Home
last modified time | relevance | path

Searched refs:vk01c0x4567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p0p1c3x8-neon-2x2.c232 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
236 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
237 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk01c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
241 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk01c0x4567, vget_low_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
242 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk01c0x4567, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
636 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
640 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
641 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
924 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
928 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neonfma-2x2.c234 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
238 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk01c0x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
239 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk01c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
243 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk01c0x4567, vget_high_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
244 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk01c0x4567, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
643 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
647 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c0x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
648 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
929 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
933 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c0x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c234 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
238 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
239 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk01c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
243 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk01c0x4567, vget_low_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
244 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk01c0x4567, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
636 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
640 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
641 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
922 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
926 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c232 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
236 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk01c0x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
237 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk01c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
241 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk01c0x4567, vget_high_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
242 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk01c0x4567, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
643 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
647 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
648 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
931 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
935 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neon-2x1.c176 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
180 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
181 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
465 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
469 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
470 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neonfma-2x1.c171 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
175 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
176 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
458 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
462 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
463 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neonfma-2x1.c178 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
182 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c0x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
183 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
465 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
469 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c0x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
470 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x1.c169 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
173 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
174 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
458 const float32x4_t vk01c0x4567 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
462 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
463 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()