* profile reduce kernel for fp16 and reduceHigherdim

* use reinterpret_cast

* fix for CI on ROCm

* add Macro for ROCm

* ROCm CI config

* ROCm CI config

* unit test repair

* pull

* add common_funcs.h

* reduceType

* Update reduce_function.h

* not higher

* rename

* implement of matmul using cublasLt instead of cublas

* cublasLt bugfix

* Update matmul_kernel_impl.h

* Update matmul_kernel_impl_via_blasLt.h

* for-loop-algo

* PR comments changes

* add macro

* ci unused variable isCublasLt

* ci unused variable isCublasLt macro

* split matmul to autotune

* rewrite the split kernel with segmented_array

* rewrite the split kernel with segmented_array

* rewrite the split kernel with segmented_array

* add some method for cuda_graph

* fix bugs for rocm

* change for ci-error

* i dont know why ci-model-benchmark gives a shit error, so i recover codes with original one to see if original codes work.

* add some changes for passing mode_benchmark and coverage ci

* fix ci error

* fix ci-rocm error

* add some changes for header

---------
Co-authored-by: Nzhangbopd <1299246947@qq.com>
Co-authored-by: NBo Zhang <105368690+zhangbopd@users.noreply.github.com>

* Unify the gpu implementation of stack and unstack to reuse the optimization.

* Optimize the cuda implementation of unstack.

* Use GpuMemcpyAsync instead of memory::Copy.

* Fix error of calculating the index.

* Use FastDivMod to further imporve the performance of unstack.