<br><br><div class="gmail_quote"><div dir="ltr">On Thu, Oct 12, 2017, 2:03 PM Jan Vesely <<a href="mailto:jan.vesely@rutgers.edu">jan.vesely@rutgers.edu</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">On Sun, 2017-10-01 at 13:01 -0400, Jan Vesely wrote:<br>
> On Sun, 2017-10-01 at 13:16 +0000, Aaron Watry wrote:<br>
> > I haven't forgotten about these two... I'm just trying to figure out some<br>
> > ambiguity in both the 1.2/2.0/2.2 spec related to whether a scalar version<br>
> > (vec-size 1 with no numeric suffix) is needed. The CTS tests for<br>
> > vloada_half being supported, while the spec's language changes a bit<br>
> > between versions, and never gets to what I'd call a consistent state.<br>
><br>
> yeah, it was weird that ctx expects a scalar version. since non-aligned<br>
> vload_half/vstore_half expect the pointer to be 16bit aligned, scalar<br>
> vloada_half/vstorea_half would be identical to non-aligned version.<br>
> The specs seem to always mention the aligned variants with 'n' suffix,<br>
> so I just considered it a CTS bug.<br>
><br>
> I don't mind going out of specs and adding a scalar version if you<br>
> think it's useful beyond appeasing the CTS.<br>
<br>
Hi,<br>
<br>
I'm not sure what the final consensus is here. Do you prefer I added<br>
scalar versions of vloada_half/vstorea_half even though they are<br>
identical to non-aligned versions?<br>
<br>
I could not find any support for it in the specs, only the CTS expects<br>
it.<br></blockquote></div><div><br></div><div>I'd say at this point that all conformant implementations already probably support it unless they've all gotten a waiver due to it being a possible spec bug.</div><div><br></div><div>Since it's basically just an alias for vload_half/vstore_half, I wouldn't be against adding it in for now. Maybe eventually we can get some clarification or a test fix put in.</div><div><br></div><div>Aaron</div><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
Jan<br>
<br>
><br>
> Jan<br>
><br>
> ><br>
> > --Aaron<br>
> ><br>
> > On Sun, Sep 24, 2017, 4:03 PM Jan Vesely <<a href="mailto:jan.vesely@rutgers.edu" target="_blank">jan.vesely@rutgers.edu</a>> wrote:<br>
> ><br>
> > > Passes newly posted piglits on Turks.<br>
> > ><br>
> > > Signed-off-by: Jan Vesely <<a href="mailto:jan.vesely@rutgers.edu" target="_blank">jan.vesely@rutgers.edu</a>><br>
> > > ---<br>
> > > generic/include/clc/shared/vload.h | 30 +++++++++++++++---------------<br>
> > > generic/lib/shared/<a href="http://vload.cl" rel="noreferrer" target="_blank">vload.cl</a> | 10 ++++++++--<br>
> > > generic/lib/shared/vload_half.inc | 26 ++++++++++++++++++--------<br>
> > > 3 files changed, 41 insertions(+), 25 deletions(-)<br>
> > ><br>
> > > diff --git a/generic/include/clc/shared/vload.h<br>
> > > b/generic/include/clc/shared/vload.h<br>
> > > index 8c262dd..f6ae917 100644<br>
> > > --- a/generic/include/clc/shared/vload.h<br>
> > > +++ b/generic/include/clc/shared/vload.h<br>
> > > @@ -12,22 +12,24 @@<br>
> > > _CLC_VECTOR_VLOAD_DECL(SUFFIX, MEM_TYPE, PRIM_TYPE, __private) \<br>
> > > _CLC_VECTOR_VLOAD_DECL(SUFFIX, MEM_TYPE, PRIM_TYPE, __local) \<br>
> > > _CLC_VECTOR_VLOAD_DECL(SUFFIX, MEM_TYPE, PRIM_TYPE, __constant) \<br>
> > > - _CLC_VECTOR_VLOAD_DECL(SUFFIX, MEM_TYPE, PRIM_TYPE, __global) \<br>
> > > + _CLC_VECTOR_VLOAD_DECL(SUFFIX, MEM_TYPE, PRIM_TYPE, __global)<br>
> > ><br>
> > > #define _CLC_VECTOR_VLOAD_PRIM1(PRIM_TYPE) \<br>
> > > - _CLC_VECTOR_VLOAD_PRIM3(, PRIM_TYPE, PRIM_TYPE) \<br>
> > > + _CLC_VECTOR_VLOAD_PRIM3(, PRIM_TYPE, PRIM_TYPE)<br>
> > ><br>
> > > -#define _CLC_VECTOR_VLOAD_PRIM() \<br>
> > > - _CLC_VECTOR_VLOAD_PRIM1(char) \<br>
> > > - _CLC_VECTOR_VLOAD_PRIM1(uchar) \<br>
> > > - _CLC_VECTOR_VLOAD_PRIM1(short) \<br>
> > > - _CLC_VECTOR_VLOAD_PRIM1(ushort) \<br>
> > > - _CLC_VECTOR_VLOAD_PRIM1(int) \<br>
> > > - _CLC_VECTOR_VLOAD_PRIM1(uint) \<br>
> > > - _CLC_VECTOR_VLOAD_PRIM1(long) \<br>
> > > - _CLC_VECTOR_VLOAD_PRIM1(ulong) \<br>
> > > - _CLC_VECTOR_VLOAD_PRIM1(float) \<br>
> > > - _CLC_VECTOR_VLOAD_PRIM3(_half, half, float)<br>
> > > +// Declare vector load prototypes<br>
> > > +_CLC_VECTOR_VLOAD_PRIM1(char)<br>
> > > +_CLC_VECTOR_VLOAD_PRIM1(uchar)<br>
> > > +_CLC_VECTOR_VLOAD_PRIM1(short)<br>
> > > +_CLC_VECTOR_VLOAD_PRIM1(ushort)<br>
> > > +_CLC_VECTOR_VLOAD_PRIM1(int)<br>
> > > +_CLC_VECTOR_VLOAD_PRIM1(uint)<br>
> > > +_CLC_VECTOR_VLOAD_PRIM1(long)<br>
> > > +_CLC_VECTOR_VLOAD_PRIM1(ulong)<br>
> > > +_CLC_VECTOR_VLOAD_PRIM1(float)<br>
> > > +_CLC_VECTOR_VLOAD_PRIM3(_half, half, float)<br>
> > > +// Use suffix to declare aligned vloada_halfN<br>
> > > +_CLC_VECTOR_VLOAD_PRIM3(a_half, half, float)<br>
> > ><br>
> > > #ifdef cl_khr_fp64<br>
> > > #pragma OPENCL EXTENSION cl_khr_fp64: enable<br>
> > > @@ -38,7 +40,6 @@<br>
> > > _CLC_VECTOR_VLOAD_PRIM1(half)<br>
> > > #endif<br>
> > ><br>
> > > -_CLC_VECTOR_VLOAD_PRIM()<br>
> > > // Plain vload_half also needs to be declared<br>
> > > _CLC_VLOAD_DECL(_half, half, float, , __constant)<br>
> > > _CLC_VLOAD_DECL(_half, half, float, , __global)<br>
> > > @@ -49,4 +50,3 @@ _CLC_VLOAD_DECL(_half, half, float, , __private)<br>
> > > #undef _CLC_VECTOR_VLOAD_DECL<br>
> > > #undef _CLC_VECTOR_VLOAD_PRIM3<br>
> > > #undef _CLC_VECTOR_VLOAD_PRIM1<br>
> > > -#undef _CLC_VECTOR_VLOAD_PRIM<br>
> > > diff --git a/generic/lib/shared/<a href="http://vload.cl" rel="noreferrer" target="_blank">vload.cl</a> b/generic/lib/shared/<a href="http://vload.cl" rel="noreferrer" target="_blank">vload.cl</a><br>
> > > index 0892270..9c37fcf 100644<br>
> > > --- a/generic/lib/shared/<a href="http://vload.cl" rel="noreferrer" target="_blank">vload.cl</a><br>
> > > +++ b/generic/lib/shared/<a href="http://vload.cl" rel="noreferrer" target="_blank">vload.cl</a><br>
> > > @@ -85,15 +85,21 @@ float __clc_vload_half_float_helper__private(const<br>
> > > __private half *);<br>
> > > VEC_LOAD8(val.lo, AS) \<br>
> > > VEC_LOAD8(val.hi, AS)<br>
> > ><br>
> > > -#define __FUNC(SUFFIX, VEC_SIZE, TYPE, AS) \<br>
> > > +#define __FUNC(SUFFIX, VEC_SIZE, OFFSET_SIZE, TYPE, AS) \<br>
> > > _CLC_OVERLOAD _CLC_DEF TYPE vload_half##SUFFIX(size_t offset, const AS<br>
> > > half *mem) { \<br>
> > > offset *= VEC_SIZE; \<br>
> > > TYPE __tmp; \<br>
> > > VEC_LOAD##VEC_SIZE(__tmp, AS) \<br>
> > > return __tmp; \<br>
> > > + } \<br>
> > > + _CLC_OVERLOAD _CLC_DEF TYPE vloada_half##SUFFIX(size_t offset, const AS<br>
> > > half *mem) { \<br>
> > > + offset *= OFFSET_SIZE; \<br>
> > > + TYPE __tmp; \<br>
> > > + VEC_LOAD##VEC_SIZE(__tmp, AS) \<br>
> > > + return __tmp; \<br>
> > > }<br>
> > ><br>
> > > -#define FUNC(SUFFIX, VEC_SIZE, TYPE, AS) __FUNC(SUFFIX, VEC_SIZE, TYPE,<br>
> > > AS)<br>
> > > +#define FUNC(SUFFIX, VEC_SIZE, OFFSET_SIZE, TYPE, AS) __FUNC(SUFFIX,<br>
> > > VEC_SIZE, OFFSET_SIZE, TYPE, AS)<br>
> > ><br>
> > > #define __CLC_BODY "vload_half.inc"<br>
> > > #include <clc/math/gentype.inc><br>
> > > diff --git a/generic/lib/shared/vload_half.inc<br>
> > > b/generic/lib/shared/vload_half.inc<br>
> > > index 00dae8a..11b2bf7 100644<br>
> > > --- a/generic/lib/shared/vload_half.inc<br>
> > > +++ b/generic/lib/shared/vload_half.inc<br>
> > > @@ -1,13 +1,23 @@<br>
> > > #if __CLC_FPSIZE == 32<br>
> > > +<br>
> > > #ifdef __CLC_VECSIZE<br>
> > > - FUNC(__CLC_VECSIZE, __CLC_VECSIZE, __CLC_GENTYPE, __private);<br>
> > > - FUNC(__CLC_VECSIZE, __CLC_VECSIZE, __CLC_GENTYPE, __local);<br>
> > > - FUNC(__CLC_VECSIZE, __CLC_VECSIZE, __CLC_GENTYPE, __global);<br>
> > > - FUNC(__CLC_VECSIZE, __CLC_VECSIZE, __CLC_GENTYPE, __constant);<br>
> > > +<br>
> > > +#if __CLC_VECSIZE == 3<br>
> > > +# define __CLC_OFFSET 4<br>
> > > #else<br>
> > > - FUNC(, 1, __CLC_GENTYPE, __private);<br>
> > > - FUNC(, 1, __CLC_GENTYPE, __local);<br>
> > > - FUNC(, 1, __CLC_GENTYPE, __global);<br>
> > > - FUNC(, 1, __CLC_GENTYPE, __constant);<br>
> > > +# define __CLC_OFFSET __CLC_VECSIZE<br>
> > > +#endif<br>
> > > +<br>
> > > + FUNC(__CLC_VECSIZE, __CLC_VECSIZE, __CLC_OFFSET, __CLC_GENTYPE,<br>
> > > __private);<br>
> > > + FUNC(__CLC_VECSIZE, __CLC_VECSIZE, __CLC_OFFSET, __CLC_GENTYPE,<br>
> > > __local);<br>
> > > + FUNC(__CLC_VECSIZE, __CLC_VECSIZE, __CLC_OFFSET, __CLC_GENTYPE,<br>
> > > __global);<br>
> > > + FUNC(__CLC_VECSIZE, __CLC_VECSIZE, __CLC_OFFSET, __CLC_GENTYPE,<br>
> > > __constant);<br>
> > > +<br>
> > > +#undef __CLC_OFFSET<br>
> > > +#else<br>
> > > + FUNC(, 1, 1, __CLC_GENTYPE, __private);<br>
> > > + FUNC(, 1, 1, __CLC_GENTYPE, __local);<br>
> > > + FUNC(, 1, 1, __CLC_GENTYPE, __global);<br>
> > > + FUNC(, 1, 1, __CLC_GENTYPE, __constant);<br>
> > > #endif<br>
> > > #endif<br>
> > > --<br>
> > > 2.13.5<br>
> > ><br>
> > ><br>
><br>
><br>
</blockquote></div>