Home
last modified time | relevance | path

Searched refs:vi4x3 (Results 1 – 5 of 5) sorted by relevance

/external/XNNPACK/src/f32-dwconv-spchw/
D5x5p2-scalar.c91 float vi4x3; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar() local
97 vi4x3 = *i4; i4 = (const float*) ((uintptr_t) i4 + input_tuple_stride); in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
128 …const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2 + vw24 * vi4x3 + vw25 * vi4x4; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
131 vi4x2 = vi4x3; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
132 vi4x3 = vi4x4; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
158 const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2 + vw24 * vi4x3; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
161 vi4x2 = vi4x3; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
D5x5s2p2-scalar.c94 const float vi4x3 = *i4; i4 = (const float*) ((uintptr_t) i4 + input_tuple_stride); in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar() local
118 …const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2 + vw24 * vi4x3 + vw25 * vi4x4; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
120 vi4x1 = vi4x3; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
135 const float vi4x3 = *i4; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar() local
141 const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2 + vw24 * vi4x3; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
/external/XNNPACK/src/f32-conv-hwc2spchw/
D3x3s2p1c3x4-neonfma-2x2.c235 const float32x4_t vi4x3 = vld1q_f32(i4); i4 += 4; in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
254 vo1x1 = vfmaq_laneq_f32(vo1x1, vk21c2, vi4x3, 0); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
275 vo1x1 = vfmaq_laneq_f32(vo1x1, vk22c0, vi4x3, 1); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
296 vo1x1 = vfmaq_laneq_f32(vo1x1, vk22c1, vi4x3, 2); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
317 vo1x1 = vfmaq_laneq_f32(vo1x1, vk22c2, vi4x3, 3); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
323 vi4x0 = vi4x3; in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
508 float32x4_t vi4x3 = vmovq_n_f32(0.0f); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
515 vi4x3 = vld1q_lane_f32(i4 + 8, vi4x3, 0); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
535 vo1x1 = vfmaq_laneq_f32(vo1x1, vk21c2, vi4x3, 0); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
/external/XNNPACK/src/f32-conv-hwc/
D3x3s2p1c3x4-neonfma-2x2.c216 const float32x4_t vi4x3 = vld1q_f32(i4); i4 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
235 vo1x1 = vfmaq_laneq_f32(vo1x1, vk21c2, vi4x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
256 vo1x1 = vfmaq_laneq_f32(vo1x1, vk22c0, vi4x3, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
277 vo1x1 = vfmaq_laneq_f32(vo1x1, vk22c1, vi4x3, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
298 vo1x1 = vfmaq_laneq_f32(vo1x1, vk22c2, vi4x3, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
304 vi4x0 = vi4x3; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
512 float32x4_t vi4x3 = vmovq_n_f32(0.0f); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
519 vi4x3 = vld1q_lane_f32(i4 + 8, vi4x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
539 vo1x1 = vfmaq_laneq_f32(vo1x1, vk21c2, vi4x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
D3x3s2p1c3x8-neonfma-2x2.c311 const float32x4_t vi4x3 = vld1q_f32(i4); i4 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
345 vo1x1c0123 = vfmaq_laneq_f32(vo1x1c0123, vk21c2x0123, vi4x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
347 vo1x1c4567 = vfmaq_laneq_f32(vo1x1c4567, vk21c2x4567, vi4x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
384 vo1x1c0123 = vfmaq_laneq_f32(vo1x1c0123, vk22c0x0123, vi4x3, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
386 vo1x1c4567 = vfmaq_laneq_f32(vo1x1c4567, vk22c0x4567, vi4x3, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
423 vo1x1c0123 = vfmaq_laneq_f32(vo1x1c0123, vk22c1x0123, vi4x3, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
425 vo1x1c4567 = vfmaq_laneq_f32(vo1x1c4567, vk22c1x4567, vi4x3, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
462 vo1x1c0123 = vfmaq_laneq_f32(vo1x1c0123, vk22c2x0123, vi4x3, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
464 vo1x1c4567 = vfmaq_laneq_f32(vo1x1c4567, vk22c2x4567, vi4x3, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
470 vi4x0 = vi4x3; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]