Home
last modified time | relevance | path

Searched refs:vk01c0x0123 (Results 1 – 17 of 17) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x4-neonfma-2x2.c184 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
186 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
187 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
189 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c0x0123, vget_high_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
190 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c0x0123, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
451 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
453 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
454 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
642 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
644 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
D3x3s2p1c3x4-neon-2x2.c182 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
184 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
185 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
187 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c0x0123, vget_high_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
188 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c0x0123, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
451 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
453 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
454 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
644 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
646 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x2.c182 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
184 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
185 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
187 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c0x0123, vget_low_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
188 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c0x0123, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
444 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
446 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
447 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
637 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
639 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neonfma-2x2.c184 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
186 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
187 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
189 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c0x0123, vget_low_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
190 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c0x0123, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
444 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
446 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
447 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
635 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
637 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x1.c139 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
141 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
142 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
333 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
335 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
336 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
D3x3s2p0p1c3x4-neonfma-2x1.c141 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
143 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
144 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
333 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
335 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
336 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neonfma-2x1.c148 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
150 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
151 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
340 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
342 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
343 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neon-2x1.c146 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
148 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
149 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
340 const float32x4_t vk01c0x0123 = vld1q_f32(w + 40); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
342 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
343 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
D3x3s2p1c3x8-neonfma-2x2.c233 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
236 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
237 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
241 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c0x0123, vget_high_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
242 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c0x0123, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
642 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
645 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
646 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
928 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
931 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c231 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
234 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
235 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
239 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c0x0123, vget_high_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
240 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c0x0123, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
642 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
645 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
646 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
930 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
933 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c233 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
236 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
237 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
241 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c0x0123, vget_low_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
242 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c0x0123, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
635 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
638 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
639 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
921 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
924 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c231 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
234 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
235 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
239 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c0x0123, vget_low_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
240 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c0x0123, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
635 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
638 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
639 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
923 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
926 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x1.c170 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
173 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
174 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
457 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
460 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
461 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neon-2x1.c175 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
178 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
179 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
464 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
467 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
468 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neon-2x1.c168 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
171 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
172 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
457 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
460 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
461 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
D3x3s2p1c3x8-neonfma-2x1.c177 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
180 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
181 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
464 const float32x4_t vk01c0x0123 = vld1q_f32(w + 80); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
467 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
468 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
/external/XNNPACK/src/f32-conv-hwc2chw/
D3x3s2p1c3x4-sse-1x1.c117 const __m128 vk01c0x0123 = _mm_load_ps(w + 40); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
119 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk01c0x0123, vi01c0)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
244 const __m128 vk01c0x0123 = _mm_load_ps(w + 40); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
246 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk01c0x0123, vi01c0)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()