xref: /aosp_15_r20/external/XNNPACK/src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.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_x64_acc4(size_t elements,const float * input,float * sum,float max)19 void xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc4(
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   __m256 vacc1 = _mm256_setzero_ps();
44   __m256 vacc2 = _mm256_setzero_ps();
45   __m256 vacc3 = _mm256_setzero_ps();
46   for (; elements >= 64 * sizeof(float); elements -= 64 * sizeof(float)) {
47     // Load 64 (8x8) inputs at a time.
48     const __m256 vi0 = _mm256_loadu_ps(input);
49     const __m256 vi1 = _mm256_loadu_ps(input + 8);
50     const __m256 vi2 = _mm256_loadu_ps(input + 16);
51     const __m256 vi3 = _mm256_loadu_ps(input + 24);
52     const __m256 vi4 = _mm256_loadu_ps(input + 32);
53     const __m256 vi5 = _mm256_loadu_ps(input + 40);
54     const __m256 vi6 = _mm256_loadu_ps(input + 48);
55     const __m256 vi7 = _mm256_loadu_ps(input + 56);
56     input += 64;
57 
58     // Subtract maximum input x := i - i_max. This implies x <= 0.
59     const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
60     const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
61     const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
62     const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
63     const __m256 vx4 = _mm256_sub_ps(vi4, vi_max);
64     const __m256 vx5 = _mm256_sub_ps(vi5, vi_max);
65     const __m256 vx6 = _mm256_sub_ps(vi6, vi_max);
66     const __m256 vx7 = _mm256_sub_ps(vi7, vi_max);
67 
68     // Compute reduced argument elements := round(x / log(2)).
69     __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
70     __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
71     __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
72     __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
73     __m256 vn4 = _mm256_fmadd_ps(vx4, vlog2e, vmagic_bias);
74     __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias);
75     __m256 vn6 = _mm256_fmadd_ps(vx6, vlog2e, vmagic_bias);
76     __m256 vn7 = _mm256_fmadd_ps(vx7, 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 
89     // Subtract the large number back to get final elements := round(x / log(2)).
90     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
91     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
92     vn2 = _mm256_sub_ps(vn2, vmagic_bias);
93     vn3 = _mm256_sub_ps(vn3, vmagic_bias);
94     vn4 = _mm256_sub_ps(vn4, vmagic_bias);
95     vn5 = _mm256_sub_ps(vn5, vmagic_bias);
96     vn6 = _mm256_sub_ps(vn6, vmagic_bias);
97     vn7 = _mm256_sub_ps(vn7, vmagic_bias);
98 
99     // Compute reduced argument t := x - elements * log(2).
100     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
101     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
102     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
103     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
104     __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
105     __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
106     __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
107     __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
108     __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
109 
110     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
111     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
112     vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
113     vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
114     vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
115     vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
116     vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
117     vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
118 
119     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
120     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
121     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
122     __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
123     __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
124     __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
125     __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
126     __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
127     __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
128 
129     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
130     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
131     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
132     vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
133     vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
134     vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
135     vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
136     vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
137 
138     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
139     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
140     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
141     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
142     vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
143     vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
144     vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
145     vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
146 
147     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
148     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
149     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
150     vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
151     vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
152     vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
153     vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
154     vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
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     vt0 = _mm256_mul_ps(vt0, vs0);
161     vt1 = _mm256_mul_ps(vt1, vs1);
162     vt2 = _mm256_mul_ps(vt2, vs2);
163     vt3 = _mm256_mul_ps(vt3, vs3);
164     vt4 = _mm256_mul_ps(vt4, vs4);
165     vt5 = _mm256_mul_ps(vt5, vs5);
166     vt6 = _mm256_mul_ps(vt6, vs6);
167     vt7 = _mm256_mul_ps(vt7, vs7);
168 
169     __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
170     __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
171     __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
172     __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
173     __m256 vf4 = _mm256_fmadd_ps(vt4, vp4, vs4);
174     __m256 vf5 = _mm256_fmadd_ps(vt5, vp5, vs5);
175     __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6);
176     __m256 vf7 = _mm256_fmadd_ps(vt7, vp7, vs7);
177 
178     // For inputs below zero cutoff, replace output with +0.0f.
179     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
180     vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
181     vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
182     vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
183     vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
184     vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vx4, vdenorm_cutoff, _CMP_LT_OS), vf4);
185     vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vx5, vdenorm_cutoff, _CMP_LT_OS), vf5);
186     vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vx6, vdenorm_cutoff, _CMP_LT_OS), vf6);
187     vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7);
188 
189     // Accumulate computed exponents.
190     vacc0 = _mm256_add_ps(vacc0, vf0);
191     vacc1 = _mm256_add_ps(vacc1, vf1);
192     vacc2 = _mm256_add_ps(vacc2, vf2);
193     vacc3 = _mm256_add_ps(vacc3, vf3);
194     vacc0 = _mm256_add_ps(vacc0, vf4);
195     vacc1 = _mm256_add_ps(vacc1, vf5);
196     vacc2 = _mm256_add_ps(vacc2, vf6);
197     vacc3 = _mm256_add_ps(vacc3, vf7);
198   }
199   // Add up all accumulators to vacc0
200   vacc0 = _mm256_add_ps(vacc0, vacc1);
201   vacc2 = _mm256_add_ps(vacc2, vacc3);
202   vacc0 = _mm256_add_ps(vacc0, vacc2);
203 
204   __m256 vacc = vacc0;
205   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
206     // Load 8 inputs at a time.
207     const __m256 vi = _mm256_loadu_ps(input);
208     input += 8;
209 
210     // Subtract maximum input x := i - i_max. This implies x <= 0.
211     const __m256 vx = _mm256_sub_ps(vi, vi_max);
212 
213     // Compute reduced argument elements := round(x / log(2)).
214     __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
215 
216     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
217     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
218     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
219 
220     // Subtract the large number back to get final elements := round(x / log(2)).
221     vn = _mm256_sub_ps(vn, vmagic_bias);
222 
223     // Compute reduced argument t := x - elements * log(2).
224     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
225     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
226     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
227 
228     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
229     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
230     vp = _mm256_fmadd_ps(vp, vt, vc3);
231     vp = _mm256_fmadd_ps(vp, vt, vc2);
232     vp = _mm256_fmadd_ps(vp, vt, vc1);
233 
234     // Reconstruct the final f value:
235     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
236     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
237     //     = s + (t * s) * p
238     vt = _mm256_mul_ps(vt, vs);
239     __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
240 
241     // For inputs below zero cutoff, replace output with +0.0f.
242     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
243     vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
244 
245     // Accumulate computed exponents.
246     vacc = _mm256_add_ps(vacc, vf);
247   }
248   if (elements != 0) {
249     assert(elements >= 1 * sizeof(float));
250     assert(elements <= 7 * sizeof(float));
251     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
252 
253     // Load up to 7 inputs at a time.
254     const __m256 vi = _mm256_maskload_ps(input, vmask);
255 
256     // Subtract maximum input x := i - i_max. This implies x <= 0.
257     const __m256 vx = _mm256_sub_ps(vi, vi_max);
258 
259     // Compute reduced argument elements := round(x / log(2)).
260     __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
261 
262     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
263     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
264     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
265 
266     // Subtract the large number back to get final elements := round(x / log(2)).
267     vn = _mm256_sub_ps(vn, vmagic_bias);
268 
269     // Compute reduced argument t := x - elements * log(2).
270     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
271     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
272     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
273 
274     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
275     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
276     vp = _mm256_fmadd_ps(vp, vt, vc3);
277     vp = _mm256_fmadd_ps(vp, vt, vc2);
278     vp = _mm256_fmadd_ps(vp, vt, vc1);
279 
280     // Reconstruct the final f value:
281     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
282     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
283     //     = s + (t * s) * p
284     vt = _mm256_mul_ps(vt, vs);
285     __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
286 
287     // For inputs below zero cutoff, replace output with +0.0f.
288     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
289     vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
290 
291     // Accumulate computed exponents. And addend with mask to leave unmasked 32-bit lanes unchanged.
292     vacc = _mm256_add_ps(vacc, _mm256_and_ps(vf, _mm256_castsi256_ps(vmask)));
293   }
294   // Reduce 8 elements in the SIMD register
295   __m128 vacc_lo = _mm_add_ps(_mm256_castps256_ps128(vacc), _mm256_extractf128_ps(vacc, 1));
296   vacc_lo = _mm_add_ps(vacc_lo, _mm_movehl_ps(vacc_lo, vacc_lo));
297   vacc_lo = _mm_add_ss(vacc_lo, _mm_movehdup_ps(vacc_lo));
298   _mm_store_ss(sum, vacc_lo);
299   _mm256_zeroupper();
300 }
301