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_x32(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_x32(
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 >= 32 * sizeof(float); n -= 32 * 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 __m256 vx3 = _mm256_loadu_ps(x + 24);
47 x += 32;
48
49 const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
50 const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
51 const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
52 const __m256 vz3 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx3, vprescale));
53
54 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
55 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
56 __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
57 __m256 vn3 = _mm256_fmadd_ps(vz3, vlog2e, vmagic_bias);
58
59 __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
60 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
61 __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
62 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
63 __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
64 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
65 __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
66 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
67
68 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
69 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
70 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
71 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2, vz3);
72
73 __m256 vp0 = _mm256_fmadd_ps(vc6, vt0, vc5);
74 __m256 vp1 = _mm256_fmadd_ps(vc6, vt1, vc5);
75 __m256 vp2 = _mm256_fmadd_ps(vc6, vt2, vc5);
76 __m256 vp3 = _mm256_fmadd_ps(vc6, vt3, vc5);
77
78 vp0 = _mm256_fmadd_ps(vp0, vt0, vc4);
79 vp1 = _mm256_fmadd_ps(vp1, vt1, vc4);
80 vp2 = _mm256_fmadd_ps(vp2, vt2, vc4);
81 vp3 = _mm256_fmadd_ps(vp3, vt3, vc4);
82
83 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
84 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
85 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
86 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
87
88 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
89 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
90 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
91 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
92
93 vp0 = _mm256_mul_ps(vp0, vt0);
94 vt0 = _mm256_mul_ps(vt0, vs0);
95 vp1 = _mm256_mul_ps(vp1, vt1);
96 vt1 = _mm256_mul_ps(vt1, vs1);
97 vp2 = _mm256_mul_ps(vp2, vt2);
98 vt2 = _mm256_mul_ps(vt2, vs2);
99 vp3 = _mm256_mul_ps(vp3, vt3);
100 vt3 = _mm256_mul_ps(vt3, vs3);
101
102 vs0 = _mm256_fmsub_ps(vs0, valpha, valpha);
103 vp0 = _mm256_fmadd_ps(vp0, vt0, vt0);
104 vs1 = _mm256_fmsub_ps(vs1, valpha, valpha);
105 vp1 = _mm256_fmadd_ps(vp1, vt1, vt1);
106 vs2 = _mm256_fmsub_ps(vs2, valpha, valpha);
107 vp2 = _mm256_fmadd_ps(vp2, vt2, vt2);
108 vs3 = _mm256_fmsub_ps(vs3, valpha, valpha);
109 vp3 = _mm256_fmadd_ps(vp3, vt3, vt3);
110
111 const __m256 ve0 = _mm256_fmadd_ps(vp0, valpha, vs0);
112 vx0 = _mm256_mul_ps(vx0, vbeta);
113 const __m256 ve1 = _mm256_fmadd_ps(vp1, valpha, vs1);
114 vx1 = _mm256_mul_ps(vx1, vbeta);
115 const __m256 ve2 = _mm256_fmadd_ps(vp2, valpha, vs2);
116 vx2 = _mm256_mul_ps(vx2, vbeta);
117 const __m256 ve3 = _mm256_fmadd_ps(vp3, valpha, vs3);
118 vx3 = _mm256_mul_ps(vx3, vbeta);
119
120 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
121 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
122 const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
123 const __m256 vy3 = _mm256_blendv_ps(vx3, ve3, vx3);
124
125 _mm256_storeu_ps(y, vy0);
126 _mm256_storeu_ps(y + 8, vy1);
127 _mm256_storeu_ps(y + 16, vy2);
128 _mm256_storeu_ps(y + 24, vy3);
129 y += 32;
130 }
131 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
132 __m256 vx = _mm256_loadu_ps(x);
133 x += 8;
134
135 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
136
137 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
138 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
139 vn = _mm256_sub_ps(vn, vmagic_bias);
140
141 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
142
143 __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
144 vp = _mm256_fmadd_ps(vp, vt, vc4);
145 vp = _mm256_fmadd_ps(vp, vt, vc3);
146 vp = _mm256_fmadd_ps(vp, vt, vc2);
147 vp = _mm256_mul_ps(vp, vt);
148
149 vt = _mm256_mul_ps(vt, vs);
150 vs = _mm256_fmsub_ps(vs, valpha, valpha);
151 vp = _mm256_fmadd_ps(vp, vt, vt);
152 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
153
154 vx = _mm256_mul_ps(vx, vbeta);
155 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
156
157 _mm256_storeu_ps(y, vy);
158 y += 8;
159 }
160 if XNN_UNLIKELY(n != 0) {
161 assert(n >= 1 * sizeof(float));
162 assert(n <= 7 * sizeof(float));
163 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
164
165 __m256 vx = _mm256_maskload_ps(x, vmask);
166
167 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
168
169 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
170 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
171 vn = _mm256_sub_ps(vn, vmagic_bias);
172
173 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
174
175 __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
176 vp = _mm256_fmadd_ps(vp, vt, vc4);
177 vp = _mm256_fmadd_ps(vp, vt, vc3);
178 vp = _mm256_fmadd_ps(vp, vt, vc2);
179 vp = _mm256_mul_ps(vp, vt);
180
181 vt = _mm256_mul_ps(vt, vs);
182 vs = _mm256_fmsub_ps(vs, valpha, valpha);
183 vp = _mm256_fmadd_ps(vp, vt, vt);
184 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
185
186 vx = _mm256_mul_ps(vx, vbeta);
187 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
188
189 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
190 __m128 vy_lo = _mm256_castps256_ps128(vy);
191 if (n & (4 * sizeof(float))) {
192 _mm_storeu_ps(y, vy_lo);
193 vy_lo = _mm256_extractf128_ps(vy, 1);
194 y += 4;
195 }
196 if (n & (2 * sizeof(float))) {
197 _mm_storel_pi((__m64*) y, vy_lo);
198 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
199 y += 2;
200 }
201 if (n & (1 * sizeof(float))) {
202 _mm_store_ss(y, vy_lo);
203 }
204 }
205 }
206