Updated Poly 16 in Gelu Erf to double precision

Updated poly Gelu Erf precision to double to keep the error with in 1e-5 limit when compared to reference gelu_erf, which is also increased the compute to 2x compared to float.

AMD-Internal: SWLCSG-3551
This commit is contained in:
Bhaskar, Nallani
2025-07-07 14:05:40 +05:30
committed by GitHub
parent c0d33879ec
commit 9b02201b5b
5 changed files with 187 additions and 121 deletions

View File

@@ -645,8 +645,8 @@ void mat_mul_accuracy_check_driver_ ## BLAS_SFX \
GEN_FUNC_NAME(C_type,_to_float)(*( c + ( rs_c * i ) + ( cs_c * j ) ), &comp_float); \
GEN_FUNC_NAME(C_type,_to_float)(out_temp_accum, &ref_float); \
\
if ( ( ( comp_float - ref_float ) > 1.0E-5 ) || \
( ( ref_float - comp_float ) > 1.0E-5 ) ) \
if ( ( ( comp_float - ref_float ) / (ref_float) > 1.0E-5 ) || \
( ( ref_float - comp_float ) / (comp_float) > 1.0E-5 ) ) \
{ \
if ( fout ) \
{ \