1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-velu/avx2-rr1-lut8-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_lut8_p4_perm_x32(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_lut8_p4_perm_x32(
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(0x1.800000p20f);
34   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
35   const __m256i vtable = _mm256_set_epi32(
36     0x3F7AC0C7, 0x3F7744FD, 0x3F75672A, 0x3F7504F3, 0x3F75FED7, 0x3F7837F0, 0x3F7B95C2, 0x3F800000);
37   const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
38   const __m256 vc4 = _mm256_set1_ps(0x1.5558ECp-5f);
39   const __m256 vc3 = _mm256_set1_ps(0x1.555C20p-3f);
40   const __m256 vc2 = _mm256_set1_ps(0x1.000000p-1f);
41 
42   for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
43     __m256 vx0 = _mm256_loadu_ps(x);
44     __m256 vx1 = _mm256_loadu_ps(x + 8);
45     __m256 vx2 = _mm256_loadu_ps(x + 16);
46     __m256 vx3 = _mm256_loadu_ps(x + 24);
47     x += 32;
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     const __m256 vz3 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx3, vprescale));
53 
54     __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
55     __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
56     __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
57     __m256 vn3 = _mm256_fmadd_ps(vz3, vlog2e, vmagic_bias);
58 
59     const __m256i ven0 = _mm256_slli_epi32(_mm256_castps_si256(vn0), 20);
60     const __m256i vl0 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn0));
61     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
62     const __m256i ven1 = _mm256_slli_epi32(_mm256_castps_si256(vn1), 20);
63     const __m256i vl1 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn1));
64     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
65     const __m256i ven2 = _mm256_slli_epi32(_mm256_castps_si256(vn2), 20);
66     const __m256i vl2 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn2));
67     vn2 = _mm256_sub_ps(vn2, vmagic_bias);
68     const __m256i ven3 = _mm256_slli_epi32(_mm256_castps_si256(vn3), 20);
69     const __m256i vl3 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn3));
70     vn3 = _mm256_sub_ps(vn3, vmagic_bias);
71 
72     __m256 vs0 = _mm256_castsi256_ps(_mm256_add_epi32(vl0, ven0));
73     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
74     __m256 vs1 = _mm256_castsi256_ps(_mm256_add_epi32(vl1, ven1));
75     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
76     __m256 vs2 = _mm256_castsi256_ps(_mm256_add_epi32(vl2, ven2));
77     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
78     __m256 vs3 = _mm256_castsi256_ps(_mm256_add_epi32(vl3, ven3));
79     __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2, vz3);
80 
81     __m256 vp0 = _mm256_fmadd_ps(vc4, vt0, vc3);
82     __m256 vp1 = _mm256_fmadd_ps(vc4, vt1, vc3);
83     __m256 vp2 = _mm256_fmadd_ps(vc4, vt2, vc3);
84     __m256 vp3 = _mm256_fmadd_ps(vc4, vt3, vc3);
85 
86     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
87     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
88     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
89     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
90 
91     vp0 = _mm256_mul_ps(vp0, vt0);
92     vt0 = _mm256_mul_ps(vt0, vs0);
93     vp1 = _mm256_mul_ps(vp1, vt1);
94     vt1 = _mm256_mul_ps(vt1, vs1);
95     vp2 = _mm256_mul_ps(vp2, vt2);
96     vt2 = _mm256_mul_ps(vt2, vs2);
97     vp3 = _mm256_mul_ps(vp3, vt3);
98     vt3 = _mm256_mul_ps(vt3, vs3);
99 
100     vs0 = _mm256_fmsub_ps(vs0, valpha, valpha);
101     vp0 = _mm256_fmadd_ps(vp0, vt0, vt0);
102     vs1 = _mm256_fmsub_ps(vs1, valpha, valpha);
103     vp1 = _mm256_fmadd_ps(vp1, vt1, vt1);
104     vs2 = _mm256_fmsub_ps(vs2, valpha, valpha);
105     vp2 = _mm256_fmadd_ps(vp2, vt2, vt2);
106     vs3 = _mm256_fmsub_ps(vs3, valpha, valpha);
107     vp3 = _mm256_fmadd_ps(vp3, vt3, vt3);
108 
109     const __m256 ve0 = _mm256_fmadd_ps(vp0, valpha, vs0);
110     vx0 = _mm256_mul_ps(vx0, vbeta);
111     const __m256 ve1 = _mm256_fmadd_ps(vp1, valpha, vs1);
112     vx1 = _mm256_mul_ps(vx1, vbeta);
113     const __m256 ve2 = _mm256_fmadd_ps(vp2, valpha, vs2);
114     vx2 = _mm256_mul_ps(vx2, vbeta);
115     const __m256 ve3 = _mm256_fmadd_ps(vp3, valpha, vs3);
116     vx3 = _mm256_mul_ps(vx3, vbeta);
117 
118     const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
119     const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
120     const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
121     const __m256 vy3 = _mm256_blendv_ps(vx3, ve3, vx3);
122 
123     _mm256_storeu_ps(y, vy0);
124     _mm256_storeu_ps(y + 8, vy1);
125     _mm256_storeu_ps(y + 16, vy2);
126     _mm256_storeu_ps(y + 24, vy3);
127     y += 32;
128   }
129   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
130     __m256 vx = _mm256_loadu_ps(x);
131     x += 8;
132 
133     const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
134 
135     __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
136     const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 20);
137     const __m256i vl = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn));
138     __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
139     vn = _mm256_sub_ps(vn, vmagic_bias);
140 
141     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
142 
143     __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3);
144     vp = _mm256_fmadd_ps(vp, vt, vc2);
145     vp = _mm256_mul_ps(vp, vt);
146 
147     vt = _mm256_mul_ps(vt, vs);
148     vs = _mm256_fmsub_ps(vs, valpha, valpha);
149     vp = _mm256_fmadd_ps(vp, vt, vt);
150     const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
151 
152     vx = _mm256_mul_ps(vx, vbeta);
153     const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
154 
155     _mm256_storeu_ps(y, vy);
156     y += 8;
157   }
158   if XNN_UNLIKELY(n != 0) {
159     assert(n >= 1 * sizeof(float));
160     assert(n <= 7 * sizeof(float));
161     __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
162 
163     __m256 vx = _mm256_maskload_ps(x, vmask);
164 
165     const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
166 
167     __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
168     const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 20);
169     const __m256i vl = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn));
170     __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
171     vn = _mm256_sub_ps(vn, vmagic_bias);
172 
173     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
174 
175     __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3);
176     vp = _mm256_fmadd_ps(vp, vt, vc2);
177     vp = _mm256_mul_ps(vp, vt);
178 
179     vt = _mm256_mul_ps(vt, vs);
180     vs = _mm256_fmsub_ps(vs, valpha, valpha);
181     vp = _mm256_fmadd_ps(vp, vt, vt);
182     const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
183 
184     vx = _mm256_mul_ps(vx, vbeta);
185     const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
186 
187     // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
188     __m128 vy_lo = _mm256_castps256_ps128(vy);
189     if (n & (4 * sizeof(float))) {
190       _mm_storeu_ps(y, vy_lo);
191       vy_lo = _mm256_extractf128_ps(vy, 1);
192       y += 4;
193     }
194     if (n & (2 * sizeof(float))) {
195       _mm_storel_pi((__m64*) y, vy_lo);
196       vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
197       y += 2;
198     }
199     if (n & (1 * sizeof(float))) {
200       _mm_store_ss(y, vy_lo);
201     }
202   }
203 }
204