Home
last modified time | relevance | path

Searched refs:vk01c2x0123 (Results 1 – 17 of 17) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p0p1c3x4-neonfma-2x2.c232 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
234 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
235 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
237 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c2x0123, vget_high_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
238 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c2x0123, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
481 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
483 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
484 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
672 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
674 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vi0x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
[all …]
D3x3s2p1c3x4-neon-2x2.c230 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
239 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
240 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
242 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c2x0123, vget_low_f32(vi0x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
243 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c2x0123, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
481 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
483 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
484 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
674 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
676 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x2.c230 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
232 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
233 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
235 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c2x0123, vget_high_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
236 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c2x0123, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
481 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
483 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
484 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
674 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
676 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vi0x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
[all …]
D3x3s2p1c3x4-neonfma-2x2.c232 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
241 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
242 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
244 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c2x0123, vget_low_f32(vi0x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
245 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c2x0123, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
481 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
483 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
484 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
672 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
674 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x1.c176 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
178 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
179 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
370 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
372 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vi0x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
373 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vi2x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
D3x3s2p1c3x4-neonfma-2x1.c178 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
180 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
181 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
370 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
372 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
373 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neon-2x1.c176 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
178 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
179 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
370 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
372 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
373 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
D3x3s2p0p1c3x4-neonfma-2x1.c178 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
180 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
181 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
370 const float32x4_t vk01c2x0123 = vld1q_f32(w + 64); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
372 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vi0x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
373 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vi2x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x2.c309 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
312 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
313 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
317 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c2x0123, vget_high_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
318 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c2x0123, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
690 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
693 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
694 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
978 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
981 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vi0x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neonfma-2x2.c311 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
321 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
322 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
326 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c2x0123, vget_low_f32(vi0x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
327 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c2x0123, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
690 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
693 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
694 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
976 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
979 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c311 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
314 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
315 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
319 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c2x0123, vget_high_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
320 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c2x0123, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
690 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
693 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
694 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
976 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
979 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vi0x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c309 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
319 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
320 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
324 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk01c2x0123, vget_low_f32(vi0x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
325 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk01c2x0123, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
690 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
693 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
694 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
978 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
981 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neon-2x1.c223 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
226 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
227 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
512 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
515 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
516 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neonfma-2x1.c225 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
228 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
229 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
512 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
515 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vi0x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
516 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vi2x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neonfma-2x1.c225 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
228 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
229 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
512 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
515 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
516 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x1.c223 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
226 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
227 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
512 const float32x4_t vk01c2x0123 = vld1q_f32(w + 128); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
515 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk01c2x0123, vi0x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
516 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vi2x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
/external/XNNPACK/src/f32-conv-hwc2chw/
D3x3s2p1c3x4-sse-1x1.c141 const __m128 vk01c2x0123 = _mm_load_ps(w + 64); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
143 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk01c2x0123, vi01c2)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
268 const __m128 vk01c2x0123 = _mm_load_ps(w + 64); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
270 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk01c2x0123, vi01c2)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()