Lines Matching refs:__m512
28 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p19f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
29 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
30 const __m512 vtable = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
35 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
36 const __m512 vc3 = _mm512_set1_ps(0x1.55559Ap-3f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
37 const __m512 vc2 = _mm512_set1_ps(0x1.00021Ep-1f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
38 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
41 const __m512 vx0 = _mm512_loadu_ps(x); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
42 const __m512 vx1 = _mm512_loadu_ps(x + 16); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
43 const __m512 vx2 = _mm512_loadu_ps(x + 32); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
44 const __m512 vx3 = _mm512_loadu_ps(x + 48); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
45 const __m512 vx4 = _mm512_loadu_ps(x + 64); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
46 const __m512 vx5 = _mm512_loadu_ps(x + 80); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
47 const __m512 vx6 = _mm512_loadu_ps(x + 96); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
50 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
51 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
52 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
53 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
54 const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
55 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
56 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
58 __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
59 __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
60 __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
61 __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
62 __m512 vn4 = _mm512_fmadd_ps(vz4, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
63 __m512 vn5 = _mm512_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
64 __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
66 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
67 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
68 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
69 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
70 const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
71 const __m512 vl5 = _mm512_permutexvar_ps(_mm512_castps_si512(vn5), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
72 const __m512 vl6 = _mm512_permutexvar_ps(_mm512_castps_si512(vn6), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
82 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
83 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
84 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
85 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
86 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
87 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
88 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
90 __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
91 __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
92 __m512 vp2 = _mm512_fmadd_ps(vt2, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
93 __m512 vp3 = _mm512_fmadd_ps(vt3, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
94 __m512 vp4 = _mm512_fmadd_ps(vt4, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
95 __m512 vp5 = _mm512_fmadd_ps(vt5, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
96 __m512 vp6 = _mm512_fmadd_ps(vt6, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
122 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
123 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
124 const __m512 ve2 = _mm512_scalef_ps(vp2, vn2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
125 const __m512 ve3 = _mm512_scalef_ps(vp3, vn3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
126 const __m512 ve4 = _mm512_scalef_ps(vp4, vn4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
127 const __m512 ve5 = _mm512_scalef_ps(vp5, vn5); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
128 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
130 const __m512 vd0 = _mm512_add_ps(ve0, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
131 const __m512 vd1 = _mm512_add_ps(ve1, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
132 const __m512 vd2 = _mm512_add_ps(ve2, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
133 const __m512 vd3 = _mm512_add_ps(ve3, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
134 const __m512 vd4 = _mm512_add_ps(ve4, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
135 const __m512 vd5 = _mm512_add_ps(ve5, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
136 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
138 __m512 vf0 = _mm512_div_ps(ve0, vd0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
139 __m512 vf1 = _mm512_div_ps(ve1, vd1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
140 __m512 vf2 = _mm512_div_ps(ve2, vd2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
141 __m512 vf3 = _mm512_div_ps(ve3, vd3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
142 __m512 vf4 = _mm512_div_ps(ve4, vd4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
143 __m512 vf5 = _mm512_div_ps(ve5, vd5); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
144 __m512 vf6 = _mm512_div_ps(ve6, vd6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
164 const __m512 vx = _mm512_loadu_ps(x); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
167 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
169 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
170 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
173 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
175 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
180 const __m512 ve = _mm512_scalef_ps(vp, vn); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
181 const __m512 vd = _mm512_add_ps(ve, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
183 __m512 vf = _mm512_div_ps(ve, vd); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
198 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
199 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
201 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
202 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
205 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
207 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
212 const __m512 ve = _mm512_scalef_ps(vp, vn); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
213 const __m512 vd = _mm512_add_ps(ve, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
215 __m512 vf = _mm512_div_ps(ve, vd); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()