1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx512f-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/intrinsics-polyfill.h>
16 #include <xnnpack/vunary.h>
17
18
xnn_f32_velu_ukernel__avx512f_rr1_p6_x96(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_velu_ukernel__avx512f_rr1_p6_x96(
20 size_t n,
21 const float* x,
22 float* y,
23 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27
28 const __m512 vprescale = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.prescale));
29 const __m512 valpha = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.alpha));
30 const __m512 vbeta = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.beta));
31
32 const __m512 vsat_cutoff = _mm512_set1_ps(-0x1.154246p+4f);
33 const __m512 vmagic_bias = _mm512_set1_ps(0x1.8000FEp23f);
34 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
35 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f);
36 const __m512 vc6 = _mm512_set1_ps(0x1.6b7338p-10f);
37 const __m512 vc5 = _mm512_set1_ps(0x1.12278Ep-7f);
38 const __m512 vc4 = _mm512_set1_ps(0x1.555716p-5f);
39 const __m512 vc3 = _mm512_set1_ps(0x1.5554B0p-3f);
40 const __m512 vc2 = _mm512_set1_ps(0x1.FFFFFEp-2f);
41
42 for (; n >= 96 * sizeof(float); n -= 96 * sizeof(float)) {
43 __m512 vx0 = _mm512_loadu_ps(x);
44 __m512 vx1 = _mm512_loadu_ps(x + 16);
45 __m512 vx2 = _mm512_loadu_ps(x + 32);
46 __m512 vx3 = _mm512_loadu_ps(x + 48);
47 __m512 vx4 = _mm512_loadu_ps(x + 64);
48 __m512 vx5 = _mm512_loadu_ps(x + 80);
49 x += 96;
50
51 const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale));
52 const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale));
53 const __m512 vz2 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx2, vprescale));
54 const __m512 vz3 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx3, vprescale));
55 const __m512 vz4 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx4, vprescale));
56 const __m512 vz5 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx5, vprescale));
57
58 __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
59 __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
60 __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
61 __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
62 __m512 vn4 = _mm512_fmadd_ps(vz4, vlog2e, vmagic_bias);
63 __m512 vn5 = _mm512_fmadd_ps(vz5, vlog2e, vmagic_bias);
64
65 __m512 vs0 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn0), 23));
66 vn0 = _mm512_sub_ps(vn0, vmagic_bias);
67 __m512 vs1 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn1), 23));
68 vn1 = _mm512_sub_ps(vn1, vmagic_bias);
69 __m512 vs2 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn2), 23));
70 vn2 = _mm512_sub_ps(vn2, vmagic_bias);
71 __m512 vs3 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn3), 23));
72 vn3 = _mm512_sub_ps(vn3, vmagic_bias);
73 __m512 vs4 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn4), 23));
74 vn4 = _mm512_sub_ps(vn4, vmagic_bias);
75 __m512 vs5 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn5), 23));
76 vn5 = _mm512_sub_ps(vn5, vmagic_bias);
77
78 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
79 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
80 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
81 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
82 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4);
83 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5);
84
85 __m512 vp0 = _mm512_fmadd_ps(vc6, vt0, vc5);
86 __m512 vp1 = _mm512_fmadd_ps(vc6, vt1, vc5);
87 __m512 vp2 = _mm512_fmadd_ps(vc6, vt2, vc5);
88 __m512 vp3 = _mm512_fmadd_ps(vc6, vt3, vc5);
89 __m512 vp4 = _mm512_fmadd_ps(vc6, vt4, vc5);
90 __m512 vp5 = _mm512_fmadd_ps(vc6, vt5, vc5);
91
92 vp0 = _mm512_fmadd_ps(vp0, vt0, vc4);
93 vp1 = _mm512_fmadd_ps(vp1, vt1, vc4);
94 vp2 = _mm512_fmadd_ps(vp2, vt2, vc4);
95 vp3 = _mm512_fmadd_ps(vp3, vt3, vc4);
96 vp4 = _mm512_fmadd_ps(vp4, vt4, vc4);
97 vp5 = _mm512_fmadd_ps(vp5, vt5, vc4);
98
99 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
100 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
101 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
102 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
103 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
104 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
105
106 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
107 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
108 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
109 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
110 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
111 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
112
113 vp0 = _mm512_mul_ps(vp0, vt0);
114 vt0 = _mm512_mul_ps(vt0, vs0);
115 vp1 = _mm512_mul_ps(vp1, vt1);
116 vt1 = _mm512_mul_ps(vt1, vs1);
117 vp2 = _mm512_mul_ps(vp2, vt2);
118 vt2 = _mm512_mul_ps(vt2, vs2);
119 vp3 = _mm512_mul_ps(vp3, vt3);
120 vt3 = _mm512_mul_ps(vt3, vs3);
121 vp4 = _mm512_mul_ps(vp4, vt4);
122 vt4 = _mm512_mul_ps(vt4, vs4);
123 vp5 = _mm512_mul_ps(vp5, vt5);
124 vt5 = _mm512_mul_ps(vt5, vs5);
125
126 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha);
127 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha);
128 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha);
129 vs3 = _mm512_fmsub_ps(vs3, valpha, valpha);
130 vs4 = _mm512_fmsub_ps(vs4, valpha, valpha);
131 vs5 = _mm512_fmsub_ps(vs5, valpha, valpha);
132
133 vp0 = _mm512_fmadd_ps(vp0, vt0, vt0);
134 vp1 = _mm512_fmadd_ps(vp1, vt1, vt1);
135 vp2 = _mm512_fmadd_ps(vp2, vt2, vt2);
136 vp3 = _mm512_fmadd_ps(vp3, vt3, vt3);
137 vp4 = _mm512_fmadd_ps(vp4, vt4, vt4);
138 vp5 = _mm512_fmadd_ps(vp5, vt5, vt5);
139
140 const __m512 vzero = _mm512_setzero_ps();
141 __m512 vy0 = _mm512_fmadd_ps(vp0, valpha, vs0);
142 const __mmask16 vsign0 = _mm512_cmp_ps_mask(vx0, vzero, _CMP_NLT_US);
143 __m512 vy1 = _mm512_fmadd_ps(vp1, valpha, vs1);
144 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US);
145 __m512 vy2 = _mm512_fmadd_ps(vp2, valpha, vs2);
146 const __mmask16 vsign2 = _mm512_cmp_ps_mask(vx2, vzero, _CMP_NLT_US);
147 __m512 vy3 = _mm512_fmadd_ps(vp3, valpha, vs3);
148 const __mmask16 vsign3 = _mm512_cmp_ps_mask(vx3, vzero, _CMP_NLT_US);
149 __m512 vy4 = _mm512_fmadd_ps(vp4, valpha, vs4);
150 const __mmask16 vsign4 = _mm512_cmp_ps_mask(vx4, vzero, _CMP_NLT_US);
151 __m512 vy5 = _mm512_fmadd_ps(vp5, valpha, vs5);
152 const __mmask16 vsign5 = _mm512_cmp_ps_mask(vx5, vzero, _CMP_NLT_US);
153
154 vy0 = _mm512_mask_mul_ps(vy0, vsign0, vx0, vbeta);
155 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta);
156 vy2 = _mm512_mask_mul_ps(vy2, vsign2, vx2, vbeta);
157 vy3 = _mm512_mask_mul_ps(vy3, vsign3, vx3, vbeta);
158 vy4 = _mm512_mask_mul_ps(vy4, vsign4, vx4, vbeta);
159 vy5 = _mm512_mask_mul_ps(vy5, vsign5, vx5, vbeta);
160
161 _mm512_storeu_ps(y, vy0);
162 _mm512_storeu_ps(y + 16, vy1);
163 _mm512_storeu_ps(y + 32, vy2);
164 _mm512_storeu_ps(y + 48, vy3);
165 _mm512_storeu_ps(y + 64, vy4);
166 _mm512_storeu_ps(y + 80, vy5);
167 y += 96;
168 }
169 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
170 __m512 vx = _mm512_loadu_ps(x);
171 x += 16;
172
173 const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
174 const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
175
176 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
177 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23));
178 vn = _mm512_sub_ps(vn, vmagic_bias);
179
180 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
181
182 __m512 vp = _mm512_fmadd_ps(vc6, vt, vc5);
183 vp = _mm512_fmadd_ps(vp, vt, vc4);
184 vp = _mm512_fmadd_ps(vp, vt, vc3);
185 vp = _mm512_fmadd_ps(vp, vt, vc2);
186 vp = _mm512_mul_ps(vp, vt);
187
188 vt = _mm512_mul_ps(vt, vs);
189 vs = _mm512_fmsub_ps(vs, valpha, valpha);
190 vp = _mm512_fmadd_ps(vp, vt, vt);
191 __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
192
193 vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
194
195 _mm512_storeu_ps(y, vy);
196 y += 16;
197 }
198 if XNN_UNLIKELY(n != 0) {
199 assert(n >= 1 * sizeof(float));
200 assert(n <= 15 * sizeof(float));
201 // Prepare mask for valid 32-bit elements (depends on n).
202 n >>= 2 /* log2(sizeof(float)) */;
203 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
204
205 __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
206
207 const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
208 const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
209
210 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
211 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23));
212 vn = _mm512_sub_ps(vn, vmagic_bias);
213
214 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
215
216 __m512 vp = _mm512_fmadd_ps(vc6, vt, vc5);
217 vp = _mm512_fmadd_ps(vp, vt, vc4);
218 vp = _mm512_fmadd_ps(vp, vt, vc3);
219 vp = _mm512_fmadd_ps(vp, vt, vc2);
220 vp = _mm512_mul_ps(vp, vt);
221
222 vt = _mm512_mul_ps(vt, vs);
223 vs = _mm512_fmsub_ps(vs, valpha, valpha);
224 vp = _mm512_fmadd_ps(vp, vt, vt);
225 __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
226
227 vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
228
229 _mm512_mask_storeu_ps(y, vmask, vy);
230 }
231 }
232