1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-velu/avx512f-rr1-lut16-p3-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/intrinsics-polyfill.h>
16 #include <xnnpack/vunary.h>
17 
18 
xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_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_lut16_p3_perm_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(0x1.800000p19f);
34   const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
35   const __m512i vtable = _mm512_set_epi32(
36     0x3F7D257D, 0x3F7AC0C7, 0x3F78CCDF, 0x3F7744FD, 0x3F76248C, 0x3F75672A, 0x3F7508A4, 0x3F7504F3,
37     0x3F75583F, 0x3F75FED7, 0x3F76F532, 0x3F7837F0, 0x3F79C3D3, 0x3F7B95C2, 0x3F7DAAC3, 0x3F800000);
38   const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f);
39   const __m512 vc3 = _mm512_set1_ps(0x1.55561Cp-3f);
40   const __m512 vc2 = _mm512_set1_ps(0x1.0001ECp-1f);
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     const __m512i ven0 = _mm512_slli_epi32(_mm512_castps_si512(vn0), 19);
66     const __m512i vl0 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn0), vtable);
67     const __m512i ven1 = _mm512_slli_epi32(_mm512_castps_si512(vn1), 19);
68     const __m512i vl1 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn1), vtable);
69     const __m512i ven2 = _mm512_slli_epi32(_mm512_castps_si512(vn2), 19);
70     const __m512i vl2 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn2), vtable);
71     const __m512i ven3 = _mm512_slli_epi32(_mm512_castps_si512(vn3), 19);
72     const __m512i vl3 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn3), vtable);
73     const __m512i ven4 = _mm512_slli_epi32(_mm512_castps_si512(vn4), 19);
74     const __m512i vl4 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn4), vtable);
75     const __m512i ven5 = _mm512_slli_epi32(_mm512_castps_si512(vn5), 19);
76     const __m512i vl5 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn5), vtable);
77 
78     __m512 vs0 = _mm512_castsi512_ps(_mm512_add_epi32(vl0, ven0));
79     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
80     __m512 vs1 = _mm512_castsi512_ps(_mm512_add_epi32(vl1, ven1));
81     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
82     __m512 vs2 = _mm512_castsi512_ps(_mm512_add_epi32(vl2, ven2));
83     vn2 = _mm512_sub_ps(vn2, vmagic_bias);
84     __m512 vs3 = _mm512_castsi512_ps(_mm512_add_epi32(vl3, ven3));
85     vn3 = _mm512_sub_ps(vn3, vmagic_bias);
86     __m512 vs4 = _mm512_castsi512_ps(_mm512_add_epi32(vl4, ven4));
87     vn4 = _mm512_sub_ps(vn4, vmagic_bias);
88     __m512 vs5 = _mm512_castsi512_ps(_mm512_add_epi32(vl5, ven5));
89     vn5 = _mm512_sub_ps(vn5, vmagic_bias);
90 
91     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
92     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
93     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
94     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
95     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4);
96     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5);
97 
98     __m512 vp0 = _mm512_fmadd_ps(vc3, vt0, vc2);
99     __m512 vp1 = _mm512_fmadd_ps(vc3, vt1, vc2);
100     __m512 vp2 = _mm512_fmadd_ps(vc3, vt2, vc2);
101     __m512 vp3 = _mm512_fmadd_ps(vc3, vt3, vc2);
102     __m512 vp4 = _mm512_fmadd_ps(vc3, vt4, vc2);
103     __m512 vp5 = _mm512_fmadd_ps(vc3, vt5, vc2);
104 
105     vp0 = _mm512_mul_ps(vp0, vt0);
106     vt0 = _mm512_mul_ps(vt0, vs0);
107     vp1 = _mm512_mul_ps(vp1, vt1);
108     vt1 = _mm512_mul_ps(vt1, vs1);
109     vp2 = _mm512_mul_ps(vp2, vt2);
110     vt2 = _mm512_mul_ps(vt2, vs2);
111     vp3 = _mm512_mul_ps(vp3, vt3);
112     vt3 = _mm512_mul_ps(vt3, vs3);
113     vp4 = _mm512_mul_ps(vp4, vt4);
114     vt4 = _mm512_mul_ps(vt4, vs4);
115     vp5 = _mm512_mul_ps(vp5, vt5);
116     vt5 = _mm512_mul_ps(vt5, vs5);
117 
118     vs0 = _mm512_fmsub_ps(vs0, valpha, valpha);
119     vs1 = _mm512_fmsub_ps(vs1, valpha, valpha);
120     vs2 = _mm512_fmsub_ps(vs2, valpha, valpha);
121     vs3 = _mm512_fmsub_ps(vs3, valpha, valpha);
122     vs4 = _mm512_fmsub_ps(vs4, valpha, valpha);
123     vs5 = _mm512_fmsub_ps(vs5, valpha, valpha);
124 
125     vp0 = _mm512_fmadd_ps(vp0, vt0, vt0);
126     vp1 = _mm512_fmadd_ps(vp1, vt1, vt1);
127     vp2 = _mm512_fmadd_ps(vp2, vt2, vt2);
128     vp3 = _mm512_fmadd_ps(vp3, vt3, vt3);
129     vp4 = _mm512_fmadd_ps(vp4, vt4, vt4);
130     vp5 = _mm512_fmadd_ps(vp5, vt5, vt5);
131 
132     const __m512 vzero = _mm512_setzero_ps();
133     __m512 vy0 = _mm512_fmadd_ps(vp0, valpha, vs0);
134     const __mmask16 vsign0 = _mm512_cmp_ps_mask(vx0, vzero, _CMP_NLT_US);
135     __m512 vy1 = _mm512_fmadd_ps(vp1, valpha, vs1);
136     const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US);
137     __m512 vy2 = _mm512_fmadd_ps(vp2, valpha, vs2);
138     const __mmask16 vsign2 = _mm512_cmp_ps_mask(vx2, vzero, _CMP_NLT_US);
139     __m512 vy3 = _mm512_fmadd_ps(vp3, valpha, vs3);
140     const __mmask16 vsign3 = _mm512_cmp_ps_mask(vx3, vzero, _CMP_NLT_US);
141     __m512 vy4 = _mm512_fmadd_ps(vp4, valpha, vs4);
142     const __mmask16 vsign4 = _mm512_cmp_ps_mask(vx4, vzero, _CMP_NLT_US);
143     __m512 vy5 = _mm512_fmadd_ps(vp5, valpha, vs5);
144     const __mmask16 vsign5 = _mm512_cmp_ps_mask(vx5, vzero, _CMP_NLT_US);
145 
146     vy0 = _mm512_mask_mul_ps(vy0, vsign0, vx0, vbeta);
147     vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta);
148     vy2 = _mm512_mask_mul_ps(vy2, vsign2, vx2, vbeta);
149     vy3 = _mm512_mask_mul_ps(vy3, vsign3, vx3, vbeta);
150     vy4 = _mm512_mask_mul_ps(vy4, vsign4, vx4, vbeta);
151     vy5 = _mm512_mask_mul_ps(vy5, vsign5, vx5, vbeta);
152 
153     _mm512_storeu_ps(y, vy0);
154     _mm512_storeu_ps(y + 16, vy1);
155     _mm512_storeu_ps(y + 32, vy2);
156     _mm512_storeu_ps(y + 48, vy3);
157     _mm512_storeu_ps(y + 64, vy4);
158     _mm512_storeu_ps(y + 80, vy5);
159     y += 96;
160   }
161   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
162     __m512 vx = _mm512_loadu_ps(x);
163     x += 16;
164 
165     const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
166     const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
167 
168     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
169     const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
170     const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
171     __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
172     vn = _mm512_sub_ps(vn, vmagic_bias);
173 
174     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
175 
176     __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
177     vp = _mm512_mul_ps(vp, vt);
178 
179     vt = _mm512_mul_ps(vt, vs);
180     vs = _mm512_fmsub_ps(vs, valpha, valpha);
181     vp = _mm512_fmadd_ps(vp, vt, vt);
182     __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
183 
184     vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
185 
186     _mm512_storeu_ps(y, vy);
187     y += 16;
188   }
189   if XNN_UNLIKELY(n != 0) {
190     assert(n >= 1 * sizeof(float));
191     assert(n <= 15 * sizeof(float));
192     // Prepare mask for valid 32-bit elements (depends on n).
193     n >>= 2 /* log2(sizeof(float)) */;
194     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
195 
196     __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
197 
198     const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
199     const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
200 
201     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
202     const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
203     const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
204     __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
205     vn = _mm512_sub_ps(vn, vmagic_bias);
206 
207     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
208 
209     __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
210     vp = _mm512_mul_ps(vp, vt);
211 
212     vt = _mm512_mul_ps(vt, vs);
213     vs = _mm512_fmsub_ps(vs, valpha, valpha);
214     vp = _mm512_fmadd_ps(vp, vt, vt);
215     __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
216 
217     vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
218 
219     _mm512_mask_storeu_ps(y, vmask, vy);
220   }
221 }
222