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_nr1fma_x80()
29 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
30 const __m512 vtable = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
35 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
36 const __m512 vc3 = _mm512_set1_ps(0x1.55559Ap-3f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
37 const __m512 vc2 = _mm512_set1_ps(0x1.00021Ep-1f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
38 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
41 const __m512 vx0 = _mm512_loadu_ps(x); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
42 const __m512 vx1 = _mm512_loadu_ps(x + 16); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
43 const __m512 vx2 = _mm512_loadu_ps(x + 32); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
44 const __m512 vx3 = _mm512_loadu_ps(x + 48); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
45 const __m512 vx4 = _mm512_loadu_ps(x + 64); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
48 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_nr1fma_x80()
49 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_nr1fma_x80()
50 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_nr1fma_x80()
51 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_nr1fma_x80()
52 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_nr1fma_x80()
54 __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
55 __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
56 __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
57 __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
58 __m512 vn4 = _mm512_fmadd_ps(vz4, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
60 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
61 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
62 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
63 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
64 const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
72 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
73 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
74 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
75 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
76 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
78 __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
79 __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
80 __m512 vp2 = _mm512_fmadd_ps(vt2, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
81 __m512 vp3 = _mm512_fmadd_ps(vt3, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
82 __m512 vp4 = _mm512_fmadd_ps(vt4, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
102 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
103 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
104 const __m512 ve2 = _mm512_scalef_ps(vp2, vn2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
105 const __m512 ve3 = _mm512_scalef_ps(vp3, vn3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
106 const __m512 ve4 = _mm512_scalef_ps(vp4, vn4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
108 const __m512 vd0 = _mm512_add_ps(ve0, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
109 const __m512 vd1 = _mm512_add_ps(ve1, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
110 const __m512 vd2 = _mm512_add_ps(ve2, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
111 const __m512 vd3 = _mm512_add_ps(ve3, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
112 const __m512 vd4 = _mm512_add_ps(ve4, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
114 __m512 vr0 = _mm512_rcp14_ps(vd0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
115 __m512 vr1 = _mm512_rcp14_ps(vd1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
116 __m512 vr2 = _mm512_rcp14_ps(vd2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
117 __m512 vr3 = _mm512_rcp14_ps(vd3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
118 __m512 vr4 = _mm512_rcp14_ps(vd4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
126 __m512 vf0 = _mm512_mul_ps(ve0, vr0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
127 __m512 vf1 = _mm512_mul_ps(ve1, vr1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
128 __m512 vf2 = _mm512_mul_ps(ve2, vr2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
129 __m512 vf3 = _mm512_mul_ps(ve3, vr3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
130 __m512 vf4 = _mm512_mul_ps(ve4, vr4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
147 const __m512 vx = _mm512_loadu_ps(x); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
150 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_nr1fma_x80()
152 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
153 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
156 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
158 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
163 const __m512 ve = _mm512_scalef_ps(vp, vn); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
164 const __m512 vd = _mm512_add_ps(ve, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
166 __m512 vr = _mm512_rcp14_ps(vd); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
169 __m512 vf = _mm512_mul_ps(ve, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
184 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
185 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_nr1fma_x80()
187 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
188 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
191 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
193 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
198 const __m512 ve = _mm512_scalef_ps(vp, vn); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
199 const __m512 vd = _mm512_add_ps(ve, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
201 __m512 vr = _mm512_rcp14_ps(vd); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
204 __m512 vf = _mm512_mul_ps(ve, vr); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()