| /external/XNNPACK/src/f32-raddexpminusmax/gen/ |
| D | avx512f-p5-scalef-x192-acc2.c | 35 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2() local
114 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
115 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
116 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
117 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
118 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
119 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
120 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
121 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
122 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
|
| D | avx512f-p5-scalef-x192.c | 35 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192() local
113 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
114 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
115 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
116 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
117 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
118 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
119 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
120 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
121 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x192-acc3.c | 35 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3() local
115 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
116 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
117 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
118 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
119 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
120 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
121 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
122 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
123 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
[all …]
|
| D | avx512f-p5-scalef-x192-acc6.c | 35 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6() local
118 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
119 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
120 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
121 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
122 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
123 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
124 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
125 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
126 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
[all …]
|
| D | avx512f-p5-scalef-x160-acc5.c | 35 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5() local
107 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
108 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
109 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
110 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
111 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
112 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
113 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
114 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
115 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 35 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160() local
103 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
104 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
105 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
106 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
107 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
108 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
109 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
110 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
111 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x160-acc2.c | 35 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2() local
104 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
105 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
106 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
107 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
108 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
109 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
110 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
111 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
112 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 35 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144() local
98 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
99 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
100 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
101 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
102 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
103 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
104 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
105 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
106 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| D | avx512f-p5-scalef-x144-acc3.c | 35 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3() local
100 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
101 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
102 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
103 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
104 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
105 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
106 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
107 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
108 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
[all …]
|
| /external/XNNPACK/src/f32-vscaleexpminusmax/gen/ |
| D | avx512f-p5-scalef-x192.c | 36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192() local
114 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
115 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
116 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
117 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
118 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
119 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
120 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
121 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
122 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x176.c | 36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176() local
109 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
110 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
111 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
112 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
113 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
114 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
115 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
116 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
117 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160() local
104 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
105 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
106 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
107 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
108 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
109 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
110 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
111 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
112 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144() local
99 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
100 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
101 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
102 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
103 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
104 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
105 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
106 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
107 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| /external/XNNPACK/src/f32-vscaleextexp/gen/ |
| D | avx512f-p5-scalef-x192.c | 37 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192() local
101 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
102 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
103 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
104 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
105 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
106 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
107 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
108 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
109 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x176.c | 37 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176() local
97 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
98 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
99 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
100 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
101 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
102 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
103 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
104 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
105 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 37 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160() local
93 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
94 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
95 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
96 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
97 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
98 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
99 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
100 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
101 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 37 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144() local
89 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
90 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
91 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
92 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
93 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
94 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
95 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
96 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
97 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | avx512f-p5-scalef-x192-acc2.c | 36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2() local
115 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
116 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
117 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
118 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
119 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
120 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
121 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
122 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
123 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
|
| D | avx512f-p5-scalef-x192-acc3.c | 36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3() local
116 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
117 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
118 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
119 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
120 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
121 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
122 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
123 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
124 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
[all …]
|
| D | avx512f-p5-scalef-x192.c | 36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192() local
114 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
115 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
116 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
117 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
118 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
119 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
120 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
121 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
122 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x160-acc2.c | 36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2() local
105 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
106 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
107 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
108 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
109 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
110 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
111 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
112 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
113 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160() local
104 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
105 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
106 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
107 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
108 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
109 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
110 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
111 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
112 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144() local
99 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
100 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
101 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
102 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
103 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
104 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
105 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
106 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
107 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| D | avx512f-p5-scalef-x144-acc3.c | 36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3() local
101 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
102 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
103 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
104 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
105 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
106 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
107 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
108 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
109 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
[all …]
|
| /external/XNNPACK/src/f32-raddextexp/gen/ |
| D | avx512f-p5-scalef-x192.c | 36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() local
101 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
102 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
103 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
104 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
105 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
106 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
107 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
108 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
109 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
[all …]
|