fp16 and bfloat16 precision for dot#655
Conversation
| // Template to dispatch testing_gemm_ex for performance tests | ||
| // When Ti == void or complex, the test is marked invalid | ||
| // When Ti == void or complex or Ti == To == Tc == bfloat16, the test is marked invalid | ||
| template <typename Ti, typename To = Ti, typename Tc = To, typename = void> |
There was a problem hiding this comment.
I do not see a test for Ti == complex. Can the comment be updated if there is no test for complex.
There was a problem hiding this comment.
The gemm_ex and gemm_strided_batched_ex templates act on exclusion rather than inclusion like the other templates. The exclusion of complex types was removed when complex gemm was added. I added exclusion for bfloat16 types here as they are now permissible in type_dispatch.hpp, but not for gemm_ex. Changed the comment to reflect this in #f746ec6.
| ############################################# | ||
| Half bfloat single double complex real: &half_bfloat_single_double_complex_real_precisions | ||
| - *half_precision | ||
| - *bfa_precision |
There was a problem hiding this comment.
I think bf16 is more descriptive than bfa. Is it possible to make this bf16?
There was a problem hiding this comment.
Good point, changed in #f746ec6.
|
All tests (quick, pre-checkin, nightly) pass on gfx900 and gfx906. |
3 participants
Summary of proposed changes: