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_rcp_x40(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_rcp_x40(
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 >= 40 * sizeof(uint16_t); batch -= 40 * 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 i += 40;
45
46 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
47 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
48 const __m256 vz2 = _mm256_or_ps(vx2, vsign_mask);
49 const __m256 vz3 = _mm256_or_ps(vx3, vsign_mask);
50 const __m256 vz4 = _mm256_or_ps(vx4, vsign_mask);
51
52 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
53 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
54 __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
55 __m256 vn3 = _mm256_fmadd_ps(vz3, vlog2e, vmagic_bias);
56 __m256 vn4 = _mm256_fmadd_ps(vz4, vlog2e, vmagic_bias);
57
58 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
59 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
60 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
61 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
62 const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
63
64 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
65 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
66 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
67 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
68 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
69
70 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
71 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
72 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
73 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2, vz3);
74 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2, vz4);
75
76 const __m256 vp0 = _mm256_fmadd_ps(vc2, vt0, vc1);
77 const __m256 vp1 = _mm256_fmadd_ps(vc2, vt1, vc1);
78 const __m256 vp2 = _mm256_fmadd_ps(vc2, vt2, vc1);
79 const __m256 vp3 = _mm256_fmadd_ps(vc2, vt3, vc1);
80 const __m256 vp4 = _mm256_fmadd_ps(vc2, vt4, vc1);
81
82 vt0 = _mm256_mul_ps(vt0, vs0);
83 vt1 = _mm256_mul_ps(vt1, vs1);
84 vt2 = _mm256_mul_ps(vt2, vs2);
85 vt3 = _mm256_mul_ps(vt3, vs3);
86 vt4 = _mm256_mul_ps(vt4, vs4);
87
88 const __m256 ve0 = _mm256_fmadd_ps(vt0, vp0, vs0);
89 const __m256 ve1 = _mm256_fmadd_ps(vt1, vp1, vs1);
90 const __m256 ve2 = _mm256_fmadd_ps(vt2, vp2, vs2);
91 const __m256 ve3 = _mm256_fmadd_ps(vt3, vp3, vs3);
92 const __m256 ve4 = _mm256_fmadd_ps(vt4, vp4, vs4);
93
94 const __m256 vd0 = _mm256_add_ps(ve0, vone);
95 const __m256 vd1 = _mm256_add_ps(ve1, vone);
96 const __m256 vd2 = _mm256_add_ps(ve2, vone);
97 const __m256 vd3 = _mm256_add_ps(ve3, vone);
98 const __m256 vd4 = _mm256_add_ps(ve4, vone);
99
100 const __m256 vr0 = _mm256_rcp_ps(vd0);
101 const __m256 vr1 = _mm256_rcp_ps(vd1);
102 const __m256 vr2 = _mm256_rcp_ps(vd2);
103 const __m256 vr3 = _mm256_rcp_ps(vd3);
104 const __m256 vr4 = _mm256_rcp_ps(vd4);
105
106 __m256 vf0 = _mm256_mul_ps(ve0, vr0);
107 __m256 vf1 = _mm256_mul_ps(ve1, vr1);
108 __m256 vf2 = _mm256_mul_ps(ve2, vr2);
109 __m256 vf3 = _mm256_mul_ps(ve3, vr3);
110 __m256 vf4 = _mm256_mul_ps(ve4, vr4);
111
112 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
113 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
114 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vz2, vdenorm_cutoff, _CMP_LT_OS), vf2);
115 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vz3, vdenorm_cutoff, _CMP_LT_OS), vf3);
116 vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vz4, vdenorm_cutoff, _CMP_LT_OS), vf4);
117
118 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
119 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
120 vf2 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf2), vf2, vx2);
121 vf3 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf3), vf3, vx3);
122 vf4 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf4), vf4, vx4);
123
124 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf0, _MM_FROUND_NO_EXC));
125 _mm_storeu_si128((__m128i*) (o + 8), _mm256_cvtps_ph(vf1, _MM_FROUND_NO_EXC));
126 _mm_storeu_si128((__m128i*) (o + 16), _mm256_cvtps_ph(vf2, _MM_FROUND_NO_EXC));
127 _mm_storeu_si128((__m128i*) (o + 24), _mm256_cvtps_ph(vf3, _MM_FROUND_NO_EXC));
128 _mm_storeu_si128((__m128i*) (o + 32), _mm256_cvtps_ph(vf4, _MM_FROUND_NO_EXC));
129 o += 40;
130 }
131 for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
132 const __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
133 i += 8;
134
135 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
136
137 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
138 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
139 vn = _mm256_sub_ps(vn, vmagic_bias);
140
141 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
142
143 const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
144 vt = _mm256_mul_ps(vt, vs);
145 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
146
147 const __m256 vd = _mm256_add_ps(ve, vone);
148 const __m256 vr = _mm256_rcp_ps(vd);
149 __m256 vf = _mm256_mul_ps(ve, vr);
150
151 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
152 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
153
154 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC));
155 o += 8;
156 }
157 if XNN_UNLIKELY(batch != 0) {
158 assert(batch >= 1 * sizeof(uint16_t));
159 assert(batch <= 7 * sizeof(uint16_t));
160 const __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
161
162 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
163
164 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
165 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
166 vn = _mm256_sub_ps(vn, vmagic_bias);
167
168 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
169
170 const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
171 vt = _mm256_mul_ps(vt, vs);
172 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
173
174 const __m256 vd = _mm256_add_ps(ve, vone);
175 const __m256 vr = _mm256_rcp_ps(vd);
176 __m256 vf = _mm256_mul_ps(ve, vr);
177
178 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
179 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
180
181 __m128i vh = _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC);
182 if (batch & (4 * sizeof(uint16_t))) {
183 _mm_storel_epi64((__m128i*) o, vh);
184 vh = _mm_unpackhi_epi64(vh, vh);
185 o += 4;
186 }
187 if (batch & (2 * sizeof(uint16_t))) {
188 _mm_storeu_si32(o, vh);
189 vh = _mm_srli_epi64(vh, 32);
190 o += 2;
191 }
192 if (batch & (1 * sizeof(uint16_t))) {
193 *o = (uint16_t) _mm_extract_epi16(vh, 0);
194 }
195 }
196 }
197