[Libclc-dev] [PATCH 1/1] Implement log10

[Jan Vesely]

On Tue, 2014-10-21 at 17:05 -0700, Matt Arsenault wrote:
> On 10/21/2014 05:00 PM, Jan Vesely wrote:
> > Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>
> > ---
> >
> > Fixes myocyte rodinia benchmark.
> > piglit test posted
> >
> >   generic/include/clc/clc.h        |  1 +
> >   generic/include/clc/math/log10.h |  9 +++++++++
> >   generic/lib/SOURCES              |  1 +
> >   generic/lib/math/log10.cl        |  8 ++++++++
> >   generic/lib/math/log10.inc       | 10 ++++++++++
> >   5 files changed, 29 insertions(+)
> >   create mode 100644 generic/include/clc/math/log10.h
> >   create mode 100644 generic/lib/math/log10.cl
> >   create mode 100644 generic/lib/math/log10.inc
> >
> > diff --git a/generic/include/clc/clc.h b/generic/include/clc/clc.h
> > index bd92fdb..29c1855 100644
> > --- a/generic/include/clc/clc.h
> > +++ b/generic/include/clc/clc.h
> > @@ -50,6 +50,7 @@
> >   #include <clc/math/fmod.h>
> >   #include <clc/math/hypot.h>
> >   #include <clc/math/log.h>
> > +#include <clc/math/log10.h>
> >   #include <clc/math/log1p.h>
> >   #include <clc/math/log2.h>
> >   #include <clc/math/mad.h>
> > diff --git a/generic/include/clc/math/log10.h b/generic/include/clc/math/log10.h
> > new file mode 100644
> > index 0000000..ec4e4ae
> > --- /dev/null
> > +++ b/generic/include/clc/math/log10.h
> > @@ -0,0 +1,9 @@
> > +#undef log10
> > +
> > +#define __CLC_BODY <clc/math/unary_decl.inc>
> > +#define __CLC_FUNCTION log10
> > +
> > +#include <clc/math/gentype.inc>
> > +
> > +#undef __CLC_BODY
> > +#undef __CLC_FUNCTION
> > diff --git a/generic/lib/SOURCES b/generic/lib/SOURCES
> > index b76fec9..1032606 100644
> > --- a/generic/lib/SOURCES
> > +++ b/generic/lib/SOURCES
> > @@ -63,6 +63,7 @@ math/fmax.cl
> >   math/fmin.cl
> >   math/fmod.cl
> >   math/hypot.cl
> > +math/log10.cl
> >   math/log1p.cl
> >   math/mad.cl
> >   math/mix.cl
> > diff --git a/generic/lib/math/log10.cl b/generic/lib/math/log10.cl
> > new file mode 100644
> > index 0000000..d65764a
> > --- /dev/null
> > +++ b/generic/lib/math/log10.cl
> > @@ -0,0 +1,8 @@
> > +#include <clc/clc.h>
> > +
> > +#ifdef cl_khr_fp64
> > +#pragma OPENCL EXTENSION cl_khr_fp64 : enable
> > +#endif
> > +
> > +#define __CLC_BODY <log10.inc>
> > +#include <clc/math/gentype.inc>
> > diff --git a/generic/lib/math/log10.inc b/generic/lib/math/log10.inc
> > new file mode 100644
> > index 0000000..82e77bc
> > --- /dev/null
> > +++ b/generic/lib/math/log10.inc
> > @@ -0,0 +1,10 @@
> > +_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE log10(__CLC_GENTYPE val) {
> > +  // log10(x) = log2(x) / log2(10)
> > +#if __CLC_FPSIZE == 32
> > +  return log2(val) / log2(10.0f);
> > +#elif __CLC_FPSIZE == 64
> > +  return exp2(val) / log2(10.0);
             ^^^^ should have been log2
> > +#else
> > +#error unknown _CLC_FPSIZE
> > +#endif
> > +}
> 
> Last time I tried it, these functions failed the OpenCL conformance 
> test. Have you considered porting from the amd-builtins branch instead?

is precision the problem?

I could not find any information about relative errors of hw
instructions. I thought that _IEEE instructions produced correctly
rounded results.
I'm still not clear about how the relative error propagates across
operations:
if we have log1p implemented ( with error <= 2ulp ) would
log(x) = log1p(x-1.0) result in error <= 2 ulp since subtraction is
correctly rounded?
if so would log(x) / LOG_2 (correctly rounded constant) give correct
log2 (error <= 3 ulp)?

jan

> 
> 

-- 
Jan Vesely <jan.vesely at rutgers.edu>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: This is a digitally signed message part
URL: <http://lists.llvm.org/pipermail/libclc-dev/attachments/20141022/22f8f9ce/attachment.sig>