xref: /aosp_15_r20/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/sse2-rr2-p5-x4.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_x4(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_x4(
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   for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
42     // Load 4 (1x4) inputs at a time.
43     const __m128 vi0123 = _mm_loadu_ps(input);
44     input += 4;
45 
46     // Subtract maximum input x := i - i_max. This implies x <= 0.
47     const __m128 vx0123 = _mm_sub_ps(vi0123, vi_max);
48 
49     // Compute reduced argument elements := round(x / log(2)).
50     __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vx0123, vlog2e), vmagic_bias);
51 
52     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
53     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
54     const __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
55 
56     // Subtract the large number back to get final elements := round(x / log(2)).
57     vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
58 
59     // Compute reduced argument t := x - elements * log(2).
60     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
61     __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vx0123);
62 
63     vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
64 
65     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
66     __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4);
67 
68     vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
69 
70     vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
71 
72     vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1);
73 
74     // Reconstruct the final f value:
75     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
76     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
77     //     = s + (t * s) * p
78     vt0123 = _mm_mul_ps(vt0123, vs0123);
79 
80     __m128 vf0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123);
81 
82     // For inputs below zero cutoff, replace output with +0.0f.
83     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
84     vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vx0123, vdenorm_cutoff), vf0123);
85 
86     // Store 4 (1x4) outputs at a time.
87     _mm_storeu_ps(output, vf0123);
88     output += 4;
89 
90     // Accumulate computed exponents.
91     vacc0 = _mm_add_ps(vacc0, vf0123);
92   }
93 
94   __m128 vacc = vacc0;
95   for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
96     // Load 4 inputs at a time.
97     const __m128 vi = _mm_loadu_ps(input);
98     input += 4;
99 
100     // Subtract maximum input x := i - i_max. This implies x <= 0.
101     const __m128 vx = _mm_sub_ps(vi, vi_max);
102 
103     // Compute reduced argument elements := round(x / log(2)).
104     __m128 vn = _mm_add_ps(_mm_mul_ps(vx, vlog2e), vmagic_bias);
105 
106     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
107     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
108     const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
109 
110     // Subtract the large number back to get final elements := round(x / log(2)).
111     vn = _mm_sub_ps(vn, vmagic_bias);
112 
113     // Compute reduced argument t := x - elements * log(2).
114     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
115     __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vx);
116     vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
117 
118     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
119     __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
120     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
121     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
122     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
123 
124     // Reconstruct the final f value:
125     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
126     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
127     //     = s + (t * s) * p
128     vt = _mm_mul_ps(vt, vs);
129     __m128 vf = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
130 
131     // For inputs below zero cutoff, replace output with +0.0f.
132     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
133     vf = _mm_andnot_ps(_mm_cmplt_ps(vx, vdenorm_cutoff), vf);
134 
135     // Store 4 outputs at a time.
136     _mm_storeu_ps(output, vf);
137     output += 4;
138 
139     // Accumulate computed exponents.
140     vacc = _mm_add_ps(vacc, vf);
141   }
142   if (elements != 0) {
143     assert(elements >= 1 * sizeof(float));
144     assert(elements <= 3 * sizeof(float));
145     // Load 4 inputs at a time.
146     const __m128 vi = _mm_loadu_ps(input);
147 
148     // Subtract maximum input x := i - i_max. This implies x <= 0.
149     const __m128 vx = _mm_sub_ps(vi, vi_max);
150 
151     // Compute reduced argument elements := round(x / log(2)).
152     __m128 vn = _mm_add_ps(_mm_mul_ps(vx, vlog2e), vmagic_bias);
153 
154     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
155     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
156     const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
157 
158     // Subtract the large number back to get final elements := round(x / log(2)).
159     vn = _mm_sub_ps(vn, vmagic_bias);
160 
161     // Compute reduced argument t := x - elements * log(2).
162     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
163     __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vx);
164     vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
165 
166     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
167     __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4);
168     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
169     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
170     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1);
171 
172     // Reconstruct the final f value:
173     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
174     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
175     //     = s + (t * s) * p
176     vt = _mm_mul_ps(vt, vs);
177     __m128 vf = _mm_add_ps(_mm_mul_ps(vt, vp), vs);
178 
179     // For inputs below zero cutoff, replace output with +0.0f.
180     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
181     vf = _mm_andnot_ps(_mm_cmplt_ps(vx, vdenorm_cutoff), vf);
182 
183     if (elements & (2 * sizeof(float))) {
184       // Store 2 outputs at a time.
185       _mm_storel_pi((__m64*) output, vf);
186       output += 2;
187 
188       // Accumulate 2 computed exponents.
189       vacc = _mm_add_ps(vacc, _mm_movelh_ps(vf, _mm_setzero_ps()));
190 
191       vf = _mm_movehl_ps(vf, vf);
192     }
193     if (elements & (1 * sizeof(float))) {
194       // Store 1 output at a time.
195       _mm_store_ss(output, vf);
196 
197       // Accumulate 1 computed exponent.
198       vacc = _mm_add_ss(vacc, vf);
199     }
200   }
201   // Reduce 4 elements in the SIMD register
202   vacc = _mm_add_ps(vacc, _mm_movehl_ps(vacc, vacc));
203   vacc = _mm_add_ss(vacc, _mm_shuffle_ps(vacc, vacc, _MM_SHUFFLE(2, 3, 0, 1)));
204   _mm_store_ss(sum, vacc);
205 }
206