xref: /aosp_15_r20/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/wasmsimd-rr2-p5-x16-acc2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_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_x16_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 >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
43     // Load 16 (4x4) 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     input += 16;
49 
50     // Subtract maximum input x := i - i_max. This implies x <= 0.
51     const v128_t vx0123 = wasm_f32x4_sub(vi0123, vi_max);
52     const v128_t vx4567 = wasm_f32x4_sub(vi4567, vi_max);
53     const v128_t vx89AB = wasm_f32x4_sub(vi89AB, vi_max);
54     const v128_t vxCDEF = wasm_f32x4_sub(viCDEF, vi_max);
55 
56     // Compute reduced argument elements := round(x / log(2)).
57     v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx0123, vlog2e));
58     v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx4567, vlog2e));
59     v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx89AB, vlog2e));
60     v128_t vnCDEF = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vxCDEF, vlog2e));
61 
62     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
63     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
64     const v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
65     const v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
66     const v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23);
67     const v128_t vsCDEF = wasm_i32x4_shl(vnCDEF, 23);
68 
69     // Subtract the large number back to get final elements := round(x / log(2)).
70     vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
71     vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
72     vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
73     vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
74 
75     // Compute reduced argument t := x - elements * log(2).
76     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
77     v128_t vt0123 = wasm_f32x4_add(vx0123, wasm_f32x4_mul(vn0123, vminus_ln2_hi));
78     v128_t vt4567 = wasm_f32x4_add(vx4567, wasm_f32x4_mul(vn4567, vminus_ln2_hi));
79     v128_t vt89AB = wasm_f32x4_add(vx89AB, wasm_f32x4_mul(vn89AB, vminus_ln2_hi));
80     v128_t vtCDEF = wasm_f32x4_add(vxCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_hi));
81 
82     vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vminus_ln2_lo));
83     vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vminus_ln2_lo));
84     vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vminus_ln2_lo));
85     vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_lo));
86 
87     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
88     v128_t vp0123 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt0123));
89     v128_t vp4567 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt4567));
90     v128_t vp89AB = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt89AB));
91     v128_t vpCDEF = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtCDEF));
92 
93     vp0123 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp0123, vt0123));
94     vp4567 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp4567, vt4567));
95     vp89AB = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp89AB, vt89AB));
96     vpCDEF = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpCDEF, vtCDEF));
97 
98     vp0123 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp0123, vt0123));
99     vp4567 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp4567, vt4567));
100     vp89AB = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp89AB, vt89AB));
101     vpCDEF = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpCDEF, vtCDEF));
102 
103     vp0123 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp0123, vt0123));
104     vp4567 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp4567, vt4567));
105     vp89AB = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp89AB, vt89AB));
106     vpCDEF = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpCDEF, vtCDEF));
107 
108     // Reconstruct the final f value:
109     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
110     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
111     //     = s + (t * s) * p
112     vt0123 = wasm_f32x4_mul(vt0123, vs0123);
113     vt4567 = wasm_f32x4_mul(vt4567, vs4567);
114     vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
115     vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
116 
117     v128_t vf0123 = wasm_f32x4_add(vs0123, wasm_f32x4_mul(vt0123, vp0123));
118     v128_t vf4567 = wasm_f32x4_add(vs4567, wasm_f32x4_mul(vt4567, vp4567));
119     v128_t vf89AB = wasm_f32x4_add(vs89AB, wasm_f32x4_mul(vt89AB, vp89AB));
120     v128_t vfCDEF = wasm_f32x4_add(vsCDEF, wasm_f32x4_mul(vtCDEF, vpCDEF));
121 
122     // For inputs below zero cutoff, replace output with +0.0f.
123     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
124     vf0123 = wasm_v128_andnot(vf0123, wasm_f32x4_lt(vx0123, vdenorm_cutoff));
125     vf4567 = wasm_v128_andnot(vf4567, wasm_f32x4_lt(vx4567, vdenorm_cutoff));
126     vf89AB = wasm_v128_andnot(vf89AB, wasm_f32x4_lt(vx89AB, vdenorm_cutoff));
127     vfCDEF = wasm_v128_andnot(vfCDEF, wasm_f32x4_lt(vxCDEF, vdenorm_cutoff));
128 
129     // Store 16 (4x4) outputs at a time.
130     wasm_v128_store(output, vf0123);
131     wasm_v128_store(output + 4, vf4567);
132     wasm_v128_store(output + 8, vf89AB);
133     wasm_v128_store(output + 12, vfCDEF);
134     output += 16;
135 
136     // Accumulate computed exponents.
137     vacc0 = wasm_f32x4_add(vacc0, vf0123);
138     vacc0 = wasm_f32x4_add(vacc0, vf4567);
139     vacc0 = wasm_f32x4_add(vacc0, vf89AB);
140     vacc0 = wasm_f32x4_add(vacc0, vfCDEF);
141   }
142   // Add up all accumulators to vacc0
143   vacc0 = wasm_f32x4_add(vacc0, vacc1);
144 
145   v128_t vacc = vacc0;
146   for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
147     // Load 4 inputs at a time.
148     const v128_t vi = wasm_v128_load(input);
149     input += 4;
150 
151     // Subtract maximum input x := i - i_max. This implies x <= 0.
152     const v128_t vx = wasm_f32x4_sub(vi, vi_max);
153 
154     // Compute reduced argument elements := round(x / log(2)).
155     v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
156 
157     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
158     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
159     const v128_t vs = wasm_i32x4_shl(vn, 23);
160 
161     // Subtract the large number back to get final elements := round(x / log(2)).
162     vn = wasm_f32x4_sub(vn, vmagic_bias);
163 
164     // Compute reduced argument t := x - elements * log(2).
165     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
166     v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
167     vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
168 
169     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
170     v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
171     vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
172     vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
173     vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
174 
175     // Reconstruct the final f value:
176     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
177     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
178     //     = s + (t * s) * p
179     vt = wasm_f32x4_mul(vt, vs);
180     v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
181 
182     // For inputs below zero cutoff, replace output with +0.0f.
183     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
184     vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
185 
186     // Store 4 outputs at a time.
187     wasm_v128_store(output, vf);
188     output += 4;
189 
190     // Accumulate computed exponents.
191     vacc = wasm_f32x4_add(vacc, vf);
192   }
193   vacc = wasm_f32x4_add(vacc, wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3));
194   float vsum = wasm_f32x4_extract_lane(vacc, 0) + wasm_f32x4_extract_lane(vacc, 1);
195   if (elements != 0) {
196     assert(elements >= 1 * sizeof(float));
197     assert(elements <= 3 * sizeof(float));
198     // Load 4 inputs at a time.
199     const v128_t vi = wasm_v128_load(input);
200 
201     // Subtract maximum input x := i - i_max. This implies x <= 0.
202     const v128_t vx = wasm_f32x4_sub(vi, vi_max);
203 
204     // Compute reduced argument elements := round(x / log(2)).
205     v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
206 
207     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
208     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
209     const v128_t vs = wasm_i32x4_shl(vn, 23);
210 
211     // Subtract the large number back to get final elements := round(x / log(2)).
212     vn = wasm_f32x4_sub(vn, vmagic_bias);
213 
214     // Compute reduced argument t := x - elements * log(2).
215     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
216     v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
217     vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
218 
219     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
220     v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
221     vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
222     vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
223     vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
224 
225     // Reconstruct the final f value:
226     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
227     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
228     //     = s + (t * s) * p
229     vt = wasm_f32x4_mul(vt, vs);
230     v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
231 
232     // For inputs below zero cutoff, replace output with +0.0f.
233     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
234     vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
235 
236     if (elements & (2 * sizeof(float))) {
237       // Store and accumulate 2 outputs at a time.
238       const float vf0 = wasm_f32x4_extract_lane(vf, 0);
239       output[0] = vf0;
240       vsum += vf0;
241 
242       const float vf1 = wasm_f32x4_extract_lane(vf, 1);
243       output[1] = vf1;
244       vsum += vf1;
245 
246       vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3);
247       output += 2;
248     }
249     if (elements & (1 * sizeof(float))) {
250       // Store 1 output at a time.
251       const float vf0 = wasm_f32x4_extract_lane(vf, 0);
252       *output = vf0;
253       vsum += vf0;
254     }
255   }
256   // Reduce 4 elements in the SIMD register
257   *sum = vsum;
258 }
259