Home
last modified time | relevance | path

Searched refs:vk01c1x4567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p0p1c3x8-neon-2x2.c271 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
275 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk01c1x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
276 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk01c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
280 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk01c1x4567, vget_high_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
281 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk01c1x4567, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
660 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
671 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
672 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
948 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
959 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vi0x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neonfma-2x2.c273 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
277 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk01c1x4567, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
278 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk01c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
282 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk01c1x4567, vget_high_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
283 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk01c1x4567, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
667 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
671 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
672 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
953 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
957 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c273 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
277 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk01c1x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
278 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk01c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
282 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk01c1x4567, vget_high_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
283 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk01c1x4567, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
660 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
671 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
672 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
946 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
957 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c1x4567, vi0x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c271 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
275 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk01c1x4567, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
276 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk01c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
280 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk01c1x4567, vget_high_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
281 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk01c1x4567, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
667 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
671 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
672 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
955 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
959 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neon-2x1.c200 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
204 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
205 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
489 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
493 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
494 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neonfma-2x1.c195 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
206 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
207 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
482 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
493 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c1x4567, vi0x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
494 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c1x4567, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neonfma-2x1.c202 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
206 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
207 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
489 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
493 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
494 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x1.c193 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
204 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
205 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
482 const float32x4_t vk01c1x4567 = vld1q_f32(w + 108); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
493 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vi0x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
494 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c1x4567, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()