107         v128_t vo0x1 = vo0x0;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()  local
121 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk00c0, wasm_v32x4_shuffle(vi0x1, vi0x1, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
128 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk10c0, wasm_v32x4_shuffle(vi1x1, vi1x1, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
135 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk20c0, wasm_v32x4_shuffle(vi2x1, vi2x1, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
149 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk00c1, wasm_v32x4_shuffle(vi0x2, vi0x2, 0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
156 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk10c1, wasm_v32x4_shuffle(vi1x2, vi1x2, 0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
163 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk20c1, wasm_v32x4_shuffle(vi2x2, vi2x2, 0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
170 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk00c2, wasm_v32x4_shuffle(vi0x2, vi0x2, 1, 1, 1, 1))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
177 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk10c2, wasm_v32x4_shuffle(vi1x2, vi1x2, 1, 1, 1, 1))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
184 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk20c2, wasm_v32x4_shuffle(vi2x2, vi2x2, 1, 1, 1, 1))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
191 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk01c0, wasm_v32x4_shuffle(vi0x2, vi0x2, 2, 2, 2, 2))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
198 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk11c0, wasm_v32x4_shuffle(vi1x2, vi1x2, 2, 2, 2, 2))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
205 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk21c0, wasm_v32x4_shuffle(vi2x2, vi2x2, 2, 2, 2, 2))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
212 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk01c1, wasm_v32x4_shuffle(vi0x2, vi0x2, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
219 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk11c1, wasm_v32x4_shuffle(vi1x2, vi1x2, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
226 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk21c1, wasm_v32x4_shuffle(vi2x2, vi2x2, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
240 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk01c2, wasm_v32x4_shuffle(vi0x3, vi0x3, 0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
247 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk11c2, wasm_v32x4_shuffle(vi1x3, vi1x3, 0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
254 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk21c2, wasm_v32x4_shuffle(vi2x3, vi2x3, 0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
261 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk02c0, wasm_v32x4_shuffle(vi0x3, vi0x3, 1, 1, 1, 1))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
268 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk12c0, wasm_v32x4_shuffle(vi1x3, vi1x3, 1, 1, 1, 1))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
275 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk22c0, wasm_v32x4_shuffle(vi2x3, vi2x3, 1, 1, 1, 1))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
282 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk02c1, wasm_v32x4_shuffle(vi0x3, vi0x3, 2, 2, 2, 2))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
289 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk12c1, wasm_v32x4_shuffle(vi1x3, vi1x3, 2, 2, 2, 2))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
296 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk22c1, wasm_v32x4_shuffle(vi2x3, vi2x3, 2, 2, 2, 2))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
303 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk02c2, wasm_v32x4_shuffle(vi0x3, vi0x3, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
310 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk12c2, wasm_v32x4_shuffle(vi1x3, vi1x3, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
317 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk22c2, wasm_v32x4_shuffle(vi2x3, vi2x3, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
328         vo0x1 = wasm_v128_bitselect(vmin, vo0x1, wasm_f32x4_lt(vo0x1, vmin));  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
333         vo0x1 = wasm_v128_bitselect(vo0x1, vmax, wasm_f32x4_le(vo0x1, vmax));  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
336         const v128_t vo0c01 = wasm_v32x4_shuffle(vo0x0, vo0x1, 0, 4 + 0, 1, 4 + 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
337         const v128_t vo0c23 = wasm_v32x4_shuffle(vo0x0, vo0x1, 2, 4 + 2, 3, 4 + 3);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
355         v128_t vo0x1 = vo0x0;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()  local
370 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk00c0, wasm_v32x4_shuffle(vi0x1, vi0x1, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
379 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk10c0, wasm_v32x4_shuffle(vi1x1, vi1x1, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
388 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk20c0, wasm_v32x4_shuffle(vi2x1, vi2x1, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
410 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk00c1, wasm_v32x4_shuffle(vi0x2, vi0x2, 0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
417 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk10c1, wasm_v32x4_shuffle(vi1x2, vi1x2, 0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
424 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk20c1, wasm_v32x4_shuffle(vi2x2, vi2x2, 0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
431 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk00c2, wasm_v32x4_shuffle(vi0x2, vi0x2, 1, 1, 1, 1))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
438 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk10c2, wasm_v32x4_shuffle(vi1x2, vi1x2, 1, 1, 1, 1))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
445 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk20c2, wasm_v32x4_shuffle(vi2x2, vi2x2, 1, 1, 1, 1))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
453 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk01c0, wasm_v32x4_shuffle(vi0x2, vi0x2, 2, 2, 2, 2))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
462 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk11c0, wasm_v32x4_shuffle(vi1x2, vi1x2, 2, 2, 2, 2))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
471 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk21c0, wasm_v32x4_shuffle(vi2x2, vi2x2, 2, 2, 2, 2))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
480 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk01c1, wasm_v32x4_shuffle(vi0x2, vi0x2, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
489 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk11c1, wasm_v32x4_shuffle(vi1x2, vi1x2, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
498 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk21c1, wasm_v32x4_shuffle(vi2x2, vi2x2, 3, 3, 3, 3))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
520 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk01c2, wasm_v32x4_shuffle(vi0x3, vi0x3, 0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
527 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk11c2, wasm_v32x4_shuffle(vi1x3, vi1x3, 0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
534 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk21c2, wasm_v32x4_shuffle(vi2x3, vi2x3, 0, 0, 0, 0))…  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
586         vo0x1 = wasm_v128_bitselect(vmin, vo0x1, wasm_f32x4_lt(vo0x1, vmin));  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
591         vo0x1 = wasm_v128_bitselect(vo0x1, vmax, wasm_f32x4_le(vo0x1, vmax));  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
596           const v128_t vo0c01 = wasm_v32x4_shuffle(vo0x0, vo0x1, 0, 4 + 0, 1, 4 + 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
597           const v128_t vo0c23 = wasm_v32x4_shuffle(vo0x0, vo0x1, 2, 4 + 2, 3, 4 + 3);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()