mirror of
https://github.com/llvm/llvm-project.git
synced 2025-04-17 04:16:46 +00:00
13a313fe58
Additionally, these builtins are now vectorized. This also moves the native_recip and native_divide builtins as they are used by the tanpi builtin.
117 lines
3.1 KiB
C++
117 lines
3.1 KiB
C++
//===----------------------------------------------------------------------===//
|
|
//
|
|
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
|
|
// See https://llvm.org/LICENSE.txt for license information.
|
|
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
#if __CLC_FPSIZE == 32
|
|
|
|
_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_cospi(__CLC_GENTYPE x) {
|
|
__CLC_GENTYPE absx = __clc_fabs(x);
|
|
__CLC_INTN ix = __CLC_AS_INTN(absx);
|
|
__CLC_INTN iax = __CLC_CONVERT_INTN(absx);
|
|
__CLC_GENTYPE r = absx - __CLC_CONVERT_GENTYPE(iax);
|
|
__CLC_INTN xodd = (iax & 0x1) != 0 ? (__CLC_INTN)0x80000000 : (__CLC_INTN)0;
|
|
|
|
// Initialize with return for +-Inf and NaN
|
|
__CLC_INTN ir = QNANBITPATT_SP32;
|
|
|
|
// 2^24 <= |x| < Inf, the result is always even integer
|
|
ir = ix < PINFBITPATT_SP32 ? 0x3f800000 : ir;
|
|
|
|
// 2^23 <= |x| < 2^24, the result is always integer
|
|
ir = ix < 0x4b800000 ? xodd | 0x3f800000 : ir;
|
|
|
|
// 0x1.0p-7 <= |x| < 2^23, result depends on which 0.25 interval
|
|
|
|
// r < 1.0
|
|
__CLC_GENTYPE a = 1.0f - r;
|
|
__CLC_INTN e = 1;
|
|
__CLC_INTN s = xodd ^ (__CLC_INTN)0x80000000;
|
|
|
|
// r <= 0.75
|
|
__CLC_INTN c = r <= 0.75f;
|
|
a = c ? r - 0.5f : a;
|
|
e = c ? 0 : e;
|
|
|
|
// r < 0.5
|
|
c = r < 0.5f;
|
|
a = c ? 0.5f - r : a;
|
|
s = c ? xodd : s;
|
|
|
|
// r <= 0.25
|
|
c = r <= 0.25f;
|
|
a = c ? r : a;
|
|
e = c ? 1 : e;
|
|
|
|
__CLC_GENTYPE sinval, cosval;
|
|
__clc_sincos_piby4(a * M_PI_F, &sinval, &cosval);
|
|
__CLC_INTN jr = s ^ __CLC_AS_INTN(e != 0 ? cosval : sinval);
|
|
|
|
ir = ix < 0x4b000000 ? jr : ir;
|
|
|
|
return __CLC_AS_GENTYPE(ir);
|
|
}
|
|
|
|
#elif __CLC_FPSIZE == 64
|
|
|
|
_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_cospi(__CLC_GENTYPE x) {
|
|
__CLC_GENTYPE absx = __clc_fabs(x);
|
|
__CLC_LONGN ix = __CLC_AS_LONGN(absx);
|
|
__CLC_LONGN iax = __CLC_CONVERT_LONGN(absx);
|
|
__CLC_GENTYPE r = absx - __CLC_CONVERT_GENTYPE(iax);
|
|
__CLC_LONGN xodd =
|
|
(iax & 0x1L) != 0 ? (__CLC_LONGN)0x8000000000000000L : (__CLC_LONGN)0L;
|
|
|
|
// Initialize with return for +-Inf and NaN
|
|
__CLC_LONGN ir = QNANBITPATT_DP64;
|
|
|
|
// 2^53 <= |x| < Inf, the result is always even integer
|
|
ir = ix < PINFBITPATT_DP64 ? 0x3ff0000000000000L : ir;
|
|
|
|
// 2^52 <= |x| < 2^53, the result is always integer
|
|
ir = absx < 0x1.0p+53 ? xodd | 0x3ff0000000000000L : ir;
|
|
|
|
// 0x1.0p-7 <= |x| < 2^52, result depends on which 0.25 interval
|
|
|
|
// r < 1.0
|
|
__CLC_GENTYPE a = 1.0 - r;
|
|
__CLC_LONGN e = 1;
|
|
__CLC_LONGN s = xodd ^ (__CLC_LONGN)0x8000000000000000L;
|
|
|
|
// r <= 0.75
|
|
__CLC_LONGN c = r <= 0.75;
|
|
__CLC_GENTYPE t = r - 0.5;
|
|
a = c ? t : a;
|
|
e = c ? 0 : e;
|
|
|
|
// r < 0.5
|
|
c = r < 0.5;
|
|
t = 0.5 - r;
|
|
a = c ? t : a;
|
|
s = c ? xodd : s;
|
|
|
|
// r <= 0.25
|
|
c = r <= 0.25;
|
|
a = c ? r : a;
|
|
e = c ? 1 : e;
|
|
|
|
__CLC_GENTYPE sinval, cosval;
|
|
__clc_sincos_piby4(a * M_PI, 0.0, &sinval, &cosval);
|
|
__CLC_LONGN jr = s ^ __CLC_AS_LONGN(e != 0 ? cosval : sinval);
|
|
|
|
ir = absx < 0x1.0p+52 ? jr : ir;
|
|
|
|
return __CLC_AS_GENTYPE(ir);
|
|
}
|
|
|
|
#elif __CLC_FPSIZE == 16
|
|
|
|
_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_cospi(__CLC_GENTYPE x) {
|
|
return __CLC_CONVERT_GENTYPE(__clc_cospi(__CLC_CONVERT_FLOATN(x)));
|
|
}
|
|
|
|
#endif
|