ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-26 16:04:58 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit '1b1c46e508c1fd40a03f54114b6b78629032fb4f' into develop

Browse Source
This commit is contained in:
assistant-librarian[bot]
2025-11-11 17:12:49 +00:00
parent 0b000816a4
commit db12c41b56
65 changed files with 845 additions and 455 deletions
Download Patch File Download Diff File Expand all files Collapse all files

98
include/ck_tile/host/tensor_shuffle_utils.hpp
View File

@@ -24,16 +24,43 @@ 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});
int n_ = t.get_lengths()[1];
int k_ = t.get_lengths()[0];
if(ck_tile::is_gfx12_supported())
{
constexpr int divisor = 2;
constexpr int kABK1PerLane = 8;
constexpr int kABK0PerLane = GemmConfig::K_Warp_Tile / divisor / kABK1PerLane;
ck_tile::HostTensor<T> t_view({n_ / GemmConfig::N_Warp_Tile,
GemmConfig::N_Warp_Tile,
k_ / GemmConfig::K_Warp_Tile,
kABK0PerLane,
divisor,
kABK1PerLane});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 2, 4, 1, 3, 5});
}
else
{
int divisor = 1;
if(ck_tile::is_gfx11_supported())
{
divisor = 1;
}
else
{
assert(is_wave32() == false);
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>
@@ -55,21 +82,46 @@ 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});
if(ck_tile::is_gfx12_supported())
{
constexpr int divisor = 2;
constexpr int kABK1PerLane = 8;
constexpr int kABK0PerLane = GemmConfig::K_Warp_Tile / divisor / kABK1PerLane;
ck_tile::HostTensor<T> t_view({n_ / GemmConfig::N_Tile,
GemmConfig::N_Warp,
GemmConfig::N_Warp_Tile,
NRepeat,
k_ / GemmConfig::K_Warp_Tile,
kABK0PerLane,
divisor,
kABK1PerLane});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 3, 1, 4, 6, 5, 2, 7});
}
else
{
int divisor = 1;
if(ck_tile::is_gfx11_supported())
{
divisor = 1;
}
else
{
assert(is_wave32() == false);
divisor = GemmConfig::N_Warp_Tile == 32 ? 2 : 4;
}
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