Lines Matching refs:vmax
66 float32x4_t vmax = vi0; in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local
69 const uint32x4_t vm1 = vcgtq_f32(vi1, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
70 vmax = vbslq_f32(vm1, vi1, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
73 const uint32x4_t vm2 = vcgtq_f32(vi2, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
74 vmax = vbslq_f32(vm2, vi2, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
77 const uint32x4_t vm3 = vcgtq_f32(vi3, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
78 vmax = vbslq_f32(vm3, vi3, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
81 const uint32x4_t vm4 = vcgtq_f32(vi4, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
82 vmax = vbslq_f32(vm4, vi4, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
85 const uint32x4_t vm5 = vcgtq_f32(vi5, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
86 vmax = vbslq_f32(vm5, vi5, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
89 const uint32x4_t vm6 = vcgtq_f32(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
90 vmax = vbslq_f32(vm6, vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
93 const uint32x4_t vm7 = vcgtq_f32(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
94 vmax = vbslq_f32(vm7, vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
97 const uint32x4_t vm8 = vcgtq_f32(vi8, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
98 vmax = vbslq_f32(vm8, vi8, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
101 vst1q_f32(ab, vmax); ab += 4; in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
141 float32x4_t vmax = vld1q_f32(ab); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local
144 const uint32x4_t vm0 = vcgtq_f32(vi0, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
145 vmax = vbslq_f32(vm0, vi0, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
148 const uint32x4_t vm1 = vcgtq_f32(vi1, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
150 vmax = vbslq_f32(vm1, vi1, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
153 const uint32x4_t vm2 = vcgtq_f32(vi2, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
155 vmax = vbslq_f32(vm2, vi2, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
158 const uint32x4_t vm3 = vcgtq_f32(vi3, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
160 vmax = vbslq_f32(vm3, vi3, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
163 const uint32x4_t vm4 = vcgtq_f32(vi4, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
165 vmax = vbslq_f32(vm4, vi4, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
168 const uint32x4_t vm5 = vcgtq_f32(vi5, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
170 vmax = vbslq_f32(vm5, vi5, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
173 const uint32x4_t vm6 = vcgtq_f32(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
175 vmax = vbslq_f32(vm6, vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
178 const uint32x4_t vm7 = vcgtq_f32(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
180 vmax = vbslq_f32(vm7, vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
183 vst1q_f32(ab, vmax); ab += 4; in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
244 float32x4_t vmax = vld1q_f32(ab); ab += 4; in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local
247 const uint32x4_t vm0 = vcgtq_f32(vi0, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
248 vmax = vbslq_f32(vm0, vi0, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
251 const uint32x4_t vm1 = vcgtq_f32(vi1, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
253 vmax = vbslq_f32(vm1, vi1, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
256 const uint32x4_t vm2 = vcgtq_f32(vi2, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
258 vmax = vbslq_f32(vm2, vi2, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
261 const uint32x4_t vm3 = vcgtq_f32(vi3, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
263 vmax = vbslq_f32(vm3, vi3, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
266 const uint32x4_t vm4 = vcgtq_f32(vi4, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
268 vmax = vbslq_f32(vm4, vi4, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
271 const uint32x4_t vm5 = vcgtq_f32(vi5, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
273 vmax = vbslq_f32(vm5, vi5, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
276 const uint32x4_t vm6 = vcgtq_f32(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
278 vmax = vbslq_f32(vm6, vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
281 const uint32x4_t vm7 = vcgtq_f32(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
283 vmax = vbslq_f32(vm7, vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
286 vst1q_f32(o, vmax); o += 4; in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
299 float32x4_t vmax = vld1q_f32(ab); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local
302 const uint32x4_t vm0 = vcgtq_f32(vi0, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
303 vmax = vbslq_f32(vm0, vi0, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
306 const uint32x4_t vm1 = vcgtq_f32(vi1, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
308 vmax = vbslq_f32(vm1, vi1, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
311 const uint32x4_t vm2 = vcgtq_f32(vi2, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
313 vmax = vbslq_f32(vm2, vi2, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
316 const uint32x4_t vm3 = vcgtq_f32(vi3, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
318 vmax = vbslq_f32(vm3, vi3, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
321 const uint32x4_t vm4 = vcgtq_f32(vi4, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
323 vmax = vbslq_f32(vm4, vi4, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
326 const uint32x4_t vm5 = vcgtq_f32(vi5, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
328 vmax = vbslq_f32(vm5, vi5, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
331 const uint32x4_t vm6 = vcgtq_f32(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
333 vmax = vbslq_f32(vm6, vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
336 const uint32x4_t vm7 = vcgtq_f32(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
338 vmax = vbslq_f32(vm7, vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
341 float32x2_t vmax_lo = vget_low_f32(vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
346 vmax_lo = vget_high_f32(vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()