1 // Auto-generated file. Do not edit!
2 // Template: src/f32-sigmoid/avx2-p5.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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__avx2_rr1_p5_nr2fma_x8(size_t n,const float * x,float * y,const void * params)20 void xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_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 = _mm256_set1_ps(-0x1.62E43p-1f);
32 const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
33 const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
34 const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
35 const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
36 const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
37 const __m256 vone = _mm256_set1_ps(1.0f);
38 const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep+6f);
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_fmadd_ps(vz, vlog2e, vmagic_bias);
47 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
48 vn = _mm256_sub_ps(vn, vmagic_bias);
49
50 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
51
52 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
53 vp = _mm256_fmadd_ps(vp, vt, vc3);
54 vp = _mm256_fmadd_ps(vp, vt, vc2);
55 vp = _mm256_fmadd_ps(vp, vt, vc1);
56
57 vt = _mm256_mul_ps(vt, vs);
58 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
59
60 const __m256 vd = _mm256_add_ps(ve, vone);
61 __m256 vr = _mm256_rcp_ps(vd);
62 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
63 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
64 __m256 vf = _mm256_mul_ps(ve, vr);
65
66 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
67 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
68
69 _mm256_storeu_ps(y, vf);
70 y += 8;
71 }
72 if XNN_UNLIKELY(n != 0) {
73 assert(n >= 1 * sizeof(float));
74 assert(n <= 7 * sizeof(float));
75 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
76
77 const __m256 vx = _mm256_maskload_ps(x, vmask);
78
79 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
80
81 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
82 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
83 vn = _mm256_sub_ps(vn, vmagic_bias);
84
85 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
86
87 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
88 vp = _mm256_fmadd_ps(vp, vt, vc3);
89 vp = _mm256_fmadd_ps(vp, vt, vc2);
90 vp = _mm256_fmadd_ps(vp, vt, vc1);
91
92 vt = _mm256_mul_ps(vt, vs);
93 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
94
95 const __m256 vd = _mm256_add_ps(ve, vone);
96 __m256 vr = _mm256_rcp_ps(vd);
97 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
98 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
99 __m256 vf = _mm256_mul_ps(ve, vr);
100
101 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
102 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
103
104 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
105 __m128 vf_lo = _mm256_castps256_ps128(vf);
106 if (n & (4 * sizeof(float))) {
107 _mm_storeu_ps(y, vf_lo);
108 vf_lo = _mm256_extractf128_ps(vf, 1);
109 y += 4;
110 }
111 if (n & (2 * sizeof(float))) {
112 _mm_storel_pi((__m64*) y, vf_lo);
113 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
114 y += 2;
115 }
116 if (n & (1 * sizeof(float))) {
117 _mm_store_ss(y, vf_lo);
118 }
119 }
120 }
121