xref: /aosp_15_r20/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/wasmsimd-rr2-p5-x12.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_x12(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_x12(
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 >= 12 * sizeof(float); elements -= 12 * sizeof(float)) {
42     // Load 12 (3x4) 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     input += 12;
47 
48     // Subtract maximum input x := i - i_max. This implies x <= 0.
49     const v128_t vx0123 = wasm_f32x4_sub(vi0123, vi_max);
50     const v128_t vx4567 = wasm_f32x4_sub(vi4567, vi_max);
51     const v128_t vx89AB = wasm_f32x4_sub(vi89AB, vi_max);
52 
53     // Compute reduced argument elements := round(x / log(2)).
54     v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx0123, vlog2e));
55     v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx4567, vlog2e));
56     v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx89AB, vlog2e));
57 
58     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
59     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
60     const v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
61     const v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
62     const v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23);
63 
64     // Subtract the large number back to get final elements := round(x / log(2)).
65     vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
66     vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
67     vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
68 
69     // Compute reduced argument t := x - elements * log(2).
70     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
71     v128_t vt0123 = wasm_f32x4_add(vx0123, wasm_f32x4_mul(vn0123, vminus_ln2_hi));
72     v128_t vt4567 = wasm_f32x4_add(vx4567, wasm_f32x4_mul(vn4567, vminus_ln2_hi));
73     v128_t vt89AB = wasm_f32x4_add(vx89AB, wasm_f32x4_mul(vn89AB, vminus_ln2_hi));
74 
75     vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vminus_ln2_lo));
76     vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vminus_ln2_lo));
77     vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vminus_ln2_lo));
78 
79     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
80     v128_t vp0123 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt0123));
81     v128_t vp4567 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt4567));
82     v128_t vp89AB = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt89AB));
83 
84     vp0123 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp0123, vt0123));
85     vp4567 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp4567, vt4567));
86     vp89AB = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp89AB, vt89AB));
87 
88     vp0123 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp0123, vt0123));
89     vp4567 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp4567, vt4567));
90     vp89AB = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp89AB, vt89AB));
91 
92     vp0123 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp0123, vt0123));
93     vp4567 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp4567, vt4567));
94     vp89AB = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp89AB, vt89AB));
95 
96     // Reconstruct the final f value:
97     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
98     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
99     //     = s + (t * s) * p
100     vt0123 = wasm_f32x4_mul(vt0123, vs0123);
101     vt4567 = wasm_f32x4_mul(vt4567, vs4567);
102     vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
103 
104     v128_t vf0123 = wasm_f32x4_add(vs0123, wasm_f32x4_mul(vt0123, vp0123));
105     v128_t vf4567 = wasm_f32x4_add(vs4567, wasm_f32x4_mul(vt4567, vp4567));
106     v128_t vf89AB = wasm_f32x4_add(vs89AB, wasm_f32x4_mul(vt89AB, vp89AB));
107 
108     // For inputs below zero cutoff, replace output with +0.0f.
109     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
110     vf0123 = wasm_v128_andnot(vf0123, wasm_f32x4_lt(vx0123, vdenorm_cutoff));
111     vf4567 = wasm_v128_andnot(vf4567, wasm_f32x4_lt(vx4567, vdenorm_cutoff));
112     vf89AB = wasm_v128_andnot(vf89AB, wasm_f32x4_lt(vx89AB, vdenorm_cutoff));
113 
114     // Store 12 (3x4) outputs at a time.
115     wasm_v128_store(output, vf0123);
116     wasm_v128_store(output + 4, vf4567);
117     wasm_v128_store(output + 8, vf89AB);
118     output += 12;
119 
120     // Accumulate computed exponents.
121     vacc0 = wasm_f32x4_add(vacc0, vf0123);
122     vacc0 = wasm_f32x4_add(vacc0, vf4567);
123     vacc0 = wasm_f32x4_add(vacc0, vf89AB);
124   }
125 
126   v128_t vacc = vacc0;
127   for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
128     // Load 4 inputs at a time.
129     const v128_t vi = wasm_v128_load(input);
130     input += 4;
131 
132     // Subtract maximum input x := i - i_max. This implies x <= 0.
133     const v128_t vx = wasm_f32x4_sub(vi, vi_max);
134 
135     // Compute reduced argument elements := round(x / log(2)).
136     v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
137 
138     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
139     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
140     const v128_t vs = wasm_i32x4_shl(vn, 23);
141 
142     // Subtract the large number back to get final elements := round(x / log(2)).
143     vn = wasm_f32x4_sub(vn, vmagic_bias);
144 
145     // Compute reduced argument t := x - elements * log(2).
146     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
147     v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
148     vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
149 
150     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
151     v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
152     vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
153     vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
154     vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
155 
156     // Reconstruct the final f value:
157     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
158     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
159     //     = s + (t * s) * p
160     vt = wasm_f32x4_mul(vt, vs);
161     v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
162 
163     // For inputs below zero cutoff, replace output with +0.0f.
164     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
165     vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
166 
167     // Store 4 outputs at a time.
168     wasm_v128_store(output, vf);
169     output += 4;
170 
171     // Accumulate computed exponents.
172     vacc = wasm_f32x4_add(vacc, vf);
173   }
174   vacc = wasm_f32x4_add(vacc, wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3));
175   float vsum = wasm_f32x4_extract_lane(vacc, 0) + wasm_f32x4_extract_lane(vacc, 1);
176   if (elements != 0) {
177     assert(elements >= 1 * sizeof(float));
178     assert(elements <= 3 * sizeof(float));
179     // Load 4 inputs at a time.
180     const v128_t vi = wasm_v128_load(input);
181 
182     // Subtract maximum input x := i - i_max. This implies x <= 0.
183     const v128_t vx = wasm_f32x4_sub(vi, vi_max);
184 
185     // Compute reduced argument elements := round(x / log(2)).
186     v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
187 
188     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
189     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
190     const v128_t vs = wasm_i32x4_shl(vn, 23);
191 
192     // Subtract the large number back to get final elements := round(x / log(2)).
193     vn = wasm_f32x4_sub(vn, vmagic_bias);
194 
195     // Compute reduced argument t := x - elements * log(2).
196     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
197     v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
198     vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
199 
200     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
201     v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
202     vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
203     vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
204     vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
205 
206     // Reconstruct the final f value:
207     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
208     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
209     //     = s + (t * s) * p
210     vt = wasm_f32x4_mul(vt, vs);
211     v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
212 
213     // For inputs below zero cutoff, replace output with +0.0f.
214     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
215     vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
216 
217     if (elements & (2 * sizeof(float))) {
218       // Store and accumulate 2 outputs at a time.
219       const float vf0 = wasm_f32x4_extract_lane(vf, 0);
220       output[0] = vf0;
221       vsum += vf0;
222 
223       const float vf1 = wasm_f32x4_extract_lane(vf, 1);
224       output[1] = vf1;
225       vsum += vf1;
226 
227       vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3);
228       output += 2;
229     }
230     if (elements & (1 * sizeof(float))) {
231       // Store 1 output at a time.
232       const float vf0 = wasm_f32x4_extract_lane(vf, 0);
233       *output = vf0;
234       vsum += vf0;
235     }
236   }
237   // Reduce 4 elements in the SIMD register
238   *sum = vsum;
239 }
240