Home
last modified time | relevance | path

Searched refs:vo1x1c0123 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p0p1c3x4-neonfma-2x2.c96 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
111 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
119 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
127 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
135 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
143 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
151 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
166 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c2x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
174 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c2x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
182 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c2x0123, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
[all …]
D3x3s2p1c3x4-neon-2x2.c94 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
109 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
117 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
125 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
140 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
148 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
156 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c1x0123, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
164 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c2x0123, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
172 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c2x0123, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
180 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c2x0123, vget_low_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x2.c94 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
109 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
117 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
125 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
133 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
141 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
149 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
164 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c2x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
172 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c2x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
180 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c2x0123, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
[all …]
D3x3s2p1c3x4-neonfma-2x2.c96 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
111 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
119 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
127 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
142 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
150 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
158 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c1x0123, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
166 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c2x0123, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
174 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c2x0123, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
182 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c2x0123, vget_low_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c97 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
116 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
129 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
142 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
155 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
168 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
181 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
201 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c2x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
214 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c2x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
227 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c2x0123, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neonfma-2x2.c99 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
118 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
131 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
144 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
164 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
177 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
190 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c1x0123, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
203 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c2x0123, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
216 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c2x0123, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
229 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c2x0123, vget_low_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c99 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
118 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
131 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
144 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
157 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
170 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
183 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
203 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c2x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
216 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c2x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
229 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c2x0123, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c97 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
116 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
129 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
142 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
162 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
175 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
188 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c1x0123, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
201 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c2x0123, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
214 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c2x0123, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
227 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c2x0123, vget_low_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]