Home
last modified time | relevance | path

Searched refs:vk20c1x4567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p0p1c3x8-neon-2x2.c173 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
177 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
178 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk20c1x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
182 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c1x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
183 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk20c1x4567, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
604 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
608 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
609 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c1x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
892 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
896 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neonfma-2x2.c182 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
186 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
187 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk20c1x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
191 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c1x4567, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
192 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk20c1x4567, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
604 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
608 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
609 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c1x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
890 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
894 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c175 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
179 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
180 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk20c1x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
184 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c1x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
185 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk20c1x4567, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
604 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
608 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
609 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c1x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
890 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
894 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c180 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
184 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
185 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk20c1x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
189 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c1x4567, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
190 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk20c1x4567, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
604 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
608 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
609 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c1x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
892 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
896 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neon-2x1.c137 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
141 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
142 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c1x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
426 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
430 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
431 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c1x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neonfma-2x1.c139 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
143 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
144 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c1x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
426 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
430 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
431 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c1x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neonfma-2x1.c139 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
143 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
144 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c1x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
426 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
430 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
431 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c1x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x1.c137 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
141 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
142 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c1x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
426 const float32x4_t vk20c1x4567 = vld1q_f32(w + 52); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
430 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
431 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c1x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()