• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2020 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 <stddef.h>
8 
9 #include <arm_neon.h>
10 
11 #include <xnnpack/math-stubs.h>
12 
13 
xnn_math_f32_expm1minus__neonfma_rr1_p6(size_t n,const float * input,float * output)14 void xnn_math_f32_expm1minus__neonfma_rr1_p6(
15     size_t n,
16     const float* input,
17     float* output)
18 {
19   assert(n % (4 * sizeof(float)) == 0);
20 
21   // The largest x for which expm1f(x) is saturated at -1.0f.
22   const float32x4_t vsat_cutoff = vmovq_n_f32(-0x1.154246p+4f);
23   // Large number such that ulp(magic bias) == 1 and magic bias === 127 mod 2**22.
24   const float32x4_t vmagic_bias = vmovq_n_f32(0x1.8000FEp23f);
25   const float32x4_t vlog2e = vmovq_n_f32(0x1.715476p+0f);
26   const float32x4_t vminus_ln2 = vmovq_n_f32(-0x1.62E430p-1f);
27   // Coefficient of polynomial approximation
28   //   exp(t) - 1 ~ t * (1 + t * (c2 + t * (c3 + t * (c4 + t * (c5 + t * c6)))))
29   // on [-log(2)/2, log(2)/2]
30   const float32x4_t vc6 = vmovq_n_f32(0x1.6b7338p-10f);
31   const float32x4_t vc5 = vmovq_n_f32(0x1.12278Ep-7f);
32   const float32x4_t vc4 = vmovq_n_f32(0x1.555716p-5f);
33   const float32x4_t vc3 = vmovq_n_f32(0x1.5554B0p-3f);
34   const float32x4_t vc2 = vmovq_n_f32(0x1.FFFFFEp-2f);
35   const float32x4_t vone = vmovq_n_f32(1.0f);
36 
37   for (; n != 0; n -= 4 * sizeof(float)) {
38     float32x4_t vx = vld1q_f32(input); input += 4;
39 
40     // The function saturates at -1 for large negative inputs: expm1f(x) == -1.0f for x <= sat_cutoff ~= -17.328680.
41     // To guarantee this behaviour, we clip input at sat_cutoff, and leverage the fact that for our implementation
42     // expm1f(sat_cutoff) == -1.0f. NaN inputs are passed unchanged.
43     vx = vmaxq_f32(vx, vsat_cutoff);
44 
45     // Compute reduced argument n := round(x / log(2)).
46     // We do it by adding a large number (magic bias), which cause rounding of the result to integer, then subtracing
47     // the large number back. The addition is combined with multiplication by log2e into a single FMA instruction. The
48     // trick with adding large number is valid only within certain bounds (|x / log(2)| <= 2**22, i.e.
49     // |x| <= 0x1.62E43p+21 = 2907270.0), but that is acceptable, because inputs x are restricted to [-17.328680, 0].
50     // Note that addition-subtraction of the large number doesn't cause overflow for inputs in this range.
51     float32x4_t vn = vfmaq_f32(vmagic_bias, vx, vlog2e);
52 
53     // Create a floating-point number s (scale) such that s == 2**n for valid inputs, i.e.
54     // -17.328680 <= x <= 0.0, and -25 <= n <= 0 accordingly.
55     // For NaN inputs, s would have zero mantissa and can have arbitrary sign and exponent, depending on the input
56     // NaN payload. In these cases, n and t are NaNs with the same payload as input while s is non-NaN, and thus
57     // input payload would be propagated in all computations.
58     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
59 
60     // Subtract the large number back to get final n := round(x / log(2)).
61     vn = vsubq_f32(vn, vmagic_bias);
62 
63     // Compute reduced argument t := x - n * log(2).
64     float32x4_t vt = vfmaq_f32(vx, vn, vminus_ln2);
65 
66     // Compute degree-6 polynomial approximation for exp(t) - 1 on [-log(2)/2, log(2)/2].
67     //   P(t) = t * (1 + t * (c2 + t * (c3 + t * (c4 + t * (c5 + t * c6)))))
68     //        = t + t * (t * (c2 + t * (c3 + t * (c4 + t * (c5 + t * c6))))) = t + t * p
69     float32x4_t vp = vfmaq_f32(vc5, vc6, vt);
70     vp = vfmaq_f32(vc4, vp, vt);
71     vp = vfmaq_f32(vc3, vp, vt);
72     vp = vfmaq_f32(vc2, vp, vt);
73     vp = vmulq_f32(vp, vt);
74 
75     // Reconstruct the exp(x) - 1 value:
76     //   exp(x) - 1 = s * (1 + t * (1 + t * (c2 + t * (c3 + t * (c4 + t * (c5 + t * c6)))))) - 1
77     //              = (s - 1) + s * (t + t * p)
78     //              = ((t * s) + (t * s) * p) + (s - 1)
79     vt = vmulq_f32(vt, vs);
80     const float32x4_t vsm1 = vsubq_f32(vs, vone);
81     vp = vfmaq_f32(vt, vp, vt);
82     const float32x4_t vf = vaddq_f32(vp, vsm1);
83 
84     vst1q_f32(output, vf); output += 4;
85   }
86 }
87