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_x20_acc2(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_x20_acc2(
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 v128_t vacc1 = vacc0;
42 for (; elements >= 20 * sizeof(float); elements -= 20 * sizeof(float)) {
43 // Load 20 (5x4) inputs at a time.
44 const v128_t vi0123 = wasm_v128_load(input);
45 const v128_t vi4567 = wasm_v128_load(input + 4);
46 const v128_t vi89AB = wasm_v128_load(input + 8);
47 const v128_t viCDEF = wasm_v128_load(input + 12);
48 const v128_t viGHIJ = wasm_v128_load(input + 16);
49 input += 20;
50
51 // Subtract maximum input x := i - i_max. This implies x <= 0.
52 const v128_t vx0123 = wasm_f32x4_sub(vi0123, vi_max);
53 const v128_t vx4567 = wasm_f32x4_sub(vi4567, vi_max);
54 const v128_t vx89AB = wasm_f32x4_sub(vi89AB, vi_max);
55 const v128_t vxCDEF = wasm_f32x4_sub(viCDEF, vi_max);
56 const v128_t vxGHIJ = wasm_f32x4_sub(viGHIJ, vi_max);
57
58 // Compute reduced argument elements := round(x / log(2)).
59 v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx0123, vlog2e));
60 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx4567, vlog2e));
61 v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx89AB, vlog2e));
62 v128_t vnCDEF = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vxCDEF, vlog2e));
63 v128_t vnGHIJ = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vxGHIJ, vlog2e));
64
65 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
66 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
67 const v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
68 const v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
69 const v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23);
70 const v128_t vsCDEF = wasm_i32x4_shl(vnCDEF, 23);
71 const v128_t vsGHIJ = wasm_i32x4_shl(vnGHIJ, 23);
72
73 // Subtract the large number back to get final elements := round(x / log(2)).
74 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
75 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
76 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
77 vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
78 vnGHIJ = wasm_f32x4_sub(vnGHIJ, vmagic_bias);
79
80 // Compute reduced argument t := x - elements * log(2).
81 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
82 v128_t vt0123 = wasm_f32x4_add(vx0123, wasm_f32x4_mul(vn0123, vminus_ln2_hi));
83 v128_t vt4567 = wasm_f32x4_add(vx4567, wasm_f32x4_mul(vn4567, vminus_ln2_hi));
84 v128_t vt89AB = wasm_f32x4_add(vx89AB, wasm_f32x4_mul(vn89AB, vminus_ln2_hi));
85 v128_t vtCDEF = wasm_f32x4_add(vxCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_hi));
86 v128_t vtGHIJ = wasm_f32x4_add(vxGHIJ, wasm_f32x4_mul(vnGHIJ, vminus_ln2_hi));
87
88 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vminus_ln2_lo));
89 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vminus_ln2_lo));
90 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vminus_ln2_lo));
91 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_lo));
92 vtGHIJ = wasm_f32x4_add(vtGHIJ, wasm_f32x4_mul(vnGHIJ, vminus_ln2_lo));
93
94 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
95 v128_t vp0123 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt0123));
96 v128_t vp4567 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt4567));
97 v128_t vp89AB = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt89AB));
98 v128_t vpCDEF = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtCDEF));
99 v128_t vpGHIJ = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtGHIJ));
100
101 vp0123 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp0123, vt0123));
102 vp4567 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp4567, vt4567));
103 vp89AB = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp89AB, vt89AB));
104 vpCDEF = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpCDEF, vtCDEF));
105 vpGHIJ = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpGHIJ, vtGHIJ));
106
107 vp0123 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp0123, vt0123));
108 vp4567 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp4567, vt4567));
109 vp89AB = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp89AB, vt89AB));
110 vpCDEF = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpCDEF, vtCDEF));
111 vpGHIJ = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpGHIJ, vtGHIJ));
112
113 vp0123 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp0123, vt0123));
114 vp4567 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp4567, vt4567));
115 vp89AB = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp89AB, vt89AB));
116 vpCDEF = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpCDEF, vtCDEF));
117 vpGHIJ = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpGHIJ, vtGHIJ));
118
119 // Reconstruct the final f value:
120 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
121 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
122 // = s + (t * s) * p
123 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
124 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
125 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
126 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
127 vtGHIJ = wasm_f32x4_mul(vtGHIJ, vsGHIJ);
128
129 v128_t vf0123 = wasm_f32x4_add(vs0123, wasm_f32x4_mul(vt0123, vp0123));
130 v128_t vf4567 = wasm_f32x4_add(vs4567, wasm_f32x4_mul(vt4567, vp4567));
131 v128_t vf89AB = wasm_f32x4_add(vs89AB, wasm_f32x4_mul(vt89AB, vp89AB));
132 v128_t vfCDEF = wasm_f32x4_add(vsCDEF, wasm_f32x4_mul(vtCDEF, vpCDEF));
133 v128_t vfGHIJ = wasm_f32x4_add(vsGHIJ, wasm_f32x4_mul(vtGHIJ, vpGHIJ));
134
135 // For inputs below zero cutoff, replace output with +0.0f.
136 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
137 vf0123 = wasm_v128_andnot(vf0123, wasm_f32x4_lt(vx0123, vdenorm_cutoff));
138 vf4567 = wasm_v128_andnot(vf4567, wasm_f32x4_lt(vx4567, vdenorm_cutoff));
139 vf89AB = wasm_v128_andnot(vf89AB, wasm_f32x4_lt(vx89AB, vdenorm_cutoff));
140 vfCDEF = wasm_v128_andnot(vfCDEF, wasm_f32x4_lt(vxCDEF, vdenorm_cutoff));
141 vfGHIJ = wasm_v128_andnot(vfGHIJ, wasm_f32x4_lt(vxGHIJ, vdenorm_cutoff));
142
143 // Store 20 (5x4) outputs at a time.
144 wasm_v128_store(output, vf0123);
145 wasm_v128_store(output + 4, vf4567);
146 wasm_v128_store(output + 8, vf89AB);
147 wasm_v128_store(output + 12, vfCDEF);
148 wasm_v128_store(output + 16, vfGHIJ);
149 output += 20;
150
151 // Accumulate computed exponents.
152 vacc0 = wasm_f32x4_add(vacc0, vf0123);
153 vacc0 = wasm_f32x4_add(vacc0, vf4567);
154 vacc0 = wasm_f32x4_add(vacc0, vf89AB);
155 vacc0 = wasm_f32x4_add(vacc0, vfCDEF);
156 vacc0 = wasm_f32x4_add(vacc0, vfGHIJ);
157 }
158 // Add up all accumulators to vacc0
159 vacc0 = wasm_f32x4_add(vacc0, vacc1);
160
161 v128_t vacc = vacc0;
162 for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
163 // Load 4 inputs at a time.
164 const v128_t vi = wasm_v128_load(input);
165 input += 4;
166
167 // Subtract maximum input x := i - i_max. This implies x <= 0.
168 const v128_t vx = wasm_f32x4_sub(vi, vi_max);
169
170 // Compute reduced argument elements := round(x / log(2)).
171 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
172
173 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
174 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
175 const v128_t vs = wasm_i32x4_shl(vn, 23);
176
177 // Subtract the large number back to get final elements := round(x / log(2)).
178 vn = wasm_f32x4_sub(vn, vmagic_bias);
179
180 // Compute reduced argument t := x - elements * log(2).
181 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
182 v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
183 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
184
185 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
186 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
187 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
188 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
189 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
190
191 // Reconstruct the final f value:
192 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
193 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
194 // = s + (t * s) * p
195 vt = wasm_f32x4_mul(vt, vs);
196 v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
197
198 // For inputs below zero cutoff, replace output with +0.0f.
199 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
200 vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
201
202 // Store 4 outputs at a time.
203 wasm_v128_store(output, vf);
204 output += 4;
205
206 // Accumulate computed exponents.
207 vacc = wasm_f32x4_add(vacc, vf);
208 }
209 vacc = wasm_f32x4_add(vacc, wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3));
210 float vsum = wasm_f32x4_extract_lane(vacc, 0) + wasm_f32x4_extract_lane(vacc, 1);
211 if (elements != 0) {
212 assert(elements >= 1 * sizeof(float));
213 assert(elements <= 3 * sizeof(float));
214 // Load 4 inputs at a time.
215 const v128_t vi = wasm_v128_load(input);
216
217 // Subtract maximum input x := i - i_max. This implies x <= 0.
218 const v128_t vx = wasm_f32x4_sub(vi, vi_max);
219
220 // Compute reduced argument elements := round(x / log(2)).
221 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
222
223 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
224 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
225 const v128_t vs = wasm_i32x4_shl(vn, 23);
226
227 // Subtract the large number back to get final elements := round(x / log(2)).
228 vn = wasm_f32x4_sub(vn, vmagic_bias);
229
230 // Compute reduced argument t := x - elements * log(2).
231 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
232 v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
233 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
234
235 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
236 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
237 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
238 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
239 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
240
241 // Reconstruct the final f value:
242 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
243 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
244 // = s + (t * s) * p
245 vt = wasm_f32x4_mul(vt, vs);
246 v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
247
248 // For inputs below zero cutoff, replace output with +0.0f.
249 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
250 vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
251
252 if (elements & (2 * sizeof(float))) {
253 // Store and accumulate 2 outputs at a time.
254 const float vf0 = wasm_f32x4_extract_lane(vf, 0);
255 output[0] = vf0;
256 vsum += vf0;
257
258 const float vf1 = wasm_f32x4_extract_lane(vf, 1);
259 output[1] = vf1;
260 vsum += vf1;
261
262 vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3);
263 output += 2;
264 }
265 if (elements & (1 * sizeof(float))) {
266 // Store 1 output at a time.
267 const float vf0 = wasm_f32x4_extract_lane(vf, 0);
268 *output = vf0;
269 vsum += vf0;
270 }
271 }
272 // Reduce 4 elements in the SIMD register
273 *sum = vsum;
274 }
275