1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx2-rr1-lut16-p3-gather.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 extern XNN_INTERNAL const int xnn_table_exp2minus_k_over_16[16];
19
20 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
21
xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x24(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])22 void xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x24(
23 size_t n,
24 const float* x,
25 float* y,
26 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(n % sizeof(float) == 0);
29
30 const __m256 vprescale = _mm256_broadcast_ps((const __m128*) params->sse.prescale);
31 const __m256 valpha = _mm256_broadcast_ps((const __m128*) params->sse.alpha);
32 const __m256 vbeta = _mm256_broadcast_ps((const __m128*) params->sse.beta);
33
34 const __m256 vsat_cutoff = _mm256_set1_ps(-0x1.154246p+4f);
35 const __m256 vmagic_bias = _mm256_set1_ps(0x1.800000p19f);
36 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
37 const __m256i vindex_mask = _mm256_set1_epi32(0xF);
38 const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
39 const __m256 vc3 = _mm256_set1_ps(0x1.55561Cp-3f);
40 const __m256 vc2 = _mm256_set1_ps(0x1.0001ECp-1f);
41
42 for (; n >= 24 * sizeof(float); n -= 24 * 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 x += 24;
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 const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
51
52 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
53 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
54 __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
55
56 const __m256i vidx0 = _mm256_and_si256(_mm256_castps_si256(vn0), vindex_mask);
57 const __m256i vl0 = _mm256_i32gather_epi32(xnn_table_exp2minus_k_over_16, vidx0, sizeof(float));
58 const __m256i vidx1 = _mm256_and_si256(_mm256_castps_si256(vn1), vindex_mask);
59 const __m256i vl1 = _mm256_i32gather_epi32(xnn_table_exp2minus_k_over_16, vidx1, sizeof(float));
60 const __m256i vidx2 = _mm256_and_si256(_mm256_castps_si256(vn2), vindex_mask);
61 const __m256i vl2 = _mm256_i32gather_epi32(xnn_table_exp2minus_k_over_16, vidx2, sizeof(float));
62
63 const __m256i ven0 = _mm256_slli_epi32(_mm256_castps_si256(vn0), 19);
64 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
65 const __m256i ven1 = _mm256_slli_epi32(_mm256_castps_si256(vn1), 19);
66 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
67 const __m256i ven2 = _mm256_slli_epi32(_mm256_castps_si256(vn2), 19);
68 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
69
70 __m256 vs0 = _mm256_castsi256_ps(_mm256_add_epi32(vl0, ven0));
71 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
72 __m256 vs1 = _mm256_castsi256_ps(_mm256_add_epi32(vl1, ven1));
73 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
74 __m256 vs2 = _mm256_castsi256_ps(_mm256_add_epi32(vl2, ven2));
75 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
76
77 __m256 vp0 = _mm256_fmadd_ps(vc3, vt0, vc2);
78 __m256 vp1 = _mm256_fmadd_ps(vc3, vt1, vc2);
79 __m256 vp2 = _mm256_fmadd_ps(vc3, vt2, vc2);
80
81 vp0 = _mm256_mul_ps(vp0, vt0);
82 vt0 = _mm256_mul_ps(vt0, vs0);
83 vp1 = _mm256_mul_ps(vp1, vt1);
84 vt1 = _mm256_mul_ps(vt1, vs1);
85 vp2 = _mm256_mul_ps(vp2, vt2);
86 vt2 = _mm256_mul_ps(vt2, vs2);
87
88 vs0 = _mm256_fmsub_ps(vs0, valpha, valpha);
89 vp0 = _mm256_fmadd_ps(vp0, vt0, vt0);
90 vs1 = _mm256_fmsub_ps(vs1, valpha, valpha);
91 vp1 = _mm256_fmadd_ps(vp1, vt1, vt1);
92 vs2 = _mm256_fmsub_ps(vs2, valpha, valpha);
93 vp2 = _mm256_fmadd_ps(vp2, vt2, vt2);
94
95 const __m256 ve0 = _mm256_fmadd_ps(vp0, valpha, vs0);
96 vx0 = _mm256_mul_ps(vx0, vbeta);
97 const __m256 ve1 = _mm256_fmadd_ps(vp1, valpha, vs1);
98 vx1 = _mm256_mul_ps(vx1, vbeta);
99 const __m256 ve2 = _mm256_fmadd_ps(vp2, valpha, vs2);
100 vx2 = _mm256_mul_ps(vx2, vbeta);
101
102 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
103 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
104 const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
105
106 _mm256_storeu_ps(y, vy0);
107 _mm256_storeu_ps(y + 8, vy1);
108 _mm256_storeu_ps(y + 16, vy2);
109 y += 24;
110 }
111 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
112 __m256 vx = _mm256_loadu_ps(x);
113 x += 8;
114
115 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
116
117 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
118 const __m256i vidx = _mm256_and_si256(_mm256_castps_si256(vn), vindex_mask);
119 const __m256i vl = _mm256_i32gather_epi32(xnn_table_exp2minus_k_over_16, vidx, sizeof(float));
120
121 const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 19);
122 vn = _mm256_sub_ps(vn, vmagic_bias);
123
124 __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
125 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
126
127 __m256 vp = _mm256_fmadd_ps(vc3, 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 vidx = _mm256_and_si256(_mm256_castps_si256(vn), vindex_mask);
152 const __m256i vl = _mm256_i32gather_epi32(xnn_table_exp2minus_k_over_16, vidx, sizeof(float));
153
154 const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 19);
155 vn = _mm256_sub_ps(vn, vmagic_bias);
156
157 __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
158 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
159
160 __m256 vp = _mm256_fmadd_ps(vc3, vt, vc2);
161 vp = _mm256_mul_ps(vp, vt);
162
163 vt = _mm256_mul_ps(vt, vs);
164 vs = _mm256_fmsub_ps(vs, valpha, valpha);
165 vp = _mm256_fmadd_ps(vp, vt, vt);
166 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
167
168 vx = _mm256_mul_ps(vx, vbeta);
169 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
170
171 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
172 __m128 vy_lo = _mm256_castps256_ps128(vy);
173 if (n & (4 * sizeof(float))) {
174 _mm_storeu_ps(y, vy_lo);
175 vy_lo = _mm256_extractf128_ps(vy, 1);
176 y += 4;
177 }
178 if (n & (2 * sizeof(float))) {
179 _mm_storel_pi((__m64*) y, vy_lo);
180 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
181 y += 2;
182 }
183 if (n & (1 * sizeof(float))) {
184 _mm_store_ss(y, vy_lo);
185 }
186 }
187 }
188