xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-wasmsimd-rr2-lut64-p2-div-x24.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_x24(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_x24(
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 >= 24 * sizeof(float); n -= 24 * 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     const v128_t vxKLMN = wasm_v128_load(x + 20);
44     x += 24;
45 
46     const v128_t vz0123 = wasm_f32x4_abs(vx0123);
47     const v128_t vz4567 = wasm_f32x4_abs(vx4567);
48     const v128_t vz89AB = wasm_f32x4_abs(vx89AB);
49     const v128_t vzCDEF = wasm_f32x4_abs(vxCDEF);
50     const v128_t vzGHIJ = wasm_f32x4_abs(vxGHIJ);
51     const v128_t vzKLMN = wasm_f32x4_abs(vxKLMN);
52 
53     v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz0123, vminus_log2e));
54     v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz4567, vminus_log2e));
55     v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz89AB, vminus_log2e));
56     v128_t vnCDEF = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzCDEF, vminus_log2e));
57     v128_t vnGHIJ = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzGHIJ, vminus_log2e));
58     v128_t vnKLMN = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzKLMN, vminus_log2e));
59 
60     const v128_t ve0123 = wasm_i32x4_shl(vn0123, 17);
61     const v128_t ve4567 = wasm_i32x4_shl(vn4567, 17);
62     const v128_t ve89AB = wasm_i32x4_shl(vn89AB, 17);
63     const v128_t veCDEF = wasm_i32x4_shl(vnCDEF, 17);
64     const v128_t veGHIJ = wasm_i32x4_shl(vnGHIJ, 17);
65     const v128_t veKLMN = wasm_i32x4_shl(vnKLMN, 17);
66 
67     const v128_t vidx0123 = wasm_i32x4_shl(wasm_v128_and(vn0123, vindex_mask), 2);
68     const v128_t vidx4567 = wasm_i32x4_shl(wasm_v128_and(vn4567, vindex_mask), 2);
69     const v128_t vidx89AB = wasm_i32x4_shl(wasm_v128_and(vn89AB, vindex_mask), 2);
70     const v128_t vidxCDEF = wasm_i32x4_shl(wasm_v128_and(vnCDEF, vindex_mask), 2);
71     const v128_t vidxGHIJ = wasm_i32x4_shl(wasm_v128_and(vnGHIJ, vindex_mask), 2);
72     const v128_t vidxKLMN = wasm_i32x4_shl(wasm_v128_and(vnKLMN, vindex_mask), 2);
73 
74     const uint64_t vidx01 = wasm_i64x2_extract_lane(vidx0123, 0);
75     const uint64_t vidx23 = wasm_i64x2_extract_lane(vidx0123, 1);
76     const float vl0   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx01));
77     const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32)));
78     const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx23));
79     const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32)));
80     const v128_t vl0123 = wasm_f32x4_make(vl0, vl1, vl2, vl3);
81     const uint64_t vidx45 = wasm_i64x2_extract_lane(vidx4567, 0);
82     const uint64_t vidx67 = wasm_i64x2_extract_lane(vidx4567, 1);
83     const float vl4   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45));
84     const float vl5 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32)));
85     const float vl6 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67));
86     const float vl7 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32)));
87     const v128_t vl4567 = wasm_f32x4_make(vl4, vl5, vl6, vl7);
88     const uint64_t vidx89 = wasm_i64x2_extract_lane(vidx89AB, 0);
89     const uint64_t vidxAB = wasm_i64x2_extract_lane(vidx89AB, 1);
90     const float vl8   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx89));
91     const float vl9 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx89 >> 32)));
92     const float vlA = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxAB));
93     const float vlB = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxAB >> 32)));
94     const v128_t vl89AB = wasm_f32x4_make(vl8, vl9, vlA, vlB);
95     const uint64_t vidxCD = wasm_i64x2_extract_lane(vidxCDEF, 0);
96     const uint64_t vidxEF = wasm_i64x2_extract_lane(vidxCDEF, 1);
97     const float vlC   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxCD));
98     const float vlD = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxCD >> 32)));
99     const float vlE = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxEF));
100     const float vlF = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxEF >> 32)));
101     const v128_t vlCDEF = wasm_f32x4_make(vlC, vlD, vlE, vlF);
102     const uint64_t vidxGH = wasm_i64x2_extract_lane(vidxGHIJ, 0);
103     const uint64_t vidxIJ = wasm_i64x2_extract_lane(vidxGHIJ, 1);
104     const float vlG   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxGH));
105     const float vlH = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxGH >> 32)));
106     const float vlI = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxIJ));
107     const float vlJ = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxIJ >> 32)));
108     const v128_t vlGHIJ = wasm_f32x4_make(vlG, vlH, vlI, vlJ);
109     const uint64_t vidxKL = wasm_i64x2_extract_lane(vidxKLMN, 0);
110     const uint64_t vidxMN = wasm_i64x2_extract_lane(vidxKLMN, 1);
111     const float vlK   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxKL));
112     const float vlL = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxKL >> 32)));
113     const float vlM = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxMN));
114     const float vlN = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxMN >> 32)));
115     const v128_t vlKLMN = wasm_f32x4_make(vlK, vlL, vlM, vlN);
116 
117     const v128_t vs0123 = wasm_i32x4_add(vl0123, ve0123);
118     const v128_t vs4567 = wasm_i32x4_add(vl4567, ve4567);
119     const v128_t vs89AB = wasm_i32x4_add(vl89AB, ve89AB);
120     const v128_t vsCDEF = wasm_i32x4_add(vlCDEF, veCDEF);
121     const v128_t vsGHIJ = wasm_i32x4_add(vlGHIJ, veGHIJ);
122     const v128_t vsKLMN = wasm_i32x4_add(vlKLMN, veKLMN);
123 
124     vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
125     vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
126     vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
127     vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
128     vnGHIJ = wasm_f32x4_sub(vnGHIJ, vmagic_bias);
129     vnKLMN = wasm_f32x4_sub(vnKLMN, vmagic_bias);
130 
131     v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi));
132     v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi));
133     v128_t vt89AB = wasm_f32x4_add(vz89AB, wasm_f32x4_mul(vn89AB, vln2_hi));
134     v128_t vtCDEF = wasm_f32x4_add(vzCDEF, wasm_f32x4_mul(vnCDEF, vln2_hi));
135     v128_t vtGHIJ = wasm_f32x4_add(vzGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_hi));
136     v128_t vtKLMN = wasm_f32x4_add(vzKLMN, wasm_f32x4_mul(vnKLMN, vln2_hi));
137 
138     vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo));
139     vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo));
140     vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vln2_lo));
141     vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vln2_lo));
142     vtGHIJ = wasm_f32x4_add(vtGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_lo));
143     vtKLMN = wasm_f32x4_add(vtKLMN, wasm_f32x4_mul(vnKLMN, vln2_lo));
144 
145     v128_t vp0123 = wasm_f32x4_mul(vt0123, vc2);
146     v128_t vp4567 = wasm_f32x4_mul(vt4567, vc2);
147     v128_t vp89AB = wasm_f32x4_mul(vt89AB, vc2);
148     v128_t vpCDEF = wasm_f32x4_mul(vtCDEF, vc2);
149     v128_t vpGHIJ = wasm_f32x4_mul(vtGHIJ, vc2);
150     v128_t vpKLMN = wasm_f32x4_mul(vtKLMN, vc2);
151 
152     vp0123 = wasm_f32x4_sub(vt0123, wasm_f32x4_mul(vp0123, vt0123));
153     vp4567 = wasm_f32x4_sub(vt4567, wasm_f32x4_mul(vp4567, vt4567));
154     vp89AB = wasm_f32x4_sub(vt89AB, wasm_f32x4_mul(vp89AB, vt89AB));
155     vpCDEF = wasm_f32x4_sub(vtCDEF, wasm_f32x4_mul(vpCDEF, vtCDEF));
156     vpGHIJ = wasm_f32x4_sub(vtGHIJ, wasm_f32x4_mul(vpGHIJ, vtGHIJ));
157     vpKLMN = wasm_f32x4_sub(vtKLMN, wasm_f32x4_mul(vpKLMN, vtKLMN));
158 
159     const v128_t vy0123 = wasm_f32x4_sub(vs0123, wasm_f32x4_mul(vs0123, vp0123));
160     const v128_t vy4567 = wasm_f32x4_sub(vs4567, wasm_f32x4_mul(vs4567, vp4567));
161     const v128_t vy89AB = wasm_f32x4_sub(vs89AB, wasm_f32x4_mul(vs89AB, vp89AB));
162     const v128_t vyCDEF = wasm_f32x4_sub(vsCDEF, wasm_f32x4_mul(vsCDEF, vpCDEF));
163     const v128_t vyGHIJ = wasm_f32x4_sub(vsGHIJ, wasm_f32x4_mul(vsGHIJ, vpGHIJ));
164     const v128_t vyKLMN = wasm_f32x4_sub(vsKLMN, wasm_f32x4_mul(vsKLMN, vpKLMN));
165 
166     const v128_t vd0123 = wasm_f32x4_add(vy0123, vone);
167     const v128_t vd4567 = wasm_f32x4_add(vy4567, vone);
168     const v128_t vd89AB = wasm_f32x4_add(vy89AB, vone);
169     const v128_t vdCDEF = wasm_f32x4_add(vyCDEF, vone);
170     const v128_t vdGHIJ = wasm_f32x4_add(vyGHIJ, vone);
171     const v128_t vdKLMN = wasm_f32x4_add(vyKLMN, vone);
172 
173     v128_t vf0123 = wasm_f32x4_div(vy0123, vd0123);
174     v128_t vf4567 = wasm_f32x4_div(vy4567, vd4567);
175     v128_t vf89AB = wasm_f32x4_div(vy89AB, vd89AB);
176     v128_t vfCDEF = wasm_f32x4_div(vyCDEF, vdCDEF);
177     v128_t vfGHIJ = wasm_f32x4_div(vyGHIJ, vdGHIJ);
178     v128_t vfKLMN = wasm_f32x4_div(vyKLMN, vdKLMN);
179 
180     vf0123 = wasm_v128_andnot(vf0123, wasm_f32x4_gt(vz0123, vdenorm_cutoff));
181     vf4567 = wasm_v128_andnot(vf4567, wasm_f32x4_gt(vz4567, vdenorm_cutoff));
182     vf89AB = wasm_v128_andnot(vf89AB, wasm_f32x4_gt(vz89AB, vdenorm_cutoff));
183     vfCDEF = wasm_v128_andnot(vfCDEF, wasm_f32x4_gt(vzCDEF, vdenorm_cutoff));
184     vfGHIJ = wasm_v128_andnot(vfGHIJ, wasm_f32x4_gt(vzGHIJ, vdenorm_cutoff));
185     vfKLMN = wasm_v128_andnot(vfKLMN, wasm_f32x4_gt(vzKLMN, vdenorm_cutoff));
186 
187     vf0123 = wasm_v128_bitselect(vf0123, wasm_f32x4_sub(vone, vf0123), wasm_i32x4_shr(vx0123, 31));
188     vf4567 = wasm_v128_bitselect(vf4567, wasm_f32x4_sub(vone, vf4567), wasm_i32x4_shr(vx4567, 31));
189     vf89AB = wasm_v128_bitselect(vf89AB, wasm_f32x4_sub(vone, vf89AB), wasm_i32x4_shr(vx89AB, 31));
190     vfCDEF = wasm_v128_bitselect(vfCDEF, wasm_f32x4_sub(vone, vfCDEF), wasm_i32x4_shr(vxCDEF, 31));
191     vfGHIJ = wasm_v128_bitselect(vfGHIJ, wasm_f32x4_sub(vone, vfGHIJ), wasm_i32x4_shr(vxGHIJ, 31));
192     vfKLMN = wasm_v128_bitselect(vfKLMN, wasm_f32x4_sub(vone, vfKLMN), wasm_i32x4_shr(vxKLMN, 31));
193 
194     wasm_v128_store(y, vf0123);
195     wasm_v128_store(y + 4, vf4567);
196     wasm_v128_store(y + 8, vf89AB);
197     wasm_v128_store(y + 12, vfCDEF);
198     wasm_v128_store(y + 16, vfGHIJ);
199     wasm_v128_store(y + 20, vfKLMN);
200     y += 24;
201   }
202   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
203     const v128_t vx = wasm_v128_load(x);
204     x += 4;
205 
206     const v128_t vz = wasm_f32x4_abs(vx);
207 
208     v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e));
209     const v128_t ve = wasm_i32x4_shl(vn, 17);
210 
211     const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
212     const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
213     const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
214     const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo));
215     const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32)));
216     const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi));
217     const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32)));
218     const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);
219 
220     const v128_t vs = wasm_i32x4_add(vl, ve);
221     vn = wasm_f32x4_sub(vn, vmagic_bias);
222 
223     v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi));
224     vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo));
225 
226     v128_t vp = wasm_f32x4_mul(vt, vc2);
227     vp = wasm_f32x4_sub(vt, wasm_f32x4_mul(vp, vt));
228 
229     const v128_t vy = wasm_f32x4_sub(vs, wasm_f32x4_mul(vs, vp));
230     const v128_t vd = wasm_f32x4_add(vy, vone);
231 
232     v128_t vf = wasm_f32x4_div(vy, vd);
233     vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
234     vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));
235 
236     wasm_v128_store(y, vf);
237     y += 4;
238   }
239   if XNN_UNLIKELY(n != 0) {
240     const v128_t vx = wasm_v128_load(x);
241 
242     const v128_t vz = wasm_f32x4_abs(vx);
243 
244     v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e));
245     const v128_t ve = wasm_i32x4_shl(vn, 17);
246 
247     const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
248     const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
249     const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
250     const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo));
251     const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32)));
252     const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi));
253     const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32)));
254     const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);
255 
256     const v128_t vs = wasm_i32x4_add(vl, ve);
257     vn = wasm_f32x4_sub(vn, vmagic_bias);
258 
259     v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi));
260     vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo));
261 
262     v128_t vp = wasm_f32x4_mul(vt, vc2);
263     vp = wasm_f32x4_sub(vt, wasm_f32x4_mul(vp, vt));
264 
265     const v128_t vy = wasm_f32x4_sub(vs, wasm_f32x4_mul(vs, vp));
266     const v128_t vd = wasm_f32x4_add(vy, vone);
267 
268     v128_t vf = wasm_f32x4_div(vy, vd);
269     vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
270     vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));
271 
272     if (n & (2 * sizeof(float))) {
273       *((double*) y) = wasm_f64x2_extract_lane(vf, 0);
274       vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3);
275       y += 2;
276     }
277     if (n & (1 * sizeof(float))) {
278       *y = wasm_f32x4_extract_lane(vf, 0);
279     }
280   }
281 }
282