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_x8(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_x8(
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
46 __m128 vx0123 = _mm_loadu_ps(x);
47 __m128 vx4567 = _mm_loadu_ps(x + 4);
48 x += 8;
49
50 const __m128 vz0123 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx0123, vprescale));
51 const __m128 vz4567 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx4567, vprescale));
52
53 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
54 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
55
56 __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
57 __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
58
59 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
60 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
61
62 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
63 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
64
65 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
66 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
67
68 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc6, vt0123), vc5);
69 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5);
70
71 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc4);
72 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4);
73
74 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
75 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
76
77 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
78 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
79
80 vp0123 = _mm_mul_ps(vp0123, vt0123);
81 vp4567 = _mm_mul_ps(vp4567, vt4567);
82
83 vt0123 = _mm_mul_ps(vt0123, vs0123);
84 vs0123 = _mm_sub_ps(vs0123, vone);
85 vt4567 = _mm_mul_ps(vt4567, vs4567);
86 vs4567 = _mm_sub_ps(vs4567, vone);
87
88 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
89 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
90
91 const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
92 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
93
94 vx0123 = _mm_mul_ps(vx0123, vbeta);
95 vx4567 = _mm_mul_ps(vx4567, vbeta);
96
97 const __m128 vy0123 = _mm_blendv_ps(vx0123, ve0123, vx0123);
98 const __m128 vy4567 = _mm_blendv_ps(vx4567, ve4567, vx4567);
99
100 _mm_storeu_ps(y, vy0123);
101 _mm_storeu_ps(y + 4, vy4567);
102 y += 8;
103 }
104 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
105 __m128 vx = _mm_loadu_ps(x);
106 x += 4;
107
108 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
109
110 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
111 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
112 vn = _mm_sub_ps(vn, vmagic_bias);
113
114 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
115 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
116
117 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
118 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
119 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
120 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
121 vp = _mm_mul_ps(vp, vt);
122
123 vt = _mm_mul_ps(vt, vs);
124 vs = _mm_sub_ps(vs, vone);
125 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
126 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
127
128 vx = _mm_mul_ps(vx, vbeta);
129 const __m128 vy = _mm_blendv_ps(vx, ve, vx);
130
131 _mm_storeu_ps(y, vy);
132 y += 4;
133 }
134 if XNN_UNLIKELY(n != 0) {
135 __m128 vx = _mm_loadu_ps(x);
136
137 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
138
139 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
140 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
141 vn = _mm_sub_ps(vn, vmagic_bias);
142
143 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
144 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
145
146 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
147 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
148 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
149 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
150 vp = _mm_mul_ps(vp, vt);
151
152 vt = _mm_mul_ps(vt, vs);
153 vs = _mm_sub_ps(vs, vone);
154 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
155 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
156
157 vx = _mm_mul_ps(vx, vbeta);
158 __m128 vy = _mm_blendv_ps(vx, ve, vx);
159
160 if (n & (2 * sizeof(float))) {
161 _mm_storel_pi((__m64*) y, vy);
162 vy = _mm_movehl_ps(vy, vy);
163 y += 2;
164 }
165 if (n & (1 * sizeof(float))) {
166 _mm_store_ss(y, vy);
167 }
168 }
169 }
170