1 // Auto-generated file. Do not edit!
2 // Template: src/f16-vsigmoid/avx2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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/intrinsics-polyfill.h>
16 #include <xnnpack/vunary.h>
17
18
xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x24(size_t batch,const void * input,void * output,const union xnn_f16_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x24(
20 size_t batch,
21 const void* input,
22 void* output,
23 const union xnn_f16_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(batch % sizeof(uint16_t) == 0);
26
27 const __m256 vsign_mask = _mm256_load_ps(params->avx2_rr1_p2.sign_mask);
28 const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p2.magic_bias);
29 const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p2.log2e);
30 const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p2.minus_ln2);
31 const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p2.c2);
32 const __m256 vc1 = _mm256_load_ps(params->avx2_rr1_p2.c1);
33 const __m256 vone = _mm256_load_ps(params->avx2_rr1_p2.one);
34 const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx2_rr1_p2.denorm_cutoff);
35
36 const uint16_t* i = (const uint16_t*) input;
37 uint16_t* o = (uint16_t*) output;
38 for (; batch >= 24 * sizeof(uint16_t); batch -= 24 * sizeof(uint16_t)) {
39 const __m256 vx0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
40 const __m256 vx1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 8)));
41 const __m256 vx2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 16)));
42 i += 24;
43
44 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
45 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
46 const __m256 vz2 = _mm256_or_ps(vx2, vsign_mask);
47
48 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
49 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
50 __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
51
52 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
53 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
54 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
55
56 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
57 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
58 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
59
60 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
61 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
62 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
63
64 const __m256 vp0 = _mm256_fmadd_ps(vc2, vt0, vc1);
65 const __m256 vp1 = _mm256_fmadd_ps(vc2, vt1, vc1);
66 const __m256 vp2 = _mm256_fmadd_ps(vc2, vt2, vc1);
67
68 vt0 = _mm256_mul_ps(vt0, vs0);
69 vt1 = _mm256_mul_ps(vt1, vs1);
70 vt2 = _mm256_mul_ps(vt2, vs2);
71
72 const __m256 ve0 = _mm256_fmadd_ps(vt0, vp0, vs0);
73 const __m256 ve1 = _mm256_fmadd_ps(vt1, vp1, vs1);
74 const __m256 ve2 = _mm256_fmadd_ps(vt2, vp2, vs2);
75
76 const __m256 vd0 = _mm256_add_ps(ve0, vone);
77 const __m256 vd1 = _mm256_add_ps(ve1, vone);
78 const __m256 vd2 = _mm256_add_ps(ve2, vone);
79
80 __m256 vf0 = _mm256_div_ps(ve0, vd0);
81 __m256 vf1 = _mm256_div_ps(ve1, vd1);
82 __m256 vf2 = _mm256_div_ps(ve2, vd2);
83
84 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
85 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
86 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vz2, vdenorm_cutoff, _CMP_LT_OS), vf2);
87
88 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
89 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
90 vf2 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf2), vf2, vx2);
91
92 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf0, _MM_FROUND_NO_EXC));
93 _mm_storeu_si128((__m128i*) (o + 8), _mm256_cvtps_ph(vf1, _MM_FROUND_NO_EXC));
94 _mm_storeu_si128((__m128i*) (o + 16), _mm256_cvtps_ph(vf2, _MM_FROUND_NO_EXC));
95 o += 24;
96 }
97 for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
98 const __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
99 i += 8;
100
101 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
102
103 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
104 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
105 vn = _mm256_sub_ps(vn, vmagic_bias);
106
107 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
108
109 const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
110 vt = _mm256_mul_ps(vt, vs);
111 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
112
113 const __m256 vd = _mm256_add_ps(ve, vone);
114 __m256 vf = _mm256_div_ps(ve, vd);
115
116 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
117 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
118
119 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC));
120 o += 8;
121 }
122 if XNN_UNLIKELY(batch != 0) {
123 assert(batch >= 1 * sizeof(uint16_t));
124 assert(batch <= 7 * sizeof(uint16_t));
125 const __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
126
127 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
128
129 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
130 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
131 vn = _mm256_sub_ps(vn, vmagic_bias);
132
133 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
134
135 const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
136 vt = _mm256_mul_ps(vt, vs);
137 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
138
139 const __m256 vd = _mm256_add_ps(ve, vone);
140 __m256 vf = _mm256_div_ps(ve, vd);
141
142 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
143 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
144
145 __m128i vh = _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC);
146 if (batch & (4 * sizeof(uint16_t))) {
147 _mm_storel_epi64((__m128i*) o, vh);
148 vh = _mm_unpackhi_epi64(vh, vh);
149 o += 4;
150 }
151 if (batch & (2 * sizeof(uint16_t))) {
152 _mm_storeu_si32(o, vh);
153 vh = _mm_srli_epi64(vh, 32);
154 o += 2;
155 }
156 if (batch & (1 * sizeof(uint16_t))) {
157 *o = (uint16_t) _mm_extract_epi16(vh, 0);
158 }
159 }
160 }
161