| /external/XNNPACK/src/f32-dwconv-spchw/ |
| D | 5x5p2-scalar.c | 75 float vi4x0 = 0.0f; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar() local
128 …const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2 + vw24 * vi4x3 + vw25 * vi4x4; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
129 vi4x0 = vi4x1; 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()
159 vi4x0 = vi4x1; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
177 const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
|
| D | 5x5s2p2-scalar.c | 75 float vi4x0 = 0.0f; 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()
119 vi4x0 = vi4x2; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
141 const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2 + vw24 * vi4x3; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
155 const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
|
| /external/XNNPACK/src/f32-conv-hwc/ |
| D | 3x3s2p1c3x8-neonfma-2x2.c | 81 float32x4_t vi4x0 = vmovq_n_f32(0.0f); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
131 vo1x0c0123 = vfmaq_laneq_f32(vo1x0c0123, vk20c0x0123, vi4x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
133 vo1x0c4567 = vfmaq_laneq_f32(vo1x0c4567, vk20c0x4567, vi4x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
177 vo1x0c0123 = vfmaq_laneq_f32(vo1x0c0123, vk20c1x0123, vi4x0, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
179 vo1x0c4567 = vfmaq_laneq_f32(vo1x0c4567, vk20c1x4567, vi4x0, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
216 vo1x0c0123 = vfmaq_laneq_f32(vo1x0c0123, vk20c2x0123, vi4x0, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
218 vo1x0c4567 = vfmaq_laneq_f32(vo1x0c4567, vk20c2x4567, vi4x0, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
470 vi4x0 = vi4x3; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
598 vo1x0c0123 = vfmaq_laneq_f32(vo1x0c0123, vk20c0x0123, vi4x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
600 vo1x0c4567 = vfmaq_laneq_f32(vo1x0c4567, vk20c0x4567, vi4x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
|
| D | 3x3s2p1c3x4-neonfma-2x2.c | 81 float32x4_t vi4x0 = vmovq_n_f32(0.0f); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
114 vo1x0 = vfmaq_laneq_f32(vo1x0, vk20c0, vi4x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
142 vo1x0 = vfmaq_laneq_f32(vo1x0, vk20c1, vi4x0, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
163 vo1x0 = vfmaq_laneq_f32(vo1x0, vk20c2, vi4x0, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
304 vi4x0 = vi4x3; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
390 vo1x0 = vfmaq_laneq_f32(vo1x0, vk20c0, vi4x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
427 vo1x0 = vfmaq_laneq_f32(vo1x0, vk20c1, vi4x0, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
448 vo1x0 = vfmaq_laneq_f32(vo1x0, vk20c2, vi4x0, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
|
| /external/XNNPACK/src/f32-conv-hwc2spchw/ |
| D | 3x3s2p1c3x4-neonfma-2x2.c | 100 float32x4_t vi4x0 = vmovq_n_f32(0.0f); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
133 vo1x0 = vfmaq_laneq_f32(vo1x0, vk20c0, vi4x0, 1); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
161 vo1x0 = vfmaq_laneq_f32(vo1x0, vk20c1, vi4x0, 2); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
182 vo1x0 = vfmaq_laneq_f32(vo1x0, vk20c2, vi4x0, 3); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
323 vi4x0 = vi4x3; in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
386 vo1x0 = vfmaq_laneq_f32(vo1x0, vk20c0, vi4x0, 1); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
423 vo1x0 = vfmaq_laneq_f32(vo1x0, vk20c1, vi4x0, 2); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
444 vo1x0 = vfmaq_laneq_f32(vo1x0, vk20c2, vi4x0, 3); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
|
| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up2x9-wasm.c | 95 const float vi4x0 = i4[0]; in xnn_f32_dwconv_ukernel_up2x9__wasm() local
100 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x9__wasm()
|
| D | up2x9-scalar.c | 95 const float vi4x0 = i4[0]; in xnn_f32_dwconv_ukernel_up2x9__scalar() local
100 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x9__scalar()
|
| D | up2x9-scalar-acc2.c | 95 const float vi4x0 = i4[0]; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2() local
100 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2()
|
| D | up2x9-wasm-acc2.c | 95 const float vi4x0 = i4[0]; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2() local
100 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2()
|
| D | up2x25-scalar.c | 127 const float vi4x0 = i4[0]; in xnn_f32_dwconv_ukernel_up2x25__scalar() local
132 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x25__scalar()
|
| D | up2x25-scalar-acc2.c | 127 const float vi4x0 = i4[0]; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2() local
132 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2()
|
| D | up2x25-wasm-acc2.c | 127 const float vi4x0 = i4[0]; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2() local
132 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2()
|
| D | up2x25-wasm.c | 127 const float vi4x0 = i4[0]; in xnn_f32_dwconv_ukernel_up2x25__wasm() local
132 vacc0p0 += vi4x0 * vk4x0; in xnn_f32_dwconv_ukernel_up2x25__wasm()
|