1 // Auto-generated file. Do not edit!
2 // Template: src/f32-sigmoid/avx-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
18 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19
xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x8(size_t n,const float * x,float * y,const void * params)20 void xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x8(
21 size_t n,
22 const float* x,
23 float* y,
24 const void* params)
25 {
26 assert(n % sizeof(float) == 0);
27
28 const __m256 vsign_mask = _mm256_set1_ps(-0.0f);
29 const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
30 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p0f);
31 const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
32 const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
33 const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
34 const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
35 const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
36 const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
37 const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
38 const __m256 vone = _mm256_set1_ps(1.0f);
39 const __m256 vtwo = _mm256_set1_ps(2.0f);
40 const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep+6f);
41
42 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
43 const __m256 vx = _mm256_loadu_ps(x);
44 x += 8;
45
46 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
47
48 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
49
50 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
51 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
52 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
53
54 vn = _mm256_sub_ps(vn, vmagic_bias);
55
56 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
57 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
58
59 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
60 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
61 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
62 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
63
64 vt = _mm256_mul_ps(vt, vs);
65 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
66
67 const __m256 vd = _mm256_add_ps(ve, vone);
68 __m256 vr = _mm256_rcp_ps(vd);
69 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
70 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
71 __m256 vf = _mm256_mul_ps(ve, vr);
72
73 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
74 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
75
76 _mm256_storeu_ps(y, vf);
77 y += 8;
78 }
79 if XNN_UNLIKELY(n != 0) {
80 assert(n >= 1 * sizeof(float));
81 assert(n <= 7 * sizeof(float));
82 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
83
84 const __m256 vx = _mm256_maskload_ps(x, vmask);
85
86 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
87
88 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
89 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
90 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
91 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
92
93 vn = _mm256_sub_ps(vn, vmagic_bias);
94
95 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
96 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
97
98 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
99 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
100 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
101 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
102
103 vt = _mm256_mul_ps(vt, vs);
104 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
105
106 const __m256 vd = _mm256_add_ps(ve, vone);
107 __m256 vr = _mm256_rcp_ps(vd);
108 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
109 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
110 __m256 vf = _mm256_mul_ps(ve, vr);
111
112 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
113 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
114
115 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
116 __m128 vf_lo = _mm256_castps256_ps128(vf);
117 if (n & (4 * sizeof(float))) {
118 _mm_storeu_ps(y, vf_lo);
119 vf_lo = _mm256_extractf128_ps(vf, 1);
120 y += 4;
121 }
122 if (n & (2 * sizeof(float))) {
123 _mm_storel_pi((__m64*) y, vf_lo);
124 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
125 y += 2;
126 }
127 if (n & (1 * sizeof(float))) {
128 _mm_store_ss(y, vf_lo);
129 }
130 }
131 }
132