Home
last modified time | relevance | path

Searched refs:vk20c0x4567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x8-neonfma-2x2.c136 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
140 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
141 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk20c0x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
145 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
146 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk20c0x4567, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
580 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
584 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
585 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
866 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
870 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c134 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
138 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
139 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk20c0x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
143 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
144 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk20c0x4567, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
580 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
584 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
585 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
868 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
872 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c136 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
140 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
141 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk20c0x4567, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
145 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
146 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk20c0x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
580 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
584 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
585 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
866 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
870 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c134 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
138 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
139 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk20c0x4567, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
143 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
144 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk20c0x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
580 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
584 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
585 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
868 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
872 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x1.c115 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
119 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
120 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
402 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
406 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
407 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neon-2x1.c113 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
117 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
118 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
402 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
406 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
407 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neon-2x1.c113 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
117 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
118 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
402 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
406 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
407 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
D3x3s2p1c3x8-neonfma-2x1.c115 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
119 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
120 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
402 const float32x4_t vk20c0x4567 = vld1q_f32(w + 28); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
406 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
407 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()