Skip to content

GitLab

Commit b5a9f642 authored by Jianfeng yan's avatar Jianfeng yan
Browse files

fix compile errors

parent 73daa6d6
Hide whitespace changes
Inline Side-by-side
Showing with 131 additions and 131 deletions
profiler/include/profile_gemm_impl.hpp
View file @ b5a9f642
...@@ -28,51 +28,51 @@ void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNo ...@@ -28,51 +28,51 @@ void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNo
void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
// std::vector<DeviceGemmNoOpPtr>&); std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
// std::vector<DeviceGemmNoOpPtr>&); std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
// std::vector<DeviceGemmNoOpPtr>&); std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
// std::vector<DeviceGemmNoOpPtr>&); std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_kn_mn_instances(
// std::vector<DeviceGemmNoOpPtr>&); std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances(
// std::vector<DeviceGemmNoOpPtr>&); std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_kn_mn_instances(
// std::vector<DeviceGemmNoOpPtr>&); std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_nk_mn_instances(
// std::vector<DeviceGemmNoOpPtr>&); std::vector<DeviceGemmNoOpPtr>&);
//
// void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
// std::vector<DeviceGemmNoOpPtr>&); std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
// void add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance } // namespace device_gemm_instance
} // namespace device } // namespace device
...@@ -174,8 +174,8 @@ void profile_gemm_impl(int do_verification, ...@@ -174,8 +174,8 @@ void profile_gemm_impl(int do_verification,
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
// //
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
} }
} }
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
...@@ -191,9 +191,9 @@ void profile_gemm_impl(int do_verification, ...@@ -191,9 +191,9 @@ void profile_gemm_impl(int do_verification,
{ {
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
//
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
} }
} }
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
...@@ -209,9 +209,9 @@ void profile_gemm_impl(int do_verification, ...@@ -209,9 +209,9 @@ void profile_gemm_impl(int do_verification,
{ {
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
//
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
} }
} }
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
...@@ -227,9 +227,9 @@ void profile_gemm_impl(int do_verification, ...@@ -227,9 +227,9 @@ void profile_gemm_impl(int do_verification,
{ {
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
//
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
} }
} }
} }
...@@ -242,16 +242,16 @@ void profile_gemm_impl(int do_verification, ...@@ -242,16 +242,16 @@ void profile_gemm_impl(int do_verification,
{ {
if(KBatch > 1) if(KBatch > 1)
{ {
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
} }
else else
{ {
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
} }
} }
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
...@@ -260,19 +260,19 @@ void profile_gemm_impl(int do_verification, ...@@ -260,19 +260,19 @@ void profile_gemm_impl(int do_verification,
{ {
if(KBatch > 1) if(KBatch > 1)
{ {
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
} }
else else
{ {
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
//
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
} }
} }
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
...@@ -281,16 +281,16 @@ void profile_gemm_impl(int do_verification, ...@@ -281,16 +281,16 @@ void profile_gemm_impl(int do_verification,
{ {
if(KBatch > 1) if(KBatch > 1)
{ {
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
} }
else else
{ {
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
} }
} }
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
...@@ -299,16 +299,16 @@ void profile_gemm_impl(int do_verification, ...@@ -299,16 +299,16 @@ void profile_gemm_impl(int do_verification,
{ {
if(KBatch > 1) if(KBatch > 1)
{ {
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_splitk_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
} }
else else
{ {
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
} }
} }
} }
...@@ -316,66 +316,66 @@ void profile_gemm_impl(int do_verification, ...@@ -316,66 +316,66 @@ void profile_gemm_impl(int do_verification,
is_same<BDataType, ck::bhalf_t>::value && is_same<BDataType, ck::bhalf_t>::value &&
is_same<CDataType, ck::bhalf_t>::value) is_same<CDataType, ck::bhalf_t>::value)
{ {
// if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value && if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
// is_same<BLayout, tensor_layout::gemm::RowMajor>::value && is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
// is_same<CLayout, tensor_layout::gemm::RowMajor>::value) is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
// { {
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(gemm_ptrs);
// } }
// else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
// is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value && is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
// is_same<CLayout, tensor_layout::gemm::RowMajor>::value) is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
// { {
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(gemm_ptrs);
// } }
// else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
// is_same<BLayout, tensor_layout::gemm::RowMajor>::value && is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
// is_same<CLayout, tensor_layout::gemm::RowMajor>::value) is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
// { {
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(gemm_ptrs);
// } }
// else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
// is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value && is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
// is_same<CLayout, tensor_layout::gemm::RowMajor>::value) is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
// { {
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(gemm_ptrs);
// } }
} }
else if constexpr(is_same<ADataType, int8_t>::value && is_same<BDataType, int8_t>::value && else if constexpr(is_same<ADataType, int8_t>::value && is_same<BDataType, int8_t>::value &&
is_same<CDataType, int8_t>::value) is_same<CDataType, int8_t>::value)
{ {
// if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value && if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
// is_same<BLayout, tensor_layout::gemm::RowMajor>::value && is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
// is_same<CLayout, tensor_layout::gemm::RowMajor>::value) is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
// { {
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_kn_mn_instances(gemm_ptrs);
// } }
// else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
// is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value && is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
// is_same<CLayout, tensor_layout::gemm::RowMajor>::value) is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
// { {
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances(gemm_ptrs);
// } }
// else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
// is_same<BLayout, tensor_layout::gemm::RowMajor>::value && is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
// is_same<CLayout, tensor_layout::gemm::RowMajor>::value) is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
// { {
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_kn_mn_instances(gemm_ptrs);
// } }
// else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
// is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value && is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
// is_same<CLayout, tensor_layout::gemm::RowMajor>::value) is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
// { {
// ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
// add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_nk_mn_instances(gemm_ptrs);
// } }
} }
if(gemm_ptrs.size() <= 0) if(gemm_ptrs.size() <= 0)
......
test/gemm/gemm_fp16.cpp
View file @ b5a9f642
...@@ -33,10 +33,10 @@ void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNo ...@@ -33,10 +33,10 @@ void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNo
void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); // void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); // void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); // void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); // void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
...@@ -63,8 +63,8 @@ int main() ...@@ -63,8 +63,8 @@ int main()
std::vector<DeviceGemmNoOpPtr> gemmPtrs; std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemmPtrs); add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance:: // ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(gemmPtrs); // add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemmPtrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemmPtrs);
...@@ -85,8 +85,8 @@ int main() ...@@ -85,8 +85,8 @@ int main()
gemmPtrs.clear(); gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemmPtrs); add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance:: // ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(gemmPtrs); // add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemmPtrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemmPtrs);
...@@ -107,8 +107,8 @@ int main() ...@@ -107,8 +107,8 @@ int main()
gemmPtrs.clear(); gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemmPtrs); add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance:: // ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(gemmPtrs); // add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemmPtrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemmPtrs);
...@@ -129,8 +129,8 @@ int main() ...@@ -129,8 +129,8 @@ int main()
gemmPtrs.clear(); gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemmPtrs); add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance:: // ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(gemmPtrs); // add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemmPtrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment