Home
last modified time | relevance | path

Searched refs:vi3x3 (Results 1 – 5 of 5) sorted by relevance

/external/XNNPACK/src/f32-dwconv-spchw/
D5x5p2-scalar.c90 float vi3x3; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar() local
96 vi3x3 = *i3; i3 = (const float*) ((uintptr_t) i3 + input_tuple_stride); in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
123 …const float vrow3_accum = vw16 * vi3x0 + vw17 * vi3x1 + vw18 * vi3x2 + vw19 * vi3x3 + vw20 * vi3x4; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
126 vi3x2 = vi3x3; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
127 vi3x3 = vi3x4; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
154 const float vrow3_accum = vw16 * vi3x0 + vw17 * vi3x1 + vw18 * vi3x2 + vw19 * vi3x3; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
157 vi3x2 = vi3x3; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
D5x5s2p2-scalar.c93 const float vi3x3 = *i3; i3 = (const float*) ((uintptr_t) i3 + input_tuple_stride); in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar() local
114 …const float vrow3_accum = vw16 * vi3x0 + vw17 * vi3x1 + vw18 * vi3x2 + vw19 * vi3x3 + vw20 * vi3x4; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
116 vi3x1 = vi3x3; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
134 const float vi3x3 = *i3; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar() local
140 const float vrow3_accum = vw16 * vi3x0 + vw17 * vi3x1 + vw18 * vi3x2 + vw19 * vi3x3; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
/external/XNNPACK/src/f32-conv-hwc2spchw/
D3x3s2p1c3x4-neonfma-2x2.c234 const float32x4_t vi3x3 = vld1q_f32(i3); i3 += 4; in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
247 vo1x1 = vfmaq_laneq_f32(vo1x1, vk11c2, vi3x3, 0); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
268 vo1x1 = vfmaq_laneq_f32(vo1x1, vk12c0, vi3x3, 1); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
289 vo1x1 = vfmaq_laneq_f32(vo1x1, vk12c1, vi3x3, 2); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
310 vo1x1 = vfmaq_laneq_f32(vo1x1, vk12c2, vi3x3, 3); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
322 vi3x0 = vi3x3; in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
507 float32x4_t vi3x3 = vmovq_n_f32(0.0f); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
514 vi3x3 = vld1q_lane_f32(i3 + 8, vi3x3, 0); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
528 vo1x1 = vfmaq_laneq_f32(vo1x1, vk11c2, vi3x3, 0); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
/external/XNNPACK/src/f32-conv-hwc/
D3x3s2p1c3x4-neonfma-2x2.c215 const float32x4_t vi3x3 = vld1q_f32(i3); i3 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
228 vo1x1 = vfmaq_laneq_f32(vo1x1, vk11c2, vi3x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
249 vo1x1 = vfmaq_laneq_f32(vo1x1, vk12c0, vi3x3, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
270 vo1x1 = vfmaq_laneq_f32(vo1x1, vk12c1, vi3x3, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
291 vo1x1 = vfmaq_laneq_f32(vo1x1, vk12c2, vi3x3, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
303 vi3x0 = vi3x3; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
511 float32x4_t vi3x3 = vmovq_n_f32(0.0f); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
518 vi3x3 = vld1q_lane_f32(i3 + 8, vi3x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
532 vo1x1 = vfmaq_laneq_f32(vo1x1, vk11c2, vi3x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
D3x3s2p1c3x8-neonfma-2x2.c310 const float32x4_t vi3x3 = vld1q_f32(i3); i3 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
332 vo1x1c0123 = vfmaq_laneq_f32(vo1x1c0123, vk11c2x0123, vi3x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
334 vo1x1c4567 = vfmaq_laneq_f32(vo1x1c4567, vk11c2x4567, vi3x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
371 vo1x1c0123 = vfmaq_laneq_f32(vo1x1c0123, vk12c0x0123, vi3x3, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
373 vo1x1c4567 = vfmaq_laneq_f32(vo1x1c4567, vk12c0x4567, vi3x3, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
410 vo1x1c0123 = vfmaq_laneq_f32(vo1x1c0123, vk12c1x0123, vi3x3, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
412 vo1x1c4567 = vfmaq_laneq_f32(vo1x1c4567, vk12c1x4567, vi3x3, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
449 vo1x1c0123 = vfmaq_laneq_f32(vo1x1c0123, vk12c2x0123, vi3x3, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
451 vo1x1c4567 = vfmaq_laneq_f32(vo1x1c4567, vk12c2x4567, vi3x3, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
469 vi3x0 = vi3x3; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]