1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/sse-rr2-lut16-p3.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 <emmintrin.h>
13
14 #include <xnnpack/vunary.h>
15 #include <xnnpack/common.h>
16
17
18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16];
19
xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8(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__sse2_rr2_lut16_p3_x8(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
25 {
26 assert(n != 0);
27 assert(n % sizeof(float) == 0);
28 assert(x != NULL);
29 assert(y != NULL);
30
31 const __m128 vprescale = _mm_load_ps(params->sse.prescale);
32 const __m128 valpha = _mm_load_ps(params->sse.alpha);
33 const __m128 vbeta = _mm_load_ps(params->sse.beta);
34
35 const __m128 vsat_cutoff = _mm_set1_ps(-0x1.154246p+4f);
36 const __m128 vmagic_bias = _mm_set1_ps(0x1.800000p19f);
37 const __m128 vlog2e = _mm_set1_ps(0x1.715476p+0f);
38 const __m128i vindex_mask = _mm_set1_epi32(0xF);
39 const __m128 vminus_ln2_hi = _mm_set1_ps(-0x1.62E400p-1f);
40 const __m128 vminus_ln2_lo = _mm_set1_ps(-0x1.7F7D1Cp-20f);
41 const __m128 vc3 = _mm_set1_ps(0x1.55561Cp-3f);
42 const __m128 vc2 = _mm_set1_ps(0x1.0001ECp-1f);
43 const __m128 vone = _mm_set1_ps(1.0f);
44
45 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
46 __m128 vx0123 = _mm_loadu_ps(x);
47 __m128 vx4567 = _mm_loadu_ps(x + 4);
48 x += 8;
49
50 const __m128 vz0123 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx0123, vprescale));
51 const __m128 vz4567 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx4567, vprescale));
52
53 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
54 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
55
56 const __m128i vidx0123 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn0123), vindex_mask), 2);
57 const __m128i ven0123 = _mm_slli_epi32(_mm_castps_si128(vn0123), 19);
58 const __m128i vidx4567 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn4567), vindex_mask), 2);
59 const __m128i ven4567 = _mm_slli_epi32(_mm_castps_si128(vn4567), 19);
60
61 #if XNN_ARCH_X86_64
62 const uint64_t vidx01 = (uint64_t) _mm_cvtsi128_si64(vidx0123);
63 const uint64_t vidx23 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx0123, vidx0123));
64 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx01)));
65 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23)));
66 const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01 >> 32))));
67 const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
68 const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23 >> 32))));
69 const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
70 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
71 const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
72 const uint64_t vidx67 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx4567, vidx4567));
73 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx45)));
74 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67)));
75 const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45 >> 32))));
76 const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
77 const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67 >> 32))));
78 const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
79 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
80 #else // !XNN_ARCH_X86_64
81 const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
82 const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
83 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
84 const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
85 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx0)));
86 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx2)));
87 const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx1)));
88 const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
89 const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx3)));
90 const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
91 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
92 const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567);
93 const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2);
94 const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4);
95 const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6);
96 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx4)));
97 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx6)));
98 const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx5)));
99 const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
100 const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx7)));
101 const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
102 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
103 #endif // XNN_ARCH_X86_64
104
105 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
106 __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ven0123));
107 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
108 __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ven4567));
109
110 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
111 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
112
113 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
114 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
115
116 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc3, vt0123), vc2);
117 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc3, vt4567), vc2);
118
119 vp0123 = _mm_mul_ps(vp0123, vt0123);
120 vp4567 = _mm_mul_ps(vp4567, vt4567);
121
122 vt0123 = _mm_mul_ps(vt0123, vs0123);
123 vs0123 = _mm_sub_ps(vs0123, vone);
124 vt4567 = _mm_mul_ps(vt4567, vs4567);
125 vs4567 = _mm_sub_ps(vs4567, vone);
126
127 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
128 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
129
130 const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
131 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
132
133 const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123)));
134 vx0123 = _mm_mul_ps(vx0123, vbeta);
135 const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567)));
136 vx4567 = _mm_mul_ps(vx4567, vbeta);
137
138 const __m128 vy0123 = _mm_or_ps(_mm_and_ps(ve0123, vm0123), _mm_andnot_ps(vm0123, vx0123));
139 const __m128 vy4567 = _mm_or_ps(_mm_and_ps(ve4567, vm4567), _mm_andnot_ps(vm4567, vx4567));
140
141 _mm_storeu_ps(y, vy0123);
142 _mm_storeu_ps(y + 4, vy4567);
143 y += 8;
144 }
145 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
146 __m128 vx = _mm_loadu_ps(x);
147 x += 4;
148
149 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
150
151 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
152
153 const __m128i ven = _mm_slli_epi32(_mm_castps_si128(vn), 19);
154 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
155 #if XNN_ARCH_X86_64
156 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
157 const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
158 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo)));
159 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi)));
160 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))));
161 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
162 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))));
163 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
164 #else // !XNN_ARCH_X86_64
165 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx))));
166 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 4))));
167 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2))));
168 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
169 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6))));
170 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
171 #endif // XNN_ARCH_X86_64
172 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
173 __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ven));
174 vn = _mm_sub_ps(vn, vmagic_bias);
175
176 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
177 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
178
179 __m128 vp = _mm_add_ps(_mm_mul_ps(vc3, vt), vc2);
180 vp = _mm_mul_ps(vp, vt);
181
182 vt = _mm_mul_ps(vt, vs);
183 vs = _mm_sub_ps(vs, vone);
184 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
185 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
186
187 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
188 vx = _mm_mul_ps(vx, vbeta);
189 const __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));
190
191 _mm_storeu_ps(y, vy);
192 y += 4;
193 }
194 if XNN_UNLIKELY(n != 0) {
195 __m128 vx = _mm_loadu_ps(x);
196
197 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
198
199 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
200
201 const __m128i ven = _mm_slli_epi32(_mm_castps_si128(vn), 19);
202 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
203 #if XNN_ARCH_X86_64
204 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
205 const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
206 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo)));
207 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi)));
208 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))));
209 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
210 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))));
211 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
212 #else // !XNN_ARCH_X86_64
213 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx))));
214 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 4))));
215 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2))));
216 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
217 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6))));
218 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
219 #endif // XNN_ARCH_X86_64
220 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
221 __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ven));
222 vn = _mm_sub_ps(vn, vmagic_bias);
223
224 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
225 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
226
227 __m128 vp = _mm_add_ps(_mm_mul_ps(vc3, vt), vc2);
228 vp = _mm_mul_ps(vp, vt);
229
230 vt = _mm_mul_ps(vt, vs);
231 vs = _mm_sub_ps(vs, vone);
232 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
233 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
234
235 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
236 vx = _mm_mul_ps(vx, vbeta);
237 __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));
238
239 if (n & (2 * sizeof(float))) {
240 _mm_storel_pi((__m64*) y, vy);
241 vy = _mm_movehl_ps(vy, vy);
242 y += 2;
243 }
244 if (n & (1 * sizeof(float))) {
245 _mm_store_ss(y, vy);
246 }
247 }
248 }
249