Home
last modified time | relevance | path

Searched refs:vk20c2x4567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x8-neonfma-2x2.c221 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
225 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
226 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk20c2x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
230 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c2x4567, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
231 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk20c2x4567, vget_low_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
628 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
632 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
633 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
914 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
918 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c219 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
223 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
224 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk20c2x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
228 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c2x4567, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
229 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk20c2x4567, vget_low_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
628 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
632 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
633 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
916 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
920 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c221 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
225 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
226 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk20c2x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
230 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c2x4567, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
231 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk20c2x4567, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
628 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
632 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
633 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
914 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
918 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c219 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
223 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
224 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk20c2x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
228 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c2x4567, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
229 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk20c2x4567, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
628 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
632 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
633 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
916 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
920 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x1.c163 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
167 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
168 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
450 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
454 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
455 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neon-2x1.c161 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
165 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
166 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
450 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
454 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
455 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neon-2x1.c161 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
165 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
166 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
450 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
454 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
455 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
D3x3s2p1c3x8-neonfma-2x1.c163 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
167 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
168 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
450 const float32x4_t vk20c2x4567 = vld1q_f32(w + 76); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
454 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
455 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()