Home
last modified time | relevance | path

Searched refs:vk11c0x0123 (Results 1 – 17 of 17) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x4-neonfma-2x2.c192 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
194 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
195 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
197 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c0x0123, vget_high_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
198 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c0x0123, vget_high_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
456 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
458 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
459 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
647 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
649 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
D3x3s2p1c3x4-neon-2x2.c190 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
192 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
193 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
195 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c0x0123, vget_high_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
196 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c0x0123, vget_high_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
456 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
458 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
459 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
649 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
651 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x2.c190 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
192 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
193 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
195 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c0x0123, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
196 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c0x0123, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
449 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
451 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
452 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
642 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
644 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neonfma-2x2.c192 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
194 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
195 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
197 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c0x0123, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
198 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c0x0123, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
449 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
451 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
452 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
640 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
642 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x1.c144 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
146 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
147 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
338 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
340 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
341 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
D3x3s2p0p1c3x4-neonfma-2x1.c146 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
148 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
149 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
338 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
340 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
341 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neonfma-2x1.c153 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
155 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
156 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
345 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
347 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
348 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neon-2x1.c151 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
153 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
154 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
345 const float32x4_t vk11c0x0123 = vld1q_f32(w + 44); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
347 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
348 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
D3x3s2p1c3x8-neonfma-2x2.c246 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
249 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
250 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
254 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c0x0123, vget_high_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
255 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c0x0123, vget_high_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
650 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
653 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
654 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
936 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
939 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c244 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
247 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
248 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
252 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c0x0123, vget_high_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
253 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c0x0123, vget_high_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
650 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
653 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
654 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
938 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
941 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c246 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
249 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
250 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
254 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c0x0123, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
255 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c0x0123, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
643 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
646 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
647 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
929 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
932 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c244 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
247 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
248 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
252 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c0x0123, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
253 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c0x0123, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
643 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
646 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
647 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
931 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
934 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x1.c178 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
181 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
182 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
465 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
468 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
469 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neon-2x1.c183 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
186 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
187 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
472 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
475 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
476 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neon-2x1.c176 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
179 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
180 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
465 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
468 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
469 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
D3x3s2p1c3x8-neonfma-2x1.c185 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
188 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
189 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
472 const float32x4_t vk11c0x0123 = vld1q_f32(w + 88); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
475 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
476 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
/external/XNNPACK/src/f32-conv-hwc2chw/
D3x3s2p1c3x4-sse-1x1.c121 const __m128 vk11c0x0123 = _mm_load_ps(w + 44); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
123 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk11c0x0123, vi11c0)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
248 const __m128 vk11c0x0123 = _mm_load_ps(w + 44); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
250 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk11c0x0123, vi11c0)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()