1 // Auto-generated file. Do not edit!
2 // Template: src/f32-raddstoreexpminusmax/wasmsimd-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_p5_x20_acc2(size_t elements,const float * input,float * output,float * sum,float max)18 void xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20_acc2(
19 size_t elements,
20 const float* input,
21 float* output,
22 float* sum,
23 float max) XNN_DISABLE_TSAN
24 {
25 assert(elements % sizeof(float) == 0);
26
27 const v128_t vmagic_bias = wasm_f32x4_splat(0x1.8000FEp23f);
28 // The smallest x for which expf(x) is normalized.
29 const v128_t vdenorm_cutoff = wasm_f32x4_splat(-0x1.5D589Ep6f);
30 const v128_t vlog2e = wasm_f32x4_splat(0x1.715476p+0f);
31 // Last 7 bits are zeroes
32 const v128_t vminus_ln2_hi = wasm_f32x4_splat(-0x1.62E400p-1f);
33 const v128_t vminus_ln2_lo = wasm_f32x4_splat(-0x1.7F7D1Cp-20f);
34
35 const v128_t vc1 = wasm_f32x4_splat(0x1.FFFFF6p-1f);
36 const v128_t vc2 = wasm_f32x4_splat(0x1.FFFDC6p-2f);
37 const v128_t vc3 = wasm_f32x4_splat(0x1.555A80p-3f);
38 const v128_t vc4 = wasm_f32x4_splat(0x1.573A1Ap-5f);
39 const v128_t vc5 = wasm_f32x4_splat(0x1.0F9F9Cp-7f);
40
41 const v128_t vi_max = wasm_f32x4_splat(max);
42
43 v128_t vacc0 = wasm_f32x4_splat(0.0f);
44 v128_t vacc1 = vacc0;
45 for (; elements >= 20 * sizeof(float); elements -= 20 * sizeof(float)) {
46 // Load 20 (5x4) inputs at a time.
47 const v128_t vi0123 = wasm_v128_load(input);
48 const v128_t vi4567 = wasm_v128_load(input + 4);
49 const v128_t vi89AB = wasm_v128_load(input + 8);
50 const v128_t viCDEF = wasm_v128_load(input + 12);
51 const v128_t viGHIJ = wasm_v128_load(input + 16);
52 input += 20;
53
54 // Subtract maximum input x := i - i_max. This implies x <= 0.
55 const v128_t vx0123 = wasm_f32x4_sub(vi0123, vi_max);
56 const v128_t vx4567 = wasm_f32x4_sub(vi4567, vi_max);
57 const v128_t vx89AB = wasm_f32x4_sub(vi89AB, vi_max);
58 const v128_t vxCDEF = wasm_f32x4_sub(viCDEF, vi_max);
59 const v128_t vxGHIJ = wasm_f32x4_sub(viGHIJ, vi_max);
60
61 // Compute reduced argument elements := round(x / log(2)).
62 v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx0123, vlog2e));
63 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx4567, vlog2e));
64 v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx89AB, vlog2e));
65 v128_t vnCDEF = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vxCDEF, vlog2e));
66 v128_t vnGHIJ = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vxGHIJ, vlog2e));
67
68 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
69 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
70 const v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
71 const v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
72 const v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23);
73 const v128_t vsCDEF = wasm_i32x4_shl(vnCDEF, 23);
74 const v128_t vsGHIJ = wasm_i32x4_shl(vnGHIJ, 23);
75
76 // Subtract the large number back to get final elements := round(x / log(2)).
77 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
78 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
79 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
80 vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
81 vnGHIJ = wasm_f32x4_sub(vnGHIJ, vmagic_bias);
82
83 // Compute reduced argument t := x - elements * log(2).
84 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
85 v128_t vt0123 = wasm_f32x4_add(vx0123, wasm_f32x4_mul(vn0123, vminus_ln2_hi));
86 v128_t vt4567 = wasm_f32x4_add(vx4567, wasm_f32x4_mul(vn4567, vminus_ln2_hi));
87 v128_t vt89AB = wasm_f32x4_add(vx89AB, wasm_f32x4_mul(vn89AB, vminus_ln2_hi));
88 v128_t vtCDEF = wasm_f32x4_add(vxCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_hi));
89 v128_t vtGHIJ = wasm_f32x4_add(vxGHIJ, wasm_f32x4_mul(vnGHIJ, vminus_ln2_hi));
90
91 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vminus_ln2_lo));
92 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vminus_ln2_lo));
93 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vminus_ln2_lo));
94 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_lo));
95 vtGHIJ = wasm_f32x4_add(vtGHIJ, wasm_f32x4_mul(vnGHIJ, vminus_ln2_lo));
96
97 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
98 v128_t vp0123 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt0123));
99 v128_t vp4567 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt4567));
100 v128_t vp89AB = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt89AB));
101 v128_t vpCDEF = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtCDEF));
102 v128_t vpGHIJ = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtGHIJ));
103
104 vp0123 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp0123, vt0123));
105 vp4567 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp4567, vt4567));
106 vp89AB = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp89AB, vt89AB));
107 vpCDEF = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpCDEF, vtCDEF));
108 vpGHIJ = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpGHIJ, vtGHIJ));
109
110 vp0123 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp0123, vt0123));
111 vp4567 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp4567, vt4567));
112 vp89AB = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp89AB, vt89AB));
113 vpCDEF = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpCDEF, vtCDEF));
114 vpGHIJ = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpGHIJ, vtGHIJ));
115
116 vp0123 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp0123, vt0123));
117 vp4567 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp4567, vt4567));
118 vp89AB = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp89AB, vt89AB));
119 vpCDEF = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpCDEF, vtCDEF));
120 vpGHIJ = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpGHIJ, vtGHIJ));
121
122 // Reconstruct the final f value:
123 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
124 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
125 // = s + (t * s) * p
126 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
127 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
128 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
129 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
130 vtGHIJ = wasm_f32x4_mul(vtGHIJ, vsGHIJ);
131
132 v128_t vf0123 = wasm_f32x4_add(vs0123, wasm_f32x4_mul(vt0123, vp0123));
133 v128_t vf4567 = wasm_f32x4_add(vs4567, wasm_f32x4_mul(vt4567, vp4567));
134 v128_t vf89AB = wasm_f32x4_add(vs89AB, wasm_f32x4_mul(vt89AB, vp89AB));
135 v128_t vfCDEF = wasm_f32x4_add(vsCDEF, wasm_f32x4_mul(vtCDEF, vpCDEF));
136 v128_t vfGHIJ = wasm_f32x4_add(vsGHIJ, wasm_f32x4_mul(vtGHIJ, vpGHIJ));
137
138 // For inputs below zero cutoff, replace output with +0.0f.
139 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
140 vf0123 = wasm_v128_andnot(vf0123, wasm_f32x4_lt(vx0123, vdenorm_cutoff));
141 vf4567 = wasm_v128_andnot(vf4567, wasm_f32x4_lt(vx4567, vdenorm_cutoff));
142 vf89AB = wasm_v128_andnot(vf89AB, wasm_f32x4_lt(vx89AB, vdenorm_cutoff));
143 vfCDEF = wasm_v128_andnot(vfCDEF, wasm_f32x4_lt(vxCDEF, vdenorm_cutoff));
144 vfGHIJ = wasm_v128_andnot(vfGHIJ, wasm_f32x4_lt(vxGHIJ, vdenorm_cutoff));
145
146 // Store 20 (5x4) outputs at a time.
147 wasm_v128_store(output, vf0123);
148 wasm_v128_store(output + 4, vf4567);
149 wasm_v128_store(output + 8, vf89AB);
150 wasm_v128_store(output + 12, vfCDEF);
151 wasm_v128_store(output + 16, vfGHIJ);
152 output += 20;
153
154 // Accumulate computed exponents.
155 vacc0 = wasm_f32x4_add(vacc0, vf0123);
156 vacc0 = wasm_f32x4_add(vacc0, vf4567);
157 vacc0 = wasm_f32x4_add(vacc0, vf89AB);
158 vacc0 = wasm_f32x4_add(vacc0, vfCDEF);
159 vacc0 = wasm_f32x4_add(vacc0, vfGHIJ);
160 }
161 // Add up all accumulators to vacc0
162 vacc0 = wasm_f32x4_add(vacc0, vacc1);
163
164 v128_t vacc = vacc0;
165 for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
166 // Load 4 inputs at a time.
167 const v128_t vi = wasm_v128_load(input);
168 input += 4;
169
170 // Subtract maximum input x := i - i_max. This implies x <= 0.
171 const v128_t vx = wasm_f32x4_sub(vi, vi_max);
172
173 // Compute reduced argument elements := round(x / log(2)).
174 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
175
176 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
177 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
178 const v128_t vs = wasm_i32x4_shl(vn, 23);
179
180 // Subtract the large number back to get final elements := round(x / log(2)).
181 vn = wasm_f32x4_sub(vn, vmagic_bias);
182
183 // Compute reduced argument t := x - elements * log(2).
184 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
185 v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
186 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
187
188 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
189 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
190 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
191 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
192 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
193
194 // Reconstruct the final f value:
195 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
196 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
197 // = s + (t * s) * p
198 vt = wasm_f32x4_mul(vt, vs);
199 v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
200
201 // For inputs below zero cutoff, replace output with +0.0f.
202 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
203 vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
204
205 // Store 4 outputs at a time.
206 wasm_v128_store(output, vf);
207 output += 4;
208
209 // Accumulate computed exponents.
210 vacc = wasm_f32x4_add(vacc, vf);
211 }
212 vacc = wasm_f32x4_add(vacc, wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3));
213 float vsum = wasm_f32x4_extract_lane(vacc, 0) + wasm_f32x4_extract_lane(vacc, 1);
214 if (elements != 0) {
215 assert(elements >= 1 * sizeof(float));
216 assert(elements <= 3 * sizeof(float));
217 // Load 4 inputs at a time.
218 const v128_t vi = wasm_v128_load(input);
219
220 // Subtract maximum input x := i - i_max. This implies x <= 0.
221 const v128_t vx = wasm_f32x4_sub(vi, vi_max);
222
223 // Compute reduced argument elements := round(x / log(2)).
224 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
225
226 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
227 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
228 const v128_t vs = wasm_i32x4_shl(vn, 23);
229
230 // Subtract the large number back to get final elements := round(x / log(2)).
231 vn = wasm_f32x4_sub(vn, vmagic_bias);
232
233 // Compute reduced argument t := x - elements * log(2).
234 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
235 v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
236 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
237
238 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
239 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
240 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
241 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
242 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
243
244 // Reconstruct the final f value:
245 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
246 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
247 // = s + (t * s) * p
248 vt = wasm_f32x4_mul(vt, vs);
249 v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
250
251 // For inputs below zero cutoff, replace output with +0.0f.
252 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
253 vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
254
255 if (elements & (2 * sizeof(float))) {
256 // Store and accumulate 2 outputs at a time.
257 const float vf0 = wasm_f32x4_extract_lane(vf, 0);
258 output[0] = vf0;
259 vsum += vf0;
260
261 const float vf1 = wasm_f32x4_extract_lane(vf, 1);
262 output[1] = vf1;
263 vsum += vf1;
264
265 vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3);
266 output += 2;
267 }
268 if (elements & (1 * sizeof(float))) {
269 // Store 1 output at a time.
270 const float vf0 = wasm_f32x4_extract_lane(vf, 0);
271 *output = vf0;
272 vsum += vf0;
273 }
274 }
275 // Reduce 4 elements in the SIMD register
276 *sum = vsum;
277 }
278