xref: /aosp_15_r20/external/XNNPACK/src/math/expminus-f16-avx2-rr1-p2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Copyright 2022 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_f16_expminus__avx2_rr1_p2(size_t n,const void * input,void * output)14 void xnn_math_f16_expminus__avx2_rr1_p2(
15     size_t n,
16     const void* input,
17     void* output)
18 {
19   assert(n % (8 * sizeof(uint16_t)) == 0);
20 
21   // Large number such that ulp(magic bias) == 1 and magic bias === 127 mod 2**22.
22   const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
23   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p0f);
24   const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
25   // Coefficient of polynomial approximation of
26   // exp(t) ~ 1 + t * (c1 + t * c2) on [-log(2)/2, log(2)/2]
27   const __m256 vc2 = _mm256_set1_ps(0x1.FF3A32p-2f);
28   const __m256 vc1 = _mm256_set1_ps(0x1.039E10p+0f);
29   // The smallest x for which exph(x) is normalized.
30   const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.368000p+3f);
31 
32   const uint16_t* i = (const uint16_t*) input;
33   uint16_t* o = (uint16_t*) output;
34   for (; n != 0; n -= 8 * sizeof(float)) {
35     const __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
36     i += 8;
37 
38     // Compute reduced argument n := round(x / log(2)).
39     // We do it by adding a large number (magic bias) to the product x * (1/log(2)), which cause rounding of the
40     // result to an integer, then subtracing the large number back. The first addition is combined with multiplication
41     // by log2e into a single FMA instruction. The trick with adding large number is valid only within certain bounds
42     // (|x / log(2)| <= 2**9, i.e. |x| <= 0x1.630p+8 = 355.0), but that is acceptable, because inputs x outside
43     // of [-9.703125, 0] underflow expf(x). We fixup the result for such inputs at the very end of the algorithm.
44     __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
45 
46     // Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
47     // -9.703125 <= x <= 0.0, and -14 <= n <= 0 accordingly.
48     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
49 
50     // Subtract the large number back to get final n := round(x / log(2)) as a floating-point number.
51     vn = _mm256_sub_ps(vn, vmagic_bias);
52 
53     // Compute reduced argument t := x - n * log(2).
54     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vx);
55 
56     // Compute degree-2 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]:
57     //   P(t) = 1 + t * (c1 + t * c2) = 1 + t * p
58     const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
59 
60     // Reconstruct the exp(x) value:
61     //   exp(x) = s * (1 + t * (c1 + t * c2))
62     //          = s + (t * s) * (c1 + t * c2)
63     //          = s + (t * s) * p
64     vt = _mm256_mul_ps(vt, vs);
65     __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
66 
67     // For inputs below denormal cutoff, replace output with +0.0f.
68     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
69     vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
70     _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC));
71     o += 8;
72   }
73 }
74