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_arm_rr2_lut16_p3_x12(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_arm_rr2_lut16_p3_x12(
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 >= 12 * sizeof(float); n -= 12 * 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 x += 12;
50
51 const v128_t vz0123 = wasm_f32x4_max(wasm_f32x4_mul(vx0123, vprescale), vsat_cutoff);
52 const v128_t vz4567 = wasm_f32x4_max(wasm_f32x4_mul(vx4567, vprescale), vsat_cutoff);
53 const v128_t vz89AB = wasm_f32x4_max(wasm_f32x4_mul(vx89AB, vprescale), vsat_cutoff);
54
55 v128_t vn0123 = wasm_f32x4_add(wasm_f32x4_mul(vz0123, vlog2e), vmagic_bias);
56 v128_t vn4567 = wasm_f32x4_add(wasm_f32x4_mul(vz4567, vlog2e), vmagic_bias);
57 v128_t vn89AB = wasm_f32x4_add(wasm_f32x4_mul(vz89AB, vlog2e), vmagic_bias);
58
59 const v128_t vidx0123 = wasm_i32x4_shl(wasm_v128_and(vn0123, vindex_mask), 2);
60 const v128_t ven0123 = wasm_i32x4_shl(vn0123, 19);
61 const v128_t vidx4567 = wasm_i32x4_shl(wasm_v128_and(vn4567, vindex_mask), 2);
62 const v128_t ven4567 = wasm_i32x4_shl(vn4567, 19);
63 const v128_t vidx89AB = wasm_i32x4_shl(wasm_v128_and(vn89AB, vindex_mask), 2);
64 const v128_t ven89AB = wasm_i32x4_shl(vn89AB, 19);
65
66 const uint64_t vidx01 = wasm_i64x2_extract_lane(vidx0123, 0);
67 const uint64_t vidx23 = wasm_i64x2_extract_lane(vidx0123, 1);
68 const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx01));
69 const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01 >> 32)));
70 const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23));
71 const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23 >> 32)));
72 const v128_t vl0123 = wasm_f32x4_make(vl0, vl1, vl2, vl3);
73 const uint64_t vidx45 = wasm_i64x2_extract_lane(vidx4567, 0);
74 const uint64_t vidx67 = wasm_i64x2_extract_lane(vidx4567, 1);
75 const float vl4 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx45));
76 const float vl5 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45 >> 32)));
77 const float vl6 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67));
78 const float vl7 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67 >> 32)));
79 const v128_t vl4567 = wasm_f32x4_make(vl4, vl5, vl6, vl7);
80 const uint64_t vidx89 = wasm_i64x2_extract_lane(vidx89AB, 0);
81 const uint64_t vidxAB = wasm_i64x2_extract_lane(vidx89AB, 1);
82 const float vl8 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx89));
83 const float vl9 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx89 >> 32)));
84 const float vlA = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxAB));
85 const float vlB = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxAB >> 32)));
86 const v128_t vl89AB = wasm_f32x4_make(vl8, vl9, vlA, vlB);
87
88 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
89 v128_t vs0123 = wasm_i32x4_add(vl0123, ven0123);
90 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
91 v128_t vs4567 = wasm_i32x4_add(vl4567, ven4567);
92 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
93 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB);
94
95 v128_t vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_hi), vz0123);
96 v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567);
97 v128_t vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_hi), vz89AB);
98
99 vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_lo), vt0123);
100 vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567);
101 vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_lo), vt89AB);
102
103 v128_t vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt0123), vc2);
104 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt4567), vc2);
105 v128_t vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt89AB), vc2);
106
107 vp0123 = wasm_f32x4_mul(vp0123, vt0123);
108 vp4567 = wasm_f32x4_mul(vp4567, vt4567);
109 vp89AB = wasm_f32x4_mul(vp89AB, vt89AB);
110
111 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
112 vs0123 = wasm_f32x4_sub(vs0123, vone);
113 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
114 vs4567 = wasm_f32x4_sub(vs4567, vone);
115 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
116 vs89AB = wasm_f32x4_sub(vs89AB, vone);
117
118 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vt0123);
119 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567);
120 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vt89AB);
121
122 const v128_t ve0123 = wasm_f32x4_mul(wasm_f32x4_add(vp0123, vs0123), valpha);
123 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha);
124 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha);
125
126 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31);
127 vx0123 = wasm_f32x4_mul(vx0123, vbeta);
128 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31);
129 vx4567 = wasm_f32x4_mul(vx4567, vbeta);
130 const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31);
131 vx89AB = wasm_f32x4_mul(vx89AB, vbeta);
132
133 const v128_t vy0123 = wasm_v128_bitselect(ve0123, vx0123, vsignm0123);
134 const v128_t vy4567 = wasm_v128_bitselect(ve4567, vx4567, vsignm4567);
135 const v128_t vy89AB = wasm_v128_bitselect(ve89AB, vx89AB, vsignm89AB);
136
137 wasm_v128_store(y, vy0123);
138 wasm_v128_store(y + 4, vy4567);
139 wasm_v128_store(y + 8, vy89AB);
140 y += 12;
141 }
142 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
143 v128_t vx = wasm_v128_load(x);
144 x += 4;
145
146 const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);
147
148 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
149 const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
150 const v128_t ven = wasm_i32x4_shl(vn, 19);
151
152 const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
153 const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
154 const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
155 const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)));
156 const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
157 const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)));
158 const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);
159
160 v128_t vs = wasm_i32x4_add(vl, ven);
161 vn = wasm_f32x4_sub(vn, vmagic_bias);
162
163 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
164 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
165
166 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt), vc2);
167 vp = wasm_f32x4_mul(vp, vt);
168
169 vt = wasm_f32x4_mul(vt, vs);
170 vs = wasm_f32x4_sub(vs, vone);
171 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
172 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
173
174 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
175 vx = wasm_f32x4_mul(vx, vbeta);
176 const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
177
178 wasm_v128_store(y, vy);
179 y += 4;
180 }
181 if XNN_UNLIKELY(n != 0) {
182 v128_t vx = wasm_v128_load(x);
183
184 const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);
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 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
203
204 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt), vc2);
205 vp = wasm_f32x4_mul(vp, vt);
206
207 vt = wasm_f32x4_mul(vt, vs);
208 vs = wasm_f32x4_sub(vs, vone);
209 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
210 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
211
212 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
213 vx = wasm_f32x4_mul(vx, vbeta);
214 v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
215
216 if (n & (2 * sizeof(float))) {
217 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
218 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
219 y += 2;
220 }
221 if (n & (1 * sizeof(float))) {
222 *y = wasm_f32x4_extract_lane(vy, 0);
223 }
224 }
225 }
226