ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 14:59:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] Adding support for TiledPermuteN on preshuffle Block Scale Gemm (#3019)

Browse Source 
* Adding support for TiledPermuteN

* Adding test

* resolving remod.py

---------

Co-authored-by: root <root@banff-cyxtera-s73-2.ctr.dcgpu>
This commit is contained in:
Khushbu Agarwal
2025-10-24 11:06:51 -07:00
committed by GitHub
parent f39626fcf7
commit 0584399571
8 changed files with 161 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files

75
include/ck_tile/host/shuffle_utils.hpp Normal file
View File

@@ -0,0 +1,75 @@
#pragma once
#include <stdexcept>
namespace ck_tile {
template <typename T>
auto shuffle_aq(const ck_tile::HostTensor<T>* t, int block_aq_k)
{
if(t->get_lengths().size() != 2)
{
throw std::runtime_error("Host tensor is not rank 2 tensor.");
}
int m_ = t->get_lengths()[0];
int aqk_ = t->get_lengths()[1];
if(aqk_ % block_aq_k != 0)
{
throw std::runtime_error("shuffle_aq needs a aqk of multiple times of block_aq_k.");
}
ck_tile::HostTensor<T> t_view({m_, aqk_ / block_aq_k, block_aq_k});
std::copy(t->begin(), t->end(), t_view.begin());
return ck_tile::reference_permute(t_view, {1, 0, 2});
}
template <typename GemmConfig, typename T>
auto shuffle_b(const ck_tile::HostTensor<T>& t)
{
assert(t.get_lengths().size() == 2);
int n_ = t.get_lengths()[1];
int k_ = t.get_lengths()[0];
constexpr int divisor = GemmConfig::N_Warp_Tile == 32 ? 2 : 4;
ck_tile::HostTensor<T> t_view({n_ / GemmConfig::N_Warp_Tile,
GemmConfig::N_Warp_Tile,
k_ / GemmConfig::K_Warp_Tile,
divisor,
GemmConfig::K_Warp_Tile / divisor});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 2, 3, 1, 4});
}
template <typename GemmConfig, typename T>
auto shuffle_bq_permuteN(const ck_tile::HostTensor<T>& t)
{
assert(t.get_lengths().size() == 2);
int n_ = t.get_lengths()[1];
int bqk_ = t.get_lengths()[0];
constexpr int NRepeat = GemmConfig::N_Tile / GemmConfig::N_Warp_Tile / GemmConfig::N_Warp;
ck_tile::HostTensor<T> t_view(
{n_ / GemmConfig::N_Tile, GemmConfig::N_Warp, GemmConfig::N_Warp_Tile, NRepeat, bqk_});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 3, 1, 2, 4});
}
template <typename GemmConfig, typename T>
auto shuffle_b_permuteN(const ck_tile::HostTensor<T>& t)
{
assert(t.get_lengths().size() == 2);
int n_ = t.get_lengths()[1];
int k_ = t.get_lengths()[0];
constexpr int divisor = GemmConfig::N_Warp_Tile == 32 ? 2 : 4;
constexpr int NRepeat = GemmConfig::N_Tile / GemmConfig::N_Warp_Tile / GemmConfig::N_Warp;
ck_tile::HostTensor<T> t_view({n_ / GemmConfig::N_Tile,
GemmConfig::N_Warp,
GemmConfig::N_Warp_Tile,
NRepeat,
k_ / GemmConfig::K_Warp_Tile,
divisor,
GemmConfig::K_Warp_Tile / divisor});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 3, 1, 4, 5, 2, 6});
}
} // namespace ck_tile