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_x24(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_x24(
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 >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
43 __m256 vx0 = _mm256_loadu_ps(x);
44 __m256 vx1 = _mm256_loadu_ps(x + 8);
45 __m256 vx2 = _mm256_loadu_ps(x + 16);
46 x += 24;
47
48 const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
49 const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
50 const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
51
52 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
53 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
54 __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
55
56 __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
57 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
58 __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
59 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
60 __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
61 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
62
63 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
64 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
65 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
66
67 __m256 vp0 = _mm256_fmadd_ps(vc6, vt0, vc5);
68 __m256 vp1 = _mm256_fmadd_ps(vc6, vt1, vc5);
69 __m256 vp2 = _mm256_fmadd_ps(vc6, vt2, vc5);
70
71 vp0 = _mm256_fmadd_ps(vp0, vt0, vc4);
72 vp1 = _mm256_fmadd_ps(vp1, vt1, vc4);
73 vp2 = _mm256_fmadd_ps(vp2, vt2, vc4);
74
75 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
76 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
77 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
78
79 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
80 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
81 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
82
83 vp0 = _mm256_mul_ps(vp0, vt0);
84 vt0 = _mm256_mul_ps(vt0, vs0);
85 vp1 = _mm256_mul_ps(vp1, vt1);
86 vt1 = _mm256_mul_ps(vt1, vs1);
87 vp2 = _mm256_mul_ps(vp2, vt2);
88 vt2 = _mm256_mul_ps(vt2, vs2);
89
90 vs0 = _mm256_fmsub_ps(vs0, valpha, valpha);
91 vp0 = _mm256_fmadd_ps(vp0, vt0, vt0);
92 vs1 = _mm256_fmsub_ps(vs1, valpha, valpha);
93 vp1 = _mm256_fmadd_ps(vp1, vt1, vt1);
94 vs2 = _mm256_fmsub_ps(vs2, valpha, valpha);
95 vp2 = _mm256_fmadd_ps(vp2, vt2, vt2);
96
97 const __m256 ve0 = _mm256_fmadd_ps(vp0, valpha, vs0);
98 vx0 = _mm256_mul_ps(vx0, vbeta);
99 const __m256 ve1 = _mm256_fmadd_ps(vp1, valpha, vs1);
100 vx1 = _mm256_mul_ps(vx1, vbeta);
101 const __m256 ve2 = _mm256_fmadd_ps(vp2, valpha, vs2);
102 vx2 = _mm256_mul_ps(vx2, vbeta);
103
104 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
105 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
106 const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
107
108 _mm256_storeu_ps(y, vy0);
109 _mm256_storeu_ps(y + 8, vy1);
110 _mm256_storeu_ps(y + 16, vy2);
111 y += 24;
112 }
113 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
114 __m256 vx = _mm256_loadu_ps(x);
115 x += 8;
116
117 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
118
119 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
120 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
121 vn = _mm256_sub_ps(vn, vmagic_bias);
122
123 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
124
125 __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
126 vp = _mm256_fmadd_ps(vp, vt, vc4);
127 vp = _mm256_fmadd_ps(vp, vt, vc3);
128 vp = _mm256_fmadd_ps(vp, vt, vc2);
129 vp = _mm256_mul_ps(vp, vt);
130
131 vt = _mm256_mul_ps(vt, vs);
132 vs = _mm256_fmsub_ps(vs, valpha, valpha);
133 vp = _mm256_fmadd_ps(vp, vt, vt);
134 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
135
136 vx = _mm256_mul_ps(vx, vbeta);
137 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
138
139 _mm256_storeu_ps(y, vy);
140 y += 8;
141 }
142 if XNN_UNLIKELY(n != 0) {
143 assert(n >= 1 * sizeof(float));
144 assert(n <= 7 * sizeof(float));
145 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
146
147 __m256 vx = _mm256_maskload_ps(x, vmask);
148
149 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
150
151 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
152 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
153 vn = _mm256_sub_ps(vn, vmagic_bias);
154
155 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
156
157 __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
158 vp = _mm256_fmadd_ps(vp, vt, vc4);
159 vp = _mm256_fmadd_ps(vp, vt, vc3);
160 vp = _mm256_fmadd_ps(vp, vt, vc2);
161 vp = _mm256_mul_ps(vp, vt);
162
163 vt = _mm256_mul_ps(vt, vs);
164 vs = _mm256_fmsub_ps(vs, valpha, valpha);
165 vp = _mm256_fmadd_ps(vp, vt, vt);
166 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
167
168 vx = _mm256_mul_ps(vx, vbeta);
169 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
170
171 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
172 __m128 vy_lo = _mm256_castps256_ps128(vy);
173 if (n & (4 * sizeof(float))) {
174 _mm_storeu_ps(y, vy_lo);
175 vy_lo = _mm256_extractf128_ps(vy, 1);
176 y += 4;
177 }
178 if (n & (2 * sizeof(float))) {
179 _mm_storel_pi((__m64*) y, vy_lo);
180 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
181 y += 2;
182 }
183 if (n & (1 * sizeof(float))) {
184 _mm_store_ss(y, vy_lo);
185 }
186 }
187 }
188