xref: /aosp_15_r20/external/XNNPACK/src/f32-raddexpminusmax/gen/avx2-p5-x72.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-raddexpminusmax/avx2-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 <immintrin.h>
13 
14 #include <xnnpack/raddexpminusmax.h>
15 
16 
17 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
18 
xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72(size_t elements,const float * input,float * sum,float max)19 void xnn_f32_raddexpminusmax_ukernel__avx2_p5_x72(
20     size_t elements,
21     const float* input,
22     float* sum,
23     float max)
24 {
25   assert(elements % sizeof(float) == 0);
26 
27   const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
28   // The smallest x for which expf(x) is normalized.
29   const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep6f);
30   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
31   const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
32   const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
33 
34   const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
35   const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
36   const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
37   const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
38   const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
39 
40   const __m256 vi_max = _mm256_set1_ps(max);
41 
42   __m256 vacc0 = _mm256_setzero_ps();
43   for (; elements >= 72 * sizeof(float); elements -= 72 * sizeof(float)) {
44     // Load 72 (9x8) inputs at a time.
45     const __m256 vi0 = _mm256_loadu_ps(input);
46     const __m256 vi1 = _mm256_loadu_ps(input + 8);
47     const __m256 vi2 = _mm256_loadu_ps(input + 16);
48     const __m256 vi3 = _mm256_loadu_ps(input + 24);
49     const __m256 vi4 = _mm256_loadu_ps(input + 32);
50     const __m256 vi5 = _mm256_loadu_ps(input + 40);
51     const __m256 vi6 = _mm256_loadu_ps(input + 48);
52     const __m256 vi7 = _mm256_loadu_ps(input + 56);
53     const __m256 vi8 = _mm256_loadu_ps(input + 64);
54     input += 72;
55 
56     // Subtract maximum input x := i - i_max. This implies x <= 0.
57     const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
58     const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
59     const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
60     const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
61     const __m256 vx4 = _mm256_sub_ps(vi4, vi_max);
62     const __m256 vx5 = _mm256_sub_ps(vi5, vi_max);
63     const __m256 vx6 = _mm256_sub_ps(vi6, vi_max);
64     const __m256 vx7 = _mm256_sub_ps(vi7, vi_max);
65     const __m256 vx8 = _mm256_sub_ps(vi8, vi_max);
66 
67     // Compute reduced argument elements := round(x / log(2)).
68     __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
69     __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
70     __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
71     __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
72     __m256 vn4 = _mm256_fmadd_ps(vx4, vlog2e, vmagic_bias);
73     __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias);
74     __m256 vn6 = _mm256_fmadd_ps(vx6, vlog2e, vmagic_bias);
75     __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias);
76     __m256 vn8 = _mm256_fmadd_ps(vx8, vlog2e, vmagic_bias);
77 
78     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
79     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
80     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
81     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
82     const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
83     const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
84     const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
85     const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
86     const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
87     const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23));
88     const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn8), 23));
89 
90     // Subtract the large number back to get final elements := round(x / log(2)).
91     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
92     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
93     vn2 = _mm256_sub_ps(vn2, vmagic_bias);
94     vn3 = _mm256_sub_ps(vn3, vmagic_bias);
95     vn4 = _mm256_sub_ps(vn4, vmagic_bias);
96     vn5 = _mm256_sub_ps(vn5, vmagic_bias);
97     vn6 = _mm256_sub_ps(vn6, vmagic_bias);
98     vn7 = _mm256_sub_ps(vn7, vmagic_bias);
99     vn8 = _mm256_sub_ps(vn8, vmagic_bias);
100 
101     // Compute reduced argument t := x - elements * log(2).
102     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
103     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
104     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
105     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
106     __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
107     __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
108     __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
109     __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
110     __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
111     __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_hi, vx8);
112 
113     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
114     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
115     vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
116     vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
117     vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
118     vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
119     vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
120     vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
121     vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8);
122 
123     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
124     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
125     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
126     __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
127     __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
128     __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
129     __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
130     __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
131     __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
132     __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
133 
134     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
135     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
136     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
137     vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
138     vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
139     vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
140     vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
141     vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
142     vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
143 
144     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
145     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
146     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
147     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
148     vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
149     vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
150     vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
151     vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
152     vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
153 
154     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
155     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
156     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
157     vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
158     vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
159     vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
160     vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
161     vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
162     vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
163 
164     // Reconstruct the final f value:
165     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
166     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
167     //     = s + (t * s) * p
168     vt0 = _mm256_mul_ps(vt0, vs0);
169     vt1 = _mm256_mul_ps(vt1, vs1);
170     vt2 = _mm256_mul_ps(vt2, vs2);
171     vt3 = _mm256_mul_ps(vt3, vs3);
172     vt4 = _mm256_mul_ps(vt4, vs4);
173     vt5 = _mm256_mul_ps(vt5, vs5);
174     vt6 = _mm256_mul_ps(vt6, vs6);
175     vt7 = _mm256_mul_ps(vt7, vs7);
176     vt8 = _mm256_mul_ps(vt8, vs8);
177 
178     __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
179     __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
180     __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
181     __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
182     __m256 vf4 = _mm256_fmadd_ps(vt4, vp4, vs4);
183     __m256 vf5 = _mm256_fmadd_ps(vt5, vp5, vs5);
184     __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6);
185     __m256 vf7 = _mm256_fmadd_ps(vt7, vp7, vs7);
186     __m256 vf8 = _mm256_fmadd_ps(vt8, vp8, vs8);
187 
188     // For inputs below zero cutoff, replace output with +0.0f.
189     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
190     vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
191     vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
192     vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
193     vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
194     vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vx4, vdenorm_cutoff, _CMP_LT_OS), vf4);
195     vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vx5, vdenorm_cutoff, _CMP_LT_OS), vf5);
196     vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vx6, vdenorm_cutoff, _CMP_LT_OS), vf6);
197     vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7);
198     vf8 = _mm256_andnot_ps(_mm256_cmp_ps(vx8, vdenorm_cutoff, _CMP_LT_OS), vf8);
199 
200     // Accumulate computed exponents.
201     vacc0 = _mm256_add_ps(vacc0, vf0);
202     vacc0 = _mm256_add_ps(vacc0, vf1);
203     vacc0 = _mm256_add_ps(vacc0, vf2);
204     vacc0 = _mm256_add_ps(vacc0, vf3);
205     vacc0 = _mm256_add_ps(vacc0, vf4);
206     vacc0 = _mm256_add_ps(vacc0, vf5);
207     vacc0 = _mm256_add_ps(vacc0, vf6);
208     vacc0 = _mm256_add_ps(vacc0, vf7);
209     vacc0 = _mm256_add_ps(vacc0, vf8);
210   }
211 
212   __m256 vacc = vacc0;
213   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
214     // Load 8 inputs at a time.
215     const __m256 vi = _mm256_loadu_ps(input);
216     input += 8;
217 
218     // Subtract maximum input x := i - i_max. This implies x <= 0.
219     const __m256 vx = _mm256_sub_ps(vi, vi_max);
220 
221     // Compute reduced argument elements := round(x / log(2)).
222     __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
223 
224     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
225     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
226     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
227 
228     // Subtract the large number back to get final elements := round(x / log(2)).
229     vn = _mm256_sub_ps(vn, vmagic_bias);
230 
231     // Compute reduced argument t := x - elements * log(2).
232     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
233     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
234     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
235 
236     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
237     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
238     vp = _mm256_fmadd_ps(vp, vt, vc3);
239     vp = _mm256_fmadd_ps(vp, vt, vc2);
240     vp = _mm256_fmadd_ps(vp, vt, vc1);
241 
242     // Reconstruct the final f value:
243     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
244     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
245     //     = s + (t * s) * p
246     vt = _mm256_mul_ps(vt, vs);
247     __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
248 
249     // For inputs below zero cutoff, replace output with +0.0f.
250     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
251     vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
252 
253     // Accumulate computed exponents.
254     vacc = _mm256_add_ps(vacc, vf);
255   }
256   if (elements != 0) {
257     assert(elements >= 1 * sizeof(float));
258     assert(elements <= 7 * sizeof(float));
259     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
260 
261     // Load up to 7 inputs at a time.
262     const __m256 vi = _mm256_maskload_ps(input, vmask);
263 
264     // Subtract maximum input x := i - i_max. This implies x <= 0.
265     const __m256 vx = _mm256_sub_ps(vi, vi_max);
266 
267     // Compute reduced argument elements := round(x / log(2)).
268     __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
269 
270     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
271     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
272     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
273 
274     // Subtract the large number back to get final elements := round(x / log(2)).
275     vn = _mm256_sub_ps(vn, vmagic_bias);
276 
277     // Compute reduced argument t := x - elements * log(2).
278     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
279     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
280     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
281 
282     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
283     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
284     vp = _mm256_fmadd_ps(vp, vt, vc3);
285     vp = _mm256_fmadd_ps(vp, vt, vc2);
286     vp = _mm256_fmadd_ps(vp, vt, vc1);
287 
288     // Reconstruct the final f value:
289     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
290     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
291     //     = s + (t * s) * p
292     vt = _mm256_mul_ps(vt, vs);
293     __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
294 
295     // For inputs below zero cutoff, replace output with +0.0f.
296     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
297     vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
298 
299     // Accumulate computed exponents. And addend with mask to leave unmasked 32-bit lanes unchanged.
300     vacc = _mm256_add_ps(vacc, _mm256_and_ps(vf, _mm256_castsi256_ps(vmask)));
301   }
302   // Reduce 8 elements in the SIMD register
303   __m128 vacc_lo = _mm_add_ps(_mm256_castps256_ps128(vacc), _mm256_extractf128_ps(vacc, 1));
304   vacc_lo = _mm_add_ps(vacc_lo, _mm_movehl_ps(vacc_lo, vacc_lo));
305   vacc_lo = _mm_add_ss(vacc_lo, _mm_movehdup_ps(vacc_lo));
306   _mm_store_ss(sum, vacc_lo);
307   _mm256_zeroupper();
308 }
309