| /external/XNNPACK/src/f32-vsqrt/gen/ |
| D | avx512f-nr1fma1adj-x128.c | 49 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
50 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
51 __m512 vsqrtx1 = _mm512_mul_ps(vrsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
52 __m512 vhalfrsqrtx1 = _mm512_mul_ps(vrsqrtx1, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
53 __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
54 __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
55 __m512 vsqrtx3 = _mm512_mul_ps(vrsqrtx3, vx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
56 __m512 vhalfrsqrtx3 = _mm512_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
57 __m512 vsqrtx4 = _mm512_mul_ps(vrsqrtx4, vx4); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
58 __m512 vhalfrsqrtx4 = _mm512_mul_ps(vrsqrtx4, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
[all …]
|
| D | avx512f-nr1fma1adj-x112.c | 47 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
48 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
49 __m512 vsqrtx1 = _mm512_mul_ps(vrsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
50 __m512 vhalfrsqrtx1 = _mm512_mul_ps(vrsqrtx1, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
51 __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
52 __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
53 __m512 vsqrtx3 = _mm512_mul_ps(vrsqrtx3, vx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
54 __m512 vhalfrsqrtx3 = _mm512_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
55 __m512 vsqrtx4 = _mm512_mul_ps(vrsqrtx4, vx4); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
56 __m512 vhalfrsqrtx4 = _mm512_mul_ps(vrsqrtx4, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
[all …]
|
| D | avx512f-nr1fma1adj-x96.c | 45 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
46 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
47 __m512 vsqrtx1 = _mm512_mul_ps(vrsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
48 __m512 vhalfrsqrtx1 = _mm512_mul_ps(vrsqrtx1, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
49 __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
50 __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
51 __m512 vsqrtx3 = _mm512_mul_ps(vrsqrtx3, vx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
52 __m512 vhalfrsqrtx3 = _mm512_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
53 __m512 vsqrtx4 = _mm512_mul_ps(vrsqrtx4, vx4); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
54 __m512 vhalfrsqrtx4 = _mm512_mul_ps(vrsqrtx4, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
[all …]
|
| D | avx512f-nr1fma1adj-x80.c | 43 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
44 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
45 __m512 vsqrtx1 = _mm512_mul_ps(vrsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
46 __m512 vhalfrsqrtx1 = _mm512_mul_ps(vrsqrtx1, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
47 __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
48 __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
49 __m512 vsqrtx3 = _mm512_mul_ps(vrsqrtx3, vx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
50 __m512 vhalfrsqrtx3 = _mm512_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
51 __m512 vsqrtx4 = _mm512_mul_ps(vrsqrtx4, vx4); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
52 __m512 vhalfrsqrtx4 = _mm512_mul_ps(vrsqrtx4, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
[all …]
|
| D | avx512f-nr1fma1adj-x64.c | 41 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
42 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
43 __m512 vsqrtx1 = _mm512_mul_ps(vrsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
44 __m512 vhalfrsqrtx1 = _mm512_mul_ps(vrsqrtx1, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
45 __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
46 __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
47 __m512 vsqrtx3 = _mm512_mul_ps(vrsqrtx3, vx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
48 __m512 vhalfrsqrtx3 = _mm512_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
85 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
86 __m512 vhalfrsqrtx = _mm512_mul_ps(vrsqrtx, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
[all …]
|
| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-avx512f-rr1-lut16-p3-perm-x128.c | 53 const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
54 const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
55 const __m512 vz2 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx2, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
56 const __m512 vz3 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx3, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
57 const __m512 vz4 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx4, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
58 const __m512 vz5 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx5, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
59 const __m512 vz6 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx6, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
60 const __m512 vz7 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
123 vp0 = _mm512_mul_ps(vp0, vt0); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
124 vt0 = _mm512_mul_ps(vt0, vs0); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
[all …]
|
| D | velu-avx512f-rr1-lut16-p3-perm-x112.c | 52 const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
53 const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
54 const __m512 vz2 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx2, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
55 const __m512 vz3 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx3, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
56 const __m512 vz4 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx4, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
57 const __m512 vz5 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx5, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
58 const __m512 vz6 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx6, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
114 vp0 = _mm512_mul_ps(vp0, vt0); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
115 vt0 = _mm512_mul_ps(vt0, vs0); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
116 vp1 = _mm512_mul_ps(vp1, vt1); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
[all …]
|
| D | velu-avx512f-rr1-p6-x128.c | 53 const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
54 const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
55 const __m512 vz2 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx2, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
56 const __m512 vz3 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx3, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
57 const __m512 vz4 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx4, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
58 const __m512 vz5 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx5, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
59 const __m512 vz6 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx6, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
60 const __m512 vz7 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx7, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
133 vp0 = _mm512_mul_ps(vp0, vt0); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
134 vt0 = _mm512_mul_ps(vt0, vs0); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
[all …]
|
| D | velu-avx512f-rr1-lut16-p3-perm-x80.c | 50 const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
51 const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
52 const __m512 vz2 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx2, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
53 const __m512 vz3 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx3, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
54 const __m512 vz4 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx4, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
96 vp0 = _mm512_mul_ps(vp0, vt0); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
97 vt0 = _mm512_mul_ps(vt0, vs0); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
98 vp1 = _mm512_mul_ps(vp1, vt1); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
99 vt1 = _mm512_mul_ps(vt1, vs1); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
100 vp2 = _mm512_mul_ps(vp2, vt2); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
[all …]
|
| D | velu-avx512f-rr1-lut16-p3-perm-x96.c | 51 const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
52 const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
53 const __m512 vz2 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx2, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
54 const __m512 vz3 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx3, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
55 const __m512 vz4 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx4, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
56 const __m512 vz5 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx5, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
105 vp0 = _mm512_mul_ps(vp0, vt0); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
106 vt0 = _mm512_mul_ps(vt0, vs0); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
107 vp1 = _mm512_mul_ps(vp1, vt1); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
108 vt1 = _mm512_mul_ps(vt1, vs1); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
[all …]
|
| D | velu-avx512f-rr1-p6-x96.c | 51 const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
52 const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
53 const __m512 vz2 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx2, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
54 const __m512 vz3 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx3, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
55 const __m512 vz4 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx4, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
56 const __m512 vz5 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx5, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
113 vp0 = _mm512_mul_ps(vp0, vt0); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
114 vt0 = _mm512_mul_ps(vt0, vs0); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
115 vp1 = _mm512_mul_ps(vp1, vt1); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
116 vt1 = _mm512_mul_ps(vt1, vs1); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
[all …]
|
| D | velu-avx512f-rr1-p6-x112.c | 52 const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
53 const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
54 const __m512 vz2 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx2, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
55 const __m512 vz3 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx3, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
56 const __m512 vz4 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx4, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
57 const __m512 vz5 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx5, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
58 const __m512 vz6 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx6, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
123 vp0 = _mm512_mul_ps(vp0, vt0); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
124 vt0 = _mm512_mul_ps(vt0, vs0); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
125 vp1 = _mm512_mul_ps(vp1, vt1); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
[all …]
|
| D | velu-avx512f-rr1-p6-x80.c | 50 const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
51 const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
52 const __m512 vz2 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx2, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
53 const __m512 vz3 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx3, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
54 const __m512 vz4 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx4, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
103 vp0 = _mm512_mul_ps(vp0, vt0); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
104 vt0 = _mm512_mul_ps(vt0, vs0); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
105 vp1 = _mm512_mul_ps(vp1, vt1); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
106 vt1 = _mm512_mul_ps(vt1, vs1); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
107 vp2 = _mm512_mul_ps(vp2, vt2); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
[all …]
|
| D | velu-avx512f-rr1-p6-x64.c | 49 const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
50 const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
51 const __m512 vz2 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx2, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
52 const __m512 vz3 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx3, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
93 vp0 = _mm512_mul_ps(vp0, vt0); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
94 vt0 = _mm512_mul_ps(vt0, vs0); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
95 vp1 = _mm512_mul_ps(vp1, vt1); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
96 vt1 = _mm512_mul_ps(vt1, vs1); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
97 vp2 = _mm512_mul_ps(vp2, vt2); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
98 vt2 = _mm512_mul_ps(vt2, vs2); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
[all …]
|
| /external/XNNPACK/src/f32-vscaleexpminusmax/gen/ |
| D | avx512f-p5-scalef-x192.c | 72 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
73 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
74 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
75 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
76 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
77 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
78 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
79 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
80 __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
81 __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x176.c | 70 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
71 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
72 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
73 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
74 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
75 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
76 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
77 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
78 __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
79 __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 68 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
69 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
70 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
71 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
72 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
73 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
74 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
75 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
76 __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
77 __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 66 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
67 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
68 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
69 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
70 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
71 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
72 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
73 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
74 __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
163 vf0 = _mm512_mul_ps(vf0, vscale); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| D | avx512f-p5-scalef-x128.c | 64 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
65 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
66 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
67 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
68 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
69 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
70 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
71 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
152 vf0 = _mm512_mul_ps(vf0, vscale); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
153 vf1 = _mm512_mul_ps(vf1, vscale); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128()
[all …]
|
| /external/XNNPACK/src/f32-vscaleextexp/gen/ |
| D | avx512f-p5-scalef-x192.c | 59 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
60 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
61 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
62 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
63 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
64 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
65 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
66 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
67 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
68 const __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x176.c | 58 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
59 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
60 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
61 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
62 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
63 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
64 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
65 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
66 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
67 const __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 57 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
58 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
59 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
60 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
61 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
62 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
63 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
64 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
65 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
66 const __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 56 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
57 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
58 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
59 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
60 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
61 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
62 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
63 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
64 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
145 __m512 vf0 = _mm512_mul_ps(vp0, vscalev); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| D | avx512f-p5-scalef-x128.c | 55 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
56 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
57 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
58 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
59 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
60 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
61 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
62 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
136 __m512 vf0 = _mm512_mul_ps(vp0, vscalev); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
137 __m512 vf1 = _mm512_mul_ps(vp1, vscalev); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x128()
[all …]
|
| D | avx512f-p5-scalef-x112.c | 54 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
55 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
56 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
57 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
58 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
59 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
60 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
127 __m512 vf0 = _mm512_mul_ps(vp0, vscalev); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
128 __m512 vf1 = _mm512_mul_ps(vp1, vscalev); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
129 __m512 vf2 = _mm512_mul_ps(vp2, vscalev); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
[all …]
|