| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | neonfma-rr1-lut64-p2-div-x24.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24() local
162 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
163 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
164 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
165 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
166 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
167 float32x4_t vtKLMN = vfmaq_f32(vzKLMN, vnKLMN, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
293 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
372 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
|
| D | neonfma-rr1-lut64-p2-div-x16.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16() local
134 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
135 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
136 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
137 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
245 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
324 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
|
| D | neonfma-rr1-lut64-p2-div-x12.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x12() local
120 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x12()
121 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x12()
122 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x12()
221 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x12()
300 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x12()
|
| D | neonfma-rr1-lut64-p2-div-x20.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20() local
148 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
149 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
150 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
151 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
152 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
269 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
348 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
|
| D | neonfma-rr1-lut64-p2-div-x8.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x8() local
106 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x8()
107 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x8()
197 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x8()
276 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x8()
|
| D | neonfma-rr1-lut64-p2-nr1recps1fma-x24.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24() local
162 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
163 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
164 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
165 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
166 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
167 float32x4_t vtKLMN = vfmaq_f32(vzKLMN, vnKLMN, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
317 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
405 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
|
| D | neonfma-rr1-lut64-p2-nr2fma-x24.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24() local
162 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
163 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
164 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
165 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
166 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
167 float32x4_t vtKLMN = vfmaq_f32(vzKLMN, vnKLMN, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
317 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
405 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
|
| D | neonfma-rr1-lut64-p2-nr2recps-x24.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24() local
162 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
163 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
164 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
165 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
166 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
167 float32x4_t vtKLMN = vfmaq_f32(vzKLMN, vnKLMN, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
317 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
405 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
|
| D | neonfma-rr1-lut64-p2-nr1recps1fma-x20.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() local
148 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
149 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
150 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
151 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
152 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
290 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
378 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
|
| D | neonfma-rr1-lut64-p2-nr2fma-x20.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() local
148 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
149 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
150 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
151 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
152 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
290 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
378 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
|
| D | neonfma-rr1-lut64-p2-nr2recps-x20.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20() local
148 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20()
149 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20()
150 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20()
151 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20()
152 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20()
290 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20()
378 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20()
|
| D | neonfma-rr1-lut64-p2-nr2fma-x16.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() local
134 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16()
135 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16()
136 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16()
137 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16()
263 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16()
351 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16()
|
| D | neonfma-rr1-lut64-p2-nr1recps1fma-x16.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() local
134 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16()
135 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16()
136 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16()
137 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16()
263 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16()
351 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16()
|
| D | neonfma-rr1-lut64-p2-nr2recps-x16.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16() local
134 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16()
135 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16()
136 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16()
137 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16()
263 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16()
351 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16()
|
| D | neonfma-rr1-lut64-p2-nr1recps1fma-x8.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8() local
106 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8()
107 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8()
209 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8()
297 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8()
|
| D | neonfma-rr1-lut64-p2-nr2recps-x8.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8() local
106 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8()
107 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8()
209 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8()
297 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8()
|
| D | neonfma-rr1-lut64-p2-nr2fma-x12.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12() local
120 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12()
121 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12()
122 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12()
236 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12()
324 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12()
|
| D | neonfma-rr1-lut64-p2-nr1recps1fma-x12.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12() local
120 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12()
121 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12()
122 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12()
236 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12()
324 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12()
|
| D | neonfma-rr1-lut64-p2-nr2recps-x12.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12() local
120 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12()
121 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12()
122 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12()
236 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12()
324 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12()
|
| D | neonfma-rr1-lut64-p2-nr2fma-x8.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x8() local
106 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x8()
107 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x8()
209 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x8()
297 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x8()
|
| D | neonfma-rr1-lut64-p2-div-x4.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x4() local
89 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x4()
168 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x4()
|
| D | neonfma-rr1-lut64-p2-nr1recps1fma-x4.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x4() local
89 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x4()
177 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x4()
|
| D | neonfma-rr1-lut64-p2-nr2recps-x4.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x4() local
89 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x4()
177 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x4()
|
| D | neonfma-rr1-lut64-p2-nr2fma-x4.c | 33 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x4() local
89 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x4()
177 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x4()
|
| /external/XNNPACK/src/math/ |
| D | sigmoid-neonfma-rr1-lut64-p2-div.c | 30 const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f); in xnn_math_f32_sigmoid__neonfma_rr1_lut64_p2_div() local
86 float32x4_t vt = vfmaq_f32(vz, vn, vln2_o64); in xnn_math_f32_sigmoid__neonfma_rr1_lut64_p2_div()
|