1 | //===-- Collection of utils for sinf16/cosf16 -------------------*- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | #ifndef LLVM_LIBC_SRC_MATH_GENERIC_SINCOSF16_UTILS_H |
10 | #define LLVM_LIBC_SRC_MATH_GENERIC_SINCOSF16_UTILS_H |
11 | |
12 | #include "src/__support/FPUtil/PolyEval.h" |
13 | #include "src/__support/FPUtil/nearest_integer.h" |
14 | #include "src/__support/common.h" |
15 | #include "src/__support/macros/config.h" |
16 | |
17 | namespace LIBC_NAMESPACE_DECL { |
18 | |
19 | // Lookup table for sin(k * pi / 32) with k = 0, ..., 63. |
20 | // Table is generated with Sollya as follows: |
21 | // > display = hexadecimmal; |
22 | // > for k from 0 to 63 do { round(sin(k * pi/32), SG, RN); }; |
23 | constexpr float SIN_K_PI_OVER_32[64] = { |
24 | 0x0.0p0, 0x1.917a6cp-4, 0x1.8f8b84p-3, 0x1.294062p-2, |
25 | 0x1.87de2ap-2, 0x1.e2b5d4p-2, 0x1.1c73b4p-1, 0x1.44cf32p-1, |
26 | 0x1.6a09e6p-1, 0x1.8bc806p-1, 0x1.a9b662p-1, 0x1.c38b3p-1, |
27 | 0x1.d906bcp-1, 0x1.e9f416p-1, 0x1.f6297cp-1, 0x1.fd88dap-1, |
28 | 0x1p0, 0x1.fd88dap-1, 0x1.f6297cp-1, 0x1.e9f416p-1, |
29 | 0x1.d906bcp-1, 0x1.c38b3p-1, 0x1.a9b662p-1, 0x1.8bc806p-1, |
30 | 0x1.6a09e6p-1, 0x1.44cf32p-1, 0x1.1c73b4p-1, 0x1.e2b5d4p-2, |
31 | 0x1.87de2ap-2, 0x1.294062p-2, 0x1.8f8b84p-3, 0x1.917a6cp-4, |
32 | 0x0.0p0, -0x1.917a6cp-4, -0x1.8f8b84p-3, -0x1.294062p-2, |
33 | -0x1.87de2ap-2, -0x1.e2b5d4p-2, -0x1.1c73b4p-1, -0x1.44cf32p-1, |
34 | -0x1.6a09e6p-1, -0x1.8bc806p-1, -0x1.a9b662p-1, -0x1.c38b3p-1, |
35 | -0x1.d906bcp-1, -0x1.e9f416p-1, -0x1.f6297ep-1, -0x1.fd88dap-1, |
36 | -0x1p0, -0x1.fd88dap-1, -0x1.f6297cp-1, -0x1.e9f416p-1, |
37 | -0x1.d906bcp-1, -0x1.c38b3p-1, -0x1.a9b662p-1, -0x1.8bc806p-1, |
38 | -0x1.6a09e6p-1, -0x1.44cf32p-1, -0x1.1c73b4p-1, -0x1.e2b5d4p-2, |
39 | -0x1.87de2ap-2, -0x1.294062p-2, -0x1.8f8b84p-3, -0x1.917a6cp-4}; |
40 | |
41 | LIBC_INLINE int32_t range_reduction_sincospif16(float x, float &y) { |
42 | float kf = fputil::nearest_integer(x * 32); |
43 | y = fputil::multiply_add(x, 32.0f, -kf); |
44 | |
45 | return static_cast<int32_t>(kf); |
46 | } |
47 | |
48 | // Recall, range reduction: |
49 | // k = round(x * 32/pi) |
50 | // |
51 | // The precision choice of 'double' in the following function is to minimize |
52 | // rounding errors in this initial scaling step, |
53 | // preserving enough bits so errors accumulated while computing the subtraction: |
54 | // y = x * 32/pi - round(x * 32/pi) |
55 | // are beyond the least-significant bit of single-precision used during |
56 | // further intermediate computation. |
57 | LIBC_INLINE int32_t range_reduction_sincosf16(float x, float &y) { |
58 | // Generated by Sollya with: |
59 | // > D(32/pi); |
60 | constexpr double THIRTYTWO_OVER_PI = 0x1.45f306dc9c883p3; |
61 | |
62 | double prod = x * THIRTYTWO_OVER_PI; |
63 | double kd = fputil::nearest_integer(prod); |
64 | y = static_cast<float>(prod - kd); |
65 | |
66 | return static_cast<int32_t>(kd); |
67 | } |
68 | |
69 | static LIBC_INLINE void sincosf16_poly_eval(int32_t k, float y, float &sin_k, |
70 | float &cos_k, float &sin_y, |
71 | float &cosm1_y) { |
72 | |
73 | sin_k = SIN_K_PI_OVER_32[k & 63]; |
74 | cos_k = SIN_K_PI_OVER_32[(k + 16) & 63]; |
75 | |
76 | // Recall, after range reduction, -0.5 <= y <= 0.5. For very small values of |
77 | // y, calculating sin(y * p/32) can be inaccurate. Generating a polynomial for |
78 | // sin(y * p/32)/y instead significantly reduces the relative errors. |
79 | float ysq = y * y; |
80 | |
81 | // Degree-6 minimax even polynomial for sin(y*pi/32)/y generated by Sollya |
82 | // with: |
83 | // > Q = fpminimax(sin(y * pi/32)/y, [|0, 2, 4, 6|], [|SG...|], [0, 0.5]); |
84 | sin_y = y * fputil::polyeval(ysq, 0x1.921fb6p-4f, -0x1.4aeabcp-13f, |
85 | 0x1.a03354p-21f, -0x1.ad02d2p-20f); |
86 | |
87 | // Degree-6 minimax even polynomial for cos(y*pi/32) generated by Sollya |
88 | // with: |
89 | // > P = fpminimax(cos(y * pi/32), [|0, 2, 4, 6|],[|1, SG...|], [0, 0.5]); |
90 | cosm1_y = ysq * fputil::polyeval(ysq, -0x1.3bd3ccp-8f, 0x1.03a61ap-18f, |
91 | 0x1.a6f7a2p-29f); |
92 | } |
93 | |
94 | LIBC_INLINE void sincosf16_eval(float xf, float &sin_k, float &cos_k, |
95 | float &sin_y, float &cosm1_y) { |
96 | float y; |
97 | int32_t k = range_reduction_sincosf16(xf, y); |
98 | |
99 | sincosf16_poly_eval(k, y, sin_k, cos_k, sin_y, cosm1_y); |
100 | } |
101 | |
102 | LIBC_INLINE void sincospif16_eval(float xf, float &sin_k, float &cos_k, |
103 | float &sin_y, float &cosm1_y) { |
104 | float y; |
105 | int32_t k = range_reduction_sincospif16(xf, y); |
106 | |
107 | sincosf16_poly_eval(k, y, sin_k, cos_k, sin_y, cosm1_y); |
108 | } |
109 | |
110 | } // namespace LIBC_NAMESPACE_DECL |
111 | |
112 | #endif // LLVM_LIBC_SRC_MATH_GENERIC_SINCOSF16_UTILS_H |
113 | |