brintos

brintos / llvm-project-archived public Read only

0
0
Text · 4.6 KiB · 74f21fd Raw
117 lines · c
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-exception6//7//===----------------------------------------------------------------------===//8 9#ifndef LLVM_LIBC_SRC_MATH_GENERIC_SINCOSF16_UTILS_H10#define LLVM_LIBC_SRC_MATH_GENERIC_SINCOSF16_UTILS_H11 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 17namespace LIBC_NAMESPACE_DECL {18 19namespace sincosf16_internal {20 21// Lookup table for sin(k * pi / 32) with k = 0, ..., 63.22// Table is generated with Sollya as follows:23// > display = hexadecimmal;24// > for k from 0 to 63 do { round(sin(k * pi/32), SG, RN); };25constexpr float SIN_K_PI_OVER_32[64] = {26    0x0.0p0,        0x1.917a6cp-4,  0x1.8f8b84p-3,  0x1.294062p-2,27    0x1.87de2ap-2,  0x1.e2b5d4p-2,  0x1.1c73b4p-1,  0x1.44cf32p-1,28    0x1.6a09e6p-1,  0x1.8bc806p-1,  0x1.a9b662p-1,  0x1.c38b3p-1,29    0x1.d906bcp-1,  0x1.e9f416p-1,  0x1.f6297cp-1,  0x1.fd88dap-1,30    0x1p0,          0x1.fd88dap-1,  0x1.f6297cp-1,  0x1.e9f416p-1,31    0x1.d906bcp-1,  0x1.c38b3p-1,   0x1.a9b662p-1,  0x1.8bc806p-1,32    0x1.6a09e6p-1,  0x1.44cf32p-1,  0x1.1c73b4p-1,  0x1.e2b5d4p-2,33    0x1.87de2ap-2,  0x1.294062p-2,  0x1.8f8b84p-3,  0x1.917a6cp-4,34    0x0.0p0,        -0x1.917a6cp-4, -0x1.8f8b84p-3, -0x1.294062p-2,35    -0x1.87de2ap-2, -0x1.e2b5d4p-2, -0x1.1c73b4p-1, -0x1.44cf32p-1,36    -0x1.6a09e6p-1, -0x1.8bc806p-1, -0x1.a9b662p-1, -0x1.c38b3p-1,37    -0x1.d906bcp-1, -0x1.e9f416p-1, -0x1.f6297ep-1, -0x1.fd88dap-1,38    -0x1p0,         -0x1.fd88dap-1, -0x1.f6297cp-1, -0x1.e9f416p-1,39    -0x1.d906bcp-1, -0x1.c38b3p-1,  -0x1.a9b662p-1, -0x1.8bc806p-1,40    -0x1.6a09e6p-1, -0x1.44cf32p-1, -0x1.1c73b4p-1, -0x1.e2b5d4p-2,41    -0x1.87de2ap-2, -0x1.294062p-2, -0x1.8f8b84p-3, -0x1.917a6cp-4};42 43LIBC_INLINE int32_t range_reduction_sincospif16(float x, float &y) {44  float kf = fputil::nearest_integer(x * 32);45  y = fputil::multiply_add(x, 32.0f, -kf);46 47  return static_cast<int32_t>(kf);48}49 50// Recall, range reduction:51//   k = round(x * 32/pi)52//53// The precision choice of 'double' in the following function is to minimize54// rounding errors in this initial scaling step,55// preserving enough bits so errors accumulated while computing the subtraction:56// y = x * 32/pi - round(x * 32/pi)57// are beyond the least-significant bit of single-precision used during58// further intermediate computation.59LIBC_INLINE int32_t range_reduction_sincosf16(float x, float &y) {60  // Generated by Sollya with:61  // > D(32/pi);62  constexpr double THIRTYTWO_OVER_PI = 0x1.45f306dc9c883p3;63 64  double prod = x * THIRTYTWO_OVER_PI;65  double kd = fputil::nearest_integer(prod);66  y = static_cast<float>(prod - kd);67 68  return static_cast<int32_t>(kd);69}70 71LIBC_INLINE static void sincosf16_poly_eval(int32_t k, float y, float &sin_k,72                                            float &cos_k, float &sin_y,73                                            float &cosm1_y) {74 75  sin_k = SIN_K_PI_OVER_32[k & 63];76  cos_k = SIN_K_PI_OVER_32[(k + 16) & 63];77 78  // Recall, after range reduction, -0.5 <= y <= 0.5. For very small values of79  // y, calculating sin(y * p/32) can be inaccurate. Generating a polynomial for80  // sin(y * p/32)/y instead significantly reduces the relative errors.81  float ysq = y * y;82 83  // Degree-6 minimax even polynomial for sin(y*pi/32)/y generated by Sollya84  // with:85  // > Q = fpminimax(sin(y * pi/32)/y, [|0, 2, 4, 6|], [|SG...|], [0, 0.5]);86  sin_y = y * fputil::polyeval(ysq, 0x1.921fb6p-4f, -0x1.4aeabcp-13f,87                               0x1.a03354p-21f, -0x1.ad02d2p-20f);88 89  // Degree-6 minimax even polynomial for cos(y*pi/32) generated by Sollya90  // with:91  // > P = fpminimax(cos(y * pi/32), [|0, 2, 4, 6|],[|1, SG...|], [0, 0.5]);92  cosm1_y = ysq * fputil::polyeval(ysq, -0x1.3bd3ccp-8f, 0x1.03a61ap-18f,93                                   0x1.a6f7a2p-29f);94}95 96LIBC_INLINE void sincosf16_eval(float xf, float &sin_k, float &cos_k,97                                float &sin_y, float &cosm1_y) {98  float y;99  int32_t k = range_reduction_sincosf16(xf, y);100 101  sincosf16_poly_eval(k, y, sin_k, cos_k, sin_y, cosm1_y);102}103 104LIBC_INLINE void sincospif16_eval(float xf, float &sin_k, float &cos_k,105                                  float &sin_y, float &cosm1_y) {106  float y;107  int32_t k = range_reduction_sincospif16(xf, y);108 109  sincosf16_poly_eval(k, y, sin_k, cos_k, sin_y, cosm1_y);110}111 112} // namespace sincosf16_internal113 114} // namespace LIBC_NAMESPACE_DECL115 116#endif // LLVM_LIBC_SRC_MATH_GENERIC_SINCOSF16_UTILS_H117