Home
last modified time | relevance | path

Searched refs:vo0x0c0123 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x4-neonfma-2x2.c93 float32x4_t vo0x0c0123 = vld1q_f32(w); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
94 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
95 float32x4_t vo0x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
96 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
107 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk00c0x0123, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
115 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
123 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
138 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
146 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
154 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
D3x3s2p1c3x4-neon-2x2.c91 float32x4_t vo0x0c0123 = vld1q_f32(w); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
92 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
93 float32x4_t vo0x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
94 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
105 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c0x0123, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
113 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
121 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
136 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
144 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
152 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x2.c91 float32x4_t vo0x0c0123 = vld1q_f32(w); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
92 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
93 float32x4_t vo0x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
94 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
105 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c0x0123, vget_low_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
113 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
121 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
129 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
137 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
145 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neonfma-2x2.c93 float32x4_t vo0x0c0123 = vld1q_f32(w); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
94 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
95 float32x4_t vo0x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
96 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
107 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk00c0x0123, vget_low_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
115 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
123 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
131 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
139 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
147 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neonfma-2x2.c93 float32x4_t vo0x0c0123 = vld1q_f32(w); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
95 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
97 float32x4_t vo0x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
99 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
112 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk00c0x0123, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
125 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
138 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
158 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
171 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
184 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c91 float32x4_t vo0x0c0123 = vld1q_f32(w); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
93 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
95 float32x4_t vo0x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
97 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
110 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c0x0123, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
123 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
136 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
156 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
169 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
182 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c93 float32x4_t vo0x0c0123 = vld1q_f32(w); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
95 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
97 float32x4_t vo0x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
99 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
112 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk00c0x0123, vget_low_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
125 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
138 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
151 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
164 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
177 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c91 float32x4_t vo0x0c0123 = vld1q_f32(w); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
93 float32x4_t vo1x0c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
95 float32x4_t vo0x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
97 float32x4_t vo1x1c0123 = vo0x0c0123; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
110 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c0x0123, vget_low_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
123 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
136 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
149 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
162 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
175 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]