1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx2-rr1-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 <immintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
18 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19
xnn_f32_velu_ukernel__avx2_rr1_p6_x16(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_velu_ukernel__avx2_rr1_p6_x16(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26 assert(n % sizeof(float) == 0);
27
28 const __m256 vprescale = _mm256_broadcast_ps((const __m128*) params->sse.prescale);
29 const __m256 valpha = _mm256_broadcast_ps((const __m128*) params->sse.alpha);
30 const __m256 vbeta = _mm256_broadcast_ps((const __m128*) params->sse.beta);
31
32 const __m256 vsat_cutoff = _mm256_set1_ps(-0x1.154246p+4f);
33 const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
34 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
35 const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
36 const __m256 vc6 = _mm256_set1_ps(0x1.6b7338p-10f);
37 const __m256 vc5 = _mm256_set1_ps(0x1.12278Ep-7f);
38 const __m256 vc4 = _mm256_set1_ps(0x1.555716p-5f);
39 const __m256 vc3 = _mm256_set1_ps(0x1.5554B0p-3f);
40 const __m256 vc2 = _mm256_set1_ps(0x1.FFFFFEp-2f);
41
42 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
43 __m256 vx0 = _mm256_loadu_ps(x);
44 __m256 vx1 = _mm256_loadu_ps(x + 8);
45 x += 16;
46
47 const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
48 const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
49
50 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
51 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
52
53 __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
54 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
55 __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
56 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
57
58 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
59 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
60
61 __m256 vp0 = _mm256_fmadd_ps(vc6, vt0, vc5);
62 __m256 vp1 = _mm256_fmadd_ps(vc6, vt1, vc5);
63
64 vp0 = _mm256_fmadd_ps(vp0, vt0, vc4);
65 vp1 = _mm256_fmadd_ps(vp1, vt1, vc4);
66
67 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
68 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
69
70 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
71 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
72
73 vp0 = _mm256_mul_ps(vp0, vt0);
74 vt0 = _mm256_mul_ps(vt0, vs0);
75 vp1 = _mm256_mul_ps(vp1, vt1);
76 vt1 = _mm256_mul_ps(vt1, vs1);
77
78 vs0 = _mm256_fmsub_ps(vs0, valpha, valpha);
79 vp0 = _mm256_fmadd_ps(vp0, vt0, vt0);
80 vs1 = _mm256_fmsub_ps(vs1, valpha, valpha);
81 vp1 = _mm256_fmadd_ps(vp1, vt1, vt1);
82
83 const __m256 ve0 = _mm256_fmadd_ps(vp0, valpha, vs0);
84 vx0 = _mm256_mul_ps(vx0, vbeta);
85 const __m256 ve1 = _mm256_fmadd_ps(vp1, valpha, vs1);
86 vx1 = _mm256_mul_ps(vx1, vbeta);
87
88 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
89 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
90
91 _mm256_storeu_ps(y, vy0);
92 _mm256_storeu_ps(y + 8, vy1);
93 y += 16;
94 }
95 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
96 __m256 vx = _mm256_loadu_ps(x);
97 x += 8;
98
99 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
100
101 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
102 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
103 vn = _mm256_sub_ps(vn, vmagic_bias);
104
105 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
106
107 __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
108 vp = _mm256_fmadd_ps(vp, vt, vc4);
109 vp = _mm256_fmadd_ps(vp, vt, vc3);
110 vp = _mm256_fmadd_ps(vp, vt, vc2);
111 vp = _mm256_mul_ps(vp, vt);
112
113 vt = _mm256_mul_ps(vt, vs);
114 vs = _mm256_fmsub_ps(vs, valpha, valpha);
115 vp = _mm256_fmadd_ps(vp, vt, vt);
116 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
117
118 vx = _mm256_mul_ps(vx, vbeta);
119 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
120
121 _mm256_storeu_ps(y, vy);
122 y += 8;
123 }
124 if XNN_UNLIKELY(n != 0) {
125 assert(n >= 1 * sizeof(float));
126 assert(n <= 7 * sizeof(float));
127 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
128
129 __m256 vx = _mm256_maskload_ps(x, vmask);
130
131 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
132
133 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
134 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
135 vn = _mm256_sub_ps(vn, vmagic_bias);
136
137 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
138
139 __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
140 vp = _mm256_fmadd_ps(vp, vt, vc4);
141 vp = _mm256_fmadd_ps(vp, vt, vc3);
142 vp = _mm256_fmadd_ps(vp, vt, vc2);
143 vp = _mm256_mul_ps(vp, vt);
144
145 vt = _mm256_mul_ps(vt, vs);
146 vs = _mm256_fmsub_ps(vs, valpha, valpha);
147 vp = _mm256_fmadd_ps(vp, vt, vt);
148 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
149
150 vx = _mm256_mul_ps(vx, vbeta);
151 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
152
153 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
154 __m128 vy_lo = _mm256_castps256_ps128(vy);
155 if (n & (4 * sizeof(float))) {
156 _mm_storeu_ps(y, vy_lo);
157 vy_lo = _mm256_extractf128_ps(vy, 1);
158 y += 4;
159 }
160 if (n & (2 * sizeof(float))) {
161 _mm_storel_pi((__m64*) y, vy_lo);
162 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
163 y += 2;
164 }
165 if (n & (1 * sizeof(float))) {
166 _mm_store_ss(y, vy_lo);
167 }
168 }
169 }
170