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_div_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_div_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 vdenorm_cutoff = _mm256_load_ps(params->avx_rr2_p5.denorm_cutoff);
38
39 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
40 const __m256 vx = _mm256_loadu_ps(x);
41 x += 8;
42
43 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
44
45 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
46
47 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
48 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
49 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
50
51 vn = _mm256_sub_ps(vn, vmagic_bias);
52
53 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
54 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
55
56 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
57 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
58 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
59 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
60
61 vt = _mm256_mul_ps(vt, vs);
62 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
63
64 const __m256 vd = _mm256_add_ps(ve, vone);
65 __m256 vf = _mm256_div_ps(ve, vd);
66
67 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
68 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
69
70 _mm256_storeu_ps(y, vf);
71 y += 8;
72 }
73 if XNN_UNLIKELY(n != 0) {
74 assert(n >= 1 * sizeof(float));
75 assert(n <= 7 * sizeof(float));
76 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx_rr2_p5.mask_table[7] - n));
77
78 const __m256 vx = _mm256_maskload_ps(x, vmask);
79
80 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
81
82 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
83 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
84 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
85 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
86
87 vn = _mm256_sub_ps(vn, vmagic_bias);
88
89 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
90 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
91
92 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
93 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
94 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
95 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
96
97 vt = _mm256_mul_ps(vt, vs);
98 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
99
100 const __m256 vd = _mm256_add_ps(ve, vone);
101 __m256 vf = _mm256_div_ps(ve, vd);
102
103 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
104 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
105
106 __m128 vf_lo = _mm256_castps256_ps128(vf);
107 if (n & (4 * sizeof(float))) {
108 _mm_storeu_ps(y, vf_lo);
109 vf_lo = _mm256_extractf128_ps(vf, 1);
110 y += 4;
111 }
112 if (n & (2 * sizeof(float))) {
113 _mm_storel_pi((__m64*) y, vf_lo);
114 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
115 y += 2;
116 }
117 if (n & (1 * sizeof(float))) {
118 _mm_store_ss(y, vf_lo);
119 }
120 }
121 }
122