1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/sse-rr2-lut64-p2-div.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 <smmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_64[64];
19
xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x4(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x4(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(n % sizeof(float) == 0);
27
28 const __m128 vsign_mask = _mm_load_ps(params->sse2_rr2_lut64_p2.sign_mask);
29 const __m128 vmagic_bias = _mm_load_ps(params->sse2_rr2_lut64_p2.magic_bias);
30 const __m128 vlog2e = _mm_load_ps(params->sse2_rr2_lut64_p2.log2e);
31 const __m128i vindex_mask = _mm_load_si128((const __m128i*) params->sse2_rr2_lut64_p2.index_mask);
32 const __m128 vminus_ln2_hi = _mm_load_ps(params->sse2_rr2_lut64_p2.minus_ln2_hi);
33 const __m128 vminus_ln2_lo = _mm_load_ps(params->sse2_rr2_lut64_p2.minus_ln2_lo);
34 const __m128 vc2 = _mm_load_ps(params->sse2_rr2_lut64_p2.c2);
35 const __m128 vone = _mm_load_ps(params->sse2_rr2_lut64_p2.one);
36 const __m128 vdenorm_cutoff = _mm_load_ps(params->sse2_rr2_lut64_p2.denorm_cutoff);
37
38 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
39 const __m128 vx = _mm_loadu_ps(x);
40 x += 4;
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 __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
46
47 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
48 #if XNN_ARCH_X86_64
49 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
50 const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
51 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
52 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
53 const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))), 1);
54 const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))), 1);
55 #else // !XNN_ARCH_X86_64
56 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
57 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
58 const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))), 1);
59 const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))), 1);
60 #endif // XNN_ARCH_X86_64
61 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
62
63 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
64 vn = _mm_sub_ps(vn, vmagic_bias);
65
66 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
67 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
68
69 __m128 vp = _mm_mul_ps(vt, vc2);
70 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
71
72 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
73
74 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
75 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
76 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
77
78 _mm_storeu_ps(y, vf);
79 y += 4;
80 }
81 if XNN_UNLIKELY(n != 0) {
82 const __m128 vx = _mm_loadu_ps(x);
83
84 const __m128 vz = _mm_or_ps(vx, vsign_mask);
85
86 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
87 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
88
89 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
90 #if XNN_ARCH_X86_64
91 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
92 const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
93 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
94 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
95 const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))), 1);
96 const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))), 1);
97 #else // !XNN_ARCH_X86_64
98 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
99 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
100 const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))), 1);
101 const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))), 1);
102 #endif // XNN_ARCH_X86_64
103 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
104
105 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
106 vn = _mm_sub_ps(vn, vmagic_bias);
107
108 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
109 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
110
111 __m128 vp = _mm_mul_ps(vt, vc2);
112 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
113
114 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
115
116 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
117 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
118 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
119
120 if (n & (2 * sizeof(float))) {
121 _mm_storel_pi((__m64*) y, vf);
122 vf = _mm_movehl_ps(vf, vf);
123 y += 2;
124 }
125 if (n & (1 * sizeof(float))) {
126 _mm_store_ss(y, vf);
127 }
128 }
129 }
130