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_x8(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_x8(
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
46 v128_t vx0123 = wasm_v128_load(x);
47 v128_t vx4567 = wasm_v128_load(x + 4);
48 x += 8;
49
50 const v128_t vz0123 = wasm_f32x4_max(wasm_f32x4_mul(vx0123, vprescale), vsat_cutoff);
51 const v128_t vz4567 = wasm_f32x4_max(wasm_f32x4_mul(vx4567, vprescale), vsat_cutoff);
52
53 v128_t vn0123 = wasm_f32x4_add(wasm_f32x4_mul(vz0123, vlog2e), vmagic_bias);
54 v128_t vn4567 = wasm_f32x4_add(wasm_f32x4_mul(vz4567, vlog2e), vmagic_bias);
55
56 const v128_t vidx0123 = wasm_i32x4_shl(wasm_v128_and(vn0123, vindex_mask), 2);
57 const v128_t ven0123 = wasm_i32x4_shl(vn0123, 19);
58 const v128_t vidx4567 = wasm_i32x4_shl(wasm_v128_and(vn4567, vindex_mask), 2);
59 const v128_t ven4567 = wasm_i32x4_shl(vn4567, 19);
60
61 const uint64_t vidx01 = wasm_i64x2_extract_lane(vidx0123, 0);
62 const uint64_t vidx23 = wasm_i64x2_extract_lane(vidx0123, 1);
63 const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx01));
64 const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01 >> 32)));
65 const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23));
66 const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23 >> 32)));
67 const v128_t vl0123 = wasm_f32x4_make(vl0, vl1, vl2, vl3);
68 const uint64_t vidx45 = wasm_i64x2_extract_lane(vidx4567, 0);
69 const uint64_t vidx67 = wasm_i64x2_extract_lane(vidx4567, 1);
70 const float vl4 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx45));
71 const float vl5 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45 >> 32)));
72 const float vl6 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67));
73 const float vl7 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67 >> 32)));
74 const v128_t vl4567 = wasm_f32x4_make(vl4, vl5, vl6, vl7);
75
76 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
77 v128_t vs0123 = wasm_i32x4_add(vl0123, ven0123);
78 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
79 v128_t vs4567 = wasm_i32x4_add(vl4567, ven4567);
80
81 v128_t vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_hi), vz0123);
82 v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567);
83
84 vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_lo), vt0123);
85 vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567);
86
87 v128_t vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt0123), vc2);
88 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt4567), vc2);
89
90 vp0123 = wasm_f32x4_mul(vp0123, vt0123);
91 vp4567 = wasm_f32x4_mul(vp4567, vt4567);
92
93 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
94 vs0123 = wasm_f32x4_sub(vs0123, vone);
95 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
96 vs4567 = wasm_f32x4_sub(vs4567, vone);
97
98 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vt0123);
99 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567);
100
101 const v128_t ve0123 = wasm_f32x4_mul(wasm_f32x4_add(vp0123, vs0123), valpha);
102 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha);
103
104 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31);
105 vx0123 = wasm_f32x4_mul(vx0123, vbeta);
106 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31);
107 vx4567 = wasm_f32x4_mul(vx4567, vbeta);
108
109 const v128_t vy0123 = wasm_v128_bitselect(ve0123, vx0123, vsignm0123);
110 const v128_t vy4567 = wasm_v128_bitselect(ve4567, vx4567, vsignm4567);
111
112 wasm_v128_store(y, vy0123);
113 wasm_v128_store(y + 4, vy4567);
114 y += 8;
115 }
116 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
117 v128_t vx = wasm_v128_load(x);
118 x += 4;
119
120 const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);
121
122 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
123 const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
124 const v128_t ven = wasm_i32x4_shl(vn, 19);
125
126 const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
127 const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
128 const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
129 const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)));
130 const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
131 const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)));
132 const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);
133
134 v128_t vs = wasm_i32x4_add(vl, ven);
135 vn = wasm_f32x4_sub(vn, vmagic_bias);
136
137 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
138 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
139
140 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt), vc2);
141 vp = wasm_f32x4_mul(vp, vt);
142
143 vt = wasm_f32x4_mul(vt, vs);
144 vs = wasm_f32x4_sub(vs, vone);
145 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
146 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
147
148 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
149 vx = wasm_f32x4_mul(vx, vbeta);
150 const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
151
152 wasm_v128_store(y, vy);
153 y += 4;
154 }
155 if XNN_UNLIKELY(n != 0) {
156 v128_t vx = wasm_v128_load(x);
157
158 const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);
159
160 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
161 const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2);
162 const v128_t ven = wasm_i32x4_shl(vn, 19);
163
164 const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0);
165 const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1);
166 const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo));
167 const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)));
168 const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi));
169 const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)));
170 const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3);
171
172 v128_t vs = wasm_i32x4_add(vl, ven);
173 vn = wasm_f32x4_sub(vn, vmagic_bias);
174
175 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
176 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
177
178 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt), vc2);
179 vp = wasm_f32x4_mul(vp, vt);
180
181 vt = wasm_f32x4_mul(vt, vs);
182 vs = wasm_f32x4_sub(vs, vone);
183 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
184 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
185
186 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
187 vx = wasm_f32x4_mul(vx, vbeta);
188 v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
189
190 if (n & (2 * sizeof(float))) {
191 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
192 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
193 y += 2;
194 }
195 if (n & (1 * sizeof(float))) {
196 *y = wasm_f32x4_extract_lane(vy, 0);
197 }
198 }
199 }
200