Home
last modified time | relevance | path

Searched refs:vk21c0x4567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x8-neonfma-2x2.c260 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
264 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
265 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk21c0x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
269 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk21c0x4567, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
270 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk21c0x4567, vget_high_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
659 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
663 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
664 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c0x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
945 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
949 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c258 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
262 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
263 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk21c0x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
267 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk21c0x4567, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
268 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk21c0x4567, vget_high_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
659 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
663 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
664 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c0x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
947 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
951 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c260 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
264 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
265 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk21c0x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
269 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk21c0x4567, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
270 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk21c0x4567, vget_low_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
652 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
656 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
657 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c0x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
938 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
942 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c258 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
262 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
263 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk21c0x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
267 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk21c0x4567, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
268 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk21c0x4567, vget_low_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
652 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
656 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
657 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c0x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
940 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
944 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x1.c187 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
191 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
192 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c0x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
474 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
478 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
479 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c0x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neon-2x1.c192 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
196 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
197 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c0x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
481 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
485 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
486 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c0x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neon-2x1.c185 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
189 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
190 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c0x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
474 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
478 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
479 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c0x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
D3x3s2p1c3x8-neonfma-2x1.c194 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
198 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
199 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c0x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
481 const float32x4_t vk21c0x4567 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
485 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
486 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c0x4567, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()