[Libclc-dev] [PATCH v2 16/16] native_powr: Switch implementation to native_exp2 and native_log2

[Jan Vesely via Libclc-dev]

On Tue, 2017-11-14 at 21:42 +0100, Jeroen Ketema via Libclc-dev wrote:
> Hi Jan,
> 
> Thanks for the clarification. Looking back at the code
> 
> const __CLC_GENTYPE zero = (__CLC_GENTYPE)0.0f;
> const __CLC_GENTYPE res = native_exp2(y * native_log2(x));
> const __CLC_INTN condnan = isless(x, zero);
> return select(res, (__CLC_GENTYPE)NAN, condnan);
> 
> and with your comments in mind, can’t this be simplified to:
> 
> return native_exp2(y * native_log2(x));
> 
> ?
> 
> If x < 0, then native_log2 yields NaN, and the NaN propagates through the operations.

yeah, that works as well. I guess the special cases were also crafted
for this optimization.

Jan

> 
> Jeroen
> 
> > On 13 Nov 2017, at 21:09, Jan Vesely <jan.vesely at rutgers.edu> wrote:
> > 
> > On Mon, 2017-11-13 at 20:50 +0100, Jeroen Ketema wrote:
> > > This one is not immediately obvious to me. Maybe add some appropriate comment?
> > 
> > it exploits that x == 2^{log2 x} and log2 {x^y} == y * log2 x
> > thus x^y == 2^{log2 x^y} == 2^{y * log2 x}.
> > However, it only works for x > 0, which is OK since powr(<0, ...)
> > should return NaN anyway.
> > IMO powr was added to allow this kind of expansion instead of more
> > expensive full pow (which needs to handle negative x).
> > 
> > Jan
> > 
> > > 
> > > Jeroen
> > > 
> > > > On 6 Nov 2017, at 23:15, Jan Vesely via Libclc-dev <libclc-dev at lists.llvm.org> wrote:
> > > > 
> > > > Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>
> > > > ---
> > > > sent wrong patch the first time. This one passes the CTS.
> > > > 
> > > > Jan
> > > > 
> > > > generic/include/clc/math/native_powr.h | 8 +++++++-
> > > > generic/lib/SOURCES                    | 1 +
> > > > generic/lib/math/native_powr.cl        | 5 +++++
> > > > generic/lib/math/native_powr.inc       | 6 ++++++
> > > > 4 files changed, 19 insertions(+), 1 deletion(-)
> > > > create mode 100644 generic/lib/math/native_powr.cl
> > > > create mode 100644 generic/lib/math/native_powr.inc
> > > > 
> > > > diff --git a/generic/include/clc/math/native_powr.h b/generic/include/clc/math/native_powr.h
> > > > index e8a37d9..c31161a 100644
> > > > --- a/generic/include/clc/math/native_powr.h
> > > > +++ b/generic/include/clc/math/native_powr.h
> > > > @@ -1 +1,7 @@
> > > > -#define native_powr pow
> > > > +#define __CLC_BODY <clc/math/binary_decl_tt.inc>
> > > > +#define __CLC_FUNCTION native_powr
> > > > +
> > > > +#include <clc/math/gentype.inc>
> > > > +
> > > > +#undef __CLC_BODY
> > > > +#undef __CLC_FUNCTION
> > > > diff --git a/generic/lib/SOURCES b/generic/lib/SOURCES
> > > > index 355741c..ad5a743 100644
> > > > --- a/generic/lib/SOURCES
> > > > +++ b/generic/lib/SOURCES
> > > > @@ -127,6 +127,7 @@ math/native_exp2.cl
> > > > math/native_log.cl
> > > > math/native_log10.cl
> > > > math/native_log2.cl
> > > > +math/native_powr.cl
> > > > math/native_recip.cl
> > > > math/native_rsqrt.cl
> > > > math/native_sin.cl
> > > > diff --git a/generic/lib/math/native_powr.cl b/generic/lib/math/native_powr.cl
> > > > new file mode 100644
> > > > index 0000000..452bc6f
> > > > --- /dev/null
> > > > +++ b/generic/lib/math/native_powr.cl
> > > > @@ -0,0 +1,5 @@
> > > > +#include <clc/clc.h>
> > > > +
> > > > +#define __CLC_BODY <native_powr.inc>
> > > > +#define __FLOAT_ONLY
> > > > +#include <clc/math/gentype.inc>
> > > > diff --git a/generic/lib/math/native_powr.inc b/generic/lib/math/native_powr.inc
> > > > new file mode 100644
> > > > index 0000000..841b1ff
> > > > --- /dev/null
> > > > +++ b/generic/lib/math/native_powr.inc
> > > > @@ -0,0 +1,6 @@
> > > > +_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE native_powr(__CLC_GENTYPE x, __CLC_GENTYPE y) {
> > > > +  const __CLC_GENTYPE zero = (__CLC_GENTYPE)0.0f;
> > > > +  const __CLC_GENTYPE res = native_exp2(y * native_log2(x));
> > > > +  const __CLC_INTN condnan = isless(x, zero);
> > > > +  return select(res, (__CLC_GENTYPE)NAN, condnan);
> > > > +}
> > > > -- 
> > > > 2.13.6
> > > > 
> > > > _______________________________________________
> > > > Libclc-dev mailing list
> > > > Libclc-dev at lists.llvm.org
> > > > http://lists.llvm.org/cgi-bin/mailman/listinfo/libclc-dev
> > > 
> > > 
> 
> _______________________________________________
> Libclc-dev mailing list
> Libclc-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/libclc-dev
-------------- 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/20171114/955184f5/attachment.sig>