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/common.h>
12 #include <xnnpack/math-stubs.h>
13
14
15 // Table of exp2(k / 64) values decremented (as integer) by (k << 17), k = 0..63
16 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_64[64];
17
xnn_math_f32_sigmoid__neon_rr2_lut64_p2_nr2recps(size_t n,const float * input,float * output)18 void xnn_math_f32_sigmoid__neon_rr2_lut64_p2_nr2recps(
19 size_t n,
20 const float* input,
21 float* output)
22 {
23 assert(n % (4 * sizeof(float)) == 0);
24
25 // Large number such that ulp(magic bias) == exp2(-6)
26 const float32x4_t vmagic_bias = vmovq_n_f32(0x1.800000p17f);
27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p0f);
28 // Mask for the lowest 6 bits
29 const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x3F));
30 // Last 13 bits are zeroes
31 const float32x4_t vln2_hi = vmovq_n_f32(0x1.630000p-1f);
32 const float32x4_t vln2_lo = vmovq_n_f32(-0x1.BD0106p-13f);
33 // Coefficient of polynomial approximation of exp(-t) ~ 1 + t * (1 + t * c2) on [-log(2)/128, log(2)/128]
34 const float32x4_t vc2 = vmovq_n_f32(0x1.FFFF0Ap-2f);
35 const float32x4_t vone = vmovq_n_f32(1.0f);
36 // The largest z for which sigmoidf(-z) is normalized.
37 // This number is also the largest z for which expf(-z) is normalized.
38 const float32x4_t vdenorm_cutoff = vmovq_n_f32(-0x1.5D589Ep+6f);
39
40 for (; n != 0; n -= 4 * sizeof(float)) {
41 const float32x4_t vx = vld1q_f32(input); input += 4;
42
43 // General structure of the algorithm:
44 //
45 // / exp(x) / (1 + exp(x)) if x <= 0
46 // f[x] :=
47 // \ 1 - f[-x] if x >= 0
48 //
49 // First we compute f[-z] := exp(-z) / (1 + exp(-z)) where z = abs(x),
50 // then replace result with 1 - f[-z] if x >= 0.
51 const float32x4_t vz = vabsq_f32(vx);
52
53 // Compute reduced argument n := round(-z / log(2), 6).
54 // We do it by adding a large number (magic bias), which cause rounding of the result to integer, then subtracing
55 // the large number back. The trick with adding large number is valid only within certain bounds
56 // (|-z / log(2)| <= 2**16, i.e. |z| <= 0x1.62E43p+15 = 5814540.0), but that is acceptable, because inputs x
57 // outside of [-87.336544, 17.328678] (i.e. z outsize [0, 87.336544]) underflow or saturate sigmoidf(x). We fixup
58 // the result for such inputs at the very end of the algorithm.
59 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
60
61 // Create a floating-point number s (scale) such that s := 2**n for such inputs that sigmoidf(-z) is normalized,
62 // i.e. 0 <= z <= 87.33642. As n has 6 fractional bits, we split s == 2**n = 2**int(n) * 2**frac(n). We create s
63 // in two steps:
64 // 1. Fetch 2**frac(n) from the table using the 6 low bits of n, as integer. Note that the fetched values are in
65 // the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
66 // 2. Adjust fecthed value by addition of int(n) to its floating-point exponent. The result is always a normalized
67 // number, because for 0 <= z <= 87.33642 (inputs for which sigmoidf(z) is normalized) we have
68 // -126 <= int(n) <= 0, and thus the adjusted exponent is not lower than -126.
69 //
70 // Shift bits 6:14 into 23:31 (position of floating-point exponent).
71 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
72
73 // Use bits 0:6 of n, as integer, as an index for table lookup of l := 2**frac(n).
74 const uint64x2_t vidx = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask), 2));
75 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
76 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
77 float32x2_t vl_lo = vld1_dup_f32((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo));
78 float32x2_t vl_hi = vld1_dup_f32((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi));
79 vl_lo = vld1_lane_f32((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32)), vl_lo, 1);
80 vl_hi = vld1_lane_f32((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32)), vl_hi, 1);
81 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
82 // Adjust exponent of the value l fetched from the table to get the final s value.
83 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
84
85 // Subtract the large number back to get the final n := round(-z / log(2), 6) as a floating-point number.
86 vn = vsubq_f32(vn, vmagic_bias);
87
88 // Compute reduced argument t := (z + n * log(2)). Note that -t = -z - n * log(2).
89 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
90 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
91 vt = vmlaq_f32(vt, vn, vln2_lo);
92
93 // Compute degree-2 polynomial approximation for exp(-t) on [-log(2)/128, log(2)/128].
94 // P(t) = 1 + t * (-1 + t * c2) = 1 - (t - t * (t * c2)) = 1 - p
95 float32x4_t vp = vmulq_f32(vt, vc2);
96 vp = vmlsq_f32(vt, vp, vt);
97
98 // Reconstruct the exp(-z) value:
99 // e = s * (1 + t * (-1 + t * c2))
100 // = s * (1 - p)
101 // = s - s * p
102 const float32x4_t vy = vmlsq_f32(vs, vs, vp);
103
104 // Denominator of the sigmoid fraction: 1.0 + exp(-z)
105 const float32x4_t vd = vaddq_f32(vy, vone);
106
107 // Use Newton-Raphson method (2 iterations) to compute reciprocal of denominator.
108 // Note: 1 < d <= 2, because z >= 0.0 and 0 < exp(-z) <= 1.0.
109 // Thus the reciprocal of the denominator never overflows.
110 float32x4_t vr = vrecpeq_f32(vd);
111 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
112 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
113
114 // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
115 float32x4_t vf = vmulq_f32(vy, vr);
116
117 // For inputs below denormal cutoff, replace output with +0.0f.
118 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
119 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
120
121 // Reconstruct sigmoid(x) = x < 0 ? sigmoid(-z) : 1.0 - sigmoid(-z)
122 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
123 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
124
125 vst1q_f32(output, vf); output += 4;
126 }
127 }
128