Update hip-clang flags for hip kernels (ROCm#202)

* update hip clang flags

* tidy up

* tidy up

* move some flags under hip-clang only

Update hip-clang flags for hip kernels (ROCm#202)

* update hip clang flags

* tidy up

* tidy up

* move some flags under hip-clang only

Limit the number of iGEMM solver selected (ROCm#2520)

* enable first solver for fwd,bwd; two for wrw
* move ConvHipImplicitGemmV4WrW ahead and disable it for gfx908
* revert the order of solver
* Update mlo_dir_conv.cpp

Merge tag '2.3.0'

Workaround hipModuleLoadData (ROCm#2501)

* replace inmem module load
* fix TmpDir bug
* update temp file location
* satisfy ROCm#2331
* Update rocBLAS to get around tensile issue. Added additional package to docker file to support this change.
Co-authored-by: Daniel Lowell <daniel.lowell@amd.com>

Fix for SWDEV-220166 (ROCm#2387)

* check for hcc
* add comment for ticket reference
* address PR comments

Fix for SWDEV-220166 (ROCm#2387)

* check for hcc
* add comment for ticket reference
* address PR comments

Merge branch '2.2.x'

Resolve merge failures.

Workaround for SWDEV-203031: set default limit for MP Winograd worksp…

…ace usage (ROCm#2090)

* wa-swdev-203031(01) MP Winograd: set the default limit of workspace usage.
* wa-swdev-203031(03) Lower limit for gfx900 and gfx906/60 down to ~1.862 GiB. Remove limit for other devices.