1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/sse-rr2-p6.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/vunary.h>
15 #include <xnnpack/common.h>
16
17
xnn_f32_velu_ukernel__sse41_rr2_p6_x16(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_velu_ukernel__sse41_rr2_p6_x16(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const __m128 vprescale = _mm_load_ps(params->sse.prescale);
30 const __m128 valpha = _mm_load_ps(params->sse.alpha);
31 const __m128 vbeta = _mm_load_ps(params->sse.beta);
32
33 const __m128 vsat_cutoff = _mm_set1_ps(-0x1.154246p+4f);
34 const __m128 vmagic_bias = _mm_set1_ps(0x1.8000FEp23f);
35 const __m128 vlog2e = _mm_set1_ps(0x1.715476p+0f);
36 const __m128 vminus_ln2_hi = _mm_set1_ps(-0x1.62E440p-1f);
37 const __m128 vminus_ln2_lo = _mm_set1_ps(0x1.0105C6p-21f);
38 const __m128 vc6 = _mm_set1_ps(0x1.6b7338p-10f);
39 const __m128 vc5 = _mm_set1_ps(0x1.12278Ep-7f);
40 const __m128 vc4 = _mm_set1_ps(0x1.555716p-5f);
41 const __m128 vc3 = _mm_set1_ps(0x1.5554B0p-3f);
42 const __m128 vc2 = _mm_set1_ps(0x1.FFFFFEp-2f);
43 const __m128 vone = _mm_set1_ps(1.0f);
44
45 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
46 __m128 vx0123 = _mm_loadu_ps(x);
47 __m128 vx4567 = _mm_loadu_ps(x + 4);
48 __m128 vx89AB = _mm_loadu_ps(x + 8);
49 __m128 vxCDEF = _mm_loadu_ps(x + 12);
50 x += 16;
51
52 const __m128 vz0123 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx0123, vprescale));
53 const __m128 vz4567 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx4567, vprescale));
54 const __m128 vz89AB = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx89AB, vprescale));
55 const __m128 vzCDEF = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxCDEF, vprescale));
56
57 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
58 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
59 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
60 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
61
62 __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
63 __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
64 __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23));
65 __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23));
66
67 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
68 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
69 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
70 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
71
72 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
73 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
74 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
75 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF);
76
77 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
78 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
79 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
80 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF);
81
82 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc6, vt0123), vc5);
83 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5);
84 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc6, vt89AB), vc5);
85 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc6, vtCDEF), vc5);
86
87 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc4);
88 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4);
89 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc4);
90 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc4);
91
92 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
93 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
94 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3);
95 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3);
96
97 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
98 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
99 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2);
100 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2);
101
102 vp0123 = _mm_mul_ps(vp0123, vt0123);
103 vp4567 = _mm_mul_ps(vp4567, vt4567);
104 vp89AB = _mm_mul_ps(vp89AB, vt89AB);
105 vpCDEF = _mm_mul_ps(vpCDEF, vtCDEF);
106
107 vt0123 = _mm_mul_ps(vt0123, vs0123);
108 vs0123 = _mm_sub_ps(vs0123, vone);
109 vt4567 = _mm_mul_ps(vt4567, vs4567);
110 vs4567 = _mm_sub_ps(vs4567, vone);
111 vt89AB = _mm_mul_ps(vt89AB, vs89AB);
112 vs89AB = _mm_sub_ps(vs89AB, vone);
113 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF);
114 vsCDEF = _mm_sub_ps(vsCDEF, vone);
115
116 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
117 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
118 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vt89AB);
119 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vtCDEF);
120
121 const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
122 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
123 const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha);
124 const __m128 veCDEF = _mm_mul_ps(_mm_add_ps(vpCDEF, vsCDEF), valpha);
125
126 vx0123 = _mm_mul_ps(vx0123, vbeta);
127 vx4567 = _mm_mul_ps(vx4567, vbeta);
128 vx89AB = _mm_mul_ps(vx89AB, vbeta);
129 vxCDEF = _mm_mul_ps(vxCDEF, vbeta);
130
131 const __m128 vy0123 = _mm_blendv_ps(vx0123, ve0123, vx0123);
132 const __m128 vy4567 = _mm_blendv_ps(vx4567, ve4567, vx4567);
133 const __m128 vy89AB = _mm_blendv_ps(vx89AB, ve89AB, vx89AB);
134 const __m128 vyCDEF = _mm_blendv_ps(vxCDEF, veCDEF, vxCDEF);
135
136 _mm_storeu_ps(y, vy0123);
137 _mm_storeu_ps(y + 4, vy4567);
138 _mm_storeu_ps(y + 8, vy89AB);
139 _mm_storeu_ps(y + 12, vyCDEF);
140 y += 16;
141 }
142 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
143 __m128 vx = _mm_loadu_ps(x);
144 x += 4;
145
146 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
147
148 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
149 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
150 vn = _mm_sub_ps(vn, vmagic_bias);
151
152 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
153 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
154
155 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
156 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
157 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
158 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
159 vp = _mm_mul_ps(vp, vt);
160
161 vt = _mm_mul_ps(vt, vs);
162 vs = _mm_sub_ps(vs, vone);
163 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
164 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
165
166 vx = _mm_mul_ps(vx, vbeta);
167 const __m128 vy = _mm_blendv_ps(vx, ve, vx);
168
169 _mm_storeu_ps(y, vy);
170 y += 4;
171 }
172 if XNN_UNLIKELY(n != 0) {
173 __m128 vx = _mm_loadu_ps(x);
174
175 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
176
177 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
178 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
179 vn = _mm_sub_ps(vn, vmagic_bias);
180
181 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
182 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
183
184 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
185 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
186 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
187 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
188 vp = _mm_mul_ps(vp, vt);
189
190 vt = _mm_mul_ps(vt, vs);
191 vs = _mm_sub_ps(vs, vone);
192 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
193 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
194
195 vx = _mm_mul_ps(vx, vbeta);
196 __m128 vy = _mm_blendv_ps(vx, ve, vx);
197
198 if (n & (2 * sizeof(float))) {
199 _mm_storel_pi((__m64*) y, vy);
200 vy = _mm_movehl_ps(vy, vy);
201 y += 2;
202 }
203 if (n & (1 * sizeof(float))) {
204 _mm_store_ss(y, vy);
205 }
206 }
207 }
208