Lines Matching refs:vo0p0
92 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
94 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x8ACE9BDF.val[0], vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
96 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x8ACE9BDF.val[0], vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
98 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x8ACE9BDF.val[0], vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
100 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x8ACE9BDF.val[0], vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
102 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x8ACE9BDF.val[0], vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
104 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x8ACE9BDF.val[1], vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
106 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x8ACE9BDF.val[1], vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
108 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x8ACE9BDF.val[1], vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
110 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x8ACE9BDF.val[1], vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
112 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x8ACE9BDF.val[1], vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
125 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
127 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
129 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
131 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
133 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x68AC, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
152 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
154 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
156 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
158 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
160 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
173 vo0p0 = vmlaq_lane_f32(vo0p0, vi0xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
175 vo0p0 = vmlaq_lane_f32(vo0p0, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
177 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
179 vo0p0 = vmlaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
181 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
184 float32x4_t vo0 = vmaxq_f32(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
194 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
208 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x8ACE, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
210 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x8ACE, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
212 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x8ACE, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
214 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x8ACE, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
216 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x8ACE, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
218 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x9BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
220 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x9BDF, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
222 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x9BDF, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
224 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x9BDF, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
226 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x9BDF, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
234 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x68AC, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
236 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
238 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x68AC, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
240 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
242 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x68AC, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
250 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
252 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
254 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
256 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
258 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
267 vo0p0 = vmlaq_lane_f32(vo0p0, vi0xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
269 vo0p0 = vmlaq_lane_f32(vo0p0, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
271 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
273 vo0p0 = vmlaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
275 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
278 float32x4_t vo0 = vmaxq_f32(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()