Lines Matching refs:vo0p2
117 v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, vk12); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3() local
121 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x4567, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
133 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
139 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
156 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
168 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
174 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
191 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
197 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
200 vo0p0 = wasm_f32x4_add(vo0p0, vo0p2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
225 v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, vk12); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3() local
229 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x4567, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
241 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
247 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
264 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
276 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
282 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
299 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
305 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
308 vo0p0 = wasm_f32x4_add(vo0p0, vo0p2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
330 v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, vk12); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3() local
334 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x4567, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
346 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
352 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
364 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
376 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
382 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
394 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
400 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()
403 vo0p0 = wasm_f32x4_add(vo0p0, vo0p2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_1x4_acc3()