• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5 
6 #include <assert.h>
7 
8 #include <immintrin.h>
9 
10 #include <xnnpack/math-stubs.h>
11 
12 
xnn_math_f32_exp__avx512f_p5_scalef(size_t n,const float * input,float * output)13 void xnn_math_f32_exp__avx512f_p5_scalef(
14     size_t n,
15     const float* input,
16     float* output)
17 {
18   assert(n % (16 * sizeof(float)) == 0);
19 
20   const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
21   const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);
22   const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);
23 
24   const __m512 vc0 = _mm512_set1_ps(1.0f);
25   const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f);
26   const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f);
27   const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f);
28   const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f);
29   const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f);
30 
31   for (; n != 0; n -= 16 * sizeof(float)) {
32     const __m512 vx = _mm512_loadu_ps(input);
33 
34     // Compute reduced argument n := round(x / log(2)).
35     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
36     // Compute reduced argument t := x - n * log(2).
37     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
38     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
39     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
40 
41     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
42     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
43     vp = _mm512_fmadd_ps(vp, vt, vc3);
44     vp = _mm512_fmadd_ps(vp, vt, vc2);
45     vp = _mm512_fmadd_ps(vp, vt, vc1);
46     vp = _mm512_fmadd_ps(vp, vt, vc0);
47 
48     // Reconstruct the final value as f = exp2(n) * p.
49     const __m512 vf = _mm512_scalef_ps(vp, vn);
50     _mm512_storeu_ps(output, vf);
51 
52     input += 16;
53     output += 16;
54   }
55 }
56