1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-velu/avx2-rr1-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__avx2_rr1_lut4_p4_perm_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__avx2_rr1_lut4_p4_perm_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(0x1.800000p21f);
34   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
35   const __m256 vtable = _mm256_set_ps(
36     0x1.EE89FAp-1f, 0x1.EA09E6p-1f, 0x1.F06FE0p-1f, 0x1.000000p+0f,
37     0x1.EE89FAp-1f, 0x1.EA09E6p-1f, 0x1.F06FE0p-1f, 0x1.000000p+0f);
38   const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
39   const __m256 vc4 = _mm256_set1_ps(0x1.554F9Ap-5f);
40   const __m256 vc3 = _mm256_set1_ps(0x1.557082p-3f);
41   const __m256 vc2 = _mm256_set1_ps(0x1.000002p-1f);
42 
43   for (; n >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
44     __m256 vx0 = _mm256_loadu_ps(x);
45     __m256 vx1 = _mm256_loadu_ps(x + 8);
46     __m256 vx2 = _mm256_loadu_ps(x + 16);
47     x += 24;
48 
49     const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
50     const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
51     const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
52 
53     __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
54     __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
55     __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
56 
57     const __m256i ven0 = _mm256_slli_epi32(_mm256_castps_si256(vn0), 21);
58     const __m256i vl0 = _mm256_castps_si256(_mm256_permutevar_ps(vtable, _mm256_castps_si256(vn0)));
59     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
60     const __m256i ven1 = _mm256_slli_epi32(_mm256_castps_si256(vn1), 21);
61     const __m256i vl1 = _mm256_castps_si256(_mm256_permutevar_ps(vtable, _mm256_castps_si256(vn1)));
62     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
63     const __m256i ven2 = _mm256_slli_epi32(_mm256_castps_si256(vn2), 21);
64     const __m256i vl2 = _mm256_castps_si256(_mm256_permutevar_ps(vtable, _mm256_castps_si256(vn2)));
65     vn2 = _mm256_sub_ps(vn2, vmagic_bias);
66 
67     __m256 vs0 = _mm256_castsi256_ps(_mm256_add_epi32(vl0, ven0));
68     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
69     __m256 vs1 = _mm256_castsi256_ps(_mm256_add_epi32(vl1, ven1));
70     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
71     __m256 vs2 = _mm256_castsi256_ps(_mm256_add_epi32(vl2, ven2));
72     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
73 
74     __m256 vp0 = _mm256_fmadd_ps(vc4, vt0, vc3);
75     __m256 vp1 = _mm256_fmadd_ps(vc4, vt1, vc3);
76     __m256 vp2 = _mm256_fmadd_ps(vc4, vt2, vc3);
77 
78     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
79     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
80     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
81 
82     vp0 = _mm256_mul_ps(vp0, vt0);
83     vt0 = _mm256_mul_ps(vt0, vs0);
84     vp1 = _mm256_mul_ps(vp1, vt1);
85     vt1 = _mm256_mul_ps(vt1, vs1);
86     vp2 = _mm256_mul_ps(vp2, vt2);
87     vt2 = _mm256_mul_ps(vt2, vs2);
88 
89     vs0 = _mm256_fmsub_ps(vs0, valpha, valpha);
90     vp0 = _mm256_fmadd_ps(vp0, vt0, vt0);
91     vs1 = _mm256_fmsub_ps(vs1, valpha, valpha);
92     vp1 = _mm256_fmadd_ps(vp1, vt1, vt1);
93     vs2 = _mm256_fmsub_ps(vs2, valpha, valpha);
94     vp2 = _mm256_fmadd_ps(vp2, vt2, vt2);
95 
96     const __m256 ve0 = _mm256_fmadd_ps(vp0, valpha, vs0);
97     vx0 = _mm256_mul_ps(vx0, vbeta);
98     const __m256 ve1 = _mm256_fmadd_ps(vp1, valpha, vs1);
99     vx1 = _mm256_mul_ps(vx1, vbeta);
100     const __m256 ve2 = _mm256_fmadd_ps(vp2, valpha, vs2);
101     vx2 = _mm256_mul_ps(vx2, vbeta);
102 
103     const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
104     const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
105     const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
106 
107     _mm256_storeu_ps(y, vy0);
108     _mm256_storeu_ps(y + 8, vy1);
109     _mm256_storeu_ps(y + 16, vy2);
110     y += 24;
111   }
112   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
113     __m256 vx = _mm256_loadu_ps(x);
114     x += 8;
115 
116     const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
117 
118     __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
119     const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 21);
120     const __m256i vl = _mm256_castps_si256(_mm256_permutevar_ps(vtable, _mm256_castps_si256(vn)));
121     __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
122     vn = _mm256_sub_ps(vn, vmagic_bias);
123 
124     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
125 
126     __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3);
127     vp = _mm256_fmadd_ps(vp, vt, vc2);
128     vp = _mm256_mul_ps(vp, vt);
129 
130     vt = _mm256_mul_ps(vt, vs);
131     vs = _mm256_fmsub_ps(vs, valpha, valpha);
132     vp = _mm256_fmadd_ps(vp, vt, vt);
133     const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
134 
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_fmadd_ps(vz, vlog2e, vmagic_bias);
151     const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 21);
152     const __m256i vl = _mm256_castps_si256(_mm256_permutevar_ps(vtable, _mm256_castps_si256(vn)));
153     __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
154     vn = _mm256_sub_ps(vn, vmagic_bias);
155 
156     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
157 
158     __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3);
159     vp = _mm256_fmadd_ps(vp, vt, vc2);
160     vp = _mm256_mul_ps(vp, vt);
161 
162     vt = _mm256_mul_ps(vt, vs);
163     vs = _mm256_fmsub_ps(vs, valpha, valpha);
164     vp = _mm256_fmadd_ps(vp, vt, vt);
165     const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
166 
167     vx = _mm256_mul_ps(vx, vbeta);
168     const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
169 
170     // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
171     __m128 vy_lo = _mm256_castps256_ps128(vy);
172     if (n & (4 * sizeof(float))) {
173       _mm_storeu_ps(y, vy_lo);
174       vy_lo = _mm256_extractf128_ps(vy, 1);
175       y += 4;
176     }
177     if (n & (2 * sizeof(float))) {
178       _mm_storel_pi((__m64*) y, vy_lo);
179       vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
180       y += 2;
181     }
182     if (n & (1 * sizeof(float))) {
183       _mm_store_ss(y, vy_lo);
184     }
185   }
186 }
187