1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/avx-rr2-p5.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x8(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x8(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n % sizeof(float) == 0);
25
26 const __m256 vsign_mask = _mm256_load_ps(params->avx_rr2_p5.sign_mask);
27 const __m256 vmagic_bias = _mm256_load_ps(params->avx_rr2_p5.magic_bias);
28 const __m256 vlog2e = _mm256_load_ps(params->avx_rr2_p5.log2e);
29 const __m256 vminus_ln2_hi = _mm256_load_ps(params->avx_rr2_p5.minus_ln2_hi);
30 const __m256 vminus_ln2_lo = _mm256_load_ps(params->avx_rr2_p5.minus_ln2_lo);
31 const __m256 vc5 = _mm256_load_ps(params->avx_rr2_p5.c5);
32 const __m256 vc4 = _mm256_load_ps(params->avx_rr2_p5.c4);
33 const __m256 vc3 = _mm256_load_ps(params->avx_rr2_p5.c3);
34 const __m256 vc2 = _mm256_load_ps(params->avx_rr2_p5.c2);
35 const __m256 vc1 = _mm256_load_ps(params->avx_rr2_p5.c1);
36 const __m256 vone = _mm256_load_ps(params->avx_rr2_p5.one);
37 const __m256 vtwo = _mm256_load_ps(params->avx_rr2_p5.two);
38 const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx_rr2_p5.denorm_cutoff);
39
40 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
41 const __m256 vx = _mm256_loadu_ps(x);
42 x += 8;
43
44 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
45
46 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
47
48 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
49 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
50 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
51
52 vn = _mm256_sub_ps(vn, vmagic_bias);
53
54 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
55 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
56
57 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
58 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
59 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
60 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
61
62 vt = _mm256_mul_ps(vt, vs);
63 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
64
65 const __m256 vd = _mm256_add_ps(ve, vone);
66 __m256 vr = _mm256_rcp_ps(vd);
67 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
68 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
69 __m256 vf = _mm256_mul_ps(ve, vr);
70
71 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
72 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
73
74 _mm256_storeu_ps(y, vf);
75 y += 8;
76 }
77 if XNN_UNLIKELY(n != 0) {
78 assert(n >= 1 * sizeof(float));
79 assert(n <= 7 * sizeof(float));
80 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx_rr2_p5.mask_table[7] - n));
81
82 const __m256 vx = _mm256_maskload_ps(x, vmask);
83
84 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
85
86 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
87 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
88 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
89 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
90
91 vn = _mm256_sub_ps(vn, vmagic_bias);
92
93 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
94 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
95
96 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
97 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
98 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
99 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
100
101 vt = _mm256_mul_ps(vt, vs);
102 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
103
104 const __m256 vd = _mm256_add_ps(ve, vone);
105 __m256 vr = _mm256_rcp_ps(vd);
106 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
107 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
108 __m256 vf = _mm256_mul_ps(ve, vr);
109
110 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
111 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
112
113 __m128 vf_lo = _mm256_castps256_ps128(vf);
114 if (n & (4 * sizeof(float))) {
115 _mm_storeu_ps(y, vf_lo);
116 vf_lo = _mm256_extractf128_ps(vf, 1);
117 y += 4;
118 }
119 if (n & (2 * sizeof(float))) {
120 _mm_storel_pi((__m64*) y, vf_lo);
121 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
122 y += 2;
123 }
124 if (n & (1 * sizeof(float))) {
125 _mm_store_ss(y, vf_lo);
126 }
127 }
128 }
129