| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | scalar-lut2048-p1-div-x4.c | 34 const float vminus_log2e_x2048 = -0x1.715476p11f; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4() local
70 float vn0 = vz0 * vminus_log2e_x2048 + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
71 float vn1 = vz1 * vminus_log2e_x2048 + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
72 float vn2 = vz2 * vminus_log2e_x2048 + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
73 float vn3 = vz3 * vminus_log2e_x2048 + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
196 float vn = vz * vminus_log2e_x2048 + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4()
|
| D | scalar-lut2048-p1-div-x2.c | 34 const float vminus_log2e_x2048 = -0x1.715476p11f; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2() local
66 float vn0 = vz0 * vminus_log2e_x2048 + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2()
67 float vn1 = vz1 * vminus_log2e_x2048 + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2()
157 float vn = vz * vminus_log2e_x2048 + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2()
|
| D | neonfma-rr1-lut2048-p1-div-x24.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24() local
69 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
70 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
71 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
72 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
73 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
74 float32x4_t vnKLMN = vfmaq_f32(vmagic_bias, vzKLMN, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
255 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
332 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
|
| D | neonfma-rr1-lut2048-p1-div-x20.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20() local
67 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
68 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
69 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
70 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
71 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
232 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
309 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
|
| D | neonfma-rr1-lut2048-p1-div-x16.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16() local
65 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
66 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
67 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
68 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
209 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
286 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
|
| D | neonfma-rr1-lut2048-p1-div-x12.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x12() local
63 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x12()
64 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x12()
65 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x12()
186 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x12()
263 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x12()
|
| D | neonfma-rr1-lut2048-p1-div-x8.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x8() local
61 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x8()
62 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x8()
163 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x8()
240 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x8()
|
| D | neonfma-rr1-lut2048-p1-nr2recps-x16.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16() local
65 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16()
66 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16()
67 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16()
68 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16()
227 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16()
313 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16()
|
| D | neonfma-rr1-lut2048-p1-nr2fma-x24.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24() local
69 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
70 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
71 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
72 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
73 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
74 float32x4_t vnKLMN = vfmaq_f32(vmagic_bias, vzKLMN, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
279 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
365 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
|
| D | neonfma-rr1-lut2048-p1-nr2recps-x24.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24() local
69 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
70 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
71 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
72 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
73 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
74 float32x4_t vnKLMN = vfmaq_f32(vmagic_bias, vzKLMN, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
279 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
365 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
|
| D | neon-rr2-lut2048-p1-nr2recps-x24.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24() local
71 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24()
72 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24()
73 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24()
74 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24()
75 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24()
76 float32x4_t vnKLMN = vmlaq_f32(vmagic_bias, vzKLMN, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24()
289 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24()
377 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24()
|
| D | neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24() local
69 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
70 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
71 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
72 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
73 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
74 float32x4_t vnKLMN = vfmaq_f32(vmagic_bias, vzKLMN, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
279 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
365 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
|
| D | scalar-lut2048-p1-div-x1.c | 34 const float vminus_log2e_x2048 = -0x1.715476p11f; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x1() local
63 float vn = vz * vminus_log2e_x2048 + vmagic_bias; in xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x1()
|
| D | neonfma-rr1-lut2048-p1-nr2recps-x20.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20() local
67 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
68 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
69 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
70 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
71 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
253 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
339 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
|
| D | neonfma-rr1-lut2048-p1-nr2fma-x20.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() local
67 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
68 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
69 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
70 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
71 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
253 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
339 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
|
| D | neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20() local
67 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
68 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
69 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
70 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
71 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
253 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
339 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
|
| D | neon-rr2-lut2048-p1-nr2recps-x20.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20() local
69 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20()
70 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20()
71 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20()
72 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20()
73 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20()
262 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20()
350 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20()
|
| D | neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12() local
63 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12()
64 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12()
65 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12()
201 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12()
287 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12()
|
| D | neonfma-rr1-lut2048-p1-nr2recps-x12.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12() local
63 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12()
64 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12()
65 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12()
201 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12()
287 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12()
|
| D | neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() local
65 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16()
66 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16()
67 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16()
68 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16()
227 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16()
313 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16()
|
| D | neon-rr2-lut2048-p1-nr2recps-x16.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16() local
67 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16()
68 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16()
69 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16()
70 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16()
235 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16()
323 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16()
|
| D | neonfma-rr1-lut2048-p1-nr2fma-x16.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16() local
65 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16()
66 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16()
67 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16()
68 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16()
227 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16()
313 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16()
|
| D | neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8() local
61 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8()
62 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8()
175 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8()
261 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8()
|
| D | neonfma-rr1-lut2048-p1-nr2recps-x8.c | 32 const float32x4_t vminus_log2e_x2048 = vmovq_n_f32(-0x1.715476p11f); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8() local
61 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8()
62 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8()
175 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8()
261 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8()
|
| /external/XNNPACK/src/math/ |
| D | sigmoid-scalar-lut2048-p1-div.c | 544 const float vminus_log2e_x2048 = -0x1.715476p11f; in xnn_math_f32_sigmoid__scalar_lut2048_p1_div() local
573 float vn = vz * vminus_log2e_x2048 + vmagic_bias; in xnn_math_f32_sigmoid__scalar_lut2048_p1_div()
|