1 // Auto-generated file. Do not edit!
2 // Template: src/f32-raddstoreexpminusmax/wasmsimd-rr2-p5.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/raddstoreexpminusmax.h>
16
17
xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x16(size_t elements,const float * input,const float * max,float * output,float * sum,const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x16(
19 size_t elements,
20 const float* input,
21 const float* max,
22 float* output,
23 float* sum,
24 const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(elements % sizeof(float) == 0);
27
28 const v128_t vi_max = wasm_v128_load32_splat(max);
29 const v128_t vlog2e = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.log2e);
30 const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.magic_bias);
31 const v128_t vminus_ln2_hi = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.minus_ln2_hi);
32 const v128_t vminus_ln2_lo = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.minus_ln2_lo);
33 const v128_t vc5 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c5);
34 const v128_t vc4 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c4);
35 const v128_t vc3 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c3);
36 const v128_t vc2 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c2);
37 const v128_t vc1 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c1);
38 const v128_t vdenorm_cutoff = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.denorm_cutoff);
39
40 v128_t vacc0 = wasm_f32x4_const_splat(0.0f);
41 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
42 // Load 16 (4x4) inputs at a time.
43 const v128_t vi0123 = wasm_v128_load(input);
44 const v128_t vi4567 = wasm_v128_load(input + 4);
45 const v128_t vi89AB = wasm_v128_load(input + 8);
46 const v128_t viCDEF = wasm_v128_load(input + 12);
47 input += 16;
48
49 // Subtract maximum input x := i - i_max. This implies x <= 0.
50 const v128_t vx0123 = wasm_f32x4_sub(vi0123, vi_max);
51 const v128_t vx4567 = wasm_f32x4_sub(vi4567, vi_max);
52 const v128_t vx89AB = wasm_f32x4_sub(vi89AB, vi_max);
53 const v128_t vxCDEF = wasm_f32x4_sub(viCDEF, vi_max);
54
55 // Compute reduced argument elements := round(x / log(2)).
56 v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx0123, vlog2e));
57 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx4567, vlog2e));
58 v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx89AB, vlog2e));
59 v128_t vnCDEF = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vxCDEF, vlog2e));
60
61 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
62 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
63 const v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
64 const v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
65 const v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23);
66 const v128_t vsCDEF = wasm_i32x4_shl(vnCDEF, 23);
67
68 // Subtract the large number back to get final elements := round(x / log(2)).
69 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
70 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
71 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
72 vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
73
74 // Compute reduced argument t := x - elements * log(2).
75 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
76 v128_t vt0123 = wasm_f32x4_add(vx0123, wasm_f32x4_mul(vn0123, vminus_ln2_hi));
77 v128_t vt4567 = wasm_f32x4_add(vx4567, wasm_f32x4_mul(vn4567, vminus_ln2_hi));
78 v128_t vt89AB = wasm_f32x4_add(vx89AB, wasm_f32x4_mul(vn89AB, vminus_ln2_hi));
79 v128_t vtCDEF = wasm_f32x4_add(vxCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_hi));
80
81 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vminus_ln2_lo));
82 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vminus_ln2_lo));
83 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vminus_ln2_lo));
84 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_lo));
85
86 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
87 v128_t vp0123 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt0123));
88 v128_t vp4567 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt4567));
89 v128_t vp89AB = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt89AB));
90 v128_t vpCDEF = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtCDEF));
91
92 vp0123 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp0123, vt0123));
93 vp4567 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp4567, vt4567));
94 vp89AB = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp89AB, vt89AB));
95 vpCDEF = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpCDEF, vtCDEF));
96
97 vp0123 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp0123, vt0123));
98 vp4567 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp4567, vt4567));
99 vp89AB = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp89AB, vt89AB));
100 vpCDEF = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpCDEF, vtCDEF));
101
102 vp0123 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp0123, vt0123));
103 vp4567 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp4567, vt4567));
104 vp89AB = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp89AB, vt89AB));
105 vpCDEF = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpCDEF, vtCDEF));
106
107 // Reconstruct the final f value:
108 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
109 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
110 // = s + (t * s) * p
111 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
112 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
113 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
114 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
115
116 v128_t vf0123 = wasm_f32x4_add(vs0123, wasm_f32x4_mul(vt0123, vp0123));
117 v128_t vf4567 = wasm_f32x4_add(vs4567, wasm_f32x4_mul(vt4567, vp4567));
118 v128_t vf89AB = wasm_f32x4_add(vs89AB, wasm_f32x4_mul(vt89AB, vp89AB));
119 v128_t vfCDEF = wasm_f32x4_add(vsCDEF, wasm_f32x4_mul(vtCDEF, vpCDEF));
120
121 // For inputs below zero cutoff, replace output with +0.0f.
122 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
123 vf0123 = wasm_v128_andnot(vf0123, wasm_f32x4_lt(vx0123, vdenorm_cutoff));
124 vf4567 = wasm_v128_andnot(vf4567, wasm_f32x4_lt(vx4567, vdenorm_cutoff));
125 vf89AB = wasm_v128_andnot(vf89AB, wasm_f32x4_lt(vx89AB, vdenorm_cutoff));
126 vfCDEF = wasm_v128_andnot(vfCDEF, wasm_f32x4_lt(vxCDEF, vdenorm_cutoff));
127
128 // Store 16 (4x4) outputs at a time.
129 wasm_v128_store(output, vf0123);
130 wasm_v128_store(output + 4, vf4567);
131 wasm_v128_store(output + 8, vf89AB);
132 wasm_v128_store(output + 12, vfCDEF);
133 output += 16;
134
135 // Accumulate computed exponents.
136 vacc0 = wasm_f32x4_add(vacc0, vf0123);
137 vacc0 = wasm_f32x4_add(vacc0, vf4567);
138 vacc0 = wasm_f32x4_add(vacc0, vf89AB);
139 vacc0 = wasm_f32x4_add(vacc0, vfCDEF);
140 }
141
142 v128_t vacc = vacc0;
143 for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
144 // Load 4 inputs at a time.
145 const v128_t vi = wasm_v128_load(input);
146 input += 4;
147
148 // Subtract maximum input x := i - i_max. This implies x <= 0.
149 const v128_t vx = wasm_f32x4_sub(vi, vi_max);
150
151 // Compute reduced argument elements := round(x / log(2)).
152 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
153
154 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
155 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
156 const v128_t vs = wasm_i32x4_shl(vn, 23);
157
158 // Subtract the large number back to get final elements := round(x / log(2)).
159 vn = wasm_f32x4_sub(vn, vmagic_bias);
160
161 // Compute reduced argument t := x - elements * log(2).
162 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
163 v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
164 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
165
166 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
167 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
168 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
169 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
170 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
171
172 // Reconstruct the final f value:
173 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
174 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
175 // = s + (t * s) * p
176 vt = wasm_f32x4_mul(vt, vs);
177 v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
178
179 // For inputs below zero cutoff, replace output with +0.0f.
180 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
181 vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
182
183 // Store 4 outputs at a time.
184 wasm_v128_store(output, vf);
185 output += 4;
186
187 // Accumulate computed exponents.
188 vacc = wasm_f32x4_add(vacc, vf);
189 }
190 vacc = wasm_f32x4_add(vacc, wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3));
191 float vsum = wasm_f32x4_extract_lane(vacc, 0) + wasm_f32x4_extract_lane(vacc, 1);
192 if (elements != 0) {
193 assert(elements >= 1 * sizeof(float));
194 assert(elements <= 3 * sizeof(float));
195 // Load 4 inputs at a time.
196 const v128_t vi = wasm_v128_load(input);
197
198 // Subtract maximum input x := i - i_max. This implies x <= 0.
199 const v128_t vx = wasm_f32x4_sub(vi, vi_max);
200
201 // Compute reduced argument elements := round(x / log(2)).
202 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
203
204 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
205 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
206 const v128_t vs = wasm_i32x4_shl(vn, 23);
207
208 // Subtract the large number back to get final elements := round(x / log(2)).
209 vn = wasm_f32x4_sub(vn, vmagic_bias);
210
211 // Compute reduced argument t := x - elements * log(2).
212 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
213 v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
214 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
215
216 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
217 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
218 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
219 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
220 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
221
222 // Reconstruct the final f value:
223 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
224 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
225 // = s + (t * s) * p
226 vt = wasm_f32x4_mul(vt, vs);
227 v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
228
229 // For inputs below zero cutoff, replace output with +0.0f.
230 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
231 vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
232
233 if (elements & (2 * sizeof(float))) {
234 // Store and accumulate 2 outputs at a time.
235 const float vf0 = wasm_f32x4_extract_lane(vf, 0);
236 output[0] = vf0;
237 vsum += vf0;
238
239 const float vf1 = wasm_f32x4_extract_lane(vf, 1);
240 output[1] = vf1;
241 vsum += vf1;
242
243 vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3);
244 output += 2;
245 }
246 if (elements & (1 * sizeof(float))) {
247 // Store 1 output at a time.
248 const float vf0 = wasm_f32x4_extract_lane(vf, 0);
249 *output = vf0;
250 vsum += vf0;
251 }
252 }
253 // Reduce 4 elements in the SIMD register
254 *sum = vsum;
255 }
256