Home
last modified time | relevance | path

Searched refs:vk21c1x4567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p0p1c3x8-neon-2x2.c297 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
301 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
302 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk21c1x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
306 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk21c1x4567, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
307 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk21c1x4567, vget_high_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
683 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
687 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
688 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c1x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
971 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
975 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neonfma-2x2.c299 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
303 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
304 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk21c1x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
308 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk21c1x4567, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
309 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk21c1x4567, vget_high_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
683 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
687 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
688 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c1x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
969 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
973 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c299 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
303 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
304 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk21c1x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
308 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk21c1x4567, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
309 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk21c1x4567, vget_high_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
683 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
687 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
688 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c1x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
969 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
973 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c297 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
301 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
302 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk21c1x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
306 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk21c1x4567, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
307 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk21c1x4567, vget_high_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
683 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
687 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
688 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c1x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
971 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
975 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neon-2x1.c216 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
220 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
221 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c1x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
505 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
509 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
510 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c1x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neonfma-2x1.c218 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
222 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
223 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c1x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
505 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
509 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
510 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c1x4567, vi4x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neonfma-2x1.c218 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
222 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
223 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c1x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
505 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
509 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
510 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c1x4567, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x1.c216 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
220 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
221 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c1x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
505 const float32x4_t vk21c1x4567 = vld1q_f32(w + 124); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
509 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
510 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c1x4567, vi4x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()