composable_kernel
composable_kernel copied to clipboard
Refactor f8_t and bf8_t as custom types, enable use of custom types
- Refactor f8_t and bf8_t
- Update conversion methods
- Update load method
- Add dynamic buffer custom types support
- Update threadwise conversion
Right now custom types are supported as source types, their support as destination types is coming soon...
Could you do a performance check to make sure the new custom data type has no impact on the performance of fp8_gemm?
@zjing14 There is no perf drop on example_gemm_xdl_fp8, nice!
Current work is covered in https://github.com/ROCm/composable_kernel/pull/1333