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_x64(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_x64(
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 >= 64 * sizeof(uint16_t); batch -= 64 * 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 const __m256 vx3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 24)));
43 const __m256 vx4 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 32)));
44 const __m256 vx5 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 40)));
45 const __m256 vx6 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 48)));
46 const __m256 vx7 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 56)));
47 i += 64;
48
49 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
50 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
51 const __m256 vz2 = _mm256_or_ps(vx2, vsign_mask);
52 const __m256 vz3 = _mm256_or_ps(vx3, vsign_mask);
53 const __m256 vz4 = _mm256_or_ps(vx4, vsign_mask);
54 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask);
55 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask);
56 const __m256 vz7 = _mm256_or_ps(vx7, vsign_mask);
57
58 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
59 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
60 __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
61 __m256 vn3 = _mm256_fmadd_ps(vz3, vlog2e, vmagic_bias);
62 __m256 vn4 = _mm256_fmadd_ps(vz4, vlog2e, vmagic_bias);
63 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias);
64 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias);
65 __m256 vn7 = _mm256_fmadd_ps(vz7, vlog2e, vmagic_bias);
66
67 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
68 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
69 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
70 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
71 const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
72 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
73 const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
74 const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23));
75
76 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
77 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
78 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
79 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
80 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
81 vn5 = _mm256_sub_ps(vn5, vmagic_bias);
82 vn6 = _mm256_sub_ps(vn6, vmagic_bias);
83 vn7 = _mm256_sub_ps(vn7, vmagic_bias);
84
85 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
86 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
87 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
88 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2, vz3);
89 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2, vz4);
90 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5);
91 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6);
92 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2, vz7);
93
94 const __m256 vp0 = _mm256_fmadd_ps(vc2, vt0, vc1);
95 const __m256 vp1 = _mm256_fmadd_ps(vc2, vt1, vc1);
96 const __m256 vp2 = _mm256_fmadd_ps(vc2, vt2, vc1);
97 const __m256 vp3 = _mm256_fmadd_ps(vc2, vt3, vc1);
98 const __m256 vp4 = _mm256_fmadd_ps(vc2, vt4, vc1);
99 const __m256 vp5 = _mm256_fmadd_ps(vc2, vt5, vc1);
100 const __m256 vp6 = _mm256_fmadd_ps(vc2, vt6, vc1);
101 const __m256 vp7 = _mm256_fmadd_ps(vc2, vt7, vc1);
102
103 vt0 = _mm256_mul_ps(vt0, vs0);
104 vt1 = _mm256_mul_ps(vt1, vs1);
105 vt2 = _mm256_mul_ps(vt2, vs2);
106 vt3 = _mm256_mul_ps(vt3, vs3);
107 vt4 = _mm256_mul_ps(vt4, vs4);
108 vt5 = _mm256_mul_ps(vt5, vs5);
109 vt6 = _mm256_mul_ps(vt6, vs6);
110 vt7 = _mm256_mul_ps(vt7, vs7);
111
112 const __m256 ve0 = _mm256_fmadd_ps(vt0, vp0, vs0);
113 const __m256 ve1 = _mm256_fmadd_ps(vt1, vp1, vs1);
114 const __m256 ve2 = _mm256_fmadd_ps(vt2, vp2, vs2);
115 const __m256 ve3 = _mm256_fmadd_ps(vt3, vp3, vs3);
116 const __m256 ve4 = _mm256_fmadd_ps(vt4, vp4, vs4);
117 const __m256 ve5 = _mm256_fmadd_ps(vt5, vp5, vs5);
118 const __m256 ve6 = _mm256_fmadd_ps(vt6, vp6, vs6);
119 const __m256 ve7 = _mm256_fmadd_ps(vt7, vp7, vs7);
120
121 const __m256 vd0 = _mm256_add_ps(ve0, vone);
122 const __m256 vd1 = _mm256_add_ps(ve1, vone);
123 const __m256 vd2 = _mm256_add_ps(ve2, vone);
124 const __m256 vd3 = _mm256_add_ps(ve3, vone);
125 const __m256 vd4 = _mm256_add_ps(ve4, vone);
126 const __m256 vd5 = _mm256_add_ps(ve5, vone);
127 const __m256 vd6 = _mm256_add_ps(ve6, vone);
128 const __m256 vd7 = _mm256_add_ps(ve7, vone);
129
130 __m256 vf0 = _mm256_div_ps(ve0, vd0);
131 __m256 vf1 = _mm256_div_ps(ve1, vd1);
132 __m256 vf2 = _mm256_div_ps(ve2, vd2);
133 __m256 vf3 = _mm256_div_ps(ve3, vd3);
134 __m256 vf4 = _mm256_div_ps(ve4, vd4);
135 __m256 vf5 = _mm256_div_ps(ve5, vd5);
136 __m256 vf6 = _mm256_div_ps(ve6, vd6);
137 __m256 vf7 = _mm256_div_ps(ve7, vd7);
138
139 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
140 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
141 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vz2, vdenorm_cutoff, _CMP_LT_OS), vf2);
142 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vz3, vdenorm_cutoff, _CMP_LT_OS), vf3);
143 vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vz4, vdenorm_cutoff, _CMP_LT_OS), vf4);
144 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5);
145 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6);
146 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vz7, vdenorm_cutoff, _CMP_LT_OS), vf7);
147
148 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
149 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
150 vf2 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf2), vf2, vx2);
151 vf3 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf3), vf3, vx3);
152 vf4 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf4), vf4, vx4);
153 vf5 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf5), vf5, vx5);
154 vf6 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf6), vf6, vx6);
155 vf7 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf7), vf7, vx7);
156
157 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf0, _MM_FROUND_NO_EXC));
158 _mm_storeu_si128((__m128i*) (o + 8), _mm256_cvtps_ph(vf1, _MM_FROUND_NO_EXC));
159 _mm_storeu_si128((__m128i*) (o + 16), _mm256_cvtps_ph(vf2, _MM_FROUND_NO_EXC));
160 _mm_storeu_si128((__m128i*) (o + 24), _mm256_cvtps_ph(vf3, _MM_FROUND_NO_EXC));
161 _mm_storeu_si128((__m128i*) (o + 32), _mm256_cvtps_ph(vf4, _MM_FROUND_NO_EXC));
162 _mm_storeu_si128((__m128i*) (o + 40), _mm256_cvtps_ph(vf5, _MM_FROUND_NO_EXC));
163 _mm_storeu_si128((__m128i*) (o + 48), _mm256_cvtps_ph(vf6, _MM_FROUND_NO_EXC));
164 _mm_storeu_si128((__m128i*) (o + 56), _mm256_cvtps_ph(vf7, _MM_FROUND_NO_EXC));
165 o += 64;
166 }
167 for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
168 const __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
169 i += 8;
170
171 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
172
173 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
174 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
175 vn = _mm256_sub_ps(vn, vmagic_bias);
176
177 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
178
179 const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
180 vt = _mm256_mul_ps(vt, vs);
181 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
182
183 const __m256 vd = _mm256_add_ps(ve, vone);
184 __m256 vf = _mm256_div_ps(ve, vd);
185
186 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
187 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
188
189 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC));
190 o += 8;
191 }
192 if XNN_UNLIKELY(batch != 0) {
193 assert(batch >= 1 * sizeof(uint16_t));
194 assert(batch <= 7 * sizeof(uint16_t));
195 const __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
196
197 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
198
199 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
200 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
201 vn = _mm256_sub_ps(vn, vmagic_bias);
202
203 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
204
205 const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
206 vt = _mm256_mul_ps(vt, vs);
207 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
208
209 const __m256 vd = _mm256_add_ps(ve, vone);
210 __m256 vf = _mm256_div_ps(ve, vd);
211
212 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
213 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
214
215 __m128i vh = _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC);
216 if (batch & (4 * sizeof(uint16_t))) {
217 _mm_storel_epi64((__m128i*) o, vh);
218 vh = _mm_unpackhi_epi64(vh, vh);
219 o += 4;
220 }
221 if (batch & (2 * sizeof(uint16_t))) {
222 _mm_storeu_si32(o, vh);
223 vh = _mm_srli_epi64(vh, 32);
224 o += 2;
225 }
226 if (batch & (1 * sizeof(uint16_t))) {
227 *o = (uint16_t) _mm_extract_epi16(vh, 0);
228 }
229 }
230 }
231