| /external/XNNPACK/src/f32-dwconv-spchw/ |
| D | 3x3s2p1-scalar.c | 53 float vi2x0 = 0.0f; in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__scalar() local
68 const float vrow2_accum = vw7 * vi2x0 + vw8 * vi2x1 + vw9 * vi2x2; in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__scalar()
69 vi2x0 = vi2x2; in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__scalar()
86 const float vrow2_accum = vw7 * vi2x0 + vw8 * vi2x1; in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__scalar()
|
| D | 3x3p1-scalar.c | 53 float vi2x0 = 0.0f; in xnn_f32_dwconv_spchw_ukernel_3x3p1__scalar() local
70 const float vrow2_accum = vw7 * vi2x0 + vw8 * vi2x1 + vw9 * vi2x2; in xnn_f32_dwconv_spchw_ukernel_3x3p1__scalar()
71 vi2x0 = vi2x1; in xnn_f32_dwconv_spchw_ukernel_3x3p1__scalar()
86 const float vrow2_accum = vw7 * vi2x0 + vw8 * vi2x1; in xnn_f32_dwconv_spchw_ukernel_3x3p1__scalar()
|
| D | 5x5p2-scalar.c | 73 float vi2x0 = 0.0f; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar() local
118 …const float vrow2_accum = vw11 * vi2x0 + vw12 * vi2x1 + vw13 * vi2x2 + vw14 * vi2x3 + vw15 * vi2x4; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
119 vi2x0 = vi2x1; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
150 const float vrow2_accum = vw11 * vi2x0 + vw12 * vi2x1 + vw13 * vi2x2 + vw14 * vi2x3; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
151 vi2x0 = vi2x1; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
175 const float vrow2_accum = vw11 * vi2x0 + vw12 * vi2x1 + vw13 * vi2x2; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
|
| D | 5x5s2p2-scalar.c | 73 float vi2x0 = 0.0f; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar() local
110 …const float vrow2_accum = vw11 * vi2x0 + vw12 * vi2x1 + vw13 * vi2x2 + vw14 * vi2x3 + vw15 * vi2x4; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
111 vi2x0 = vi2x2; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
139 const float vrow2_accum = vw11 * vi2x0 + vw12 * vi2x1 + vw13 * vi2x2 + vw14 * vi2x3; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
153 const float vrow2_accum = vw11 * vi2x0 + vw12 * vi2x1 + vw13 * vi2x2; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
|
| /external/XNNPACK/src/f32-conv-hwc/ |
| D | 3x3s2p1c3x8-neonfma-2x2.c | 79 float32x4_t vi2x0 = vmovq_n_f32(0.0f); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
105 vo1x0c0123 = vfmaq_laneq_f32(vo1x0c0123, vk00c0x0123, vi2x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
107 vo1x0c4567 = vfmaq_laneq_f32(vo1x0c4567, vk00c0x4567, vi2x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
130 vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk20c0x0123, vi2x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
132 vo0x0c4567 = vfmaq_laneq_f32(vo0x0c4567, vk20c0x4567, vi2x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
151 vo1x0c0123 = vfmaq_laneq_f32(vo1x0c0123, vk00c1x0123, vi2x0, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
153 vo1x0c4567 = vfmaq_laneq_f32(vo1x0c4567, vk00c1x4567, vi2x0, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
176 vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk20c1x0123, vi2x0, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
178 vo0x0c4567 = vfmaq_laneq_f32(vo0x0c4567, vk20c1x4567, vi2x0, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
190 vo1x0c0123 = vfmaq_laneq_f32(vo1x0c0123, vk00c2x0123, vi2x0, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
|
| D | 3x3s2p1c3x4-neonfma-2x2.c | 79 float32x4_t vi2x0 = vmovq_n_f32(0.0f); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
100 vo1x0 = vfmaq_laneq_f32(vo1x0, vk00c0, vi2x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
113 vo0x0 = vfmaq_laneq_f32(vo0x0, vk20c0, vi2x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
128 vo1x0 = vfmaq_laneq_f32(vo1x0, vk00c1, vi2x0, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
141 vo0x0 = vfmaq_laneq_f32(vo0x0, vk20c1, vi2x0, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
149 vo1x0 = vfmaq_laneq_f32(vo1x0, vk00c2, vi2x0, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
162 vo0x0 = vfmaq_laneq_f32(vo0x0, vk20c2, vi2x0, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
302 vi2x0 = vi2x3; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
372 vo1x0 = vfmaq_laneq_f32(vo1x0, vk00c0, vi2x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
389 vo0x0 = vfmaq_laneq_f32(vo0x0, vk20c0, vi2x0, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
|
| /external/XNNPACK/src/f32-conv-hwc2spchw/ |
| D | 3x3s2p1c3x4-neonfma-2x2.c | 98 float32x4_t vi2x0 = vmovq_n_f32(0.0f); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
119 vo1x0 = vfmaq_laneq_f32(vo1x0, vk00c0, vi2x0, 1); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
132 vo0x0 = vfmaq_laneq_f32(vo0x0, vk20c0, vi2x0, 1); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
147 vo1x0 = vfmaq_laneq_f32(vo1x0, vk00c1, vi2x0, 2); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
160 vo0x0 = vfmaq_laneq_f32(vo0x0, vk20c1, vi2x0, 2); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
168 vo1x0 = vfmaq_laneq_f32(vo1x0, vk00c2, vi2x0, 3); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
181 vo0x0 = vfmaq_laneq_f32(vo0x0, vk20c2, vi2x0, 3); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
321 vi2x0 = vi2x3; in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
368 vo1x0 = vfmaq_laneq_f32(vo1x0, vk00c0, vi2x0, 1); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
385 vo0x0 = vfmaq_laneq_f32(vo0x0, vk20c0, vi2x0, 1); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
|
| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up2x4-scalar.c | 67 const float vi2x0 = i2[0]; in xnn_f32_dwconv_ukernel_up2x4__scalar() local
72 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x4__scalar()
|
| D | up2x4-wasm.c | 67 const float vi2x0 = i2[0]; in xnn_f32_dwconv_ukernel_up2x4__wasm() local
72 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x4__wasm()
|
| D | up2x4-wasm-acc2.c | 67 const float vi2x0 = i2[0]; in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2() local
72 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2()
|
| D | up2x4-scalar-acc2.c | 67 const float vi2x0 = i2[0]; in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2() local
72 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2()
|
| D | up2x9-wasm.c | 77 const float vi2x0 = i2[0]; in xnn_f32_dwconv_ukernel_up2x9__wasm() local
82 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x9__wasm()
|
| D | up2x9-scalar.c | 77 const float vi2x0 = i2[0]; in xnn_f32_dwconv_ukernel_up2x9__scalar() local
82 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x9__scalar()
|
| D | up2x9-scalar-acc2.c | 77 const float vi2x0 = i2[0]; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2() local
82 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2()
|
| D | up2x9-wasm-acc2.c | 77 const float vi2x0 = i2[0]; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2() local
82 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2()
|
| D | up2x25-scalar.c | 109 const float vi2x0 = i2[0]; in xnn_f32_dwconv_ukernel_up2x25__scalar() local
114 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x25__scalar()
|
| D | up2x25-scalar-acc2.c | 109 const float vi2x0 = i2[0]; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2() local
114 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2()
|
| D | up2x25-wasm-acc2.c | 109 const float vi2x0 = i2[0]; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2() local
114 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2()
|
| D | up2x25-wasm.c | 109 const float vi2x0 = i2[0]; in xnn_f32_dwconv_ukernel_up2x25__wasm() local
114 vacc0p0 += vi2x0 * vk2x0; in xnn_f32_dwconv_ukernel_up2x25__wasm()
|