1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-velu/avx-rr2-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__avx_rr2_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__avx_rr2_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(0x1.8003F8p21f);
34   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
35   const __m256 vindex_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x3));
36   const __m256 vtable = _mm256_set_ps(
37     0x1.AE89FAp+0f, 0x1.6A09E6p+0f, 0x1.306FE0p+0f, 0x1.000000p+0f,
38     0x1.AE89FAp+0f, 0x1.6A09E6p+0f, 0x1.306FE0p+0f, 0x1.000000p+0f);
39   const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E400p-1f);
40   const __m256 vminus_ln2_lo = _mm256_set1_ps(-0x1.7F7D1Cp-20f);
41   const __m256 vc4 = _mm256_set1_ps(0x1.554F9Ap-5f);
42   const __m256 vc3 = _mm256_set1_ps(0x1.557082p-3f);
43   const __m256 vc2 = _mm256_set1_ps(0x1.000002p-1f);
44   const __m256 vone = _mm256_set1_ps(1.0f);
45 
46   for (; n >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
47     __m256 vx0 = _mm256_loadu_ps(x);
48     __m256 vx1 = _mm256_loadu_ps(x + 8);
49     __m256 vx2 = _mm256_loadu_ps(x + 16);
50     x += 24;
51 
52     const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
53     const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
54     const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
55 
56     __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
57     __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
58     __m256 vn2 = _mm256_add_ps(_mm256_mul_ps(vz2, vlog2e), vmagic_bias);
59 
60     __m256 ven0 = _mm256_andnot_ps(vindex_mask, vn0);
61     const __m256 vl0 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn0));
62     const __m128 ven0_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven0)), 21));
63     __m256 ven1 = _mm256_andnot_ps(vindex_mask, vn1);
64     const __m256 vl1 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn1));
65     const __m128 ven1_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven1)), 21));
66     __m256 ven2 = _mm256_andnot_ps(vindex_mask, vn2);
67     const __m256 vl2 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn2));
68     const __m128 ven2_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven2)), 21));
69 
70     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
71     const __m128 ven0_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven0, 1)), 21));
72     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
73     const __m128 ven1_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven1, 1)), 21));
74     vn2 = _mm256_sub_ps(vn2, vmagic_bias);
75     const __m128 ven2_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven2, 1)), 21));
76 
77     __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
78     ven0 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven0_lo), ven0_hi, 1);
79     __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
80     ven1 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven1_lo), ven1_hi, 1);
81     __m256 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_hi), vz2);
82     ven2 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven2_lo), ven2_hi, 1);
83 
84     vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
85     __m256 vs0 = _mm256_mul_ps(vl0, ven0);
86     vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
87     __m256 vs1 = _mm256_mul_ps(vl1, ven1);
88     vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_lo), vt2);
89     __m256 vs2 = _mm256_mul_ps(vl2, ven2);
90 
91     __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc4, vt0), vc3);
92     __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc4, vt1), vc3);
93     __m256 vp2 = _mm256_add_ps(_mm256_mul_ps(vc4, vt2), vc3);
94 
95     vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
96     vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
97     vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc2);
98 
99     vp0 = _mm256_mul_ps(vp0, vt0);
100     vp1 = _mm256_mul_ps(vp1, vt1);
101     vp2 = _mm256_mul_ps(vp2, vt2);
102 
103     vt0 = _mm256_mul_ps(vt0, vs0);
104     vs0 = _mm256_sub_ps(vs0, vone);
105     vt1 = _mm256_mul_ps(vt1, vs1);
106     vs1 = _mm256_sub_ps(vs1, vone);
107     vt2 = _mm256_mul_ps(vt2, vs2);
108     vs2 = _mm256_sub_ps(vs2, vone);
109 
110     vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vt0);
111     vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vt1);
112     vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vt2);
113 
114     const __m256 ve0 = _mm256_mul_ps(_mm256_add_ps(vp0, vs0), valpha);
115     vx0 = _mm256_mul_ps(vx0, vbeta);
116     const __m256 ve1 = _mm256_mul_ps(_mm256_add_ps(vp1, vs1), valpha);
117     vx1 = _mm256_mul_ps(vx1, vbeta);
118     const __m256 ve2 = _mm256_mul_ps(_mm256_add_ps(vp2, vs2), valpha);
119     vx2 = _mm256_mul_ps(vx2, vbeta);
120 
121     const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
122     const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
123     const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
124 
125     _mm256_storeu_ps(y, vy0);
126     _mm256_storeu_ps(y + 8, vy1);
127     _mm256_storeu_ps(y + 16, vy2);
128     y += 24;
129   }
130   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
131     __m256 vx = _mm256_loadu_ps(x);
132     x += 8;
133 
134     const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
135 
136     __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
137     __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
138     const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
139     const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
140     vn = _mm256_sub_ps(vn, vmagic_bias);
141     const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
142 
143     __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
144     ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
145     vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
146     __m256 vs = _mm256_mul_ps(vl, ven);
147 
148     __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
149     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
150     vp = _mm256_mul_ps(vp, vt);
151 
152     vt = _mm256_mul_ps(vt, vs);
153     vs = _mm256_sub_ps(vs, vone);
154     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
155 
156     const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
157     vx = _mm256_mul_ps(vx, vbeta);
158     const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
159 
160     _mm256_storeu_ps(y, vy);
161     y += 8;
162   }
163   if XNN_UNLIKELY(n != 0) {
164     assert(n >= 1 * sizeof(float));
165     assert(n <= 7 * sizeof(float));
166     __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
167 
168     __m256 vx = _mm256_maskload_ps(x, vmask);
169 
170     const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
171 
172     __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
173     __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
174     const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
175     const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
176     vn = _mm256_sub_ps(vn, vmagic_bias);
177     const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
178 
179     __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
180     ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
181     vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
182     __m256 vs = _mm256_mul_ps(vl, ven);
183 
184     __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
185     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
186     vp = _mm256_mul_ps(vp, vt);
187 
188     vt = _mm256_mul_ps(vt, vs);
189     vs = _mm256_sub_ps(vs, vone);
190     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
191 
192     const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
193     vx = _mm256_mul_ps(vx, vbeta);
194     const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
195 
196     // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
197     __m128 vy_lo = _mm256_castps256_ps128(vy);
198     if (n & (4 * sizeof(float))) {
199       _mm_storeu_ps(y, vy_lo);
200       vy_lo = _mm256_extractf128_ps(vy, 1);
201       y += 4;
202     }
203     if (n & (2 * sizeof(float))) {
204       _mm_storel_pi((__m64*) y, vy_lo);
205       vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
206       y += 2;
207     }
208     if (n & (1 * sizeof(float))) {
209       _mm_store_ss(y, vy_lo);
210     }
211   }
212 }
213