Jan Vesely writes:
> On Thu, 2018-02-15 at 22:50 -0800, Francisco Jerez wrote:
>> Jan Vesely writes:
>>
>> > On Thu, 2018-02-15 at 20:36 -0800, Francisco Jerez wrote:
>> > > Jan Vesely writes:
>> > >
>> > > > On Thu,
On Thu, 2018-02-15 at 22:50 -0800, Francisco Jerez wrote:
> Jan Vesely writes:
>
> > On Thu, 2018-02-15 at 20:36 -0800, Francisco Jerez wrote:
> > > Jan Vesely writes:
> > >
> > > > On Thu, 2018-02-08 at 15:56 -0800, Francisco Jerez wrote:
> > >
Jan Vesely writes:
> On Thu, 2018-02-15 at 20:36 -0800, Francisco Jerez wrote:
>> Jan Vesely writes:
>>
>> > On Thu, 2018-02-08 at 15:56 -0800, Francisco Jerez wrote:
>> > > Jan Vesely writes:
>> > >
>> > > > On Thu,
On Thu, 2018-02-15 at 20:36 -0800, Francisco Jerez wrote:
> Jan Vesely writes:
>
> > On Thu, 2018-02-08 at 15:56 -0800, Francisco Jerez wrote:
> > > Jan Vesely writes:
> > >
> > > > On Thu, 2018-02-08 at 23:16 +0100, Pierre Moreau wrote:
> > > >
Jan Vesely writes:
> On Thu, 2018-02-08 at 15:56 -0800, Francisco Jerez wrote:
>> Jan Vesely writes:
>>
>> > On Thu, 2018-02-08 at 23:16 +0100, Pierre Moreau wrote:
>> > > (Moving the conversation to its own thread.)
>> > >
>> > > > target
On Thu, 2018-02-08 at 15:56 -0800, Francisco Jerez wrote:
> Jan Vesely writes:
>
> > On Thu, 2018-02-08 at 23:16 +0100, Pierre Moreau wrote:
> > > (Moving the conversation to its own thread.)
> > >
> > > > target agnostic libclc is rather difficult to do. CLC includes 3
Jan Vesely writes:
> On Thu, 2018-02-08 at 23:16 +0100, Pierre Moreau wrote:
>> (Moving the conversation to its own thread.)
>>
>> > target agnostic libclc is rather difficult to do. CLC includes 3 levels
>> > of precision on float (fp32) operands; full, half, native.
On Thu, 2018-02-08 at 23:16 +0100, Pierre Moreau wrote:
> (Moving the conversation to its own thread.)
>
> > target agnostic libclc is rather difficult to do. CLC includes 3 levels
> > of precision on float (fp32) operands; full, half, native. The
> > implementation of each depends on capabilites
(Moving the conversation to its own thread.)
> target agnostic libclc is rather difficult to do. CLC includes 3 levels
> of precision on float (fp32) operands; full, half, native. The
> implementation of each depends on capabilites of specific device (e.g.
> vega(VI+?) can do 1 ULP log2/exp2 in