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_acc4(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_acc4(
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 v128_t vacc2 = vacc0;
43 v128_t vacc3 = vacc0;
44 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
45 // Load 16 (4x4) inputs at a time.
46 const v128_t vi0123 = wasm_v128_load(input);
47 const v128_t vi4567 = wasm_v128_load(input + 4);
48 const v128_t vi89AB = wasm_v128_load(input + 8);
49 const v128_t viCDEF = wasm_v128_load(input + 12);
50 input += 16;
51
52 // Subtract maximum input x := i - i_max. This implies x <= 0.
53 const v128_t vx0123 = wasm_f32x4_sub(vi0123, vi_max);
54 const v128_t vx4567 = wasm_f32x4_sub(vi4567, vi_max);
55 const v128_t vx89AB = wasm_f32x4_sub(vi89AB, vi_max);
56 const v128_t vxCDEF = wasm_f32x4_sub(viCDEF, 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
64 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
65 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
66 const v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
67 const v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
68 const v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23);
69 const v128_t vsCDEF = wasm_i32x4_shl(vnCDEF, 23);
70
71 // Subtract the large number back to get final elements := round(x / log(2)).
72 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
73 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
74 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
75 vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
76
77 // Compute reduced argument t := x - elements * log(2).
78 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
79 v128_t vt0123 = wasm_f32x4_add(vx0123, wasm_f32x4_mul(vn0123, vminus_ln2_hi));
80 v128_t vt4567 = wasm_f32x4_add(vx4567, wasm_f32x4_mul(vn4567, vminus_ln2_hi));
81 v128_t vt89AB = wasm_f32x4_add(vx89AB, wasm_f32x4_mul(vn89AB, vminus_ln2_hi));
82 v128_t vtCDEF = wasm_f32x4_add(vxCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_hi));
83
84 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vminus_ln2_lo));
85 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vminus_ln2_lo));
86 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vminus_ln2_lo));
87 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_lo));
88
89 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
90 v128_t vp0123 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt0123));
91 v128_t vp4567 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt4567));
92 v128_t vp89AB = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt89AB));
93 v128_t vpCDEF = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtCDEF));
94
95 vp0123 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp0123, vt0123));
96 vp4567 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp4567, vt4567));
97 vp89AB = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp89AB, vt89AB));
98 vpCDEF = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpCDEF, vtCDEF));
99
100 vp0123 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp0123, vt0123));
101 vp4567 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp4567, vt4567));
102 vp89AB = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp89AB, vt89AB));
103 vpCDEF = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpCDEF, vtCDEF));
104
105 vp0123 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp0123, vt0123));
106 vp4567 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp4567, vt4567));
107 vp89AB = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp89AB, vt89AB));
108 vpCDEF = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpCDEF, vtCDEF));
109
110 // Reconstruct the final f value:
111 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
112 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
113 // = s + (t * s) * p
114 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
115 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
116 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
117 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
118
119 v128_t vf0123 = wasm_f32x4_add(vs0123, wasm_f32x4_mul(vt0123, vp0123));
120 v128_t vf4567 = wasm_f32x4_add(vs4567, wasm_f32x4_mul(vt4567, vp4567));
121 v128_t vf89AB = wasm_f32x4_add(vs89AB, wasm_f32x4_mul(vt89AB, vp89AB));
122 v128_t vfCDEF = wasm_f32x4_add(vsCDEF, wasm_f32x4_mul(vtCDEF, vpCDEF));
123
124 // For inputs below zero cutoff, replace output with +0.0f.
125 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
126 vf0123 = wasm_v128_andnot(vf0123, wasm_f32x4_lt(vx0123, vdenorm_cutoff));
127 vf4567 = wasm_v128_andnot(vf4567, wasm_f32x4_lt(vx4567, vdenorm_cutoff));
128 vf89AB = wasm_v128_andnot(vf89AB, wasm_f32x4_lt(vx89AB, vdenorm_cutoff));
129 vfCDEF = wasm_v128_andnot(vfCDEF, wasm_f32x4_lt(vxCDEF, vdenorm_cutoff));
130
131 // Store 16 (4x4) outputs at a time.
132 wasm_v128_store(output, vf0123);
133 wasm_v128_store(output + 4, vf4567);
134 wasm_v128_store(output + 8, vf89AB);
135 wasm_v128_store(output + 12, vfCDEF);
136 output += 16;
137
138 // Accumulate computed exponents.
139 vacc0 = wasm_f32x4_add(vacc0, vf0123);
140 vacc0 = wasm_f32x4_add(vacc0, vf4567);
141 vacc0 = wasm_f32x4_add(vacc0, vf89AB);
142 vacc0 = wasm_f32x4_add(vacc0, vfCDEF);
143 }
144 // Add up all accumulators to vacc0
145 vacc0 = wasm_f32x4_add(vacc0, vacc1);
146 vacc2 = wasm_f32x4_add(vacc2, vacc3);
147 vacc0 = wasm_f32x4_add(vacc0, vacc2);
148
149 v128_t vacc = vacc0;
150 for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
151 // Load 4 inputs at a time.
152 const v128_t vi = wasm_v128_load(input);
153 input += 4;
154
155 // Subtract maximum input x := i - i_max. This implies x <= 0.
156 const v128_t vx = wasm_f32x4_sub(vi, vi_max);
157
158 // Compute reduced argument elements := round(x / log(2)).
159 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
160
161 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
162 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
163 const v128_t vs = wasm_i32x4_shl(vn, 23);
164
165 // Subtract the large number back to get final elements := round(x / log(2)).
166 vn = wasm_f32x4_sub(vn, vmagic_bias);
167
168 // Compute reduced argument t := x - elements * log(2).
169 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
170 v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
171 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
172
173 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
174 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
175 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
176 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
177 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
178
179 // Reconstruct the final f value:
180 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
181 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
182 // = s + (t * s) * p
183 vt = wasm_f32x4_mul(vt, vs);
184 v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
185
186 // For inputs below zero cutoff, replace output with +0.0f.
187 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
188 vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
189
190 // Store 4 outputs at a time.
191 wasm_v128_store(output, vf);
192 output += 4;
193
194 // Accumulate computed exponents.
195 vacc = wasm_f32x4_add(vacc, vf);
196 }
197 vacc = wasm_f32x4_add(vacc, wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3));
198 float vsum = wasm_f32x4_extract_lane(vacc, 0) + wasm_f32x4_extract_lane(vacc, 1);
199 if (elements != 0) {
200 assert(elements >= 1 * sizeof(float));
201 assert(elements <= 3 * sizeof(float));
202 // Load 4 inputs at a time.
203 const v128_t vi = wasm_v128_load(input);
204
205 // Subtract maximum input x := i - i_max. This implies x <= 0.
206 const v128_t vx = wasm_f32x4_sub(vi, vi_max);
207
208 // Compute reduced argument elements := round(x / log(2)).
209 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
210
211 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
212 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
213 const v128_t vs = wasm_i32x4_shl(vn, 23);
214
215 // Subtract the large number back to get final elements := round(x / log(2)).
216 vn = wasm_f32x4_sub(vn, vmagic_bias);
217
218 // Compute reduced argument t := x - elements * log(2).
219 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
220 v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
221 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
222
223 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
224 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
225 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
226 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
227 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
228
229 // Reconstruct the final f value:
230 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
231 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
232 // = s + (t * s) * p
233 vt = wasm_f32x4_mul(vt, vs);
234 v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
235
236 // For inputs below zero cutoff, replace output with +0.0f.
237 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
238 vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
239
240 if (elements & (2 * sizeof(float))) {
241 // Store and accumulate 2 outputs at a time.
242 const float vf0 = wasm_f32x4_extract_lane(vf, 0);
243 output[0] = vf0;
244 vsum += vf0;
245
246 const float vf1 = wasm_f32x4_extract_lane(vf, 1);
247 output[1] = vf1;
248 vsum += vf1;
249
250 vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3);
251 output += 2;
252 }
253 if (elements & (1 * sizeof(float))) {
254 // Store 1 output at a time.
255 const float vf0 = wasm_f32x4_extract_lane(vf, 0);
256 *output = vf0;
257 vsum += vf0;
258 }
259 }
260 // Reduce 4 elements in the SIMD register
261 *sum = vsum;
262 }
263