implicit gemm v1r2: adding support for nchw

This commit is contained in:
Chao Liu
2019-04-18 11:49:09 -05:00
parent 17f3d2d4bc
commit 19f17df47a
16 changed files with 1624 additions and 220 deletions

View File

@@ -1,4 +1,6 @@
#pragma once
#include "Sequence.hip.hpp"
#include "functional.hip.hpp"
template <class TData, index_t NSize>
struct Array
@@ -18,3 +20,35 @@ struct Array
__host__ __device__ TData& operator[](index_t i) { return mData[i]; }
};
template <class TData, index_t NSize, index_t... IRs>
__host__ __device__ auto reorder_array_given_new2old(const Array<TData, NSize>& old_array,
Sequence<IRs...> new2old)
{
Array<TData, NSize> new_array;
static_assert(NSize == sizeof...(IRs), "NSize not consistent");
static_for<0, NSize, 1>{}([&](auto IDim) {
constexpr index_t idim = IDim.Get();
new_array[idim] = old_array[new2old.Get(IDim)];
});
return new_array;
}
template <class TData, index_t NSize, index_t... IRs>
__host__ __device__ auto reorder_array_given_old2new(const Array<TData, NSize>& old_array,
Sequence<IRs...> old2new)
{
Array<TData, NSize> new_array;
static_assert(NSize == sizeof...(IRs), "NSize not consistent");
static_for<0, NSize, 1>{}([&](auto IDim) {
constexpr index_t idim = IDim.Get();
new_array[old2new.Get(IDim)] = old_array[idim];
});
return new_array;
}