF, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
134 …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_1x4()
136 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
149 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x68AC, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
151 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x68AC, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
153 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
155 …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_1x4()
157 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x68AC, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
197 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
199 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x79BD, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
201 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x79BD, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
203 …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_1x4()
205 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x79BD, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
223 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0xACEG, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
225 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
227 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
229 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
231 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1))…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
234       v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin));  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
243       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()  local
257 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
259 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
261 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x8ACE, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
263 …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_1x4()
265 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
267 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
269 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
271 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x9BDF, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
273 …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_1x4()
275 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
283 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x68AC, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
285 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x68AC, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
287 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
289 …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_1x4()
291 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x68AC, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
299 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
301 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x79BD, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
303 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x79BD, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
305 …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_1x4()
307 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x79BD, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
315 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0xACEG, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
317 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
319 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
321 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
323 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1))…  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
326       v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin));  in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()