[Libclc-dev] [PATCH] libclc/math: Add cospi
Tom Stellard
tom at stellard.net
Tue Dec 16 18:33:05 PST 2014
On Tue, Dec 16, 2014 at 03:41:52PM +0000, Jeroen Ketema wrote:
>
> > On 16 Dec 2014, at 15:29, Aaron Watry <awatry at gmail.com> wrote:
> >
> > On Mon, Dec 15, 2014 at 5:25 PM, Jeroen Ketema <j.ketema at imperial.ac.uk> wrote:
> >>
> >> Hi Aaron,
> >>
> >> Comments inline.
> >>
> >>> On 15 Dec 2014, at 23:14, Aaron Watry <awatry at gmail.com> wrote:
> >>>
> >>> Ported from the libclc/amd-builtins branch
> >>>
> >>> Signed-off-by: Aaron Watry <awatry at gmail.com>
> >>> ---
> >>> generic/include/clc/clc.h | 1 +
> >>> generic/include/clc/math/cospi.h | 3 ++
> >>> generic/include/clc/math/cospi.inc | 1 +
> >>> generic/lib/SOURCES | 1 +
> >>> generic/lib/math/cospi.cl | 96 ++++++++++++++++++++++++++++++++++++++
> >>> generic/lib/math/sincospiF_piby4.h | 57 ++++++++++++++++++++++
> >>> 6 files changed, 159 insertions(+)
> >>> create mode 100644 generic/include/clc/math/cospi.h
> >>> create mode 100644 generic/include/clc/math/cospi.inc
> >>> create mode 100644 generic/lib/math/cospi.cl
> >>> create mode 100644 generic/lib/math/sincospiF_piby4.h
> >>>
> >>> diff --git a/generic/include/clc/clc.h b/generic/include/clc/clc.h
> >>> index bd92fdb..bff6ddd 100644
> >>> --- a/generic/include/clc/clc.h
> >>> +++ b/generic/include/clc/clc.h
> >>> @@ -38,6 +38,7 @@
> >>> #include <clc/math/atan2.h>
> >>> #include <clc/math/copysign.h>
> >>> #include <clc/math/cos.h>
> >>> +#include <clc/math/cospi.h>
> >>> #include <clc/math/ceil.h>
> >>> #include <clc/math/exp.h>
> >>> #include <clc/math/exp10.h>
> >>> diff --git a/generic/include/clc/math/cospi.h b/generic/include/clc/math/cospi.h
> >>> new file mode 100644
> >>> index 0000000..427733b
> >>> --- /dev/null
> >>> +++ b/generic/include/clc/math/cospi.h
> >>> @@ -0,0 +1,3 @@
> >>> +#define __CLC_BODY <clc/math/cospi.inc>
> >>> +#include <clc/math/gentype.inc>
> >>> +#undef __CLC_BODY
> >>> diff --git a/generic/include/clc/math/cospi.inc b/generic/include/clc/math/cospi.inc
> >>> new file mode 100644
> >>> index 0000000..1e786cf
> >>> --- /dev/null
> >>> +++ b/generic/include/clc/math/cospi.inc
> >>> @@ -0,0 +1 @@
> >>> +_CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE cospi(__CLC_GENTYPE a);
> >>> diff --git a/generic/lib/SOURCES b/generic/lib/SOURCES
> >>> index b76fec9..2b9426e 100644
> >>> --- a/generic/lib/SOURCES
> >>> +++ b/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/generic/lib/math/cospi.cl b/generic/lib/math/cospi.cl
> >>> new file mode 100644
> >>> index 0000000..666deb7
> >>> --- /dev/null
> >>> +++ b/generic/lib/math/cospi.cl
> >>> @@ -0,0 +1,96 @@
> >>> +/*
> >>> + * 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 <clc/clc.h>
> >>> +
> >>> +#include "math.h"
> >>> +#include "sincos_helpers.h"
> >>> +#include "sincospiF_piby4.h"
> >>> +#include "../clcmacro.h"
> >>> +
> >>> +_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 = 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
> >>> +
> >>> +#define __CLC_FUNCTION __clc_cos_intrinsic
> >>> +#define __CLC_INTRINSIC “llvm.cos"
> >>
> >> Does llvm.cos give sufficient precision for double, or is the intention to leave double as it is for now?
> >>
> >
> > I honestly don't know if llvm.cos gives sufficient precision for
> > doubles here. I didn't see any precision notes in the SI ISA guide,
> > so until we either have some unit tests for doubles or maybe if Tom
> > sees this, I can't really answer that. There is a cospi(double)
> > implementation in the amd builtins that were released recently, but I
> > haven't attempted to port that yet. I was assuming that because
> > cos(double) used __clc_cos_intrinsic, it was probably safe to use it
> > for cospi(double) as well... although it's also possible that
> > cos(double) isn't precise enough either.
>
> AFAIK, Tom didn’t try to fix-up the double implementation of cos, but
> only the float one. So, whatever applies there is likely to apply here
> as well.
>
The only reason I didn't fix the double implementation was that I didn't
have a good way to test it. I think it's fine to leave the double
version as is for cospi too.
-Tom
> > One of these days, we need to extend the generated piglit tests to
> > test double for inputs/outputs for builtins (that should give us
> > decent test coverage of doubles without too much effort, at least for
> > devices that support doubles). I have been focusing on other things
> > first (basically getting all of the float builtins implemented so that
> > cppamp-driver-ng, clBLAS, clFFT, and their unit/precision tests start
> > working).
> >
> >>> +#include <clc/math/unary_intrin.inc>
> >>> +#undef __CLC_FUNCTION
> >>> +#undef __CLC_INTRINSIC
> >>> +
> >>> +_CLC_OVERLOAD _CLC_DEF double cospi(double x) {
> >>> + return __clc_cos_intrinsic(x*M_PI);
> >>> +}
> >>> +
> >>> +_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, double, cospi, double);
> >>> +
> >>> +#endif
> >>> diff --git a/generic/lib/math/sincospiF_piby4.h b/generic/lib/math/sincospiF_piby4.h
> >>> new file mode 100644
> >>> index 0000000..3274d90
> >>> --- /dev/null
> >>> +++ b/generic/lib/math/sincospiF_piby4.h
> >>> @@ -0,0 +1,57 @@
> >>> +/*
> >>> + * 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]
> >>> +inline float2
> >>> +sincosf_piby4(float x)
> >>
> >>
> >> Would it be possible to start prefixing the internal functions like these with some standard prefix (__libclc_ or something like that). Just the be sure they do not clash with the user’s namespace?
> >
> > Sounds reasonable.
> >
> >>
> >> Also by C99’s definition of inline, you would also need to provide a non-inlined variant to make sure linking always succeeds; did you also meant to add __attribute__((always_inline)) ?
> >
> > Yeah, I did mean to add the attribute. Instead of coding that
> > directly, I'd like to use the _CLC_INLINE macro from <clc/clcfunc.h>.
> > Sound good?
>
> Sounds good to me.
>
> Jeroen
>
> >
> > --Aaron
> >
> >>
> >> Jeroen
> >>
> >>
> >>> +{
> >>> + // 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;
> >>> +}
> >>> +
> >>> --
> >>> 2.1.0
> >>>
> >>>
> >>> _______________________________________________
> >>> Libclc-dev mailing list
> >>> Libclc-dev at pcc.me.uk
> >>> http://www.pcc.me.uk/cgi-bin/mailman/listinfo/libclc-dev
>
>
> _______________________________________________
> Libclc-dev mailing list
> Libclc-dev at pcc.me.uk
> http://www.pcc.me.uk/cgi-bin/mailman/listinfo/libclc-dev
More information about the Libclc-dev
mailing list