| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | scalar-lut2048-p1-div-x4.c | 34 const float vln2_lo = 0x1.7217F8p-8f; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4() local
80 vt0 = vn0 * vln2_lo + vt0; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
81 vt1 = vn1 * vln2_lo + vt1; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
82 vt2 = vn2 * vln2_lo + vt2; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
83 vt3 = vn3 * vln2_lo + vt3; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
150 vt = vn * vln2_lo + vt; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
|
| D | scalar-lut64-p2-div-x4.c | 34 const float vln2_lo = -0x1.BD0106p-13f; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4() local
80 vt0 = vn0 * vln2_lo + vt0; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
81 vt1 = vn1 * vln2_lo + vt1; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
82 vt2 = vn2 * vln2_lo + vt2; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
83 vt3 = vn3 * vln2_lo + vt3; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
155 vt = vn * vln2_lo + vt; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
|
| D | scalar-p5-div-x4.c | 30 const float vln2_lo = 0x1.7F7D1Cp-20f; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4() local
71 vt0 = vn0 * vln2_lo + vt0; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
72 vt1 = vn1 * vln2_lo + vt1; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
73 vt2 = vn2 * vln2_lo + vt2; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
74 vt3 = vn3 * vln2_lo + vt3; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
159 vt = vn * vln2_lo + vt; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
|
| D | scalar-lut2048-p1-div-x2.c | 34 const float vln2_lo = 0x1.7217F8p-8f; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2() local
64 vt0 = vn0 * vln2_lo + vt0; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2()
65 vt1 = vn1 * vln2_lo + vt1; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2()
109 vt = vn * vln2_lo + vt; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2()
|
| D | scalar-p5-div-x2.c | 30 const float vln2_lo = 0x1.7F7D1Cp-20f; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x2() local
59 vt0 = vn0 * vln2_lo + vt0; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x2()
60 vt1 = vn1 * vln2_lo + vt1; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x2()
114 vt = vn * vln2_lo + vt; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x2()
|
| D | scalar-lut64-p2-div-x2.c | 34 const float vln2_lo = -0x1.BD0106p-13f; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2() local
64 vt0 = vn0 * vln2_lo + vt0; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2()
65 vt1 = vn1 * vln2_lo + vt1; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2()
112 vt = vn * vln2_lo + vt; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2()
|
| D | wasmsimd-p5-div-x24.c | 29 const v128_t vln2_lo = wasm_f32x4_splat(0x1.7F7D1Cp-20f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24() local
82 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
83 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
84 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
85 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
86 vtGHIJ = wasm_f32x4_add(vtGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
87 vtKLMN = wasm_f32x4_add(vtKLMN, wasm_f32x4_mul(vnKLMN, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
178 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
206 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
|
| D | wasmsimd-p5-div-x16.c | 29 const v128_t vln2_lo = wasm_f32x4_splat(0x1.7F7D1Cp-20f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16() local
70 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
71 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
72 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
73 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
142 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
170 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
|
| D | wasmsimd-p5-div-x20.c | 29 const v128_t vln2_lo = wasm_f32x4_splat(0x1.7F7D1Cp-20f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20() local
76 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
77 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
78 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
79 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
80 vtGHIJ = wasm_f32x4_add(vtGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
160 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
188 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
|
| D | wasmsimd-p5-div-x12.c | 29 const v128_t vln2_lo = wasm_f32x4_splat(0x1.7F7D1Cp-20f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12() local
64 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
65 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
66 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
124 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
152 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
|
| D | wasmsimd-p5-div-x8.c | 29 const v128_t vln2_lo = wasm_f32x4_splat(0x1.7F7D1Cp-20f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8() local
58 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
59 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
106 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
134 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
|
| D | neon-rr2-p5-nr2recps-x24.c | 29 const float32x4_t vln2_lo = vmovq_n_f32(0x1.7F7D1Cp-20f); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24() local
81 vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
82 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
83 vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
84 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
85 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
86 vtKLMN = vmlaq_f32(vtKLMN, vnKLMN, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
202 vt = vmlaq_f32(vt, vn, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
233 vt = vmlaq_f32(vt, vn, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
|
| D | neon-rr2-p5-nr2recps-x20.c | 29 const float32x4_t vln2_lo = vmovq_n_f32(0x1.7F7D1Cp-20f); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20() local
75 vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
76 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
77 vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
78 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
79 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
180 vt = vmlaq_f32(vt, vn, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
211 vt = vmlaq_f32(vt, vn, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
|
| D | neon-rr2-p5-nr2recps-x16.c | 29 const float32x4_t vln2_lo = vmovq_n_f32(0x1.7F7D1Cp-20f); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16() local
69 vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
70 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
71 vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
72 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
158 vt = vmlaq_f32(vt, vn, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
189 vt = vmlaq_f32(vt, vn, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
|
| D | neon-rr2-p5-nr2recps-x12.c | 29 const float32x4_t vln2_lo = vmovq_n_f32(0x1.7F7D1Cp-20f); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12() local
63 vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
64 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
65 vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
136 vt = vmlaq_f32(vt, vn, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
167 vt = vmlaq_f32(vt, vn, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
|
| D | wasmsimd-lut64-p2-div-x24.c | 32 const v128_t vln2_lo = wasm_f32x4_splat(-0x1.BD0106p-13f); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24() local
138 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
139 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
140 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
141 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
142 vtGHIJ = wasm_f32x4_add(vtGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
143 vtKLMN = wasm_f32x4_add(vtKLMN, wasm_f32x4_mul(vnKLMN, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
224 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
260 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
|
| D | wasmsimd-lut64-p2-div-x20.c | 32 const v128_t vln2_lo = wasm_f32x4_splat(-0x1.BD0106p-13f); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20() local
123 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
124 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
125 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
126 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
127 vtGHIJ = wasm_f32x4_add(vtGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
200 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
236 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
|
| D | wasmsimd-lut64-p2-div-x16.c | 32 const v128_t vln2_lo = wasm_f32x4_splat(-0x1.BD0106p-13f); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16() local
108 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
109 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
110 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
111 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
176 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
212 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
|
| D | neon-rr2-p5-nr2recps-x8.c | 29 const float32x4_t vln2_lo = vmovq_n_f32(0x1.7F7D1Cp-20f); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8() local
57 vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
58 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
114 vt = vmlaq_f32(vt, vn, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
145 vt = vmlaq_f32(vt, vn, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
|
| D | wasmsimd-lut64-p2-div-x12.c | 32 const v128_t vln2_lo = wasm_f32x4_splat(-0x1.BD0106p-13f); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12() local
93 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12()
94 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12()
95 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12()
152 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12()
188 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12()
|
| D | wasmsimd-lut64-p2-div-x8.c | 32 const v128_t vln2_lo = wasm_f32x4_splat(-0x1.BD0106p-13f); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8() local
78 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8()
79 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8()
128 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8()
164 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8()
|
| D | scalar-p5-div-x1.c | 30 const float vln2_lo = 0x1.7F7D1Cp-20f; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x1() local
49 vt = vn * vln2_lo + vt; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x1()
|
| /external/XNNPACK/src/f32-sigmoid/ |
| D | scalar-lut2048-p1-div.c.in | 32 const float vln2_lo = 0x1.7217F8p-8f; variable
63 vt${N} = vn${N} * vln2_lo + vt${N};
104 vt = vn * vln2_lo + vt;
135 vt = vn * vln2_lo + vt;
165 vt = vn * vln2_lo + vt;
|
| /external/XNNPACK/src/math/ |
| D | sigmoid-scalar-rr2-lut2048-p1-div.c | 34 const float vln2_lo = 0x1.7217F8p-8f; in xnn_math_f32_sigmoid__scalar_rr2_lut2048_p1_div() local
86 vt = vn * vln2_lo + vt; in xnn_math_f32_sigmoid__scalar_rr2_lut2048_p1_div()
|
| D | sigmoid-scalar-rr2-lut64-p2-div.c | 34 const float vln2_lo = -0x1.BD0106p-13f; in xnn_math_f32_sigmoid__scalar_rr2_lut64_p2_div() local
86 vt = vn * vln2_lo + vt; in xnn_math_f32_sigmoid__scalar_rr2_lut64_p2_div()
|