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 #include <math.h>
8
9 #include <immintrin.h>
10
11 #include <xnnpack/math-stubs.h>
12
13
xnn_math_f32_extexp__avx2_p5(size_t n,const float * input,float * output_mantissa,float * output_exponent)14 void xnn_math_f32_extexp__avx2_p5(
15 size_t n,
16 const float* input,
17 float* output_mantissa,
18 float* output_exponent)
19 {
20 assert(n % (8 * sizeof(float)) == 0);
21
22 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
23 const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
24 const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
25
26 const __m256 vc0 = _mm256_set1_ps(1.0f);
27 const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
28 const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
29 const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
30 const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
31 const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
32
33 for (; n != 0; n -= 8 * sizeof(float)) {
34 const __m256 vx = _mm256_loadu_ps(input);
35
36 // Compute reduced argument n := round(x / log(2)).
37 const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
38
39 // Compute reduced argument t := x - n * log(2).
40 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
41 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
42 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
43
44 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
45 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
46 vp = _mm256_fmadd_ps(vp, vt, vc3);
47 vp = _mm256_fmadd_ps(vp, vt, vc2);
48 vp = _mm256_fmadd_ps(vp, vt, vc1);
49 vp = _mm256_fmadd_ps(vp, vt, vc0);
50
51 // Skip reconstruction step and separately store "mantissa" and "exponent" of the output.
52 _mm256_storeu_ps(output_mantissa, vp);
53 _mm256_storeu_ps(output_exponent, vn);
54
55 input += 8;
56 output_mantissa += 8;
57 output_exponent += 8;
58 }
59 }
60