| /external/XNNPACK/src/f32-raddexpminusmax/gen/ |
| D | avx512f-p5-scalef-x192-acc2.c | 26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2() local
72 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
73 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
74 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
75 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
76 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
77 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
78 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
79 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
80 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
|
| D | avx512f-p5-scalef-x192.c | 26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192() local
71 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
72 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
73 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
74 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
75 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
76 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
77 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
78 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
79 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x192-acc3.c | 26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3() local
73 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
74 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
75 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
76 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
77 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
78 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
79 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
80 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
81 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
[all …]
|
| D | avx512f-p5-scalef-x192-acc6.c | 26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6() local
76 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
77 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
78 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
79 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
80 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
81 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
82 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
83 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
84 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
[all …]
|
| D | avx512f-p5-scalef-x160-acc5.c | 26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5() local
71 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
72 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
73 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
74 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
75 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
76 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
77 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
78 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
79 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160() local
67 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
68 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
69 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
70 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
71 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
72 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
73 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
74 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
75 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x160-acc2.c | 26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2() local
68 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
69 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
70 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
71 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
72 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
73 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
74 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
75 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
76 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144() local
65 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
66 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
67 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
68 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
69 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
70 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
71 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
72 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
73 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| D | avx512f-p5-scalef-x144-acc3.c | 26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3() local
67 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
68 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
69 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
70 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
71 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
72 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
73 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
74 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
75 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
[all …]
|
| /external/XNNPACK/src/f32-vscaleexpminusmax/gen/ |
| D | avx512f-p5-scalef-x192.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192() local
72 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
73 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
74 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
75 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
76 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
77 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
78 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
79 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
80 __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x176.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176() local
70 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
71 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
72 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
73 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
74 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
75 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
76 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
77 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
78 __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160() local
68 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
69 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
70 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
71 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
72 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
73 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
74 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
75 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
76 __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144() local
66 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
67 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
68 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
69 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
70 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
71 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
72 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
73 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
74 __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| /external/XNNPACK/src/f32-vscaleextexp/gen/ |
| D | avx512f-p5-scalef-x192.c | 28 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192() local
59 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
60 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
61 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
62 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
63 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
64 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
65 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
66 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
67 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x176.c | 28 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176() local
58 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
59 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
60 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
61 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
62 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
63 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
64 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
65 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
66 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 28 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160() local
57 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
58 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
59 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
60 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
61 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
62 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
63 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
64 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
65 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 28 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144() local
56 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
57 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
58 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
59 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
60 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
61 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
62 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
63 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
64 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | avx512f-p5-scalef-x192-acc2.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2() local
73 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
74 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
75 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
76 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
77 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
78 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
79 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
80 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
81 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
|
| D | avx512f-p5-scalef-x192-acc3.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3() local
74 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
75 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
76 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
77 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
78 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
79 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
80 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
81 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
82 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
[all …]
|
| D | avx512f-p5-scalef-x192.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192() local
72 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
73 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
74 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
75 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
76 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
77 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
78 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
79 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
80 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x160-acc2.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2() local
69 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
70 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
71 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
72 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
73 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
74 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
75 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
76 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
77 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160() local
68 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
69 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
70 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
71 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
72 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
73 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
74 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
75 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
76 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144() local
66 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
67 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
68 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
69 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
70 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
71 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
72 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
73 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
74 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| D | avx512f-p5-scalef-x144-acc3.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3() local
68 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
69 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
70 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
71 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
72 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
73 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
74 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
75 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
76 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
[all …]
|
| /external/XNNPACK/src/f32-raddextexp/gen/ |
| D | avx512f-p5-scalef-x192.c | 27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() local
59 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
60 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
61 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
62 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
63 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
64 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
65 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
66 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
67 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
[all …]
|