[Libclc-dev] [PATCH] libclc/asin: Switch to amd builtins version of asin
Jan Vesely via Libclc-dev
libclc-dev at lists.llvm.org
Mon Feb 3 19:41:16 PST 2020
(re-adding libclc-dev)
On Mon, 2020-02-03 at 20:37 -0600, Aaron Watry wrote:
> On Mon, Feb 3, 2020 at 4:49 PM Jan Vesely <jan.vesely at rutgers.edu> wrote:
> > On Sun, 2020-02-02 at 18:23 -0600, Aaron Watry wrote:
> > > On Sun, Feb 2, 2020 at 2:07 PM Jan Vesely <jan.vesely at rutgers.edu> wrote:
> > > > On Sat, 2020-02-01 at 19:54 -0600, Aaron Watry via Libclc-dev wrote:
> > > > > Fixes a wimpy-mode CTS failure for asin(float).
> > > > >
> > > > > Passes non-wimpy for both float/double on RX580.
> > > > >
> > > > > Signed-off-by: Aaron Watry <awatry at gmail.com>
> > > >
> > > > Nice, thanks!
> > > > I presume the failure appeared on llvm-11? the test passes ok on llvm-
> > > > 10 on my Carrizo machine. Although this patch reduces max ulp from
> > > > 2.91 to 0.91. It also fixes the same test on Turks.
> > > > Tested-by: Jan Vesely <jan.vesely at rutgers.edu>
> > > >
> > >
> > > Actually the issue was present on llvm 9.0,0, which is what my current
> > > stack has installed.
> >
> > That's interesting. I wonder whether it's because of different code
> > generated by LLVM or different accuracy of used instructions.
> > Does acos pass on your setup?
>
> Yup, acos passes for both float/double.
>
> I do have a crash in cos(double), but I planned on dealing with that separately.
>
> With that done, other than a failure to build some kernel related to a
> NOT operation, all of the math_brute_force CTS tests at least pass
> wimpy mode for me.
>
> Did you want me to hold off pushing the acos fix for a bit, or treat
> any other follow-up work as a separate thing?
No, please push asap.
math_brute_force has been passing for some time on my carrizo setup,
that's why I was surprised to see the failure.
Scalar versions of the tests even pass on Turks/EG, with the exception
of asin/acos and anythin using divide or sqrt.
Jan
>
> --Aaron
>
> > Jan
> >
> > > I'll get rid of the TODO comment before pushing.
> > >
> > > --Aaron
> > >
> > > > One small nit below, otherwise
> > > > Reviewed-by: Jan Vesely <jan.vesely at rutgers.edu>
> > > >
> > > > > ---
> > > > > libclc/generic/lib/math/asin.cl | 168 ++++++++++++++++++++++++++++++-
> > > > > libclc/generic/lib/math/asin.inc | 18 ----
> > > > > 2 files changed, 166 insertions(+), 20 deletions(-)
> > > > > delete mode 100644 libclc/generic/lib/math/asin.inc
> > > > >
> > > > > diff --git a/libclc/generic/lib/math/asin.cl b/libclc/generic/lib/math/asin.cl
> > > > > index 7581a504093..e07dd298d9c 100644
> > > > > --- a/libclc/generic/lib/math/asin.cl
> > > > > +++ b/libclc/generic/lib/math/asin.cl
> > > > > @@ -1,4 +1,168 @@
> > > > > +/*
> > > > > + * 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.
> > > > > + */
> > > > > +
> > > > > +// TODO: Enable half precision when atan2 is implemented
> > > >
> > > > The above TODO seems redundant since you switched the implementation
> > > > away from atan2.
> > > >
> > > > > #include <clc/clc.h>
> > > > >
> > > > > -#define __CLC_BODY <asin.inc>
> > > > > -#include <clc/math/gentype.inc>
> > > > > +#include "math.h"
> > > > > +#include "../clcmacro.h"
> > > > > +
> > > > > +_CLC_OVERLOAD _CLC_DEF float asin(float x) {
> > > > > + // Computes arcsin(x).
> > > > > + // The argument is first reduced by noting that arcsin(x)
> > > > > + // is invalid for abs(x) > 1 and arcsin(-x) = -arcsin(x).
> > > > > + // For denormal and small arguments arcsin(x) = x to machine
> > > > > + // accuracy. Remaining argument ranges are handled as follows.
> > > > > + // For abs(x) <= 0.5 use
> > > > > + // arcsin(x) = x + x^3*R(x^2)
> > > > > + // where R(x^2) is a rational minimax approximation to
> > > > > + // (arcsin(x) - x)/x^3.
> > > > > + // For abs(x) > 0.5 exploit the identity:
> > > > > + // arcsin(x) = pi/2 - 2*arcsin(sqrt(1-x)/2)
> > > > > + // together with the above rational approximation, and
> > > > > + // reconstruct the terms carefully.
> > > > > +
> > > > > + const float piby2_tail = 7.5497894159e-08F; /* 0x33a22168 */
> > > > > + const float hpiby2_head = 7.8539812565e-01F; /* 0x3f490fda */
> > > > > + const float piby2 = 1.5707963705e+00F; /* 0x3fc90fdb */
> > > > > +
> > > > > + uint ux = as_uint(x);
> > > > > + uint aux = ux & EXSIGNBIT_SP32;
> > > > > + uint xs = ux ^ aux;
> > > > > + float spiby2 = as_float(xs | as_uint(piby2));
> > > > > + int xexp = (int)(aux >> EXPSHIFTBITS_SP32) - EXPBIAS_SP32;
> > > > > + float y = as_float(aux);
> > > > > +
> > > > > + // abs(x) >= 0.5
> > > > > + int transform = xexp >= -1;
> > > > > +
> > > > > + float y2 = y * y;
> > > > > + float rt = 0.5f * (1.0f - y);
> > > > > + float r = transform ? rt : y2;
> > > > > +
> > > > > + // Use a rational approximation for [0.0, 0.5]
> > > > > + float a = mad(r,
> > > > > + mad(r,
> > > > > + mad(r, -0.00396137437848476485201154797087F, -0.0133819288943925804214011424456F),
> > > > > + -0.0565298683201845211985026327361F),
> > > > > + 0.184161606965100694821398249421F);
> > > > > +
> > > > > + float b = mad(r, -0.836411276854206731913362287293F, 1.10496961524520294485512696706F);
> > > > > + float u = r * MATH_DIVIDE(a, b);
> > > > > +
> > > > > + float s = MATH_SQRT(r);
> > > > > + float s1 = as_float(as_uint(s) & 0xffff0000);
> > > > > + float c = MATH_DIVIDE(mad(-s1, s1, r), s + s1);
> > > > > + float p = mad(2.0f*s, u, -mad(c, -2.0f, piby2_tail));
> > > > > + float q = mad(s1, -2.0f, hpiby2_head);
> > > > > + float vt = hpiby2_head - (p - q);
> > > > > + float v = mad(y, u, y);
> > > > > + v = transform ? vt : v;
> > > > > +
> > > > > + float ret = as_float(xs | as_uint(v));
> > > > > + ret = aux > 0x3f800000U ? as_float(QNANBITPATT_SP32) : ret;
> > > > > + ret = aux == 0x3f800000U ? spiby2 : ret;
> > > > > + ret = xexp < -14 ? x : ret;
> > > > > +
> > > > > + return ret;
> > > > > +}
> > > > > +
> > > > > +_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, float, asin, float);
> > > > > +
> > > > > +#ifdef cl_khr_fp64
> > > > > +
> > > > > +#pragma OPENCL EXTENSION cl_khr_fp64 : enable
> > > > > +
> > > > > +_CLC_OVERLOAD _CLC_DEF double asin(double x) {
> > > > > + // Computes arcsin(x).
> > > > > + // The argument is first reduced by noting that arcsin(x)
> > > > > + // is invalid for abs(x) > 1 and arcsin(-x) = -arcsin(x).
> > > > > + // For denormal and small arguments arcsin(x) = x to machine
> > > > > + // accuracy. Remaining argument ranges are handled as follows.
> > > > > + // For abs(x) <= 0.5 use
> > > > > + // arcsin(x) = x + x^3*R(x^2)
> > > > > + // where R(x^2) is a rational minimax approximation to
> > > > > + // (arcsin(x) - x)/x^3.
> > > > > + // For abs(x) > 0.5 exploit the identity:
> > > > > + // arcsin(x) = pi/2 - 2*arcsin(sqrt(1-x)/2)
> > > > > + // together with the above rational approximation, and
> > > > > + // reconstruct the terms carefully.
> > > > > +
> > > > > + const double piby2_tail = 6.1232339957367660e-17; /* 0x3c91a62633145c07 */
> > > > > + const double hpiby2_head = 7.8539816339744831e-01; /* 0x3fe921fb54442d18 */
> > > > > + const double piby2 = 1.5707963267948965e+00; /* 0x3ff921fb54442d18 */
> > > > > +
> > > > > + double y = fabs(x);
> > > > > + int xneg = as_int2(x).hi < 0;
> > > > > + int xexp = (as_int2(y).hi >> 20) - EXPBIAS_DP64;
> > > > > +
> > > > > + // abs(x) >= 0.5
> > > > > + int transform = xexp >= -1;
> > > > > +
> > > > > + double rt = 0.5 * (1.0 - y);
> > > > > + double y2 = y * y;
> > > > > + double r = transform ? rt : y2;
> > > > > +
> > > > > + // Use a rational approximation for [0.0, 0.5]
> > > > > +
> > > > > + double un = fma(r,
> > > > > + fma(r,
> > > > > + fma(r,
> > > > > + fma(r,
> > > > > + fma(r, 0.0000482901920344786991880522822991,
> > > > > + 0.00109242697235074662306043804220),
> > > > > + -0.0549989809235685841612020091328),
> > > > > + 0.275558175256937652532686256258),
> > > > > + -0.445017216867635649900123110649),
> > > > > + 0.227485835556935010735943483075);
> > > > > +
> > > > > + double ud = fma(r,
> > > > > + fma(r,
> > > > > + fma(r,
> > > > > + fma(r, 0.105869422087204370341222318533,
> > > > > + -0.943639137032492685763471240072),
> > > > > + 2.76568859157270989520376345954),
> > > > > + -3.28431505720958658909889444194),
> > > > > + 1.36491501334161032038194214209);
> > > > > +
> > > > > + double u = r * MATH_DIVIDE(un, ud);
> > > > > +
> > > > > + // Reconstruct asin carefully in transformed region
> > > > > + double s = sqrt(r);
> > > > > + double sh = as_double(as_ulong(s) & 0xffffffff00000000UL);
> > > > > + double c = MATH_DIVIDE(fma(-sh, sh, r), s + sh);
> > > > > + double p = fma(2.0*s, u, -fma(-2.0, c, piby2_tail));
> > > > > + double q = fma(-2.0, sh, hpiby2_head);
> > > > > + double vt = hpiby2_head - (p - q);
> > > > > + double v = fma(y, u, y);
> > > > > + v = transform ? vt : v;
> > > > > +
> > > > > + v = xexp < -28 ? y : v;
> > > > > + v = xexp >= 0 ? as_double(QNANBITPATT_DP64) : v;
> > > > > + v = y == 1.0 ? piby2 : v;
> > > > > +
> > > > > + return xneg ? -v : v;
> > > > > +}
> > > > > +
> > > > > +_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, double, asin, double);
> > > > > +
> > > > > +#endif // cl_khr_fp64
> > > > > diff --git a/libclc/generic/lib/math/asin.inc b/libclc/generic/lib/math/asin.inc
> > > > > deleted file mode 100644
> > > > > index b08c7bd29ce..00000000000
> > > > > --- a/libclc/generic/lib/math/asin.inc
> > > > > +++ /dev/null
> > > > > @@ -1,18 +0,0 @@
> > > > > -// TODO: Enable half precision when atan2 is implemented
> > > > > -#if __CLC_FPSIZE > 16
> > > > > -
> > > > > -#if __CLC_FPSIZE == 64
> > > > > -#define __CLC_CONST(x) x
> > > > > -#elif __CLC_FPSIZE == 32
> > > > > -#define __CLC_CONST(x) x ## f
> > > > > -#elif __CLC_FPSIZE == 16
> > > > > -#define __CLC_CONST(x) x ## h
> > > > > -#endif
> > > > > -
> > > > > -_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE asin(__CLC_GENTYPE x) {
> > > > > - return atan2(x, sqrt( (__CLC_GENTYPE)__CLC_CONST(1.0) - (x*x) ));
> > > > > -}
> > > > > -
> > > > > -#undef __CLC_CONST
> > > > > -
> > > > > -#endif
> > > >
> > > > Jan
> > > >
> > > > --
> > > > Jan Vesely <jan.vesely at rutgers.edu>
> >
> > --
> > Jan Vesely <jan.vesely at rutgers.edu>
--
Jan Vesely <jan.vesely at rutgers.edu>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: This is a digitally signed message part
URL: <http://lists.llvm.org/pipermail/libclc-dev/attachments/20200203/9b3644d7/attachment.sig>
More information about the Libclc-dev
mailing list