[Libclc-dev] [PATCH 1/2] frexp: Reuse types provided by gentype.inc
Aaron Watry via Libclc-dev
libclc-dev at lists.llvm.org
Thu Sep 28 19:37:07 PDT 2017
On Thu, Sep 28, 2017 at 9:18 PM, Aaron Watry <awatry at gmail.com> wrote:
> On Fri, Sep 22, 2017 at 8:25 PM, Jan Vesely via Libclc-dev
> <libclc-dev at lists.llvm.org> wrote:
>> Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>
>> ---
>> generic/lib/math/frexp.cl | 14 ++++++---
>> generic/lib/math/frexp.inc | 75 +++++++++++++++++++---------------------------
>> 2 files changed, 40 insertions(+), 49 deletions(-)
>>
>> diff --git a/generic/lib/math/frexp.cl b/generic/lib/math/frexp.cl
>> index acd5d93..b331ff8 100644
>> --- a/generic/lib/math/frexp.cl
>> +++ b/generic/lib/math/frexp.cl
>> @@ -1,10 +1,16 @@
>> #include <clc/clc.h>
>>
>> -#include "math.h"
>> +#define __CLC_BODY <frexp.inc>
>> +#define __CLC_ADDRESS_SPACE private
>> +#include <clc/math/gentype.inc>
>> +#undef __CLC_ADDRESS_SPACE
>>
>> -#ifdef cl_khr_fp64
>> -#pragma OPENCL EXTENSION cl_khr_fp64 : enable
>> -#endif
>
> Pretty sure we need to keep this hunk at the top of the file,
> otherwise the double variants fail to compile for me.
>
Otherwise, this looks ok to me. No changes in CTS pass/fail rate.
We've got a failure in double, but that was there before this patch
for the same input value.
--Aaron
> --Aaron
>
>> +#define __CLC_BODY <frexp.inc>
>> +#define __CLC_ADDRESS_SPACE global
>> +#include <clc/math/gentype.inc>
>> +#undef __CLC_ADDRESS_SPACE
>>
>> #define __CLC_BODY <frexp.inc>
>> +#define __CLC_ADDRESS_SPACE local
>> #include <clc/math/gentype.inc>
>> +#undef __CLC_ADDRESS_SPACE
>> diff --git a/generic/lib/math/frexp.inc b/generic/lib/math/frexp.inc
>> index 0f5ddea..cf3841c 100644
>> --- a/generic/lib/math/frexp.inc
>> +++ b/generic/lib/math/frexp.inc
>> @@ -22,45 +22,44 @@
>> */
>> #if __CLC_FPSIZE == 32
>> #ifdef __CLC_SCALAR
>> -_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE frexp(float x, private int *ep) {
>> - int i = as_int(x);
>> - int ai = i & 0x7fffffff;
>> - int d = ai > 0 & ai < 0x00800000;
>> +_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE frexp(float x, __CLC_ADDRESS_SPACE __CLC_INTN *ep) {
>> + __CLC_INTN i = as_int(x);
>> + __CLC_INTN ai = i & 0x7fffffff;
>> + __CLC_INTN d = ai > 0 & ai < 0x00800000;
>> // scale subnormal by 2^26 without multiplying
>> - float s = as_float(ai | 0x0d800000) - 0x1.0p-100F;
>> + __CLC_GENTYPE s = as_float(ai | 0x0d800000) - 0x1.0p-100F;
>> ai = d ? as_int(s) : ai;
>> - int e = (ai >> 23) - 126 - (d ? 26 : 0);
>> - int t = ai == 0 | e == 129;
>> + __CLC_INTN e = (ai >> 23) - 126 - (d ? 26 : 0);
>> + __CLC_INTN t = ai == 0 | e == 129;
>> i = (i & 0x80000000) | 0x3f000000 | (ai & 0x007fffff);
>> *ep = t ? 0 : e;
>> return t ? x : as_float(i);
>> }
>> +#else
>> #define __CLC_FREXP_VEC(width) \
>> -_CLC_OVERLOAD _CLC_DEF float##width frexp(float##width x, private int##width *ep) { \
>> - int##width i = as_int##width(x); \
>> - int##width ai = i & 0x7fffffff; \
>> - int##width d = ai > 0 & ai < 0x00800000; \
>> +_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE frexp(__CLC_GENTYPE x, __CLC_ADDRESS_SPACE __CLC_INTN *ep) { \
>> + __CLC_INTN i = as_int##width(x); \
>> + __CLC_INTN ai = i & 0x7fffffff; \
>> + __CLC_INTN d = ai > 0 & ai < 0x00800000; \
>> /* scale subnormal by 2^26 without multiplying */ \
>> - float##width s = as_float##width(ai | 0x0d800000) - 0x1.0p-100F; \
>> + __CLC_GENTYPE s = as_float##width(ai | 0x0d800000) - 0x1.0p-100F; \
>> ai = bitselect(ai, as_int##width(s), d); \
>> - int##width e = (ai >> 23) - 126 - bitselect((int##width)0, (int##width)26, d); \
>> - int##width t = ai == (int##width)0 | e == (int##width)129; \
>> - i = (i & (int##width)0x80000000) | (int##width)0x3f000000 | (ai & 0x007fffff); \
>> - *ep = bitselect(e, (int##width)0, t); \
>> + __CLC_INTN e = (ai >> 23) - 126 - bitselect((__CLC_INTN)0, (__CLC_INTN)26, d); \
>> + __CLC_INTN t = ai == (__CLC_INTN)0 | e == (__CLC_INTN)129; \
>> + i = (i & (__CLC_INTN)0x80000000) | (__CLC_INTN)0x3f000000 | (ai & 0x007fffff); \
>> + *ep = bitselect(e, (__CLC_INTN)0, t); \
>> return bitselect(as_float##width(i), x, as_float##width(t)); \
>> }
>> -__CLC_FREXP_VEC(2)
>> -__CLC_FREXP_VEC(3)
>> -__CLC_FREXP_VEC(4)
>> -__CLC_FREXP_VEC(8)
>> -__CLC_FREXP_VEC(16)
>> +#define __CLC_FREXP_VEC2(x) __CLC_FREXP_VEC(x)
>> +__CLC_FREXP_VEC2(__CLC_VECSIZE)
>> +#undef __CLC_FREXP_VEC2
>> #undef __CLC_FREXP_VEC
>> #endif
>> #endif
>>
>> #if __CLC_FPSIZE == 64
>> #ifdef __CLC_SCALAR
>> -_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE frexp(__CLC_GENTYPE x, private __CLC_INTN *ep) {
>> +_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE frexp(__CLC_GENTYPE x, __CLC_ADDRESS_SPACE __CLC_INTN *ep) {
>> long i = as_long(x);
>> long ai = i & 0x7fffffffffffffffL;
>> int d = ai > 0 & ai < 0x0010000000000000L;
>> @@ -73,38 +72,24 @@ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE frexp(__CLC_GENTYPE x, private __CLC_INTN *
>> *ep = t ? 0 : e;
>> return t ? x : as_double(i);
>> }
>> +#else
>> #define __CLC_FREXP_VEC(width) \
>> -_CLC_OVERLOAD _CLC_DEF double##width frexp(double##width x, private int##width *ep) { \
>> +_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE frexp(__CLC_GENTYPE x, __CLC_ADDRESS_SPACE __CLC_INTN *ep) { \
>> long##width i = as_long##width(x); \
>> long##width ai = i & 0x7fffffffffffffffL; \
>> long##width d = ai > 0 & ai < 0x0010000000000000L; \
>> /* scale subnormal by 2^54 without multiplying */ \
>> - double##width s = as_double##width(ai | 0x0370000000000000L) - 0x1.0p-968; \
>> + __CLC_GENTYPE s = as_double##width(ai | 0x0370000000000000L) - 0x1.0p-968; \
>> ai = bitselect(ai, as_long##width(s), d); \
>> - int##width e = convert_int##width(ai >> 52) - 1022 - bitselect((int##width)0, (int##width)54, convert_int##width(d)); \
>> - int##width t = convert_int##width(ai == (long##width)0) | (e == (int##width)129); \
>> + __CLC_INTN e = convert_int##width(ai >> 52) - 1022 - bitselect((__CLC_INTN)0, (__CLC_INTN)54, convert_int##width(d)); \
>> + __CLC_INTN t = convert_int##width(ai == (long##width)0) | (e == (__CLC_INTN)129); \
>> i = (i & (long##width)0x8000000000000000L) | (long##width)0x3fe0000000000000L | (ai & 0x000fffffffffffffL); \
>> - *ep = bitselect(e, (int##width)0, t); \
>> + *ep = bitselect(e, (__CLC_INTN)0, t); \
>> return bitselect(as_double##width(i), x, as_double##width(convert_long##width(t))); \
>> }
>> -__CLC_FREXP_VEC(2)
>> -__CLC_FREXP_VEC(3)
>> -__CLC_FREXP_VEC(4)
>> -__CLC_FREXP_VEC(8)
>> -__CLC_FREXP_VEC(16)
>> +#define __CLC_FREXP_VEC2(x) __CLC_FREXP_VEC(x)
>> +__CLC_FREXP_VEC2(__CLC_VECSIZE)
>> +#undef __CLC_FREXP_VEC2
>> #undef __CLC_FREXP_VEC
>> #endif
>> #endif
>> -
>> -#define __CLC_FREXP_DEF(addrspace) \
>> - _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE frexp(__CLC_GENTYPE x, addrspace __CLC_INTN *iptr) { \
>> - __CLC_INTN private_iptr; \
>> - __CLC_GENTYPE ret = frexp(x, &private_iptr); \
>> - *iptr = private_iptr; \
>> - return ret; \
>> -}
>> -
>> -__CLC_FREXP_DEF(local);
>> -__CLC_FREXP_DEF(global);
>> -
>> -#undef __CLC_FREXP_DEF
>> --
>> 2.13.5
>>
>> _______________________________________________
>> Libclc-dev mailing list
>> Libclc-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/libclc-dev
More information about the Libclc-dev
mailing list