ahmed/OpenCL-CTS
1
0
Fork 0
mirror of https://github.com/KhronosGroup/OpenCL-CTS.git synced 2026-03-19 06:09:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix testing of half-precision fma. (#1882)

Browse Source 
Half-precision functions are generally tested against the
single-precision reference. This causes double rounding: first to single
precision, then from there to half precision. For the most part, it is
good enough, but specifically in the case of fma, a correctly rounded
result is required and is not obtained, for instance for arguments
0x1.eacp+7, 0x1.3f4p+4, 0x1.c04p+14, which produce an exact result of
0x1.065fffp+15 which should be rounded to half-prefcision 0x1.064p+15,
but was previously first rounded to single-precision 0x1.066p+15, and
from there to half-precision 0x1.068p+15. Testing against reference_fmal
gives us sufficient precision that double rounding does not cause
issues.

The f_fma(..., FLUSHED) calls for FTZ testing cannot be updated the same
way but do not need to be: these calls all have at least one constant
operand of zero. If one operand is zero, double rounding cannot be an
issue.
This commit is contained in:
Harald van Dijk
2024-02-06 17:25:31 +00:00
committed by GitHub
parent 79fc236e4a
commit d338b42e8f
4 changed files with 10 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
test_conformance/math_brute_force/ternary_half.cpp
View File

@@ -237,8 +237,8 @@ int TestFunc_Half_Half_Half_Half(const Func *f, MTdata d, bool relaxedMode)
for (size_t j = 0; j < bufferElements; j++)
{
feclearexcept(FE_OVERFLOW);
res[j] = HFF((float)f->func.f_fma(
HTF(hp0[j]), HTF(hp1[j]), HTF(hp2[j]), CORRECTLY_ROUNDED));
res[j] = HFD((double)f->dfunc.f_fff(HTF(hp0[j]), HTF(hp1[j]),
HTF(hp2[j])));
overflow[j] =
FE_OVERFLOW == (FE_OVERFLOW & fetestexcept(FE_OVERFLOW));
}
@@ -246,8 +246,8 @@ int TestFunc_Half_Half_Half_Half(const Func *f, MTdata d, bool relaxedMode)
else
{
for (size_t j = 0; j < bufferElements; j++)
res[j] = HFF((float)f->func.f_fma(
HTF(hp0[j]), HTF(hp1[j]), HTF(hp2[j]), CORRECTLY_ROUNDED));
res[j] = HFD((double)f->dfunc.f_fff(HTF(hp0[j]), HTF(hp1[j]),
HTF(hp2[j])));
}
// Read the data back
@@ -277,9 +277,9 @@ int TestFunc_Half_Half_Half_Half(const Func *f, MTdata d, bool relaxedMode)
{
int fail;
cl_half test = ((cl_half *)q)[j];
float ref1 = f->func.f_fma(HTF(hp0[j]), HTF(hp1[j]),
HTF(hp2[j]), CORRECTLY_ROUNDED);
cl_half correct = HFF(ref1);
double ref1 = (double)f->dfunc.f_fff(
HTF(hp0[j]), HTF(hp1[j]), HTF(hp2[j]));
cl_half correct = HFD(ref1);
// Per section 10 paragraph 6, accept any result if an input
// or output is a infinity or NaN or overflow

1
test_conformance/math_brute_force/utility.h
View File

@@ -75,6 +75,7 @@ extern RoundingMode gFloatToHalfRoundingMode;
extern cl_half_rounding_mode gHalfRoundingMode;
#define HFF(num) cl_half_from_float(num, gHalfRoundingMode)
#define HFD(num) cl_half_from_double(num, gHalfRoundingMode)
#define HTF(num) cl_half_to_float(num)
#define LOWER_IS_BETTER 0