Skip to content

GitLab

Commit 05fd7ff8 authored by Jakub Piasecki's avatar Jakub Piasecki
Browse files

Merge remote-tracking branch 'origin/develop' into gemm_f16_int8

parents 2784b516 84832fc4
Hide whitespace changes
Inline Side-by-side
Showing with 1099 additions and 114 deletions
library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
View file @ 05fd7ff8
......@@ -27,6 +27,7 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmKPadding = ck::tensor_operation::device::GemmSpecialization::KPadding;
static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
......@@ -110,17 +111,39 @@ using device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances = std::tuple<
// clang-format on
>;
template <ck::tensor_operation::device::GemmSpecialization GemmSpec>
template <ck::tensor_operation::device::GemmSpecialization GemmSpec,
ck::PipelineVersion PipVer,
ck::LoopScheduler LoopSche>
using device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances = std::tuple<
// clang-format off
//#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 4, 8, 16, 16, 1, 4, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 256, 4, 8, 16, 16, 1, 8, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 4, 8, 16, 16, 1, 4, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipelineVersion::v1, LoopScheduler::Interwave>
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 4, 8, 16, 16, 1, 4, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 256, 4, 8, 16, 16, 1, 8, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 4, 8, 16, 16, 1, 4, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 16>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 512, 4, 8, 16, 16, 1, 8, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 16>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 4, 8, 16, 16, 4, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 32, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 16, 4, 8, 16, 16, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 32, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 4, 8, 16, 16, 4, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 64, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 512, 16, 4, 8, 16, 16, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 64, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 8, 8, 16, 16, 1, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 16, 8, 16, 16, 1, 1, S<1, 16, 4, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 8, 8, 16, 16, 1, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 8, 8, 16, 16, 1, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 8, 8, 16, 16, 1, 4, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 256, 8, 8, 16, 16, 1, 8, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 8, 8, 16, 16, 1, 4, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 16>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 8, 8, 16, 16, 1, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 32, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 8, 8, 16, 16, 2, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 32, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 8, 8, 16, 16, 4, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 32, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 16, 8, 8, 16, 16, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 32, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 8, 8, 16, 16, 4, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 64, 1, 4>, 4, F16, PipVer, LoopSche>
// clang-format on
>;
......@@ -141,9 +164,51 @@ void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
add_device_operation_instances(
instances, device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances<GemmMNKPadding>{});
add_device_operation_instances(
instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances<GemmMNKPadding>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances<
GemmDefault,
ck::PipelineVersion::v1,
ck::LoopScheduler::Default>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances<
GemmDefault,
ck::PipelineVersion::v2,
ck::LoopScheduler::Default>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances<
GemmDefault,
ck::PipelineVersion::v1,
ck::LoopScheduler::Interwave>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances<
GemmKPadding,
ck::PipelineVersion::v1,
ck::LoopScheduler::Default>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances<
GemmKPadding,
ck::PipelineVersion::v2,
ck::LoopScheduler::Default>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances<
GemmKPadding,
ck::PipelineVersion::v1,
ck::LoopScheduler::Interwave>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances<
GemmMNKPadding,
ck::PipelineVersion::v1,
ck::LoopScheduler::Default>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances<
GemmMNKPadding,
ck::PipelineVersion::v2,
ck::LoopScheduler::Default>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances<
GemmMNKPadding,
ck::PipelineVersion::v1,
ck::LoopScheduler::Interwave>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
View file @ 05fd7ff8
......@@ -27,6 +27,7 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmKPadding = ck::tensor_operation::device::GemmSpecialization::KPadding;
static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
......@@ -95,6 +96,41 @@ using device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances = std::tuple<
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F16, PipelineVersion::v2>
// clang-format on
>;
template <ck::tensor_operation::device::GemmSpecialization GemmSpec,
ck::PipelineVersion PipVer,
ck::LoopScheduler LoopSche>
using device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances = std::tuple<
// clang-format off
//#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 4, 8, 16, 16, 1, 4, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 256, 4, 8, 16, 16, 1, 8, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 4, 8, 16, 16, 1, 4, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 16>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 512, 4, 8, 16, 16, 1, 8, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 16>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 4, 8, 16, 16, 4, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 16, 4, 8, 16, 16, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 4, 8, 16, 16, 4, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 64, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 512, 16, 4, 8, 16, 16, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 64, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 8, 8, 16, 16, 1, 1, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 8, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 16, 8, 16, 16, 1, 1, S<1, 16, 4, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 16, 4, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 8, 8, 16, 16, 1, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 8, 8, 16, 16, 1, 2, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 8, 8, 16, 16, 1, 4, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 256, 8, 8, 16, 16, 1, 8, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 8, 8, 16, 16, 1, 4, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 16>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 8, 8, 16, 16, 1, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 8, 8, 16, 16, 2, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 8, 8, 16, 16, 4, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 16, 8, 8, 16, 16, 8, 1, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 4, F16, PipVer, LoopSche>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 8, 8, 16, 16, 4, 1, S<1, 8, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 64, 1, 4>, 4, F16, PipVer, LoopSche>
// clang-format on
>;
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
......@@ -112,6 +148,52 @@ void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
add_device_operation_instances(
instances, device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances<GemmMNKPadding>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances<
GemmDefault,
ck::PipelineVersion::v1,
ck::LoopScheduler::Default>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances<
GemmDefault,
ck::PipelineVersion::v2,
ck::LoopScheduler::Default>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances<
GemmDefault,
ck::PipelineVersion::v1,
ck::LoopScheduler::Interwave>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances<
GemmKPadding,
ck::PipelineVersion::v1,
ck::LoopScheduler::Default>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances<
GemmKPadding,
ck::PipelineVersion::v2,
ck::LoopScheduler::Default>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances<
GemmKPadding,
ck::PipelineVersion::v1,
ck::LoopScheduler::Interwave>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances<
GemmMNKPadding,
ck::PipelineVersion::v1,
ck::LoopScheduler::Default>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances<
GemmMNKPadding,
ck::PipelineVersion::v2,
ck::LoopScheduler::Default>{});
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances<
GemmMNKPadding,
ck::PipelineVersion::v1,
ck::LoopScheduler::Interwave>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_layernorm2d_bwd_gamma_beta_f16_instance.cpp
View file @ 05fd7ff8
......@@ -8,7 +8,7 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_layernorm2d_bwd_gamma_beta_rank_2_1_f16_instances(
void add_device_layernorm2d_bwd_gamma_beta_f16_instances(
std::vector<std::unique_ptr<DeviceNormalizationBwdGammaBeta<F16, F16, F16, F16, F16, 2, 1>>>&
instances)
{
......
library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_layernorm2d_bwd_gamma_beta_f32_instance.cpp
View file @ 05fd7ff8
......@@ -8,7 +8,7 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_layernorm2d_bwd_gamma_beta_rank_2_1_f32_instances(
void add_device_layernorm2d_bwd_gamma_beta_f32_instances(
std::vector<std::unique_ptr<DeviceNormalizationBwdGammaBeta<F32, F32, F32, F32, F32, 2, 1>>>&
instances)
{
......
profiler/include/profiler/profile_gemm_impl.hpp
View file @ 05fd7ff8
......@@ -298,7 +298,7 @@ int profile_gemm_impl(int do_verification,
}
}
return pass ? 0 : 1;
return pass;
}
} // namespace profiler
......
profiler/include/profiler/profile_gemm_splitk_impl.hpp
View file @ 05fd7ff8
......@@ -145,7 +145,7 @@ bool profile_gemm_splitk_impl(int do_verification,
// profile device GEMM instances
for(auto& op_ptr : op_ptrs)
{
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 32, 36, 40, 64, 96, 128};
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 19, 20, 32, 38};
if(KBatch > 0)
{
......
profiler/include/profiler/profile_groupnorm_bwd_gamma_beta_impl.hpp 0 → 100644
View file @ 05fd7ff8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/groupnorm_bwd_gamma_beta.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_groupnorm_bwd.hpp"
namespace ck {
namespace profiler {
template <typename DYDataType,
typename XDataType,
typename MeanInvStdDataType,
typename ComputeDataType,
typename DGammaDataType,
typename DBetaDataType>
bool profile_groupnorm_bwd_gamma_beta_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
std::vector<index_t> length)
{
// we don't need GammaDataType and DXDataType here, just for reference class
using GammaDataType = DYDataType;
using DXDataType = DYDataType;
if(length.size() != 5)
return false;
index_t N = length[0];
index_t G = length[3];
index_t C = length[4];
std::vector<index_t> reduce_dim = {0, 1, 2};
std::vector<index_t> gamma_beta_length = {G, C};
Tensor<DYDataType> dy(length);
Tensor<XDataType> x(length);
Tensor<GammaDataType> gamma(gamma_beta_length); // dummy tensor, for reference
Tensor<MeanInvStdDataType> mean({N, G});
Tensor<MeanInvStdDataType> inv_std({N, G});
Tensor<DGammaDataType> dgamma(gamma_beta_length);
Tensor<DBetaDataType> dbeta(gamma_beta_length);
Tensor<DXDataType> host_dx(length); // dummy tensor, for reference
Tensor<DGammaDataType> host_dgamma(gamma_beta_length);
Tensor<DBetaDataType> host_dbeta(gamma_beta_length);
std::vector<index_t> strideDy =
std::vector<ck::index_t>{dy.mDesc.GetStrides().begin(), dy.mDesc.GetStrides().end()};
std::vector<index_t> strideX =
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()};
std::vector<index_t> strideDGamma{dgamma.mDesc.GetStrides().begin(),
dgamma.mDesc.GetStrides().end()};
std::vector<index_t> strideDBeta{dbeta.mDesc.GetStrides().begin(),
dbeta.mDesc.GetStrides().end()};
std::vector<index_t> strideMeanInvStd = {G, 0, 0, 1, 0};
switch(init_method)
{
case 0:
dy.GenerateTensorValue(GeneratorTensor_1<DYDataType>{});
x.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
mean.GenerateTensorValue(GeneratorTensor_1<MeanInvStdDataType>{});
inv_std.GenerateTensorValue(GeneratorTensor_1<MeanInvStdDataType>{});
dgamma.GenerateTensorValue(GeneratorTensor_1<DGammaDataType>{});
dbeta.GenerateTensorValue(GeneratorTensor_1<DBetaDataType>{});
break;
case 1:
dy.GenerateTensorValue(GeneratorTensor_2<DYDataType>{-5, 5});
x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5});
mean.GenerateTensorValue(GeneratorTensor_2<MeanInvStdDataType>{-5, 5});
inv_std.GenerateTensorValue(GeneratorTensor_2<MeanInvStdDataType>{0, 5});
dgamma.GenerateTensorValue(GeneratorTensor_2<DGammaDataType>{-5, 5});
dbeta.GenerateTensorValue(GeneratorTensor_2<DBetaDataType>{-5, 5});
break;
default:
dy.GenerateTensorValue(GeneratorTensor_3<DYDataType>{0, 1});
x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1});
mean.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{-0.5, 0.5});
inv_std.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0, 0.5});
dgamma.GenerateTensorValue(GeneratorTensor_3<DGammaDataType>{-0.5, 0.5});
dbeta.GenerateTensorValue(GeneratorTensor_3<DBetaDataType>{-0.5, 0.5});
}
DeviceMem dy_dev(sizeof(DYDataType) * dy.mDesc.GetElementSpaceSize());
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem mean_dev(sizeof(MeanInvStdDataType) * mean.mDesc.GetElementSpaceSize());
DeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * inv_std.mDesc.GetElementSpaceSize());
DeviceMem dgamma_dev(sizeof(DGammaDataType) * dgamma.mDesc.GetElementSpaceSize());
DeviceMem dbeta_dev(sizeof(DBetaDataType) * dbeta.mDesc.GetElementSpaceSize());
dy_dev.ToDevice(dy.mData.data());
x_dev.ToDevice(x.mData.data());
mean_dev.ToDevice(mean.mData.data());
inv_std_dev.ToDevice(inv_std.mData.data());
// add device normalization instances
using DeviceOp =
ck::tensor_operation::device::DeviceNormalizationBwdGammaBeta<DYDataType,
XDataType,
MeanInvStdDataType,
DGammaDataType,
DBetaDataType,
5,
3>;
// get device op instances
const auto instance_ptrs =
ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
std::string best_instance_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
if(do_verification)
{
using ReferenceInstance =
ck::tensor_operation::host::ReferenceGroupnormBwd<DYDataType,
XDataType,
GammaDataType,
MeanInvStdDataType,
DGammaDataType,
DBetaDataType,
DXDataType,
ComputeDataType>;
ReferenceInstance ref;
auto ref_argument =
ref.MakeArgument(dy, x, gamma, mean, inv_std, host_dgamma, host_dbeta, host_dx, length);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
}
std::size_t num_bytes = dy.mDesc.GetElementSize() * sizeof(DYDataType) +
x.mDesc.GetElementSize() * sizeof(XDataType) +
mean.mDesc.GetElementSize() * sizeof(MeanInvStdDataType) +
inv_std.mDesc.GetElementSize() * sizeof(MeanInvStdDataType) +
dgamma.mDesc.GetElementSize() * sizeof(DGammaDataType) +
dbeta.mDesc.GetElementSize() * sizeof(DBetaDataType);
int num_kernel = 0;
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = inst_ptr->MakeArgumentPointer(length,
strideDy,
strideX,
strideMeanInvStd,
strideMeanInvStd,
gamma_beta_length,
strideDGamma,
strideDBeta,
reduce_dim,
dy_dev.GetDeviceBuffer(),
x_dev.GetDeviceBuffer(),
mean_dev.GetDeviceBuffer(),
inv_std_dev.GetDeviceBuffer(),
dgamma_dev.GetDeviceBuffer(),
dbeta_dev.GetDeviceBuffer());
if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
++num_kernel;
}
else
{
if(time_kernel)
{
std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
LogRange(std::cout << "input lengths = ", length, ", ") << std::endl;
}
continue;
}
size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
float gb_per_sec = num_bytes / 1.E6 / avg_time;
if(time_kernel)
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< inst_ptr->GetTypeString() << std::endl;
if(avg_time < best_avg_time)
{
best_instance_name = inst_ptr->GetTypeString();
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
dgamma_dev.FromDevice(dgamma.mData.data());
dbeta_dev.FromDevice(dbeta.mData.data());
bool pass =
ck::utils::check_err(dgamma, host_dgamma, "Error: Incorrect dgamma", 1e-3, 1e-3);
pass &= ck::utils::check_err(dbeta, host_dbeta, "Error: Incorrect dbeta", 1e-3, 1e-3);
if(do_log)
{
LogRangeAsType<float>(std::cout << "dy : ", dy.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "host_dgamma : ", host_dgamma.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "dgamma : ", dgamma.mData, ",") << std::endl;
}
if(!pass)
{
std::cout << inst_ptr->GetTypeString() << " failed verification: ";
LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
return false;
}
else
{
if(time_kernel)
std::cout << "pass" << std::endl;
}
}
}
if(time_kernel)
{
LogRange(std::cout << "length = ", length, ",") << ", ";
LogRange(std::cout << "reduce dims ", reduce_dim, ",") << std::endl;
std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s,"
<< best_instance_name << std::endl;
}
if(num_kernel == 0)
{
std::cout << "Error: No kernel is applicable" << std::endl;
return false;
}
return true;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_layernorm_bwd_gamma_beta_impl.hpp 0 → 100644
View file @ 05fd7ff8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/layernorm_bwd_gamma_beta.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm_bwd.hpp"
namespace ck {
namespace profiler {
template <typename DYDataType,
typename XDataType,
typename MeanInvStdDataType,
typename ComputeDataType,
typename DGammaDataType,
typename DBetaDataType,
index_t Rank>
bool profile_layernorm_bwd_gamma_beta_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
std::vector<index_t> length)
{
// we don't need GammaDataType and DXDataType here, just for reference class
using GammaDataType = DYDataType;
using DXDataType = DYDataType;
if(length.size() != Rank || Rank < 2)
return false;
// Assume normalize dimension for first dimension
// Layernorm 2D, input = [M, K], reduce on M axis
// Layernorm 4D, input = [N, H, W, C], redice on N axis
constexpr int NumReduceDim = Rank - 1;
std::vector<index_t> reduce_dim = {0};
std::vector<index_t> invarient_length{length.begin() + 1, length.end()};
Tensor<DYDataType> dy(length);
Tensor<XDataType> x(length);
Tensor<GammaDataType> gamma(invarient_length); // dummy tensor, for reference
Tensor<MeanInvStdDataType> mean({length[0]});
Tensor<MeanInvStdDataType> inv_std({length[0]});
Tensor<DGammaDataType> dgamma(invarient_length);
Tensor<DBetaDataType> dbeta(invarient_length);
Tensor<DXDataType> host_dx(length); // dummy tensor, for reference
Tensor<DGammaDataType> host_dgamma(invarient_length);
Tensor<DBetaDataType> host_dbeta(invarient_length);
std::vector<index_t> strideDy =
std::vector<ck::index_t>{dy.mDesc.GetStrides().begin(), dy.mDesc.GetStrides().end()};
std::vector<index_t> strideX = strideDy;
std::vector<index_t> strideDGamma{dgamma.mDesc.GetStrides().begin(),
dgamma.mDesc.GetStrides().end()};
std::vector<index_t> strideDBeta{dbeta.mDesc.GetStrides().begin(),
dbeta.mDesc.GetStrides().end()};
std::vector<index_t> strideMeanInvStd{Rank, 0};
strideMeanInvStd[0] = 1;
switch(init_method)
{
case 0:
dy.GenerateTensorValue(GeneratorTensor_1<DYDataType>{});
x.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
mean.GenerateTensorValue(GeneratorTensor_1<MeanInvStdDataType>{});
inv_std.GenerateTensorValue(GeneratorTensor_1<MeanInvStdDataType>{});
dgamma.GenerateTensorValue(GeneratorTensor_1<DGammaDataType>{});
dbeta.GenerateTensorValue(GeneratorTensor_1<DBetaDataType>{});
break;
case 1:
dy.GenerateTensorValue(GeneratorTensor_2<DYDataType>{-5, 5});
x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5});
mean.GenerateTensorValue(GeneratorTensor_2<MeanInvStdDataType>{-5, 5});
inv_std.GenerateTensorValue(GeneratorTensor_2<MeanInvStdDataType>{0, 5});
dgamma.GenerateTensorValue(GeneratorTensor_2<DGammaDataType>{-5, 5});
dbeta.GenerateTensorValue(GeneratorTensor_2<DBetaDataType>{-5, 5});
break;
default:
dy.GenerateTensorValue(GeneratorTensor_3<DYDataType>{0, 1});
x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1});
mean.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{-0.5, 0.5});
inv_std.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0, 0.5});
dgamma.GenerateTensorValue(GeneratorTensor_3<DGammaDataType>{-0.5, 0.5});
dbeta.GenerateTensorValue(GeneratorTensor_3<DBetaDataType>{-0.5, 0.5});
}
DeviceMem dy_dev(sizeof(DYDataType) * dy.mDesc.GetElementSpaceSize());
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem mean_dev(sizeof(MeanInvStdDataType) * mean.mDesc.GetElementSpaceSize());
DeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * inv_std.mDesc.GetElementSpaceSize());
DeviceMem dgamma_dev(sizeof(DGammaDataType) * dgamma.mDesc.GetElementSpaceSize());
DeviceMem dbeta_dev(sizeof(DBetaDataType) * dbeta.mDesc.GetElementSpaceSize());
dy_dev.ToDevice(dy.mData.data());
x_dev.ToDevice(x.mData.data());
mean_dev.ToDevice(mean.mData.data());
inv_std_dev.ToDevice(inv_std.mData.data());
// add device normalization instances
using DeviceOp =
ck::tensor_operation::device::DeviceNormalizationBwdGammaBeta<DYDataType,
XDataType,
MeanInvStdDataType,
DGammaDataType,
DBetaDataType,
Rank,
NumReduceDim>;
// get device op instances
const auto instance_ptrs =
ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
std::string best_instance_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
if(do_verification)
{
using ReferenceInstance =
ck::tensor_operation::host::ReferenceLayernormBwd<DYDataType,
XDataType,
GammaDataType,
MeanInvStdDataType,
DGammaDataType,
DBetaDataType,
DXDataType,
ComputeDataType>;
ReferenceInstance ref;
auto ref_argument =
ref.MakeArgument(dy, x, gamma, mean, inv_std, host_dgamma, host_dbeta, host_dx, length);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
}
std::size_t num_bytes = dy.mDesc.GetElementSize() * sizeof(DYDataType) +
x.mDesc.GetElementSize() * sizeof(XDataType) +
mean.mDesc.GetElementSize() * sizeof(MeanInvStdDataType) +
inv_std.mDesc.GetElementSize() * sizeof(MeanInvStdDataType) +
dgamma.mDesc.GetElementSize() * sizeof(DGammaDataType) +
dbeta.mDesc.GetElementSize() * sizeof(DBetaDataType);
int num_kernel = 0;
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = inst_ptr->MakeArgumentPointer(length,
strideDy,
strideX,
strideMeanInvStd,
strideMeanInvStd,
invarient_length,
strideDGamma,
strideDBeta,
reduce_dim,
dy_dev.GetDeviceBuffer(),
x_dev.GetDeviceBuffer(),
mean_dev.GetDeviceBuffer(),
inv_std_dev.GetDeviceBuffer(),
dgamma_dev.GetDeviceBuffer(),
dbeta_dev.GetDeviceBuffer());
if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
++num_kernel;
}
else
{
if(time_kernel)
{
std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
LogRange(std::cout << "input lengths = ", length, ", ") << std::endl;
}
continue;
}
size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
float gb_per_sec = num_bytes / 1.E6 / avg_time;
if(time_kernel)
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< inst_ptr->GetTypeString() << std::endl;
if(avg_time < best_avg_time)
{
best_instance_name = inst_ptr->GetTypeString();
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
dgamma_dev.FromDevice(dgamma.mData.data());
dbeta_dev.FromDevice(dbeta.mData.data());
bool pass =
ck::utils::check_err(dgamma, host_dgamma, "Error: Incorrect dgamma", 1e-3, 1e-3);
pass &= ck::utils::check_err(dbeta, host_dbeta, "Error: Incorrect dbeta", 1e-3, 1e-3);
if(do_log)
{
LogRangeAsType<float>(std::cout << "dy : ", dy.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "host_dgamma : ", host_dgamma.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "dgamma : ", dgamma.mData, ",") << std::endl;
}
if(!pass)
{
std::cout << inst_ptr->GetTypeString() << " failed verification: ";
LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
return false;
}
else
{
if(time_kernel)
std::cout << "pass" << std::endl;
}
}
}
if(time_kernel)
{
LogRange(std::cout << "length = ", length, ",") << ", ";
LogRange(std::cout << "reduce dims ", reduce_dim, ",") << std::endl;
std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s,"
<< best_instance_name << std::endl;
}
if(num_kernel == 0)
{
std::cout << "Error: No kernel is applicable" << std::endl;
return false;
}
return true;
}
} // namespace profiler
} // namespace ck
profiler/src/CMakeLists.txt
View file @ 05fd7ff8
......@@ -19,6 +19,8 @@ set(PROFILER_SOURCES
profile_groupnorm_bwd_data.cpp
profile_groupnorm_fwd.cpp
profile_layernorm_bwd_data.cpp
profile_layernorm_bwd_gamma_beta.cpp
profile_groupnorm_bwd_gamma_beta.cpp
profile_layernorm_fwd.cpp
profile_max_pool3d_fwd.cpp
profile_avg_pool3d_bwd.cpp
......@@ -85,6 +87,7 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
......
profiler/src/profile_gemm.cpp
View file @ 05fd7ff8
......@@ -137,8 +137,14 @@ int profile_gemm(int argc, char* argv[])
return pass ? 0 : 1;
};
if(false)
;
if(data_type != GemmDataType::F32_F32_F32 && data_type != GemmDataType::F16_F16_F16 &&
data_type != GemmDataType::BF16_BF16_BF16 && data_type != GemmDataType::INT8_INT8_INT8 &&
data_type != GemmDataType::F8_F8_F8)
{
// dummy clause before the else clauses for different data types
std::cout << "Gemm: this data_type is not implemented" << std::endl;
return 1;
}
#ifdef CK_ENABLE_FP32
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{
......@@ -231,7 +237,7 @@ int profile_gemm(int argc, char* argv[])
#endif
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
std::cout << "Gemm: this data_type & layout is not implemented" << std::endl;
return 1;
}
......
profiler/src/profile_groupnorm_bwd_gamma_beta.cpp 0 → 100644
View file @ 05fd7ff8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/data_type_enum.hpp"
#include "profiler/profile_groupnorm_bwd_gamma_beta_impl.hpp"
#include "profiler_operation_registry.hpp"
using ck::index_t;
struct groupnormBwdGammaBetaArgParser
{
std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
bool parse_opt(int argc, char* argv[], const std::string& key, int i)
{
if(std::string("--") + key == argv[i])
{
int pos = i;
while(++i < argc && argv[i][0] != '-') {}
int end = i;
for(int j = pos + 1; j < end; j++)
{
long_opts[key].push_back(std::stoi(argv[j]));
}
return true;
}
return false;
}
void operator()(int argc, char* argv[])
{
for(auto& kv : long_opts)
{
for(int i = 1; i < argc; i++)
{
if(parse_opt(argc, argv, kv.first, i))
break;
}
}
}
};
void print_help_groupnorm_bwd_gamma_beta()
{
// eg: ckProfiler groupnorm_bwd_gamma_beta 1 0 2 0 1 --length 1 16 16 32 40
std::cout << "arg1: data type (0: fp16; 1: fp32)\n"
<< "arg2: verification (0: no; 1: yes)\n"
<< "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg4: print tensor value (0: no; 1: yes)\n"
<< "arg5: time kernel (0=no, 1=yes)\n"
<< "--length: tensor extents (e.g, --length 1 16 16 32 40) \n"
<< std::endl;
}
int profile_groupnorm_bwd_gamma_beta(int argc, char* argv[])
{
if(argc <= 2)
{
print_help_groupnorm_bwd_gamma_beta();
return 0;
}
groupnormBwdGammaBetaArgParser arg_parser;
// short unnamed options
const ck::DataTypeEnum data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
const bool do_verification = std::stoi(argv[3]);
const int init_method = std::stoi(argv[4]);
const bool do_log = std::stoi(argv[5]);
const bool time_kernel = std::stoi(argv[6]);
// parse the long options
arg_parser(argc, argv);
const std::vector<index_t> length = arg_parser.long_opts["length"];
using F32 = float;
if(length.size() == 5)
{
if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_groupnorm_bwd_gamma_beta_impl<F32, F32, F32, F32, F32, F32>(
do_verification, init_method, do_log, time_kernel, length);
}
else
{
throw std::runtime_error("not implemented yet");
}
}
else
{
throw std::runtime_error("length should be 5");
}
return 0;
}
REGISTER_PROFILER_OPERATION("groupnorm_bwd_gamma_beta",
"Group Normalization",
profile_groupnorm_bwd_gamma_beta);
profiler/src/profile_layernorm_bwd_gamma_beta.cpp 0 → 100644
View file @ 05fd7ff8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/data_type_enum.hpp"
#include "profiler/profile_layernorm_bwd_gamma_beta_impl.hpp"
#include "profiler_operation_registry.hpp"
using ck::index_t;
struct layernormBwdGammaBetaArgParser
{
std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
bool parse_opt(int argc, char* argv[], const std::string& key, int i)
{
if(std::string("--") + key == argv[i])
{
int pos = i;
while(++i < argc && argv[i][0] != '-') {}
int end = i;
for(int j = pos + 1; j < end; j++)
{
long_opts[key].push_back(std::stoi(argv[j]));
}
return true;
}
return false;
}
void operator()(int argc, char* argv[])
{
for(auto& kv : long_opts)
{
for(int i = 1; i < argc; i++)
{
if(parse_opt(argc, argv, kv.first, i))
break;
}
}
}
};
void print_help_layernorm_bwd_gamma_beta()
{
// eg: ckProfiler layernorm_bwd_gamma_beta 0 0 2 0 1 --length 1502 4096
std::cout << "arg1: data type (0: fp16; 1: fp32)\n"
<< "arg2: verification (0: no; 1: yes)\n"
<< "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg4: print tensor value (0: no; 1: yes)\n"
<< "arg5: time kernel (0=no, 1=yes)\n"
<< "--length: tensor extents (e.g, --length 1024 1024) \n"
<< std::endl;
}
int profile_layernorm_bwd_gamma_beta(int argc, char* argv[])
{
if(argc <= 2)
{
print_help_layernorm_bwd_gamma_beta();
return 0;
}
layernormBwdGammaBetaArgParser arg_parser;
// short unnamed options
const ck::DataTypeEnum data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
const bool do_verification = std::stoi(argv[3]);
const int init_method = std::stoi(argv[4]);
const bool do_log = std::stoi(argv[5]);
const bool time_kernel = std::stoi(argv[6]);
// parse the long options
arg_parser(argc, argv);
const std::vector<index_t> length = arg_parser.long_opts["length"];
using F16 = ck::half_t;
using F32 = float;
if(length.size() == 2)
{
constexpr int rank = 2;
if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_layernorm_bwd_gamma_beta_impl<F16, F16, F16, F32, F16, F16, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_layernorm_bwd_gamma_beta_impl<F32, F32, F32, F32, F32, F32, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else
{
throw std::runtime_error("not implemented yet");
}
}
else
{
throw std::runtime_error("not implemented yet");
}
return 0;
}
REGISTER_PROFILER_OPERATION("layernorm_bwd_gamma_beta",
"Layer Normalization",
profile_layernorm_bwd_gamma_beta);
script/clang-format-overwrite.sh
View file @ 05fd7ff8
#find . -name deps -prune -o -name build -prune -o -iname '*.h' -o -iname '*.hpp' -o -iname '*.cpp' -o -iname '*.h.in' -o -iname '*.hpp.in' -o -iname '*.cpp.in' -o -iname '*.cl' -o -iname '*.cuh' -o -iname '*.cu' -o -iname '*.inc' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-12 -i -style=file {}'
find . -name deps -prune -o -name build -prune -o -iname '*.h' -o -iname '*.hpp' -o -iname '*.cpp' -o -iname '*.h.in' -o -iname '*.hpp.in' -o -iname '*.cpp.in' -o -iname '*.cl' -o -iname '*.cuh' -o -iname '*.cu' -o -iname '*.inc' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-12 -i -style=file {}'
git status --porcelain | awk '$1 != "D" && (match($2, "\\.cpp|hpp|inc")) {print $2}' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-12 -i -style=file {}'
test/CMakeLists.txt
View file @ 05fd7ff8
......@@ -140,6 +140,7 @@ add_subdirectory(block_to_ctile_map)
add_subdirectory(softmax)
add_subdirectory(normalization_fwd)
add_subdirectory(normalization_bwd_data)
add_subdirectory(normalization_bwd_gamma_beta)
add_subdirectory(data_type)
add_subdirectory(elementwise_normalization)
add_subdirectory(batchnorm)
......
test/conv_tensor_rearrange/test_conv_tensor_rearrange_interface.cpp
View file @ 05fd7ff8
......@@ -135,6 +135,8 @@ class TestConvTensorRearrangeInterface : public ::testing::Test
return col2img.IsSupportedArgument(argument);
}
throw std::runtime_error("Conv_tensor_rearrange: problem with tensor rearrange operator. ");
return 1;
}
};
......
test/normalization_bwd_gamma_beta/CMakeLists.txt 0 → 100644
View file @ 05fd7ff8
add_custom_target(test_normalization_bwd_gamma_beta)
add_gtest_executable(test_layernorm2d_bwd_gamma_beta_fp32 test_layernorm2d_bwd_gamma_beta_fp32.cpp)
if(result EQUAL 0)
target_link_libraries(test_layernorm2d_bwd_gamma_beta_fp32 PRIVATE utility device_normalization_bwd_gamma_beta_instance)
add_dependencies(test_normalization_bwd_gamma_beta test_layernorm2d_bwd_gamma_beta_fp32)
endif()
add_gtest_executable(test_groupnorm_bwd_gamma_beta_fp32 test_groupnorm_bwd_gamma_beta_fp32.cpp)
if(result EQUAL 0)
target_link_libraries(test_groupnorm_bwd_gamma_beta_fp32 PRIVATE utility device_normalization_bwd_gamma_beta_instance)
add_dependencies(test_normalization_bwd_gamma_beta test_groupnorm_bwd_gamma_beta_fp32)
endif()
test/normalization_bwd_gamma_beta/test_groupnorm_bwd_gamma_beta_fp32.cpp 0 → 100644
View file @ 05fd7ff8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "profiler/profile_groupnorm_bwd_gamma_beta_impl.hpp"
using F16 = ck::half_t;
using F32 = float;
using ck::index_t;
template <typename Tuple>
class TestgroupnormBwdGammaBeta : public ::testing::Test
{
protected:
using DYDataType = std::tuple_element_t<0, Tuple>;
using XDataType = std::tuple_element_t<1, Tuple>;
using MeanInvStdDataType = std::tuple_element_t<2, Tuple>;
using ComputeDataType = std::tuple_element_t<3, Tuple>;
using DGammaDataType = std::tuple_element_t<4, Tuple>;
using DBetaDataType = std::tuple_element_t<5, Tuple>;
void Run()
{
// Bwd data: [N, H, W, G, C], reduce H, W, C
std::vector<std::vector<ck::index_t>> lengths = {{1, 1, 1, 1, 1},
{1, 2, 3, 4, 5},
{256, 9, 9, 9, 9},
{1, 64, 64, 32, 10},
{1, 32, 32, 32, 20},
{1, 16, 16, 32, 40}};
for(auto length : lengths)
{
bool success = ck::profiler::profile_groupnorm_bwd_gamma_beta_impl<DYDataType,
XDataType,
MeanInvStdDataType,
ComputeDataType,
DGammaDataType,
DBetaDataType>(
true, 2, false, false, length);
EXPECT_TRUE(success);
}
}
};
using KernelTypes = ::testing::Types<
// DYDataType XDataType, MeanInvStdDataType, ComputeDataType, DGammaDataType, DBetaDataType>
std::tuple<F32, F32, F32, F32, F32, F32>>;
TYPED_TEST_SUITE(TestgroupnormBwdGammaBeta, KernelTypes);
TYPED_TEST(TestgroupnormBwdGammaBeta, Test_FP32) { this->Run(); }
test/normalization_bwd_gamma_beta/test_layernorm2d_bwd_gamma_beta_fp32.cpp 0 → 100644
View file @ 05fd7ff8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "profiler/profile_layernorm_bwd_gamma_beta_impl.hpp"
using F16 = ck::half_t;
using F32 = float;
using ck::index_t;
template <typename Tuple>
class TestLayernorm2dBwdGammaBeta : public ::testing::Test
{
protected:
using DYDataType = std::tuple_element_t<0, Tuple>;
using XDataType = std::tuple_element_t<1, Tuple>;
using MeanInvStdDataType = std::tuple_element_t<2, Tuple>;
using ComputeDataType = std::tuple_element_t<3, Tuple>;
using DGammaDataType = std::tuple_element_t<4, Tuple>;
using DBetaDataType = std::tuple_element_t<5, Tuple>;
void Run()
{
// Bwd data: [N, D], reduce D
std::vector<std::vector<ck::index_t>> lengths = {
{4, 256}, {8, 511}, {9, 1032}, {4, 2048}, {1, 8192}, {4000, 2000}};
for(auto length : lengths)
{
bool success = ck::profiler::profile_layernorm_bwd_gamma_beta_impl<DYDataType,
XDataType,
MeanInvStdDataType,
ComputeDataType,
DGammaDataType,
DBetaDataType,
2>(
true, 2, false, false, length);
EXPECT_TRUE(success);
}
}
};
using KernelTypes = ::testing::Types<
// DYDataType XDataType, MeanInvStdDataType, ComputeDataType, DGammaDataType, DBetaDataType>
std::tuple<F32, F32, F32, F32, F32, F32>>;
TYPED_TEST_SUITE(TestLayernorm2dBwdGammaBeta, KernelTypes);
TYPED_TEST(TestLayernorm2dBwdGammaBeta, Test_FP32) { this->Run(); }
test/wrapper/test_copy.cpp
View file @ 05fd7ff8
......@@ -21,49 +21,59 @@ template <typename InputTensor,
typename OutputTensor,
typename BlockShape,
typename ThreadLayoutShape,
typename LocalTileSteps,
typename LocalPartitionSteps>
bool UseOptimizedCopy>
__global__ void TestCopyDevice(const InputTensor input_tensor,
OutputTensor output_tensor,
const BlockShape tile_shape,
const ThreadLayoutShape thread_layout,
const LocalTileSteps block_steps,
const LocalPartitionSteps thread_steps)
const ThreadLayoutShape thread_layout)
{
__shared__ ck::index_t p_shared[ck::wrapper::size(tile_shape)];
auto tensor_lds = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Lds>(
const auto tensor_lds = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Lds>(
p_shared, ck::wrapper::make_layout(tile_shape));
const auto block_idxs = ck::make_tuple(ck::make_tuple(0, 0), blockIdx.x);
const auto block_idx = static_cast<ck::index_t>(blockIdx.x);
// Get local tiles for global memory
const auto input_local_tile =
ck::wrapper::make_local_tile(input_tensor, tile_shape, block_idxs, block_steps);
const auto input_local_tile = ck::wrapper::make_local_tile(input_tensor, tile_shape, block_idx);
const auto output_local_tile =
ck::wrapper::make_local_tile(output_tensor, tile_shape, block_idxs, block_steps);
ck::wrapper::make_local_tile(output_tensor, tile_shape, block_idx);
// Get partition per thread
const auto input_local_partition = ck::wrapper::make_local_partition(
input_local_tile, thread_layout, threadIdx.x, thread_steps);
const auto input_local_partition =
ck::wrapper::make_local_partition(input_local_tile, thread_layout, threadIdx.x);
auto lds_local_partition =
ck::wrapper::make_local_partition(tensor_lds, thread_layout, threadIdx.x, thread_steps);
auto output_local_partition = ck::wrapper::make_local_partition(
output_local_tile, thread_layout, threadIdx.x, thread_steps);
ck::wrapper::make_local_partition(tensor_lds, thread_layout, threadIdx.x);
auto output_local_partition =
ck::wrapper::make_local_partition(output_local_tile, thread_layout, threadIdx.x);
// Allocate VGPR
constexpr ck::index_t scalar_per_vector = 1;
constexpr ck::index_t vgpr_size = ck::wrapper::size(lds_local_partition);
auto tensor_vgpr = ck::wrapper::make_register_tensor<ck::wrapper::MemoryTypeEnum::Vgpr,
vgpr_size,
scalar_per_vector,
ck::index_t>();
auto tensor_vgpr =
ck::wrapper::make_register_tensor<ck::wrapper::MemoryTypeEnum::Vgpr, ck::index_t>(
layout(lds_local_partition));
// Perform copy
ck::wrapper::copy(input_local_partition, lds_local_partition);
ck::wrapper::copy(lds_local_partition, tensor_vgpr);
ck::wrapper::copy(tensor_vgpr, output_local_partition);
if constexpr(UseOptimizedCopy)
{
using DimAccessOrder = ck::Tuple<ck::Number<1>, ck::Number<0>>;
constexpr ck::index_t vector_dim = 0;
constexpr ck::index_t scalar_per_vector = 2;
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(input_local_partition,
lds_local_partition);
// TODO: Enable optimized copy for static buffers
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(lds_local_partition,
tensor_vgpr);
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(tensor_vgpr,
output_local_partition);
}
else
{
ck::wrapper::copy(input_local_partition, lds_local_partition);
ck::wrapper::copy(lds_local_partition, tensor_vgpr);
ck::wrapper::copy(tensor_vgpr, output_local_partition);
}
}
template <bool UseOptimizedCopy>
void PerformCopyGlobalToGlobalViaLDS()
{
const auto shape =
......@@ -89,15 +99,8 @@ void PerformCopyGlobalToGlobalViaLDS()
auto output_tensor_global = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Global>(
static_cast<ck::index_t*>(out_buf.GetDeviceBuffer()), layout);
const auto thread_layout =
ck::make_tuple(ck::make_tuple(ck::Number<1>{}, ck::Number<1>{}), ck::Number<32>{});
const auto tile_shape =
ck::make_tuple(ck::make_tuple(ck::Number<2>{}, ck::Number<2>{}), ck::Number<64>{});
const auto thread_steps =
ck::make_tuple(ck::make_tuple(ck::Number<1>{}, ck::Number<1>{}), ck::Number<2>{});
const auto block_steps =
ck::make_tuple(ck::make_tuple(ck::Number<1>{}, ck::Number<1>{}), ck::Number<64>{});
const auto thread_layout = ck::make_tuple(ck::Number<1>{}, ck::Number<32>{});
const auto tile_shape = ck::make_tuple(ck::Number<4>{}, ck::Number<64>{});
const ck::index_t grid_size = ck::math::integer_divide_ceil(
ck::wrapper::size(input_tensor_global), ck::wrapper::size(tile_shape));
......@@ -106,8 +109,7 @@ void PerformCopyGlobalToGlobalViaLDS()
decltype(output_tensor_global),
decltype(tile_shape),
decltype(thread_layout),
decltype(block_steps),
decltype(thread_steps)>;
UseOptimizedCopy>;
launch_and_time_kernel(StreamConfig{},
kernel,
dim3(grid_size),
......@@ -116,9 +118,7 @@ void PerformCopyGlobalToGlobalViaLDS()
input_tensor_global,
output_tensor_global,
tile_shape,
thread_layout,
block_steps,
thread_steps);
thread_layout);
// Verify results
std::vector<ck::index_t> output_data(ck::wrapper::size(shape));
......@@ -126,4 +126,5 @@ void PerformCopyGlobalToGlobalViaLDS()
EXPECT_TRUE(ck::utils::check_err(output_data, input_data));
}
TEST(TestCopy, CopyGlobalToGlobalViaLDS) { PerformCopyGlobalToGlobalViaLDS(); }
TEST(TestCopyGlobalToGlobalViaLDS, GenericCopy) { PerformCopyGlobalToGlobalViaLDS<false>(); }
TEST(TestCopyGlobalToGlobalViaLDS, OptimizedCopy) { PerformCopyGlobalToGlobalViaLDS<true>(); }
test/wrapper/test_partition.cpp
View file @ 05fd7ff8
......@@ -29,42 +29,29 @@ TEST(TestPartition, LocalPartition)
const auto tensor =
ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Generic>(data.data(), layout);
const auto thread_steps =
ck::make_tuple(ck::make_tuple(ck::Number<2>{}, ck::Number<1>{}), ck::Number<1>{});
const auto thread_layout =
ck::make_tuple(ck::make_tuple(ck::Number<8>{}, ck::Number<1>{}), ck::Number<1>{});
for(ck::index_t thread_id = 0; thread_id < ck::wrapper::size(thread_layout); thread_id++)
{
const auto raked_partition =
ck::wrapper::make_local_partition(tensor, thread_layout, thread_id);
const auto expected_partition_size =
ck::wrapper::size(tensor) / ck::wrapper::size(thread_layout);
EXPECT_EQ(ck::wrapper::size(raked_partition), expected_partition_size);
EXPECT_EQ(raked_partition(0), thread_id);
}
const auto thread_steps = ck::make_tuple(ck::Number<8>{}, ck::Number<1>{});
const auto thread_layout = ck::make_tuple(ck::Number<8>{}, ck::Number<1>{});
for(ck::index_t thread_id = 0; thread_id < ck::wrapper::size(thread_layout); thread_id++)
{
const auto packed_partition =
ck::wrapper::make_local_partition(tensor, thread_layout, thread_id, thread_steps);
ck::wrapper::make_local_partition(tensor, thread_layout, thread_id);
const auto expected_partition_size =
ck::wrapper::size(tensor) / ck::wrapper::size(thread_layout);
const auto expected_partition_first_val = thread_id * ck::wrapper::size<0, 0>(thread_steps);
const auto expected_partition_first_val = thread_id * ck::wrapper::size<0>(thread_steps);
const auto expected_partition_second_val = expected_partition_first_val + 1;
EXPECT_EQ(ck::wrapper::size(packed_partition), expected_partition_size);
EXPECT_EQ(packed_partition(0), expected_partition_first_val);
EXPECT_EQ(packed_partition(1), expected_partition_second_val);
}
}
TEST(TestPartition, LocalTile)
{
const auto shape =
ck::make_tuple(ck::make_tuple(ck::Number<16>{}, ck::Number<4>{}), ck::Number<4>{});
const auto strides =
ck::make_tuple(ck::make_tuple(ck::Number<1>{}, ck::Number<16>{}), ck::Number<64>{});
const auto layout = ck::wrapper::make_layout(shape, strides);
const auto shape = ck::make_tuple(ck::Number<16>{}, ck::Number<4>{}, ck::Number<4>{});
const auto strides = ck::make_tuple(ck::Number<1>{}, ck::Number<16>{}, ck::Number<64>{});
const auto layout = ck::wrapper::make_layout(shape, strides);
std::vector<ck::index_t> data(ck::wrapper::size(layout));
std::iota(data.begin(), data.end(), 0);
......@@ -72,48 +59,34 @@ TEST(TestPartition, LocalTile)
const auto tensor =
ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Generic>(data.data(), layout);
const auto block_steps =
ck::make_tuple(ck::make_tuple(ck::Number<4>{}, ck::Number<2>{}), ck::Number<2>{});
const auto block_shape =
ck::make_tuple(ck::make_tuple(ck::Number<4>{}, ck::Number<2>{}), ck::Number<2>{});
const auto block_layout =
ck::make_tuple(ck::make_tuple(ck::Number<4>{}, ck::Number<2>{}), ck::Number<2>{});
std::vector<ck::Tuple<ck::Tuple<ck::index_t, ck::index_t>, ck::index_t>> block_idxs;
for(ck::index_t x = 0; x < ck::wrapper::size<0, 0>(block_layout); x++)
{
for(ck::index_t y = 0; y < ck::wrapper::size<0, 1>(block_layout); y++)
{
for(ck::index_t z = 0; z < ck::wrapper::size<1>(block_layout); z++)
{
block_idxs.emplace_back(ck::make_tuple(x, y), z);
}
}
}
for(const auto& block_idx : block_idxs)
{
const auto raked_tile = ck::wrapper::make_local_tile(tensor, block_shape, block_idx);
const auto expected_tile_size = ck::wrapper::size(block_shape);
EXPECT_EQ(ck::wrapper::size(raked_tile), expected_tile_size);
EXPECT_EQ(raked_tile(0), layout(block_idx));
}
const auto block_shape = ck::make_tuple(ck::Number<2>{}, ck::Number<4>{}, ck::Number<2>{});
const auto num_blocks =
ck::make_tuple(ck::wrapper::size<0>(shape) / ck::wrapper::size<0>(block_shape),
ck::wrapper::size<1>(shape) / ck::wrapper::size<1>(block_shape),
ck::wrapper::size<2>(shape) / ck::wrapper::size<2>(block_shape));
std::vector<ck::index_t> block_idxs(ck::wrapper::size(num_blocks));
std::iota(block_idxs.begin(), block_idxs.end(), 0);
for(const auto& block_idx : block_idxs)
for(auto block_idx : block_idxs)
{
const auto packed_tile =
ck::wrapper::make_local_tile(tensor, block_shape, block_idx, block_steps);
const auto packed_tile = ck::wrapper::make_local_tile(tensor, block_shape, block_idx);
const auto expected_tile_size = ck::wrapper::size(block_shape);
const auto expected_tile_first_val =
ck::wrapper::size<0, 0>(block_idx) * ck::wrapper::size<0, 0>(block_shape) *
ck::wrapper::size<0, 0>(strides) +
ck::wrapper::size<0, 1>(block_idx) * ck::wrapper::size<0, 1>(block_shape) *
ck::wrapper::size<0, 1>(strides) +
ck::wrapper::size<1>(block_idx) * ck::wrapper::size<1>(block_shape) *
ck::wrapper::size<1>(strides);
auto expected_tile_first_val = (block_idx % ck::wrapper::size<2>(num_blocks)) *
ck::wrapper::size<2>(block_shape) *
ck::wrapper::size<2>(strides);
block_idx /= ck::wrapper::size<2>(num_blocks);
expected_tile_first_val += (block_idx % ck::wrapper::size<1>(num_blocks)) *
ck::wrapper::size<1>(block_shape) *
ck::wrapper::size<1>(strides);
block_idx /= ck::wrapper::size<1>(num_blocks);
expected_tile_first_val += (block_idx % ck::wrapper::size<0>(num_blocks)) *
ck::wrapper::size<0>(block_shape) *
ck::wrapper::size<0>(strides);
const auto expected_tile_second_val = expected_tile_first_val + 1;
EXPECT_EQ(ck::wrapper::size(packed_tile), expected_tile_size);
EXPECT_EQ(packed_tile(0), expected_tile_first_val);
EXPECT_EQ(packed_tile(1), expected_tile_second_val);
}
}
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