Skip to content

GitLab

Commit c2e13d9b authored by Haocong WANG's avatar Haocong WANG
Browse files

Add rocm6.2 related gemm optimization

parent 7d69eb3b
Hide whitespace changes
Inline Side-by-side
Showing with 129 additions and 151 deletions
CMakeLists.txt
View file @ c2e13d9b
...@@ -183,8 +183,9 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600140090) ...@@ -183,8 +183,9 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600140090)
add_compile_options("SHELL: -mllvm -enable-post-misched=0") add_compile_options("SHELL: -mllvm -enable-post-misched=0")
endif() endif()
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132 AND ${hip_VERSION_FLAT} LESS 600300000) if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132 AND ${hip_VERSION_FLAT} LESS 600300000)
message("Adding the amdgpu-coerce-illegal-types=1") message("Adding the amdgpu-coerce-illegal-types=1 and greedy-reverse-local-assignment=1")
add_compile_options("SHELL: -mllvm -amdgpu-coerce-illegal-types=1") add_compile_options("SHELL: -mllvm -amdgpu-coerce-illegal-types=1")
add_compile_options("SHELL: -mllvm -greedy-reverse-local-assignment=1")
endif() endif()
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132) if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
message("Adding -amdgpu-early-inline-all=true and -amdgpu-function-calls=false") message("Adding -amdgpu-early-inline-all=true and -amdgpu-function-calls=false")
......
example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp
View file @ c2e13d9b
...@@ -72,7 +72,7 @@ using CDEElementOp = MultiplyMultiply; ...@@ -72,7 +72,7 @@ using CDEElementOp = MultiplyMultiply;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNPadding; static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNPadding;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3 using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3
// clang-format off // clang-format off
///######| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| ///######| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
///######| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| ///######| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
///######| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| ///######| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3.hpp
View file @ c2e13d9b
...@@ -417,6 +417,13 @@ struct GridwiseGemm_xdl_cshuffle_v3 ...@@ -417,6 +417,13 @@ struct GridwiseGemm_xdl_cshuffle_v3
} }
}(); }();
// pad M and N
return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(M, MPad - M),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
#if 0
using GemmSpecialization = tensor_operation::device::GemmSpecialization; using GemmSpecialization = tensor_operation::device::GemmSpecialization;
if constexpr(GemmSpec == GemmSpecialization::MNPadding || if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
...@@ -454,6 +461,7 @@ struct GridwiseGemm_xdl_cshuffle_v3 ...@@ -454,6 +461,7 @@ struct GridwiseGemm_xdl_cshuffle_v3
// not pad M or N // not pad M or N
return c_grid_desc_mraw_nraw; return c_grid_desc_mraw_nraw;
} }
#endif
} }
struct Problem struct Problem
...@@ -953,7 +961,8 @@ struct GridwiseGemm_xdl_cshuffle_v3 ...@@ -953,7 +961,8 @@ struct GridwiseGemm_xdl_cshuffle_v3
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding || if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding || GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)) GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
!(is_same<tensor_layout::gemm::RowMajor, ALayout>::value))
{ {
if(!(karg.M % MPerBlock == 0)) if(!(karg.M % MPerBlock == 0))
{ {
...@@ -970,7 +979,8 @@ struct GridwiseGemm_xdl_cshuffle_v3 ...@@ -970,7 +979,8 @@ struct GridwiseGemm_xdl_cshuffle_v3
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding || if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding || GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)) GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
(is_same<tensor_layout::gemm::RowMajor, BLayout>::value))
{ {
if(!(karg.N % NPerBlock == 0)) if(!(karg.N % NPerBlock == 0))
{ {
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp
View file @ c2e13d9b
...@@ -452,6 +452,13 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -452,6 +452,13 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
} }
}(); }();
// pad M and N
return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(M, MPad - M),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
#if 0
using GemmSpecialization = tensor_operation::device::GemmSpecialization; using GemmSpecialization = tensor_operation::device::GemmSpecialization;
if constexpr(GemmSpec == GemmSpecialization::MNPadding || if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
...@@ -489,6 +496,7 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -489,6 +496,7 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
// not pad M or N // not pad M or N
return c_grid_desc_mraw_nraw; return c_grid_desc_mraw_nraw;
} }
#endif
} }
__host__ __device__ static auto MakeDsGridDescriptor_M_N( __host__ __device__ static auto MakeDsGridDescriptor_M_N(
...@@ -1014,7 +1022,8 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -1014,7 +1022,8 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding || if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding || GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)) GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
!(is_same<tensor_layout::gemm::RowMajor, ALayout>::value))
{ {
if(!(karg.M % MPerBlock == 0)) if(!(karg.M % MPerBlock == 0))
{ {
...@@ -1031,7 +1040,8 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -1031,7 +1040,8 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding || if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding || GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)) GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
(is_same<tensor_layout::gemm::RowMajor, BLayout>::value))
{ {
if(!(karg.N % NPerBlock == 0)) if(!(karg.N % NPerBlock == 0))
{ {
......
library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply.hpp
View file @ c2e13d9b
...@@ -43,32 +43,6 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_ins ...@@ -43,32 +43,6 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_ins
PassThrough, PassThrough,
MultiplyMultiply>>>& instances); MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_default_instances( void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row, std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col, Col,
...@@ -95,19 +69,6 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_i ...@@ -95,19 +69,6 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_i
PassThrough, PassThrough,
MultiplyMultiply>>>& instances); MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_default_instances( void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row, std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col, Col,
...@@ -133,19 +94,6 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_i ...@@ -133,19 +94,6 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_i
PassThrough, PassThrough,
PassThrough, PassThrough,
MultiplyMultiply>>>& instances); MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
F8,
F8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
#endif #endif
template <typename ADataType, template <typename ADataType,
...@@ -195,24 +143,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu ...@@ -195,24 +143,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
op_ptrs); op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances( add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances(
op_ptrs); op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mnpadding_instances(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_default_instances( add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_default_instances(
op_ptrs); op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_instances( add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_instances(
op_ptrs); op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_default_instances( add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_default_instances(
op_ptrs); op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instances( add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instances(
op_ptrs); op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_mnkpadding_instances(
op_ptrs);
} }
} }
#endif #endif
......
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/CMakeLists.txt
View file @ c2e13d9b
...@@ -4,14 +4,10 @@ set(GEMM_MULTIPLY_MULTIPLY_INSTANCES) ...@@ -4,14 +4,10 @@ set(GEMM_MULTIPLY_MULTIPLY_INSTANCES)
list(APPEND GEMM_MULTIPLY_MULTIPLY_INSTANCES list(APPEND GEMM_MULTIPLY_MULTIPLY_INSTANCES
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance.cpp device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance.cpp device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mnpadding_instance.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mnkpadding_instance.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_default_instance.cpp device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_default_instance.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_instance.cpp device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_instance.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_mnkpadding_instance.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_default_instance.cpp device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_default_instance.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instance.cpp device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instance.cpp
device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_mnkpadding_instance.cpp
) )
add_instance_library(device_gemm_multiply_multiply_instance ${GEMM_MULTIPLY_MULTIPLY_INSTANCES}) add_instance_library(device_gemm_multiply_multiply_instance ${GEMM_MULTIPLY_MULTIPLY_INSTANCES})
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp
View file @ c2e13d9b
...@@ -46,8 +46,7 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances = std ...@@ -46,8 +46,7 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances = std
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 16, 16, 8, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 16, 16, 16, 16, 8, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 224, 128, 16, 16, 16, 16, 8, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 224, 128, 16, 16, 16, 16, 8, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
......
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn.hpp
View file @ c2e13d9b
...@@ -45,8 +45,7 @@ using device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_instances = std::tuple< ...@@ -45,8 +45,7 @@ using device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_instances = std::tuple<
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 64, 16, 16, 16, 16, 8, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 64, 16, 16, 16, 16, 8, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 224, 128, 16, 16, 16, 16, 8, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 224, 128, 16, 16, 16, 16, 8, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
......
profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp
View file @ c2e13d9b
...@@ -183,105 +183,128 @@ bool profile_gemm_multiply_multiply_impl(int do_verification, ...@@ -183,105 +183,128 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
float best_ave_time = 0; float best_ave_time = 0;
float best_tflops = 0; float best_tflops = 0;
float best_gb_per_sec = 0; float best_gb_per_sec = 0;
float best_kbatch = 0;
// profile device GEMM instances // profile device GEMM instances
for(auto& op_ptr : op_ptrs) for(auto& op_ptr : op_ptrs)
{ {
auto argument_ptr = std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38};
op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
std::array<const void*, 2>{d0_device_buf.GetDeviceBuffer(),
d1_device_buf.GetDeviceBuffer()},
static_cast<EDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 2>{StrideD0, StrideD1},
StrideE,
KBatch,
a_element_op,
b_element_op,
c_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init C to zero before profiling next kernel
c_device_buf.SetZero();
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false, 0, n_warmup, n_iter}); if(KBatch > 0)
{
kbatch_list = {KBatch};
}
if(do_verification) for(std::size_t i = 0; i < kbatch_list.size(); i++)
{
auto kbatch_curr = kbatch_list[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
std::array<const void*, 2>{d0_device_buf.GetDeviceBuffer(),
d1_device_buf.GetDeviceBuffer()},
static_cast<EDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 2>{StrideD0, StrideD1},
StrideE,
kbatch_curr,
a_element_op,
b_element_op,
c_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{ {
c_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result); // re-init C to zero before profiling next kernel
c_device_buf.SetZero();
invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr, false, 0, n_warmup, n_iter});
if(do_log) if(do_verification)
{ {
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl; c_device_buf.FromDevice(e_m_n_device_result.mData.data());
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host : ", e_m_n_host_result.mData, ",") pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", e_m_n_device_result.mData, ",") if(do_log)
<< std::endl; {
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "c_host : ", e_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", e_m_n_device_result.mData, ",")
<< std::endl;
}
} }
}
std::string op_name = op_ptr->GetTypeString(); std::string op_name = op_ptr->GetTypeString();
float ave_time = invoker_ptr->Run( float ave_time = invoker_ptr->Run(argument_ptr.get(),
argument_ptr.get(), StreamConfig{nullptr,
StreamConfig{ time_kernel,
nullptr, time_kernel, 0, n_warmup, n_iter, rotating_count > 1, rotating_count}); 0,
n_warmup,
n_iter,
rotating_count > 1,
rotating_count});
std::size_t flop = std::size_t(2) * M * N * K; std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N; sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time; float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, " std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
<< gb_per_sec << " GB/s, " << op_name << std::endl; << " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
<< kbatch_curr << std::endl;
#if defined CK_ENABLE_FP8 #if defined CK_ENABLE_FP8
// set softer tolerances for fp8 // set softer tolerances for fp8
if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> || if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
is_same_v<EDataType, f8_t>) is_same_v<EDataType, f8_t>)
{ {
std::string msg = "Error: Incorrect results!"; std::string msg = "Error: Incorrect results!";
double rtol = 1e-1; double rtol = 1e-1;
double atol = 1e-1; double atol = 1e-1;
pass = pass & ck::utils::check_err( pass = pass & ck::utils::check_err(
e_m_n_device_result, e_m_n_host_result, msg, rtol, atol); e_m_n_device_result, e_m_n_host_result, msg, rtol, atol);
} }
else else
{ {
#endif #endif
pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result); pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
#if defined CK_ENABLE_FP8 #if defined CK_ENABLE_FP8
} }
#endif #endif
if(tflops > best_tflops) if(tflops > best_tflops && ave_time > 1e-10)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
best_kbatch = kbatch_curr;
}
}
else
{ {
best_op_name = op_name; std::cout << op_ptr->GetTypeString() << " does not support this problem"
best_tflops = tflops; << std::endl;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
} }
} }
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
}
} }
if constexpr(is_same<EDataType, float>::value) if constexpr(is_same<EDataType, float>::value)
...@@ -320,9 +343,9 @@ bool profile_gemm_multiply_multiply_impl(int do_verification, ...@@ -320,9 +343,9 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
} }
std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
<< " StrideB = " << StrideB << " StrideE = " << StrideE << " : " << best_ave_time << " StrideB = " << StrideB << " StrideE = " << StrideE << " KBatch = " << best_kbatch
<< " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, " << " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
<< best_op_name << std::endl; << " GB/s, " << best_op_name << std::endl;
return pass; return pass;
} }
......
profiler/include/profiler/profile_gemm_universal_impl.hpp
View file @ c2e13d9b
...@@ -152,7 +152,7 @@ bool profile_gemm_universal_impl(int do_verification, ...@@ -152,7 +152,7 @@ bool profile_gemm_universal_impl(int do_verification,
// profile device GEMM instances // profile device GEMM instances
for(auto& op_ptr : op_ptrs) for(auto& op_ptr : op_ptrs)
{ {
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 19, 20, 32, 38}; std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38};
if(KBatch > 0) if(KBatch > 0)
{ {
...@@ -249,7 +249,7 @@ bool profile_gemm_universal_impl(int do_verification, ...@@ -249,7 +249,7 @@ bool profile_gemm_universal_impl(int do_verification,
<< " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch " << " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
<< kbatch_curr << std::endl; << kbatch_curr << std::endl;
if(tflops > best_tflops) if(tflops > best_tflops && ave_time > 1e-10)
{ {
best_op_name = op_name; best_op_name = op_name;
best_tflops = tflops; best_tflops = tflops;
......
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