• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-velu/avx-rr2-p6.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_p6_x48(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_p6_x48(
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(0x1.8000FEp23f);
34   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
35   const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E440p-1f);
36   const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.0105C6p-21f);
37   const __m256 vc6 = _mm256_set1_ps(0x1.6b7338p-10f);
38   const __m256 vc5 = _mm256_set1_ps(0x1.12278Ep-7f);
39   const __m256 vc4 = _mm256_set1_ps(0x1.555716p-5f);
40   const __m256 vc3 = _mm256_set1_ps(0x1.5554B0p-3f);
41   const __m256 vc2 = _mm256_set1_ps(0x1.FFFFFEp-2f);
42   const __m256 vone = _mm256_set1_ps(1.0f);
43 
44   for (; n >= 48 * sizeof(float); n -= 48 * sizeof(float)) {
45     __m256 vx0 = _mm256_loadu_ps(x);
46     __m256 vx1 = _mm256_loadu_ps(x + 8);
47     __m256 vx2 = _mm256_loadu_ps(x + 16);
48     __m256 vx3 = _mm256_loadu_ps(x + 24);
49     __m256 vx4 = _mm256_loadu_ps(x + 32);
50     __m256 vx5 = _mm256_loadu_ps(x + 40);
51     x += 48;
52 
53     const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
54     const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
55     const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
56     const __m256 vz3 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx3, vprescale));
57     const __m256 vz4 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx4, vprescale));
58     const __m256 vz5 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx5, vprescale));
59 
60     __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
61     __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
62     __m256 vn2 = _mm256_add_ps(_mm256_mul_ps(vz2, vlog2e), vmagic_bias);
63     __m256 vn3 = _mm256_add_ps(_mm256_mul_ps(vz3, vlog2e), vmagic_bias);
64     __m256 vn4 = _mm256_add_ps(_mm256_mul_ps(vz4, vlog2e), vmagic_bias);
65     __m256 vn5 = _mm256_add_ps(_mm256_mul_ps(vz5, vlog2e), vmagic_bias);
66 
67     const __m128 vs0_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)), 23));
68     const __m128 vs0_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, 1)), 23));
69     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
70     const __m128 vs1_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)), 23));
71     const __m128 vs1_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, 1)), 23));
72     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
73     const __m128 vs2_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn2)), 23));
74     const __m128 vs2_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn2, 1)), 23));
75     vn2 = _mm256_sub_ps(vn2, vmagic_bias);
76     const __m128 vs3_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn3)), 23));
77     const __m128 vs3_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn3, 1)), 23));
78     vn3 = _mm256_sub_ps(vn3, vmagic_bias);
79     const __m128 vs4_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn4)), 23));
80     const __m128 vs4_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn4, 1)), 23));
81     vn4 = _mm256_sub_ps(vn4, vmagic_bias);
82     const __m128 vs5_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn5)), 23));
83     const __m128 vs5_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn5, 1)), 23));
84     vn5 = _mm256_sub_ps(vn5, vmagic_bias);
85 
86     __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
87     __m256 vs0 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs0_lo), vs0_hi, 1);
88     __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
89     __m256 vs1 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs1_lo), vs1_hi, 1);
90     __m256 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_hi), vz2);
91     __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs2_lo), vs2_hi, 1);
92     __m256 vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_hi), vz3);
93     __m256 vs3 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs3_lo), vs3_hi, 1);
94     __m256 vt4 = _mm256_add_ps(_mm256_mul_ps(vn4, vminus_ln2_hi), vz4);
95     __m256 vs4 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs4_lo), vs4_hi, 1);
96     __m256 vt5 = _mm256_add_ps(_mm256_mul_ps(vn5, vminus_ln2_hi), vz5);
97     __m256 vs5 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs5_lo), vs5_hi, 1);
98 
99     vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
100     vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
101     vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_lo), vt2);
102     vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_lo), vt3);
103     vt4 = _mm256_add_ps(_mm256_mul_ps(vn4, vminus_ln2_lo), vt4);
104     vt5 = _mm256_add_ps(_mm256_mul_ps(vn5, vminus_ln2_lo), vt5);
105 
106     __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc6, vt0), vc5);
107     __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc6, vt1), vc5);
108     __m256 vp2 = _mm256_add_ps(_mm256_mul_ps(vc6, vt2), vc5);
109     __m256 vp3 = _mm256_add_ps(_mm256_mul_ps(vc6, vt3), vc5);
110     __m256 vp4 = _mm256_add_ps(_mm256_mul_ps(vc6, vt4), vc5);
111     __m256 vp5 = _mm256_add_ps(_mm256_mul_ps(vc6, vt5), vc5);
112 
113     vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc4);
114     vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc4);
115     vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc4);
116     vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc4);
117     vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vc4);
118     vp5 = _mm256_add_ps(_mm256_mul_ps(vp5, vt5), vc4);
119 
120     vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc3);
121     vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc3);
122     vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc3);
123     vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc3);
124     vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vc3);
125     vp5 = _mm256_add_ps(_mm256_mul_ps(vp5, vt5), vc3);
126 
127     vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
128     vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
129     vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc2);
130     vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc2);
131     vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vc2);
132     vp5 = _mm256_add_ps(_mm256_mul_ps(vp5, vt5), vc2);
133 
134     vp0 = _mm256_mul_ps(vp0, vt0);
135     vp1 = _mm256_mul_ps(vp1, vt1);
136     vp2 = _mm256_mul_ps(vp2, vt2);
137     vp3 = _mm256_mul_ps(vp3, vt3);
138     vp4 = _mm256_mul_ps(vp4, vt4);
139     vp5 = _mm256_mul_ps(vp5, vt5);
140 
141     vt0 = _mm256_mul_ps(vt0, vs0);
142     vs0 = _mm256_sub_ps(vs0, vone);
143     vt1 = _mm256_mul_ps(vt1, vs1);
144     vs1 = _mm256_sub_ps(vs1, vone);
145     vt2 = _mm256_mul_ps(vt2, vs2);
146     vs2 = _mm256_sub_ps(vs2, vone);
147     vt3 = _mm256_mul_ps(vt3, vs3);
148     vs3 = _mm256_sub_ps(vs3, vone);
149     vt4 = _mm256_mul_ps(vt4, vs4);
150     vs4 = _mm256_sub_ps(vs4, vone);
151     vt5 = _mm256_mul_ps(vt5, vs5);
152     vs5 = _mm256_sub_ps(vs5, vone);
153 
154     vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vt0);
155     vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vt1);
156     vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vt2);
157     vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vt3);
158     vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vt4);
159     vp5 = _mm256_add_ps(_mm256_mul_ps(vp5, vt5), vt5);
160 
161     const __m256 ve0 = _mm256_mul_ps(_mm256_add_ps(vp0, vs0), valpha);
162     vx0 = _mm256_mul_ps(vx0, vbeta);
163     const __m256 ve1 = _mm256_mul_ps(_mm256_add_ps(vp1, vs1), valpha);
164     vx1 = _mm256_mul_ps(vx1, vbeta);
165     const __m256 ve2 = _mm256_mul_ps(_mm256_add_ps(vp2, vs2), valpha);
166     vx2 = _mm256_mul_ps(vx2, vbeta);
167     const __m256 ve3 = _mm256_mul_ps(_mm256_add_ps(vp3, vs3), valpha);
168     vx3 = _mm256_mul_ps(vx3, vbeta);
169     const __m256 ve4 = _mm256_mul_ps(_mm256_add_ps(vp4, vs4), valpha);
170     vx4 = _mm256_mul_ps(vx4, vbeta);
171     const __m256 ve5 = _mm256_mul_ps(_mm256_add_ps(vp5, vs5), valpha);
172     vx5 = _mm256_mul_ps(vx5, vbeta);
173 
174     const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
175     const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
176     const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
177     const __m256 vy3 = _mm256_blendv_ps(vx3, ve3, vx3);
178     const __m256 vy4 = _mm256_blendv_ps(vx4, ve4, vx4);
179     const __m256 vy5 = _mm256_blendv_ps(vx5, ve5, vx5);
180 
181     _mm256_storeu_ps(y, vy0);
182     _mm256_storeu_ps(y + 8, vy1);
183     _mm256_storeu_ps(y + 16, vy2);
184     _mm256_storeu_ps(y + 24, vy3);
185     _mm256_storeu_ps(y + 32, vy4);
186     _mm256_storeu_ps(y + 40, vy5);
187     y += 48;
188   }
189   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
190     __m256 vx = _mm256_loadu_ps(x);
191     x += 8;
192 
193     const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
194 
195     __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
196     const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
197     const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
198     vn = _mm256_sub_ps(vn, vmagic_bias);
199 
200     __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
201     __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
202     vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
203 
204     __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
205     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
206     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
207     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
208     vp = _mm256_mul_ps(vp, vt);
209 
210     vt = _mm256_mul_ps(vt, vs);
211     vs = _mm256_sub_ps(vs, vone);
212     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
213 
214     const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
215     vx = _mm256_mul_ps(vx, vbeta);
216     const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
217 
218     _mm256_storeu_ps(y, vy);
219     y += 8;
220   }
221   if XNN_UNLIKELY(n != 0) {
222     assert(n >= 1 * sizeof(float));
223     assert(n <= 7 * sizeof(float));
224     __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
225 
226     __m256 vx = _mm256_maskload_ps(x, vmask);
227 
228     const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
229 
230     __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
231     const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
232     const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
233     vn = _mm256_sub_ps(vn, vmagic_bias);
234 
235     __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
236     __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
237     vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
238 
239     __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
240     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
241     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
242     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
243     vp = _mm256_mul_ps(vp, vt);
244 
245     vt = _mm256_mul_ps(vt, vs);
246     vs = _mm256_sub_ps(vs, vone);
247     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
248 
249     const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
250     vx = _mm256_mul_ps(vx, vbeta);
251     const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
252 
253     // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
254     __m128 vy_lo = _mm256_castps256_ps128(vy);
255     if (n & (4 * sizeof(float))) {
256       _mm_storeu_ps(y, vy_lo);
257       vy_lo = _mm256_extractf128_ps(vy, 1);
258       y += 4;
259     }
260     if (n & (2 * sizeof(float))) {
261       _mm_storel_pi((__m64*) y, vy_lo);
262       vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
263       y += 2;
264     }
265     if (n & (1 * sizeof(float))) {
266       _mm_store_ss(y, vy_lo);
267     }
268   }
269 }
270