• 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 <math.h>
8 #include <stddef.h>
9 
10 #include <arm_neon.h>
11 
12 #include <xnnpack/math-stubs.h>
13 
14 
15 // Table of exp2(k / 64) values, k = 0..63
16 static const float exp2_table[64] = {
17   0x1.000000p+0f, 0x1.02C9A4p+0f, 0x1.059B0Ep+0f, 0x1.087452p+0f,
18   0x1.0B5586p+0f, 0x1.0E3EC4p+0f, 0x1.11301Ep+0f, 0x1.1429AAp+0f,
19   0x1.172B84p+0f, 0x1.1A35BEp+0f, 0x1.1D4874p+0f, 0x1.2063B8p+0f,
20   0x1.2387A6p+0f, 0x1.26B456p+0f, 0x1.29E9E0p+0f, 0x1.2D285Ap+0f,
21   0x1.306FE0p+0f, 0x1.33C08Cp+0f, 0x1.371A74p+0f, 0x1.3A7DB4p+0f,
22   0x1.3DEA64p+0f, 0x1.4160A2p+0f, 0x1.44E086p+0f, 0x1.486A2Cp+0f,
23   0x1.4BFDAEp+0f, 0x1.4F9B28p+0f, 0x1.5342B6p+0f, 0x1.56F474p+0f,
24   0x1.5AB07Ep+0f, 0x1.5E76F2p+0f, 0x1.6247ECp+0f, 0x1.662388p+0f,
25   0x1.6A09E6p+0f, 0x1.6DFB24p+0f, 0x1.71F75Ep+0f, 0x1.75FEB6p+0f,
26   0x1.7A1148p+0f, 0x1.7E2F34p+0f, 0x1.82589Ap+0f, 0x1.868D9Ap+0f,
27   0x1.8ACE54p+0f, 0x1.8F1AEAp+0f, 0x1.93737Cp+0f, 0x1.97D82Ap+0f,
28   0x1.9C4918p+0f, 0x1.A0C668p+0f, 0x1.A5503Cp+0f, 0x1.A9E6B6p+0f,
29   0x1.AE89FAp+0f, 0x1.B33A2Cp+0f, 0x1.B7F770p+0f, 0x1.BCC1EAp+0f,
30   0x1.C199BEp+0f, 0x1.C67F12p+0f, 0x1.CB720Ep+0f, 0x1.D072D4p+0f,
31   0x1.D5818Ep+0f, 0x1.DA9E60p+0f, 0x1.DFC974p+0f, 0x1.E502EEp+0f,
32   0x1.EA4AFAp+0f, 0x1.EFA1BEp+0f, 0x1.F50766p+0f, 0x1.FA7C18p+0f,
33 };
34 
xnn_math_f32_exp__neonfma_lut64_p2(size_t n,const float * input,float * output)35 void xnn_math_f32_exp__neonfma_lut64_p2(
36     size_t n,
37     const float* input,
38     float* output)
39 {
40   assert(n % (4 * sizeof(float)) == 0);
41 
42   const float32x4_t vmagic_bias = vmovq_n_f32(0x1.800000p23f);
43   // The smallest x for which expf(x) is non-zero.
44   const float32x4_t vzero_cutoff = vmovq_n_f32(-0x1.9FE368p6f);
45   // The largest x for which expf(x) is finite.
46   const float32x4_t vinf_cutoff = vmovq_n_f32(0x1.62E42Ep6f);
47   const float32x4_t vlog2e_x64  = vmovq_n_f32(0x1.715476p6f);
48   const float32x4_t vminus_ln2_o64_hi = vmovq_n_f32(-0x1.62e43p-7f);
49   const float32x4_t vminus_ln2_o64_lo = vmovq_n_f32(0x1.05c61p-35f);
50   const float32x4_t vplus_inf = vmovq_n_f32(INFINITY);
51 
52   const float32x4_t vc2 = vmovq_n_f32(0x1.FFFF0Ap-2f);
53 
54   const int32x4_t vmin_exponent = vmovq_n_s32(INT32_C(0xC1000000));
55   const int32x4_t vmax_exponent = vmovq_n_s32(INT32_C(0x3F800000));
56   const int32x4_t vdefault_exponent = vmax_exponent;
57   const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x3F));
58 
59   for (; n != 0; n -= 4 * sizeof(float)) {
60     const float32x4_t vx = vld1q_f32(input); input += 4;
61 
62     // Compute reduced argument n := round(x * 64 / log(2)).
63     // We do it by adding a large number (magic bias), which cause rounding of result to an integer, then subtracing the
64     // large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
65     // The trick with adding large number is valid only within certain bounds (|x| <= 2**22), but thats ok, because
66     // inputs outside of [-103.97207, 88.72283] underflow or overflow expf(x) anyway. We fixup the result for such
67     // inputs at the very end of the algorithm.
68     float32x4_t vn = vfmaq_f32(vmagic_bias, vx, vlog2e_x64);
69 
70     // Create two floating-point numbers, sn (scale, normal) and so (scale, overflow) such that sn * so == 2**n
71     // for inputs which don't cause overflow, i.e. -103.97207 <= x <= 88.72283, and -150 <= n <= 128 accordingly.
72     // We need to use two numbers rather than one because a normalized single-precision exponent must be in [-127, 126]
73     // range, which is insufficient to cover [-150, 128] range of n.
74     // - When n is within [-127, 126], sn == 2**n and so == 1.0.
75     // - When n < -127, sn == 2**(-127) and so == 2**(n + 127).
76     // - When n > 126, sn == 2**126 and so == 2**(n - 126).
77     // While we explicitly compute sn, the so is fused into the value l fetched from a table by adjusting its exponential.
78     int32x4_t veo = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))), 17);
79     int32x4_t ven = vmaxq_s32(veo, vmin_exponent);
80     ven = vminq_s32(ven, vmax_exponent);
81     veo = vsubq_s32(veo, ven);
82     const float32x4_t vsn = vreinterpretq_f32_s32(vaddq_s32(ven, vdefault_exponent));
83 
84     // Use the low 6 bits of n (as integer) for table lookup.
85     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
86     const uint64_t vidx01 = vgetq_lane_u64(vidx, 0);
87     const uint64_t vidx23 = vgetq_lane_u64(vidx, 1);
88     float32x2_t vl01 = vld1_dup_f32(&exp2_table[(uint32_t) vidx01]);
89     float32x2_t vl23 = vld1_dup_f32(&exp2_table[(uint32_t) vidx23]);
90     vl01 = vld1_lane_f32(&exp2_table[(uint32_t) (vidx01 >> 32)], vl01, 1);
91     vl23 = vld1_lane_f32(&exp2_table[(uint32_t) (vidx23 >> 32)], vl23, 1);
92     float32x4_t vl = vcombine_f32(vl01, vl23);
93     // Fuse so into the value l fetched from a table by adjusting its exponential.
94     vl = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), veo));
95 
96     // Subtract the large number back to get final n := round(x * 64 / log(2)).
97     vn = vsubq_f32(vn, vmagic_bias);
98 
99     // Compute reduced argument t := x - n * log(2) / 64.
100     // Use Cody-Waite range reduction method (note two constants to represent log(2) / 64) to improve accuracy.
101     float32x4_t vt = vfmaq_f32(vx, vn, vminus_ln2_o64_hi);
102     vt = vfmaq_f32(vt, vn, vminus_ln2_o64_lo);
103 
104     // Compute degree-2 polynomial approxiatmion for exp(t) on [-log(2)/128, log(2)/128].
105     float32x4_t vp = vmulq_f32(vt, vc2);
106     vp = vfmaq_f32(vt, vt, vp);
107 
108     // Reconstruct the final f value:
109     //   f = sn * (so * l) * (1 + t * (1 + t * c2))
110     //     = sn * (so * l) * (1 + t + t * (t * c2))
111     //     = sn * ((so * l) + (so * l) * (t + t * (t * c2)))
112     float32x4_t vf = vfmaq_f32(vl, vl, vp);
113     vf = vmulq_f32(vf, vsn);
114 
115     // For inputs below zero cutoff, replace output with +0.0f.
116     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
117     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vzero_cutoff)));
118     // For inputs above inf cutoff, replace output with +inf.
119     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
120     vf = vbslq_f32(vcgtq_f32(vx, vinf_cutoff), vplus_inf, vf);
121     vst1q_f32(output, vf); output += 4;
122   }
123 }
124