1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx-rr2-lut4-p4-perm.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_lut4_p4_perm_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_lut4_p4_perm_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(0x1.8003F8p21f);
34 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
35 const __m256 vindex_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x3));
36 const __m256 vtable = _mm256_set_ps(
37 0x1.AE89FAp+0f, 0x1.6A09E6p+0f, 0x1.306FE0p+0f, 0x1.000000p+0f,
38 0x1.AE89FAp+0f, 0x1.6A09E6p+0f, 0x1.306FE0p+0f, 0x1.000000p+0f);
39 const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E400p-1f);
40 const __m256 vminus_ln2_lo = _mm256_set1_ps(-0x1.7F7D1Cp-20f);
41 const __m256 vc4 = _mm256_set1_ps(0x1.554F9Ap-5f);
42 const __m256 vc3 = _mm256_set1_ps(0x1.557082p-3f);
43 const __m256 vc2 = _mm256_set1_ps(0x1.000002p-1f);
44 const __m256 vone = _mm256_set1_ps(1.0f);
45
46 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
47 __m256 vx0 = _mm256_loadu_ps(x);
48 __m256 vx1 = _mm256_loadu_ps(x + 8);
49 x += 16;
50
51 const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
52 const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, 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
57 __m256 ven0 = _mm256_andnot_ps(vindex_mask, vn0);
58 const __m256 vl0 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn0));
59 const __m128 ven0_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven0)), 21));
60 __m256 ven1 = _mm256_andnot_ps(vindex_mask, vn1);
61 const __m256 vl1 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn1));
62 const __m128 ven1_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven1)), 21));
63
64 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
65 const __m128 ven0_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven0, 1)), 21));
66 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
67 const __m128 ven1_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven1, 1)), 21));
68
69 __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
70 ven0 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven0_lo), ven0_hi, 1);
71 __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
72 ven1 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven1_lo), ven1_hi, 1);
73
74 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
75 __m256 vs0 = _mm256_mul_ps(vl0, ven0);
76 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
77 __m256 vs1 = _mm256_mul_ps(vl1, ven1);
78
79 __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc4, vt0), vc3);
80 __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc4, vt1), vc3);
81
82 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
83 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
84
85 vp0 = _mm256_mul_ps(vp0, vt0);
86 vp1 = _mm256_mul_ps(vp1, vt1);
87
88 vt0 = _mm256_mul_ps(vt0, vs0);
89 vs0 = _mm256_sub_ps(vs0, vone);
90 vt1 = _mm256_mul_ps(vt1, vs1);
91 vs1 = _mm256_sub_ps(vs1, vone);
92
93 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vt0);
94 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vt1);
95
96 const __m256 ve0 = _mm256_mul_ps(_mm256_add_ps(vp0, vs0), valpha);
97 vx0 = _mm256_mul_ps(vx0, vbeta);
98 const __m256 ve1 = _mm256_mul_ps(_mm256_add_ps(vp1, vs1), valpha);
99 vx1 = _mm256_mul_ps(vx1, vbeta);
100
101 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
102 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
103
104 _mm256_storeu_ps(y, vy0);
105 _mm256_storeu_ps(y + 8, vy1);
106 y += 16;
107 }
108 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
109 __m256 vx = _mm256_loadu_ps(x);
110 x += 8;
111
112 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
113
114 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
115 __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
116 const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
117 const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
118 vn = _mm256_sub_ps(vn, vmagic_bias);
119 const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
120
121 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
122 ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
123 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
124 __m256 vs = _mm256_mul_ps(vl, ven);
125
126 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
127 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
128 vp = _mm256_mul_ps(vp, vt);
129
130 vt = _mm256_mul_ps(vt, vs);
131 vs = _mm256_sub_ps(vs, vone);
132 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
133
134 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
135 vx = _mm256_mul_ps(vx, vbeta);
136 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
137
138 _mm256_storeu_ps(y, vy);
139 y += 8;
140 }
141 if XNN_UNLIKELY(n != 0) {
142 assert(n >= 1 * sizeof(float));
143 assert(n <= 7 * sizeof(float));
144 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
145
146 __m256 vx = _mm256_maskload_ps(x, vmask);
147
148 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
149
150 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
151 __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
152 const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
153 const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
154 vn = _mm256_sub_ps(vn, vmagic_bias);
155 const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
156
157 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
158 ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
159 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
160 __m256 vs = _mm256_mul_ps(vl, ven);
161
162 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
163 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
164 vp = _mm256_mul_ps(vp, vt);
165
166 vt = _mm256_mul_ps(vt, vs);
167 vs = _mm256_sub_ps(vs, vone);
168 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
169
170 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
171 vx = _mm256_mul_ps(vx, vbeta);
172 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
173
174 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
175 __m128 vy_lo = _mm256_castps256_ps128(vy);
176 if (n & (4 * sizeof(float))) {
177 _mm_storeu_ps(y, vy_lo);
178 vy_lo = _mm256_extractf128_ps(vy, 1);
179 y += 4;
180 }
181 if (n & (2 * sizeof(float))) {
182 _mm_storel_pi((__m64*) y, vy_lo);
183 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
184 y += 2;
185 }
186 if (n & (1 * sizeof(float))) {
187 _mm_store_ss(y, vy_lo);
188 }
189 }
190 }
191