[Libclc-dev] [PATCH v2 1/2] tan: Port from amd_builtins
Aaron Watry via Libclc-dev
libclc-dev at lists.llvm.org
Thu Jan 18 18:50:47 PST 2018
On Thu, Jan 18, 2018 at 6:15 PM, Jan Vesely via Libclc-dev
<libclc-dev at lists.llvm.org> wrote:
> v2: fixup constant precision
> Passes piglit on turks and carrizo.
> Passes CTS on carrizo
> Fixes half_tan to pass CTS on carrizo
>
> Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>
> ---
> generic/include/math/clc_tan.h | 5 +++
> generic/lib/SOURCES | 1 +
> generic/lib/math/clc_sw_unary.inc | 9 +++++
> generic/lib/math/clc_tan.cl | 71 ++++++++++++++++++++++++++++++++++++++
> generic/lib/math/sincosD_piby4.h | 55 +++++++++++++++++++++++++++++
> generic/lib/math/sincos_helpers.cl | 17 +++++++++
> generic/lib/math/sincos_helpers.h | 3 ++
> generic/lib/math/tan.cl | 7 ++--
> generic/lib/math/tan.inc | 17 ---------
> 9 files changed, 164 insertions(+), 21 deletions(-)
> create mode 100644 generic/include/math/clc_tan.h
> create mode 100644 generic/lib/math/clc_sw_unary.inc
> create mode 100644 generic/lib/math/clc_tan.cl
> delete mode 100644 generic/lib/math/tan.inc
>
> diff --git a/generic/include/math/clc_tan.h b/generic/include/math/clc_tan.h
> new file mode 100644
> index 0000000..bc97902
> --- /dev/null
> +++ b/generic/include/math/clc_tan.h
> @@ -0,0 +1,5 @@
> +#define __CLC_FUNCTION __clc_tan
> +#define __CLC_BODY <clc/math/unary_decl.inc>
> +#include <clc/math/gentype.inc>
> +#undef __CLC_BODY
> +#undef __CLC_FUNCTION
> diff --git a/generic/lib/SOURCES b/generic/lib/SOURCES
> index 4c319a5..624180f 100644
> --- a/generic/lib/SOURCES
> +++ b/generic/lib/SOURCES
> @@ -166,6 +166,7 @@ math/sinh.cl
> math/sinpi.cl
> math/clc_sqrt.cl
> math/sqrt.cl
> +math/clc_tan.cl
> math/tan.cl
> math/tanh.cl
> math/tgamma.cl
> diff --git a/generic/lib/math/clc_sw_unary.inc b/generic/lib/math/clc_sw_unary.inc
> new file mode 100644
> index 0000000..0f8467c
> --- /dev/null
> +++ b/generic/lib/math/clc_sw_unary.inc
> @@ -0,0 +1,9 @@
> +#include <utils.h>
> +
> +#define __CLC_SW_FUNC(x) __CLC_CONCAT(__clc_, x)
> +
> +_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNC(__CLC_GENTYPE x) {
> + return __CLC_SW_FUNC(__CLC_FUNC)(x);
> +}
> +
> +#undef __CLC_SW_FUNC
> diff --git a/generic/lib/math/clc_tan.cl b/generic/lib/math/clc_tan.cl
> new file mode 100644
> index 0000000..b7ad8f8
> --- /dev/null
> +++ b/generic/lib/math/clc_tan.cl
> @@ -0,0 +1,71 @@
> +/*
> + * 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 "../clcmacro.h"
> +#include "tables.h"
> +
> +_CLC_DEF _CLC_OVERLOAD float __clc_tan(float x)
> +{
> + int ix = as_int(x);
> + int ax = ix & 0x7fffffff;
> + float dx = as_float(ax);
> +
> + float r0, r1;
> + int regn = __clc_argReductionS(&r0, &r1, dx);
> +
> + float t = __clc_tanf_piby4(r0 + r1, regn);
> + t = as_float(as_int(t) ^ (ix ^ ax));
> +
> + t = ax >= PINFBITPATT_SP32 ? as_float(QNANBITPATT_SP32) : t;
> + //Take care of subnormals
> + t = (x == 0.0f) ? x : t;
> + return t;
> +}
> +_CLC_UNARY_VECTORIZE(_CLC_DEF _CLC_OVERLOAD, float, __clc_tan, float);
> +
> +#ifdef cl_khr_fp64
> +#include "sincosD_piby4.h"
> +
> +_CLC_DEF _CLC_OVERLOAD double __clc_tan(double x)
> +{
> + double y = fabs(x);
> +
> + double r, rr;
> + int regn;
> +
> + if (y < 0x1.0p+30)
> + __clc_remainder_piby2_medium(y, &r, &rr, ®n);
> + else
Trailing whitespace here.
Otherwise:
Acked-By: Aaron Watry <awatry at gmail.com>
Tested-By: Aaron Watry <awatry at gmail.com>
(Tested on a Radeon RX 580)
I had confirmed some tan(x) wimpy-mode CTS failures before this patch
was applied, they're gone now.
Both tan and tanpi have passed a full CTS run on my system for float and double.
I don't feel sufficiently qualified to review the AMD tan/tanpi
implementations themselves, but given that CTS tests all possible
input values for the float implementation, I think we're ok here.
--Aaron
> + __clc_remainder_piby2_large(y, &r, &rr, ®n);
> +
> + double2 tt = __clc_tan_piby4(r, rr);
> +
> + int2 t = as_int2(regn & 1 ? tt.y : tt.x);
> + t.hi ^= (x < 0.0) << 31;
> +
> + return isnan(x) || isinf(x) ? as_double(QNANBITPATT_DP64) : as_double(t);
> +}
> +_CLC_UNARY_VECTORIZE(_CLC_DEF _CLC_OVERLOAD, double, __clc_tan, double);
> +#endif
> diff --git a/generic/lib/math/sincosD_piby4.h b/generic/lib/math/sincosD_piby4.h
> index a00db85..c98488b 100644
> --- a/generic/lib/math/sincosD_piby4.h
> +++ b/generic/lib/math/sincosD_piby4.h
> @@ -76,3 +76,58 @@ __libclc__sincos_piby4(double x, double xx)
>
> return ret;
> }
> +
> +_CLC_INLINE double2
> +__clc_tan_piby4(double x, double xx)
> +{
> + const double piby4_lead = 7.85398163397448278999e-01; // 0x3fe921fb54442d18
> + const double piby4_tail = 3.06161699786838240164e-17; // 0x3c81a62633145c06
> +
> + // In order to maintain relative precision transform using the identity:
> + // tan(pi/4-x) = (1-tan(x))/(1+tan(x)) for arguments close to pi/4.
> + // Similarly use tan(x-pi/4) = (tan(x)-1)/(tan(x)+1) close to -pi/4.
> +
> + int ca = x > 0.68;
> + int cb = x < -0.68;
> + double transform = ca ? 1.0 : 0.0;
> + transform = cb ? -1.0 : transform;
> +
> + double tx = fma(-transform, x, piby4_lead) + fma(-transform, xx, piby4_tail);
> + int c = ca | cb;
> + x = c ? tx : x;
> + xx = c ? 0.0 : xx;
> +
> + // Core Remez [2,3] approximation to tan(x+xx) on the interval [0,0.68].
> + double t1 = x;
> + double r = fma(2.0, x*xx, x*x);
> +
> + double a = fma(r,
> + fma(r, 0.224044448537022097264602535574e-3, -0.229345080057565662883358588111e-1),
> + 0.372379159759792203640806338901e0);
> +
> + double b = fma(r,
> + fma(r,
> + fma(r, -0.232371494088563558304549252913e-3, 0.260656620398645407524064091208e-1),
> + -0.515658515729031149329237816945e0),
> + 0.111713747927937668539901657944e1);
> +
> + double t2 = fma(MATH_DIVIDE(a, b), x*r, xx);
> +
> + double tp = t1 + t2;
> +
> + // Compute -1.0/(t1 + t2) accurately
> + double z1 = as_double(as_long(tp) & 0xffffffff00000000L);
> + double z2 = t2 - (z1 - t1);
> + double trec = -MATH_RECIP(tp);
> + double trec_top = as_double(as_long(trec) & 0xffffffff00000000L);
> +
> + double tpr = fma(fma(trec_top, z2, fma(trec_top, z1, 1.0)), trec, trec_top);
> +
> + double tpt = transform * (1.0 - MATH_DIVIDE(2.0*tp, 1.0 + tp));
> + double tptr = transform * (MATH_DIVIDE(2.0*tp, tp - 1.0) - 1.0);
> +
> + double2 ret;
> + ret.lo = c ? tpt : tp;
> + ret.hi = c ? tptr : tpr;
> + return ret;
> +}
> diff --git a/generic/lib/math/sincos_helpers.cl b/generic/lib/math/sincos_helpers.cl
> index 251b7f9..3c466bc 100644
> --- a/generic/lib/math/sincos_helpers.cl
> +++ b/generic/lib/math/sincos_helpers.cl
> @@ -90,6 +90,23 @@ _CLC_DEF float __clc_cosf_piby4(float x, float y) {
> return ret;
> }
>
> +_CLC_DEF float __clc_tanf_piby4(float x, int regn)
> +{
> + // Core Remez [1,2] approximation to tan(x) on the interval [0,pi/4].
> + float r = x * x;
> +
> + float a = mad(r, -0.0172032480471481694693109f, 0.385296071263995406715129f);
> +
> + float b = mad(r,
> + mad(r, 0.01844239256901656082986661f, -0.51396505478854532132342f),
> + 1.15588821434688393452299f);
> +
> + float t = mad(x*r, native_divide(a, b), x);
> + float tr = -MATH_RECIP(t);
> +
> + return regn & 1 ? tr : t;
> +}
> +
> _CLC_DEF void __clc_fullMulS(float *hi, float *lo, float a, float b, float bh, float bt)
> {
> if (HAVE_HW_FMA32()) {
> diff --git a/generic/lib/math/sincos_helpers.h b/generic/lib/math/sincos_helpers.h
> index 2565d44..e307abc 100644
> --- a/generic/lib/math/sincos_helpers.h
> +++ b/generic/lib/math/sincos_helpers.h
> @@ -20,8 +20,11 @@
> * THE SOFTWARE.
> */
>
> +#include "clc/clcfunc.h"
> +
> _CLC_DECL float __clc_sinf_piby4(float x, float y);
> _CLC_DECL float __clc_cosf_piby4(float x, float y);
> +_CLC_DECL float __clc_tanf_piby4(float x, int y);
> _CLC_DECL int __clc_argReductionS(float *r, float *rr, float x);
>
> #ifdef cl_khr_fp64
> diff --git a/generic/lib/math/tan.cl b/generic/lib/math/tan.cl
> index a447999..380db67 100644
> --- a/generic/lib/math/tan.cl
> +++ b/generic/lib/math/tan.cl
> @@ -1,8 +1,7 @@
> #include <clc/clc.h>
>
> -#ifdef cl_khr_fp64
> -#pragma OPENCL EXTENSION cl_khr_fp64 : enable
> -#endif
> +#include <math/clc_tan.h>
>
> -#define __CLC_BODY <tan.inc>
> +#define __CLC_FUNC tan
> +#define __CLC_BODY <clc_sw_unary.inc>
> #include <clc/math/gentype.inc>
> diff --git a/generic/lib/math/tan.inc b/generic/lib/math/tan.inc
> deleted file mode 100644
> index b9ce33e..0000000
> --- a/generic/lib/math/tan.inc
> +++ /dev/null
> @@ -1,17 +0,0 @@
> -/*
> - * Note: tan(x) = sin(x)/cos(x) also, but the final assembly ends up being
> - * twice as long for R600 (maybe for others as well).
> - */
> -
> -#if __CLC_FPSIZE == 32
> -#define __CLC_CONST(x) x ## f
> -#else
> -#define __CLC_CONST(x) x
> -#endif
> -
> -_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE tan(__CLC_GENTYPE x) {
> - __CLC_GENTYPE sinx = sin(x);
> - return sinx / sqrt( (__CLC_GENTYPE) __CLC_CONST(1.0) - (sinx*sinx) );
> -}
> -
> -#undef __CLC_CONST
> --
> 2.14.3
>
> _______________________________________________
> Libclc-dev mailing list
> Libclc-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/libclc-dev
More information about the Libclc-dev
mailing list