ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-24 14:54:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

General 2D Reduction Kernel (#2535)

Browse Source 
* General 2D Reduction Kernel

* Move the reduction kernel from the example
* Split the code and add the necessary policy, problem, shape files as
per ck_tile convention
* Add/modify the headers
* Modified the example to work with the 'new' kernel
* Added tests for the kernel
* N-D refernce reduce
* Added support for N-D input with transform to 2D
* Added padding to support various input sized tensors
* Bug fix in the thread buffer constructor
* Some comments to explain the reduce2d block kernel

* comments resolution

* clang-format

* comments resolution

* clang-format

* clang-format

* comments resolution

* clang-format

[ROCm/composable_kernel commit: 4750b293fe]
This commit is contained in:
Yashvardhan Agarwal
2025-08-06 16:36:59 +03:00
committed by GitHub
parent b7659e284a
commit b6f0e98da6
14 changed files with 905 additions and 199 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
include/ck_tile/core/container/thread_buffer.hpp
View File

@@ -42,7 +42,11 @@ struct thread_buffer {
// TODO: this ctor can't ignore
CK_TILE_HOST_DEVICE constexpr thread_buffer() : data{} {}
CK_TILE_HOST_DEVICE constexpr thread_buffer(const value_type & o) : data{o} {}
CK_TILE_HOST_DEVICE constexpr thread_buffer(const value_type & o) : data{} {
static_for<0, N, 1>{}(
[&](auto i) { data[i] = o; }
);
}
CK_TILE_HOST_DEVICE static constexpr auto size() { return N; }
CK_TILE_HOST_DEVICE auto & get() {return data; }

57
include/ck_tile/core/utility/reduce_operator.hpp
View File

@@ -26,7 +26,8 @@ struct Add
}
template <typename T,
typename = std::enable_if_t<std::is_same_v<T, half_t> || std::is_same_v<T, bf16_t>>>
typename = std::enable_if_t<std::is_same_v<T, half_t> || std::is_same_v<T, bf16_t> ||
std::is_same_v<T, fp8_t> || std::is_same_v<T, bf8_t>>>
CK_TILE_HOST_DEVICE constexpr T operator()(T& y, T x) const
{
float y_ = type_convert<float>(y);
@@ -34,6 +35,8 @@ struct Add
return type_convert<T>(y_ + x_);
}
static constexpr bool requires_special_combine = false;
};
struct SquareAdd
@@ -51,13 +54,47 @@ struct SquareAdd
{
return y + (x * x);
}
template <typename T,
typename = std::enable_if_t<std::is_same_v<T, half_t> || std::is_same_v<T, bf16_t> ||
std::is_same_v<T, fp8_t> || std::is_same_v<T, bf8_t>>>
CK_TILE_HOST_DEVICE constexpr T operator()(T& y, T x) const
{
float y_ = type_convert<float>(y);
float x_ = type_convert<float>(x);
return type_convert<T>(y_ + (x_ * x_));
}
// For combining partial results
template <typename T,
typename = std::enable_if_t<std::is_same_v<T, float> || std::is_same_v<T, double> ||
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>>>
CK_TILE_HOST_DEVICE constexpr T combine_partial_results(const T& partial1,
const T& partial2) const
{
return partial1 + partial2; // Just add the partial sums, don't square again
}
template <typename T,
typename = std::enable_if_t<std::is_same_v<T, half_t> || std::is_same_v<T, bf16_t> ||
std::is_same_v<T, fp8_t> || std::is_same_v<T, bf8_t>>>
CK_TILE_HOST_DEVICE constexpr T combine_partial_results(T& partial1, T& partial2) const
{
float partial1_ = type_convert<float>(partial1);
float partial2_ = type_convert<float>(partial2);
return type_convert<T>(partial1_ + partial2_);
}
static constexpr bool requires_special_combine = true;
};
struct Max
{
template <typename T,
typename = std::enable_if_t<std::is_same_v<T, float> || std::is_same_v<T, double> ||
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>>>
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t> ||
std::is_same_v<T, half_t> || std::is_same_v<T, bf16_t> ||
std::is_same_v<T, fp8_t> || std::is_same_v<T, bf8_t>>>
CK_TILE_HOST_DEVICE static constexpr T GetIdentityValue()
{
return numeric<T>::min();
@@ -65,18 +102,24 @@ struct Max
template <typename T,
typename = std::enable_if_t<std::is_same_v<T, float> || std::is_same_v<T, double> ||
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>>>
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t> ||
std::is_same_v<T, half_t> || std::is_same_v<T, bf16_t> ||
std::is_same_v<T, fp8_t> || std::is_same_v<T, bf8_t>>>
CK_TILE_HOST_DEVICE constexpr T operator()(const T& y, const T x) const
{
return max(y, x);
}
static constexpr bool requires_special_combine = false;
};
struct AbsMax
{
template <typename T,
typename = std::enable_if_t<std::is_same_v<T, float> || std::is_same_v<T, double> ||
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>>>
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t> ||
std::is_same_v<T, half_t> || std::is_same_v<T, bf16_t> ||
std::is_same_v<T, fp8_t> || std::is_same_v<T, bf8_t>>>
CK_TILE_HOST_DEVICE static constexpr T GetIdentityValue()
{
return numeric<T>::min();
@@ -84,11 +127,15 @@ struct AbsMax
template <typename T,
typename = std::enable_if_t<std::is_same_v<T, float> || std::is_same_v<T, double> ||
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>>>
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t> ||
std::is_same_v<T, half_t> || std::is_same_v<T, bf16_t> ||
std::is_same_v<T, fp8_t> || std::is_same_v<T, bf8_t>>>
CK_TILE_HOST_DEVICE constexpr T operator()(const T& y, const T x) const
{
return max(y, abs(x));
}
static constexpr bool requires_special_combine = false;
};
} // namespace ReduceOp