Skip to content

GitLab

Commit 2c1ed8b2 authored by Anthony Chang's avatar Anthony Chang
Browse files

Merge remote-tracking branch 'upstream/develop' into gemm-layernorm-4

parents b86b318b 56adf7e9
Hide whitespace changes
Inline Side-by-side
Showing with 1577 additions and 116 deletions
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp 0 → 100644
View file @ 2c1ed8b2
#include <stdlib.h>
#include "config.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
#include "device_gemm_multiple_d_xdl_cshuffle.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16_F16 = ck::Tuple<F16, F16>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// e = elementwise((a * b), d)
// outout: e[m, n]
// input: a[m, k], b[k, n], d[m, n]
using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::tuple<
// clang-format off
//##############################| ALayout| BLayout| CLayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| 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| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| 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| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmMultipleDPtr<2, PassThrough, PassThrough, AddAddFastGelu>>& instances)
{
add_device_operation_instances(
instances, device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp 0 → 100644
View file @ 2c1ed8b2
#include <stdlib.h>
#include "config.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
#include "device_gemm_multiple_d_xdl_cshuffle.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16_F16 = ck::Tuple<F16, F16>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// e = elementwise((a * b), d)
// outout: e[m, n]
// input: a[m, k], b[n, k], d[m, n]
using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances = std::tuple<
// clang-format off
//##############################| ALayout| BLayout| CLayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| 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| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| 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| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
// clang-format on
>;
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmMultipleDPtr<2, PassThrough, PassThrough, AddAddFastGelu>>& instances)
{
add_device_operation_instances(
instances, device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt 0 → 100644
View file @ 2c1ed8b2
set(DEVICE_GEMM_REDUCE_INSTANCE_SOURCE
device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
)
add_instance_library(device_gemm_bias_add_reduce_instance ${DEVICE_GEMM_REDUCE_INSTANCE_SOURCE})
install(TARGETS device_gemm_bias_add_reduce_instance LIBRARY DESTINATION lib)
clang_tidy_check(device_gemm_bias_add_reduce_instance)
library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp 0 → 100644
View file @ 2c1ed8b2
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_bias_add_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reduction_operator.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[m, n] = a[k, m] * b[k, n]
using device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| C1| Dxs| DxsInEleOp| DxsAccEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 2, 2, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
std::vector<DeviceGemmBiasAddReducePtr<PassThrough,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp 0 → 100644
View file @ 2c1ed8b2
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_bias_add_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reduction_operator.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[m, n] = a[k, m] * b[n, k]
using device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| C1| Dxs| DxsInEleOp| DxsAccEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
std::vector<DeviceGemmBiasAddReducePtr<PassThrough,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp 0 → 100644
View file @ 2c1ed8b2
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_bias_add_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reduction_operator.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[m, n] = a[m, k] * b[n, k]
using device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout| AData| BData| CData|C0Data|C1Data| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| C1| Dxs| DxsInEleOp| DxsAccEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmBiasAddReducePtr<PassThrough,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp 0 → 100644
View file @ 2c1ed8b2
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_bias_add_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reduction_operator.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[m, n] = a[m, k] * b[n, k]
using device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| C1| Dxs| DxsInEleOp| DxsAccEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>,
DeviceGemmBiasAddReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>
// clang-format on
>;
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
std::vector<DeviceGemmBiasAddReducePtr<PassThrough,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
View file @ 2c1ed8b2
...@@ -21,12 +21,12 @@ template <ck::index_t... Is> ...@@ -21,12 +21,12 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>; using ReduceSum = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>; using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>; using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>; using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>; using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>; using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>; using DOutElementOps = ck::Tuple<Div, Div>;
...@@ -62,12 +62,9 @@ using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances = s ...@@ -62,12 +62,9 @@ using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances = s
>; >;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances( void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances(
std::vector<DeviceGemmReducePtr<DPtrsGlobal, std::vector<
PassThrough, DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, DInElementOps, DOutElementOps>>&
PassThrough, instances)
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{ {
add_device_operation_instances( add_device_operation_instances(
instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances{}); instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances{});
......
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
View file @ 2c1ed8b2
...@@ -21,12 +21,12 @@ template <ck::index_t... Is> ...@@ -21,12 +21,12 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>; using ReduceSum = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>; using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>; using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>; using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>; using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>; using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>; using DOutElementOps = ck::Tuple<Div, Div>;
...@@ -62,12 +62,9 @@ using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances = s ...@@ -62,12 +62,9 @@ using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances = s
>; >;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances( void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances(
std::vector<DeviceGemmReducePtr<DPtrsGlobal, std::vector<
PassThrough, DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, DInElementOps, DOutElementOps>>&
PassThrough, instances)
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{ {
add_device_operation_instances( add_device_operation_instances(
instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances{}); instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances{});
......
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
View file @ 2c1ed8b2
...@@ -21,12 +21,12 @@ template <ck::index_t... Is> ...@@ -21,12 +21,12 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>; using ReduceSum = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>; using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>; using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>; using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>; using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>; using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>; using DOutElementOps = ck::Tuple<Div, Div>;
...@@ -62,12 +62,9 @@ using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances = s ...@@ -62,12 +62,9 @@ using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances = s
>; >;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances( void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmReducePtr<DPtrsGlobal, std::vector<
PassThrough, DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, DInElementOps, DOutElementOps>>&
PassThrough, instances)
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{ {
add_device_operation_instances( add_device_operation_instances(
instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances{}); instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances{});
......
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
View file @ 2c1ed8b2
...@@ -21,12 +21,12 @@ template <ck::index_t... Is> ...@@ -21,12 +21,12 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>; using ReduceSum = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>; using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>; using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>; using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>; using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>; using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>; using DOutElementOps = ck::Tuple<Div, Div>;
...@@ -59,12 +59,9 @@ using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances = s ...@@ -59,12 +59,9 @@ using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances = s
>; >;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances( void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances(
std::vector<DeviceGemmReducePtr<DPtrsGlobal, std::vector<
PassThrough, DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, DInElementOps, DOutElementOps>>&
PassThrough, instances)
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{ {
add_device_operation_instances( add_device_operation_instances(
instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances{}); instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances{});
......
profiler/CMakeLists.txt
View file @ 2c1ed8b2
...@@ -29,6 +29,7 @@ set(PROFILER_SOURCE ...@@ -29,6 +29,7 @@ set(PROFILER_SOURCE
src/profile_gemm_bias_relu.cpp src/profile_gemm_bias_relu.cpp
src/profile_gemm_bias_relu_add.cpp src/profile_gemm_bias_relu_add.cpp
src/profile_gemm_reduce.cpp src/profile_gemm_reduce.cpp
src/profile_gemm_bias_add_reduce.cpp
src/profile_batched_gemm.cpp src/profile_batched_gemm.cpp
src/profile_conv_fwd_bias_relu.cpp src/profile_conv_fwd_bias_relu.cpp
src/profile_conv_fwd_bias_relu_add.cpp src/profile_conv_fwd_bias_relu_add.cpp
...@@ -39,6 +40,7 @@ set(PROFILER_SOURCE ...@@ -39,6 +40,7 @@ set(PROFILER_SOURCE
src/profile_grouped_gemm.cpp src/profile_grouped_gemm.cpp
src/profile_conv_bwd_weight.cpp src/profile_conv_bwd_weight.cpp
src/profile_batched_gemm_reduce.cpp src/profile_batched_gemm_reduce.cpp
src/profile_gemm_add_add_fastgelu.cpp
) )
add_executable(ckProfiler ${PROFILER_SOURCE}) add_executable(ckProfiler ${PROFILER_SOURCE})
...@@ -46,6 +48,7 @@ add_executable(ckProfiler ${PROFILER_SOURCE}) ...@@ -46,6 +48,7 @@ add_executable(ckProfiler ${PROFILER_SOURCE})
target_link_libraries(ckProfiler PRIVATE host_tensor) target_link_libraries(ckProfiler PRIVATE host_tensor)
target_link_libraries(ckProfiler PRIVATE conv_util) target_link_libraries(ckProfiler PRIVATE conv_util)
target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_add_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias2d_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_bias2d_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_instance)
...@@ -62,3 +65,4 @@ target_link_libraries(ckProfiler PRIVATE device_reduce_instance) ...@@ -62,3 +65,4 @@ target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance) target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_weight_instance) target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_weight_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance) target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_add_add_fastgelu_instance)
profiler/include/profile_batched_gemm_reduce_impl.hpp
View file @ 2c1ed8b2
...@@ -20,13 +20,12 @@ namespace device_gemm_instance { ...@@ -20,13 +20,12 @@ namespace device_gemm_instance {
using F32 = float; using F32 = float;
using F16 = ck::half_t; using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>; using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>; using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>; using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>; using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>; using DOutElementOps = ck::Tuple<Identity, Identity>;
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr< using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
DPtrsGlobal,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
...@@ -129,17 +128,15 @@ bool profile_batched_gemm_reduce_impl(int do_verification, ...@@ -129,17 +128,15 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread); b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
} }
using AElementOp = ck::tensor_operation::element_wise::PassThrough; using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough; using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough; using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>; using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add<float>; using D1ReduceOp = ck::reduce::Add;
using UnaryIdenticElementOp = using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
ck::tensor_operation::element_wise::UnaryIdentic<float, float, false>; using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using UnarySquareElementOp = using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
ck::tensor_operation::element_wise::UnarySquare<float, float, false>; using DxsOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
const auto a_element_op = AElementOp{}; const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{}; const auto b_element_op = BElementOp{};
...@@ -171,8 +168,8 @@ bool profile_batched_gemm_reduce_impl(int do_verification, ...@@ -171,8 +168,8 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
{ {
for(int m = 0; m < M; ++m) for(int m = 0; m < M; ++m)
{ {
float d0_acc = d0_reduce_op.GetIdentityValue(); float d0_acc = d0_reduce_op.GetIdentityValue<float>();
float d1_acc = d1_reduce_op.GetIdentityValue(); float d1_acc = d1_reduce_op.GetIdentityValue<float>();
for(int n = 0; n < N; ++n) for(int n = 0; n < N; ++n)
{ {
...@@ -260,7 +257,7 @@ bool profile_batched_gemm_reduce_impl(int do_verification, ...@@ -260,7 +257,7 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()), gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()), static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()), static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
dxs_global, &dxs_global,
M, M,
N, N,
K, K,
......
profiler/include/profile_gemm_add_add_fastgelu_impl.hpp 0 → 100644
View file @ 2c1ed8b2
#pragma once
#include <iomanip>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "device_gemm_multiple_d.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using DeviceGemmAddAddFastGeluPtr = ck::tensor_operation::device::DeviceGemmMultipleDPtr<
2,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddAddFastGelu>;
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmAddAddFastGeluPtr>&);
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmAddAddFastGeluPtr>&);
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmAddAddFastGeluPtr>&);
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmAddAddFastGeluPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename D0DataType,
typename D1DataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename D0Layout,
typename D1Layout,
typename ELayout>
int profile_gemm_add_add_fastgelu_impl(int do_verification,
int init_method,
bool /*do_log*/,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideD0,
int StrideD1,
int StrideE)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddAddFastGelu;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmAddAddFastGeluPtr>
device_op_ptrs;
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<EDataType, half_t>)
{
if constexpr(is_same_v<ALayout, tensor_layout::gemm::RowMajor> &&
is_same_v<BLayout, tensor_layout::gemm::RowMajor> &&
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
device_op_ptrs);
}
else if constexpr(is_same_v<ALayout, tensor_layout::gemm::RowMajor> &&
is_same_v<BLayout, tensor_layout::gemm::ColumnMajor> &&
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
device_op_ptrs);
}
else if constexpr(is_same_v<ALayout, tensor_layout::gemm::ColumnMajor> &&
is_same_v<BLayout, tensor_layout::gemm::RowMajor> &&
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
device_op_ptrs);
}
else if constexpr(is_same_v<ALayout, tensor_layout::gemm::ColumnMajor> &&
is_same_v<BLayout, tensor_layout::gemm::ColumnMajor> &&
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
device_op_ptrs);
}
}
std::cout << "found " << device_op_ptrs.size() << " instances" << std::endl;
// run reference
if(do_verification)
{
Tensor<AccDataType> c_m_n(HostTensorDescriptor(
std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpace());
DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpace());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
std::string best_device_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
bool pass = true;
// profile device operation instances
for(auto& device_op_ptr : device_op_ptrs)
{
auto argument_ptr = device_op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
d1_m_n_device_buf.GetDeviceBuffer()},
static_cast<EDataType*>(e_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 2>{StrideD0, StrideD1},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = device_op_ptr->MakeInvokerPointer();
std::string device_op_name = device_op_ptr->GetTypeString();
if(device_op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init E to zero before profiling a kernel
e_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << device_op_name << std::endl;
if(tflops > best_tflops)
{
best_device_op_name = device_op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = pass &&
ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData);
}
}
else
{
std::cout << device_op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_device_op_name << std::endl;
return pass ? 0 : 1;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_gemm_bias_add_reduce_impl.hpp 0 → 100644
View file @ 2c1ed8b2
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "reduction_operator.hpp"
#include "device_gemm_reduce.hpp"
#include "reference_gemm.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmBiasAddReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmBiasAddReducePtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
DInElementOps,
DOutElementOps>;
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename C0DataType,
typename C1DataType,
typename DDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_gemm_bias_add_reduce_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int StrideC1)
{
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}),
std::vector<std::size_t>({stride}));
};
auto f_host_tensor_descriptor2d =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<C0DataType> bias_n(f_host_tensor_descriptor1d(N, 1));
Tensor<C1DataType> c1_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl;
std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl;
std::size_t num_thread = 1;
switch(init_method)
{
case 0: break;
case 1:
std::srand(0);
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
bias_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
c1_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
break;
default:
std::srand(0);
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
bias_n.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5}, num_thread);
c1_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
using C1ElementOp = PassThrough;
using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto c1_element_op = C1ElementOp{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{N, N};
if(do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
DDataType,
AElementOp,
BElementOp,
CElementOp>;
using ReduceAccDataType = DDataType;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
for(int n = 0; n < N; ++n)
{
ReduceAccDataType acc = static_cast<ReduceAccDataType>(c_m_n_host_result(m, n)) +
static_cast<ReduceAccDataType>(bias_n(n));
ReduceAccDataType c1 = static_cast<ReduceAccDataType>(c1_m_n(m, n));
c_element_op(acc, acc);
c1_element_op(c1, c1);
acc += c1;
c_m_n_host_result(m, n) = static_cast<CDataType>(acc);
}
for(int m = 0; m < M; ++m)
{
auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
ReduceAccDataType c_val =
ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
ReduceAccDataType d0_val;
ReduceAccDataType d1_val;
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
}
dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem bias_device_buf(sizeof(C0DataType) * bias_n.mDesc.GetElementSpace());
DeviceMem c1_device_buf(sizeof(C1DataType) * c1_m_n.mDesc.GetElementSpace());
DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
bias_device_buf.ToDevice(bias_n.mData.data());
c1_device_buf.ToDevice(c1_m_n.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmBiasAddReduceNoOpPtr>
gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
{
auto argument_ptr = gemm_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<C0DataType*>(bias_device_buf.GetDeviceBuffer()),
static_cast<C1DataType*>(c1_device_buf.GetDeviceBuffer()),
&dxs_global,
M,
N,
K,
StrideA,
StrideB,
StrideC,
StrideC1,
a_element_op,
b_element_op,
c_element_op,
c1_element_op,
dxs_in_element_op,
dxs_out_element_op);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::string gemm_name = gemm_ptr->GetTypeString();
std::size_t flop = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
std::size_t num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(C0DataType) * M * N +
sizeof(C1DataType) * M * N + sizeof(DDataType) * M +
sizeof(DDataType) * M;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
d0_device_buf.FromDevice(d0_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_m_device_result.mData.data());
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
ck::utils::check_err(d0_m_device_result.mData, d0_m_host_result.mData);
ck::utils::check_err(d1_m_device_result.mData, d1_m_host_result.mData);
if(do_log)
{
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: ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d0_host: ", d0_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d0_device: ", d0_m_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d1_host: ", d1_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d1_device: ", d1_m_device_result.mData, ",")
<< std::endl;
}
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_gemm_reduce_impl.hpp
View file @ 2c1ed8b2
#pragma once #pragma once
#include "check_err.hpp"
#include "config.hpp" #include "config.hpp"
#include "device.hpp" #include "device.hpp"
#include "host_tensor.hpp" #include "host_tensor.hpp"
...@@ -19,14 +20,13 @@ namespace device_gemm_instance { ...@@ -19,14 +20,13 @@ namespace device_gemm_instance {
using F32 = float; using F32 = float;
using F16 = ck::half_t; using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>; using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Div = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>; using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>; using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>; using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>; using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>; using DOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr< using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
DPtrsGlobal,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
...@@ -123,18 +123,16 @@ bool profile_gemm_reduce_impl(int do_verification, ...@@ -123,18 +123,16 @@ bool profile_gemm_reduce_impl(int do_verification,
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread); b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
} }
using AElementOp = ck::tensor_operation::element_wise::PassThrough; using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough; using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough; using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>; using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add<float>; using D1ReduceOp = ck::reduce::Add;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryIdentic<float, float, true>; using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using UnaryIdenticElementOp = using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
ck::tensor_operation::element_wise::UnaryIdentic<float, float, false>; using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using UnarySquareElementOp = using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
ck::tensor_operation::element_wise::UnarySquare<float, float, false>; using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
const auto a_element_op = AElementOp{}; const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{}; const auto b_element_op = BElementOp{};
...@@ -143,7 +141,7 @@ bool profile_gemm_reduce_impl(int do_verification, ...@@ -143,7 +141,7 @@ bool profile_gemm_reduce_impl(int do_verification,
const auto d1_reduce_op = D1ReduceOp{}; const auto d1_reduce_op = D1ReduceOp{};
auto dxs_in_element_op = DxsInElementOps{}; auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{M, M}; auto dxs_out_element_op = DxsOutElementOps{N, N};
if(do_verification) if(do_verification)
{ {
...@@ -155,6 +153,8 @@ bool profile_gemm_reduce_impl(int do_verification, ...@@ -155,6 +153,8 @@ bool profile_gemm_reduce_impl(int do_verification,
BElementOp, BElementOp,
CElementOp>; CElementOp>;
using ReduceAccDataType = DDataType;
auto ref_gemm = ReferenceGemmInstance{}; auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
...@@ -165,14 +165,15 @@ bool profile_gemm_reduce_impl(int do_verification, ...@@ -165,14 +165,15 @@ bool profile_gemm_reduce_impl(int do_verification,
for(int m = 0; m < M; ++m) for(int m = 0; m < M; ++m)
{ {
float d0_acc = d0_reduce_op.GetIdentityValue(); auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
float d1_acc = d1_reduce_op.GetIdentityValue(); auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n) for(int n = 0; n < N; ++n)
{ {
float c_val = ck::type_convert<float>(c_m_n_host_result(m, n)); ReduceAccDataType c_val =
float d0_val = 0; ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
float d1_val = 0; ReduceAccDataType d0_val;
ReduceAccDataType d1_val;
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val); dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val); dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
...@@ -257,7 +258,7 @@ bool profile_gemm_reduce_impl(int do_verification, ...@@ -257,7 +258,7 @@ bool profile_gemm_reduce_impl(int do_verification,
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()), gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()), static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()), static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
dxs_global, &dxs_global,
M, M,
N, N,
K, K,
...@@ -309,13 +310,9 @@ bool profile_gemm_reduce_impl(int do_verification, ...@@ -309,13 +310,9 @@ bool profile_gemm_reduce_impl(int do_verification,
d0_device_buf.FromDevice(d0_m_device_result.mData.data()); d0_device_buf.FromDevice(d0_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_m_device_result.mData.data()); d1_device_buf.FromDevice(d1_m_device_result.mData.data());
float c_error = check_error(c_m_n_host_result, c_m_n_device_result); ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
float d0_error = check_error(d0_m_host_result, d0_m_device_result); ck::utils::check_err(d0_m_device_result.mData, d0_m_host_result.mData);
float d1_error = check_error(d1_m_host_result, d1_m_device_result); ck::utils::check_err(d1_m_device_result.mData, d1_m_host_result.mData);
pass = pass && (c_error < 1E-6);
pass = pass && (d0_error < 1E-6);
pass = pass && (d1_error < 1E-6);
if(do_log) if(do_log)
{ {
......
profiler/include/profile_reduce_impl.hpp
View file @ 2c1ed8b2
...@@ -261,13 +261,18 @@ bool profile_reduce_impl_impl(bool do_verification, ...@@ -261,13 +261,18 @@ bool profile_reduce_impl_impl(bool do_verification,
float best_gb_per_sec = 0; float best_gb_per_sec = 0;
using InElementwiseOperation = using InElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>:: typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
InElementwiseOperation;
using AccElementwiseOperation = using AccElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>:: typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
AccElementwiseOperation;
using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType; using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
InElementwiseOperation in_elementwise_op;
AccElementwiseOperation acc_elementwise_op;
std::tie(in_elementwise_op, acc_elementwise_op) =
reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
static_cast<int32_t>(reduce_total_length));
using DeviceReduceInstPtr0 = using DeviceReduceInstPtr0 =
DeviceReducePtr<InElementwiseOperation, AccElementwiseOperation>; DeviceReducePtr<InElementwiseOperation, AccElementwiseOperation>;
...@@ -323,8 +328,13 @@ bool profile_reduce_impl_impl(bool do_verification, ...@@ -323,8 +328,13 @@ bool profile_reduce_impl_impl(bool do_verification,
OutputIndex> OutputIndex>
hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims); hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run( hostReduce.Run(alpha,
alpha, in.mData.data(), beta, out_ref.mData.data(), out_indices_ref.mData.data()); in.mData.data(),
beta,
out_ref.mData.data(),
out_indices_ref.mData.data(),
in_elementwise_op,
acc_elementwise_op);
}; };
std::vector<ck::index_t> i_inLengths; std::vector<ck::index_t> i_inLengths;
...@@ -339,10 +349,6 @@ bool profile_reduce_impl_impl(bool do_verification, ...@@ -339,10 +349,6 @@ bool profile_reduce_impl_impl(bool do_verification,
for(auto& reduce_ptr : reduce0_ptrs) for(auto& reduce_ptr : reduce0_ptrs)
{ {
InElementwiseOperation in_elementwise_op(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation acc_elementwise_op(static_cast<int32_t>(reduce_total_length));
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths, auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
i_inStrides, i_inStrides,
i_outLengths, i_outLengths,
......
profiler/src/profile_gemm_add_add_fastgelu.cpp 0 → 100644
View file @ 2c1ed8b2
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include "profile_gemm_add_add_fastgelu_impl.hpp"
int profile_gemm_add_add_fastgelu(int argc, char* argv[])
{
enum struct MatrixLayout
{
MK_KN_MN_MN_MN, // 0
MK_NK_MN_MN_MN, // 1
KM_KN_MN_MN_MN, // 2
KM_NK_MN_MN_MN, // 3
MK_KN_NM_MN_MN, // 4
MK_NK_NM_MN_MN, // 5
KM_KN_NM_MN_MN, // 6
KM_NK_NM_MN_MN, // 7
};
enum struct MatrixDataType
{
F32_F32_F32_F32_F32, // 0
F16_F16_F16_F16_F16, // 1
BF16_BF16_BF16_BF16_BF16, // 2
INT8_INT8_INT8_INT8_INT8, // 3
};
if(argc != 16)
{
// clang-format off
printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+GeLU)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n] + D1[m, n]);\n");
printf(" 1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n] + D1[m, n]);\n");
printf(" 2: E[m, n] = FastGeLU(A[k, m] * B[k, n] + D0[m, n] + D1[m, n]);\n");
printf(" 3: E[m, n] = FastGeLU(A[k, m] * B[n, k] + D0[m, n] + D1[m, n]))\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
// clang-format on
exit(1);
}
const auto data_type = static_cast<MatrixDataType>(std::stoi(argv[2]));
const auto layout = static_cast<MatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideD0 = std::stoi(argv[13]);
const int StrideD1 = std::stoi(argv[14]);
const int StrideE = std::stoi(argv[15]);
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto acc_type,
auto d0_type,
auto d1_type,
auto e_type,
auto a_layout,
auto b_layout,
auto d0_layout,
auto d1_layout,
auto e_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
using D0DataType = decltype(d0_type);
using D1DataType = decltype(d1_type);
using EDataType = decltype(e_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using D0Layout = decltype(d0_layout);
using D1Layout = decltype(d1_layout);
using ELayout = decltype(e_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
return ck::profiler::profile_gemm_add_add_fastgelu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
D1DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
D1Layout,
ELayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
(StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
(StrideE < 0) ? DefaultStrideE : StrideE);
};
if(data_type == MatrixDataType::F16_F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::MK_NK_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::KM_KN_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::KM_NK_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 0;
}
}
profiler/src/profile_gemm_bias_add_reduce.cpp 0 → 100644
View file @ 2c1ed8b2
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "profile_gemm_bias_add_reduce_impl.hpp"
int profile_gemm_bias_add_reduce(int argc, char* argv[])
{
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
enum struct GemmReduceDataType
{
F32_F32_F32_F32_F32_F32_F32, // 0
F16_F16_F16_F16_F16_F32_F32, // 1
};
if(!(argc == 14 || argc == 15))
{
printf("arg1: tensor operation (gemm: GEMM+bias+add+Reduce)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, StrideC1\n");
exit(1);
}
const auto data_type = static_cast<GemmReduceDataType>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
const int StrideC1 = std::stoi(argv[14]);
if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else
{
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 0;
}
profiler/src/profiler.cpp
View file @ 2c1ed8b2
...@@ -11,6 +11,7 @@ int profile_gemm_bias_2d(int, char*[]); ...@@ -11,6 +11,7 @@ int profile_gemm_bias_2d(int, char*[]);
int profile_gemm_bias_relu(int, char*[]); int profile_gemm_bias_relu(int, char*[]);
int profile_gemm_bias_relu_add(int, char*[]); int profile_gemm_bias_relu_add(int, char*[]);
int profile_gemm_reduce(int, char*[]); int profile_gemm_reduce(int, char*[]);
int profile_gemm_bias_add_reduce(int, char*[]);
int profile_batched_gemm(int, char*[]); int profile_batched_gemm(int, char*[]);
int profile_grouped_gemm(int, char*[]); int profile_grouped_gemm(int, char*[]);
int profile_conv_fwd(int, char*[]); int profile_conv_fwd(int, char*[]);
...@@ -21,9 +22,39 @@ int profile_convnd_bwd_data(int, char*[], int); ...@@ -21,9 +22,39 @@ int profile_convnd_bwd_data(int, char*[], int);
int profile_reduce(int, char*[]); int profile_reduce(int, char*[]);
int profile_conv_bwd_weight(int, char*[]); int profile_conv_bwd_weight(int, char*[]);
int profile_batched_gemm_reduce(int, char*[]); int profile_batched_gemm_reduce(int, char*[]);
int profile_gemm_add_add_fastgelu(int, char*[]);
static void print_helper_message()
{
// clang-format off
printf("arg1: tensor operation (gemm: GEMM\n"
" gemm_bias_2d: GEMM+Bias(2D)\n"
" gemm_bias_relu: GEMM+Bias+ReLU\n"
" gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n"
" gemm_reduce: GEMM+Reduce\n"
" grouped_gemm: Grouped GEMM\n"
" conv_fwd: ForwardConvolution\n"
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
" conv_fwd_bias_relu_atomic_add: ForwardConvolution+Bias+ReLU+AtomicAdd\n"
" conv1d_bwd_data: BackwardConvolution data 1 dim\n"
" conv2d_bwd_data: BackwardConvolution data 2 dim\n"
" conv3d_bwd_data: BackwardConvolution data 3 dim\n"
" reduce: Reduce\n"
" conv2d_bwd_weight: Backward Weight Convolution 2d\n"
" gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n");
// clang-format on
}
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
if(argc == 1)
{
print_helper_message();
return 0;
}
if(strcmp(argv[1], "gemm") == 0) if(strcmp(argv[1], "gemm") == 0)
{ {
return profile_gemm(argc, argv); return profile_gemm(argc, argv);
...@@ -44,6 +75,10 @@ int main(int argc, char* argv[]) ...@@ -44,6 +75,10 @@ int main(int argc, char* argv[])
{ {
return profile_gemm_reduce(argc, argv); return profile_gemm_reduce(argc, argv);
} }
else if(strcmp(argv[1], "gemm_bias_add_reduce") == 0)
{
return profile_gemm_bias_add_reduce(argc, argv);
}
else if(strcmp(argv[1], "batched_gemm") == 0) else if(strcmp(argv[1], "batched_gemm") == 0)
{ {
return profile_batched_gemm(argc, argv); return profile_batched_gemm(argc, argv);
...@@ -92,25 +127,14 @@ int main(int argc, char* argv[]) ...@@ -92,25 +127,14 @@ int main(int argc, char* argv[])
{ {
return profile_conv_bwd_weight(argc, argv); return profile_conv_bwd_weight(argc, argv);
} }
else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0)
{
return profile_gemm_add_add_fastgelu(argc, argv);
}
else else
{ {
// clang-format off print_helper_message();
printf("arg1: tensor operation (gemm: GEMM\n"
" gemm_bias_2d: GEMM+Bias(2D)\n" return 0;
" gemm_bias_relu: GEMM+Bias+ReLU\n"
" gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n"
" gemm_reduce: GEMM+Reduce\n"
" grouped_gemm: Grouped GEMM\n"
" conv_fwd: ForwardConvolution\n"
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
" conv_fwd_bias_relu_atomic_add: ForwardConvolution+Bias+ReLU+AtomicAdd\n"
" conv1d_bwd_data: BackwardConvolution data 1 dim\n"
" conv2d_bwd_data: BackwardConvolution data 2 dim\n"
" conv3d_bwd_data: BackwardConvolution data 3 dim\n"
" reduce: Reduce\n"
" conv2d_bwd_weight: Backward Weight Convolution 2d\n");
// clang-format on
} }
return 0;
} }
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