|
| 1 | +//===-- Half-precision cospif function ------------------------------------===// |
| 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 | +#include "src/math/cospif16.h" |
| 10 | +#include "src/__support/FPUtil/FEnvImpl.h" |
| 11 | +#include "src/__support/FPUtil/FPBits.h" |
| 12 | +#include "src/__support/FPUtil/PolyEval.h" |
| 13 | +#include "src/__support/FPUtil/cast.h" |
| 14 | +#include "src/__support/FPUtil/multiply_add.h" |
| 15 | +#include "src/__support/FPUtil/nearest_integer.h" |
| 16 | +#include "src/__support/common.h" |
| 17 | +#include "src/__support/macros/config.h" |
| 18 | + |
| 19 | +namespace LIBC_NAMESPACE_DECL { |
| 20 | +// Lookup table for sin(k * pi / 32) with k = 0, ..., 63. |
| 21 | +// Table is generated with Sollya as follows: |
| 22 | +// > display = hexadecimal; |
| 23 | +// > for k from 0 to 63 do { round(sin(k * pi/32), SG, RN); }; |
| 24 | +static constexpr float SIN_K_PI_OVER_32[64] = { |
| 25 | + 0x0.0p0, 0x1.917a6cp-4, 0x1.8f8b84p-3, 0x1.294062p-2, |
| 26 | + 0x1.87de2ap-2, 0x1.e2b5d4p-2, 0x1.1c73b4p-1, 0x1.44cf32p-1, |
| 27 | + 0x1.6a09e6p-1, 0x1.8bc806p-1, 0x1.a9b662p-1, 0x1.c38b3p-1, |
| 28 | + 0x1.d906bcp-1, 0x1.e9f416p-1, 0x1.f6297cp-1, 0x1.fd88dap-1, |
| 29 | + 0x1p0, 0x1.fd88dap-1, 0x1.f6297cp-1, 0x1.e9f416p-1, |
| 30 | + 0x1.d906bcp-1, 0x1.c38b3p-1, 0x1.a9b662p-1, 0x1.8bc806p-1, |
| 31 | + 0x1.6a09e6p-1, 0x1.44cf32p-1, 0x1.1c73b4p-1, 0x1.e2b5d4p-2, |
| 32 | + 0x1.87de2ap-2, 0x1.294062p-2, 0x1.8f8b84p-3, 0x1.917a6cp-4, |
| 33 | + 0x0.0p0, -0x1.917a6cp-4, -0x1.8f8b84p-3, -0x1.294062p-2, |
| 34 | + -0x1.87de2ap-2, -0x1.e2b5d4p-2, -0x1.1c73b4p-1, -0x1.44cf32p-1, |
| 35 | + -0x1.6a09e6p-1, -0x1.8bc806p-1, -0x1.a9b662p-1, -0x1.c38b3p-1, |
| 36 | + -0x1.d906bcp-1, -0x1.e9f416p-1, -0x1.f6297ep-1, -0x1.fd88dap-1, |
| 37 | + -0x1p0, -0x1.fd88dap-1, -0x1.f6297cp-1, -0x1.e9f416p-1, |
| 38 | + -0x1.d906bcp-1, -0x1.c38b3p-1, -0x1.a9b662p-1, -0x1.8bc806p-1, |
| 39 | + -0x1.6a09e6p-1, -0x1.44cf32p-1, -0x1.1c73b4p-1, -0x1.e2b5d4p-2, |
| 40 | + -0x1.87de2ap-2, -0x1.294062p-2, -0x1.8f8b84p-3, -0x1.917a6cp-4}; |
| 41 | + |
| 42 | +static LIBC_INLINE int32_t range_reduction(float x, float &y) { |
| 43 | + float kf = fputil::nearest_integer(x * 32); |
| 44 | + y = fputil::multiply_add<float>(x, 32.0, -kf); |
| 45 | + |
| 46 | + return static_cast<int32_t>(kf); |
| 47 | +} |
| 48 | + |
| 49 | +LLVM_LIBC_FUNCTION(float16, cospif16, (float16 x)) { |
| 50 | + using FPBits = typename fputil::FPBits<float16>; |
| 51 | + FPBits xbits(x); |
| 52 | + |
| 53 | + uint16_t x_u = xbits.uintval(); |
| 54 | + uint16_t x_abs = x_u & 0x7fff; |
| 55 | + |
| 56 | + // Range reduction: |
| 57 | + // For |x| > 1/32, we perform range reduction as follows: |
| 58 | + // Find k and y such that: |
| 59 | + // x = (k + y) * 1/32 |
| 60 | + // k is an integer |
| 61 | + // |y| < 0.5 |
| 62 | + // |
| 63 | + // This is done by performing: |
| 64 | + // k = round(x * 32) |
| 65 | + // y = x * 32 - k |
| 66 | + // |
| 67 | + // Once k and y are computed, we then deduce the answer by the sine of sum |
| 68 | + // formula: |
| 69 | + // sin(x * pi) = sin((k + y) * pi/32) |
| 70 | + // = sin(k * pi/32) * cos(y * pi/32) + sin (y * pi/32) * cos (k * |
| 71 | + // pi/32) |
| 72 | + // The values of sin(k * pi/32) and cos (k * pi/32) for k = 0...63 are |
| 73 | + // precomputed and stored using a vector of 64 single precision floats. sin(y |
| 74 | + // * pi/32) and cos(y * pi/32) are computed using degree-9 chebyshev |
| 75 | + // polynomials generated by Sollya. |
| 76 | + |
| 77 | + // For signed zeros |
| 78 | + if (LIBC_UNLIKELY(x_abs == 0U)) return fputil::cast<float16>(1.0f); |
| 79 | + |
| 80 | + // Numbers greater or equal to 2^10 are integers, or infinity, or NaN |
| 81 | + if (LIBC_UNLIKELY(x_abs >= 0x6400)) { |
| 82 | + if (LIBC_UNLIKELY(x_abs <= 0x67FF)) { |
| 83 | + return fputil::cast<float16>((x_abs & 0x1) ? -1.0f : 1.0f); |
| 84 | + } |
| 85 | + |
| 86 | + // Check for NaN or infintiy values |
| 87 | + if (LIBC_UNLIKELY(x_abs >= 0x7c00)) { |
| 88 | + // If value is equal to infinity |
| 89 | + if (x_abs == 0x7c00) { |
| 90 | + fputil::set_errno_if_required(EDOM); |
| 91 | + fputil::raise_except_if_required(FE_INVALID); |
| 92 | + } |
| 93 | + |
| 94 | + return x + FPBits::quiet_nan().get_val(); |
| 95 | + } |
| 96 | + |
| 97 | + return fputil::cast<float16>(1.0f); |
| 98 | + } |
| 99 | + |
| 100 | + |
| 101 | + float f32 = x; |
| 102 | + float y; |
| 103 | + int32_t k = range_reduction(f32, y); |
| 104 | + |
| 105 | + float sin_k = SIN_K_PI_OVER_32[k & 63]; |
| 106 | + float cos_k = SIN_K_PI_OVER_32[(k + 16) & 63]; |
| 107 | + |
| 108 | + // Recall; |
| 109 | + // cos(x * pi/32) = cos((k + y) * pi/32) |
| 110 | + // = cos(y * pi/32) * cos(k * pi/32) |
| 111 | + // - sin(y * pi/32) * sin(k * pi/32) |
| 112 | + // Recall, after range reduction, -0.5 <= y <= 0.5. For very small |
| 113 | + // values of y, calculating sin(y * p/32) can be inaccurate. Generating a |
| 114 | + // polynomial for sin(y * p/32)/y instead significantly reduces the relative |
| 115 | + // errors. |
| 116 | + float ysq = y * y; |
| 117 | + |
| 118 | + // Degree-6 minimax even polynomial for sin(y*pi/32)/y generated by Sollya |
| 119 | + // with: |
| 120 | + // > Q = fpminimax(sin(y*pi/32)/y, [|0, 2, 4, 6|], [|SG...|], [0, 0.5]); |
| 121 | + float sin_y = y * fputil::polyeval(ysq, 0x1.921fb6p-4f, -0x1.4aeabcp-13f, |
| 122 | + 0x1.a03354p-21f, -0x1.ad02d2p-20f); |
| 123 | + |
| 124 | + // Note that cosm1_y = cos(y*pi/32) - 1 = cos_y - 1 |
| 125 | + // Derivation: |
| 126 | + // cos(x * pi) = cos((k + y) * pi/32) |
| 127 | + // = cos_k * cos_y + sin_k * sin_y |
| 128 | + // = cos_k * (1 + cos_y - 1) + sin_k * sin_y |
| 129 | + // Degree-6 minimax even polynomial for cos(y*pi/32) generated by Sollya with: |
| 130 | + // > P = fpminimax(cos(y*pi/32), [|0, 2, 4, 6|],[|1, SG...|], [0, 0.5]); |
| 131 | + float cosm1_y = ysq * fputil::polyeval(ysq, -0x1.3bd3ccp-8f, 0x1.03a61ap-18f, |
| 132 | + 0x1.a6f7a2p-29f); |
| 133 | + |
| 134 | + if (LIBC_UNLIKELY(sin_y == 0 && cos_k == 0)) |
| 135 | + return fputil::cast<float16>(0.0f); |
| 136 | + |
| 137 | + // Since, cosm1_y = cos_y - 1, therefore: |
| 138 | + // cos(x * pi) = cos_k(cosm1_y) + cos_k - sin_k * sin_y |
| 139 | + return fputil::cast<float16>(fputil::multiply_add(cos_k, cosm1_y, fputil::multiply_add(-sin_k, sin_y, cos_k))); |
| 140 | +} |
| 141 | +} // namespace LIBC_NAMESPACE_DECL |
0 commit comments