| /external/XNNPACK/src/f32-prelu/gen/ |
| D | wasm-2x4.c | 55 float vi0x3 = i0[3]; in xnn_f32_prelu_ukernel__wasm_2x4() local
69 float vacc0x3 = __builtin_wasm_max_f32(vi0x3, vzero); in xnn_f32_prelu_ukernel__wasm_2x4()
70 vi0x3 = __builtin_wasm_min_f32(vi0x3, vzero); in xnn_f32_prelu_ukernel__wasm_2x4()
83 vacc0x3 += vi0x3 * vw3; in xnn_f32_prelu_ukernel__wasm_2x4()
|
| D | scalar-2x4.c | 54 const float vi0x3 = i0[3]; in xnn_f32_prelu_ukernel__scalar_2x4() local
65 const float vacc0x3 = XNN_UNPREDICTABLE(vi0x3 < 0.0f) ? vi0x3 * vw3 : vi0x3; in xnn_f32_prelu_ukernel__scalar_2x4()
|
| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-scalar-1x1-acc4.c | 98 float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local
141 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
147 vo0p3 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
153 vi0x3 = vi0x4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
207 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
213 vo0p3 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
|
| D | 5x5p2-minmax-scalar-1x1-acc3.c | 98 float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3() local
141 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
147 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
153 vi0x3 = vi0x4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
206 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
212 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
|
| D | 5x5p2-minmax-scalar-1x1.c | 98 float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1() local
141 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
147 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
153 vi0x3 = vi0x4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
204 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
210 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
|
| D | 5x5p2-minmax-scalar-1x1-acc2.c | 98 float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2() local
141 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
147 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
153 vi0x3 = vi0x4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
205 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
211 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
|
| D | 5x5p2-minmax-scalar-1x1-acc5.c | 98 float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local
141 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
147 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
153 vi0x3 = vi0x4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
208 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
214 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc2.c | 106 const float vi0x3 = i0[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2() local
141 vi0x1 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
159 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
180 const float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2() local
204 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc4.c | 106 const float vi0x3 = i0[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local
141 vi0x1 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
159 vo0p3 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
182 const float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local
206 vo0p3 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
|
| D | 5x5s2p2-minmax-scalar-1x1.c | 106 const float vi0x3 = i0[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1() local
141 vi0x1 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
159 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
179 const float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1() local
203 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc3.c | 106 const float vi0x3 = i0[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3() local
141 vi0x1 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3()
159 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3()
181 const float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3() local
205 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc5.c | 106 const float vi0x3 = i0[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local
141 vi0x1 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5()
159 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5()
183 const float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local
207 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5()
|
| D | 5x5p2-minmax-scalar-2x1.c | 107 float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
169 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
176 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
187 vi0x3 = vi0x4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
264 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
271 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
|
| D | 5x5p2-minmax-scalar-2x1-acc3.c | 107 float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
169 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
176 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
187 vi0x3 = vi0x4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
268 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
275 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
|
| D | 5x5p2-minmax-scalar-2x1-acc2.c | 107 float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
169 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
176 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
187 vi0x3 = vi0x4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
266 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
273 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
|
| D | 5x5p2-minmax-scalar-3x1.c | 116 float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
197 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
205 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
221 vi0x3 = vi0x4; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
324 vi0x2 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
332 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
|
| D | 5x5s2p2-minmax-scalar-2x1-acc3.c | 123 const float vi0x3 = i0[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
176 vi0x1 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
203 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
240 const float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
281 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
|
| D | 5x5s2p2-minmax-scalar-2x1.c | 123 const float vi0x3 = i0[0]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1() local
176 vi0x1 = vi0x3; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
203 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
236 const float vi0x3 = *i0++; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1() local
277 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
|
| /external/XNNPACK/src/f32-conv-hwc2chw/ |
| D | 3x3s2p1c3x4-wasmsimd-2x2.c | 231 const v128_t vi0x3 = wasm_v128_load(i0); i0 += 4; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2() local
239 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk01c2, wasm_v32x4_shuffle(vi0x3, vi0x3, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
260 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk02c0, wasm_v32x4_shuffle(vi0x3, vi0x3, 1, 1, 1, 1))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
281 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk02c1, wasm_v32x4_shuffle(vi0x3, vi0x3, 2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
302 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk02c2, wasm_v32x4_shuffle(vi0x3, vi0x3, 3, 3, 3, 3))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
319 vi0x0 = vi0x3; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
503 v128_t vi0x3 = wasm_f32x4_const_splat(0.0f); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2() local
510 vi0x3 = wasm_v128_load32_splat(i0 + 8); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
519 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk01c2, wasm_v32x4_shuffle(vi0x3, vi0x3, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
|
| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7p7x-minmax-fp32-scalar-fmagic-c4.c | 44 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_fmagic_c4() local
53 int32_t vacc3 = vi0x3 + vinit_bias; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_fmagic_c4()
133 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_fmagic_c4() local
142 vacc3 += vi0x3; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_fmagic_c4()
243 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_fmagic_c4() local
253 vacc3 += vi0x3; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_fmagic_c4()
|
| D | 7p7x-minmax-fp32-scalar-lrintf-c4.c | 45 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c4() local
54 int32_t vacc3 = vi0x3 + vinit_bias; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c4()
134 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c4() local
143 vacc3 += vi0x3; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c4()
243 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c4() local
253 vacc3 += vi0x3; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c4()
|
| D | 7p7x-minmax-fp32-scalar-imagic-c4.c | 44 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4() local
53 int32_t vacc3 = vi0x3 + vinit_bias; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4()
133 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4() local
142 vacc3 += vi0x3; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4()
243 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4() local
253 vacc3 += vi0x3; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4()
|
| /external/XNNPACK/src/qu8-gavgpool/gen/ |
| D | 7p7x-minmax-fp32-scalar-imagic-c4.c | 44 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4() local
53 int32_t vacc3 = vi0x3 + vinit_bias; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4()
133 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4() local
142 vacc3 += vi0x3; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4()
243 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4() local
253 vacc3 += vi0x3; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4()
|
| D | 7p7x-minmax-fp32-scalar-fmagic-c4.c | 44 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_fmagic_c4() local
53 int32_t vacc3 = vi0x3 + vinit_bias; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_fmagic_c4()
133 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_fmagic_c4() local
142 vacc3 += vi0x3; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_fmagic_c4()
243 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_fmagic_c4() local
253 vacc3 += vi0x3; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_fmagic_c4()
|
| D | 7p7x-minmax-fp32-scalar-lrintf-c4.c | 45 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c4() local
54 int32_t vacc3 = vi0x3 + vinit_bias; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c4()
134 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c4() local
143 vacc3 += vi0x3; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c4()
243 const int32_t vi0x3 = (int32_t) i0[3]; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c4() local
253 vacc3 += vi0x3; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c4()
|