Lines Matching refs:__m512
28 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
29 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
30 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
31 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
32 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
33 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
34 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
35 const __m512 vone = _mm512_set1_ps(1.0f); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
38 const __m512 vx0 = _mm512_loadu_ps(x); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
39 const __m512 vx1 = _mm512_loadu_ps(x + 16); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
40 const __m512 vx2 = _mm512_loadu_ps(x + 32); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
41 const __m512 vx3 = _mm512_loadu_ps(x + 48); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
42 const __m512 vx4 = _mm512_loadu_ps(x + 64); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
43 const __m512 vx5 = _mm512_loadu_ps(x + 80); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
44 const __m512 vx6 = _mm512_loadu_ps(x + 96); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
45 const __m512 vx7 = _mm512_loadu_ps(x + 112); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
48 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
49 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
50 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
51 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
52 const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
53 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
54 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
55 const __m512 vz7 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx7), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
57 __m512 vn0 = _mm512_mul_ps(vz0, vlog2e); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
58 __m512 vn1 = _mm512_mul_ps(vz1, vlog2e); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
59 __m512 vn2 = _mm512_mul_ps(vz2, vlog2e); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
60 __m512 vn3 = _mm512_mul_ps(vz3, vlog2e); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
61 __m512 vn4 = _mm512_mul_ps(vz4, vlog2e); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
62 __m512 vn5 = _mm512_mul_ps(vz5, vlog2e); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
63 __m512 vn6 = _mm512_mul_ps(vz6, vlog2e); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
64 __m512 vn7 = _mm512_mul_ps(vz7, vlog2e); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
75 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
76 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
77 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
78 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
79 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
80 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
81 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
82 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2, vz7); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
84 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
85 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
86 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
87 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
88 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
89 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
90 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
91 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
129 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
130 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
131 const __m512 ve2 = _mm512_scalef_ps(vp2, vn2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
132 const __m512 ve3 = _mm512_scalef_ps(vp3, vn3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
133 const __m512 ve4 = _mm512_scalef_ps(vp4, vn4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
134 const __m512 ve5 = _mm512_scalef_ps(vp5, vn5); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
135 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
136 const __m512 ve7 = _mm512_scalef_ps(vp7, vn7); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
138 const __m512 vd0 = _mm512_add_ps(ve0, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
139 const __m512 vd1 = _mm512_add_ps(ve1, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
140 const __m512 vd2 = _mm512_add_ps(ve2, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
141 const __m512 vd3 = _mm512_add_ps(ve3, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
142 const __m512 vd4 = _mm512_add_ps(ve4, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
143 const __m512 vd5 = _mm512_add_ps(ve5, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
144 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
145 const __m512 vd7 = _mm512_add_ps(ve7, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
147 __m512 vf0 = _mm512_div_ps(ve0, vd0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
148 __m512 vf1 = _mm512_div_ps(ve1, vd1); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
149 __m512 vf2 = _mm512_div_ps(ve2, vd2); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
150 __m512 vf3 = _mm512_div_ps(ve3, vd3); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
151 __m512 vf4 = _mm512_div_ps(ve4, vd4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
152 __m512 vf5 = _mm512_div_ps(ve5, vd5); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
153 __m512 vf6 = _mm512_div_ps(ve6, vd6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
154 __m512 vf7 = _mm512_div_ps(ve7, vd7); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
176 const __m512 vx = _mm512_loadu_ps(x); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
179 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
181 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
183 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
185 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
191 const __m512 ve = _mm512_scalef_ps(vp, vn); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
192 const __m512 vd = _mm512_add_ps(ve, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
194 __m512 vf = _mm512_div_ps(ve, vd); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
209 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
210 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
212 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
214 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
216 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
222 const __m512 ve = _mm512_scalef_ps(vp, vn); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
223 const __m512 vd = _mm512_add_ps(ve, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
225 __m512 vf = _mm512_div_ps(ve, vd); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()