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_x24(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_x24(
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 >= 24 * sizeof(float); n -= 24 * 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 __m128 vxKLMN = _mm_loadu_ps(x + 20);
52 x += 24;
53
54 const __m128 vz0123 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx0123, vprescale));
55 const __m128 vz4567 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx4567, vprescale));
56 const __m128 vz89AB = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx89AB, vprescale));
57 const __m128 vzCDEF = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxCDEF, vprescale));
58 const __m128 vzGHIJ = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxGHIJ, vprescale));
59 const __m128 vzKLMN = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxKLMN, vprescale));
60
61 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
62 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
63 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
64 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
65 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias);
66 __m128 vnKLMN = _mm_add_ps(_mm_mul_ps(vzKLMN, vlog2e), vmagic_bias);
67
68 __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
69 __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
70 __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23));
71 __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23));
72 __m128 vsGHIJ = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnGHIJ), 23));
73 __m128 vsKLMN = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnKLMN), 23));
74
75 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
76 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
77 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
78 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
79 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias);
80 vnKLMN = _mm_sub_ps(vnKLMN, vmagic_bias);
81
82 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
83 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
84 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
85 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF);
86 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ);
87 __m128 vtKLMN = _mm_add_ps(_mm_mul_ps(vnKLMN, vminus_ln2_hi), vzKLMN);
88
89 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
90 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
91 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
92 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF);
93 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ);
94 vtKLMN = _mm_add_ps(_mm_mul_ps(vnKLMN, vminus_ln2_lo), vtKLMN);
95
96 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc6, vt0123), vc5);
97 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5);
98 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc6, vt89AB), vc5);
99 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc6, vtCDEF), vc5);
100 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc6, vtGHIJ), vc5);
101 __m128 vpKLMN = _mm_add_ps(_mm_mul_ps(vc6, vtKLMN), vc5);
102
103 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc4);
104 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4);
105 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc4);
106 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc4);
107 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc4);
108 vpKLMN = _mm_add_ps(_mm_mul_ps(vpKLMN, vtKLMN), vc4);
109
110 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
111 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
112 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3);
113 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3);
114 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3);
115 vpKLMN = _mm_add_ps(_mm_mul_ps(vpKLMN, vtKLMN), vc3);
116
117 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
118 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
119 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2);
120 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2);
121 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2);
122 vpKLMN = _mm_add_ps(_mm_mul_ps(vpKLMN, vtKLMN), vc2);
123
124 vp0123 = _mm_mul_ps(vp0123, vt0123);
125 vp4567 = _mm_mul_ps(vp4567, vt4567);
126 vp89AB = _mm_mul_ps(vp89AB, vt89AB);
127 vpCDEF = _mm_mul_ps(vpCDEF, vtCDEF);
128 vpGHIJ = _mm_mul_ps(vpGHIJ, vtGHIJ);
129 vpKLMN = _mm_mul_ps(vpKLMN, vtKLMN);
130
131 vt0123 = _mm_mul_ps(vt0123, vs0123);
132 vs0123 = _mm_sub_ps(vs0123, vone);
133 vt4567 = _mm_mul_ps(vt4567, vs4567);
134 vs4567 = _mm_sub_ps(vs4567, vone);
135 vt89AB = _mm_mul_ps(vt89AB, vs89AB);
136 vs89AB = _mm_sub_ps(vs89AB, vone);
137 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF);
138 vsCDEF = _mm_sub_ps(vsCDEF, vone);
139 vtGHIJ = _mm_mul_ps(vtGHIJ, vsGHIJ);
140 vsGHIJ = _mm_sub_ps(vsGHIJ, vone);
141 vtKLMN = _mm_mul_ps(vtKLMN, vsKLMN);
142 vsKLMN = _mm_sub_ps(vsKLMN, vone);
143
144 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
145 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
146 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vt89AB);
147 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vtCDEF);
148 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vtGHIJ);
149 vpKLMN = _mm_add_ps(_mm_mul_ps(vpKLMN, vtKLMN), vtKLMN);
150
151 const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
152 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
153 const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha);
154 const __m128 veCDEF = _mm_mul_ps(_mm_add_ps(vpCDEF, vsCDEF), valpha);
155 const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha);
156 const __m128 veKLMN = _mm_mul_ps(_mm_add_ps(vpKLMN, vsKLMN), valpha);
157
158 vx0123 = _mm_mul_ps(vx0123, vbeta);
159 vx4567 = _mm_mul_ps(vx4567, vbeta);
160 vx89AB = _mm_mul_ps(vx89AB, vbeta);
161 vxCDEF = _mm_mul_ps(vxCDEF, vbeta);
162 vxGHIJ = _mm_mul_ps(vxGHIJ, vbeta);
163 vxKLMN = _mm_mul_ps(vxKLMN, vbeta);
164
165 const __m128 vy0123 = _mm_blendv_ps(vx0123, ve0123, vx0123);
166 const __m128 vy4567 = _mm_blendv_ps(vx4567, ve4567, vx4567);
167 const __m128 vy89AB = _mm_blendv_ps(vx89AB, ve89AB, vx89AB);
168 const __m128 vyCDEF = _mm_blendv_ps(vxCDEF, veCDEF, vxCDEF);
169 const __m128 vyGHIJ = _mm_blendv_ps(vxGHIJ, veGHIJ, vxGHIJ);
170 const __m128 vyKLMN = _mm_blendv_ps(vxKLMN, veKLMN, vxKLMN);
171
172 _mm_storeu_ps(y, vy0123);
173 _mm_storeu_ps(y + 4, vy4567);
174 _mm_storeu_ps(y + 8, vy89AB);
175 _mm_storeu_ps(y + 12, vyCDEF);
176 _mm_storeu_ps(y + 16, vyGHIJ);
177 _mm_storeu_ps(y + 20, vyKLMN);
178 y += 24;
179 }
180 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
181 __m128 vx = _mm_loadu_ps(x);
182 x += 4;
183
184 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
185
186 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
187 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
188 vn = _mm_sub_ps(vn, vmagic_bias);
189
190 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
191 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
192
193 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
194 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
195 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
196 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
197 vp = _mm_mul_ps(vp, vt);
198
199 vt = _mm_mul_ps(vt, vs);
200 vs = _mm_sub_ps(vs, vone);
201 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
202 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
203
204 vx = _mm_mul_ps(vx, vbeta);
205 const __m128 vy = _mm_blendv_ps(vx, ve, vx);
206
207 _mm_storeu_ps(y, vy);
208 y += 4;
209 }
210 if XNN_UNLIKELY(n != 0) {
211 __m128 vx = _mm_loadu_ps(x);
212
213 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
214
215 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
216 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
217 vn = _mm_sub_ps(vn, vmagic_bias);
218
219 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
220 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
221
222 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
223 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
224 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
225 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
226 vp = _mm_mul_ps(vp, vt);
227
228 vt = _mm_mul_ps(vt, vs);
229 vs = _mm_sub_ps(vs, vone);
230 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
231 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
232
233 vx = _mm_mul_ps(vx, vbeta);
234 __m128 vy = _mm_blendv_ps(vx, ve, vx);
235
236 if (n & (2 * sizeof(float))) {
237 _mm_storel_pi((__m64*) y, vy);
238 vy = _mm_movehl_ps(vy, vy);
239 y += 2;
240 }
241 if (n & (1 * sizeof(float))) {
242 _mm_store_ss(y, vy);
243 }
244 }
245 }
246