/external/XNNPACK/src/f32-conv-hwc/gen/ |
D | 3x3s2p0p1c3x8-neon-2x2.c | 395 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 399 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk02c1x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 400 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk02c1x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 404 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk02c1x4567, vget_low_f32(vi0x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 405 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk02c1x4567, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 739 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 743 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c1x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 744 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c1x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
|
D | 3x3s2p1c3x8-neonfma-2x2.c | 397 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 401 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk02c1x4567, vget_low_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 402 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk02c1x4567, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 406 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk02c1x4567, vget_high_f32(vi0x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 407 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk02c1x4567, vget_high_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 746 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 750 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c1x4567, vi0x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 751 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk02c1x4567, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
|
D | 3x3s2p0p1c3x8-neonfma-2x2.c | 397 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 401 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk02c1x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 402 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk02c1x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 406 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk02c1x4567, vget_low_f32(vi0x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 407 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk02c1x4567, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 739 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 743 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c1x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 744 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk02c1x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
|
D | 3x3s2p1c3x8-neon-2x2.c | 395 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 399 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk02c1x4567, vget_low_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 400 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk02c1x4567, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 404 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk02c1x4567, vget_high_f32(vi0x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 405 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk02c1x4567, vget_high_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 746 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 750 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c1x4567, vi0x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 751 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c1x4567, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
|
D | 3x3s2p1c3x8-neon-2x1.c | 279 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local 283 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c1x4567, vi0x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 284 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c1x4567, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
|
D | 3x3s2p0p1c3x8-neonfma-2x1.c | 274 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local 278 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c1x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() 279 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk02c1x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
|
D | 3x3s2p1c3x8-neonfma-2x1.c | 281 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local 285 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c1x4567, vi0x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 286 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk02c1x4567, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
|
D | 3x3s2p0p1c3x8-neon-2x1.c | 272 const float32x4_t vk02c1x4567 = vld1q_f32(w + 180); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local 276 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c1x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() 277 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c1x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
|