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_x20(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_x20(
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 >= 20 * sizeof(float); n -= 20 * 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 __m128 vxGHIJ = _mm_loadu_ps(x + 16);
51 x += 20;
52
53 const __m128 vz0123 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx0123, vprescale));
54 const __m128 vz4567 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx4567, vprescale));
55 const __m128 vz89AB = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx89AB, vprescale));
56 const __m128 vzCDEF = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxCDEF, vprescale));
57 const __m128 vzGHIJ = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxGHIJ, vprescale));
58
59 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
60 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
61 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
62 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
63 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias);
64
65 __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
66 __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
67 __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23));
68 __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23));
69 __m128 vsGHIJ = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnGHIJ), 23));
70
71 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
72 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
73 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
74 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
75 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias);
76
77 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
78 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
79 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
80 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF);
81 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ);
82
83 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
84 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
85 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
86 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF);
87 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ);
88
89 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc6, vt0123), vc5);
90 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5);
91 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc6, vt89AB), vc5);
92 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc6, vtCDEF), vc5);
93 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc6, vtGHIJ), vc5);
94
95 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc4);
96 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4);
97 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc4);
98 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc4);
99 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc4);
100
101 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
102 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
103 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3);
104 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3);
105 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3);
106
107 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
108 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
109 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2);
110 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2);
111 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2);
112
113 vp0123 = _mm_mul_ps(vp0123, vt0123);
114 vp4567 = _mm_mul_ps(vp4567, vt4567);
115 vp89AB = _mm_mul_ps(vp89AB, vt89AB);
116 vpCDEF = _mm_mul_ps(vpCDEF, vtCDEF);
117 vpGHIJ = _mm_mul_ps(vpGHIJ, vtGHIJ);
118
119 vt0123 = _mm_mul_ps(vt0123, vs0123);
120 vs0123 = _mm_sub_ps(vs0123, vone);
121 vt4567 = _mm_mul_ps(vt4567, vs4567);
122 vs4567 = _mm_sub_ps(vs4567, vone);
123 vt89AB = _mm_mul_ps(vt89AB, vs89AB);
124 vs89AB = _mm_sub_ps(vs89AB, vone);
125 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF);
126 vsCDEF = _mm_sub_ps(vsCDEF, vone);
127 vtGHIJ = _mm_mul_ps(vtGHIJ, vsGHIJ);
128 vsGHIJ = _mm_sub_ps(vsGHIJ, vone);
129
130 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
131 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
132 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vt89AB);
133 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vtCDEF);
134 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vtGHIJ);
135
136 const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
137 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
138 const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha);
139 const __m128 veCDEF = _mm_mul_ps(_mm_add_ps(vpCDEF, vsCDEF), valpha);
140 const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha);
141
142 vx0123 = _mm_mul_ps(vx0123, vbeta);
143 vx4567 = _mm_mul_ps(vx4567, vbeta);
144 vx89AB = _mm_mul_ps(vx89AB, vbeta);
145 vxCDEF = _mm_mul_ps(vxCDEF, vbeta);
146 vxGHIJ = _mm_mul_ps(vxGHIJ, vbeta);
147
148 const __m128 vy0123 = _mm_blendv_ps(vx0123, ve0123, vx0123);
149 const __m128 vy4567 = _mm_blendv_ps(vx4567, ve4567, vx4567);
150 const __m128 vy89AB = _mm_blendv_ps(vx89AB, ve89AB, vx89AB);
151 const __m128 vyCDEF = _mm_blendv_ps(vxCDEF, veCDEF, vxCDEF);
152 const __m128 vyGHIJ = _mm_blendv_ps(vxGHIJ, veGHIJ, vxGHIJ);
153
154 _mm_storeu_ps(y, vy0123);
155 _mm_storeu_ps(y + 4, vy4567);
156 _mm_storeu_ps(y + 8, vy89AB);
157 _mm_storeu_ps(y + 12, vyCDEF);
158 _mm_storeu_ps(y + 16, vyGHIJ);
159 y += 20;
160 }
161 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
162 __m128 vx = _mm_loadu_ps(x);
163 x += 4;
164
165 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
166
167 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
168 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
169 vn = _mm_sub_ps(vn, vmagic_bias);
170
171 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
172 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
173
174 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
175 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
176 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
177 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
178 vp = _mm_mul_ps(vp, vt);
179
180 vt = _mm_mul_ps(vt, vs);
181 vs = _mm_sub_ps(vs, vone);
182 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
183 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
184
185 vx = _mm_mul_ps(vx, vbeta);
186 const __m128 vy = _mm_blendv_ps(vx, ve, vx);
187
188 _mm_storeu_ps(y, vy);
189 y += 4;
190 }
191 if XNN_UNLIKELY(n != 0) {
192 __m128 vx = _mm_loadu_ps(x);
193
194 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
195
196 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
197 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
198 vn = _mm_sub_ps(vn, vmagic_bias);
199
200 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
201 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
202
203 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
204 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
205 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
206 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
207 vp = _mm_mul_ps(vp, vt);
208
209 vt = _mm_mul_ps(vt, vs);
210 vs = _mm_sub_ps(vs, vone);
211 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
212 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
213
214 vx = _mm_mul_ps(vx, vbeta);
215 __m128 vy = _mm_blendv_ps(vx, ve, vx);
216
217 if (n & (2 * sizeof(float))) {
218 _mm_storel_pi((__m64*) y, vy);
219 vy = _mm_movehl_ps(vy, vy);
220 y += 2;
221 }
222 if (n & (1 * sizeof(float))) {
223 _mm_store_ss(y, vy);
224 }
225 }
226 }
227