1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx-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 <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__avx_rr2_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__avx_rr2_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_hi = _mm256_set1_ps(-0x1.62E440p-1f);
36 const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.0105C6p-21f);
37 const __m256 vc6 = _mm256_set1_ps(0x1.6b7338p-10f);
38 const __m256 vc5 = _mm256_set1_ps(0x1.12278Ep-7f);
39 const __m256 vc4 = _mm256_set1_ps(0x1.555716p-5f);
40 const __m256 vc3 = _mm256_set1_ps(0x1.5554B0p-3f);
41 const __m256 vc2 = _mm256_set1_ps(0x1.FFFFFEp-2f);
42 const __m256 vone = _mm256_set1_ps(1.0f);
43
44 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
45 __m256 vx0 = _mm256_loadu_ps(x);
46 __m256 vx1 = _mm256_loadu_ps(x + 8);
47 x += 16;
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
52 __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
53 __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
54
55 const __m128 vs0_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)), 23));
56 const __m128 vs0_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, 1)), 23));
57 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
58 const __m128 vs1_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)), 23));
59 const __m128 vs1_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, 1)), 23));
60 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
61
62 __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
63 __m256 vs0 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs0_lo), vs0_hi, 1);
64 __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
65 __m256 vs1 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs1_lo), vs1_hi, 1);
66
67 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
68 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
69
70 __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc6, vt0), vc5);
71 __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc6, vt1), vc5);
72
73 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc4);
74 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc4);
75
76 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc3);
77 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc3);
78
79 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
80 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
81
82 vp0 = _mm256_mul_ps(vp0, vt0);
83 vp1 = _mm256_mul_ps(vp1, vt1);
84
85 vt0 = _mm256_mul_ps(vt0, vs0);
86 vs0 = _mm256_sub_ps(vs0, vone);
87 vt1 = _mm256_mul_ps(vt1, vs1);
88 vs1 = _mm256_sub_ps(vs1, vone);
89
90 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vt0);
91 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vt1);
92
93 const __m256 ve0 = _mm256_mul_ps(_mm256_add_ps(vp0, vs0), valpha);
94 vx0 = _mm256_mul_ps(vx0, vbeta);
95 const __m256 ve1 = _mm256_mul_ps(_mm256_add_ps(vp1, vs1), valpha);
96 vx1 = _mm256_mul_ps(vx1, vbeta);
97
98 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
99 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
100
101 _mm256_storeu_ps(y, vy0);
102 _mm256_storeu_ps(y + 8, vy1);
103 y += 16;
104 }
105 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
106 __m256 vx = _mm256_loadu_ps(x);
107 x += 8;
108
109 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
110
111 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
112 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
113 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
114 vn = _mm256_sub_ps(vn, vmagic_bias);
115
116 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
117 __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
118 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
119
120 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
121 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
122 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
123 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
124 vp = _mm256_mul_ps(vp, vt);
125
126 vt = _mm256_mul_ps(vt, vs);
127 vs = _mm256_sub_ps(vs, vone);
128 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
129
130 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
131 vx = _mm256_mul_ps(vx, vbeta);
132 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
133
134 _mm256_storeu_ps(y, vy);
135 y += 8;
136 }
137 if XNN_UNLIKELY(n != 0) {
138 assert(n >= 1 * sizeof(float));
139 assert(n <= 7 * sizeof(float));
140 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
141
142 __m256 vx = _mm256_maskload_ps(x, vmask);
143
144 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
145
146 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
147 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
148 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
149 vn = _mm256_sub_ps(vn, vmagic_bias);
150
151 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
152 __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
153 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
154
155 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
156 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
157 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
158 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
159 vp = _mm256_mul_ps(vp, vt);
160
161 vt = _mm256_mul_ps(vt, vs);
162 vs = _mm256_sub_ps(vs, vone);
163 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
164
165 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
166 vx = _mm256_mul_ps(vx, vbeta);
167 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
168
169 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
170 __m128 vy_lo = _mm256_castps256_ps128(vy);
171 if (n & (4 * sizeof(float))) {
172 _mm_storeu_ps(y, vy_lo);
173 vy_lo = _mm256_extractf128_ps(vy, 1);
174 y += 4;
175 }
176 if (n & (2 * sizeof(float))) {
177 _mm_storel_pi((__m64*) y, vy_lo);
178 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
179 y += 2;
180 }
181 if (n & (1 * sizeof(float))) {
182 _mm_store_ss(y, vy_lo);
183 }
184 }
185 }
186