Skip to content

GitLab

Commit 563c1e23 authored by Astha Rai's avatar Astha Rai
Browse files

Merge branch 'codegen_hiprtc' of github.com:ROCm/composable_kernel into codegen_hiprtc

parents fb1bc1bf b14b1973
Hide whitespace changes
Inline Side-by-side
Showing with 746 additions and 996 deletions
library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
View file @ 563c1e23
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, 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_xdl_splitk_cshuffle.hpp"
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_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 Empty_Tuple = 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_splitk_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| 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|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 192, 64, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 48, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 192, 32, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 256, 32, 8, 8, 32, 32, 1, 4, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 64, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 32, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
// clang-format on
>;
void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Col,
......@@ -72,7 +23,7 @@ void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances
PassThrough>>>& instances)
{
add_device_operation_instances(
instances, device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_tile_instances{});
instances, device_grouped_gemm_xdl_splitk_2Bt_rcr_instances<F16, GemmMNKPadding>{});
}
} // namespace instance
......
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 deleted 100644 → 0
View file @ fb1bc1bf
// 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 GemmDefault = GemmSpecialization::Default;
static constexpr auto GemmKPadding = GemmSpecialization::KPadding;
static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
template <typename DsLayout,
typename DsDataType,
typename CDEElementwiseOp,
GemmSpecialization GemmSpec = GemmMNKPadding>
using device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_comp_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|
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | S<C,D0...,D_N|
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 1, 256, 256, 256, 32, 8, 4, 32, 32, 4, 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>, S<8,8,1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<8,8,1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 1, 256, 256, 256, 32, 8, 4, 32, 32, 4, 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>, S<8,8,1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 1, 256, 256, 256, 32, 8, 4, 32, 32, 4, 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>, S<8,8,1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 1, 256, 224, 256, 64, 8, 4, 16, 16, 7, 8, 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, 2, S<1, 32, 1, 8>, S<8,8,1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<8,8,1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<8,8,1>, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<8,8,1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
// clang-format on
>;
template <typename DsLayout,
typename DsDataType,
typename CDEElementwiseOp,
GemmSpecialization GemmSpec = GemmMNKPadding,
BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave>
using device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_mem_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|
//###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | S<C,D0...,D_N|
// Latency friendly
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<4,4,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<4,4,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
// Memory friendly
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<4,4,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<4,4,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<4,4,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<8,8,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<4,4,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<8,8,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<4,4,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
// DeviceGroupedGemmMultipleDXdlCShuffleTileLoop< Row, Row, DsLayout, Row, BF16, I8, F32, F32, DsDataType, BF16, PassThrough, PassThrough, CDEElementwiseOp, GemmSpec, 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>, S<8,8,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
// 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)
{
// comp
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_comp_instances<ck::Tuple<Row>,
ck::Tuple<BF16>,
Multiply,
GemmDefault>{});
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_comp_instances<ck::Tuple<Row>,
ck::Tuple<BF16>,
Multiply,
GemmMNKPadding>{});
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_comp_instances<ck::Tuple<Row>,
ck::Tuple<BF16>,
Multiply,
GemmMNPadding>{});
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_comp_instances<ck::Tuple<Row>,
ck::Tuple<BF16>,
Multiply,
GemmKPadding>{});
// mem
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_mem_instances<ck::Tuple<Row>,
ck::Tuple<BF16>,
Multiply,
GemmDefault,
Intrawave>{});
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_mem_instances<ck::Tuple<Row>,
ck::Tuple<BF16>,
Multiply,
GemmMNKPadding,
Intrawave>{});
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_mem_instances<ck::Tuple<Row>,
ck::Tuple<BF16>,
Multiply,
GemmMNPadding,
Intrawave>{});
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_mem_instances<ck::Tuple<Row>,
ck::Tuple<BF16>,
Multiply,
GemmKPadding,
Intrawave>{});
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_mem_instances<ck::Tuple<Row>,
ck::Tuple<BF16>,
Multiply,
GemmDefault,
Interwave>{});
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_mem_instances<ck::Tuple<Row>,
ck::Tuple<BF16>,
Multiply,
GemmMNKPadding,
Interwave>{});
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_mem_instances<ck::Tuple<Row>,
ck::Tuple<BF16>,
Multiply,
GemmMNPadding,
Interwave>{});
add_device_operation_instances(
instances,
device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_mem_instances<ck::Tuple<Row>,
ck::Tuple<BF16>,
Multiply,
GemmKPadding,
Interwave>{});
}
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
modified_files.txt 0 → 100755
View file @ 563c1e23
example/01_gemm/gemm_xdl_fp8_streamk_v3.cpp
example/01_gemm/run_gemm_example_streamk_v2.inc
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_streamk_v3.hpp
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f8_f16/device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f8_f16/device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_comp_mnpadding_instance.cpp
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f8_f16_f16/device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f8_f16_f16/device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_comp_mnpadding_instance.cpp
profiler/src/profile_gemm_universal_streamk.cpp
modified_files.txt
profiler/include/profiler/profile_gemm_universal_batched_impl.hpp 0 → 100644
View file @ 563c1e23
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <memory>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.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_batched_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementOp,
typename BElementOp,
typename CElementOp,
typename DeviceOp>
bool profile_gemm_universal_batched_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int BatchStrideA,
int BatchStrideB,
int BatchStrideC,
int StrideA,
int StrideB,
int StrideC,
int BatchCount,
int n_warmup,
int n_iter,
uint64_t rotating = 0)
{
bool pass = true;
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
}
else
{
return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
}
};
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
Tensor<BDataType> b_g_k_n(
f_host_tensor_descriptor(BatchCount, K, N, StrideB, BatchStrideB, BLayout{}));
Tensor<CDataType> c_g_m_n_host_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, BatchStrideC, CLayout{}));
Tensor<CDataType> c_g_m_n_device_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, BatchStrideC, CLayout{}));
int total_gemm_needed =
a_g_m_k.GetElementSpaceSizeInBytes() + b_g_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_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
std::cout << "rotating count: " << rotating_count << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
}
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
if(do_verification)
{
using ReferenceBatchedGemmInstance =
ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
BDataType,
CDataType,
float,
AElementOp,
BElementOp,
CElementOp>;
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
auto ref_argument = ref_batched_gemm.MakeArgument(
a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
c_device_buf.ToDevice(c_g_m_n_device_result.mData.data());
// 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;
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device op instances
for(auto& op_ptr : op_ptrs)
{
std::unique_ptr<tensor_operation::device::BaseArgument> argument_ptr;
// false branch for multi d dl kernel
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,
BatchCount,
StrideA,
StrideB,
{},
StrideC,
BatchStrideA,
BatchStrideB,
{},
BatchStrideC,
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{});
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();
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, true, rotating_count});
std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N) *
BatchCount;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
pass = pass & ck::utils::check_err(c_g_m_n_device_result, c_g_m_n_host_result);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_g_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host: ", c_g_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
<< std::endl;
}
}
}
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 << " B = " << BatchCount << " M = " << M << " N = " << N << " K = " << K
<< " StrideA = " << StrideA << " StrideB = " << StrideB << " StrideC = " << StrideC
<< ": " << 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_grouped_gemm_impl.hpp
View file @ 563c1e23
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -17,7 +17,6 @@
#include "ck/library/utility/convolution_parameter.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/utility/fill.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
......@@ -42,11 +41,14 @@ bool profile_grouped_gemm_impl(int do_verification,
const std::vector<int>& StrideAs,
const std::vector<int>& StrideBs,
const std::vector<int>& StrideCs,
int kbatch = 1,
int n_warmup = 1,
int n_iter = 10)
const std::vector<int>& kbatches = {},
int n_warmup = 1,
int n_iter = 10)
{
bool pass = true;
// TODO: Fixme - we do not pass compute data type here but need it
// to compute error thresholds.
using ComputeDataType = ADataType;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
......@@ -75,6 +77,7 @@ bool profile_grouped_gemm_impl(int do_verification,
std::vector<Tensor<CDataType>> c_m_n_host_results;
std::vector<Tensor<CDataType>> c_m_n_device_results;
ComputeDataType max_abs_in_val = 0.f;
for(std::size_t i = 0; i < group_count; i++)
{
a_m_k.push_back(
......@@ -93,17 +96,18 @@ bool profile_grouped_gemm_impl(int do_verification,
<< i << "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
<< "]:" << c_m_n_device_results[i].mDesc << std::endl;
}
std::size_t num_thread = 1;
switch(init_method)
{
case 0: break;
case 1:
a_m_k[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-2.f, 2.f}(a_m_k[i]);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-2.f, 2.f}(b_k_n[i]);
max_abs_in_val = 2.f;
break;
default:
a_m_k[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_k_n[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
ck::utils::FillUniformDistribution<ADataType>{-0.5f, 0.5f}(a_m_k[i]);
ck::utils::FillUniformDistribution<BDataType>{-0.5f, 0.5f}(b_k_n[i]);
max_abs_in_val = 0.5f;
}
}
......@@ -164,7 +168,20 @@ bool profile_grouped_gemm_impl(int do_verification,
BElementOp,
CElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
// If kbatch would be bigger than 1, then we will use SplitK version.
using DeviceOpSplitK = ck::tensor_operation::device::DeviceGroupedGemmSplitK<ALayout,
BLayout,
ck::Tuple<>,
CLayout,
ADataType,
BDataType,
ck::Tuple<>,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
if(op_ptrs.size() <= 0)
......@@ -205,7 +222,6 @@ bool profile_grouped_gemm_impl(int do_verification,
ref_invoker.Run(ref_argument);
}
}
// profile device GEMM instances
for(auto& gemm_ptr : op_ptrs)
{
......@@ -221,43 +237,44 @@ bool profile_grouped_gemm_impl(int do_verification,
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
std::size_t workspace_size = gemm_ptr->GetWorkSpaceSize(argument_ptr.get());
std::size_t kargs_size = gemm_ptr->GetDeviceKernelArgSize(argument_ptr.get());
gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
std::string gemm_name = gemm_ptr->GetTypeString();
DeviceMem gemm_workspace, gemm_kargs;
using DeviceOpSplitK = ck::tensor_operation::device::DeviceGroupedGemmSplitK<ALayout,
BLayout,
ck::Tuple<>,
CLayout,
ADataType,
BDataType,
ck::Tuple<>,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
// skip non-splitk grouped_gemm
if(dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get()) == nullptr)
// The following is necessary since TwoStage kernel is using additional memory both
// for Workspace and kernel arguments.
if(kargs_size > 0)
{
continue;
gemm_kargs.Realloc(kargs_size);
gemm_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kargs.GetDeviceBuffer());
}
if(workspace_size > 0 && workspace_size != kargs_size)
{
gemm_workspace.Realloc(workspace_size);
gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_workspace.GetDeviceBuffer());
}
std::string gemm_name = gemm_ptr->GetTypeString();
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 48, 64};
if(kbatch > 0)
// If the user will provide not empty kbatches list, then we test predefined set of kbatch
// values.
if(!kbatches.empty())
{
kbatch_list = {kbatch};
kbatch_list = kbatches;
}
for(std::size_t j = 0; j < kbatch_list.size(); j++)
{
auto kbatch_curr = kbatch_list[j];
dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())
->SetKBatchSize(argument_ptr.get(), kbatch_curr);
if(kbatch_curr > 1 && dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get()) != nullptr)
{
dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())
->SetKBatchSize(argument_ptr.get(), kbatch_curr);
}
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
......@@ -272,23 +289,18 @@ bool profile_grouped_gemm_impl(int 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());
if(std::is_same_v<CDataType, ck::half_t> && kbatch_curr > 1)
{
instance_pass =
instance_pass && ck::utils::check_err(c_m_n_device_results[i],
c_m_n_host_results[i],
"Error: Incorrect results!",
0.06);
}
else
{
instance_pass =
instance_pass && ck::utils::check_err(c_m_n_device_results[i],
c_m_n_host_results[i]);
}
auto atol = ck::utils::get_absolute_threshold<ComputeDataType, CDataType>(
max_abs_in_val, gemm_descs[i].K_);
auto rtol = ck::utils::get_relative_threshold<ComputeDataType, CDataType>(
gemm_descs[i].K_);
instance_pass =
instance_pass && ck::utils::check_err(c_m_n_device_results[i],
c_m_n_host_results[i],
"Error: Incorrect results!",
rtol,
atol);
if(do_log)
{
......@@ -311,11 +323,12 @@ bool profile_grouped_gemm_impl(int do_verification,
pass = pass && instance_pass;
}
float ave_time = invoker_ptr->Run(
argument_ptr.get(), StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter});
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++)
{
......
profiler/include/profiler/profile_grouped_gemm_multiply_tile_loop_impl.hpp
View file @ 563c1e23
......@@ -143,8 +143,7 @@ bool profile_grouped_gemm_multiply_tile_loop_impl(int do_verification,
p_ds.reserve(group_count);
p_e.reserve(group_count);
using KernelArguments =
ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDTensor>;
using KernelArguments = ck::tensor_operation::device::GroupedGemmKernelArgument<NumDTensor>;
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
std::vector<KernelArguments> gemm_kargs;
......
profiler/include/profiler/profile_grouped_gemm_tile_loop_impl.hpp
View file @ 563c1e23
......@@ -127,7 +127,7 @@ bool profile_grouped_gemm_tile_loop_impl(int do_verification,
p_b.reserve(group_count);
p_c.reserve(group_count);
using KernelArguments = ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<>;
using KernelArguments = ck::tensor_operation::device::GroupedGemmKernelArgument<>;
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
std::vector<KernelArguments> gemm_kargs;
......
profiler/include/profiler/profile_grouped_gemm_two_stage_impl.hpp deleted 100644 → 0
View file @ fb1bc1bf
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_multiple_d_splitk.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_parameter.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/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_two_stage_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 kbatch = 1,
int n_warmup = 1,
int n_iter = 10)
{
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(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
{
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;
}
std::size_t num_thread = 1;
switch(init_method)
{
case 0: break;
case 1:
a_m_k[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
break;
default:
a_m_k[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_k_n[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
}
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);
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
gemm_descs.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());
gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
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());
}
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemm<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;
float best_kbatch = 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();
DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
std::string gemm_name = gemm_ptr->GetTypeString();
using DeviceOpSplitK =
ck::tensor_operation::device::DeviceGroupedGemmMultipleDSplitK<ALayout,
BLayout,
ck::Tuple<>,
CLayout,
ADataType,
BDataType,
ck::Tuple<>,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
// skip non-splitk grouped_gemm
if(dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get()) == nullptr)
{
continue;
}
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 48, 64};
if(kbatch > 0)
{
kbatch_list = {kbatch};
}
for(std::size_t j = 0; j < kbatch_list.size(); j++)
{
auto kbatch_curr = kbatch_list[j];
dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())
->SetKBatchSize(argument_ptr.get(), kbatch_curr);
DeviceMem gemm_arg_dev_mem(dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())
->GetDeviceKernelArgSize(argument_ptr.get()));
dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())
->SetDeviceKernelArgs(argument_ptr.get(), gemm_arg_dev_mem.GetDeviceBuffer());
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
gemm_desc_workspace.SetZero();
for(std::size_t i = 0; i < gemm_descs.size(); i++)
c_device_buf[i]->SetZero();
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());
if(std::is_same_v<CDataType, ck::half_t> && kbatch_curr > 1)
{
instance_pass =
instance_pass && ck::utils::check_err(c_m_n_device_results[i],
c_m_n_host_results[i],
"Error: Incorrect results!",
0.06);
}
else
{
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;
}
float ave_time = invoker_ptr->Run(
argument_ptr.get(), StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter});
if(time_kernel)
{
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 << ", KBatch "
<< kbatch_curr << 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;
best_kbatch = kbatch_curr;
}
}
}
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 << ", KBatch = " << best_kbatch
<< std::endl;
}
return pass;
}
} // namespace profiler
} // namespace ck
profiler/src/CMakeLists.txt
View file @ 563c1e23
......@@ -43,7 +43,6 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.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_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
......@@ -59,6 +58,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal.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_streamk.cpp)
list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
......@@ -141,6 +141,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance)
target_link_libraries(${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_add_multiply_instance)
......
profiler/src/profile_gemm_universal.cpp
View file @ 563c1e23
......@@ -101,7 +101,7 @@ int profile_gemm_universal(int argc, char* argv[])
using F32 = float;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
using F8 = ck::f8_t;
#endif
......@@ -164,7 +164,7 @@ int profile_gemm_universal(int argc, char* argv[])
{
return profile(F16{}, F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
}
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F16{}, F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
......@@ -198,7 +198,7 @@ int profile_gemm_universal(int argc, char* argv[])
{
return profile(BF16{}, BF16{}, BF16{}, F32{}, BF16{}, Col{}, Row{}, Row{});
}
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
else if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F8{}, F8{}, F8{}, F32{}, BF16{}, Row{}, Row{}, Row{});
......
profiler/src/profile_gemm_universal_batched.cpp 0 → 100644
View file @ 563c1e23
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdint>
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_gemm_universal_batched_impl.hpp"
#include "profiler_operation_registry.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_universal_batched.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
enum struct GemmDataType
{
BF16_BF16_BF16, // 0
F8_F8_BF16, // 1
};
#define OP_NAME "gemm_universal_batched"
#define OP_DESC "Batched GEMM Universal"
int profile_batched_gemm_universal(int argc, char* argv[])
{
if(argc != 18 && argc != 21)
{
// clang-format off
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: bf16, 1: fp8->bf16)\n");
printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n");
printf(" 1: A[g, m, k] * B[g, n, k] = C[g, m, n];\n");
printf(" 2: A[g, k, m] * B[g, k, n] = C[g, m, n];\n");
printf(" 3: A[g, k, m] * B[g, n, k] = C[g, m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount\n");
printf("optional:\n");
printf("arg18: number of warm-up cycles (default 1)\n");
printf("arg19: number of iterations (default 10)\n");
printf("arg20: memory for rotating buffer (default 0, size in MB)\n");
// clang-format on
exit(1);
}
int n_warmup = 1;
int n_iter = 10;
uint64_t rotating = 0;
if(argc == 21)
{
n_warmup = std::stoi(argv[18]);
n_iter = std::stoi(argv[19]);
rotating = std::stoull(argv[20]) * 1024 * 1024;
}
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 int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
const int BatchStrideA = std::stoi(argv[14]);
const int BatchStrideB = std::stoi(argv[15]);
const int BatchStrideC = std::stoi(argv[16]);
const int BatchCount = std::stoi(argv[17]);
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
using F8 = ck::f8_t;
#endif
using BF16 = ck::bhalf_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile =
[&](auto a_type, auto b_type, auto c_type, auto a_layout, auto b_layout, auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using DsDataType = ck::Tuple<>;
using CDataType = decltype(c_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using DsLayout = ck::Tuple<>;
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;
const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA;
const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB;
const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC;
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
const int DefaultBatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA;
const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB;
const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemmV2MultiD<ALayout,
BLayout,
DsLayout,
CLayout,
ADataType,
BDataType,
DsDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
bool pass = ck::profiler::profile_gemm_universal_batched_impl<ADataType,
BDataType,
CDataType,
ALayout,
BLayout,
CLayout,
AElementOp,
BElementOp,
CElementOp,
DeviceOp>(do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
BatchStrideA_,
BatchStrideB_,
BatchStrideC_,
StrideA_,
StrideB_,
StrideC_,
BatchCount,
n_warmup,
n_iter,
rotating);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(BF16{}, BF16{}, BF16{}, Row{}, Col{}, Row{});
}
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
else if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F8{}, F8{}, BF16{}, Row{}, Col{}, Row{});
}
#endif
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_universal);
profiler/src/profile_gemm_universal_streamk.cpp 100644 → 100755
View file @ 563c1e23
......@@ -85,8 +85,10 @@ int profile_gemm_universal_streamk(int argc, char* argv[])
using F32 = float;
using F16 = ck::half_t;
// using BF16 = ck::bhalf_t;
// using F8 = ck::f8_t;
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
using F8 = ck::f8_t;
#endif
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -145,6 +147,24 @@ int profile_gemm_universal_streamk(int argc, char* argv[])
{
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
}
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
}
#endif
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
......
profiler/src/profile_grouped_gemm.cpp
View file @ 563c1e23
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
......@@ -39,16 +39,13 @@ 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;
}
......@@ -69,7 +66,7 @@ int profile_grouped_gemm(int argc, char* argv[])
<< "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"
<< "arg15: kbatch value (default 1)\n"
<< "arg15: kbatch values (default 1)\n"
<< "optional:\n"
<< "arg16: number of warm-up cycles (default 1)\n"
<< "arg17: number of iterations (default 10)\n"
......@@ -92,7 +89,7 @@ int profile_grouped_gemm(int argc, char* argv[])
const auto StrideAs = argToIntArray(argv[11]);
const auto StrideBs = argToIntArray(argv[12]);
const auto StrideCs = argToIntArray(argv[13]);
const int kbatch = argc == 15 ? std::stoi(argv[14]) : 1;
const auto kbatches = argc >= 15 ? argToIntArray(argv[14]) : std::vector<int>{};
int n_warmup = 1;
int n_iter = 10;
......@@ -102,7 +99,6 @@ int profile_grouped_gemm(int argc, char* argv[])
n_iter = std::stoi(argv[16]);
}
#ifdef CK_ENABLE_FP16
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::half_t,
......@@ -121,7 +117,7 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch,
kbatches,
n_warmup,
n_iter);
}
......@@ -143,7 +139,7 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch,
kbatches,
n_warmup,
n_iter);
}
......@@ -165,7 +161,7 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch,
kbatches,
n_warmup,
n_iter);
}
......@@ -187,7 +183,7 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch,
kbatches,
n_warmup,
n_iter);
}
......@@ -209,7 +205,7 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch,
kbatches,
n_warmup,
n_iter);
}
......@@ -231,7 +227,73 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch,
kbatches,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_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,
kbatches,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_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,
kbatches,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
time_kernel,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs,
kbatches,
n_warmup,
n_iter);
}
......@@ -239,7 +301,6 @@ int profile_grouped_gemm(int argc, char* argv[])
{
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
#endif
return 0;
}
......
profiler/src/profile_grouped_gemm_fixed_nk.cpp
View file @ 563c1e23
......@@ -32,9 +32,7 @@ namespace {
std::vector<int> argToIntArray(char* input)
{
std::vector<int> out;
std::istringstream in(input);
std::string item;
while(std::getline(in, item, ','))
......@@ -83,7 +81,7 @@ int profile_grouped_gemm_fixed_nk(int argc, char* argv[])
const auto StrideAs = argToIntArray(argv[11]);
const auto StrideBs = argToIntArray(argv[12]);
const auto StrideCs = argToIntArray(argv[13]);
const int kbatch = argc == 15 ? std::stoi(argv[14]) : 1;
const int kbatch = argc >= 15 ? std::stoi(argv[14]) : 1;
using F32 = float;
using F16 = ck::half_t;
......@@ -97,8 +95,8 @@ int profile_grouped_gemm_fixed_nk(int argc, char* argv[])
int n_iter = 10;
if(argc == 17)
{
n_warmup = std::stoi(argv[16]);
n_iter = std::stoi(argv[17]);
n_warmup = std::stoi(argv[15]);
n_iter = std::stoi(argv[16]);
}
#if defined(CK_ENABLE_BF16) && defined(CK_ENABLE_INT8)
......
profiler/src/profile_grouped_gemm_two_stage.cpp deleted 100644 → 0
View file @ fb1bc1bf
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_grouped_gemm_two_stage_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
};
enum struct GemmDataType
{
F16_F16_F16, // 0
BF16_INT8_BF16, // 1
BF16_BF16_BF16 // 2
};
#define OP_NAME "grouped_gemm_two_stage"
#define OP_DESC "Grouped GEMM TwoStage"
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_two_stage(int argc, char* argv[])
{
if(argc < 14)
{
std::cout
<< "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp16; 1: bf16@int8; 2: bf16)\n"
<< "arg3: matrix layout (0: A[m, k] * B[k, n] = 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"
<< "arg15: kbatch value (default 1)\n"
<< "optional:\n"
<< "arg16: number of warm-up cycles (default 1)\n"
<< "arg17: 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 kbatch = argc == 15 ? std::stoi(argv[14]) : 1;
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 = 1;
int n_iter = 10;
if(argc == 17)
{
n_warmup = std::stoi(argv[16]);
n_iter = std::stoi(argv[17]);
}
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_two_stage_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,
kbatch,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::BF16_INT8_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_two_stage_impl<ck::bhalf_t,
int8_t,
ck::bhalf_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,
kbatch,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::BF16_INT8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_grouped_gemm_two_stage_impl<ck::bhalf_t,
int8_t,
ck::bhalf_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,
kbatch,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_two_stage_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_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,
kbatch,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_grouped_gemm_two_stage_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_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,
kbatch,
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_two_stage);
test/ck_tile/gemm/test_gemm_mem_pipeline.cpp
View file @ 563c1e23
......@@ -11,8 +11,20 @@
using F16 = ck_tile::half_t;
using F32 = float;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
static constexpr auto Intrawave = ck_tile::GemmPipelineScheduler::Intrawave;
static constexpr auto Interwave = ck_tile::GemmPipelineScheduler::Interwave;
template <typename Tuple>
class TestCkTileGemmMemPipelineIntrawave : public TestCkTileGemmMemPipeline<Tuple, Intrawave>
{
};
template <typename Tuple>
class TestCkTileGemmMemPipelineInterwave : public TestCkTileGemmMemPipeline<Tuple, Interwave>
{
};
// clang-format off
using KernelTypes = ::testing::Types<
......@@ -24,6 +36,7 @@ using KernelTypes = ::testing::Types<
>;
// clang-format on
TYPED_TEST_SUITE(TestCkTileGemmMemPipeline, KernelTypes);
TYPED_TEST_SUITE(TestCkTileGemmMemPipelineIntrawave, KernelTypes);
TYPED_TEST_SUITE(TestCkTileGemmMemPipelineInterwave, KernelTypes);
#include "test_gemm_mem_pipeline_ut_cases.inc"
test/ck_tile/gemm/test_gemm_mem_pipeline_ut_cases.inc
View file @ 563c1e23
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
TYPED_TEST(TestCkTileGemmMemPipeline, SmallM)
//------------------------------------------------------------------------------------------------
// INTERWAVE SCHEDULER
//------------------------------------------------------------------------------------------------
TYPED_TEST(TestCkTileGemmMemPipelineInterwave, SmallM)
{
std::vector<int> Ms{1, 2, 3, 4, 5, 6};
constexpr int N = 1024;
constexpr int K = 320;
for(int M : Ms)
this->Run(M, N, K);
}
TYPED_TEST(TestCkTileGemmMemPipelineInterwave, MidLargeM)
{
std::vector<int> Ms{127, 255, 312, 799, 1573};
constexpr int N = 1024;
constexpr int K = 320;
for(int M : Ms)
this->Run(M, N, K);
}
TYPED_TEST(TestCkTileGemmMemPipelineInterwave, PaddK)
{
std::vector<int> Ms{127};
constexpr int N = 1024;
constexpr int K = 432;
for(int M : Ms)
this->Run(M, N, K);
}
TYPED_TEST(TestCkTileGemmMemPipelineInterwave, Regular)
{
std::vector<int> Ms{512};
constexpr int N = 1024;
constexpr int K = 512;
for(int M : Ms)
this->Run(M, N, K);
}
//------------------------------------------------------------------------------------------------
// INTRAWAVE SCHEDULER
//------------------------------------------------------------------------------------------------
TYPED_TEST(TestCkTileGemmMemPipelineIntrawave, SmallM)
{
std::vector<int> Ms{1, 2, 3, 4, 5, 6};
constexpr int N = 1024;
......@@ -10,7 +61,7 @@ TYPED_TEST(TestCkTileGemmMemPipeline, SmallM)
this->Run(M, N, K);
}
TYPED_TEST(TestCkTileGemmMemPipeline, MidLargeM)
TYPED_TEST(TestCkTileGemmMemPipelineIntrawave, MidLargeM)
{
std::vector<int> Ms{127, 255, 312, 799, 1573};
constexpr int N = 1024;
......@@ -20,7 +71,7 @@ TYPED_TEST(TestCkTileGemmMemPipeline, MidLargeM)
this->Run(M, N, K);
}
TYPED_TEST(TestCkTileGemmMemPipeline, PaddK)
TYPED_TEST(TestCkTileGemmMemPipelineIntrawave, PaddK)
{
std::vector<int> Ms{127};
constexpr int N = 1024;
......@@ -30,7 +81,7 @@ TYPED_TEST(TestCkTileGemmMemPipeline, PaddK)
this->Run(M, N, K);
}
TYPED_TEST(TestCkTileGemmMemPipeline, Regular)
TYPED_TEST(TestCkTileGemmMemPipelineIntrawave, Regular)
{
std::vector<int> Ms{512};
constexpr int N = 1024;
......
test/ck_tile/gemm/test_gemm_mem_pipeline_util.hpp
View file @ 563c1e23
......@@ -11,20 +11,21 @@
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/gemm.hpp"
template <typename Tuple>
template <typename Tuple, ck_tile::GemmPipelineScheduler Scheduler_>
class TestCkTileGemmMemPipeline : public ::testing::Test
{
protected:
using ALayout = std::tuple_element_t<0, Tuple>;
using BLayout = std::tuple_element_t<1, Tuple>;
using CLayout = std::tuple_element_t<2, Tuple>;
using ADataType = std::tuple_element_t<3, Tuple>;
using BDataType = std::tuple_element_t<4, Tuple>;
using AccDataType = std::tuple_element_t<5, Tuple>;
using CDataType = std::tuple_element_t<6, Tuple>;
using ALayout = std::tuple_element_t<0, Tuple>;
using BLayout = std::tuple_element_t<1, Tuple>;
using CLayout = std::tuple_element_t<2, Tuple>;
using ADataType = std::tuple_element_t<3, Tuple>;
using BDataType = std::tuple_element_t<4, Tuple>;
using AccDataType = std::tuple_element_t<5, Tuple>;
using CDataType = std::tuple_element_t<6, Tuple>;
static constexpr auto Scheduler = Scheduler_;
// TODO: expose tile size through test t-param ?
struct gemm_basic_args
struct gemm_args
{
const void* p_a;
const void* p_b;
......@@ -38,7 +39,7 @@ class TestCkTileGemmMemPipeline : public ::testing::Test
ck_tile::index_t stride_C;
};
void invoke_gemm(const gemm_basic_args& args, const ck_tile::stream_config& s)
void invoke_gemm(const gemm_args& args, const ck_tile::stream_config& s)
{
// TODO: This should be parameterized in tests
constexpr ck_tile::index_t M_Tile = 128;
......@@ -53,9 +54,9 @@ class TestCkTileGemmMemPipeline : public ::testing::Test
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 8;
constexpr bool kPadA = true;
constexpr bool kPadB = true;
constexpr bool kPadC = true;
constexpr bool kPadM = true;
constexpr bool kPadN = true;
constexpr bool kPadK = true;
constexpr int kBlockPerCu = 1;
......@@ -68,9 +69,9 @@ class TestCkTileGemmMemPipeline : public ::testing::Test
using TilePartitioner = ck_tile::GemmTilePartitioner<GemmShape>;
using GemmEpilogue = ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, false, kPadC>>;
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>;
using Traits = ck_tile::TileGemmTraits<kPadA, kPadB, kPadC, ALayout, BLayout, CLayout>;
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using BaseGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrMem<
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>;
......@@ -89,7 +90,7 @@ class TestCkTileGemmMemPipeline : public ::testing::Test
AccDataType,
GemmShape,
Traits,
ck_tile::GemmPipelineScheduler::Intrawave,
Scheduler,
has_hot_loop_v,
tail_number_v>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
......@@ -108,7 +109,7 @@ class TestCkTileGemmMemPipeline : public ::testing::Test
if(s.log_level_ > 0)
{
std::cout << "Lunching kernel with args:"
std::cout << "Launching kernel with args:"
<< " grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z
<< "}" << std::endl;
......@@ -288,7 +289,7 @@ class TestCkTileGemmMemPipeline : public ::testing::Test
c_m_n_dev_buf.SetZero();
c_m_n_dev_result.SetZero();
gemm_basic_args args;
gemm_args args;
args.p_a = a_m_k_dev_buf.GetDeviceBuffer();
args.p_b = b_k_n_dev_buf.GetDeviceBuffer();
args.p_c = c_m_n_dev_buf.GetDeviceBuffer();
......
test/gemm_universal/test_gemm_universal_xdl.cpp
View file @ 563c1e23
......@@ -56,7 +56,7 @@ class TestGemmUniversal_KM_NK
using KernelTypes_MK_KN = ::testing::Types<
// ADataType, BDataType, ComputeDataType, CDataType
std::tuple< F16, F16, F16, F16>,
#if defined(CK_ENABLE_FP8) && defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
#if defined(CK_ENABLE_FP8) && (defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94))
std::tuple< F16, F8, F16, F16>,
std::tuple< F8, F16, F16, F16>,
std::tuple< F8, F8, F8, BF16>,
......@@ -66,7 +66,7 @@ using KernelTypes_MK_KN = ::testing::Types<
using KernelTypes_MK_NK = ::testing::Types<
// ADataType, BDataType, ComputeDataType, CDataType
std::tuple< F16, F16, F16, F16>,
#if defined(CK_ENABLE_FP8) && defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
#if defined(CK_ENABLE_FP8) && (defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94))
std::tuple< F16, F8, F16, F16>,
std::tuple< F8, F16, F16, F16>,
std::tuple< F8, F8, F8, BF16>,
......
test/grouped_gemm/CMakeLists.txt
View file @ 563c1e23
......@@ -6,12 +6,6 @@ if(result EQUAL 0)
add_dependencies(test_grouped_gemm test_grouped_gemm_splitk)
endif()
add_gtest_executable(test_grouped_gemm_two_stage_splitk test_grouped_gemm_two_stage_multiple_d_splitk_xdl.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_gemm_two_stage_splitk PRIVATE utility device_grouped_gemm_instance)
add_dependencies(test_grouped_gemm test_grouped_gemm_two_stage_splitk)
endif()
add_gtest_executable(test_grouped_gemm_interface test_grouped_gemm_interface_xdl.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_gemm_interface PRIVATE utility device_grouped_gemm_instance)
......
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