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_x32(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_x32(
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 >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
43     __m512 vx0 = _mm512_loadu_ps(x);
44     __m512 vx1 = _mm512_loadu_ps(x + 16);
45     x += 32;
46 
47     const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale));
48     const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale));
49 
50     __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
51     __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
52 
53     const __m512i ven0 = _mm512_slli_epi32(_mm512_castps_si512(vn0), 19);
54     const __m512i vl0 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn0), vtable);
55     const __m512i ven1 = _mm512_slli_epi32(_mm512_castps_si512(vn1), 19);
56     const __m512i vl1 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn1), vtable);
57 
58     __m512 vs0 = _mm512_castsi512_ps(_mm512_add_epi32(vl0, ven0));
59     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
60     __m512 vs1 = _mm512_castsi512_ps(_mm512_add_epi32(vl1, ven1));
61     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
62 
63     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
64     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
65 
66     __m512 vp0 = _mm512_fmadd_ps(vc3, vt0, vc2);
67     __m512 vp1 = _mm512_fmadd_ps(vc3, vt1, vc2);
68 
69     vp0 = _mm512_mul_ps(vp0, vt0);
70     vt0 = _mm512_mul_ps(vt0, vs0);
71     vp1 = _mm512_mul_ps(vp1, vt1);
72     vt1 = _mm512_mul_ps(vt1, vs1);
73 
74     vs0 = _mm512_fmsub_ps(vs0, valpha, valpha);
75     vs1 = _mm512_fmsub_ps(vs1, valpha, valpha);
76 
77     vp0 = _mm512_fmadd_ps(vp0, vt0, vt0);
78     vp1 = _mm512_fmadd_ps(vp1, vt1, vt1);
79 
80     const __m512 vzero = _mm512_setzero_ps();
81     __m512 vy0 = _mm512_fmadd_ps(vp0, valpha, vs0);
82     const __mmask16 vsign0 = _mm512_cmp_ps_mask(vx0, vzero, _CMP_NLT_US);
83     __m512 vy1 = _mm512_fmadd_ps(vp1, valpha, vs1);
84     const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US);
85 
86     vy0 = _mm512_mask_mul_ps(vy0, vsign0, vx0, vbeta);
87     vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta);
88 
89     _mm512_storeu_ps(y, vy0);
90     _mm512_storeu_ps(y + 16, vy1);
91     y += 32;
92   }
93   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
94     __m512 vx = _mm512_loadu_ps(x);
95     x += 16;
96 
97     const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
98     const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
99 
100     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
101     const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
102     const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
103     __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
104     vn = _mm512_sub_ps(vn, vmagic_bias);
105 
106     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
107 
108     __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
109     vp = _mm512_mul_ps(vp, vt);
110 
111     vt = _mm512_mul_ps(vt, vs);
112     vs = _mm512_fmsub_ps(vs, valpha, valpha);
113     vp = _mm512_fmadd_ps(vp, vt, vt);
114     __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
115 
116     vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
117 
118     _mm512_storeu_ps(y, vy);
119     y += 16;
120   }
121   if XNN_UNLIKELY(n != 0) {
122     assert(n >= 1 * sizeof(float));
123     assert(n <= 15 * sizeof(float));
124     // Prepare mask for valid 32-bit elements (depends on n).
125     n >>= 2 /* log2(sizeof(float)) */;
126     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
127 
128     __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
129 
130     const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
131     const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
132 
133     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
134     const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
135     const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
136     __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
137     vn = _mm512_sub_ps(vn, vmagic_bias);
138 
139     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
140 
141     __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
142     vp = _mm512_mul_ps(vp, vt);
143 
144     vt = _mm512_mul_ps(vt, vs);
145     vs = _mm512_fmsub_ps(vs, valpha, valpha);
146     vp = _mm512_fmadd_ps(vp, vt, vt);
147     __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
148 
149     vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
150 
151     _mm512_mask_storeu_ps(y, vmask, vy);
152   }
153 }
154