Lines Matching refs:__m512
26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
27 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
30 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
31 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
32 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
33 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
34 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
35 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
37 const __m512 vi_max = _mm512_set1_ps(max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
39 __m512 vacc0 = _mm512_setzero_ps(); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
40 __m512 vacc1 = _mm512_setzero_ps(); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
41 __m512 vacc2 = _mm512_setzero_ps(); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
42 __m512 vacc3 = _mm512_setzero_ps(); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
43 __m512 vacc4 = _mm512_setzero_ps(); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
44 __m512 vacc5 = _mm512_setzero_ps(); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
47 const __m512 vi0 = _mm512_loadu_ps(input); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
48 const __m512 vi1 = _mm512_loadu_ps(input + 16); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
49 const __m512 vi2 = _mm512_loadu_ps(input + 32); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
50 const __m512 vi3 = _mm512_loadu_ps(input + 48); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
51 const __m512 vi4 = _mm512_loadu_ps(input + 64); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
52 const __m512 vi5 = _mm512_loadu_ps(input + 80); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
53 const __m512 vi6 = _mm512_loadu_ps(input + 96); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
54 const __m512 vi7 = _mm512_loadu_ps(input + 112); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
55 const __m512 vi8 = _mm512_loadu_ps(input + 128); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
56 const __m512 vi9 = _mm512_loadu_ps(input + 144); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
57 const __m512 vi10 = _mm512_loadu_ps(input + 160); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
58 const __m512 vi11 = _mm512_loadu_ps(input + 176); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
62 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
63 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
64 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
65 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
66 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
67 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
68 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
69 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
70 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
71 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
72 const __m512 vx10 = _mm512_sub_ps(vi10, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
73 const __m512 vx11 = _mm512_sub_ps(vi11, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
76 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
77 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
78 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
79 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
80 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
81 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
82 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
83 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
84 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
85 const __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
86 const __m512 vn10 = _mm512_roundscale_ps(_mm512_mul_ps(vx10, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
87 const __m512 vn11 = _mm512_roundscale_ps(_mm512_mul_ps(vx11, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
91 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
92 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
93 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
94 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
95 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
96 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
97 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
98 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
99 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
100 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
101 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
102 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
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()
127 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
128 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
129 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
186 const __m512 vf0 = _mm512_scalef_ps(vp0, vn0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
187 const __m512 vf1 = _mm512_scalef_ps(vp1, vn1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
188 const __m512 vf2 = _mm512_scalef_ps(vp2, vn2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
189 const __m512 vf3 = _mm512_scalef_ps(vp3, vn3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
190 const __m512 vf4 = _mm512_scalef_ps(vp4, vn4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
191 const __m512 vf5 = _mm512_scalef_ps(vp5, vn5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
192 const __m512 vf6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
193 const __m512 vf7 = _mm512_scalef_ps(vp7, vn7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
194 const __m512 vf8 = _mm512_scalef_ps(vp8, vn8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
195 const __m512 vf9 = _mm512_scalef_ps(vp9, vn9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
196 const __m512 vf10 = _mm512_scalef_ps(vp10, vn10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
197 const __m512 vf11 = _mm512_scalef_ps(vp11, vn11); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
220 __m512 vacc = vacc0; in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
223 const __m512 vi = _mm512_loadu_ps(input); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
227 const __m512 vx = _mm512_sub_ps(vi, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
230 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
234 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
238 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
247 const __m512 vf = _mm512_scalef_ps(vp, vn); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
258 const __m512 vi = _mm512_maskz_loadu_ps(vmask, input); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
261 const __m512 vx = _mm512_sub_ps(vi, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
264 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
268 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
272 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
281 const __m512 vf = _mm512_scalef_ps(vp, vn); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()