From f89bcca0b783ab1003ac8e1f1af2424c751b0361 Mon Sep 17 00:00:00 2001 From: Aaron Watry Date: Thu, 26 Feb 2015 15:42:00 +0000 Subject: [PATCH] libclc/math: Add cospi Ported from the libclc/amd-builtins branch v2: Rename sincos_f_piby4 to __libclc__sincosf_piby4 Add cospi(double) implementation instead of using llvm.cos Notes: The sincosD_piby4.h file is mostly the same as the builtin implementation released by AMD. The inline attribute declaration is changed, and M_PI is used instead of a constant double. Otherwise, the only difference is that the header explicitly enables the fp64 pragma. Signed-off-by: Aaron Watry Reviewed-by: Jeroen Ketema CC: Tom Stellard CC: Matt Arsenault llvm-svn: 230641 --- libclc/generic/include/clc/clc.h | 1 + libclc/generic/include/clc/math/cospi.h | 3 + libclc/generic/include/clc/math/cospi.inc | 1 + libclc/generic/lib/SOURCES | 1 + libclc/generic/lib/math/cospi.cl | 136 ++++++++++++++++++++++ libclc/generic/lib/math/sincosD_piby4.h | 78 +++++++++++++ libclc/generic/lib/math/sincospiF_piby4.h | 56 +++++++++ 7 files changed, 276 insertions(+) create mode 100644 libclc/generic/include/clc/math/cospi.h create mode 100644 libclc/generic/include/clc/math/cospi.inc create mode 100644 libclc/generic/lib/math/cospi.cl create mode 100644 libclc/generic/lib/math/sincosD_piby4.h create mode 100644 libclc/generic/lib/math/sincospiF_piby4.h diff --git a/libclc/generic/include/clc/clc.h b/libclc/generic/include/clc/clc.h index 29c185500c28..a324b8f68bd7 100644 --- a/libclc/generic/include/clc/clc.h +++ b/libclc/generic/include/clc/clc.h @@ -38,6 +38,7 @@ #include #include #include +#include #include #include #include diff --git a/libclc/generic/include/clc/math/cospi.h b/libclc/generic/include/clc/math/cospi.h new file mode 100644 index 000000000000..427733bccda4 --- /dev/null +++ b/libclc/generic/include/clc/math/cospi.h @@ -0,0 +1,3 @@ +#define __CLC_BODY +#include +#undef __CLC_BODY diff --git a/libclc/generic/include/clc/math/cospi.inc b/libclc/generic/include/clc/math/cospi.inc new file mode 100644 index 000000000000..1e786cfda6b5 --- /dev/null +++ b/libclc/generic/include/clc/math/cospi.inc @@ -0,0 +1 @@ +_CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE cospi(__CLC_GENTYPE a); diff --git a/libclc/generic/lib/SOURCES b/libclc/generic/lib/SOURCES index 1032606c518a..af5f97acaaf6 100644 --- a/libclc/generic/lib/SOURCES +++ b/libclc/generic/lib/SOURCES @@ -57,6 +57,7 @@ math/atan.cl math/atan2.cl math/copysign.cl math/cos.cl +math/cospi.cl math/exp.cl math/exp10.cl math/fmax.cl diff --git a/libclc/generic/lib/math/cospi.cl b/libclc/generic/lib/math/cospi.cl new file mode 100644 index 000000000000..108b637c9abb --- /dev/null +++ b/libclc/generic/lib/math/cospi.cl @@ -0,0 +1,136 @@ +/* + * Copyright (c) 2014 Advanced Micro Devices, Inc. + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * in the Software without restriction, including without limitation the rights + * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell + * copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN + * THE SOFTWARE. + */ + +#include + +#include "math.h" +#include "sincos_helpers.h" +#include "sincospiF_piby4.h" +#include "../clcmacro.h" +#ifdef cl_khr_fp64 +#include "sincosD_piby4.h" +#endif + +_CLC_OVERLOAD _CLC_DEF float cospi(float x) +{ + int ix = as_int(x) & 0x7fffffff; + float ax = as_float(ix); + int iax = (int)ax; + float r = ax - iax; + int xodd = iax & 0x1 ? 0x80000000 : 0; + + // Initialize with return for +-Inf and NaN + int ir = 0x7fc00000; + + // 2^24 <= |x| < Inf, the result is always even integer + ir = ix < 0x7f800000 ? 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 + float a = 1.0f - r; + int e = 1; + int s = xodd ^ 0x80000000; + + // r <= 0.75 + int 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; + + float2 t = __libclc__sincosf_piby4(a * M_PI_F); + int jr = s ^ as_int(e ? t.hi : t.lo); + + ir = ix < 0x4b000000 ? jr : ir; + + return as_float(ir); +} + + +_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, float, cospi, float); + +#ifdef cl_khr_fp64 + +#pragma OPENCL EXTENSION cl_khr_fp64 : enable + +_CLC_OVERLOAD _CLC_DEF double cospi(double x) { + + long ix = as_long(x) & 0x7fffffffffffffffL; + double ax = as_double(ix); + long iax = (long)ax; + double r = ax - (double)iax; + long xodd = iax & 0x1L ? 0x8000000000000000L : 0L; + + // Initialize with return for +-Inf and NaN + long ir = 0x7ff8000000000000L; + + // 2^53 <= |x| < Inf, the result is always even integer + ir = ix < 0x7ff0000000000000 ? 0x3ff0000000000000L : ir; + + // 2^52 <= |x| < 2^53, the result is always integer + ir = ax < 0x1.0p+53 ? xodd | 0x3ff0000000000000L : ir; + + // 0x1.0p-7 <= |x| < 2^52, result depends on which 0.25 interval + + // r < 1.0 + double a = 1.0 - r; + int e = 1; + long s = xodd ^ 0x8000000000000000L; + + // r <= 0.75 + int c = r <= 0.75; + double 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; + + double2 sc = __libclc__sincos_piby4(a * M_PI, 0.0); + long jr = s ^ as_long(e ? sc.hi : sc.lo); + + ir = ax < 0x1.0p+52 ? jr : ir; + + return as_double(ir); +} +_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, double, cospi, double); +#endif diff --git a/libclc/generic/lib/math/sincosD_piby4.h b/libclc/generic/lib/math/sincosD_piby4.h new file mode 100644 index 000000000000..a00db854891a --- /dev/null +++ b/libclc/generic/lib/math/sincosD_piby4.h @@ -0,0 +1,78 @@ +/* + * Copyright (c) 2014 Advanced Micro Devices, Inc. + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * in the Software without restriction, including without limitation the rights + * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell + * copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN + * THE SOFTWARE. + */ + +#pragma OPENCL EXTENSION cl_khr_fp64 : enable + +_CLC_INLINE double2 +__libclc__sincos_piby4(double x, double xx) +{ + // Taylor series for sin(x) is x - x^3/3! + x^5/5! - x^7/7! ... + // = x * (1 - x^2/3! + x^4/5! - x^6/7! ... + // = x * f(w) + // where w = x*x and f(w) = (1 - w/3! + w^2/5! - w^3/7! ... + // We use a minimax approximation of (f(w) - 1) / w + // because this produces an expansion in even powers of x. + // If xx (the tail of x) is non-zero, we add a correction + // term g(x,xx) = (1-x*x/2)*xx to the result, where g(x,xx) + // is an approximation to cos(x)*sin(xx) valid because + // xx is tiny relative to x. + + // Taylor series for cos(x) is 1 - x^2/2! + x^4/4! - x^6/6! ... + // = f(w) + // where w = x*x and f(w) = (1 - w/2! + w^2/4! - w^3/6! ... + // We use a minimax approximation of (f(w) - 1 + w/2) / (w*w) + // because this produces an expansion in even powers of x. + // If xx (the tail of x) is non-zero, we subtract a correction + // term g(x,xx) = x*xx to the result, where g(x,xx) + // is an approximation to sin(x)*sin(xx) valid because + // xx is tiny relative to x. + + const double sc1 = -0.166666666666666646259241729; + const double sc2 = 0.833333333333095043065222816e-2; + const double sc3 = -0.19841269836761125688538679e-3; + const double sc4 = 0.275573161037288022676895908448e-5; + const double sc5 = -0.25051132068021699772257377197e-7; + const double sc6 = 0.159181443044859136852668200e-9; + + const double cc1 = 0.41666666666666665390037e-1; + const double cc2 = -0.13888888888887398280412e-2; + const double cc3 = 0.248015872987670414957399e-4; + const double cc4 = -0.275573172723441909470836e-6; + const double cc5 = 0.208761463822329611076335e-8; + const double cc6 = -0.113826398067944859590880e-10; + + double x2 = x * x; + double x3 = x2 * x; + double r = 0.5 * x2; + double t = 1.0 - r; + + double sp = fma(fma(fma(fma(sc6, x2, sc5), x2, sc4), x2, sc3), x2, sc2); + + double cp = t + fma(fma(fma(fma(fma(fma(cc6, x2, cc5), x2, cc4), x2, cc3), x2, cc2), x2, cc1), + x2*x2, fma(x, xx, (1.0 - t) - r)); + + double2 ret; + ret.lo = x - fma(-x3, sc1, fma(fma(-x3, sp, 0.5*xx), x2, -xx)); + ret.hi = cp; + + return ret; +} diff --git a/libclc/generic/lib/math/sincospiF_piby4.h b/libclc/generic/lib/math/sincospiF_piby4.h new file mode 100644 index 000000000000..90ecb1d7a636 --- /dev/null +++ b/libclc/generic/lib/math/sincospiF_piby4.h @@ -0,0 +1,56 @@ +/* + * Copyright (c) 2014 Advanced Micro Devices, Inc. + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * in the Software without restriction, including without limitation the rights + * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell + * copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN + * THE SOFTWARE. + */ + +// Evaluate single precisions in and cos of value in interval [-pi/4, pi/4] +_CLC_INLINE float2 +__libclc__sincosf_piby4(float x) +{ + // Taylor series for sin(x) is x - x^3/3! + x^5/5! - x^7/7! ... + // = x * (1 - x^2/3! + x^4/5! - x^6/7! ... + // = x * f(w) + // where w = x*x and f(w) = (1 - w/3! + w^2/5! - w^3/7! ... + // We use a minimax approximation of (f(w) - 1) / w + // because this produces an expansion in even powers of x. + + // Taylor series for cos(x) is 1 - x^2/2! + x^4/4! - x^6/6! ... + // = f(w) + // where w = x*x and f(w) = (1 - w/2! + w^2/4! - w^3/6! ... + // We use a minimax approximation of (f(w) - 1 + w/2) / (w*w) + // because this produces an expansion in even powers of x. + + const float sc1 = -0.166666666638608441788607926e0F; + const float sc2 = 0.833333187633086262120839299e-2F; + const float sc3 = -0.198400874359527693921333720e-3F; + const float sc4 = 0.272500015145584081596826911e-5F; + + const float cc1 = 0.41666666664325175238031e-1F; + const float cc2 = -0.13888887673175665567647e-2F; + const float cc3 = 0.24800600878112441958053e-4F; + const float cc4 = -0.27301013343179832472841e-6F; + + float x2 = x * x; + + float2 ret; + ret.x = mad(x*x2, mad(x2, mad(x2, mad(x2, sc4, sc3), sc2), sc1), x); + ret.y = mad(x2*x2, mad(x2, mad(x2, mad(x2, cc4, cc3), cc2), cc1), mad(x2, -0.5f, 1.0f)); + return ret; +}