Skip to content

GitLab

Commit f0759faf authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'develop' into amd-develop

parents 20ddaeba 764164b4
Hide whitespace changes
Inline Side-by-side
Showing with 951 additions and 25 deletions
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_mnkpadding_instance.cpp 0 → 100644
View file @ f0759faf
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_instances<Intrawave,
GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_default_instance.cpp 0 → 100644
View file @ f0759faf
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_instances<Interwave, GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_kpadding_instance.cpp 0 → 100644
View file @ f0759faf
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_instances<Interwave, GemmKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_mnkpadding_instance.cpp 0 → 100644
View file @ f0759faf
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_instances<Interwave,
GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
View file @ f0759faf
...@@ -8,9 +8,6 @@ namespace tensor_operation { ...@@ -8,9 +8,6 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
using F16 = ck::half_t;
using F32 = float;
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances( void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
......
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_mnpadding_instance.cpp
View file @ f0759faf
...@@ -8,9 +8,6 @@ namespace tensor_operation { ...@@ -8,9 +8,6 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
using F16 = ck::half_t;
using F32 = float;
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_mnpadding_instances( void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
......
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn.hpp
View file @ f0759faf
...@@ -46,7 +46,9 @@ using device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_instances = std::tuple< ...@@ -46,7 +46,9 @@ using device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_instances = std::tuple<
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 192, 256, 64, 16, 8, 32, 32, 3, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 192, 256, 64, 16, 8, 32, 32, 3, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 16, 8, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 16, 8, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 16, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
// We prefer following instance, however, existing compiler bug cause it failed to generate sanity code.
// DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
// clang-format on // clang-format on
>; >;
......
library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk_multi_abd/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_km_kn_mn_common.hpp
View file @ f0759faf
...@@ -47,14 +47,14 @@ using D0Layout = Row; ...@@ -47,14 +47,14 @@ using D0Layout = Row;
// using DsLayout = ck::Tuple<Row>; // using DsLayout = ck::Tuple<Row>;
using ELayout = Row; using ELayout = Row;
using Scales = ck::tensor_operation::element_wise::Scales; using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu; using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using Add = ck::tensor_operation::element_wise::Add; using Add = ck::tensor_operation::element_wise::Add;
using FastGelu = ck::tensor_operation::element_wise::FastGelu; using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AElementOp = PassThrough; using AElementOp = PassThrough;
using BElementOp = Scales; using BElementOp = Multiply;
// using CDEElementOp = AddFastGelu; // using CDEElementOp = AddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
......
library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk_multi_abd/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_common.hpp
View file @ f0759faf
...@@ -47,14 +47,14 @@ using D0Layout = Row; ...@@ -47,14 +47,14 @@ using D0Layout = Row;
// using DsLayout = ck::Tuple<Row>; // using DsLayout = ck::Tuple<Row>;
using ELayout = Row; using ELayout = Row;
using Scales = ck::tensor_operation::element_wise::Scales; using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu; using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using Add = ck::tensor_operation::element_wise::Add; using Add = ck::tensor_operation::element_wise::Add;
using FastGelu = ck::tensor_operation::element_wise::FastGelu; using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AElementOp = PassThrough; using AElementOp = PassThrough;
using BElementOp = Scales; using BElementOp = Multiply;
// using CDEElementOp = AddFastGelu; // using CDEElementOp = AddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
......
library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk_multi_abd/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_common.hpp
View file @ f0759faf
...@@ -47,14 +47,14 @@ using D0Layout = Row; ...@@ -47,14 +47,14 @@ using D0Layout = Row;
// using DsLayout = ck::Tuple<Row>; // using DsLayout = ck::Tuple<Row>;
using ELayout = Row; using ELayout = Row;
using Scales = ck::tensor_operation::element_wise::Scales; using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu; using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using Add = ck::tensor_operation::element_wise::Add; using Add = ck::tensor_operation::element_wise::Add;
using FastGelu = ck::tensor_operation::element_wise::FastGelu; using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AElementOp = PassThrough; using AElementOp = PassThrough;
using BElementOp = Scales; using BElementOp = Multiply;
// using CDEElementOp = AddFastGelu; // using CDEElementOp = AddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
......
library/src/tensor_operation_instance/gpu/grouped_gemm_tile_loop/CMakeLists.txt 0 → 100644
View file @ f0759faf
# ONLY XDL_KERNELS
set(GROUPED_GEMM_TILE_LOOP_INSTANCES)
list(APPEND GROUPED_GEMM_TILE_LOOP_INSTANCES
device_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_kn_mn_instance.cpp
device_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_nk_mn_instance.cpp
device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_instance.cpp
)
add_instance_library(device_grouped_gemm_tile_loop_instance ${GROUPED_GEMM_TILE_LOOP_INSTANCES})
library/src/tensor_operation_instance/gpu/grouped_gemm_tile_loop/device_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_kn_mn_instance.cpp 0 → 100644
View file @ f0759faf
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_xdl_cshuffle_tile_loop.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
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 DsDataType = ck::Tuple<>;
using DsLayout = ck::Tuple<>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using device_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
// clang-format off
//###########################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| C| 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|
//###########################################| Layout| Layout| Layout| Layout| 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|
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 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, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 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>
// clang-format on
>;
void add_device_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
Row,
DsLayout,
Row,
F16,
F16,
DsDataType,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_gemm_tile_loop/device_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_nk_mn_instance.cpp 0 → 100644
View file @ f0759faf
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_xdl_cshuffle_tile_loop.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
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 DsDataType = ck::Tuple<>;
using DsLayout = ck::Tuple<>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using device_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
// clang-format off
//###########################################| A| B| Ds| E| 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|
//###########################################| Layout| Layout| Layout| Layout| 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|
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 64, 8, 8, 32, 32, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 64, 8, 8, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 64, 8, 8, 32, 32, 1, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 64, 8, 8, 32, 32, 4, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 64, 8, 8, 32, 32, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 64, 8, 8, 32, 32, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 64, 8, 8, 32, 32, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 64, 8, 8, 32, 32, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 256, 64, 8, 8, 32, 32, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 64, 8, 8, 32, 32, 2, 2, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 64, 8, 8, 32, 32, 2, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Col, DsLayout, Row, F16, F16, F32, F32, DsDataType, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 64, 8, 8, 32, 32, 1, 2, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 8, 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_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
Col,
DsLayout,
Row,
F16,
F16,
DsDataType,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_nk_mn_irregular_tile_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_gemm_tile_loop/device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_instance.cpp 0 → 100644
View file @ f0759faf
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_xdl_cshuffle_tile_loop.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
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 Multiply = ck::tensor_operation::element_wise::Multiply;
using MultiplyAddFastGelu = ck::tensor_operation::element_wise::MultiplyAddFastGelu;
using MultiplyFastGelu = ck::tensor_operation::element_wise::MultiplyFastGelu;
using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <typename DsLayout, typename DsDataType, typename CDEElementwiseOp>
using device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_irregular_tile_instances = std::tuple<
// clang-format off
//###########################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| C| 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|
//###########################################| Layout| Layout| Layout| Layout| 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|
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if 1
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 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, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 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>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 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>
#endif
#if 0
//comp
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8>,
//latency
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 64, 16, 16, 256, 8, 4, 16, 16, 1, 1, S<32, 2, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<64, 1, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, 1, 1, S<1, 16, 1, 4>, 4>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 128, 16, 32, 256, 8, 4, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<64, 2, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, 1, 1, S<1, 16, 1, 8>, 4>,
//mem
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 64, 16, 16, 256, 8, 4, 16, 16, 1, 1, S<32, 2, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<64, 1, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, 1, 1, S<1, 16, 1, 4>, 4>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 128, 16, 32, 256, 8, 4, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<64, 2, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 128, 16, 64, 128, 8, 4, 16, 16, 1, 2, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 128, 32, 64, 128, 8, 4, 32, 32, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 128, 16, 128, 64, 8, 4, 16, 16, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 128, 32, 128, 64, 8, 4, 32, 32, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 256, 16, 256, 64, 8, 4, 16, 16, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmMNKPadding, 1, 256, 32, 256, 64, 8, 4, 32, 32, 1, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 4, 0, 1, 1, S<1, 16, 1, 16>, 8>
#endif
// clang-format on
>;
void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
Row,
ck::Tuple<Row>,
Row,
BF16,
I8,
ck::Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
Multiply>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_irregular_tile_instances<
ck::Tuple<Row>,
ck::Tuple<BF16>,
Multiply>{});
}
void add_device_grouped_gemm_xdl_tile_loop_multiply_bias_bf16_i8_bf16_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
Row,
ck::Tuple<Row, Row>,
Row,
BF16,
I8,
ck::Tuple<BF16, BF16>,
BF16,
PassThrough,
PassThrough,
MultiplyAdd>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_irregular_tile_instances<
ck::Tuple<Row, Row>,
ck::Tuple<BF16, BF16>,
MultiplyAdd>{});
}
void add_device_grouped_gemm_xdl_tile_loop_multiply_bias_fastgelu_bf16_i8_bf16_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
Row,
ck::Tuple<Row, Row>,
Row,
BF16,
I8,
ck::Tuple<BF16, BF16>,
BF16,
PassThrough,
PassThrough,
MultiplyAddFastGelu>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_irregular_tile_instances<
ck::Tuple<Row, Row>,
ck::Tuple<BF16, BF16>,
MultiplyAddFastGelu>{});
}
void add_device_grouped_gemm_xdl_tile_loop_multiply_fastgelu_bf16_i8_bf16_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
Row,
ck::Tuple<Row>,
Row,
BF16,
I8,
ck::Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
MultiplyFastGelu>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_irregular_tile_instances<
ck::Tuple<Row>,
ck::Tuple<BF16>,
MultiplyFastGelu>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/include/profiler/profile_gemm_universal_impl.hpp
View file @ f0759faf
...@@ -43,7 +43,8 @@ bool profile_gemm_universal_impl(int do_verification, ...@@ -43,7 +43,8 @@ bool profile_gemm_universal_impl(int do_verification,
int StrideC, int StrideC,
int KBatch, int KBatch,
int n_warmup, int n_warmup,
int n_iter) int n_iter,
uint64_t rotating = 0)
{ {
bool pass = true; bool pass = true;
...@@ -66,9 +67,16 @@ bool profile_gemm_universal_impl(int do_verification, ...@@ -66,9 +67,16 @@ bool profile_gemm_universal_impl(int do_verification,
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
int total_gemm_needed = a_m_k.GetElementSpaceSizeInBytes() + b_k_n.GetElementSpaceSizeInBytes();
int rotating_count = std::max(
1,
std::min(n_iter,
static_cast<int>(std::ceil(static_cast<double>(rotating) / total_gemm_needed))));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl; std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl; std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl; std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
std::cout << "rotating count: " << rotating_count << std::endl;
switch(init_method) switch(init_method)
{ {
...@@ -200,8 +208,14 @@ bool profile_gemm_universal_impl(int do_verification, ...@@ -200,8 +208,14 @@ bool profile_gemm_universal_impl(int do_verification,
std::string op_name = op_ptr->GetTypeString(); std::string op_name = op_ptr->GetTypeString();
float ave_time = invoker_ptr->Run( float ave_time = invoker_ptr->Run(argument_ptr.get(),
argument_ptr.get(), StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter}); StreamConfig{nullptr,
time_kernel,
0,
n_warmup,
n_iter,
rotating_count > 1,
rotating_count});
std::size_t flop = std::size_t(2) * M * N * K; std::size_t flop = std::size_t(2) * M * N * K;
......
profiler/include/profiler/profile_grouped_gemm_fixed_nk_impl.hpp
View file @ f0759faf
...@@ -73,9 +73,11 @@ bool profile_grouped_gemm_fixed_nk_impl(int do_verification, ...@@ -73,9 +73,11 @@ bool profile_grouped_gemm_fixed_nk_impl(int do_verification,
std::vector<Tensor<BDataType>> b_k_n; std::vector<Tensor<BDataType>> b_k_n;
std::vector<Tensor<CDataType>> c_m_n_host_results; std::vector<Tensor<CDataType>> c_m_n_host_results;
std::vector<Tensor<CDataType>> c_m_n_device_results; std::vector<Tensor<CDataType>> c_m_n_device_results;
int sum_of_m = 0;
for(std::size_t i = 0; i < group_count; i++) for(std::size_t i = 0; i < group_count; i++)
{ {
sum_of_m += Ms[i];
a_m_k.push_back( a_m_k.push_back(
Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{}))); Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{})));
b_k_n.push_back( b_k_n.push_back(
...@@ -146,7 +148,7 @@ bool profile_grouped_gemm_fixed_nk_impl(int do_verification, ...@@ -146,7 +148,7 @@ bool profile_grouped_gemm_fixed_nk_impl(int do_verification,
a_device_buf[i]->ToDevice(a_m_k[i].mData.data()); a_device_buf[i]->ToDevice(a_m_k[i].mData.data());
b_device_buf[i]->ToDevice(b_k_n[i].mData.data()); b_device_buf[i]->ToDevice(b_k_n[i].mData.data());
gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}}); gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
p_a.push_back(a_device_buf[i]->GetDeviceBuffer()); p_a.push_back(a_device_buf[i]->GetDeviceBuffer());
p_b.push_back(b_device_buf[i]->GetDeviceBuffer()); p_b.push_back(b_device_buf[i]->GetDeviceBuffer());
......
profiler/include/profiler/profile_grouped_gemm_tile_loop_impl.hpp 0 → 100644
View file @ f0759faf
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/host_utility/hip_check_error.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_tile_loop.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_tile_loop.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/literals.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename ALayout,
typename BLayout,
typename CLayout>
bool profile_grouped_gemm_tile_loop_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
const std::vector<int>& Ms,
const std::vector<int>& Ns,
const std::vector<int>& Ks,
const std::vector<int>& StrideAs,
const std::vector<int>& StrideBs,
const std::vector<int>& StrideCs,
int n_warmup = 10,
int n_iter = 50)
{
bool pass = true;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
std::size_t group_count = Ms.size();
if(!(group_count == Ns.size() && group_count == Ks.size() && group_count == StrideAs.size() &&
group_count == StrideBs.size() && group_count == StrideCs.size()))
{
throw std::runtime_error("wrong! inconsistent M/N/Ks, StrideA/B/Cs size\n");
}
std::vector<Tensor<ADataType>> a_m_k;
std::vector<Tensor<BDataType>> b_k_n;
std::vector<Tensor<CDataType>> c_m_n_host_results;
std::vector<Tensor<CDataType>> c_m_n_device_results;
for(std::size_t i = 0; i < group_count; i++)
{
a_m_k.push_back(
Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{})));
b_k_n.push_back(
Tensor<BDataType>(f_host_tensor_descriptor(Ks[i], Ns[i], StrideBs[i], BLayout{})));
c_m_n_device_results.push_back(
Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
c_m_n_host_results.push_back(
Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
#if DEBUG_LOG
std::cout << "group: " << i << " a_m_k[" << i << "]:" << a_m_k[i].mDesc << ", b_k_n[" << i
<< "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
<< "]:" << c_m_n_device_results[i].mDesc << std::endl;
#endif // DEBUG_LOG
switch(init_method)
{
case 0: break;
case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5, 5}(a_m_k[i]);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5, 5}(b_k_n[i]);
break;
case 2:
ck::utils::FillUniformDistribution<ADataType>{.0, 1.}(a_m_k[i]);
ck::utils::FillUniformDistribution<BDataType>{-0.5, 0.5}(b_k_n[i]);
break;
default:
ck::utils::FillConstant<ADataType>{1}(a_m_k[i]);
ck::utils::FillConstant<BDataType>{1}(b_k_n[i]);
}
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_device_buf, b_device_buf, c_device_buf;
a_device_buf.reserve(group_count);
b_device_buf.reserve(group_count);
c_device_buf.reserve(group_count);
std::vector<const void*> p_a, p_b;
std::vector<void*> p_c;
p_a.reserve(group_count);
p_b.reserve(group_count);
p_c.reserve(group_count);
using KernelArguments = ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<>;
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
std::vector<KernelArguments> gemm_kargs;
gemm_descs.reserve(group_count);
gemm_kargs.reserve(group_count);
for(std::size_t i = 0; i < group_count; i++)
{
a_device_buf.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpaceSize()));
b_device_buf.emplace_back(
std::make_unique<DeviceMem>(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSpaceSize()));
c_device_buf.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_m_n_device_results[i].mDesc.GetElementSpaceSize()));
a_device_buf[i]->ToDevice(a_m_k[i].mData.data());
b_device_buf[i]->ToDevice(b_k_n[i].mData.data());
c_device_buf[i]->SetZero();
p_a.push_back(a_device_buf[i]->GetDeviceBuffer());
p_b.push_back(b_device_buf[i]->GetDeviceBuffer());
p_c.push_back(c_device_buf[i]->GetDeviceBuffer());
gemm_descs.push_back({0, Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
gemm_kargs.push_back({a_device_buf[i]->GetDeviceBuffer(),
b_device_buf[i]->GetDeviceBuffer(),
{},
c_device_buf[i]->GetDeviceBuffer(),
Ms[i],
Ns[i],
Ks[i],
StrideAs[i],
StrideBs[i],
{},
StrideCs[i]});
}
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmTileLoop<ALayout,
BLayout,
ck::Tuple<>,
CLayout,
ADataType,
BDataType,
ck::Tuple<>,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
if(op_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;
auto p_ds = std::vector<std::array<const void*, 0>>{};
if(do_verification)
{
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_m_k[i],
b_k_n[i],
c_m_n_host_results[i],
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
}
}
// profile device GEMM instances
for(auto& gemm_ptr : op_ptrs)
{
auto argument_ptr =
gemm_ptr->MakeArgumentPointer(p_a,
p_b,
p_ds,
p_c,
gemm_descs,
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{});
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
std::string gemm_name = gemm_ptr->GetTypeString();
DeviceMem gemm_arg_dev_mem(gemm_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
hip_check_error(hipMemcpy(gemm_arg_dev_mem.GetDeviceBuffer(),
gemm_kargs.data(),
gemm_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
hipMemcpyHostToDevice));
gemm_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_arg_dev_mem.GetDeviceBuffer());
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false, 0, n_warmup, n_iter});
if(do_verification)
{
bool instance_pass = true;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
instance_pass = instance_pass && ck::utils::check_err(c_m_n_device_results[i],
c_m_n_host_results[i]);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k[i].mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n[i].mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_m_n_device_results[i].mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_results[i].mData, ",")
<< std::endl;
}
}
std::cout << "Instance: " << gemm_name << " verification "
<< (instance_pass ? "SUCCEED" : "FAILED") << std::endl;
pass = pass && instance_pass;
}
if(time_kernel)
{
float ave_time = invoker_ptr->Run(
argument_ptr.get(), StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter});
std::size_t flop = 0, num_btype = 0;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
num_btype += sizeof(ADataType) * Ms[i] * Ks[i] +
sizeof(BDataType) * Ks[i] * Ns[i] +
sizeof(CDataType) * Ms[i] * Ns[i];
}
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, " << 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;
}
}
}
else
{
std::cout << "Instance: " << gemm_name << ", does not support this GEMM problem"
<< std::endl;
}
}
if(time_kernel)
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
}
return pass;
}
} // namespace profiler
} // namespace ck
profiler/src/CMakeLists.txt
View file @ f0759faf
...@@ -42,6 +42,7 @@ if(GPU_TARGETS MATCHES "gfx9") ...@@ -42,6 +42,7 @@ if(GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp) list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_two_stage.cpp) list(APPEND PROFILER_SOURCES profile_grouped_gemm_two_stage.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp) list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp)
endif() endif()
list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
...@@ -111,6 +112,7 @@ if(GPU_TARGETS MATCHES "gfx9") ...@@ -111,6 +112,7 @@ if(GPU_TARGETS MATCHES "gfx9")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance)
endif() endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
......
profiler/src/profile_gemm_universal.cpp
View file @ f0759faf
...@@ -33,7 +33,7 @@ enum struct GemmDataType ...@@ -33,7 +33,7 @@ enum struct GemmDataType
int profile_gemm_universal(int argc, char* argv[]) int profile_gemm_universal(int argc, char* argv[])
{ {
if(argc != 15 && argc != 17) if(argc != 15 && argc != 18)
{ {
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"); printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: f16, " printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: f16, "
...@@ -51,6 +51,7 @@ int profile_gemm_universal(int argc, char* argv[]) ...@@ -51,6 +51,7 @@ int profile_gemm_universal(int argc, char* argv[])
printf("optional:\n"); printf("optional:\n");
printf("arg15: number of warm-up cycles (default 1)\n"); printf("arg15: number of warm-up cycles (default 1)\n");
printf("arg16: number of iterations (default 10)\n"); printf("arg16: number of iterations (default 10)\n");
printf("arg17: memory for rotating buffer (default 0, size in MB)\n");
exit(1); exit(1);
} }
...@@ -70,17 +71,20 @@ int profile_gemm_universal(int argc, char* argv[]) ...@@ -70,17 +71,20 @@ int profile_gemm_universal(int argc, char* argv[])
const int StrideC = std::stoi(argv[13]); const int StrideC = std::stoi(argv[13]);
const int KBatch = std::stoi(argv[14]); const int KBatch = std::stoi(argv[14]);
int n_warmup = 1; int n_warmup = 1;
int n_iter = 10; int n_iter = 10;
if(argc == 17) uint64_t rotating = 0;
if(argc == 18)
{ {
n_warmup = std::stoi(argv[15]); n_warmup = std::stoi(argv[15]);
n_iter = std::stoi(argv[16]); n_iter = std::stoi(argv[16]);
rotating = std::stoull(argv[17]) * 1024 * 1024;
} }
using F32 = float; using F32 = float;
using F16 = ck::half_t; using F16 = ck::half_t;
using F8 = ck::f8_t; using BF16 = ck::bhalf_t;
using F8 = ck::f8_t;
using Row = ck::tensor_layout::gemm::RowMajor; using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor; using Col = ck::tensor_layout::gemm::ColumnMajor;
...@@ -124,7 +128,8 @@ int profile_gemm_universal(int argc, char* argv[]) ...@@ -124,7 +128,8 @@ int profile_gemm_universal(int argc, char* argv[])
(StrideC < 0) ? DefaultStrideC : StrideC, (StrideC < 0) ? DefaultStrideC : StrideC,
KBatch, KBatch,
n_warmup, n_warmup,
n_iter); n_iter,
rotating);
return pass ? 0 : 1; return pass ? 0 : 1;
}; };
...@@ -153,6 +158,14 @@ int profile_gemm_universal(int argc, char* argv[]) ...@@ -153,6 +158,14 @@ int profile_gemm_universal(int argc, char* argv[])
{ {
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{}); return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
} }
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(BF16{}, BF16{}, F32{}, BF16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(BF16{}, BF16{}, F32{}, BF16{}, Row{}, Col{}, Row{});
}
else else
{ {
std::cout << "this data_type & layout is not implemented" << std::endl; std::cout << "this data_type & layout is not implemented" << std::endl;
......
profiler/src/profile_grouped_gemm_tile_loop.cpp 0 → 100644
View file @ f0759faf
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <sstream>
#include <string>
#include <vector>
#include "profiler/profile_grouped_gemm_tile_loop_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 0
};
enum struct GemmDataType
{
F16_F16_F16, // 0
};
#define OP_NAME "grouped_gemm_tile_loop"
#define OP_DESC "Grouped GEMM Multiple D Tile Loop"
namespace {
std::vector<int> argToIntArray(char* input)
{
std::vector<int> out;
std::istringstream in(input);
std::string item;
while(std::getline(in, item, ','))
{
out.push_back(std::stoi(item));
}
return out;
}
int profile_grouped_gemm_tile_loop(int argc, char* argv[])
{
if(argc < 14)
{
std::cout
<< "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp16)\n"
<< "arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n]);\n"
<< " 1: A[m, k] * B[n, k] = C[m, n];\n"
<< "arg4: verification (0: no; 1: yes)\n"
<< "arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: time kernel (0=n0, 1=yes)\n"
<< "arg8 to 13: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs (e.g., 256,256 128,128 64,64 "
"64,64 64,64 128,128)\n"
<< "optional:\n"
<< "arg14: number of warm-up cycles (default 1)\n"
<< "arg15: number of iterations (default 10)\n"
<< std::endl;
exit(1);
}
const auto data_type = static_cast<GemmDataType>(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 auto Ms = argToIntArray(argv[8]);
const auto Ns = argToIntArray(argv[9]);
const auto Ks = argToIntArray(argv[10]);
auto StrideAs = argToIntArray(argv[11]);
auto StrideBs = argToIntArray(argv[12]);
auto StrideCs = argToIntArray(argv[13]);
const int DefaultStrideA = Ks[0];
const int DefaultStrideB = Ns[0];
const int DefaultStrideC = Ns[0];
for(size_t i = 0; i < Ms.size(); ++i)
{
StrideAs[i] = StrideAs[i] == -1 ? DefaultStrideA : StrideAs[i];
StrideBs[i] = StrideBs[i] == -1 ? DefaultStrideB : StrideBs[i];
StrideCs[i] = StrideCs[i] == -1 ? DefaultStrideC : StrideCs[i];
}
int n_warmup = 10;
int n_iter = 50;
if(argc == 16)
{
n_warmup = std::stoi(argv[14]);
n_iter = std::stoi(argv[15]);
}
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_tile_loop_impl<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,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_grouped_gemm_tile_loop_impl<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,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs,
n_warmup,
n_iter);
}
else
{
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
return 0;
}
} // anonymous namespace
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_gemm_tile_loop);
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