Home
last modified time | relevance | path

Searched refs:vo1x0c0123 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x4-neonfma-2x2.c94 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
108 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
116 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
124 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
139 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
147 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
155 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c1x0123, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
163 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
171 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c2x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
179 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c2x0123, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
D3x3s2p1c3x4-neon-2x2.c92 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
106 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
114 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
122 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
137 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
145 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
153 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c1x0123, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
161 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
169 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c2x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
177 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c2x0123, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x2.c92 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
106 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
114 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
122 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
130 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
138 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
146 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c1x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
161 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
169 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c2x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
177 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c2x0123, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neonfma-2x2.c94 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
108 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
116 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
124 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
132 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
140 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
148 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c1x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
163 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
171 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c2x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
179 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c2x0123, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neonfma-2x2.c95 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
113 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
126 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
139 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
159 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
172 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
185 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c1x0123, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
198 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
211 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c2x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
224 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c2x0123, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c93 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
111 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
124 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
137 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
157 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
170 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
183 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c1x0123, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
196 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
209 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c2x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
222 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c2x0123, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c95 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
113 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
126 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
139 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
152 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
165 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
178 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c1x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
198 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
211 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c2x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
224 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c2x0123, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c93 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
111 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
124 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
137 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
150 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
163 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
176 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c1x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
196 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
209 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c2x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
222 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c2x0123, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]