• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-raddexpminusmax/avx512f-p5-scalef.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/intrinsics-polyfill.h>
15 #include <xnnpack/raddexpminusmax.h>
16 
17 
xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128(size_t elements,const float * input,float * sum,float max)18 void xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128(
19     size_t elements,
20     const float* input,
21     float* sum,
22     float max)
23 {
24   assert(elements % sizeof(float) == 0);
25 
26   const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
27   const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);
28   const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);
29 
30   const __m512 vc0 = _mm512_set1_ps(1.0f);
31   const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f);
32   const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f);
33   const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f);
34   const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f);
35   const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f);
36 
37   const __m512 vi_max = _mm512_set1_ps(max);
38 
39   __m512 vacc0 = _mm512_setzero_ps();
40   for (; elements >= 128 * sizeof(float); elements -= 128 * sizeof(float)) {
41     // Load 128 (8x16) inputs at a time.
42     const __m512 vi0 = _mm512_loadu_ps(input);
43     const __m512 vi1 = _mm512_loadu_ps(input + 16);
44     const __m512 vi2 = _mm512_loadu_ps(input + 32);
45     const __m512 vi3 = _mm512_loadu_ps(input + 48);
46     const __m512 vi4 = _mm512_loadu_ps(input + 64);
47     const __m512 vi5 = _mm512_loadu_ps(input + 80);
48     const __m512 vi6 = _mm512_loadu_ps(input + 96);
49     const __m512 vi7 = _mm512_loadu_ps(input + 112);
50     input += 128;
51 
52     // Subtract maximum input x := i - i_max.
53     const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
54     const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
55     const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
56     const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
57     const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
58     const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
59     const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
60     const __m512 vx7 = _mm512_sub_ps(vi7, vi_max);
61 
62     // Compute reduced argument elements := round(x / log(2)).
63     const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
64     const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
65     const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
66     const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
67     const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
68     const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
69     const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
70     const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
71 
72     // Compute reduced argument t := x - elements * log(2).
73     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
74     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
75     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
76     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
77     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
78     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
79     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
80     __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
81     __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
82 
83     vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
84     vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
85     vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
86     vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
87     vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
88     vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
89     vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
90     vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
91 
92     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
93     __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
94     __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
95     __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
96     __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
97     __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
98     __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
99     __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
100     __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
101 
102     vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
103     vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
104     vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
105     vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
106     vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
107     vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
108     vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
109     vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
110 
111     vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
112     vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
113     vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
114     vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
115     vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
116     vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
117     vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
118     vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
119 
120     vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
121     vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
122     vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
123     vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
124     vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
125     vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
126     vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
127     vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
128 
129     vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
130     vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
131     vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
132     vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
133     vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
134     vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
135     vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
136     vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
137 
138     // Reconstruct the final f value:
139     //   f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
140     //     = 2**elements * p
141     const __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
142     const __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
143     const __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
144     const __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
145     const __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
146     const __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
147     const __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
148     const __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
149 
150     // Accumulate computed exponents.
151     vacc0 = _mm512_add_ps(vacc0, vf0);
152     vacc0 = _mm512_add_ps(vacc0, vf1);
153     vacc0 = _mm512_add_ps(vacc0, vf2);
154     vacc0 = _mm512_add_ps(vacc0, vf3);
155     vacc0 = _mm512_add_ps(vacc0, vf4);
156     vacc0 = _mm512_add_ps(vacc0, vf5);
157     vacc0 = _mm512_add_ps(vacc0, vf6);
158     vacc0 = _mm512_add_ps(vacc0, vf7);
159   }
160 
161   __m512 vacc = vacc0;
162   for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
163     // Load 16 inputs at a time.
164     const __m512 vi = _mm512_loadu_ps(input);
165     input += 16;
166 
167     // Subtract maximum input x := i - i_max.
168     const __m512 vx = _mm512_sub_ps(vi, vi_max);
169 
170     // Compute reduced argument elements := round(x / log(2)).
171     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
172 
173     // Compute reduced argument t := x - elements * log(2).
174     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
175     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
176     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
177 
178     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
179     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
180     vp = _mm512_fmadd_ps(vp, vt, vc3);
181     vp = _mm512_fmadd_ps(vp, vt, vc2);
182     vp = _mm512_fmadd_ps(vp, vt, vc1);
183     vp = _mm512_fmadd_ps(vp, vt, vc0);
184 
185     // Reconstruct the final f value:
186     //   f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
187     //     = 2**elements * p
188     const __m512 vf = _mm512_scalef_ps(vp, vn);
189 
190     // Accumulate computed exponents.
191     vacc = _mm512_add_ps(vacc, vf);
192   }
193   if (elements != 0) {
194     // Prepare mask for valid 32-bit elements (depends on elements).
195     elements >>= 2 /* log2(sizeof(float)) */;
196     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
197 
198     // Load up to 15 inputs at a time.
199     const __m512 vi = _mm512_maskz_loadu_ps(vmask, input);
200 
201     // Subtract maximum input x := i - i_max.
202     const __m512 vx = _mm512_sub_ps(vi, vi_max);
203 
204     // Compute reduced argument elements := round(x / log(2)).
205     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
206 
207     // Compute reduced argument t := x - elements * log(2).
208     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
209     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
210     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
211 
212     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
213     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
214     vp = _mm512_fmadd_ps(vp, vt, vc3);
215     vp = _mm512_fmadd_ps(vp, vt, vc2);
216     vp = _mm512_fmadd_ps(vp, vt, vc1);
217     vp = _mm512_fmadd_ps(vp, vt, vc0);
218 
219     // Reconstruct the final f value:
220     //   f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
221     //     = 2**elements * p
222     const __m512 vf = _mm512_scalef_ps(vp, vn);
223 
224     // Accumulate computed exponents.
225     vacc = _mm512_mask_add_ps(vacc, vmask, vacc, vf);
226   }
227   *sum = _mm512_reduce_add_ps(vacc);
228 }
229