1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/wasmsimd-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 <wasm_simd128.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__wasmsimd_x86_rr2_lut16_p3_x20(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__wasmsimd_x86_rr2_lut16_p3_x20(
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 v128_t vprescale = wasm_v32x4_load_splat(¶ms->scalar.prescale);
32 const v128_t valpha = wasm_v32x4_load_splat(¶ms->scalar.alpha);
33 const v128_t vbeta = wasm_v32x4_load_splat(¶ms->scalar.beta);
34
35 const v128_t vsat_cutoff = wasm_f32x4_splat(-0x1.154246p+4f);
36 const v128_t vmagic_bias = wasm_f32x4_splat(0x1.800000p19f);
37 const v128_t vlog2e = wasm_f32x4_splat(0x1.715476p+0f);
38 const v128_t vindex_mask = wasm_i32x4_splat(0xF);
39 const v128_t vminus_ln2_hi = wasm_f32x4_splat(-0x1.62E400p-1f);
40 const v128_t vminus_ln2_lo = wasm_f32x4_splat(-0x1.7F7D1Cp-20f);
41 const v128_t vc3 = wasm_f32x4_splat(0x1.55561Cp-3f);
42 const v128_t vc2 = wasm_f32x4_splat(0x1.0001ECp-1f);
43 const v128_t vone = wasm_f32x4_splat(1.0f);
44
45 for (; n >= 20 * sizeof(float); n -= 20 * sizeof(float)) {
46 v128_t vx0123 = wasm_v128_load(x);
47 v128_t vx4567 = wasm_v128_load(x + 4);
48 v128_t vx89AB = wasm_v128_load(x + 8);
49 v128_t vxCDEF = wasm_v128_load(x + 12);
50 v128_t vxGHIJ = wasm_v128_load(x + 16);
51 x += 20;
52
53 const v128_t vz0123 = wasm_f32x4_mul(vx0123, vprescale);
54 const v128_t vz4567 = wasm_f32x4_mul(vx4567, vprescale);
55 const v128_t vz89AB = wasm_f32x4_mul(vx89AB, vprescale);
56 const v128_t vzCDEF = wasm_f32x4_mul(vxCDEF, vprescale);
57 const v128_t vzGHIJ = wasm_f32x4_mul(vxGHIJ, vprescale);
58
59 v128_t vn0123 = wasm_f32x4_add(wasm_f32x4_mul(vz0123, vlog2e), vmagic_bias);
60 v128_t vn4567 = wasm_f32x4_add(wasm_f32x4_mul(vz4567, vlog2e), vmagic_bias);
61 v128_t vn89AB = wasm_f32x4_add(wasm_f32x4_mul(vz89AB, vlog2e), vmagic_bias);
62 v128_t vnCDEF = wasm_f32x4_add(wasm_f32x4_mul(vzCDEF, vlog2e), vmagic_bias);
63 v128_t vnGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vzGHIJ, vlog2e), vmagic_bias);
64
65 const v128_t vidx0123 = wasm_i32x4_shl(wasm_v128_and(vn0123, vindex_mask), 2);
66 const v128_t ven0123 = wasm_i32x4_shl(vn0123, 19);
67 const v128_t vidx4567 = wasm_i32x4_shl(wasm_v128_and(vn4567, vindex_mask), 2);
68 const v128_t ven4567 = wasm_i32x4_shl(vn4567, 19);
69 const v128_t vidx89AB = wasm_i32x4_shl(wasm_v128_and(vn89AB, vindex_mask), 2);
70 const v128_t ven89AB = wasm_i32x4_shl(vn89AB, 19);
71 const v128_t vidxCDEF = wasm_i32x4_shl(wasm_v128_and(vnCDEF, vindex_mask), 2);
72 const v128_t venCDEF = wasm_i32x4_shl(vnCDEF, 19);
73 const v128_t vidxGHIJ = wasm_i32x4_shl(wasm_v128_and(vnGHIJ, vindex_mask), 2);
74 const v128_t venGHIJ = wasm_i32x4_shl(vnGHIJ, 19);
75
76 const uint64_t vidx01 = wasm_i64x2_extract_lane(vidx0123, 0);
77 const uint64_t vidx23 = wasm_i64x2_extract_lane(vidx0123, 1);
78 const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx01));
79 const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01 >> 32)));
80 const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23));
81 const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23 >> 32)));
82 const v128_t vl0123 = wasm_f32x4_make(vl0, vl1, vl2, vl3);
83 const uint64_t vidx45 = wasm_i64x2_extract_lane(vidx4567, 0);
84 const uint64_t vidx67 = wasm_i64x2_extract_lane(vidx4567, 1);
85 const float vl4 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx45));
86 const float vl5 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45 >> 32)));
87 const float vl6 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67));
88 const float vl7 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67 >> 32)));
89 const v128_t vl4567 = wasm_f32x4_make(vl4, vl5, vl6, vl7);
90 const uint64_t vidx89 = wasm_i64x2_extract_lane(vidx89AB, 0);
91 const uint64_t vidxAB = wasm_i64x2_extract_lane(vidx89AB, 1);
92 const float vl8 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx89));
93 const float vl9 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx89 >> 32)));
94 const float vlA = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxAB));
95 const float vlB = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxAB >> 32)));
96 const v128_t vl89AB = wasm_f32x4_make(vl8, vl9, vlA, vlB);
97 const uint64_t vidxCD = wasm_i64x2_extract_lane(vidxCDEF, 0);
98 const uint64_t vidxEF = wasm_i64x2_extract_lane(vidxCDEF, 1);
99 const float vlC = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxCD));
100 const float vlD = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxCD >> 32)));
101 const float vlE = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxEF));
102 const float vlF = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxEF >> 32)));
103 const v128_t vlCDEF = wasm_f32x4_make(vlC, vlD, vlE, vlF);
104 const uint64_t vidxGH = wasm_i64x2_extract_lane(vidxGHIJ, 0);
105 const uint64_t vidxIJ = wasm_i64x2_extract_lane(vidxGHIJ, 1);
106 const float vlG = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxGH));
107 const float vlH = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxGH >> 32)));
108 const float vlI = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxIJ));
109 const float vlJ = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxIJ >> 32)));
110 const v128_t vlGHIJ = wasm_f32x4_make(vlG, vlH, vlI, vlJ);
111
112 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
113 v128_t vs0123 = wasm_i32x4_add(vl0123, ven0123);
114 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
115 v128_t vs4567 = wasm_i32x4_add(vl4567, ven4567);
116 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
117 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB);
118 vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
119 v128_t vsCDEF = wasm_i32x4_add(vlCDEF, venCDEF);
120 vnGHIJ = wasm_f32x4_sub(vnGHIJ, vmagic_bias);
121 v128_t vsGHIJ = wasm_i32x4_add(vlGHIJ, venGHIJ);
122
123 v128_t vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_hi), vz0123);
124 const v128_t vsatm0123 = wasm_f32x4_le(vz0123, vsat_cutoff);
125 v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567);
126 const v128_t vsatm4567 = wasm_f32x4_le(vz4567, vsat_cutoff);
127 v128_t vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_hi), vz89AB);
128 const v128_t vsatm89AB = wasm_f32x4_le(vz89AB, vsat_cutoff);
129 v128_t vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_hi), vzCDEF);
130 const v128_t vsatmCDEF = wasm_f32x4_le(vzCDEF, vsat_cutoff);
131 v128_t vtGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vnGHIJ, vminus_ln2_hi), vzGHIJ);
132 const v128_t vsatmGHIJ = wasm_f32x4_le(vzGHIJ, vsat_cutoff);
133
134 vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_lo), vt0123);
135 vs0123 = wasm_v128_andnot(vs0123, vsatm0123);
136 vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567);
137 vs4567 = wasm_v128_andnot(vs4567, vsatm4567);
138 vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_lo), vt89AB);
139 vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB);
140 vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_lo), vtCDEF);
141 vsCDEF = wasm_v128_andnot(vsCDEF, vsatmCDEF);
142 vtGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vnGHIJ, vminus_ln2_lo), vtGHIJ);
143 vsGHIJ = wasm_v128_andnot(vsGHIJ, vsatmGHIJ);
144
145 vt0123 = wasm_v128_andnot(vt0123, vsatm0123);
146 v128_t vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt0123), vc2);
147 vt4567 = wasm_v128_andnot(vt4567, vsatm4567);
148 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt4567), vc2);
149 vt89AB = wasm_v128_andnot(vt89AB, vsatm89AB);
150 v128_t vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt89AB), vc2);
151 vtCDEF = wasm_v128_andnot(vtCDEF, vsatmCDEF);
152 v128_t vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vc3, vtCDEF), vc2);
153 vtGHIJ = wasm_v128_andnot(vtGHIJ, vsatmGHIJ);
154 v128_t vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vc3, vtGHIJ), vc2);
155
156 vp0123 = wasm_f32x4_mul(vp0123, vt0123);
157 vp4567 = wasm_f32x4_mul(vp4567, vt4567);
158 vp89AB = wasm_f32x4_mul(vp89AB, vt89AB);
159 vpCDEF = wasm_f32x4_mul(vpCDEF, vtCDEF);
160 vpGHIJ = wasm_f32x4_mul(vpGHIJ, vtGHIJ);
161
162 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
163 vs0123 = wasm_f32x4_sub(vs0123, vone);
164 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
165 vs4567 = wasm_f32x4_sub(vs4567, vone);
166 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
167 vs89AB = wasm_f32x4_sub(vs89AB, vone);
168 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
169 vsCDEF = wasm_f32x4_sub(vsCDEF, vone);
170 vtGHIJ = wasm_f32x4_mul(vtGHIJ, vsGHIJ);
171 vsGHIJ = wasm_f32x4_sub(vsGHIJ, vone);
172
173 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vt0123);
174 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567);
175 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vt89AB);
176 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vtCDEF);
177 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vtGHIJ);
178
179 const v128_t ve0123 = wasm_f32x4_mul(wasm_f32x4_add(vp0123, vs0123), valpha);
180 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha);
181 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha);
182 const v128_t veCDEF = wasm_f32x4_mul(wasm_f32x4_add(vpCDEF, vsCDEF), valpha);
183 const v128_t veGHIJ = wasm_f32x4_mul(wasm_f32x4_add(vpGHIJ, vsGHIJ), valpha);
184
185 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31);
186 vx0123 = wasm_f32x4_mul(vx0123, vbeta);
187 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31);
188 vx4567 = wasm_f32x4_mul(vx4567, vbeta);
189 const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31);
190 vx89AB = wasm_f32x4_mul(vx89AB, vbeta);
191 const v128_t vsignmCDEF = wasm_i32x4_shr(vxCDEF, 31);
192 vxCDEF = wasm_f32x4_mul(vxCDEF, vbeta);
193 const v128_t vsignmGHIJ = wasm_i32x4_shr(vxGHIJ, 31);
194 vxGHIJ = wasm_f32x4_mul(vxGHIJ, vbeta);
195
196 const v128_t vy0123 = wasm_v128_bitselect(ve0123, vx0123, vsignm0123);
197 const v128_t vy4567 = wasm_v128_bitselect(ve4567, vx4567, vsignm4567);
198 const v128_t vy89AB = wasm_v128_bitselect(ve89AB, vx89AB, vsignm89AB);
199 const v128_t vyCDEF = wasm_v128_bitselect(veCDEF, vxCDEF, vsignmCDEF);
200 const v128_t vyGHIJ = wasm_v128_bitselect(veGHIJ, vxGHIJ, vsignmGHIJ);
201
202 wasm_v128_store(y, vy0123);
203 wasm_v128_store(y + 4, vy4567);
204 wasm_v128_store(y + 8, vy89AB);
205 wasm_v128_store(y + 12, vyCDEF);
206 wasm_v128_store(y + 16, vyGHIJ);
207 y += 20;
208 }
209 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
210 v128_t vx = wasm_v128_load(x);
211 x += 4;
212
213 const v128_t vz = wasm_f32x4_mul(vx, vprescale);
214
215 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
216 const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
217 const v128_t ven = wasm_i32x4_shl(vn, 19);
218
219 const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
220 const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
221 const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
222 const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)));
223 const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
224 const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)));
225 const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);
226
227 v128_t vs = wasm_i32x4_add(vl, ven);
228 vn = wasm_f32x4_sub(vn, vmagic_bias);
229
230 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
231 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
232 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
233 vs = wasm_v128_andnot(vs, vsatm);
234 vt = wasm_v128_andnot(vt, vsatm);
235
236 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt), vc2);
237 vp = wasm_f32x4_mul(vp, vt);
238
239 vt = wasm_f32x4_mul(vt, vs);
240 vs = wasm_f32x4_sub(vs, vone);
241 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
242 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
243
244 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
245 vx = wasm_f32x4_mul(vx, vbeta);
246 const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
247
248 wasm_v128_store(y, vy);
249 y += 4;
250 }
251 if XNN_UNLIKELY(n != 0) {
252 v128_t vx = wasm_v128_load(x);
253
254 const v128_t vz = wasm_f32x4_mul(vx, vprescale);
255
256 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
257 const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
258 const v128_t ven = wasm_i32x4_shl(vn, 19);
259
260 const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
261 const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
262 const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
263 const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)));
264 const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
265 const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)));
266 const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);
267
268 v128_t vs = wasm_i32x4_add(vl, ven);
269 vn = wasm_f32x4_sub(vn, vmagic_bias);
270
271 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
272 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
273 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
274 vs = wasm_v128_andnot(vs, vsatm);
275 vt = wasm_v128_andnot(vt, vsatm);
276
277 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt), vc2);
278 vp = wasm_f32x4_mul(vp, vt);
279
280 vt = wasm_f32x4_mul(vt, vs);
281 vs = wasm_f32x4_sub(vs, vone);
282 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
283 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
284
285 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
286 vx = wasm_f32x4_mul(vx, vbeta);
287 v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
288
289 if (n & (2 * sizeof(float))) {
290 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
291 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
292 y += 2;
293 }
294 if (n & (1 * sizeof(float))) {
295 *y = wasm_f32x4_extract_lane(vy, 0);
296 }
297 }
298 }
299