Home
last modified time | relevance | path

Searched refs:vk10c1x0123 (Results 1 – 17 of 17) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x4-neonfma-2x2.c144 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
146 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_high_f32(vi1x0), 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()
149 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_low_f32(vi1x2), 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()
419 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
421 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
422 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
610 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
612 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
D3x3s2p1c3x4-neon-2x2.c142 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
144 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_high_f32(vi1x0), 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()
147 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_low_f32(vi1x2), 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()
419 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
421 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
422 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
612 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
614 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x2.c135 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
137 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_low_f32(vi1x0), 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()
140 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_high_f32(vi1x1), 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()
419 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
421 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
422 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
612 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
614 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neonfma-2x2.c137 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
139 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_low_f32(vi1x0), 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()
142 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_high_f32(vi1x1), 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()
419 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
421 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
422 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
610 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
612 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x1.c114 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
116 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
117 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
308 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
310 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
311 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
D3x3s2p0p1c3x4-neonfma-2x1.c116 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
118 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
119 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
308 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
310 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
311 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neonfma-2x1.c116 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
118 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
119 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
308 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
310 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
311 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neon-2x1.c114 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
116 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
117 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
308 const float32x4_t vk10c1x0123 = vld1q_f32(w + 20); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
310 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
311 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
D3x3s2p1c3x8-neonfma-2x2.c168 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
171 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_high_f32(vi1x0), 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()
176 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_low_f32(vi1x2), 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()
595 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
598 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
599 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
881 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
884 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c166 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
169 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_high_f32(vi1x0), 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()
174 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_low_f32(vi1x2), 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()
595 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
598 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
599 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
883 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
886 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c161 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
164 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_low_f32(vi1x0), 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()
169 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_high_f32(vi1x1), 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()
595 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
598 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
599 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
881 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
884 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c159 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
162 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_low_f32(vi1x0), 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()
167 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_high_f32(vi1x1), 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()
595 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
598 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
599 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
883 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
886 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x1.c130 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
133 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
134 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
417 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
420 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
421 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neon-2x1.c128 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
131 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
132 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
417 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
420 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
421 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neon-2x1.c128 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
131 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
132 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
417 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
420 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
421 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
D3x3s2p1c3x8-neonfma-2x1.c130 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
133 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
134 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
417 const float32x4_t vk10c1x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
420 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
421 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
/external/XNNPACK/src/f32-conv-hwc2chw/
D3x3s2p1c3x4-sse-1x1.c102 const __m128 vk10c1x0123 = _mm_load_ps(w + 20); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
103 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk10c1x0123, vi10c1)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
229 const __m128 vk10c1x0123 = _mm_load_ps(w + 20); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
230 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk10c1x0123, vi10c1)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()