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_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__avx2_rr1_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(0x1.800000p21f);
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 >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
44     __m256 vx0 = _mm256_loadu_ps(x);
45     __m256 vx1 = _mm256_loadu_ps(x + 8);
46     x += 16;
47 
48     const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
49     const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
50 
51     __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
52     __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
53 
54     const __m256i ven0 = _mm256_slli_epi32(_mm256_castps_si256(vn0), 21);
55     const __m256i vl0 = _mm256_castps_si256(_mm256_permutevar_ps(vtable, _mm256_castps_si256(vn0)));
56     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
57     const __m256i ven1 = _mm256_slli_epi32(_mm256_castps_si256(vn1), 21);
58     const __m256i vl1 = _mm256_castps_si256(_mm256_permutevar_ps(vtable, _mm256_castps_si256(vn1)));
59     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
60 
61     __m256 vs0 = _mm256_castsi256_ps(_mm256_add_epi32(vl0, ven0));
62     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
63     __m256 vs1 = _mm256_castsi256_ps(_mm256_add_epi32(vl1, ven1));
64     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
65 
66     __m256 vp0 = _mm256_fmadd_ps(vc4, vt0, vc3);
67     __m256 vp1 = _mm256_fmadd_ps(vc4, vt1, vc3);
68 
69     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
70     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
71 
72     vp0 = _mm256_mul_ps(vp0, vt0);
73     vt0 = _mm256_mul_ps(vt0, vs0);
74     vp1 = _mm256_mul_ps(vp1, vt1);
75     vt1 = _mm256_mul_ps(vt1, vs1);
76 
77     vs0 = _mm256_fmsub_ps(vs0, valpha, valpha);
78     vp0 = _mm256_fmadd_ps(vp0, vt0, vt0);
79     vs1 = _mm256_fmsub_ps(vs1, valpha, valpha);
80     vp1 = _mm256_fmadd_ps(vp1, vt1, vt1);
81 
82     const __m256 ve0 = _mm256_fmadd_ps(vp0, valpha, vs0);
83     vx0 = _mm256_mul_ps(vx0, vbeta);
84     const __m256 ve1 = _mm256_fmadd_ps(vp1, valpha, vs1);
85     vx1 = _mm256_mul_ps(vx1, vbeta);
86 
87     const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
88     const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
89 
90     _mm256_storeu_ps(y, vy0);
91     _mm256_storeu_ps(y + 8, vy1);
92     y += 16;
93   }
94   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
95     __m256 vx = _mm256_loadu_ps(x);
96     x += 8;
97 
98     const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
99 
100     __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
101     const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 21);
102     const __m256i vl = _mm256_castps_si256(_mm256_permutevar_ps(vtable, _mm256_castps_si256(vn)));
103     __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
104     vn = _mm256_sub_ps(vn, vmagic_bias);
105 
106     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
107 
108     __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3);
109     vp = _mm256_fmadd_ps(vp, vt, vc2);
110     vp = _mm256_mul_ps(vp, vt);
111 
112     vt = _mm256_mul_ps(vt, vs);
113     vs = _mm256_fmsub_ps(vs, valpha, valpha);
114     vp = _mm256_fmadd_ps(vp, vt, vt);
115     const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
116 
117     vx = _mm256_mul_ps(vx, vbeta);
118     const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
119 
120     _mm256_storeu_ps(y, vy);
121     y += 8;
122   }
123   if XNN_UNLIKELY(n != 0) {
124     assert(n >= 1 * sizeof(float));
125     assert(n <= 7 * sizeof(float));
126     __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
127 
128     __m256 vx = _mm256_maskload_ps(x, vmask);
129 
130     const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
131 
132     __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
133     const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 21);
134     const __m256i vl = _mm256_castps_si256(_mm256_permutevar_ps(vtable, _mm256_castps_si256(vn)));
135     __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
136     vn = _mm256_sub_ps(vn, vmagic_bias);
137 
138     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
139 
140     __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3);
141     vp = _mm256_fmadd_ps(vp, vt, vc2);
142     vp = _mm256_mul_ps(vp, vt);
143 
144     vt = _mm256_mul_ps(vt, vs);
145     vs = _mm256_fmsub_ps(vs, valpha, valpha);
146     vp = _mm256_fmadd_ps(vp, vt, vt);
147     const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
148 
149     vx = _mm256_mul_ps(vx, vbeta);
150     const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
151 
152     // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
153     __m128 vy_lo = _mm256_castps256_ps128(vy);
154     if (n & (4 * sizeof(float))) {
155       _mm_storeu_ps(y, vy_lo);
156       vy_lo = _mm256_extractf128_ps(vy, 1);
157       y += 4;
158     }
159     if (n & (2 * sizeof(float))) {
160       _mm_storel_pi((__m64*) y, vy_lo);
161       vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
162       y += 2;
163     }
164     if (n & (1 * sizeof(float))) {
165       _mm_store_ss(y, vy_lo);
166     }
167   }
168 }
169