Home
last modified time | relevance | path

Searched refs:vk11c2x4567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p0p1c3x8-neon-2x2.c323 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
327 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk11c2x4567, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
328 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk11c2x4567, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
332 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk11c2x4567, vget_high_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
333 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk11c2x4567, vget_high_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
699 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
703 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
704 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c2x4567, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
987 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
991 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vi1x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neonfma-2x2.c332 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
336 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk11c2x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
337 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk11c2x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
341 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk11c2x4567, vget_low_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
342 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk11c2x4567, vget_low_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
699 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
703 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c2x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
704 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk11c2x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
985 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
989 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c2x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c325 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
329 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk11c2x4567, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
330 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk11c2x4567, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
334 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk11c2x4567, vget_high_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
335 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk11c2x4567, vget_high_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
699 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
703 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c2x4567, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
704 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk11c2x4567, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
985 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
989 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c2x4567, vi1x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c330 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
334 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk11c2x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
335 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk11c2x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
339 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk11c2x4567, vget_low_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
340 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk11c2x4567, vget_low_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
699 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
703 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
704 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c2x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
987 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
991 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neon-2x1.c232 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
236 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
237 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c2x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
521 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
525 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
526 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c2x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neonfma-2x1.c234 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
238 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c2x4567, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
239 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk11c2x4567, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
521 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
525 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c2x4567, vi1x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
526 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk11c2x4567, vi3x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neonfma-2x1.c234 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
238 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c2x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
239 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk11c2x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
521 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
525 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c2x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
526 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk11c2x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x1.c232 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
236 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
237 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c2x4567, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
521 const float32x4_t vk11c2x4567 = vld1q_f32(w + 140); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
525 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vi1x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
526 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c2x4567, vi3x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()