Home
last modified time | relevance | path

Searched refs:vk10c2x4567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x8-neonfma-2x2.c208 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
212 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
213 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk10c2x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
217 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk10c2x4567, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
218 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk10c2x4567, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
620 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
624 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
625 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
906 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
910 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c206 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
210 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
211 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk10c2x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
215 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk10c2x4567, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
216 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk10c2x4567, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
620 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
624 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
625 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
908 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
912 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c208 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
212 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
213 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk10c2x4567, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
217 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk10c2x4567, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
218 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk10c2x4567, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
620 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
624 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
625 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
906 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
910 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c206 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
210 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
211 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk10c2x4567, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
215 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk10c2x4567, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
216 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk10c2x4567, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
620 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
624 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
625 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
908 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
912 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x1.c155 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
159 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
160 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
442 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
446 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
447 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neon-2x1.c153 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
157 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
158 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
442 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
446 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
447 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neon-2x1.c153 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
157 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
158 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
442 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
446 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
447 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
D3x3s2p1c3x8-neonfma-2x1.c155 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
159 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
160 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
442 const float32x4_t vk10c2x4567 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
446 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
447 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()