• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-raddstoreexpminusmax/scalar-rr2-lut64-p2.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <xnnpack/common.h>
13 #include <xnnpack/raddstoreexpminusmax.h>
14 
15 #include <fp16/bitcasts.h>
16 
17 
18 // Note redefine as uint32[] to avoid redundant bitcasts.
19 extern XNN_INTERNAL const uint32_t xnn_table_exp2_k_over_64[64];
20 
xnn_f32_raddstoreexpminusmax_ukernel__scalar_rr2_lut64_p2_x1(size_t elements,const float * input,const float * max,float * output,float * sum,const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS (1)])21 void xnn_f32_raddstoreexpminusmax_ukernel__scalar_rr2_lut64_p2_x1(
22     size_t elements,
23     const float* input,
24     const float* max,
25     float* output,
26     float* sum,
27     const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29   assert(elements % sizeof(float) == 0);
30 
31   const float vi_max = *max;
32   const float vlog2e = params->scalar_rr2_lut64_p2.log2e;
33   const float vmagic_bias = params->scalar_rr2_lut64_p2.magic_bias;
34   const uint32_t vindex_mask = UINT32_C(0x3F);
35   const float vminus_ln2_hi = params->scalar_rr2_lut64_p2.minus_ln2_hi;
36   const float vminus_ln2_lo = params->scalar_rr2_lut64_p2.minus_ln2_lo;
37   const float vc2 = params->scalar_rr2_lut64_p2.c2;
38   const float vdenorm_cutoff = params->scalar_rr2_lut64_p2.denorm_cutoff;
39 
40   float vacc = 0.0f;
41   for (; elements >= sizeof(float); elements -= sizeof(float)) {
42     // Load 1 input at a time.
43     const float vi = *input++;
44 
45     // Subtract maximum input x := i - i_max. This implies x <= 0.
46     const float vx = vi - vi_max;
47 
48     // Compute reduced argument n := round(x * 64 / log(2)).
49     // We do it by adding a large number (magic bias), which cause rounding of the result to an integer, then subtracing
50     // the large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
51     // The trick with adding large number is valid only within certain bounds (|x * 64 / log(2)| <= 2**22, i.e.
52     // |x| <= 0x1.62E43p+15 = 45426.09375), but that is acceptable, because inputs outside of [-87.336540, 0.0]
53     // result in denormalized or underflown expf(x). We fixup the result for such inputs at the very end of the
54     // algorithm.
55     float vn = vx * vlog2e + vmagic_bias;
56 
57     // Create a floating-point number s (scale) such that s := 2**(n / 64) for such inputs that expf(x) is normalized,
58     // i.e. -87.33642 <= x <= 0.0. As n has 6 fractional bits, we split s == 2**(n / 64) = 2**e * 2**(n / 64 - e), where
59     // e := int(n / 64). We create s in two steps:
60     // 1. Fetch 2**(n / 64 - e) = 2**(n % 64) from the table using the 6 low bits of n, as integer. Note that the
61     //    fetched values are in the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
62     // 2. Adjust fecthed value by addition of e to its floating-point exponent. The result is always a normalized
63     //    number, because for -87.33642 <= x <= 0.0 (inputs for which expf(x) is normalized) we have -126 <= e <= 0,
64     //    and thus the adjusted exponent is not lower than -126.
65     //
66     // Extract e from bits 6:14 of n and shift it into bits 23:31 (position of floating-point exponent).
67     const uint32_t ve = (fp32_to_bits(vn) & UINT32_C(0xFFFFFFC0)) << 17;
68 
69     // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
70     const uint32_t vidx = fp32_to_bits(vn) & vindex_mask;
71     // Adjust exponent of the value l fetched from the table to get the final s value.
72     const float vs = fp32_from_bits(xnn_table_exp2_k_over_64[vidx] + ve);
73 
74     // Subtract the large number back to get final n := round(x * 64 / log(2)) as a floating-point number.
75     vn -= vmagic_bias;
76 
77     // Compute reduced argument t := x - n * log(2) / 64.
78     // Use Cody-Waite range reduction method (note the two constants representing log(2) / 64) to improve accuracy.
79     float vt = vn * vminus_ln2_hi + vx;
80     vt = vn * vminus_ln2_lo + vt;
81 
82     // Compute degree-2 polynomial approximation for exp(t) on [-log(2)/128, log(2)/128].
83     float vp = vt * vc2;
84     vp = vp * vt + vt;
85 
86     // Reconstruct the final f value:
87     //   f = s * (1 + t * (1 + t * c2))
88     //     = s * (1 + t + t * (t * c2))
89     //     = s + s * (t + t * (t * c2))
90     //     = s + s * p
91     float vf = vp * vs + vs;
92 
93     // For inputs below denormal cutoff, replace output with +0.0f.
94     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
95     if XNN_UNPREDICTABLE(vx < vdenorm_cutoff) {
96       vf = 0.0f;
97     }
98 
99     // Store 1 output at a time.
100     *output++ = vf;
101 
102     // Accumulate computed exponents.
103     vacc += vf;
104   }
105   *sum = vacc;
106 }
107