Skip to content

Multi AB support for wave transfer #3578

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
EnricoDeg wants to merge 6 commits into
base: streamhpc/padding_support_wave_transfer
Choose a base branch
from
Open

Multi AB support for wave transfer #3578

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
7656a1c
Add multi AB support to wave transfer
EnricoDeg Jan 8, 2026
ca9e5cb
Improviments to multi ABD examples
EnricoDeg Jan 8, 2026
ea2e811
Add instances and use intrawave v1 instead of interwave
EnricoDeg Jan 9, 2026
6918f05
Apply changes to other transfers
EnricoDeg Jan 9, 2026
32dbf96
Wave transfer: add support for multiple internal vgpr buffers
EnricoDeg Jan 9, 2026
86de53f
Fix compilation error gfx11
EnricoDeg Jan 12, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
29 changes: 26 additions & 3 deletions example/60_gemm_multi_ABD/gemm_multi_ABD_wmma_bias_fastgelu_bf16_i8.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,19 +96,19 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Wmm
8,
8,
0,
S<8, 32, 1>,
S<8, 16, 1>,
S<0, 2, 1>,
S<0, 2, 1>,
1,
1,
8,
8,
0,
1,
1,
S<1, 32, 1, 8>,
S<8, 8, 8>,
ck::BlockGemmPipelineScheduler::Intrawave,
ck::BlockGemmPipelineVersion::v3>;
ck::BlockGemmPipelineVersion::v1>;

int main(int argc, char* argv[])
{
Expand Down Expand Up @@ -174,6 +174,29 @@ int main(int argc, char* argv[])
}
};

auto f_get_default_stride =
[](std::size_t row, std::size_t col, ck::index_t stride, auto layout) {
if(stride == -1 || stride == 0)
{
// give a chance if stride is -1, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return static_cast<std::size_t>(col);
}
else
{
return static_cast<std::size_t>(row);
}
}
else
return static_cast<std::size_t>(stride);
};

StrideA = f_get_default_stride(M, K, StrideA, A0Layout{});
StrideB = f_get_default_stride(K, N, StrideB, B0Layout{});
StrideD = f_get_default_stride(M, N, StrideD, D0Layout{});
StrideE = f_get_default_stride(M, N, StrideE, ELayout{});

Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, StrideB, B1Layout{}));
Expand Down
30 changes: 26 additions & 4 deletions example/60_gemm_multi_ABD/gemm_multi_ABD_wmma_fastgelu_bf16_i8.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,19 +94,19 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Wmm
8,
8,
0,
S<8, 32, 1>,
S<8, 16, 1>,
S<0, 2, 1>,
S<0, 2, 1>,
1,
1,
8,
8,
0,
1,
1,
S<1, 32, 1, 8>,
S<8, 8, 8>,
ck::BlockGemmPipelineScheduler::Intrawave,
ck::BlockGemmPipelineVersion::v3>;
ck::BlockGemmPipelineVersion::v1>;

int main(int argc, char* argv[])
{
Expand All @@ -133,7 +133,7 @@ int main(int argc, char* argv[])
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
Expand Down Expand Up @@ -170,6 +170,28 @@ int main(int argc, char* argv[])
}
};

auto f_get_default_stride =
[](std::size_t row, std::size_t col, ck::index_t stride, auto layout) {
if(stride == -1 || stride == 0)
{
// give a chance if stride is -1, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return static_cast<std::size_t>(col);
}
else
{
return static_cast<std::size_t>(row);
}
}
else
return static_cast<std::size_t>(stride);
};

StrideA = f_get_default_stride(M, K, StrideA, A0Layout{});
StrideB = f_get_default_stride(K, N, StrideB, B0Layout{});
StrideE = f_get_default_stride(M, N, StrideE, ELayout{});

Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, StrideB, B1Layout{}));
Expand Down
27 changes: 25 additions & 2 deletions example/60_gemm_multi_ABD/gemm_multi_ABD_wmma_fp16.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -141,11 +141,11 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Wmm
8,
8,
0,
S<4, 64, 1>,
S<4, 16, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
1,
1,
8,
8,
0,
1,
Expand Down Expand Up @@ -233,6 +233,29 @@ int main(int argc, char* argv[])
}
};

auto f_get_default_stride =
[](std::size_t row, std::size_t col, ck::index_t stride, auto layout) {
if(stride == -1 || stride == 0)
{
// give a chance if stride is -1, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return static_cast<std::size_t>(col);
}
else
{
return static_cast<std::size_t>(row);
}
}
else
return static_cast<std::size_t>(stride);
};

StrideA = f_get_default_stride(M, K, StrideA, ALayout{});
StrideB = f_get_default_stride(K, N, StrideB, BLayout{});
StrideD = f_get_default_stride(M, N, StrideD, DLayout{});
StrideE = f_get_default_stride(M, N, StrideE, ELayout{});

Tensor<ADataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<ADataType> a1_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Expand Down
29 changes: 26 additions & 3 deletions example/60_gemm_multi_ABD/gemm_multi_ABD_wmma_multiply_bias_fastgelu_bf16_i8.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,19 +95,19 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Wmm
8,
8,
0,
S<8, 32, 1>,
S<8, 16, 1>,
S<0, 2, 1>,
S<0, 2, 1>,
1,
1,
8,
8,
0,
1,
1,
S<1, 32, 1, 8>,
S<8, 8, 8>,
ck::BlockGemmPipelineScheduler::Intrawave,
ck::BlockGemmPipelineVersion::v3>;
ck::BlockGemmPipelineVersion::v1>;

int main(int argc, char* argv[])
{
Expand Down Expand Up @@ -173,6 +173,29 @@ int main(int argc, char* argv[])
}
};

auto f_get_default_stride =
[](std::size_t row, std::size_t col, ck::index_t stride, auto layout) {
if(stride == -1 || stride == 0)
{
// give a chance if stride is -1, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return static_cast<std::size_t>(col);
}
else
{
return static_cast<std::size_t>(row);
}
}
else
return static_cast<std::size_t>(stride);
};

StrideA = f_get_default_stride(M, K, StrideA, A0Layout{});
StrideB = f_get_default_stride(K, N, StrideB, B0Layout{});
StrideD = f_get_default_stride(M, N, StrideD, D0Layout{});
StrideE = f_get_default_stride(M, N, StrideE, ELayout{});

Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
Expand Down
Loading