| /external/XNNPACK/src/f32-dwconv-spchw/ |
| D | 3x3p1-scalar.c | 54 float vi0x1 = *i0; i0 = (const float*) ((uintptr_t) i0 + input_tuple_stride); in xnn_f32_dwconv_spchw_ukernel_3x3p1__scalar() local
64 const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1 + vw3 * vi0x2; in xnn_f32_dwconv_spchw_ukernel_3x3p1__scalar()
65 vi0x0 = vi0x1; in xnn_f32_dwconv_spchw_ukernel_3x3p1__scalar()
66 vi0x1 = vi0x2; in xnn_f32_dwconv_spchw_ukernel_3x3p1__scalar()
84 const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1; in xnn_f32_dwconv_spchw_ukernel_3x3p1__scalar()
|
| D | 5x5p2-scalar.c | 76 float vi0x1 = 0.0f; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar() local
108 …const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1 + vw3 * vi0x2 + vw4 * vi0x3 + vw5 * vi0x4; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
109 vi0x0 = vi0x1; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
110 vi0x1 = vi0x2; 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()
143 vi0x0 = vi0x1; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
144 vi0x1 = vi0x2; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
173 const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1 + vw3 * vi0x2; in xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar()
|
| D | 3x3s2p1-scalar.c | 57 const float vi0x1 = *i0; i0 = (const float*) ((uintptr_t) i0 + input_tuple_stride); in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__scalar() local
64 const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1 + vw3 * vi0x2; in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__scalar()
81 const float vi0x1 = i0[0]; in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__scalar() local
84 const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1; in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__scalar()
|
| D | 5x5s2p2-scalar.c | 76 float vi0x1 = 0.0f; 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()
137 const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1 + vw3 * vi0x2 + vw4 * vi0x3; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
151 const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1 + vw3 * vi0x2; in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar()
|
| /external/XNNPACK/src/f32-prelu/gen/ |
| D | scalar-2x4.c | 56 const float vi0x1 = i0[1]; in xnn_f32_prelu_ukernel__scalar_2x4() local
67 float vacc0x1 = signbit(vi0x1) ? vi0x1 * vw1 : vi0x1; in xnn_f32_prelu_ukernel__scalar_2x4()
|
| D | wasm-2x4.c | 56 const float vi0x1 = i0[1]; in xnn_f32_prelu_ukernel__wasm_2x4() local
67 float vacc0x1 = signbit(vi0x1) ? vi0x1 * vw1 : vi0x1; in xnn_f32_prelu_ukernel__wasm_2x4()
|
| /external/XNNPACK/src/f32-conv-hwc/ |
| D | 3x3s2p1c3x8-neonfma-2x2.c | 98 const float32x4_t vi0x1 = vld1q_f32(i0); i0 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
109 vo0x1c0123 = vfmaq_laneq_f32(vo0x1c0123, vk00c0x0123, vi0x1, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
111 vo0x1c4567 = vfmaq_laneq_f32(vo0x1c4567, vk00c0x4567, vi0x1, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
228 vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk01c0x0123, vi0x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
230 vo0x0c4567 = vfmaq_laneq_f32(vo0x0c4567, vk01c0x4567, vi0x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
267 vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk01c1x0123, vi0x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
269 vo0x0c4567 = vfmaq_laneq_f32(vo0x0c4567, vk01c1x4567, vi0x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
313 vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk01c2x0123, vi0x1, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
315 vo0x0c4567 = vfmaq_laneq_f32(vo0x0c4567, vk01c2x4567, vi0x1, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
352 vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk02c0x0123, vi0x1, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
|
| D | 3x3s2p1c3x4-neonfma-2x2.c | 93 const float32x4_t vi0x1 = vld1q_f32(i0); i0 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
101 vo0x1 = vfmaq_laneq_f32(vo0x1, vk00c0, vi0x1, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
169 vo0x0 = vfmaq_laneq_f32(vo0x0, vk01c0, vi0x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
190 vo0x0 = vfmaq_laneq_f32(vo0x0, vk01c1, vi0x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
218 vo0x0 = vfmaq_laneq_f32(vo0x0, vk01c2, vi0x1, 2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
239 vo0x0 = vfmaq_laneq_f32(vo0x0, vk02c0, vi0x1, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
365 float32x4_t vi0x1 = vld1q_f32(i0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
374 vo0x1 = vfmaq_laneq_f32(vo0x1, vk00c0, vi0x1, 3); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
454 vo0x0 = vfmaq_laneq_f32(vo0x0, vk01c0, vi0x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
481 vo0x0 = vfmaq_laneq_f32(vo0x0, vk01c1, vi0x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
|
| /external/XNNPACK/src/f32-conv-hwc2spchw/ |
| D | 3x3s2p1c3x4-neonfma-2x2.c | 112 const float32x4_t vi0x1 = vld1q_f32(i0); i0 += 4; in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
120 vo0x1 = vfmaq_laneq_f32(vo0x1, vk00c0, vi0x1, 3); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
188 vo0x0 = vfmaq_laneq_f32(vo0x0, vk01c0, vi0x1, 0); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
209 vo0x0 = vfmaq_laneq_f32(vo0x0, vk01c1, vi0x1, 1); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
237 vo0x0 = vfmaq_laneq_f32(vo0x0, vk01c2, vi0x1, 2); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
258 vo0x0 = vfmaq_laneq_f32(vo0x0, vk02c0, vi0x1, 3); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
361 float32x4_t vi0x1 = vld1q_f32(i0); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
370 vo0x1 = vfmaq_laneq_f32(vo0x1, vk00c0, vi0x1, 3); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
450 vo0x0 = vfmaq_laneq_f32(vo0x0, vk01c0, vi0x1, 0); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
477 vo0x0 = vfmaq_laneq_f32(vo0x0, vk01c1, vi0x1, 1); in xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
|
| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up2x4-scalar.c | 50 const float vi0x1 = i0[1]; in xnn_f32_dwconv_ukernel_up2x4__scalar() local
56 vacc1p0 += vi0x1 * vk0x1; in xnn_f32_dwconv_ukernel_up2x4__scalar()
|
| D | up2x4-wasm.c | 50 const float vi0x1 = i0[1]; in xnn_f32_dwconv_ukernel_up2x4__wasm() local
56 vacc1p0 += vi0x1 * vk0x1; in xnn_f32_dwconv_ukernel_up2x4__wasm()
|
| D | up2x4-wasm-acc2.c | 50 const float vi0x1 = i0[1]; in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2() local
56 vacc1p0 += vi0x1 * vk0x1; in xnn_f32_dwconv_ukernel_up2x4__wasm_acc2()
|
| D | up2x4-scalar-acc2.c | 50 const float vi0x1 = i0[1]; in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2() local
56 vacc1p0 += vi0x1 * vk0x1; in xnn_f32_dwconv_ukernel_up2x4__scalar_acc2()
|
| D | up2x9-wasm.c | 60 const float vi0x1 = i0[1]; in xnn_f32_dwconv_ukernel_up2x9__wasm() local
66 vacc1p0 += vi0x1 * vk0x1; in xnn_f32_dwconv_ukernel_up2x9__wasm()
|
| D | up2x9-scalar.c | 60 const float vi0x1 = i0[1]; in xnn_f32_dwconv_ukernel_up2x9__scalar() local
66 vacc1p0 += vi0x1 * vk0x1; in xnn_f32_dwconv_ukernel_up2x9__scalar()
|
| D | up2x9-scalar-acc2.c | 60 const float vi0x1 = i0[1]; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2() local
66 vacc1p0 += vi0x1 * vk0x1; in xnn_f32_dwconv_ukernel_up2x9__scalar_acc2()
|
| D | up2x9-wasm-acc2.c | 60 const float vi0x1 = i0[1]; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2() local
66 vacc1p0 += vi0x1 * vk0x1; in xnn_f32_dwconv_ukernel_up2x9__wasm_acc2()
|
| D | up2x25-scalar.c | 92 const float vi0x1 = i0[1]; in xnn_f32_dwconv_ukernel_up2x25__scalar() local
98 vacc1p0 += vi0x1 * vk0x1; in xnn_f32_dwconv_ukernel_up2x25__scalar()
|
| D | up2x25-scalar-acc2.c | 92 const float vi0x1 = i0[1]; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2() local
98 vacc1p0 += vi0x1 * vk0x1; in xnn_f32_dwconv_ukernel_up2x25__scalar_acc2()
|
| D | up2x25-wasm-acc2.c | 92 const float vi0x1 = i0[1]; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2() local
98 vacc1p0 += vi0x1 * vk0x1; in xnn_f32_dwconv_ukernel_up2x25__wasm_acc2()
|
| D | up2x25-wasm.c | 92 const float vi0x1 = i0[1]; in xnn_f32_dwconv_ukernel_up2x25__wasm() local
98 vacc1p0 += vi0x1 * vk0x1; in xnn_f32_dwconv_ukernel_up2x25__wasm()
|