Add optimized blockwise gemm using ck wrapper (#1157)

* Add optimized blockwise gemm using ck wrapper

* Add basic gemm example

* Update docs

* Add tutorial for gemm using ck wrapper

* Add perf note

* edits

* Fix cmake

* Fixes

---------

Co-authored-by: Lisa Delaney <lisa.delaney@amd.com>
This commit is contained in:
Bartłomiej Kocot
2024-02-13 17:04:36 +01:00
committed by GitHub
parent bf98b47697
commit 1e73adbc28
20 changed files with 1597 additions and 512 deletions

View File

@@ -172,10 +172,10 @@ __host__ __device__ constexpr auto GenerateUpperDims(const Tuple<Transforms...>&
}
}
template <typename... Ts, typename Shape, typename FlattenDescriptor>
template <typename... Ts, typename Shape, typename UnrolledDescriptor>
__host__ __device__ constexpr auto GenerateSlicedDescriptor(const Tuple<Ts...>& idx,
const Shape& shape,
const FlattenDescriptor& flatten_desc)
const UnrolledDescriptor& flatten_desc)
{
constexpr auto old_shape_dims = decltype(UnrollNestedTuple(shape))::Size();