1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-sigmoid/wasmsimd-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_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12(size_t n,const float * x,float * y,const void * params)20 void xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12(
21     size_t n,
22     const float* x,
23     float* y,
24     const void* params) XNN_DISABLE_TSAN
25 {
26   assert(n % sizeof(float) == 0);
27 
28   const v128_t vmagic_bias = wasm_f32x4_splat(0x1.800000p17f);
29   const v128_t vminus_log2e = wasm_f32x4_splat(-0x1.715476p0f);
30   const v128_t vindex_mask = wasm_i32x4_splat(INT32_C(0x3F));
31   const v128_t vln2_hi = wasm_f32x4_splat(0x1.630000p-1f);
32   const v128_t vln2_lo = wasm_f32x4_splat(-0x1.BD0106p-13f);
33   const v128_t vc2 = wasm_f32x4_splat(0x1.FFFF0Ap-2f);
34   const v128_t vone = wasm_f32x4_splat(1.0f);
35   const v128_t vdenorm_cutoff = wasm_f32x4_splat(0x1.5D589Ep+6f);
36 
37   for (; n >= 12 * sizeof(float); n -= 12 * 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     x += 12;
42 
43     const v128_t vz0123 = wasm_f32x4_abs(vx0123);
44     const v128_t vz4567 = wasm_f32x4_abs(vx4567);
45     const v128_t vz89AB = wasm_f32x4_abs(vx89AB);
46 
47     v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz0123, vminus_log2e));
48     v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz4567, vminus_log2e));
49     v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz89AB, vminus_log2e));
50 
51     const v128_t ve0123 = wasm_i32x4_shl(vn0123, 17);
52     const v128_t ve4567 = wasm_i32x4_shl(vn4567, 17);
53     const v128_t ve89AB = wasm_i32x4_shl(vn89AB, 17);
54 
55     const v128_t vidx0123 = wasm_i32x4_shl(wasm_v128_and(vn0123, vindex_mask), 2);
56     const v128_t vidx4567 = wasm_i32x4_shl(wasm_v128_and(vn4567, vindex_mask), 2);
57     const v128_t vidx89AB = wasm_i32x4_shl(wasm_v128_and(vn89AB, vindex_mask), 2);
58 
59     const uint64_t vidx01 = wasm_i64x2_extract_lane(vidx0123, 0);
60     const uint64_t vidx23 = wasm_i64x2_extract_lane(vidx0123, 1);
61     const float vl0   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx01));
62     const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32)));
63     const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx23));
64     const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32)));
65     const v128_t vl0123 = wasm_f32x4_make(vl0, vl1, vl2, vl3);
66     const uint64_t vidx45 = wasm_i64x2_extract_lane(vidx4567, 0);
67     const uint64_t vidx67 = wasm_i64x2_extract_lane(vidx4567, 1);
68     const float vl4   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45));
69     const float vl5 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32)));
70     const float vl6 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67));
71     const float vl7 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32)));
72     const v128_t vl4567 = wasm_f32x4_make(vl4, vl5, vl6, vl7);
73     const uint64_t vidx89 = wasm_i64x2_extract_lane(vidx89AB, 0);
74     const uint64_t vidxAB = wasm_i64x2_extract_lane(vidx89AB, 1);
75     const float vl8   = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx89));
76     const float vl9 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx89 >> 32)));
77     const float vlA = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxAB));
78     const float vlB = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxAB >> 32)));
79     const v128_t vl89AB = wasm_f32x4_make(vl8, vl9, vlA, vlB);
80 
81     const v128_t vs0123 = wasm_i32x4_add(vl0123, ve0123);
82     const v128_t vs4567 = wasm_i32x4_add(vl4567, ve4567);
83     const v128_t vs89AB = wasm_i32x4_add(vl89AB, ve89AB);
84 
85     vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
86     vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
87     vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
88 
89     v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi));
90     v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi));
91     v128_t vt89AB = wasm_f32x4_add(vz89AB, wasm_f32x4_mul(vn89AB, vln2_hi));
92 
93     vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo));
94     vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo));
95     vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vln2_lo));
96 
97     v128_t vp0123 = wasm_f32x4_mul(vt0123, vc2);
98     v128_t vp4567 = wasm_f32x4_mul(vt4567, vc2);
99     v128_t vp89AB = wasm_f32x4_mul(vt89AB, vc2);
100 
101     vp0123 = wasm_f32x4_sub(vt0123, wasm_f32x4_mul(vp0123, vt0123));
102     vp4567 = wasm_f32x4_sub(vt4567, wasm_f32x4_mul(vp4567, vt4567));
103     vp89AB = wasm_f32x4_sub(vt89AB, wasm_f32x4_mul(vp89AB, vt89AB));
104 
105     const v128_t vy0123 = wasm_f32x4_sub(vs0123, wasm_f32x4_mul(vs0123, vp0123));
106     const v128_t vy4567 = wasm_f32x4_sub(vs4567, wasm_f32x4_mul(vs4567, vp4567));
107     const v128_t vy89AB = wasm_f32x4_sub(vs89AB, wasm_f32x4_mul(vs89AB, vp89AB));
108 
109     const v128_t vd0123 = wasm_f32x4_add(vy0123, vone);
110     const v128_t vd4567 = wasm_f32x4_add(vy4567, vone);
111     const v128_t vd89AB = wasm_f32x4_add(vy89AB, vone);
112 
113     v128_t vf0123 = wasm_f32x4_div(vy0123, vd0123);
114     v128_t vf4567 = wasm_f32x4_div(vy4567, vd4567);
115     v128_t vf89AB = wasm_f32x4_div(vy89AB, vd89AB);
116 
117     vf0123 = wasm_v128_andnot(vf0123, wasm_f32x4_gt(vz0123, vdenorm_cutoff));
118     vf4567 = wasm_v128_andnot(vf4567, wasm_f32x4_gt(vz4567, vdenorm_cutoff));
119     vf89AB = wasm_v128_andnot(vf89AB, wasm_f32x4_gt(vz89AB, vdenorm_cutoff));
120 
121     vf0123 = wasm_v128_bitselect(vf0123, wasm_f32x4_sub(vone, vf0123), wasm_i32x4_shr(vx0123, 31));
122     vf4567 = wasm_v128_bitselect(vf4567, wasm_f32x4_sub(vone, vf4567), wasm_i32x4_shr(vx4567, 31));
123     vf89AB = wasm_v128_bitselect(vf89AB, wasm_f32x4_sub(vone, vf89AB), wasm_i32x4_shr(vx89AB, 31));
124 
125     wasm_v128_store(y, vf0123);
126     wasm_v128_store(y + 4, vf4567);
127     wasm_v128_store(y + 8, vf89AB);
128     y += 12;
129   }
130   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
131     const v128_t vx = wasm_v128_load(x);
132     x += 4;
133 
134     const v128_t vz = wasm_f32x4_abs(vx);
135 
136     v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e));
137     const v128_t ve = wasm_i32x4_shl(vn, 17);
138 
139     const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
140     const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
141     const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
142     const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo));
143     const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32)));
144     const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi));
145     const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32)));
146     const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);
147 
148     const v128_t vs = wasm_i32x4_add(vl, ve);
149     vn = wasm_f32x4_sub(vn, vmagic_bias);
150 
151     v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi));
152     vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo));
153 
154     v128_t vp = wasm_f32x4_mul(vt, vc2);
155     vp = wasm_f32x4_sub(vt, wasm_f32x4_mul(vp, vt));
156 
157     const v128_t vy = wasm_f32x4_sub(vs, wasm_f32x4_mul(vs, vp));
158     const v128_t vd = wasm_f32x4_add(vy, vone);
159 
160     v128_t vf = wasm_f32x4_div(vy, vd);
161     vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
162     vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));
163 
164     wasm_v128_store(y, vf);
165     y += 4;
166   }
167   if XNN_UNLIKELY(n != 0) {
168     const v128_t vx = wasm_v128_load(x);
169 
170     const v128_t vz = wasm_f32x4_abs(vx);
171 
172     v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e));
173     const v128_t ve = wasm_i32x4_shl(vn, 17);
174 
175     const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
176     const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
177     const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
178     const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo));
179     const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32)));
180     const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi));
181     const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32)));
182     const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);
183 
184     const v128_t vs = wasm_i32x4_add(vl, ve);
185     vn = wasm_f32x4_sub(vn, vmagic_bias);
186 
187     v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi));
188     vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo));
189 
190     v128_t vp = wasm_f32x4_mul(vt, vc2);
191     vp = wasm_f32x4_sub(vt, wasm_f32x4_mul(vp, vt));
192 
193     const v128_t vy = wasm_f32x4_sub(vs, wasm_f32x4_mul(vs, vp));
194     const v128_t vd = wasm_f32x4_add(vy, vone);
195 
196     v128_t vf = wasm_f32x4_div(vy, vd);
197     vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
198     vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));
199 
200     if (n & (2 * sizeof(float))) {
201       *((double*) y) = wasm_f64x2_extract_lane(vf, 0);
202       vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3);
203       y += 2;
204     }
205     if (n & (1 * sizeof(float))) {
206       *y = wasm_f32x4_extract_lane(vf, 0);
207     }
208   }
209 }
210