| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | scalar-lut2048-p1-div-x4.c | 33 const float vln2_hi = 0x1.600000p-1f; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4() local
75 float vt0 = vn0 * vln2_hi + vz0; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
76 float vt1 = vn1 * vln2_hi + vz1; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
77 float vt2 = vn2 * vln2_hi + vz2; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
78 float vt3 = vn3 * vln2_hi + vz3; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
149 float vt = vn * vln2_hi + vz; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
|
| D | scalar-lut64-p2-div-x4.c | 33 const float vln2_hi = 0x1.630000p-1f; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4() local
75 float vt0 = vn0 * vln2_hi + vz0; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
76 float vt1 = vn1 * vln2_hi + vz1; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
77 float vt2 = vn2 * vln2_hi + vz2; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
78 float vt3 = vn3 * vln2_hi + vz3; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
154 float vt = vn * vln2_hi + vz; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4()
|
| D | scalar-p5-div-x4.c | 29 const float vln2_hi = 0x1.62E400p-1f; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4() local
66 float vt0 = vn0 * vln2_hi + vz0; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
67 float vt1 = vn1 * vln2_hi + vz1; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
68 float vt2 = vn2 * vln2_hi + vz2; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
69 float vt3 = vn3 * vln2_hi + vz3; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
158 float vt = vn * vln2_hi + vz; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x4()
|
| D | scalar-lut2048-p1-div-x2.c | 33 const float vln2_hi = 0x1.600000p-1f; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2() local
61 float vt0 = vn0 * vln2_hi + vz0; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2()
62 float vt1 = vn1 * vln2_hi + vz1; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2()
108 float vt = vn * vln2_hi + vz; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2()
|
| D | scalar-p5-div-x2.c | 29 const float vln2_hi = 0x1.62E400p-1f; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x2() local
56 float vt0 = vn0 * vln2_hi + vz0; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x2()
57 float vt1 = vn1 * vln2_hi + vz1; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x2()
113 float vt = vn * vln2_hi + vz; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x2()
|
| D | scalar-lut64-p2-div-x2.c | 33 const float vln2_hi = 0x1.630000p-1f; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2() local
61 float vt0 = vn0 * vln2_hi + vz0; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2()
62 float vt1 = vn1 * vln2_hi + vz1; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2()
111 float vt = vn * vln2_hi + vz; in xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2()
|
| D | wasmsimd-p5-div-x24.c | 28 const v128_t vln2_hi = wasm_f32x4_splat(0x1.62E400p-1f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24() local
75 v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
76 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
77 v128_t vt89AB = wasm_f32x4_add(vz89AB, wasm_f32x4_mul(vn89AB, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
78 v128_t vtCDEF = wasm_f32x4_add(vzCDEF, wasm_f32x4_mul(vnCDEF, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
79 v128_t vtGHIJ = wasm_f32x4_add(vzGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
80 v128_t vtKLMN = wasm_f32x4_add(vzKLMN, wasm_f32x4_mul(vnKLMN, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
177 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
205 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24()
|
| D | wasmsimd-p5-div-x16.c | 28 const v128_t vln2_hi = wasm_f32x4_splat(0x1.62E400p-1f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16() local
65 v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
66 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
67 v128_t vt89AB = wasm_f32x4_add(vz89AB, wasm_f32x4_mul(vn89AB, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
68 v128_t vtCDEF = wasm_f32x4_add(vzCDEF, wasm_f32x4_mul(vnCDEF, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
141 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
169 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
|
| D | wasmsimd-p5-div-x20.c | 28 const v128_t vln2_hi = wasm_f32x4_splat(0x1.62E400p-1f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20() local
70 v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
71 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
72 v128_t vt89AB = wasm_f32x4_add(vz89AB, wasm_f32x4_mul(vn89AB, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
73 v128_t vtCDEF = wasm_f32x4_add(vzCDEF, wasm_f32x4_mul(vnCDEF, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
74 v128_t vtGHIJ = wasm_f32x4_add(vzGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
159 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
187 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
|
| D | wasmsimd-p5-div-x12.c | 28 const v128_t vln2_hi = wasm_f32x4_splat(0x1.62E400p-1f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12() local
60 v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
61 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
62 v128_t vt89AB = wasm_f32x4_add(vz89AB, wasm_f32x4_mul(vn89AB, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
123 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
151 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12()
|
| D | wasmsimd-p5-div-x8.c | 28 const v128_t vln2_hi = wasm_f32x4_splat(0x1.62E400p-1f); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8() local
55 v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
56 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
105 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
133 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8()
|
| D | neon-rr2-p5-nr2recps-x24.c | 28 const float32x4_t vln2_hi = vmovq_n_f32(0x1.62E400p-1f); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24() local
74 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
75 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
76 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
77 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
78 float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
79 float32x4_t vtKLMN = vmlaq_f32(vzKLMN, vnKLMN, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
201 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
232 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
|
| D | neon-rr2-p5-nr2recps-x20.c | 28 const float32x4_t vln2_hi = vmovq_n_f32(0x1.62E400p-1f); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20() local
69 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
70 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
71 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
72 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
73 float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
179 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
210 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
|
| D | neon-rr2-p5-nr2recps-x16.c | 28 const float32x4_t vln2_hi = vmovq_n_f32(0x1.62E400p-1f); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16() local
64 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
65 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
66 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
67 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
157 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
188 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
|
| D | neon-rr2-p5-nr2recps-x12.c | 28 const float32x4_t vln2_hi = vmovq_n_f32(0x1.62E400p-1f); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12() local
59 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
60 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
61 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
135 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
166 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
|
| D | wasmsimd-lut64-p2-div-x24.c | 31 const v128_t vln2_hi = wasm_f32x4_splat(0x1.630000p-1f); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24() local
131 v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
132 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
133 v128_t vt89AB = wasm_f32x4_add(vz89AB, wasm_f32x4_mul(vn89AB, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
134 v128_t vtCDEF = wasm_f32x4_add(vzCDEF, wasm_f32x4_mul(vnCDEF, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
135 v128_t vtGHIJ = wasm_f32x4_add(vzGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
136 v128_t vtKLMN = wasm_f32x4_add(vzKLMN, wasm_f32x4_mul(vnKLMN, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
223 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
259 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
|
| D | wasmsimd-lut64-p2-div-x20.c | 31 const v128_t vln2_hi = wasm_f32x4_splat(0x1.630000p-1f); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20() local
117 v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
118 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
119 v128_t vt89AB = wasm_f32x4_add(vz89AB, wasm_f32x4_mul(vn89AB, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
120 v128_t vtCDEF = wasm_f32x4_add(vzCDEF, wasm_f32x4_mul(vnCDEF, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
121 v128_t vtGHIJ = wasm_f32x4_add(vzGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
199 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
235 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
|
| D | wasmsimd-lut64-p2-div-x16.c | 31 const v128_t vln2_hi = wasm_f32x4_splat(0x1.630000p-1f); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16() local
103 v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
104 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
105 v128_t vt89AB = wasm_f32x4_add(vz89AB, wasm_f32x4_mul(vn89AB, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
106 v128_t vtCDEF = wasm_f32x4_add(vzCDEF, wasm_f32x4_mul(vnCDEF, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
175 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
211 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
|
| D | neon-rr2-p5-nr2recps-x8.c | 28 const float32x4_t vln2_hi = vmovq_n_f32(0x1.62E400p-1f); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8() local
54 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
55 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
113 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
144 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
|
| D | wasmsimd-lut64-p2-div-x12.c | 31 const v128_t vln2_hi = wasm_f32x4_splat(0x1.630000p-1f); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12() local
89 v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12()
90 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12()
91 v128_t vt89AB = wasm_f32x4_add(vz89AB, wasm_f32x4_mul(vn89AB, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12()
151 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12()
187 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12()
|
| D | wasmsimd-lut64-p2-div-x8.c | 31 const v128_t vln2_hi = wasm_f32x4_splat(0x1.630000p-1f); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8() local
75 v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8()
76 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8()
127 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8()
163 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8()
|
| D | scalar-p5-div-x1.c | 29 const float vln2_hi = 0x1.62E400p-1f; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x1() local
48 float vt = vn * vln2_hi + vz; in xnn_f32_sigmoid_ukernel__scalar_p5_div_x1()
|
| /external/XNNPACK/src/f32-sigmoid/ |
| D | scalar-lut2048-p1-div.c.in | 31 const float vln2_hi = 0x1.600000p-1f; variable
60 float vt${N} = vn${N} * vln2_hi + vz${N};
103 float vt = vn * vln2_hi + vz;
134 float vt = vn * vln2_hi + vz;
164 float vt = vn * vln2_hi + vz;
|
| /external/XNNPACK/src/math/ |
| D | sigmoid-scalar-rr2-lut2048-p1-div.c | 33 const float vln2_hi = 0x1.600000p-1f; in xnn_math_f32_sigmoid__scalar_rr2_lut2048_p1_div() local
85 float vt = vn * vln2_hi + vz; in xnn_math_f32_sigmoid__scalar_rr2_lut2048_p1_div()
|
| D | sigmoid-scalar-rr2-lut64-p2-div.c | 33 const float vln2_hi = 0x1.630000p-1f; in xnn_math_f32_sigmoid__scalar_rr2_lut64_p2_div() local
85 float vt = vn * vln2_hi + vz; in xnn_math_f32_sigmoid__scalar_rr2_lut64_p2_div()
|