• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 / 2048) values, k = 0..2047
16 extern XNN_INTERNAL const float xnn_table_exp2_k_over_2048[2048];
17 
xnn_math_f32_sigmoid__neonfma_rr2_lut2048_p1_div(size_t n,const float * input,float * output)18 void xnn_math_f32_sigmoid__neonfma_rr2_lut2048_p1_div(
19     size_t n,
20     const float* input,
21     float* output)
22 {
23   assert(n % (4 * sizeof(float)) == 0);
24 
25   const float32x4_t vmagic_bias = vmovq_n_f32(0x1.800000p23f);
26   // The largest z for which sigmoidf(-z) is normalized.
27   // This number is also the largest z for which expf(-z) is normalized.
28   const float32x4_t vdenorm_cutoff = vmovq_n_f32(-0x1.5D589Ep+6f);
29   const float32x4_t vminus_log2e_x2048  = vmovq_n_f32(-0x1.715476p11f);
30   const float32x4_t vln2_o2048_hi = vmovq_n_f32(0x1.62E43p-12f);
31   const float32x4_t vln2_o2048_lo = vmovq_n_f32(-0x1.05C61p-40f);
32   const float32x4_t vone = vmovq_n_f32(1.0f);
33 
34   const float32x4_t vc1 = vmovq_n_f32(-0x1.FFFFFEp-1f);
35 
36   const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x7FF));
37 
38   for (; n != 0; n -= 4 * sizeof(float)) {
39     const float32x4_t vx = vld1q_f32(input); input += 4;
40 
41     // General structure of the algorithm:
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 * 2048 / log(2)).
51     // We do it by adding a large number (magic bias), which cause rounding of the result to an integer, then subtracing
52     // the large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
53     // The trick with adding large number is valid only within certain bounds (|z * 2048 / log(2)| <= 2**22, i.e.
54     // |z| <= 0x1.62E43p+10 = 1419.5654296875), but that is acceptable, because inputs x outside of
55     // [-87.336544, 17.328678] (i.e. z outsize [0, 87.336544]) underflow or saturate sigmoidf(x). We fixup the result
56     // for such inputs at the very end of the algorithm.
57     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e_x2048);
58 
59     // Create a floating-point number s (scale) such that s := 2**(n / 2048) for such inputs that sigmoidf(-z) is
60     // normalized, i.e. 0 <= z <= 87.33642. As n has 11 fractional bits, we split s == 2**(n / 2048) =
61     // = 2**e * 2**(n / 2048 - e), where e := int(n / 2048). We create s in two steps:
62     // 1. Fetch 2**(n / 2048 - e) = 2**(n % 2048) from the table using the 11 low bits of n, as integer. Note that the
63     //    fetched values are in the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
64     // 2. Adjust fecthed value by addition of e to its floating-point exponent. The result is always a normalized
65     //    number, because for 0 <= z <= 87.33642 (inputs for which sigmoidf(-z) is normalized) we have -126 <= e <= 0,
66     //    and thus the adjusted exponent is not lower than -126.
67     //
68     // Extract e from bits 11:19 of n and shift it into bits 23:31 (position of floating-point exponent).
69     const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x7FF))), 12);
70 
71     // Use bits 0:11 bits of n, as integer, as an index for table lookup of l := 2**(n % 2048).
72     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
73     const uint64_t vidx01 = vgetq_lane_u64(vidx, 0);
74     const uint64_t vidx23 = vgetq_lane_u64(vidx, 1);
75     float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2_k_over_2048[(uint32_t) vidx01]);
76     float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2_k_over_2048[(uint32_t) vidx23]);
77     vl01 = vld1_lane_f32(&xnn_table_exp2_k_over_2048[(uint32_t) (vidx01 >> 32)], vl01, 1);
78     vl23 = vld1_lane_f32(&xnn_table_exp2_k_over_2048[(uint32_t) (vidx23 >> 32)], vl23, 1);
79     const float32x4_t vl = vcombine_f32(vl01, vl23);
80     // Adjust exponent of the value l fetched from the table to get the final s value.
81     const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
82 
83     // Subtract the large number back to get the final n := round(-z * 2048 / log(2)) as a floating-point number.
84     vn = vsubq_f32(vn, vmagic_bias);
85 
86     // Compute reduced argument t := (z + n * log(2) / 2048). Note that -t = -z - n * log(2) / 2048.
87     // Use Cody-Waite range reduction method (note two constants to represent log(2) / 2048) to improve accuracy.
88     float32x4_t vt = vfmaq_f32(vz, vn, vln2_o2048_hi);
89     vt = vfmaq_f32(vt, vn, vln2_o2048_lo);
90 
91     // Compute degree-1 polynomial approximation for exp(-t) on [-log(2)/4096, log(2)/4096]:
92     //   P1(t) = 1 + t * c1
93     const float32x4_t vp = vmulq_f32(vt, vc1);
94 
95     // Reconstruct the exp(-z) value:
96     //   y = s * (1 + t * c1)
97     //     = s + s * (t * c1))
98     //     = s + s * p
99     const float32x4_t vy = vfmaq_f32(vs, vs, vp);
100 
101     // Denominator of the sigmoid fraction: 1.0 + exp(-z)
102     const float32x4_t vd = vaddq_f32(vy, vone);
103 
104     // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
105     float32x4_t vf = vdivq_f32(vy, vd);
106 
107     // For inputs below denormal cutoff, replace output with +0.0f.
108     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
109     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
110 
111     // Reconstruct sigmoid(x) = x < 0 ? sigmoid(-z) : 1.0 - sigmoid(-z)
112     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
113     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
114 
115     vst1q_f32(output, vf); output += 4;
116   }
117 }
118