Skip to content

GitLab

Commit 67ab3896 authored by Aleksander Dudek's avatar Aleksander Dudek
Browse files

Merge branch 'develop' into gemm_getname

parents 8adaf418 d5c8a334
Hide whitespace changes
Inline Side-by-side
Showing with 1147 additions and 6 deletions
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_comp_mnpadding_instance.cpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
Row,
Row,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_comp_instances<GemmMNPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_v1_default_instance.cpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
Row,
Row,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_instances<Intrawave,
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_v1_kpadding_instance.cpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
Row,
Row,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_instances<Intrawave,
GemmKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_v1_mnkpadding_instance.cpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
Row,
Row,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_instances<Intrawave,
GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_v2_default_instance.cpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
Row,
Row,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_instances<Interwave,
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_v2_kpadding_instance.cpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
Row,
Row,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_instances<Interwave,
GemmKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
Row,
Row,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_kn_mn_mem_instances<Interwave,
GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn.hpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#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_gemm_xdl_cshuffle_streamk_v3.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using BF16 = bhalf_t;
using F32 = float;
using Row = tensor_layout::gemm::RowMajor;
using Col = tensor_layout::gemm::ColumnMajor;
template <index_t... Is>
using S = Sequence<Is...>;
using PassThrough = element_wise::PassThrough;
static constexpr auto GemmDefault = GemmSpecialization::Default;
static constexpr auto GemmKPadding = GemmSpecialization::KPadding;
static constexpr auto GemmMPadding = GemmSpecialization::MPadding;
static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding;
static constexpr auto GemmMKPadding = GemmSpecialization::MKPadding;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
template <GemmSpecialization GemmSpec>
using device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_comp_instances = std::tuple<
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | Type| 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| Pipeline| Pipeline|
//#########################| | | | | | | | | 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| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Compute friendly
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 8, 8, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 32, 32, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 32, 32, 4, 4, S<16,16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<16,16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 8, 8, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 8, 8, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
// AGPR Spill
// DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 8, 8, 16, 16, 8, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
// AGPR Spill when use permuted lds layout. so, use padding for these two.
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 64, 8, 8, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 2, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 224, 64, 8, 8, 16, 16, 8, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 2, 1, S<1, 32, 1, 8>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 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, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 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, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 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, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
// clang-format on
>;
template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
using device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_mem_instances = std::tuple<
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | Type| 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| Pipeline| Pipeline|
//#########################| | | | | | | | | 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| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Latency friendly
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
// Memory friendly
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 8, 16, 16, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 4, 4, 16, 16, 4, 1, S<16,16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 0, 1, 1, S<1, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, S<32, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 8, 16, 16, 4, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 8, 8, 16, 16, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 8, 8, 16, 16, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle_Streamk_V3< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 8, 8, 16, 16, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_comp_default_instance.cpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_comp_default_instances(
std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
Col,
Row,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_comp_instances<GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_comp_kpadding_instance.cpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
Col,
Row,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_comp_instances<GemmKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_mem_v1_default_instance.cpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
Col,
Row,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_mem_instances<Intrawave,
GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_mem_v1_kpadding_instance.cpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
Col,
Row,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_mem_instances<Intrawave,
GemmKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_mem_v2_default_instance.cpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
Col,
Row,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_streamk_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_streamk/device_gemm_xdl_universal_streamk_bf16_bf16_bf16/device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_mem_v2_kpadding_instance.cpp 0 → 100755
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
Col,
Row,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_streamk_bf16_bf16_bf16_mk_nk_mn_mem_instances<Interwave,
GemmKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/include/profiler/profile_gemm_b_scale_impl.hpp 0 → 100644
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include <iostream>
#include <typeinfo>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_b_scale.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_b_scale.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/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename BScaleDataType,
typename ComputeDataType,
typename AccDataType,
typename CDataType,
index_t ScaleBlockK,
typename ALayout,
typename BLayout,
typename CLayout>
bool profile_gemm_b_scale_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int KBatch,
int n_warmup,
int n_iter,
uint64_t rotating = 0)
{
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});
}
};
ck::index_t Scale_Stride_BN = ck::is_same_v<BLayout, ck::tensor_layout::gemm::ColumnMajor>
? ((K + ScaleBlockK - 1) / ScaleBlockK)
: N;
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<BDataType> b_k_n_permute(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<BScaleDataType> b1_k_n(f_host_tensor_descriptor(
(K + ScaleBlockK - 1) / ScaleBlockK, // K direction group size is ScaleBlockK
N, // N direction group size is 1
Scale_Stride_BN,
BLayout{}));
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{}));
int total_gemm_needed = a_m_k.GetElementSpaceSizeInBytes() +
b_k_n.GetElementSpaceSizeInBytes() +
b1_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 << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "b1_k_n: " << b1_k_n.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)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-1, 2});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
}
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{};
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n_permute.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(BScaleDataType) * b1_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b1_device_buf.ToDevice(b1_k_n.mData.data());
using DeviceOp = ck::tensor_operation::device::DeviceGemmV2BScale<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
BScaleDataType,
CDataType,
1,
ScaleBlockK,
AElementOp,
BElementOp,
CElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// Run reference GEMM
if(do_verification)
{
Tensor<float> b_k_n_dequant({K, N});
float v_b = 0;
for(int n = 0; n < N; n++)
{
for(int k = 0; k < K; k++)
{
ck::pk_i4_t i4x2 = b_k_n(k, n).data;
int8_t i4 = 0;
if(k % 2 == 1)
i4 = (i4x2.data >> 0) & 0xf;
else
i4 = (i4x2.data >> 4) & 0xf;
i4 = i4 - 8;
v_b = ck::type_convert<float>(i4);
b_k_n_dequant(k, n) = ck::type_convert<float>(v_b) *
ck::type_convert<float>(b1_k_n(k / ScaleBlockK, n));
}
}
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp,
ComputeDataType>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n_dequant, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
float best_kbatch = 0;
// profile device GEMM instances
for(auto& op_ptr : op_ptrs)
{
const int KPerBlock = op_ptr->GetKPerBlock();
if(op_ptr->GetPermuteB())
{
int K1 = KPerBlock;
int K0 = K / KPerBlock;
// int K0, N, K1
for(int j = 0; j < K0; j++)
{
for(int i = 0; i < N; i++)
{
for(int jj = 0; jj < K1; jj++)
{
b_k_n_permute(j * N * K1 + i * K1 + jj) = b_k_n(i * K + (j * K1 + jj));
}
}
}
if(is_same_v<BDataType, pk_i4_t> && is_same_v<ADataType, half_t>)
{
// vector pk_i4x4 permute
for(int i = 0; i < N; i++)
{
for(int j = 0; j < K; j += 8)
{
int input[8];
for(int k = 0; k < 4; k++)
{
int i4x2 = b_k_n_permute(j + k * 2, i).data;
input[k * 2 + 0] = (i4x2 >> 4) & 0xf;
input[k * 2 + 1] = (i4x2 >> 0) & 0xf;
}
// permute 01234567->20643175
{
int hi = input[2];
int lo = input[0];
int i4x2 = (hi << 4) | lo;
b_k_n_permute(j + 0, i) = i4x2;
}
{
int hi = input[6];
int lo = input[4];
int i4x2 = (hi << 4) | lo;
b_k_n_permute(j + 2, i) = i4x2;
}
{
int hi = input[3];
int lo = input[1];
int i4x2 = (hi << 4) | lo;
b_k_n_permute(j + 4, i) = i4x2;
}
{
int hi = input[7];
int lo = input[5];
int i4x2 = (hi << 4) | lo;
b_k_n_permute(j + 6, i) = i4x2;
}
}
}
}
}
else
{
b_k_n_permute = b_k_n;
}
b_device_buf.ToDevice(b_k_n_permute.mData.data());
std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38};
if(KBatch > 0)
{
kbatch_list = {KBatch};
}
for(std::size_t i = 0; i < kbatch_list.size(); i++)
{
auto kbatch_curr = kbatch_list[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
Scale_Stride_BN,
static_cast<BScaleDataType*>(b1_device_buf.GetDeviceBuffer()),
kbatch_curr,
a_element_op,
b_element_op,
c_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init C to zero before profiling next kernel
c_device_buf.SetZero();
invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr, false, 0, n_warmup, n_iter});
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
#if defined CK_ENABLE_FP8
// set softer tolerances for fp8
if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
is_same_v<CDataType, f8_t>)
{
std::string msg = "Error: Incorrect results!";
double rtol = 1e-1;
double atol = 1e-1;
pass = pass & ck::utils::check_err(
c_m_n_device_result, c_m_n_host_result, msg, rtol, atol);
}
else
{
#endif
pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
#if defined CK_ENABLE_FP8
}
#endif
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<int8_t>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
}
}
std::string op_name = op_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(),
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;
static constexpr index_t BPackedSize = []() {
if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
return 2;
else
return 1;
}();
std::size_t num_btype = sizeof(ADataType) * M * K +
sizeof(BDataType) * K * N / BPackedSize +
sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
<< " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
<< kbatch_curr << std::endl;
if(tflops > best_tflops && ave_time > 1e-10)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
best_kbatch = kbatch_curr;
}
}
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem"
<< std::endl;
}
}
}
if constexpr(is_same<CDataType, float>::value)
{
std::cout << "Best Perf for datatype = f32";
}
else if constexpr(is_same<CDataType, half_t>::value)
{
std::cout << "Best Perf for datatype = f16";
}
else if constexpr(is_same<CDataType, bhalf_t>::value)
{
std::cout << "Best Perf for datatype = bf16";
}
else if constexpr(is_same<CDataType, int8_t>::value)
{
std::cout << "Best Perf for datatype = int8";
}
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " ALayout = RowMajor";
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " ALayout = ColumnMajor";
}
if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " BLayout = RowMajor";
}
else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " BLayout = ColumnMajor";
}
std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
<< " StrideB = " << StrideB << " StrideC = " << StrideC << " KBatch = " << best_kbatch
<< " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
<< " GB/s, " << best_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_gemm_universal_impl.hpp
View file @ 67ab3896
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2023-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
...@@ -177,7 +177,7 @@ bool profile_gemm_universal_impl(int do_verification, ...@@ -177,7 +177,7 @@ bool profile_gemm_universal_impl(int do_verification,
} }
} }
if(is_same_v<BDataType, pk_i4_t> && is_same_v<ADataType, half_t>) if constexpr(is_same_v<BDataType, pk_i4_t> && is_same_v<ADataType, half_t>)
{ {
// vector pk_i4x4 permute // vector pk_i4x4 permute
for(int i = 0; i < N; i++) for(int i = 0; i < N; i++)
...@@ -188,7 +188,7 @@ bool profile_gemm_universal_impl(int do_verification, ...@@ -188,7 +188,7 @@ bool profile_gemm_universal_impl(int do_verification,
for(int k = 0; k < 4; k++) for(int k = 0; k < 4; k++)
{ {
int i4x2 = b_k_n_permute(j + k * 2, i); int i4x2 = b_k_n_permute(j + k * 2, i).data;
input[k * 2 + 0] = (i4x2 >> 4) & 0xf; input[k * 2 + 0] = (i4x2 >> 4) & 0xf;
input[k * 2 + 1] = (i4x2 >> 0) & 0xf; input[k * 2 + 1] = (i4x2 >> 0) & 0xf;
} }
......
profiler/src/CMakeLists.txt
View file @ 67ab3896
...@@ -58,6 +58,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") ...@@ -58,6 +58,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp) list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp) list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp) list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_b_scale.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp) list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp) list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp) list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
...@@ -141,6 +142,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") ...@@ -141,6 +142,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
endif() endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_b_scale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
......
profiler/src/profile_gemm_b_scale.cpp 0 → 100644
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <initializer_list>
#include <iostream>
#include <numeric>
#include "profiler/profile_gemm_b_scale_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
F8_F16_F16, // 4
F16_F8_F16, // 5
F16_F16_F16_F8, // 6
F8_F8_BF16, // 7
F16_I4_F16, // 8
};
enum struct BScaleBlockTile
{
K_64, // 0
K_128, // 1
};
#define OP_NAME "gemm_b_scale"
#define OP_DESC "Int4-dequant GEMM"
int profile_gemm_b_scale(int argc, char* argv[])
{
if(argc != 16 && argc != 19)
{
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->f8; 7: f8->bf16, "
"comp f8; 8: f16@i4)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: B scale block tile (0: 64, 1: 128):\n");
printf("arg5: verification (0: no; 1: yes)\n");
printf("arg6: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg7: print tensor value (0: no; 1: yes)\n");
printf("arg8: time kernel (0=no, 1=yes)\n");
printf("arg9 to 14: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg15: split k into mulitiple batch\n");
printf("optional:\n");
printf("arg16: number of warm-up cycles (default 1)\n");
printf("arg17: number of iterations (default 10)\n");
printf("arg18: memory for rotating buffer (default 0, size in MB)\n");
exit(1);
}
printf("Start profiling\n");
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const auto B_scale_block = static_cast<BScaleBlockTile>(std::stoi(argv[4]));
const bool do_verification = std::stoi(argv[5]);
const int init_method = std::stoi(argv[6]);
const bool do_log = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[8]);
const int M = std::stoi(argv[9]);
const int N = std::stoi(argv[10]);
const int K = std::stoi(argv[11]);
const int StrideA = std::stoi(argv[12]);
const int StrideB = std::stoi(argv[13]);
const int StrideC = std::stoi(argv[14]);
const int KBatch = std::stoi(argv[15]);
printf("M:%d, N:%d, K:%d, StrideA:%d, StrideB:%d, StrideC:%d, KBatch:%d\n",
M,
N,
K,
StrideA,
StrideB,
StrideC,
KBatch);
int n_warmup = 1;
int n_iter = 10;
uint64_t rotating = 0;
if(argc == 19)
{
n_warmup = std::stoi(argv[16]);
n_iter = std::stoi(argv[17]);
rotating = std::stoull(argv[18]) * 1024 * 1024;
printf("n_warmup:%d, n_iter:%d, rotating:%lu\n", n_warmup, n_iter, rotating);
}
using F32 = float;
using F16 = ck::half_t;
using I4 = ck::pk_i4_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto b_scale_type,
auto comp_type,
auto acc_type,
auto c_type,
auto scale_block_k,
auto a_layout,
auto b_layout,
auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using BScaleDataType = decltype(b_scale_type);
using ComputeDataType = decltype(comp_type);
using AccDataType = decltype(acc_type);
using CDataType = decltype(c_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CLayout = decltype(c_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_b_scale_impl<ADataType,
BDataType,
BScaleDataType,
ComputeDataType,
AccDataType,
CDataType,
scale_block_k,
ALayout,
BLayout,
CLayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideC < 0) ? DefaultStrideC : StrideC,
KBatch,
n_warmup,
n_iter,
rotating);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::F16_I4_F16 && layout == GemmMatrixLayout::MK_NK_MN &&
B_scale_block == BScaleBlockTile::K_128)
{
printf("F16_I4_F16 MK_NK_MN K_128\n");
return profile(
F16{}, I4{}, F16{}, F16{}, F32{}, F16{}, ck::Number<128>{}, Row{}, Col{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_b_scale);
profiler/src/profile_gemm_universal_streamk.cpp 100755 → 100644
View file @ 67ab3896
...@@ -83,8 +83,9 @@ int profile_gemm_universal_streamk(int argc, char* argv[]) ...@@ -83,8 +83,9 @@ int profile_gemm_universal_streamk(int argc, char* argv[])
rotating = std::stoull(argv[18]) * 1024 * 1024; rotating = std::stoull(argv[18]) * 1024 * 1024;
} }
using F32 = float; using F32 = float;
using F16 = ck::half_t; using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94) #if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
using F8 = ck::f8_t; using F8 = ck::f8_t;
...@@ -165,6 +166,22 @@ int profile_gemm_universal_streamk(int argc, char* argv[]) ...@@ -165,6 +166,22 @@ int profile_gemm_universal_streamk(int argc, char* argv[])
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{}); return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
} }
#endif #endif
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 if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
{
return profile(BF16{}, BF16{}, F32{}, BF16{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
{
return profile(BF16{}, BF16{}, F32{}, BF16{}, Col{}, 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;
......
script/cmake-ck-dev.sh
View file @ 67ab3896
...@@ -15,7 +15,7 @@ else ...@@ -15,7 +15,7 @@ else
fi fi
cmake \ cmake \
-D CMAKE_PREFIX_PATH=/opt/rocm \ -D CMAKE_PREFIX_PATH=/opt/rocm/ \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \ -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_CXX_FLAGS="-Xclang -mllvm -Xclang -enable-post-misched=0 -std=c++17 -O3 -ftemplate-backtrace-limit=0 -fPIE -Wno-gnu-line-marker" \ -D CMAKE_CXX_FLAGS="-Xclang -mllvm -Xclang -enable-post-misched=0 -std=c++17 -O3 -ftemplate-backtrace-limit=0 -fPIE -Wno-gnu-line-marker" \
-D CMAKE_BUILD_TYPE=Release \ -D CMAKE_BUILD_TYPE=Release \
......
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