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

Cleanups (#2631)

Browse Source 
* Remove some duplicate code in fmha_fwd_appendkv_kernel.hpp

* Simplify two templated operator calls by having the templated types deduced automatically

* Simplify two GemmPipeline calls

* Fix GemmPipelineAgBgCrCompV4::GetName

* Refactor use of ArgParser in CK tile GEMM examples

* Update args in README.md to match the implementation in create_args

* Remove some unnecessary include statements

* Rename two variables

* Factor out common code

* Factor out do_verify

* Add and use type aliases for memory operation integral constants

* In gemm_basic.cpp, use kPadM, kPadN, kPadK, and kBlockPerCu from GemmConfig

---------

Co-authored-by: Adam Osewski <19374865+aosewski@users.noreply.github.com>
This commit is contained in:
SamiAario-AMD
2025-08-13 11:12:08 +03:00
committed by GitHub
parent 05a6e92705
commit 28a97865f5
9 changed files with 142 additions and 178 deletions
Download Patch File Download Diff File Expand all files Collapse all files

71
example/ck_tile/03_gemm/run_gemm_example.inc
View File

@@ -272,6 +272,25 @@ auto shuffle_b(const ck_tile::HostTensor<T>& t)
return ck_tile::reference_permute(t_view, {0, 2, 3, 1, 4});
}
template <typename CDataType>
bool do_verify(const ck_tile::HostTensor<CDataType>& c_m_n_dev_result,
const ck_tile::HostTensor<CDataType>& c_m_n_ref,
const ck_tile::tuple<double, double>& rtol_atol,
const char* variant)
{
bool pass = ck_tile::check_err(c_m_n_dev_result,
c_m_n_ref,
"Error: Incorrect results!",
rtol_atol.at(ck_tile::number<0>{}),
rtol_atol.at(ck_tile::number<1>{}));
std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
<< " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{}) << std::endl;
std::cout << "The " << variant << " verification result is:" << (pass ? "correct" : "fail")
<< std::endl;
return pass;
}
template <typename GemmConfig,
typename ADataType,
typename BDataType = ADataType,
@@ -279,16 +298,11 @@ template <typename GemmConfig,
typename ALayout,
typename BLayout,
typename CLayout>
int run_gemm_example_with_layouts(int argc,
char* argv[],
int run_gemm_example_with_layouts(ck_tile::ArgParser& arg_parser,
const ALayout a_layout = ALayout{},
const BLayout b_layout = BLayout{},
[[maybe_unused]] const CLayout c_layout = CLayout{})
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
using AccDataType = typename GemmTypeConfig<ADataType, BDataType, CDataType>::AccDataType;
ck_tile::index_t M = arg_parser.get_int("m");
@@ -430,28 +444,20 @@ int run_gemm_example_with_layouts(int argc,
c_m_n_dev_buf.FromDevice(c_m_n_dev_result.data());
bool pass = true;
// memory on host to store gpu reference result
ck_tile::HostTensor<CDataType> c_m_n_ref(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
c_m_n_ref.SetZero();
if(arg_parser.get_int("v") == 1)
{
ck_tile::HostTensor<CDataType> c_m_n_host_ref(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
c_m_n_host_ref.SetZero();
ck_tile::reference_gemm<ADataType, BDataType, AccDataType, CDataType>(
a_m_k, b_k_n, c_m_n_host_ref);
a_m_k, b_k_n, c_m_n_ref);
const float max_accumulated_value =
*std::max_element(c_m_n_host_ref.mData.begin(), c_m_n_host_ref.mData.end());
*std::max_element(c_m_n_ref.mData.begin(), c_m_n_ref.mData.end());
const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(
K, kbatch, max_accumulated_value);
pass = ck_tile::check_err(c_m_n_dev_result,
c_m_n_host_ref,
"Error: Incorrect results!",
rtol_atol.at(ck_tile::number<0>{}),
rtol_atol.at(ck_tile::number<1>{}));
std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
<< " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{})
<< std::endl;
std::cout << "The CPU verification result is:" << (pass ? "correct" : "fail") << std::endl;
pass = do_verify(c_m_n_dev_result, c_m_n_ref, rtol_atol, "CPU");
}
else if(arg_parser.get_int("v") == 2)
{
@@ -465,13 +471,8 @@ int run_gemm_example_with_layouts(int argc,
b_k_n_dev_buf.ToDevice(b_k_n.data());
}
// memory on host to store gpu reference result
ck_tile::HostTensor<CDataType> c_m_n_gpu_ref(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
// memory on device to store gpu reference result
ck_tile::DeviceMem c_m_n_gpu_buf_ref(c_m_n_gpu_ref.get_element_space_size_in_bytes());
c_m_n_gpu_ref.SetZero();
ck_tile::DeviceMem c_m_n_gpu_buf_ref(c_m_n_ref.get_element_space_size_in_bytes());
c_m_n_gpu_buf_ref.SetZero();
ADataType* d_A = static_cast<ADataType*>(a_m_k_dev_buf.GetDeviceBuffer());
@@ -486,21 +487,13 @@ int run_gemm_example_with_layouts(int argc,
BLayout,
CLayout>(d_A, d_B, d_C, M, N, K, stride_A, stride_B, stride_C);
c_m_n_gpu_buf_ref.FromDevice(c_m_n_gpu_ref.data());
c_m_n_gpu_buf_ref.FromDevice(c_m_n_ref.data());
const float max_accumulated_value =
*std::max_element(c_m_n_gpu_ref.mData.begin(), c_m_n_gpu_ref.mData.end());
*std::max_element(c_m_n_ref.mData.begin(), c_m_n_ref.mData.end());
const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(
K, kbatch, max_accumulated_value);
pass = ck_tile::check_err(c_m_n_dev_result,
c_m_n_gpu_ref,
"Error: Incorrect results!",
rtol_atol.at(ck_tile::number<0>{}),
rtol_atol.at(ck_tile::number<1>{}));
std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
<< " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{})
<< std::endl;
std::cout << "The GPU verification result is: " << (pass ? "correct" : "fail") << std::endl;
pass = do_verify(c_m_n_dev_result, c_m_n_ref, rtol_atol, "GPU");
}
return pass;