Avoid prematurely casting GEMM parameters `alpha`, `beta` to `scalar_t` (#67633)
Summary:
stas00 uncovered an issue where certain half-precision GEMMs would produce outputs that looked like the result of strange rounding behavior (e.g., `10008.` in place of `10000.`). ptrblck suspected that this was due to the parameters being downcasted to the input types (which would reproduce the problematic output). Indeed, the GEMM and BGEMM cublas wrappers are currently converting the `alpha` and `beta` parameters to `scalar_t` (which potentially is reduced precision) before converting them back to `float`. This PR changes the "ARGTYPE" wrappers to use `acc_t` instead and adds a corresponding test.
CC ngimel
Pull Request resolved: https://github.com/pytorch/pytorch/pull/67633
Reviewed By: mruberry
Differential Revision: D32076474
Pulled By: ngimel
fbshipit-source-id: 2540d9b9d0195c17d07d1161374fb6a5850779d5
