Lines Matching refs:vo0p0
136 v128_t vo0p0 = vbias; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local
138 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
140 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, vk12)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
142 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x8ACE, vk22)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
144 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x8ACE, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
146 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
148 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x9BDF, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
150 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
152 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x9BDF, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
154 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x9BDF, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
156 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
169 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
171 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
173 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
175 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
177 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x68AC, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
217 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
219 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
221 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
223 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
225 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x79BD, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
243 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
245 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
247 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
249 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
251 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
254 v128_t vo0 = wasm_f32x4_max(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
263 v128_t vo0p0 = vbias; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local
277 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x8ACE, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
279 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, vk12)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
281 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x8ACE, vk22)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
283 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x8ACE, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
285 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x8ACE, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
287 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x9BDF, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
289 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
291 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x9BDF, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
293 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x9BDF, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
295 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
303 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x68AC, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
305 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
307 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
309 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
311 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x68AC, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
319 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
321 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
323 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
325 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
327 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x79BD, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
335 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
337 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
339 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
341 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
343 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
346 v128_t vo0 = wasm_f32x4_max(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()