Home
last modified time | relevance | path

Searched refs:vk21c2x4567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x8-neonfma-2x2.c345 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
349 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
350 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk21c2x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
354 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk21c2x4567, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
355 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk21c2x4567, vget_low_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
707 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
711 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
712 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c2x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
993 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
997 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c343 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
347 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
348 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk21c2x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
352 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk21c2x4567, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
353 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk21c2x4567, vget_low_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
707 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
711 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
712 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c2x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
995 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
999 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c338 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
342 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
343 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk21c2x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
347 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk21c2x4567, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
348 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk21c2x4567, vget_high_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
707 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
711 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
712 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c2x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
993 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
997 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vi2x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c336 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
340 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
341 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk21c2x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
345 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk21c2x4567, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
346 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk21c2x4567, vget_high_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
707 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
711 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
712 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c2x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
995 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
999 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vi2x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x1.c242 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
246 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
247 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c2x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
529 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
533 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vi2x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
534 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c2x4567, vi4x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neon-2x1.c240 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
244 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
245 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c2x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
529 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
533 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
534 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c2x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neon-2x1.c240 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
244 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
245 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c2x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
529 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
533 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vi2x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
534 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c2x4567, vi4x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
D3x3s2p1c3x8-neonfma-2x1.c242 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
246 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
247 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c2x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
529 const float32x4_t vk21c2x4567 = vld1q_f32(w + 148); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
533 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
534 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c2x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()