43   const v128_t vwKLMN = wasm_v128_load(weights + 20);  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()  local
111 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
112 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
133 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
134 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
161 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
162 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
208 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
209 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
256 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
257 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
278 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
279 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
306 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
307 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
353 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
354 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
397 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
398 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
419 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
420 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
441 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
442 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
482 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
483 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,…  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()