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_x16(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_x16(
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(&params->scalar.prescale);
32   const v128_t valpha = wasm_v32x4_load_splat(&params->scalar.alpha);
33   const v128_t vbeta = wasm_v32x4_load_splat(&params->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 >= 16 * sizeof(float); n -= 16 * 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     x += 16;
51 
52     const v128_t vz0123 = wasm_f32x4_mul(vx0123, vprescale);
53     const v128_t vz4567 = wasm_f32x4_mul(vx4567, vprescale);
54     const v128_t vz89AB = wasm_f32x4_mul(vx89AB, vprescale);
55     const v128_t vzCDEF = wasm_f32x4_mul(vxCDEF, vprescale);
56 
57     v128_t vn0123 = wasm_f32x4_add(wasm_f32x4_mul(vz0123, vlog2e), vmagic_bias);
58     v128_t vn4567 = wasm_f32x4_add(wasm_f32x4_mul(vz4567, vlog2e), vmagic_bias);
59     v128_t vn89AB = wasm_f32x4_add(wasm_f32x4_mul(vz89AB, vlog2e), vmagic_bias);
60     v128_t vnCDEF = wasm_f32x4_add(wasm_f32x4_mul(vzCDEF, vlog2e), vmagic_bias);
61 
62     const v128_t vidx0123 = wasm_i32x4_shl(wasm_v128_and(vn0123, vindex_mask), 2);
63     const v128_t ven0123 = wasm_i32x4_shl(vn0123, 19);
64     const v128_t vidx4567 = wasm_i32x4_shl(wasm_v128_and(vn4567, vindex_mask), 2);
65     const v128_t ven4567 = wasm_i32x4_shl(vn4567, 19);
66     const v128_t vidx89AB = wasm_i32x4_shl(wasm_v128_and(vn89AB, vindex_mask), 2);
67     const v128_t ven89AB = wasm_i32x4_shl(vn89AB, 19);
68     const v128_t vidxCDEF = wasm_i32x4_shl(wasm_v128_and(vnCDEF, vindex_mask), 2);
69     const v128_t venCDEF = wasm_i32x4_shl(vnCDEF, 19);
70 
71     const uint64_t vidx01 = wasm_i64x2_extract_lane(vidx0123, 0);
72     const uint64_t vidx23 = wasm_i64x2_extract_lane(vidx0123, 1);
73     const float vl0   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx01));
74     const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01 >> 32)));
75     const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23));
76     const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23 >> 32)));
77     const v128_t vl0123 = wasm_f32x4_make(vl0, vl1, vl2, vl3);
78     const uint64_t vidx45 = wasm_i64x2_extract_lane(vidx4567, 0);
79     const uint64_t vidx67 = wasm_i64x2_extract_lane(vidx4567, 1);
80     const float vl4   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx45));
81     const float vl5 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45 >> 32)));
82     const float vl6 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67));
83     const float vl7 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67 >> 32)));
84     const v128_t vl4567 = wasm_f32x4_make(vl4, vl5, vl6, vl7);
85     const uint64_t vidx89 = wasm_i64x2_extract_lane(vidx89AB, 0);
86     const uint64_t vidxAB = wasm_i64x2_extract_lane(vidx89AB, 1);
87     const float vl8   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx89));
88     const float vl9 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx89 >> 32)));
89     const float vlA = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxAB));
90     const float vlB = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxAB >> 32)));
91     const v128_t vl89AB = wasm_f32x4_make(vl8, vl9, vlA, vlB);
92     const uint64_t vidxCD = wasm_i64x2_extract_lane(vidxCDEF, 0);
93     const uint64_t vidxEF = wasm_i64x2_extract_lane(vidxCDEF, 1);
94     const float vlC   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxCD));
95     const float vlD = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxCD >> 32)));
96     const float vlE = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxEF));
97     const float vlF = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxEF >> 32)));
98     const v128_t vlCDEF = wasm_f32x4_make(vlC, vlD, vlE, vlF);
99 
100     vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
101     v128_t vs0123 = wasm_i32x4_add(vl0123, ven0123);
102     vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
103     v128_t vs4567 = wasm_i32x4_add(vl4567, ven4567);
104     vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
105     v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB);
106     vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
107     v128_t vsCDEF = wasm_i32x4_add(vlCDEF, venCDEF);
108 
109     v128_t vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_hi), vz0123);
110     const v128_t vsatm0123 = wasm_f32x4_le(vz0123, vsat_cutoff);
111     v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567);
112     const v128_t vsatm4567 = wasm_f32x4_le(vz4567, vsat_cutoff);
113     v128_t vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_hi), vz89AB);
114     const v128_t vsatm89AB = wasm_f32x4_le(vz89AB, vsat_cutoff);
115     v128_t vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_hi), vzCDEF);
116     const v128_t vsatmCDEF = wasm_f32x4_le(vzCDEF, vsat_cutoff);
117 
118     vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_lo), vt0123);
119     vs0123 = wasm_v128_andnot(vs0123, vsatm0123);
120     vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567);
121     vs4567 = wasm_v128_andnot(vs4567, vsatm4567);
122     vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_lo), vt89AB);
123     vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB);
124     vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_lo), vtCDEF);
125     vsCDEF = wasm_v128_andnot(vsCDEF, vsatmCDEF);
126 
127     vt0123 = wasm_v128_andnot(vt0123, vsatm0123);
128     v128_t vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt0123), vc2);
129     vt4567 = wasm_v128_andnot(vt4567, vsatm4567);
130     v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt4567), vc2);
131     vt89AB = wasm_v128_andnot(vt89AB, vsatm89AB);
132     v128_t vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt89AB), vc2);
133     vtCDEF = wasm_v128_andnot(vtCDEF, vsatmCDEF);
134     v128_t vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vc3, vtCDEF), vc2);
135 
136     vp0123 = wasm_f32x4_mul(vp0123, vt0123);
137     vp4567 = wasm_f32x4_mul(vp4567, vt4567);
138     vp89AB = wasm_f32x4_mul(vp89AB, vt89AB);
139     vpCDEF = wasm_f32x4_mul(vpCDEF, vtCDEF);
140 
141     vt0123 = wasm_f32x4_mul(vt0123, vs0123);
142     vs0123 = wasm_f32x4_sub(vs0123, vone);
143     vt4567 = wasm_f32x4_mul(vt4567, vs4567);
144     vs4567 = wasm_f32x4_sub(vs4567, vone);
145     vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
146     vs89AB = wasm_f32x4_sub(vs89AB, vone);
147     vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
148     vsCDEF = wasm_f32x4_sub(vsCDEF, vone);
149 
150     vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vt0123);
151     vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567);
152     vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vt89AB);
153     vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vtCDEF);
154 
155     const v128_t ve0123 = wasm_f32x4_mul(wasm_f32x4_add(vp0123, vs0123), valpha);
156     const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha);
157     const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha);
158     const v128_t veCDEF = wasm_f32x4_mul(wasm_f32x4_add(vpCDEF, vsCDEF), valpha);
159 
160     const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31);
161     vx0123 = wasm_f32x4_mul(vx0123, vbeta);
162     const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31);
163     vx4567 = wasm_f32x4_mul(vx4567, vbeta);
164     const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31);
165     vx89AB = wasm_f32x4_mul(vx89AB, vbeta);
166     const v128_t vsignmCDEF = wasm_i32x4_shr(vxCDEF, 31);
167     vxCDEF = wasm_f32x4_mul(vxCDEF, vbeta);
168 
169     const v128_t vy0123 = wasm_v128_bitselect(ve0123, vx0123, vsignm0123);
170     const v128_t vy4567 = wasm_v128_bitselect(ve4567, vx4567, vsignm4567);
171     const v128_t vy89AB = wasm_v128_bitselect(ve89AB, vx89AB, vsignm89AB);
172     const v128_t vyCDEF = wasm_v128_bitselect(veCDEF, vxCDEF, vsignmCDEF);
173 
174     wasm_v128_store(y, vy0123);
175     wasm_v128_store(y + 4, vy4567);
176     wasm_v128_store(y + 8, vy89AB);
177     wasm_v128_store(y + 12, vyCDEF);
178     y += 16;
179   }
180   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
181     v128_t vx = wasm_v128_load(x);
182     x += 4;
183 
184     const v128_t vz = wasm_f32x4_mul(vx, vprescale);
185 
186     v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
187     const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
188     const v128_t ven = wasm_i32x4_shl(vn, 19);
189 
190     const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
191     const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
192     const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
193     const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)));
194     const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
195     const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)));
196     const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);
197 
198     v128_t vs = wasm_i32x4_add(vl, ven);
199     vn = wasm_f32x4_sub(vn, vmagic_bias);
200 
201     v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
202     const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
203     vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
204     vs = wasm_v128_andnot(vs, vsatm);
205     vt = wasm_v128_andnot(vt, vsatm);
206 
207     v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt), vc2);
208     vp = wasm_f32x4_mul(vp, vt);
209 
210     vt = wasm_f32x4_mul(vt, vs);
211     vs = wasm_f32x4_sub(vs, vone);
212     vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
213     const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
214 
215     const v128_t vsignm = wasm_i32x4_shr(vx, 31);
216     vx = wasm_f32x4_mul(vx, vbeta);
217     const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
218 
219     wasm_v128_store(y, vy);
220     y += 4;
221   }
222   if XNN_UNLIKELY(n != 0) {
223     v128_t vx = wasm_v128_load(x);
224 
225     const v128_t vz = wasm_f32x4_mul(vx, vprescale);
226 
227     v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
228     const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
229     const v128_t ven = wasm_i32x4_shl(vn, 19);
230 
231     const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
232     const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
233     const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
234     const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)));
235     const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
236     const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)));
237     const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);
238 
239     v128_t vs = wasm_i32x4_add(vl, ven);
240     vn = wasm_f32x4_sub(vn, vmagic_bias);
241 
242     v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
243     const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
244     vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
245     vs = wasm_v128_andnot(vs, vsatm);
246     vt = wasm_v128_andnot(vt, vsatm);
247 
248     v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt), vc2);
249     vp = wasm_f32x4_mul(vp, vt);
250 
251     vt = wasm_f32x4_mul(vt, vs);
252     vs = wasm_f32x4_sub(vs, vone);
253     vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
254     const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
255 
256     const v128_t vsignm = wasm_i32x4_shr(vx, 31);
257     vx = wasm_f32x4_mul(vx, vbeta);
258     v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
259 
260     if (n & (2 * sizeof(float))) {
261       *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
262       vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
263       y += 2;
264     }
265     if (n & (1 * sizeof(float))) {
266       *y = wasm_f32x4_extract_lane(vy, 0);
267     }
268   }
269 }
270