Files
composable_kernel/include/ck/utility/statically_indexed_array_multi_index.hpp
Bartłomiej Kocot 60df79085d [rocm-libraries] ROCm/rocm-libraries#7631 (commit d591a7c)
[CK] Grouped Convolution Global Load/Store support (#7631)

## Motivation

Grouped Convolution Global Load/Store support to cover large tensor
cases.

## Technical Details

Utilize global load for grouped convolution forwad kernels. Update
Indexes to use int64.

## Test Plan

- test utils
- test conv kernels in next pr with instances

## Test Result

CI pending

## Submission Checklist

- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

AICK-1255

---------

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
2026-05-27 08:21:54 +00:00

159 lines
5.0 KiB
C++

// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#ifndef CK_STATICALLY_INDEXED_ARRAY_MULTI_INDEX_HPP
#define CK_STATICALLY_INDEXED_ARRAY_MULTI_INDEX_HPP
#include "common_header.hpp"
#include "ck/utility/math_v2.hpp"
namespace ck {
template <index_t N, typename T = index_t>
using MultiIndex = StaticallyIndexedArray<T, N>;
template <typename IdxType = index_t, typename... Xs>
__host__ __device__ constexpr auto make_multi_index(Xs&&... xs)
{
return make_statically_indexed_array<IdxType>(IdxType{xs}...);
}
template <index_t NSize, typename IdxType = index_t>
__host__ __device__ constexpr auto make_zero_multi_index()
{
return unpack([](auto... xs) { return make_multi_index<IdxType>(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);
}
// Here should use MultiIndex<NSize>, instead of Tuple<Ys...>, although the former
// is the alias of the latter. This is because compiler cannot infer the NSize if
// using MultiIndex<NSize>
// TODO: how to fix this?
template <typename... Ys,
typename X,
enable_if_t<!ck::is_integral<X>::value && !ck::is_floating_point<X>::value, bool> = false>
__host__ __device__ constexpr auto operator+=(Tuple<Ys...>& y, const X& x)
{
static_assert(X::Size() == sizeof...(Ys), "wrong! size not the same");
constexpr index_t NSize = sizeof...(Ys);
static_for<0, NSize, 1>{}([&](auto i) { y(i) += x[i]; });
return y;
}
template <typename... Ys,
typename X,
enable_if_t<!ck::is_integral<X>::value && !ck::is_floating_point<X>::value, bool> = false>
__host__ __device__ constexpr auto operator-=(Tuple<Ys...>& y, const X& x)
{
static_assert(X::Size() == sizeof...(Ys), "wrong! size not the same");
constexpr index_t NSize = sizeof...(Ys);
static_for<0, NSize, 1>{}([&](auto i) { y(i) -= x[i]; });
return y;
}
template <typename... Xs,
typename Y,
enable_if_t<!ck::is_integral<Y>::value && !ck::is_floating_point<Y>::value, bool> = false>
__host__ __device__ constexpr auto operator+(const Tuple<Xs...>& x, const Y& y)
{
static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
constexpr index_t NSize = sizeof...(Xs);
Tuple<Xs...> r;
static_for<0, NSize, 1>{}([&](auto i) { r(i) = x[i] + y[i]; });
return r;
}
template <typename... Xs,
typename Y,
enable_if_t<!ck::is_integral<Y>::value && !ck::is_floating_point<Y>::value, bool> = false>
__host__ __device__ constexpr auto operator-(const Tuple<Xs...>& x, const Y& y)
{
static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
constexpr index_t NSize = sizeof...(Xs);
Tuple<Xs...> r;
static_for<0, NSize, 1>{}([&](auto i) { r(i) = x[i] - y[i]; });
return r;
}
template <typename... Xs,
typename Y,
enable_if_t<!ck::is_integral<Y>::value && !ck::is_floating_point<Y>::value, bool> = false>
__host__ __device__ constexpr auto operator*(const Tuple<Xs...>& x, const Y& y)
{
static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
constexpr index_t NSize = sizeof...(Xs);
Tuple<Xs...> r;
static_for<0, NSize, 1>{}([&](auto i) { r(i) = x[i] * y[i]; });
return r;
}
// MultiIndex = scalar * MultiIndex
template <typename... Xs,
typename Y,
enable_if_t<ck::is_integral<Y>::value || ck::is_floating_point<Y>::value, bool> = false>
__host__ __device__ constexpr auto operator*(Y a, const Tuple<Xs...>& x)
{
constexpr index_t NSize = sizeof...(Xs);
Tuple<Xs...> r;
static_for<0, NSize, 1>{}([&](auto i) { r(i) = a * x[i]; });
return r;
}
// MultiIndex = MultiIndex * scalar
template <typename... Xs,
typename Y,
enable_if_t<ck::is_integral<Y>::value || ck::is_floating_point<Y>::value, bool> = false>
__host__ __device__ constexpr auto operator*(const Tuple<Xs...>& x, Y a)
{
return a * x;
}
namespace mathext {
template <typename... Xs>
__host__ __device__ constexpr auto exp(const Tuple<Xs...>& x)
{
constexpr index_t NSize = sizeof...(Xs);
Tuple<Xs...> r;
static_for<0, NSize, 1>{}([&](auto i) { r(i) = math::exp(x[i]); });
return r;
}
template <typename... Xs, typename Y>
__host__ __device__ constexpr auto max(const Tuple<Xs...>& x, const Y& y)
{
static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
constexpr index_t NSize = sizeof...(Xs);
Tuple<Xs...> r;
static_for<0, NSize, 1>{}([&](auto i) { r(i) = math::max(x[i], y[i]); });
return r;
}
} // namespace mathext
template <typename... Xs>
__host__ __device__ void print_multi_index(const Tuple<Xs...>& x)
{
printf("{");
printf("MultiIndex, ");
printf("size %d, ", index_t{sizeof...(Xs)});
static_for<0, sizeof...(Xs), 1>{}(
[&](auto i) { printf("%d ", static_cast<index_t>(x.At(i))); });
printf("}");
}
} // namespace ck
#endif