diff options
| author | Jan Vesely <jan.vesely@rutgers.edu> | 2018-02-06 18:44:45 +0000 |
|---|---|---|
| committer | Jan Vesely <jan.vesely@rutgers.edu> | 2018-02-06 18:44:45 +0000 |
| commit | e276f35e65ea804df07a81e9cd4f9212d9c591e1 (patch) | |
| tree | d8fabb8b57ee0df7339902bb9f5ead3fd8fc978b /libclc/generic | |
| parent | 126b643e89a7ded36983d3ef0e91d64668467e7b (diff) | |
Add vstore_half_rtn implementation
Passes CTS on carrizo
Reviewer: Jeroen Ketema <j.ketema@xs4all.nl>
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Diffstat (limited to 'libclc/generic')
| -rw-r--r-- | libclc/generic/include/clc/shared/vstore.h | 2 | ||||
| -rw-r--r-- | libclc/generic/lib/shared/vstore.cl | 42 |
2 files changed, 43 insertions, 1 deletions
diff --git a/libclc/generic/include/clc/shared/vstore.h b/libclc/generic/include/clc/shared/vstore.h index 549e6bfea64..8639be707e5 100644 --- a/libclc/generic/include/clc/shared/vstore.h +++ b/libclc/generic/include/clc/shared/vstore.h @@ -38,11 +38,13 @@ _CLC_VECTOR_VSTORE_PRIM1(float) _CLC_VECTOR_VSTORE_HALF_PRIM1(float,) _CLC_VECTOR_VSTORE_HALF_PRIM1(float, _rtz) +_CLC_VECTOR_VSTORE_HALF_PRIM1(float, _rtn) #ifdef cl_khr_fp64 _CLC_VECTOR_VSTORE_PRIM1(double) _CLC_VECTOR_VSTORE_HALF_PRIM1(double,) _CLC_VECTOR_VSTORE_HALF_PRIM1(double, _rtz) + _CLC_VECTOR_VSTORE_HALF_PRIM1(double, _rtn) #endif #ifdef cl_khr_fp16 diff --git a/libclc/generic/lib/shared/vstore.cl b/libclc/generic/lib/shared/vstore.cl index cbddd59795b..243c2f2810c 100644 --- a/libclc/generic/lib/shared/vstore.cl +++ b/libclc/generic/lib/shared/vstore.cl @@ -124,6 +124,26 @@ _CLC_DEF _CLC_OVERLOAD float __clc_rtz(float x) return x; return as_float(as_uint(x) & mask); } +_CLC_DEF _CLC_OVERLOAD float __clc_rti(float x) +{ + const float inf = copysign(INFINITY, x); + /* Set lower 13 bits */ + int mask = (1 << 13) - 1; + const int exp = (as_uint(x) >> 23 & 0xff) - 127; + /* Denormals cannot be flushed, and they use different bit for rounding */ + if (exp < -14) + mask = (1 << (13 + min(-(exp + 14), 10))) - 1; + /* Handle nan corner case */ + if (isnan(x)) + return x; + const float next = nextafter(as_float(as_uint(x) | mask), inf); + return ((as_uint(x) & mask) == 0) ? x : next; +} +_CLC_DEF _CLC_OVERLOAD float __clc_rtn(float x) +{ + return ((as_uint(x) & 0x80000000) == 0) ? __clc_rtz(x) : __clc_rti(x); +} + #ifdef cl_khr_fp64 _CLC_DEF _CLC_OVERLOAD double __clc_noop(double x) { @@ -145,11 +165,31 @@ _CLC_DEF _CLC_OVERLOAD double __clc_rtz(double x) return x; return as_double(as_ulong(x) & mask); } +_CLC_DEF _CLC_OVERLOAD double __clc_rti(double x) +{ + const double inf = copysign((double)INFINITY, x); + /* Set lower 42 bits */ + long mask = (1UL << 42UL) - 1UL; + const int exp = (as_ulong(x) >> 52 & 0x7ff) - 1023; + /* Denormals cannot be flushed, and they use different bit for rounding */ + if (exp < -14) + mask = (1UL << (42UL + min(-(exp + 14), 10))) - 1; + /* Handle nan corner case */ + if (isnan(x)) + return x; + const double next = nextafter(as_double(as_ulong(x) | mask), inf); + return ((as_ulong(x) & mask) == 0) ? x : next; +} +_CLC_DEF _CLC_OVERLOAD double __clc_rtn(double x) +{ + return ((as_ulong(x) & 0x8000000000000000UL) == 0) ? __clc_rtz(x) : __clc_rti(x); +} #endif #define __XFUNC(SUFFIX, VEC_SIZE, OFFSET, TYPE, STYPE, AS) \ __FUNC(SUFFIX, VEC_SIZE, OFFSET, TYPE, STYPE, AS, __clc_noop) \ - __FUNC(SUFFIX ## _rtz, VEC_SIZE, OFFSET, TYPE, STYPE, AS, __clc_rtz) + __FUNC(SUFFIX ## _rtz, VEC_SIZE, OFFSET, TYPE, STYPE, AS, __clc_rtz) \ + __FUNC(SUFFIX ## _rtn, VEC_SIZE, OFFSET, TYPE, STYPE, AS, __clc_rtn) #define FUNC(SUFFIX, VEC_SIZE, OFFSET, TYPE, STYPE, AS) \ __XFUNC(SUFFIX, VEC_SIZE, OFFSET, TYPE, STYPE, AS) |