1 // Auto-generated file. Do not edit!
2 // Template: src/f32-sigmoid/sse-p5-div.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 <emmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_sigmoid_ukernel__sse2_p5_div_x4(size_t n,const float * x,float * y,const void * params)18 void xnn_f32_sigmoid_ukernel__sse2_p5_div_x4(
19 size_t n,
20 const float* x,
21 float* y,
22 const void* params) XNN_DISABLE_TSAN
23 {
24 assert(n % sizeof(float) == 0);
25
26 const __m128 vsign_mask = _mm_set1_ps(-0.0f);
27 const __m128 vmagic_bias = _mm_set1_ps(0x1.8000FEp23f);
28 const __m128 vlog2e = _mm_set1_ps(0x1.715476p0f);
29 const __m128 vminus_ln2_hi = _mm_set1_ps(-0x1.62E400p-1f);
30 const __m128 vminus_ln2_lo = _mm_set1_ps(-0x1.7F7D1Cp-20f);
31 const __m128 vc5 = _mm_set1_ps(0x1.0F9F9Cp-7f);
32 const __m128 vc4 = _mm_set1_ps(0x1.573A1Ap-5f);
33 const __m128 vc3 = _mm_set1_ps(0x1.555A80p-3f);
34 const __m128 vc2 = _mm_set1_ps(0x1.FFFDC6p-2f);
35 const __m128 vc1 = _mm_set1_ps(0x1.FFFFF6p-1f);
36 const __m128 vone = _mm_set1_ps(1.0f);
37 const __m128 vdenorm_cutoff = _mm_set1_ps(-0x1.5D589Ep+6f);
38
39 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
40 const __m128 vx = _mm_loadu_ps(x);
41
42 const __m128 vz = _mm_or_ps(vx, vsign_mask);
43
44 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
45 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
46 vn = _mm_sub_ps(vn, vmagic_bias);
47
48 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
49 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
50
51 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
52 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
53 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
54 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
55
56 vt = _mm_mul_ps(vt, vs);
57 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
58
59 __m128 vd = _mm_add_ps(ve, vone);
60 __m128 vf = _mm_div_ps(ve, vd);
61
62 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
63 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
64 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
65
66 _mm_storeu_ps(y, vf);
67
68 x += 4;
69 y += 4;
70 }
71 if XNN_UNLIKELY(n != 0) {
72 const __m128 vx = _mm_loadu_ps(x);
73
74 const __m128 vz = _mm_or_ps(vx, vsign_mask);
75
76 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
77 const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
78 vn = _mm_sub_ps(vn, vmagic_bias);
79
80 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
81 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
82
83 __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
84 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
85 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
86 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
87
88 vt = _mm_mul_ps(vt, vs);
89 __m128 ve = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
90
91 __m128 vd = _mm_add_ps(ve, vone);
92 __m128 vf = _mm_div_ps(ve, vd);
93
94 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
95 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
96 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
97
98 if (n & (2 * sizeof(float))) {
99 _mm_storel_pi((__m64*) y, vf);
100 vf = _mm_movehl_ps(vf, vf);
101 y += 2;
102 }
103 if (n & (1 * sizeof(float))) {
104 _mm_store_ss(y, vf);
105 }
106 }
107 }
108