xref: /aosp_15_r20/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/sse2-rr2-p5-x12-acc3.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-raddstoreexpminusmax/sse2-rr2-p5.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 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 <emmintrin.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/raddstoreexpminusmax.h>
16 
17 
xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x12_acc3(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__sse2_rr2_p5_x12_acc3(
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 __m128 vi_max = _mm_load1_ps(max);
29   const __m128 vlog2e = _mm_load_ps(params->sse2_rr2_p5.log2e);
30   const __m128 vmagic_bias = _mm_load_ps(params->sse2_rr2_p5.magic_bias);
31   const __m128 vminus_ln2_hi = _mm_load_ps(params->sse2_rr2_p5.minus_ln2_hi);
32   const __m128 vminus_ln2_lo = _mm_load_ps(params->sse2_rr2_p5.minus_ln2_lo);
33   const __m128 vc5 = _mm_load_ps(params->sse2_rr2_p5.c5);
34   const __m128 vc4 = _mm_load_ps(params->sse2_rr2_p5.c4);
35   const __m128 vc3 = _mm_load_ps(params->sse2_rr2_p5.c3);
36   const __m128 vc2 = _mm_load_ps(params->sse2_rr2_p5.c2);
37   const __m128 vc1 = _mm_load_ps(params->sse2_rr2_p5.c1);
38   const __m128 vdenorm_cutoff = _mm_load_ps(params->sse2_rr2_p5.denorm_cutoff);
39 
40   __m128 vacc0 = _mm_setzero_ps();
41   __m128 vacc1 = _mm_setzero_ps();
42   __m128 vacc2 = _mm_setzero_ps();
43   for (; elements >= 12 * sizeof(float); elements -= 12 * sizeof(float)) {
44     // Load 12 (3x4) inputs at a time.
45     const __m128 vi0123 = _mm_loadu_ps(input);
46     const __m128 vi4567 = _mm_loadu_ps(input + 4);
47     const __m128 vi89AB = _mm_loadu_ps(input + 8);
48     input += 12;
49 
50     // Subtract maximum input x := i - i_max. This implies x <= 0.
51     const __m128 vx0123 = _mm_sub_ps(vi0123, vi_max);
52     const __m128 vx4567 = _mm_sub_ps(vi4567, vi_max);
53     const __m128 vx89AB = _mm_sub_ps(vi89AB, vi_max);
54 
55     // Compute reduced argument elements := round(x / log(2)).
56     __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vx0123, vlog2e), vmagic_bias);
57     __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vx4567, vlog2e), vmagic_bias);
58     __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vx89AB, vlog2e), vmagic_bias);
59 
60     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
61     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
62     const __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
63     const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
64     const __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23));
65 
66     // Subtract the large number back to get final elements := round(x / log(2)).
67     vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
68     vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
69     vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
70 
71     // Compute reduced argument t := x - elements * log(2).
72     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
73     __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vx0123);
74     __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vx4567);
75     __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vx89AB);
76 
77     vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
78     vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
79     vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
80 
81     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
82     __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4);
83     __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4);
84     __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc5, vt89AB), vc4);
85 
86     vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
87     vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
88     vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3);
89 
90     vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
91     vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
92     vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2);
93 
94     vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1);
95     vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1);
96     vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc1);
97 
98     // Reconstruct the final f value:
99     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
100     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
101     //     = s + (t * s) * p
102     vt0123 = _mm_mul_ps(vt0123, vs0123);
103     vt4567 = _mm_mul_ps(vt4567, vs4567);
104     vt89AB = _mm_mul_ps(vt89AB, vs89AB);
105 
106     __m128 vf0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123);
107     __m128 vf4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567);
108     __m128 vf89AB = _mm_add_ps(_mm_mul_ps(vt89AB, vp89AB), vs89AB);
109 
110     // For inputs below zero cutoff, replace output with +0.0f.
111     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
112     vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vx0123, vdenorm_cutoff), vf0123);
113     vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vx4567, vdenorm_cutoff), vf4567);
114     vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vx89AB, vdenorm_cutoff), vf89AB);
115 
116     // Store 12 (3x4) outputs at a time.
117     _mm_storeu_ps(output, vf0123);
118     _mm_storeu_ps(output + 4, vf4567);
119     _mm_storeu_ps(output + 8, vf89AB);
120     output += 12;
121 
122     // Accumulate computed exponents.
123     vacc0 = _mm_add_ps(vacc0, vf0123);
124     vacc1 = _mm_add_ps(vacc1, vf4567);
125     vacc2 = _mm_add_ps(vacc2, vf89AB);
126   }
127   // Add up all accumulators to vacc0
128   vacc0 = _mm_add_ps(vacc0, vacc1);
129   vacc0 = _mm_add_ps(vacc0, vacc2);
130 
131   __m128 vacc = vacc0;
132   for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
133     // Load 4 inputs at a time.
134     const __m128 vi = _mm_loadu_ps(input);
135     input += 4;
136 
137     // Subtract maximum input x := i - i_max. This implies x <= 0.
138     const __m128 vx = _mm_sub_ps(vi, vi_max);
139 
140     // Compute reduced argument elements := round(x / log(2)).
141     __m128 vn = _mm_add_ps(_mm_mul_ps(vx, vlog2e), vmagic_bias);
142 
143     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
144     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
145     const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
146 
147     // Subtract the large number back to get final elements := round(x / log(2)).
148     vn = _mm_sub_ps(vn, vmagic_bias);
149 
150     // Compute reduced argument t := x - elements * log(2).
151     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
152     __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vx);
153     vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
154 
155     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
156     __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
157     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
158     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
159     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
160 
161     // Reconstruct the final f value:
162     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
163     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
164     //     = s + (t * s) * p
165     vt = _mm_mul_ps(vt, vs);
166     __m128 vf = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
167 
168     // For inputs below zero cutoff, replace output with +0.0f.
169     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
170     vf = _mm_andnot_ps(_mm_cmplt_ps(vx, vdenorm_cutoff), vf);
171 
172     // Store 4 outputs at a time.
173     _mm_storeu_ps(output, vf);
174     output += 4;
175 
176     // Accumulate computed exponents.
177     vacc = _mm_add_ps(vacc, vf);
178   }
179   if (elements != 0) {
180     assert(elements >= 1 * sizeof(float));
181     assert(elements <= 3 * sizeof(float));
182     // Load 4 inputs at a time.
183     const __m128 vi = _mm_loadu_ps(input);
184 
185     // Subtract maximum input x := i - i_max. This implies x <= 0.
186     const __m128 vx = _mm_sub_ps(vi, vi_max);
187 
188     // Compute reduced argument elements := round(x / log(2)).
189     __m128 vn = _mm_add_ps(_mm_mul_ps(vx, vlog2e), vmagic_bias);
190 
191     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
192     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
193     const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
194 
195     // Subtract the large number back to get final elements := round(x / log(2)).
196     vn = _mm_sub_ps(vn, vmagic_bias);
197 
198     // Compute reduced argument t := x - elements * log(2).
199     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
200     __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vx);
201     vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
202 
203     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
204     __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
205     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
206     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
207     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
208 
209     // Reconstruct the final f value:
210     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
211     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
212     //     = s + (t * s) * p
213     vt = _mm_mul_ps(vt, vs);
214     __m128 vf = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
215 
216     // For inputs below zero cutoff, replace output with +0.0f.
217     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
218     vf = _mm_andnot_ps(_mm_cmplt_ps(vx, vdenorm_cutoff), vf);
219 
220     if (elements & (2 * sizeof(float))) {
221       // Store 2 outputs at a time.
222       _mm_storel_pi((__m64*) output, vf);
223       output += 2;
224 
225       // Accumulate 2 computed exponents.
226       vacc = _mm_add_ps(vacc, _mm_movelh_ps(vf, _mm_setzero_ps()));
227 
228       vf = _mm_movehl_ps(vf, vf);
229     }
230     if (elements & (1 * sizeof(float))) {
231       // Store 1 output at a time.
232       _mm_store_ss(output, vf);
233 
234       // Accumulate 1 computed exponent.
235       vacc = _mm_add_ss(vacc, vf);
236     }
237   }
238   // Reduce 4 elements in the SIMD register
239   vacc = _mm_add_ps(vacc, _mm_movehl_ps(vacc, vacc));
240   vacc = _mm_add_ss(vacc, _mm_shuffle_ps(vacc, vacc, _MM_SHUFFLE(2, 3, 0, 1)));
241   _mm_store_ss(sum, vacc);
242 }
243