1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/wasmsimd-rr2-lut64-p2-div.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/common.h>
15 #include <xnnpack/vunary.h>
16
17
18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_64[64];
19
xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_lut64_p2_div_x20(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_lut64_p2_div_x20(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(n % sizeof(float) == 0);
27
28 const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd_rr2_lut64_p2.magic_bias);
29 const v128_t vminus_log2e = wasm_v128_load64_splat(params->wasmsimd_rr2_lut64_p2.minus_log2e);
30 const v128_t vindex_mask = wasm_v128_load64_splat(params->wasmsimd_rr2_lut64_p2.index_mask);
31 const v128_t vln2_hi = wasm_v128_load64_splat(params->wasmsimd_rr2_lut64_p2.ln2_hi);
32 const v128_t vln2_lo = wasm_v128_load64_splat(params->wasmsimd_rr2_lut64_p2.ln2_lo);
33 const v128_t vc2 = wasm_v128_load64_splat(params->wasmsimd_rr2_lut64_p2.c2);
34 const v128_t vone = wasm_v128_load64_splat(params->wasmsimd_rr2_lut64_p2.one);
35 const v128_t vdenorm_cutoff = wasm_v128_load64_splat(params->wasmsimd_rr2_lut64_p2.denorm_cutoff);
36
37 for (; n >= 20 * sizeof(float); n -= 20 * sizeof(float)) {
38 const v128_t vx0123 = wasm_v128_load(x);
39 const v128_t vx4567 = wasm_v128_load(x + 4);
40 const v128_t vx89AB = wasm_v128_load(x + 8);
41 const v128_t vxCDEF = wasm_v128_load(x + 12);
42 const v128_t vxGHIJ = wasm_v128_load(x + 16);
43 x += 20;
44
45 const v128_t vz0123 = wasm_f32x4_abs(vx0123);
46 const v128_t vz4567 = wasm_f32x4_abs(vx4567);
47 const v128_t vz89AB = wasm_f32x4_abs(vx89AB);
48 const v128_t vzCDEF = wasm_f32x4_abs(vxCDEF);
49 const v128_t vzGHIJ = wasm_f32x4_abs(vxGHIJ);
50
51 v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz0123, vminus_log2e));
52 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz4567, vminus_log2e));
53 v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz89AB, vminus_log2e));
54 v128_t vnCDEF = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzCDEF, vminus_log2e));
55 v128_t vnGHIJ = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzGHIJ, vminus_log2e));
56
57 const v128_t ve0123 = wasm_i32x4_shl(vn0123, 17);
58 const v128_t ve4567 = wasm_i32x4_shl(vn4567, 17);
59 const v128_t ve89AB = wasm_i32x4_shl(vn89AB, 17);
60 const v128_t veCDEF = wasm_i32x4_shl(vnCDEF, 17);
61 const v128_t veGHIJ = wasm_i32x4_shl(vnGHIJ, 17);
62
63 const v128_t vidx0123 = wasm_i32x4_shl(wasm_v128_and(vn0123, vindex_mask), 2);
64 const v128_t vidx4567 = wasm_i32x4_shl(wasm_v128_and(vn4567, vindex_mask), 2);
65 const v128_t vidx89AB = wasm_i32x4_shl(wasm_v128_and(vn89AB, vindex_mask), 2);
66 const v128_t vidxCDEF = wasm_i32x4_shl(wasm_v128_and(vnCDEF, vindex_mask), 2);
67 const v128_t vidxGHIJ = wasm_i32x4_shl(wasm_v128_and(vnGHIJ, vindex_mask), 2);
68
69 const uint64_t vidx01 = wasm_i64x2_extract_lane(vidx0123, 0);
70 const uint64_t vidx23 = wasm_i64x2_extract_lane(vidx0123, 1);
71 const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx01));
72 const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32)));
73 const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx23));
74 const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32)));
75 const v128_t vl0123 = wasm_f32x4_make(vl0, vl1, vl2, vl3);
76 const uint64_t vidx45 = wasm_i64x2_extract_lane(vidx4567, 0);
77 const uint64_t vidx67 = wasm_i64x2_extract_lane(vidx4567, 1);
78 const float vl4 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45));
79 const float vl5 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32)));
80 const float vl6 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67));
81 const float vl7 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32)));
82 const v128_t vl4567 = wasm_f32x4_make(vl4, vl5, vl6, vl7);
83 const uint64_t vidx89 = wasm_i64x2_extract_lane(vidx89AB, 0);
84 const uint64_t vidxAB = wasm_i64x2_extract_lane(vidx89AB, 1);
85 const float vl8 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx89));
86 const float vl9 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx89 >> 32)));
87 const float vlA = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxAB));
88 const float vlB = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxAB >> 32)));
89 const v128_t vl89AB = wasm_f32x4_make(vl8, vl9, vlA, vlB);
90 const uint64_t vidxCD = wasm_i64x2_extract_lane(vidxCDEF, 0);
91 const uint64_t vidxEF = wasm_i64x2_extract_lane(vidxCDEF, 1);
92 const float vlC = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxCD));
93 const float vlD = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxCD >> 32)));
94 const float vlE = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxEF));
95 const float vlF = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxEF >> 32)));
96 const v128_t vlCDEF = wasm_f32x4_make(vlC, vlD, vlE, vlF);
97 const uint64_t vidxGH = wasm_i64x2_extract_lane(vidxGHIJ, 0);
98 const uint64_t vidxIJ = wasm_i64x2_extract_lane(vidxGHIJ, 1);
99 const float vlG = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxGH));
100 const float vlH = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxGH >> 32)));
101 const float vlI = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxIJ));
102 const float vlJ = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxIJ >> 32)));
103 const v128_t vlGHIJ = wasm_f32x4_make(vlG, vlH, vlI, vlJ);
104
105 const v128_t vs0123 = wasm_i32x4_add(vl0123, ve0123);
106 const v128_t vs4567 = wasm_i32x4_add(vl4567, ve4567);
107 const v128_t vs89AB = wasm_i32x4_add(vl89AB, ve89AB);
108 const v128_t vsCDEF = wasm_i32x4_add(vlCDEF, veCDEF);
109 const v128_t vsGHIJ = wasm_i32x4_add(vlGHIJ, veGHIJ);
110
111 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
112 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
113 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
114 vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
115 vnGHIJ = wasm_f32x4_sub(vnGHIJ, vmagic_bias);
116
117 v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi));
118 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi));
119 v128_t vt89AB = wasm_f32x4_add(vz89AB, wasm_f32x4_mul(vn89AB, vln2_hi));
120 v128_t vtCDEF = wasm_f32x4_add(vzCDEF, wasm_f32x4_mul(vnCDEF, vln2_hi));
121 v128_t vtGHIJ = wasm_f32x4_add(vzGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_hi));
122
123 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo));
124 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo));
125 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vln2_lo));
126 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vln2_lo));
127 vtGHIJ = wasm_f32x4_add(vtGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_lo));
128
129 v128_t vp0123 = wasm_f32x4_mul(vt0123, vc2);
130 v128_t vp4567 = wasm_f32x4_mul(vt4567, vc2);
131 v128_t vp89AB = wasm_f32x4_mul(vt89AB, vc2);
132 v128_t vpCDEF = wasm_f32x4_mul(vtCDEF, vc2);
133 v128_t vpGHIJ = wasm_f32x4_mul(vtGHIJ, vc2);
134
135 vp0123 = wasm_f32x4_sub(vt0123, wasm_f32x4_mul(vp0123, vt0123));
136 vp4567 = wasm_f32x4_sub(vt4567, wasm_f32x4_mul(vp4567, vt4567));
137 vp89AB = wasm_f32x4_sub(vt89AB, wasm_f32x4_mul(vp89AB, vt89AB));
138 vpCDEF = wasm_f32x4_sub(vtCDEF, wasm_f32x4_mul(vpCDEF, vtCDEF));
139 vpGHIJ = wasm_f32x4_sub(vtGHIJ, wasm_f32x4_mul(vpGHIJ, vtGHIJ));
140
141 const v128_t vy0123 = wasm_f32x4_sub(vs0123, wasm_f32x4_mul(vs0123, vp0123));
142 const v128_t vy4567 = wasm_f32x4_sub(vs4567, wasm_f32x4_mul(vs4567, vp4567));
143 const v128_t vy89AB = wasm_f32x4_sub(vs89AB, wasm_f32x4_mul(vs89AB, vp89AB));
144 const v128_t vyCDEF = wasm_f32x4_sub(vsCDEF, wasm_f32x4_mul(vsCDEF, vpCDEF));
145 const v128_t vyGHIJ = wasm_f32x4_sub(vsGHIJ, wasm_f32x4_mul(vsGHIJ, vpGHIJ));
146
147 const v128_t vd0123 = wasm_f32x4_add(vy0123, vone);
148 const v128_t vd4567 = wasm_f32x4_add(vy4567, vone);
149 const v128_t vd89AB = wasm_f32x4_add(vy89AB, vone);
150 const v128_t vdCDEF = wasm_f32x4_add(vyCDEF, vone);
151 const v128_t vdGHIJ = wasm_f32x4_add(vyGHIJ, vone);
152
153 v128_t vf0123 = wasm_f32x4_div(vy0123, vd0123);
154 v128_t vf4567 = wasm_f32x4_div(vy4567, vd4567);
155 v128_t vf89AB = wasm_f32x4_div(vy89AB, vd89AB);
156 v128_t vfCDEF = wasm_f32x4_div(vyCDEF, vdCDEF);
157 v128_t vfGHIJ = wasm_f32x4_div(vyGHIJ, vdGHIJ);
158
159 vf0123 = wasm_v128_andnot(vf0123, wasm_f32x4_gt(vz0123, vdenorm_cutoff));
160 vf4567 = wasm_v128_andnot(vf4567, wasm_f32x4_gt(vz4567, vdenorm_cutoff));
161 vf89AB = wasm_v128_andnot(vf89AB, wasm_f32x4_gt(vz89AB, vdenorm_cutoff));
162 vfCDEF = wasm_v128_andnot(vfCDEF, wasm_f32x4_gt(vzCDEF, vdenorm_cutoff));
163 vfGHIJ = wasm_v128_andnot(vfGHIJ, wasm_f32x4_gt(vzGHIJ, vdenorm_cutoff));
164
165 vf0123 = wasm_v128_bitselect(vf0123, wasm_f32x4_sub(vone, vf0123), wasm_i32x4_shr(vx0123, 31));
166 vf4567 = wasm_v128_bitselect(vf4567, wasm_f32x4_sub(vone, vf4567), wasm_i32x4_shr(vx4567, 31));
167 vf89AB = wasm_v128_bitselect(vf89AB, wasm_f32x4_sub(vone, vf89AB), wasm_i32x4_shr(vx89AB, 31));
168 vfCDEF = wasm_v128_bitselect(vfCDEF, wasm_f32x4_sub(vone, vfCDEF), wasm_i32x4_shr(vxCDEF, 31));
169 vfGHIJ = wasm_v128_bitselect(vfGHIJ, wasm_f32x4_sub(vone, vfGHIJ), wasm_i32x4_shr(vxGHIJ, 31));
170
171 wasm_v128_store(y, vf0123);
172 wasm_v128_store(y + 4, vf4567);
173 wasm_v128_store(y + 8, vf89AB);
174 wasm_v128_store(y + 12, vfCDEF);
175 wasm_v128_store(y + 16, vfGHIJ);
176 y += 20;
177 }
178 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
179 const v128_t vx = wasm_v128_load(x);
180 x += 4;
181
182 const v128_t vz = wasm_f32x4_abs(vx);
183
184 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e));
185 const v128_t ve = wasm_i32x4_shl(vn, 17);
186
187 const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
188 const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
189 const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
190 const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo));
191 const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32)));
192 const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi));
193 const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32)));
194 const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);
195
196 const v128_t vs = wasm_i32x4_add(vl, ve);
197 vn = wasm_f32x4_sub(vn, vmagic_bias);
198
199 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi));
200 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo));
201
202 v128_t vp = wasm_f32x4_mul(vt, vc2);
203 vp = wasm_f32x4_sub(vt, wasm_f32x4_mul(vp, vt));
204
205 const v128_t vy = wasm_f32x4_sub(vs, wasm_f32x4_mul(vs, vp));
206 const v128_t vd = wasm_f32x4_add(vy, vone);
207
208 v128_t vf = wasm_f32x4_div(vy, vd);
209 vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
210 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));
211
212 wasm_v128_store(y, vf);
213 y += 4;
214 }
215 if XNN_UNLIKELY(n != 0) {
216 const v128_t vx = wasm_v128_load(x);
217
218 const v128_t vz = wasm_f32x4_abs(vx);
219
220 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e));
221 const v128_t ve = wasm_i32x4_shl(vn, 17);
222
223 const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
224 const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
225 const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
226 const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo));
227 const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32)));
228 const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi));
229 const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32)));
230 const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);
231
232 const v128_t vs = wasm_i32x4_add(vl, ve);
233 vn = wasm_f32x4_sub(vn, vmagic_bias);
234
235 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi));
236 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo));
237
238 v128_t vp = wasm_f32x4_mul(vt, vc2);
239 vp = wasm_f32x4_sub(vt, wasm_f32x4_mul(vp, vt));
240
241 const v128_t vy = wasm_f32x4_sub(vs, wasm_f32x4_mul(vs, vp));
242 const v128_t vd = wasm_f32x4_add(vy, vone);
243
244 v128_t vf = wasm_f32x4_div(vy, vd);
245 vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
246 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));
247
248 if (n & (2 * sizeof(float))) {
249 *((double*) y) = wasm_f64x2_extract_lane(vf, 0);
250 vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3);
251 y += 2;
252 }
253 if (n & (1 * sizeof(float))) {
254 *y = wasm_f32x4_extract_lane(vf, 0);
255 }
256 }
257 }
258