| /external/XNNPACK/src/f32-dwconv-spchw/ |
| D | 5x5p2-scalar.c | 87 float vi0x3; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar() local
93 vi0x3 = *i0; i0 = (const float*) ((uintptr_t) i0 + input_tuple_stride); in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
108 …const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1 + vw3 * vi0x2 + vw4 * vi0x3 + vw5 * vi0x4; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
111 vi0x2 = vi0x3; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
112 vi0x3 = vi0x4; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
142 const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1 + vw3 * vi0x2 + vw4 * vi0x3; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
145 vi0x2 = vi0x3; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
|
| D | 5x5s2p2-scalar.c | 90 const float vi0x3 = *i0; i0 = (const float*) ((uintptr_t) i0 + input_tuple_stride); in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar() local
102 …const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1 + vw3 * vi0x2 + vw4 * vi0x3 + vw5 * vi0x4; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
104 vi0x1 = vi0x3; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
131 const float vi0x3 = *i0; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar() local
137 const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1 + vw3 * vi0x2 + vw4 * vi0x3; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
|
| /external/XNNPACK/src/f32-prelu/gen/ |
| D | scalar-2x4.c | 58 const float vi0x3 = i0[3]; in xnn_f32_prelu_ukernel__scalar_2x4() local
69 float vacc0x3 = signbit(vi0x3) ? vi0x3 * vw3 : vi0x3; in xnn_f32_prelu_ukernel__scalar_2x4()
|
| D | wasm-2x4.c | 58 const float vi0x3 = i0[3]; in xnn_f32_prelu_ukernel__wasm_2x4() local
69 float vacc0x3 = signbit(vi0x3) ? vi0x3 * vw3 : vi0x3; in xnn_f32_prelu_ukernel__wasm_2x4()
|
| /external/XNNPACK/src/f32-conv-hwc2spchw/ |
| D | 3x3s2p1c3x4-neonfma-2x2.c | 231 const float32x4_t vi0x3 = vld1q_f32(i0); i0 += 4; in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
239 vo0x1 = vfmaq_laneq_f32(vo0x1, vk01c2, vi0x3, 0); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
260 vo0x1 = vfmaq_laneq_f32(vo0x1, vk02c0, vi0x3, 1); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
281 vo0x1 = vfmaq_laneq_f32(vo0x1, vk02c1, vi0x3, 2); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
302 vo0x1 = vfmaq_laneq_f32(vo0x1, vk02c2, vi0x3, 3); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
319 vi0x0 = vi0x3; in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
504 float32x4_t vi0x3 = vmovq_n_f32(0.0f); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
511 vi0x3 = vld1q_lane_f32(i0 + 8, vi0x3, 0); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
520 vo0x1 = vfmaq_laneq_f32(vo0x1, vk01c2, vi0x3, 0); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
|
| /external/XNNPACK/src/f32-conv-hwc/ |
| D | 3x3s2p1c3x4-neonfma-2x2.c | 212 const float32x4_t vi0x3 = vld1q_f32(i0); i0 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
220 vo0x1 = vfmaq_laneq_f32(vo0x1, vk01c2, vi0x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
241 vo0x1 = vfmaq_laneq_f32(vo0x1, vk02c0, vi0x3, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
262 vo0x1 = vfmaq_laneq_f32(vo0x1, vk02c1, vi0x3, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
283 vo0x1 = vfmaq_laneq_f32(vo0x1, vk02c2, vi0x3, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
300 vi0x0 = vi0x3; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
508 float32x4_t vi0x3 = vmovq_n_f32(0.0f); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
515 vi0x3 = vld1q_lane_f32(i0 + 8, vi0x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
524 vo0x1 = vfmaq_laneq_f32(vo0x1, vk01c2, vi0x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
|
| D | 3x3s2p1c3x8-neonfma-2x2.c | 307 const float32x4_t vi0x3 = vld1q_f32(i0); i0 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
318 vo0x1c0123 = vfmaq_laneq_f32(vo0x1c0123, vk01c2x0123, vi0x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
320 vo0x1c4567 = vfmaq_laneq_f32(vo0x1c4567, vk01c2x4567, vi0x3, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
357 vo0x1c0123 = vfmaq_laneq_f32(vo0x1c0123, vk02c0x0123, vi0x3, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
359 vo0x1c4567 = vfmaq_laneq_f32(vo0x1c4567, vk02c0x4567, vi0x3, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
396 vo0x1c0123 = vfmaq_laneq_f32(vo0x1c0123, vk02c1x0123, vi0x3, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
398 vo0x1c4567 = vfmaq_laneq_f32(vo0x1c4567, vk02c1x4567, vi0x3, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
435 vo0x1c0123 = vfmaq_laneq_f32(vo0x1c0123, vk02c2x0123, vi0x3, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
437 vo0x1c4567 = vfmaq_laneq_f32(vo0x1c4567, vk02c2x4567, vi0x3, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
466 vi0x0 = vi0x3; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
|