| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | avx512f-rr1-lut16-p3-perm-scalef-div-x112.c | 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()
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()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c | 57 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
58 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
59 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
60 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
61 const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
62 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
63 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
64 const __m512 vz7 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx7), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
75 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
76 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
[all …]
|
| D | avx512f-rr1-lut16-p3-perm-scalef-div-x128.c | 51 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_x128()
52 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_x128()
53 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_x128()
54 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_x128()
55 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_x128()
56 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_x128()
57 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_x128()
58 const __m512 vz7 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx7), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128()
69 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128()
70 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c | 55 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
56 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
57 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
58 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
59 const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
60 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
69 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
70 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
71 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
72 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
[all …]
|
| D | avx512f-rr1-lut16-p3-perm-scalef-div-x96.c | 49 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_x96()
50 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_x96()
51 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_x96()
52 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_x96()
53 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_x96()
54 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_x96()
63 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96()
64 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96()
65 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96()
66 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c | 56 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
57 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
58 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
59 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
60 const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
61 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
62 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
72 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
73 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
74 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
[all …]
|
| D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c | 51 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_x128()
52 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_x128()
53 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_x128()
54 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_x128()
55 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_x128()
56 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_nr1fma_x128()
57 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_nr1fma_x128()
58 const __m512 vz7 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx7), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
69 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
70 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c | 54 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
55 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
56 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
57 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
58 const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
66 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
67 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
68 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
69 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
70 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
[all …]
|
| D | avx512f-rr1-lut16-p3-perm-scalef-div-x80.c | 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_div_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_div_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_div_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_div_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_div_x80()
60 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80()
61 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80()
62 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80()
63 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80()
64 const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80()
[all …]
|
| D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c | 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_nr1fma_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_nr1fma_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_nr1fma_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_nr1fma_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_nr1fma_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_nr1fma_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_nr1fma_x112()
66 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
67 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
68 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c | 56 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
57 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
58 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
59 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
60 const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
61 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
62 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
72 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
73 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
74 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c | 57 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
58 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
59 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
60 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
61 const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
62 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
63 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
64 const __m512 vz7 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx7), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
75 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
76 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
[all …]
|
| D | avx512f-rr1-lut16-p3-perm-scalef-div-x64.c | 47 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_x64()
48 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_x64()
49 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_x64()
50 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_x64()
57 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64()
58 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64()
59 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64()
60 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64()
107 …vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64()
108 …vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c | 53 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
54 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
55 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
56 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
63 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
64 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
65 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
66 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
113 …vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
114 …vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c | 54 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
55 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
56 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
57 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
58 const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
66 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
67 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
68 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
69 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
70 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
[all …]
|
| D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c | 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()
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()
[all …]
|
| D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c | 49 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_x96()
50 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_x96()
51 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_x96()
52 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_x96()
53 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_x96()
54 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_nr1fma_x96()
63 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96()
64 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96()
65 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96()
66 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c | 55 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
56 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
57 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
58 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
59 const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
60 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
69 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
70 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
71 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
72 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
[all …]
|
| D | avx512f-rr1-lut16-p3-perm-scalef-div-x48.c | 46 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_x48()
47 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_x48()
48 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_x48()
54 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x48()
55 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x48()
56 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x48()
94 …vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x48()
95 …vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x48()
96 …vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x48()
107 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_x48()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c | 52 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
53 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
54 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
60 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
61 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
62 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
100 …vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
101 …vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
102 …vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
113 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
[all …]
|
| D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c | 47 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_x64()
48 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_x64()
49 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_x64()
50 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_x64()
57 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64()
58 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64()
59 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64()
60 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64()
118 …vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64()
119 …vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c | 53 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
54 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
55 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
56 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
63 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
64 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
65 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
66 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
124 …vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
125 …vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c | 52 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
53 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
54 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
60 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
61 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
62 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
109 …vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
110 …vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
111 …vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
122 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
[all …]
|
| D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c | 46 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_x48()
47 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_x48()
48 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_x48()
54 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x48()
55 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x48()
56 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x48()
103 …vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x48()
104 …vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x48()
105 …vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x48()
116 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_x48()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c | 51 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
52 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
57 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
58 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
87 …vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
88 …vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone,… in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
98 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
101 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
116 …vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf… in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
130 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
[all …]
|