Lines Matching refs:__m512
26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
27 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
30 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
31 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
32 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
33 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
34 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
35 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
37 const __m512 vi_max = _mm512_set1_ps(max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
39 __m512 vacc0 = _mm512_setzero_ps(); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
40 __m512 vacc1 = _mm512_setzero_ps(); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
41 __m512 vacc2 = _mm512_setzero_ps(); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
42 __m512 vacc3 = _mm512_setzero_ps(); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
43 __m512 vacc4 = _mm512_setzero_ps(); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
46 const __m512 vi0 = _mm512_loadu_ps(input); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
47 const __m512 vi1 = _mm512_loadu_ps(input + 16); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
48 const __m512 vi2 = _mm512_loadu_ps(input + 32); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
49 const __m512 vi3 = _mm512_loadu_ps(input + 48); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
50 const __m512 vi4 = _mm512_loadu_ps(input + 64); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
51 const __m512 vi5 = _mm512_loadu_ps(input + 80); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
52 const __m512 vi6 = _mm512_loadu_ps(input + 96); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
53 const __m512 vi7 = _mm512_loadu_ps(input + 112); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
54 const __m512 vi8 = _mm512_loadu_ps(input + 128); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
55 const __m512 vi9 = _mm512_loadu_ps(input + 144); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
59 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
60 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
61 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
62 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
63 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
64 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
65 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
66 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
67 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
68 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
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()
80 const __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
84 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
85 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
86 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
87 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
88 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
89 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
90 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
91 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
92 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
93 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
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()
116 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
165 const __m512 vf0 = _mm512_scalef_ps(vp0, vn0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
166 const __m512 vf1 = _mm512_scalef_ps(vp1, vn1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
167 const __m512 vf2 = _mm512_scalef_ps(vp2, vn2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
168 const __m512 vf3 = _mm512_scalef_ps(vp3, vn3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
169 const __m512 vf4 = _mm512_scalef_ps(vp4, vn4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
170 const __m512 vf5 = _mm512_scalef_ps(vp5, vn5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
171 const __m512 vf6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
172 const __m512 vf7 = _mm512_scalef_ps(vp7, vn7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
173 const __m512 vf8 = _mm512_scalef_ps(vp8, vn8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
174 const __m512 vf9 = _mm512_scalef_ps(vp9, vn9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
194 __m512 vacc = vacc0; in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
197 const __m512 vi = _mm512_loadu_ps(input); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
201 const __m512 vx = _mm512_sub_ps(vi, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
204 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
208 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
212 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
221 const __m512 vf = _mm512_scalef_ps(vp, vn); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
232 const __m512 vi = _mm512_maskz_loadu_ps(vmask, input); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
235 const __m512 vx = _mm512_sub_ps(vi, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
238 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
242 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
246 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
255 const __m512 vf = _mm512_scalef_ps(vp, vn); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()