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 <stddef.h>
8
9 #include <arm_neon.h>
10
11 #include <xnnpack/math-stubs.h>
12
13
xnn_math_f32_sigmoid__neonfma_rr1_p5_div(size_t n,const float * input,float * output)14 void xnn_math_f32_sigmoid__neonfma_rr1_p5_div(
15 size_t n,
16 const float* input,
17 float* output)
18 {
19 assert(n % (4 * sizeof(float)) == 0);
20
21 // Large number such that ulp(magic bias) == 1 and magic bias === 127 mod 2**22.
22 const float32x4_t vmagic_bias = vmovq_n_f32(0x1.8000FEp23f);
23 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f);
24 const float32x4_t vln2 = vmovq_n_f32(0x1.62E43p-1f);
25 // Coefficient of polynomial approximation of
26 // exp(-t) ~ 1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) on [-log(2)/2, log(2)/2]
27 const float32x4_t vc5 = vmovq_n_f32(-0x1.0F9F9Cp-7f);
28 const float32x4_t vc4 = vmovq_n_f32(0x1.573A1Ap-5f);
29 const float32x4_t vc3 = vmovq_n_f32(-0x1.555A80p-3f);
30 const float32x4_t vc2 = vmovq_n_f32(0x1.FFFDC6p-2f);
31 const float32x4_t vc1 = vmovq_n_f32(-0x1.FFFFF6p-1f);
32 const float32x4_t vone = vmovq_n_f32(1.0f);
33 // The largest z for which sigmoidf(-z) is normalized.
34 // This number is also the largest z for which expf(-z) is normalized.
35 const float32x4_t vdenorm_cutoff = vmovq_n_f32(-0x1.5D589Ep+6f);
36
37 for (; n != 0; n -= 4 * sizeof(float)) {
38 const float32x4_t vx = vld1q_f32(input); input += 4;
39
40 // General structure of the algorithm:
41 //
42 // / exp(x) / (1 + exp(x)) if x <= 0
43 // f[x] :=
44 // \ 1 - f[-x] if x >= 0
45 //
46 // First we compute f[-z] := exp(-z) / (1 + exp(-z)) where z = abs(x),
47 // then replace result with 1 - f[-z] if x >= 0.
48 const float32x4_t vz = vabsq_f32(vx);
49
50 // Compute reduced argument n := round(-z / log(2)).
51 // We do it by adding a large number (magic bias), which cause rounding of the result to integer, then subtracing
52 // the large number back. The trick with adding large number is valid only within certain bounds
53 // (|-z / log(2)| <= 2**22, i.e. |z| <= 0x1.62E43p+22 = 5814540.0), but that is acceptable, because inputs x
54 // outside of [-87.336544, 17.328678] (i.e. z outsize [0, 87.336544]) underflow or saturate sigmoidf(x). We fixup
55 // the result for such inputs at the very end of the algorithm.
56 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
57
58 // Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
59 // -87.336544 <= -z <= 0.0, and -126 <= n <= 0 accordingly.
60 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
61
62 // Subtract the large number back to get the final n := round(-z / log(2)) as a floating-point number.
63 vn = vsubq_f32(vn, vmagic_bias);
64
65 // Compute reduced argument t := z + n * log(2). Note that -t = -z - n * log(2).
66 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
67
68 // Compute degree-5 polynomial approximation for exp(-t) on [-log(2)/2, log(2)/2]:
69 // P(t) = 1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) = 1 + t * p
70 float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
71 vp = vfmaq_f32(vc3, vp, vt);
72 vp = vfmaq_f32(vc2, vp, vt);
73 vp = vfmaq_f32(vc1, vp, vt);
74
75 // Reconstruct the exp(-z) value:
76 // e = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
77 // = s * (1 + t * p)
78 // = s + (t * s) * p
79 vt = vmulq_f32(vt, vs);
80 float32x4_t ve = vfmaq_f32(vs, vp, vt);
81
82 // Denominator of the sigmoid fraction: 1.0 + exp(-z)
83 float32x4_t vd = vaddq_f32(ve, vone);
84
85 // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
86 float32x4_t vf = vdivq_f32(ve, vd);
87
88 // For inputs below denormal cutoff, replace output with +0.0f.
89 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
90 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
91
92 // Reconstruct sigmoid(x) = x < 0 ? sigmoid(-z) : 1.0 - sigmoid(-z)
93 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
94 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
95
96 vst1q_f32(output, vf); output += 4;
97 }
98 }
99