Home
last modified time | relevance | path

Searched refs:vo0c4567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x8-neon-2x1.c92 float32x4_t vo0c4567 = vld1q_f32(w + 4); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
94 float32x4_t vo1c4567 = vo0c4567; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
101 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
109 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
117 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
125 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c1x4567, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
133 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c1x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
141 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
149 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
157 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
[all …]
D3x3s2p0p1c3x8-neonfma-2x1.c94 float32x4_t vo0c4567 = vld1q_f32(w + 4); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
96 float32x4_t vo1c4567 = vo0c4567; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
103 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
111 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
119 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
127 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c1x4567, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
135 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
143 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
151 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
159 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
[all …]
D3x3s2p1c3x8-neonfma-2x1.c94 float32x4_t vo0c4567 = vld1q_f32(w + 4); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
96 float32x4_t vo1c4567 = vo0c4567; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
103 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
111 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
119 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
127 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c1x4567, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
135 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c1x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
143 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
151 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
159 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
[all …]
D3x3s2p0p1c3x8-neon-2x1.c92 float32x4_t vo0c4567 = vld1q_f32(w + 4); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
94 float32x4_t vo1c4567 = vo0c4567; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
101 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
109 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
117 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
125 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c1x4567, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
133 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
141 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
149 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
157 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c559 float32x4_t vo0c4567 = vld1q_f32(w + 4); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
561 float32x4_t vo1c4567 = vo0c4567; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
568 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
576 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
584 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
592 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c1x4567, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
600 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
608 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
616 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
624 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neonfma-2x2.c559 float32x4_t vo0c4567 = vld1q_f32(w + 4); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
561 float32x4_t vo1c4567 = vo0c4567; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
568 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
576 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
584 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
592 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c1x4567, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
600 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c1x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
608 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
616 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
624 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c559 float32x4_t vo0c4567 = vld1q_f32(w + 4); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
561 float32x4_t vo1c4567 = vo0c4567; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
568 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
576 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
584 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
592 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c1x4567, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
600 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
608 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
616 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
624 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c559 float32x4_t vo0c4567 = vld1q_f32(w + 4); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
561 float32x4_t vo1c4567 = vo0c4567; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
568 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
576 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
584 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
592 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c1x4567, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
600 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c1x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
608 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
616 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
624 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]