mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-16 19:09:59 +00:00
c55129e8f5
* experimenting magic number division
* overhauling fwd-v4r4 to clearly reflect transformation graph
* added fwd-v4r5
* bug fix for make_dynamic_naive_tensor_descriptor_aligned_v2
* bug fix and added sanity-check in transform_dynamic_tensor_descriptor
* added conv_driver_v2
[ROCm/composable_kernel commit: 30072aec37]
78 lines
2.2 KiB
C++
78 lines
2.2 KiB
C++
#ifndef CK_ARRAY_MULTI_INDEX_HPP
|
|
#define CK_ARRAY_MULTI_INDEX_HPP
|
|
|
|
#include "common_header.hpp"
|
|
|
|
namespace ck {
|
|
|
|
template <index_t N>
|
|
using MultiIndex = Array<index_t, N>;
|
|
|
|
template <typename... Xs>
|
|
__host__ __device__ constexpr auto make_multi_index(Xs&&... xs)
|
|
{
|
|
return make_array<index_t>(index_t{xs}...);
|
|
}
|
|
|
|
template <index_t NSize>
|
|
__host__ __device__ constexpr auto make_zero_multi_index()
|
|
{
|
|
return unpack([](auto... xs) { return make_multi_index(xs...); },
|
|
typename uniform_sequence_gen<NSize, 0>::type{});
|
|
}
|
|
|
|
template <typename T>
|
|
__host__ __device__ constexpr auto to_multi_index(const T& x)
|
|
{
|
|
return unpack([](auto... ys) { return make_multi_index(ys...); }, x);
|
|
}
|
|
|
|
template <index_t NSize, typename X>
|
|
__host__ __device__ constexpr auto operator+=(MultiIndex<NSize>& y, const X& x)
|
|
{
|
|
static_assert(X::Size() == NSize, "wrong! size not the same");
|
|
static_for<0, NSize, 1>{}([&](auto i) { y(i) += x[i]; });
|
|
return y;
|
|
}
|
|
|
|
template <index_t NSize, typename X>
|
|
__host__ __device__ constexpr auto operator-=(MultiIndex<NSize>& y, const X& x)
|
|
{
|
|
static_assert(X::Size() == NSize, "wrong! size not the same");
|
|
static_for<0, NSize, 1>{}([&](auto i) { y(i) -= x[i]; });
|
|
return y;
|
|
}
|
|
|
|
template <index_t NSize, typename T>
|
|
__host__ __device__ constexpr auto operator+(const MultiIndex<NSize>& a, const T& b)
|
|
{
|
|
using type = MultiIndex<NSize>;
|
|
static_assert(T::Size() == NSize, "wrong! size not the same");
|
|
type r;
|
|
static_for<0, NSize, 1>{}([&](auto i) { r(i) = a[i] + b[i]; });
|
|
return r;
|
|
}
|
|
|
|
template <index_t NSize, typename T>
|
|
__host__ __device__ constexpr auto operator-(const MultiIndex<NSize>& a, const T& b)
|
|
{
|
|
using type = MultiIndex<NSize>;
|
|
static_assert(T::Size() == NSize, "wrong! size not the same");
|
|
type r;
|
|
static_for<0, NSize, 1>{}([&](auto i) { r(i) = a[i] - b[i]; });
|
|
return r;
|
|
}
|
|
|
|
template <index_t NSize, typename T>
|
|
__host__ __device__ constexpr auto operator*(const MultiIndex<NSize>& a, const T& b)
|
|
{
|
|
using type = MultiIndex<NSize>;
|
|
static_assert(T::Size() == NSize, "wrong! size not the same");
|
|
type r;
|
|
static_for<0, NSize, 1>{}([&](auto i) { r(i) = a[i] * b[i]; });
|
|
return r;
|
|
}
|
|
|
|
} // namespace ck
|
|
#endif
|