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_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__avx_rr2_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_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 >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
45 __m256 vx0 = _mm256_loadu_ps(x);
46 __m256 vx1 = _mm256_loadu_ps(x + 8);
47 __m256 vx2 = _mm256_loadu_ps(x + 16);
48 x += 24;
49
50 const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
51 const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
52 const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
53
54 __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
55 __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
56 __m256 vn2 = _mm256_add_ps(_mm256_mul_ps(vz2, vlog2e), vmagic_bias);
57
58 const __m128 vs0_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)), 23));
59 const __m128 vs0_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, 1)), 23));
60 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
61 const __m128 vs1_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)), 23));
62 const __m128 vs1_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, 1)), 23));
63 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
64 const __m128 vs2_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn2)), 23));
65 const __m128 vs2_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn2, 1)), 23));
66 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
67
68 __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
69 __m256 vs0 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs0_lo), vs0_hi, 1);
70 __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
71 __m256 vs1 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs1_lo), vs1_hi, 1);
72 __m256 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_hi), vz2);
73 __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs2_lo), vs2_hi, 1);
74
75 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
76 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
77 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_lo), vt2);
78
79 __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc6, vt0), vc5);
80 __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc6, vt1), vc5);
81 __m256 vp2 = _mm256_add_ps(_mm256_mul_ps(vc6, vt2), vc5);
82
83 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc4);
84 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc4);
85 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc4);
86
87 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc3);
88 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc3);
89 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc3);
90
91 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
92 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
93 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc2);
94
95 vp0 = _mm256_mul_ps(vp0, vt0);
96 vp1 = _mm256_mul_ps(vp1, vt1);
97 vp2 = _mm256_mul_ps(vp2, vt2);
98
99 vt0 = _mm256_mul_ps(vt0, vs0);
100 vs0 = _mm256_sub_ps(vs0, vone);
101 vt1 = _mm256_mul_ps(vt1, vs1);
102 vs1 = _mm256_sub_ps(vs1, vone);
103 vt2 = _mm256_mul_ps(vt2, vs2);
104 vs2 = _mm256_sub_ps(vs2, vone);
105
106 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vt0);
107 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vt1);
108 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vt2);
109
110 const __m256 ve0 = _mm256_mul_ps(_mm256_add_ps(vp0, vs0), valpha);
111 vx0 = _mm256_mul_ps(vx0, vbeta);
112 const __m256 ve1 = _mm256_mul_ps(_mm256_add_ps(vp1, vs1), valpha);
113 vx1 = _mm256_mul_ps(vx1, vbeta);
114 const __m256 ve2 = _mm256_mul_ps(_mm256_add_ps(vp2, vs2), valpha);
115 vx2 = _mm256_mul_ps(vx2, vbeta);
116
117 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
118 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
119 const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
120
121 _mm256_storeu_ps(y, vy0);
122 _mm256_storeu_ps(y + 8, vy1);
123 _mm256_storeu_ps(y + 16, vy2);
124 y += 24;
125 }
126 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
127 __m256 vx = _mm256_loadu_ps(x);
128 x += 8;
129
130 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
131
132 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
133 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
134 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
135 vn = _mm256_sub_ps(vn, vmagic_bias);
136
137 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
138 __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
139 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
140
141 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
142 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
143 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
144 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
145 vp = _mm256_mul_ps(vp, vt);
146
147 vt = _mm256_mul_ps(vt, vs);
148 vs = _mm256_sub_ps(vs, vone);
149 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
150
151 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
152 vx = _mm256_mul_ps(vx, vbeta);
153 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
154
155 _mm256_storeu_ps(y, vy);
156 y += 8;
157 }
158 if XNN_UNLIKELY(n != 0) {
159 assert(n >= 1 * sizeof(float));
160 assert(n <= 7 * sizeof(float));
161 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
162
163 __m256 vx = _mm256_maskload_ps(x, vmask);
164
165 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
166
167 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
168 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
169 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
170 vn = _mm256_sub_ps(vn, vmagic_bias);
171
172 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
173 __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
174 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
175
176 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
177 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
178 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
179 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
180 vp = _mm256_mul_ps(vp, vt);
181
182 vt = _mm256_mul_ps(vt, vs);
183 vs = _mm256_sub_ps(vs, vone);
184 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
185
186 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
187 vx = _mm256_mul_ps(vx, vbeta);
188 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
189
190 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
191 __m128 vy_lo = _mm256_castps256_ps128(vy);
192 if (n & (4 * sizeof(float))) {
193 _mm_storeu_ps(y, vy_lo);
194 vy_lo = _mm256_extractf128_ps(vy, 1);
195 y += 4;
196 }
197 if (n & (2 * sizeof(float))) {
198 _mm_storel_pi((__m64*) y, vy_lo);
199 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
200 y += 2;
201 }
202 if (n & (1 * sizeof(float))) {
203 _mm_store_ss(y, vy_lo);
204 }
205 }
206 }
207