[Libclc-dev] [PATCH v3 1/1] math: Don't use llvm intrinsic for pow

[Aaron Watry via Libclc-dev]

On Wed, Sep 27, 2017 at 3:29 PM, Jan Vesely via Libclc-dev
<libclc-dev at lists.llvm.org> wrote:
> the intrinsic does not work for fp64
> amdgpu targets expand the fp32 intrinsic into exp2(mul(log2)) anyway.
>
> v2: drop leftover development code
> v3: enable cl_khr_fp64 in gentype.inc
>
> Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>

Looks ok to me. I verified that the same list of inputs fail, although
now, we tend to generate positive NaN instead of negative NaN for
these invalid results... But given that they're invalid anyway, I
can't say that this is a bad thing.

I verified that the double version generates code that doesn't crash
out llvm, and there's roughly the same number of test failures in the
double variant due to invalid results (probably the same corner cases
being tested for double precision inputs).

--Aaron

> ---
> I really should wait for travis to complete all builds before sending
> out patches.
> Sorry for the noise.
>
> Jan
>
>  generic/include/clc/math/gentype.inc | 1 +
>  generic/include/clc/math/pow.h       | 8 ++------
>  generic/include/clc/math/pow.inc     | 1 +
>  generic/lib/SOURCES                  | 1 +
>  generic/lib/math/pow.cl              | 4 ++++
>  generic/lib/math/pow.inc             | 6 ++++++
>  6 files changed, 15 insertions(+), 6 deletions(-)
>  create mode 100644 generic/include/clc/math/pow.inc
>  create mode 100644 generic/lib/math/pow.cl
>  create mode 100644 generic/lib/math/pow.inc
>
> diff --git a/generic/include/clc/math/gentype.inc b/generic/include/clc/math/gentype.inc
> index e6ffad1..37a98e6 100644
> --- a/generic/include/clc/math/gentype.inc
> +++ b/generic/include/clc/math/gentype.inc
> @@ -54,6 +54,7 @@
>
>  #ifndef __FLOAT_ONLY
>  #ifdef cl_khr_fp64
> +#pragma OPENCL EXTENSION cl_khr_fp64:enable
>  #define __CLC_SCALAR_GENTYPE double
>  #define __CLC_FPSIZE 64
>
> diff --git a/generic/include/clc/math/pow.h b/generic/include/clc/math/pow.h
> index 320d341..a11dc05 100644
> --- a/generic/include/clc/math/pow.h
> +++ b/generic/include/clc/math/pow.h
> @@ -1,6 +1,2 @@
> -#undef pow
> -#define pow __clc_pow
> -
> -#define __CLC_FUNCTION __clc_pow
> -#define __CLC_INTRINSIC "llvm.pow"
> -#include <clc/math/binary_intrin.inc>
> +#define __CLC_BODY <clc/math/pow.inc>
> +#include <clc/math/gentype.inc>
> diff --git a/generic/include/clc/math/pow.inc b/generic/include/clc/math/pow.inc
> new file mode 100644
> index 0000000..1af72c1
> --- /dev/null
> +++ b/generic/include/clc/math/pow.inc
> @@ -0,0 +1 @@
> +_CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE pow(__CLC_GENTYPE x, __CLC_GENTYPE y);
> diff --git a/generic/lib/SOURCES b/generic/lib/SOURCES
> index f919bc7..285e2f4 100644
> --- a/generic/lib/SOURCES
> +++ b/generic/lib/SOURCES
> @@ -130,6 +130,7 @@ math/native_log2.cl
>  math/tables.cl
>  math/clc_nextafter.cl
>  math/nextafter.cl
> +math/pow.cl
>  math/pown.cl
>  math/sin.cl
>  math/sincos.cl
> diff --git a/generic/lib/math/pow.cl b/generic/lib/math/pow.cl
> new file mode 100644
> index 0000000..884e335
> --- /dev/null
> +++ b/generic/lib/math/pow.cl
> @@ -0,0 +1,4 @@
> +#include <clc/clc.h>
> +
> +#define __CLC_BODY <pow.inc>
> +#include <clc/math/gentype.inc>
> diff --git a/generic/lib/math/pow.inc b/generic/lib/math/pow.inc
> new file mode 100644
> index 0000000..e1030d7
> --- /dev/null
> +++ b/generic/lib/math/pow.inc
> @@ -0,0 +1,6 @@
> +// TODO: enable when fp16 exp2 and log2 are implemented
> +#if __CLC_FPSIZE > 16
> +_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE pow(__CLC_GENTYPE x, __CLC_GENTYPE y) {
> +       return exp2(y * log2(x));
> +}
> +#endif
> --
> 2.13.5
>
> _______________________________________________
> Libclc-dev mailing list
> Libclc-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/libclc-dev